WO2020073737A1 - 一种基于深度学习的定量光谱数据分析处理方法 - Google Patents
一种基于深度学习的定量光谱数据分析处理方法 Download PDFInfo
- Publication number
- WO2020073737A1 WO2020073737A1 PCT/CN2019/101573 CN2019101573W WO2020073737A1 WO 2020073737 A1 WO2020073737 A1 WO 2020073737A1 CN 2019101573 W CN2019101573 W CN 2019101573W WO 2020073737 A1 WO2020073737 A1 WO 2020073737A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- convolution
- layer
- model
- convolutional
- convolution kernel
- Prior art date
- Legal status (The legal status 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 status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/70—Denoising; Smoothing
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
- G01N21/00—Investigating or analysing materials by the use of optical means, i.e. using sub-millimetre waves, infrared, visible or ultraviolet light
- G01N21/17—Systems in which incident light is modified in accordance with the properties of the material investigated
- G01N21/25—Colour; Spectral properties, i.e. comparison of effect of material on the light at two or more different wavelengths or wavelength bands
- G01N21/31—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry
- G01N21/35—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light
- G01N21/359—Investigating relative effect of material at wavelengths characteristic of specific elements or molecules, e.g. atomic absorption spectrometry using infrared light using near infrared light
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/09—Supervised learning
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/0985—Hyperparameter optimisation; Meta-learning; Learning-to-learn
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Definitions
- the invention belongs to the field of spectrum analysis, and in particular relates to a method for analyzing and processing quantitative spectrum data based on deep learning.
- chemometrics has promoted the application of spectral analysis in the fields of agricultural products, pharmaceuticals, petroleum, and soil. It has been widely used in the qualitative and quantitative analysis of infrared spectroscopy and Raman spectroscopy.
- the traditional chemometric data analysis process includes two steps: spectral preprocessing and establishing a calibration model. Spectral preprocessing is mainly used to remove noise in spectral data and improve the prediction accuracy of the model. On the one hand, spectral preprocessing mainly has four steps: baseline correction, scattering correction, smoothing and normalization. Each step has a different data processing method. Choosing a combination of preprocessing methods through trial and error will increase the modeling process The complexity is more time-consuming.
- Deep learning is a data-driven learning method.
- the model can automatically learn the low-dimensional features and high-dimensional features contained in the data from the original data.
- the traditional artificial neural network often needs to use principal component analysis and other methods for dimensionality reduction when performing spectral data analysis, and the artificial neural network is prone to overfitting due to the large number of parameters.
- the convolutional neural network has the characteristics of local connection and weight sharing, which can fully extract local features in the data and prevent overfitting.
- Existing convolutional neural network models still require spectral preprocessing, or just as a feature extraction method. Acquarelli et al. Proposed a one-layer convolutional neural network qualitative analysis model, but the model still performs better on the preprocessed spectral data (J.
- the present invention provides a method for analyzing and processing quantitative spectral data based on deep learning.
- the method of the present invention is a data-driven modeling method that does not require data preprocessing and can extract features of different sizes through convolution kernels of different sizes without outputting background noise from the original spectral data and output prediction results To improve the accuracy of predictions.
- the technical solution adopted by the present invention is:
- Step 1) Construct a one-dimensional convolutional neural network model, and optimize the calculation to obtain the hyperparameters of the model;
- Step 2) Use the spectral data with known predicted values of the sample to input into the convolutional neural network model, adopt Adam optimization method combined with back propagation method to obtain the weight of the model, and obtain the optimal model after multiple rounds of training Later model
- Step 3) Input the spectral data of the unknown predicted value of the sample into the trained model, and output the result of obtaining the predicted value of the spectral data.
- spectral data are pre-processed to remove background information, and then use the effective information to establish a correction model using partial least squares (PLS), artificial neural network (ANN) and other methods.
- PLS partial least squares
- ANN artificial neural network
- the invention establishes a convolutional neural network model with a special structure, directly processes the complete original spectrum data without removing background information, and obtains a very good detection accuracy.
- the samples of the present invention include soil, animal feed and cereals.
- One spectral curve corresponds to one soil or one animal feed sample or one cereal sample.
- step 1) specifically:
- the convolutional neural network model is mainly composed of the input layer, the convolutional layer 1, the convolutional layer 2, the convolutional layer 3, the stretching layer, the fully connected layer, and the output layer connected in sequence; the input layer Enter the original full-band spectral curve;
- the first convolution layer contains a convolution module, using 8 convolution kernels, all convolution kernels are the same size;
- the second convolutional layer uses three parallel modules with two convolutional modules and one pooling module.
- the output of the first convolutional layer is input into two convolutional modules and one pooling module, each volume
- the convolution module uses a convolution kernel.
- the convolution kernels of the two convolution modules are different.
- Each convolution module contains four 1 ⁇ 1 ⁇ 8 convolution kernels.
- the pooling module contains four parallel maximum pools. Structural
- the third convolution layer uses four convolution modules.
- the four convolution modules use four different convolution kernels respectively.
- the first convolution module contains four 1 ⁇ 1 ⁇ 8 first convolution kernels.
- the second convolution module contains four p ⁇ 1 ⁇ 4 second convolution kernels, the third convolution module contains four q ⁇ 1 ⁇ 4 third convolution kernels, and the fourth convolution kernel Contains four 1 ⁇ 1 ⁇ 4 fourth-type convolution kernels, p and q represent the lengths of the second-type convolution kernel and the third-type convolution kernel, respectively, the output of the first-type convolution kernel is input to the third
- the first convolution module of a convolution layer, the two convolution modules of a second convolution layer and a pooling module are input into the last three convolution modules of the third convolution layer; the stretch layer Perform the operation of stretching the output of the third convolutional layer into a one-dimensional feature vector;
- the objective function loss of the convolutional neural network model is composed of the mean square error and the second norm regularization function:
- the regularization coefficient of the ⁇ objective function w is the weight of the model
- the random grid search method is used to optimize the size and step size of the convolution kernel of the convolution layer, including the size and step size of the convolution kernel of the first convolution layer, and the two convolution kernels of the third convolution layer Size and step size, the convolution kernel size and step size of the second convolutional layer are fixed values; specifically, the random grid search method is used to search the hyperparameters in the convolutional layer in the following hyperparameter search space, using Folding and cross-validation chooses to obtain a set of optimal hyperparameter combinations composed of hyperparameters;
- the size and step range of the different convolution kernels in the above three convolutional layers are as follows: the size range of the convolution kernel in the first convolutional layer is 2-19, and the range of the convolution kernel step is 2-9; The length of the first type of convolution kernel in the second convolutional layer is set to 1, the step size of the first type of convolution kernel is set to 1, and the length of the second type of convolution kernel in the second layer of convolution layer is set to 1.
- the second convolution kernel step is set to 1; the first convolution kernel length in the third convolution layer is set to 1, and the second convolution kernel length in the third convolution layer is The p size range is 2-5, the third convolution kernel length q in the second convolution layer is 6-9, and the four convolution kernel step sizes in the third convolution layer are 2-9.
- the activation function of the three convolutional layers and fully connected layers of the model is the Leaky ReLU function, and the output layer of the model does not have an activation function.
- the number of neurons in the last layer of the output layer is 1.
- step 2) the spectral data of the known reference value of the sample is input into the convolutional neural network model, and the Adam method is combined with the back propagation algorithm to obtain the weight of the model, and the number of training rounds is 5000 rounds.
- step 3 the predicted value corresponding to each spectral curve is predicted by the model output.
- the predicted value may be, for example, the organic carbon content of the soil, the protein content of the animal feed, the protein content of the cereal sample, or the like.
- all data is divided into a training set and a test set, the training set is used for model training, the trained model is saved, and the original spectral data of the test set is input into the trained model for prediction.
- output test set R 2 and RMSEP For model effect evaluation, output test set R 2 and RMSEP.
- the invention does not need to preprocess the data, can learn the effective spectrum information and background information hidden in the data, and improve the accuracy of quantitative spectrum analysis.
- the present invention extracts high-dimensional features in the spectral data through three convolutional layers, and adopts three convolution kernels of different sizes in the third convolutional layer, which can learn from the original spectral data hidden in the spectral data Features of different sizes.
- the present invention does not preprocess the data, and can directly process the original data. When the spectral noise distribution of the test set spectrum and the training set are different, the generalization ability of the present invention is high.
- Figure 1 is a modeling flowchart of the present invention
- Figure 2 shows the structure of the quantitative analysis model.
- the selected data set is a public soil data set.
- the soil samples are from the United States, Africa, Asia, South America, and Europe. Only one sample is collected from a single soil body, and the spectral data of 3793 soil bodies is divided into 2502 samples in the training set and 1291 samples in the test set.
- the minimum organic carbon content of all soil samples is 0, the maximum organic carbon content is 241.6g kg -1 , the average organic carbon content is 11.97g kg -1 , and the standard deviation of the organic carbon content is 20.87g kg -1 .
- FIG 2 Optimize the hyperparameters of the model.
- the structure of the model is shown in Figure 2, which includes: three convolutional layers, named convolutional layer 1, convolutional layer 2 and convolutional layer 3, stretch layer, full Connection layer and output layer.
- convolutional layer 1 In the first convolution layer of the model, 8 convolution kernels are used.
- the optimal convolution kernel size is 9 and the step size is 5.
- the third convolution layer uses four convolution modules, using 16 convolution kernels, which are four different sizes of convolution kernels, including a 1 ⁇ 1 ⁇ 8 convolution kernel, the second and third
- the size of the convolution kernel is 5 and 7, respectively, and the step size of the four convolution kernels is 3.
- the number of neurons in the fully connected layer is 64.
- the number of neurons in the output layer is 1 ( Figure 2).
- the traditional PLS-LDA and PCA-ANN methods are used to predict the original soil spectrum before and without pretreatment, and the best result is the prediction of PCA-ANN in the original spectrum without pretreatment.
- the RMSEP is 11.59 and R 2 is 0.72, and the original spectrum without pretreatment is predicted by the method proposed by the present invention, and the predicted RMSEP is 8.88 and R 2 is 0.84. It can be seen from the comparison that the prediction accuracy of the method on the original spectrum without data preprocessing is higher than that of traditional modeling methods with and without preprocessing, which can improve the accuracy of near infrared spectroscopy on soil sample prediction. .
- the samples of the test set and the samples of the training set come from different regions, and the prediction accuracy of the test set is still high, which reflects the better generalization ability of the present invention.
- FIG 2 Optimize the hyperparameters of the model.
- the structure of the model is shown in Figure 2, which includes: three convolutional layers, named convolutional layer 1, convolutional layer 2 and convolutional layer 3, stretch layer, full Connection layer and output layer.
- convolutional layer 1 In the first convolution layer of the model, 8 convolution kernels are used.
- the optimal convolution kernel size is 9 and the step size is 5.
- the third convolution layer uses four convolution modules, using 16 convolution kernels, which are four different sizes of convolution kernels, including a 1 ⁇ 1 ⁇ 8 convolution kernel, the second and third The size of the convolution kernel is 5 and 7, and the step size of the four convolution kernels is 3.
- the number of neurons in the fully connected layer is 64.
- the number of neurons in the output layer is 1 ( Figure 2).
- the method proposed by the present invention is used to predict the pre-processed original spectrum.
- the predicted RMSEP is 12.92 and R 2 is 0.65. It can be seen from the comparison that the prediction accuracy of the method of the present invention on the spectrum after data preprocessing is worse than that on the original spectrum.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Spectroscopy & Molecular Physics (AREA)
- Chemical & Material Sciences (AREA)
- Biochemistry (AREA)
- Analytical Chemistry (AREA)
- Pathology (AREA)
- Immunology (AREA)
- Investigating Or Analysing Materials By Optical Means (AREA)
Abstract
本发明公开了一种基于深度学习的定量光谱数据分析处理方法。本发明不需要对数据进行预处理,能够从原始光谱数据中学习到有效信息和背景信息,提高定量光谱分析的准确度。本发明通过三个卷积层提取光谱数据中的高维特征,在第二层采用1×1的卷积核,能够降维和减少计算量,并且在第三个卷积层采用三种不同大小的卷积核,能够从原始光谱数据中学习到隐含在光谱数据中的不同大小的特征。本发明没有对数据做预处理,可以直接处理原始数据,当测试集光谱与训练集的光谱噪声分布不同时,本发明的泛化能力较高。
Description
本发明属于光谱分析领域,具体涉及了一种基于深度学习的定量光谱数据分析处理方法。
化学计量学的发展促进了光谱分析在农产品、药品、石油和土壤等领域的应用,目前已经被广泛应用于红外光谱和拉曼光谱的定性与定量分析中。传统的化学计量学数据分析过程包括光谱预处理和建立校正模型两个步骤。光谱预处理主要用于去除光谱数据中的噪声,提高模型的预测精度。一方面,光谱预处理主要有基线校正、散射校正、平滑和归一化等四个步骤,每个步骤又有不同的数据处理方法,通过试错法选择预处理方法的组合会增加建模过程的复杂度,耗费更多时间。另一方面,光谱数据的采集环境、采集仪器或者样本来源发生变化时,数据中的噪声分布也会随之发生变化。原有的预处理方法应用于新的数据时不能有效去除噪声并且会引入新的噪声,致使模型的预测效果变差。
深度学习是一种数据驱动的学习方式,模型可以从原始数据中自动学习到数据中蕴藏的低维特征和高维特征。传统的人工神经网络在进行光谱数据分析时,往往需要先用主成分分析等方法进行降维,而且人工神经网络由于参数量较多容易出现过拟合现象。卷积神经网络具有局部连接、权值共享等特点可以充分提取数据中的局部特征并且防止过拟合。现有的卷积神经网络模型仍然需要进行光谱预处理,或者只是作为一种特征提取方法。Acquarelli等人提出了一种一层的卷积神经网络定性分析模型,但是该模型仍然是在经过预处理后的光谱数据上效果较好(J.Acquarelli.,T.v.,Laarhoven,J.,Gerretzen,T.N.,Tran,L.M.C.,Buydens,E.,Marchiori,Convolutional Neural Networks for Vibrational Spectroscopic Data Analysis,2017)。Malek等人提出了一种卷积神经网络的定量分析模型,但是在该模型中卷积神经网络被用于特征提取,提取的特征后在回归模型中进行训练(S.,Malek,F.,Melgani,Y.,Bazi,One-dimensional convolutional neural networks for spectroscopic signal regression,2017)。
发明内容
为了弥补现有的化学计量学建模方法的不足,本发明提出了一种基于深度学习的定量光谱数据分析处理方法。本发明方法是一种数据驱动的建模方法,不需要数据预处理,可以在不从原始光谱数据中去除背景噪声的情况下,通过 不同大小的卷积核提取不同大小的特征,输出预测结果,提高预测的准确度。
如图1所示,本发明所采用的技术方案是:
步骤1):构建一维卷积神经网络模型,并且优化计算获得模型的超参数;
步骤2):用样品已知预测值的光谱数据输入到卷积神经网络模型中,采用Adam优化方法结合反向传播方法训练获得模型的权重,经过多轮训练后得到一个最优模型,获得训练后的模型;
步骤3):将样品未知预测值的光谱数据输入到训练后的模型,输出获得光谱数据的预测值结果。
现有的光谱数据均是进行数据预处理,去除背景信息,然后利用有效信息采用偏最小二乘(PLS)、人工神经网络(ANN)等方法建立校正模型。
而本发明建立特殊结构的卷积神经网络模型,直接对未去除背景信息的完整的原始光谱数据进行处理,获得了很好的检测精度。
本发明样本是包括土壤、动物饲料和谷物等。一条光谱曲线对应一份土壤或者一份动物饲料样品或者一份谷物样品。
所述的步骤1)中,具体为:
1.1)如图2所示,卷积神经网络模型主要由输入层、卷积层1、卷积层2、卷积层3、拉伸层、全连接层和输出层依次连接构成;输入层中输入原始的全波段光谱曲线;
第一个卷积层包含一个卷积模块,使用8个卷积核,所有卷积核大小相同;
第二个卷积层采用两个卷积模块和一个池化模块的三个并列的模块,第一个卷积层的输出分别输入到两个卷积模块和一个池化模块中,每个卷积模块使用一种卷积核,两个卷积模块的卷积核不同,每个卷积模块均含有4个1×1×8的卷积核,池化模块中含有4个并列的最大池化结构;
第三个卷积层采用四个卷积模块,四个卷积模块分别使用四种不同的卷积核,第一个卷积模块包含四个1×1×8的第一种卷积核,第二个卷积模块包含四个p×1×4的第二种卷积核,第三个卷积模块包含四个q×1×4的第三种卷积核,第四个卷积模块包含四个1×1×4的第四种卷积核,p和q分别表示第二种卷积核和第三种卷积核的长度,其中第一种卷积核的输出输入到第三个卷积层的第一个卷积模块,第二个卷积层的两个卷积模块和一个池化模块分别输入到第三个卷积层的后三个卷积模块中;拉伸层进行将第三个卷积层的输出拉伸成一维特征向量的操作;
所述卷积神经网络模型的目标函数loss由均方误差和第二范数正则化函数组成:
其中,λ目标函数的正则化系数,w是模型的权重;
1.2)采用随机网格搜索方法优化卷积层的卷积核大小和步长,包括第一层卷积层的卷积核的大小和步长,第三层卷积层的两个卷积核的大小和步长,第二层卷积层的卷积核大小和步长为固定值;具体是采用随机网格搜索方法在以下超参数搜索空间内搜索卷积层中的超参数,采用五折交叉验证选择获得一组由超参数构成的最优超参数组合;
上述三个卷积层中不同卷积核的大小和步长的范围如下:第一层卷积层中的卷积核大小范围是2-19,卷积核步长范围是2-9;第二层卷积层中的第一种卷积核长度设定为1,第一种卷积核步长设定为1,第二层卷积层中的第二种卷积核长度设定为1,第二种卷积核步长设定为1;第三层卷积层中的第一种卷积核长度设定为1,第三层卷积层中的第二种卷积核长度p大小范围是2-5,第二层卷积层中的第三种卷积核长度q大小范围是6-9,第三个卷积层中四种卷积核步长范围2-9。
模型的三个卷积层和全连接层的激活函数为Leaky ReLU函数,模型的输出层不具有激活函数。
输出层的最后一层神经元个数为1。
所述步骤2)中,将样品已知参考值的光谱数据输入到卷积神经网络模型中采用Adam方法结合反向传播算方法训练获得模型的权重,训练轮数为5000轮。
所述步骤3)中,通过模型输出预测每一条光谱曲线对应的预测值。
预测值例如可以为土壤的有机碳含量,动物饲料的蛋白质含量、谷物样品的蛋白质含量等。
具体实施中,将所有数据分为训练集和测试集,用训练集进行模型训练,将训练好的模型保存,并将测试集的原始光谱数据输入到训练好的模型中进行预测。对于模型效果评价,输出测试集R
2和RMSEP。
本发明的有益效果是:
本发明不需要对数据进行预处理,能够学习到隐含在数据中的有效光谱信息和背景信息,提高定量光谱分析的准确度。本发明通过三个卷积层提取光谱数据中的高维特征,并且在第三个卷积层采用三种不同大小的卷积核,能够从原始光谱数据中学习到隐含在光谱数据中的不同大小的特征。本发明没有对数据做预处理,可以直接处理原始数据,当测试集光谱与训练集的光谱噪声分布 不同时,本发明的泛化能力较高。
图1为本发明的建模流程图;
图2为模型的定量分析结构图。
为更好理解本发明,下面结合实施例对本发明做进一步详细说明,但本发明要求保护的范围并不局限于实施例表示的范围。以下进行的实施例,在Python软件上运行。下面结合附图和实施例对本发明做进一步说明。
按照上述本发明完整方法实施的具体实施例如下:
本实施例应用于近红外光谱的定量分析,预测土壤的有机碳含量。所选取的数据集为土壤的公开数据集。土壤样本来源于美国、非洲、亚洲和南美洲和欧洲等地区,单个土体中只采集一个样本,共有3793个土体的光谱数据,分为训练集样本2502个和测试集样本1291个。使用FieldSpec Pro-FR spectrometer光谱仪采集光谱数据,光谱数据的波段范围是350到2500nm。所有土壤样本的最小有机碳含量是0,最大有机碳含量是241.6g kg
-1,平均有机碳含量是11.97g kg
-1,有机碳含量的标准差是20.87g kg
-1。
如图1所示步骤,上述实施例过程具体如下:
1)将训练集数据输入到模型中。
2)优化模型的超参数,模型的结构如图2所示,具体包括:包含三个卷积层,分别命名为卷积层1、卷积层2和卷积层3,拉伸层、全连接层和输出层。模型的第一个卷积层中使用了8个卷积核,最优卷积核大小为9,步长为5。第三个卷积层采用四个卷积模块,使用了16个卷积核,分别是四种不同大小的卷积核,包含一个1×1×8的卷积核,第二种和第三种卷积核大小分别为5和7,四种卷积核的步长都为3。全连接层的神经元个数是64。输出层中神经元的个数为1(图2)。
3)模型的超参数固定后,再次用所有的训练集数据来训练模型的参数。通过反向传播算法在训练集上训练模型的权重值。模型训练5000轮。
4)模型训练好以后,终止训练,保存最优的模型,并在测试集上测试模型的效果。输出每一条待测光谱的预测值和模型在测试集上的R
2和RMSEP。
5)利用传统的PLS-LDA和PCA-ANN方法对经过预处理和没有经过预处理的原始土壤光谱分别进行预测,得到最好的结果是PCA-ANN在没有经过预处理的原始光谱的预测,其RMSEP是11.59,R
2是0.72,而通过本发明提出的方法对没有经过预处理的原始光谱进行预测,其预测RMSEP是8.88,R
2是0.84。 通过比较可以看出,该方法在没有经过数据预处理的原始光谱上的预测精度高于传统的经过预处理和没有经过预处理的建模方法,可以提高近红外光谱对土壤样品预测的准确率。测试集样本和训练集样本来自于不同的地区,测试集的预测精度仍然较高,体现了本发明较好的泛化能力。
对比例:
对比该模型在经过预处理的数据上的结果,过程如下:
1)将训练集数据输入到模型中。
2)优化模型的超参数,模型的结构如图2所示,具体包括:包含三个卷积层,分别命名为卷积层1、卷积层2和卷积层3,拉伸层、全连接层和输出层。模型的第一个卷积层中使用了8个卷积核,最优卷积核大小为9,步长为5。第三个卷积层采用四个卷积模块,使用了16个卷积核,分别是四种不同大小的卷积核,包含一个1×1×8的卷积核,第二种和第三种卷积核大小分别为5和7,四种卷积核的步长是3。全连接层的神经元个数是64。输出层中神经元的个数为1(图2)。
3)模型的超参数固定后,再次用所有的训练集数据来训练模型的参数。通过反向传播算法在训练集上训练模型的权重值。模型训练5000轮。
4)模型训练好以后,终止训练,保存最优的模型,并在测试集上测试模型的效果。输出每一条待测光谱的预测值和模型在测试集上的R
2和RMSEP。
5)通过本发明提出的方法对经过预处理的原始光谱进行预测,其预测RMSEP是12.92,R
2是0.65。通过比较可以看出,本发明方法在经过数据预处理的光谱上的预测精度比在原始光谱的预测精度差。
Claims (5)
- 一种基于深度学习的定量光谱数据分析处理方法,其特征在于方法包含如下步骤:步骤1):构建一维卷积神经网络模型,并且优化计算获得模型的超参数;步骤2):用样品已知预测值的光谱数据输入到卷积神经网络模型中,训练获得模型的权重,经过多轮训练后得到一个最优模型,获得训练后的模型;步骤3):将样品未知预测值的光谱数据输入到训练后的模型,输出获得光谱数据的预测值结果。
- 根据权利要求1所述的一种基于深度学习的定量光谱数据分析处理方法,其特征在于:所述的步骤1)中,具体为:1.1)卷积神经网络模型主要由输入层、卷积层(1)、卷积层(2)、卷积层(3)、拉伸层、全连接层和输出层依次连接构成;输入层中输入原始的全波段光谱曲线;第一个卷积层包含一个卷积模块,使用8个卷积核,所有卷积核大小相同;第二个卷积层采用两个卷积模块和一个池化模块的三个并列的模块,第一个卷积层的输出分别输入到两个卷积模块和一个池化模块中,每个卷积模块使用一种卷积核,两个卷积模块的卷积核不同,每个卷积模块均含有4个1×1×8的卷积核,池化模块中含有4个并列的最大池化结构;第三个卷积层采用四个卷积模块,四个卷积模块分别使用四种不同的卷积核,第一个卷积模块包含四个1×1×8的第一种卷积核,第二个卷积模块包含四个p×1×4的第二种卷积核,第三个卷积模块包含四个q×1×4的第三种卷积核,第四个卷积模块包含四个1×1×4的第四种卷积核,p和q分别表示第二种卷积核和第三种卷积核的长度,其中第一种卷积核的输出输入到第三个卷积层的第一个卷积模块,第二个卷积层的两个卷积模块和一个池化模块分别输入到第三个卷积层的后三个卷积模块中;拉伸层进行将第三个卷积层的输出拉伸成一维特征向量的操作;所述卷积神经网络模型的目标函数loss由均方误差和第二范数正则化函数组成:其中,λ目标函数的正则化系数,w是模型的权重;1.2)采用随机网格搜索方法优化卷积层的卷积核大小和步长,包括第一层卷积层的卷积核的大小和步长,第三层卷积层的两个卷积核的大小和步长,第二层卷积层的卷积核大小和步长为固定值;具体是采用随机网格搜索方法在以下超参数搜索空间内搜索卷积层中的超参数,采用五折交叉验证选择获得一组由超参数构成的最优超参数组合;上述三个卷积层中不同卷积核的大小和步长的范围如下:第一层卷积层中的卷积核大小范围是2-19,卷积核步长范围是2-9;第二层卷积层中的第一种卷积核长度设定为1,第一种卷积核步长设定为1,第二层卷积层中的第二种卷积核长度设定为1,第二种卷积核步长设定为1;第三层卷积层中的第一种卷积核长度设定为1,第三层卷积层中的第二种卷积核长度p大小范围是2-5,第二层卷积层中的第三种卷积核长度q大小范围是6-9,第三个卷积层中四种卷积核步长范围2-9。
- 根据权利要求2所述的一种基于深度学习的定量光谱数据分析处理方法,其特征在于:模型的三个卷积层和全连接层的激活函数为Leaky ReLU函数,模型的输出层不具有激活函数。
- 根据权利要求1所述的一种基于深度学习的定量光谱数据分析处理方法,其特征在于:所述步骤2)中,将样品已知参考值的光谱数据输入到卷积神经网络模型中采用Adam方法结合反向传播算方法训练获得模型的权重,训练轮数为5000轮。
- 根据权利要求1所述的一种基于深度学习的定量光谱数据分析处理方法,其特征在于:所述步骤3)中,通过模型输出预测每一条光谱曲线对应的预测值。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US17/274,436 US20220051074A1 (en) | 2018-10-10 | 2019-08-20 | Quantitative spectral data analysis and processing method based on deep learning |
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811179747.2 | 2018-10-10 | ||
| CN201811179747.2A CN109493287B (zh) | 2018-10-10 | 2018-10-10 | 一种基于深度学习的定量光谱数据分析处理方法 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2020073737A1 true WO2020073737A1 (zh) | 2020-04-16 |
Family
ID=65690134
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2019/101573 Ceased WO2020073737A1 (zh) | 2018-10-10 | 2019-08-20 | 一种基于深度学习的定量光谱数据分析处理方法 |
Country Status (3)
| Country | Link |
|---|---|
| US (1) | US20220051074A1 (zh) |
| CN (1) | CN109493287B (zh) |
| WO (1) | WO2020073737A1 (zh) |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111767521A (zh) * | 2020-06-22 | 2020-10-13 | 中国石油化工股份有限公司 | 一种基于卷积神经网络与长短期记忆网络的输油泵滚动轴承状态评估方法 |
| CN111797971A (zh) * | 2020-05-27 | 2020-10-20 | 北京迈格威科技有限公司 | 应用卷积神经网络进行数据处理的方法、装置和电子系统 |
| CN111898469A (zh) * | 2020-07-09 | 2020-11-06 | 西北大学 | 一种基于缩放卷积层的信号时频特征提取方法及装置 |
| CN112986210A (zh) * | 2021-02-10 | 2021-06-18 | 四川大学 | 一种尺度自适应的微生物拉曼光谱检测方法及系统 |
| CN113959974A (zh) * | 2021-09-27 | 2022-01-21 | 桂林电子科技大学 | 一种普适性近红外光谱模型的传递方法 |
| 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 |
| CN114067169A (zh) * | 2021-10-15 | 2022-02-18 | 浙江工业大学 | 一种基于卷积神经网络的拉曼光谱分析方法 |
| CN114397297A (zh) * | 2022-01-19 | 2022-04-26 | 河南中烟工业有限责任公司 | 一种烤烟淀粉含量的快速无损检测方法 |
| CN114675140A (zh) * | 2022-03-09 | 2022-06-28 | 浙江理工大学 | 一种基于物联网的高压电网电晕监测系统 |
| CN114692773A (zh) * | 2022-04-08 | 2022-07-01 | 浙江工业大学 | 基于drs-vgg的端到端深度学习拉曼光谱数据分类方法 |
| CN114878543A (zh) * | 2022-05-27 | 2022-08-09 | 中国人民解放军32181部队 | 一种痕量物质的拉曼光谱探测识别方法及系统 |
| EP4080414A1 (en) * | 2021-04-23 | 2022-10-26 | Samsung Display Co., Ltd. | Systems and methods for identifying manufacturing defects |
| CN116561590A (zh) * | 2023-07-10 | 2023-08-08 | 之江实验室 | 基于深度学习的微纳光纤负载大小和位置预测方法及装置 |
| CN116597278A (zh) * | 2023-05-30 | 2023-08-15 | 黑龙江八一农垦大学 | 一种基于Cycle-GAN的光谱粒度效应降低方法 |
| CN117235512A (zh) * | 2022-06-02 | 2023-12-15 | 中国科学院沈阳自动化研究所 | 一种面向libs的残差学习轻量卷积神经网络定量化方法 |
| CN118209537A (zh) * | 2024-03-25 | 2024-06-18 | 哈尔滨理工大学 | 基于融合光谱与神经网络的变压器油纸绝缘老化诊断方法 |
| CN119903327A (zh) * | 2025-03-31 | 2025-04-29 | 山东大学 | 一种dfb激光器光谱参数的提取方法 |
| EP4449101A4 (en) * | 2021-12-17 | 2025-11-12 | Terramera Inc | SOIL PROPERTY PREDICTION SYSTEMS AND METHODS |
| US20250391514A1 (en) * | 2023-06-21 | 2025-12-25 | Zju-hangzhou Global Scientific And Technological Innovation Center | Method and device for spectral prediction of soil organic carbon based on spectrum-guided ensemble learning |
Families Citing this family (33)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109493287B (zh) * | 2018-10-10 | 2022-03-15 | 浙江大学 | 一种基于深度学习的定量光谱数据分析处理方法 |
| FR3088459B1 (fr) * | 2018-11-13 | 2021-09-17 | Pierre Chapelle | Procede pour l’apprentissage d’une machine automatique apte a determiner des caracteristiques d’elements vegetaux. |
| CN110208211B (zh) * | 2019-07-03 | 2021-10-22 | 南京林业大学 | 一种用于农药残留检测的近红外光谱降噪方法 |
| CN110705372B (zh) * | 2019-09-10 | 2023-05-05 | 中国科学院上海技术物理研究所 | 基于深度学习卷积神经网络的libs多成分定量反演方法 |
| CN112633307A (zh) * | 2019-10-08 | 2021-04-09 | 中强光电股份有限公司 | 用于光谱仪的自动化模型训练装置和自动化模型训练方法 |
| CN110554161B (zh) * | 2019-10-12 | 2021-10-15 | 无锡中科水质环境技术有限公司 | 一种基于鱼类行为信号的水质监测方法 |
| CN110726813B (zh) * | 2019-10-12 | 2021-04-27 | 浙江大学 | 基于双层集成神经网络的电子鼻预测方法 |
| CN111222465B (zh) * | 2019-11-07 | 2023-06-13 | 深圳云天励飞技术股份有限公司 | 基于卷积神经网络的图像分析方法及相关设备 |
| CN112326619A (zh) * | 2020-11-12 | 2021-02-05 | 中国计量大学 | 一种基于双光谱技术的微流控农药残留检测方法 |
| CN112862077B (zh) * | 2021-02-25 | 2022-08-30 | 燕山大学 | 多模光纤结合深度学习网络替代传统光谱仪的系统及方法 |
| CN112883901A (zh) * | 2021-03-12 | 2021-06-01 | 北京睿芯高通量科技有限公司 | 一种基于ResNet的天体光谱分类方法及其系统 |
| CN113903407A (zh) * | 2021-09-09 | 2022-01-07 | 中国科学院自动化研究所 | 成分识别方法、装置、电子设备和存储介质 |
| CN113945537A (zh) * | 2021-09-27 | 2022-01-18 | 桂林电子科技大学 | 一种高准确度近红外光谱定量模型建立方法 |
| CN114218985A (zh) * | 2021-12-09 | 2022-03-22 | 上海理工大学 | 用于肾移植抗排斥反应类型分类的模型系统及其构建方法 |
| CN114997229B (zh) * | 2022-05-31 | 2025-02-18 | 北京航空航天大学 | 一种基于脑电信号的运动想象分类方法和系统 |
| CN114878427B (zh) * | 2022-06-07 | 2025-06-27 | 天津大学 | 基于Trident-net的火焰中碳烟多参数场同时测量的方法和系统 |
| CN115440315B (zh) * | 2022-09-07 | 2025-10-31 | 北京工业大学 | 基于dwt-cars-mc-pls的药品api预测方法 |
| CN115859029B (zh) * | 2022-11-29 | 2023-09-15 | 长沙理工大学 | 基于二维重构的光谱定量分析方法 |
| CN116068062B (zh) * | 2022-12-13 | 2025-10-10 | 广东工业大学 | 一种基于声信号和深度学习的焊接质量在线监测方法及系统 |
| CN116380833A (zh) * | 2022-12-16 | 2023-07-04 | 江门市华讯方舟科技有限公司 | 基于太赫兹光谱的特征融合陈皮检测方法和系统 |
| CN116297311A (zh) * | 2023-01-13 | 2023-06-23 | 昆明理工大学 | 烟叶早期病害识别方法、系统、电子设备及存储介质 |
| CN115979973B (zh) * | 2023-03-20 | 2023-06-16 | 湖南大学 | 一种基于双通道压缩注意力网络的高光谱中药材鉴别方法 |
| CN117007547A (zh) * | 2023-06-25 | 2023-11-07 | 余姚市机器人研究中心 | 基于动态卷积神经网络的苹果机械损伤判别方法及装置 |
| CN117030628A (zh) * | 2023-06-27 | 2023-11-10 | 浙江大学 | 一种中药浸膏质量高光谱智能分析方法 |
| WO2025005556A1 (ko) * | 2023-06-30 | 2025-01-02 | 주식회사 클레온 | 소수샷 평생 학습을 위한 컴퓨팅 장치 |
| CN116663437B (zh) * | 2023-08-02 | 2023-11-21 | 中国人民解放军战略支援部队航天工程大学 | 一种基于深度神经网络的频谱测绘图构建方法和系统 |
| CN117056775B (zh) * | 2023-08-09 | 2025-03-04 | 南京农业大学 | 一种基于特征波段选择的多品种梨糖度协同预测方法 |
| CN117309786A (zh) * | 2023-09-04 | 2023-12-29 | 湖南省吉祥天生物科技有限公司 | 用于有机肥料的检测分析方法及其系统 |
| CN117092050B (zh) * | 2023-09-13 | 2024-02-27 | 佛山科学技术学院 | 基于光谱多模态时序学习的煤泥浮选灰分检测方法及系统 |
| CN118624559B (zh) * | 2024-05-14 | 2025-11-25 | 国家能源投资集团有限责任公司 | 煤炭检测方法及系统 |
| CN118675776B (zh) * | 2024-08-23 | 2025-07-29 | 中国计量大学 | 基于拉曼光谱检测硼酸溶液浓度的方法及系统 |
| CN120104986B (zh) * | 2025-05-07 | 2025-07-15 | 浙江农林大学 | 一种基于自监督学习的libs降噪模型的建立方法 |
| CN122024928B (zh) * | 2026-01-29 | 2026-07-07 | 中国科学院上海技术物理研究所 | 一种基于ema-cnn和加权损失函数的libs反演方法 |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107255647A (zh) * | 2017-05-10 | 2017-10-17 | 中国科学院合肥物质科学研究院 | 基于x射线荧光光谱和深度对抗学习的土壤微量元素含量分析预测方法 |
| CN108388917A (zh) * | 2018-02-26 | 2018-08-10 | 东北大学 | 一种基于改进深度学习模型的高光谱图像分类方法 |
| CN108596213A (zh) * | 2018-04-03 | 2018-09-28 | 中国地质大学(武汉) | 一种基于卷积神经网络的高光谱遥感图像分类方法及系统 |
| CN109493287A (zh) * | 2018-10-10 | 2019-03-19 | 浙江大学 | 一种基于深度学习的定量光谱数据分析处理方法 |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN104102929B (zh) * | 2014-07-25 | 2017-05-03 | 哈尔滨工业大学 | 基于深度学习的高光谱遥感数据分类方法 |
| CN105300895B (zh) * | 2015-11-05 | 2017-12-26 | 浙江大学 | 一种利用特征点切线夹角预警马铃薯发芽缺陷的方法 |
| US10397257B2 (en) * | 2016-12-07 | 2019-08-27 | General Electric Company | Multi-mode boundary selection for threat detection in industrial asset control system |
| EP3340106B1 (en) * | 2016-12-23 | 2023-02-08 | Hexagon Technology Center GmbH | Method and system for assigning particular classes of interest within measurement data |
| CN106706546A (zh) * | 2016-12-28 | 2017-05-24 | 中山市腾创贸易有限公司 | 一种基于红外和拉曼光谱数据的人工智能学习物质分析方法 |
-
2018
- 2018-10-10 CN CN201811179747.2A patent/CN109493287B/zh active Active
-
2019
- 2019-08-20 WO PCT/CN2019/101573 patent/WO2020073737A1/zh not_active Ceased
- 2019-08-20 US US17/274,436 patent/US20220051074A1/en not_active Abandoned
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107255647A (zh) * | 2017-05-10 | 2017-10-17 | 中国科学院合肥物质科学研究院 | 基于x射线荧光光谱和深度对抗学习的土壤微量元素含量分析预测方法 |
| CN108388917A (zh) * | 2018-02-26 | 2018-08-10 | 东北大学 | 一种基于改进深度学习模型的高光谱图像分类方法 |
| CN108596213A (zh) * | 2018-04-03 | 2018-09-28 | 中国地质大学(武汉) | 一种基于卷积神经网络的高光谱遥感图像分类方法及系统 |
| CN109493287A (zh) * | 2018-10-10 | 2019-03-19 | 浙江大学 | 一种基于深度学习的定量光谱数据分析处理方法 |
Non-Patent Citations (1)
| Title |
|---|
| ZHANG, XIAOLEI ET AL.: "DeepSpectra: An end-to-end deep learning approach for quantitative spectral analysis", ANALYTICA CHIMICA ACTA, vol. 1058, 13 June 2019 (2019-06-13), XP055701638 * |
Cited By (25)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111797971A (zh) * | 2020-05-27 | 2020-10-20 | 北京迈格威科技有限公司 | 应用卷积神经网络进行数据处理的方法、装置和电子系统 |
| CN111767521A (zh) * | 2020-06-22 | 2020-10-13 | 中国石油化工股份有限公司 | 一种基于卷积神经网络与长短期记忆网络的输油泵滚动轴承状态评估方法 |
| CN111898469A (zh) * | 2020-07-09 | 2020-11-06 | 西北大学 | 一种基于缩放卷积层的信号时频特征提取方法及装置 |
| CN111898469B (zh) * | 2020-07-09 | 2024-04-26 | 西北大学 | 一种基于缩放卷积层的信号时频特征提取方法及装置 |
| CN112986210A (zh) * | 2021-02-10 | 2021-06-18 | 四川大学 | 一种尺度自适应的微生物拉曼光谱检测方法及系统 |
| EP4080414A1 (en) * | 2021-04-23 | 2022-10-26 | Samsung Display Co., Ltd. | Systems and methods for identifying manufacturing defects |
| CN113959974A (zh) * | 2021-09-27 | 2022-01-21 | 桂林电子科技大学 | 一种普适性近红外光谱模型的传递方法 |
| CN113959974B (zh) * | 2021-09-27 | 2024-06-07 | 桂林电子科技大学 | 一种普适性近红外光谱模型的传递方法 |
| CN114067169A (zh) * | 2021-10-15 | 2022-02-18 | 浙江工业大学 | 一种基于卷积神经网络的拉曼光谱分析方法 |
| 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 |
| EP4449101A4 (en) * | 2021-12-17 | 2025-11-12 | Terramera Inc | SOIL PROPERTY PREDICTION SYSTEMS AND METHODS |
| CN114397297A (zh) * | 2022-01-19 | 2022-04-26 | 河南中烟工业有限责任公司 | 一种烤烟淀粉含量的快速无损检测方法 |
| CN114397297B (zh) * | 2022-01-19 | 2024-01-23 | 河南中烟工业有限责任公司 | 一种烤烟淀粉含量的快速无损检测方法 |
| CN114675140A (zh) * | 2022-03-09 | 2022-06-28 | 浙江理工大学 | 一种基于物联网的高压电网电晕监测系统 |
| CN114692773A (zh) * | 2022-04-08 | 2022-07-01 | 浙江工业大学 | 基于drs-vgg的端到端深度学习拉曼光谱数据分类方法 |
| CN114878543A (zh) * | 2022-05-27 | 2022-08-09 | 中国人民解放军32181部队 | 一种痕量物质的拉曼光谱探测识别方法及系统 |
| CN117235512A (zh) * | 2022-06-02 | 2023-12-15 | 中国科学院沈阳自动化研究所 | 一种面向libs的残差学习轻量卷积神经网络定量化方法 |
| CN117235512B (zh) * | 2022-06-02 | 2026-01-02 | 中国科学院沈阳自动化研究所 | 一种面向libs的残差学习轻量卷积神经网络定量化方法 |
| CN116597278A (zh) * | 2023-05-30 | 2023-08-15 | 黑龙江八一农垦大学 | 一种基于Cycle-GAN的光谱粒度效应降低方法 |
| US20250391514A1 (en) * | 2023-06-21 | 2025-12-25 | Zju-hangzhou Global Scientific And Technological Innovation Center | Method and device for spectral prediction of soil organic carbon based on spectrum-guided ensemble learning |
| US12537075B2 (en) * | 2023-06-21 | 2026-01-27 | Zju-hangzhou Global Scientific And Technological Innovation Center | Method and device for spectral prediction of soil organic carbon based on spectrum-guided ensemble learning |
| CN116561590B (zh) * | 2023-07-10 | 2023-10-03 | 之江实验室 | 基于深度学习的微纳光纤负载大小和位置预测方法及装置 |
| CN116561590A (zh) * | 2023-07-10 | 2023-08-08 | 之江实验室 | 基于深度学习的微纳光纤负载大小和位置预测方法及装置 |
| CN118209537A (zh) * | 2024-03-25 | 2024-06-18 | 哈尔滨理工大学 | 基于融合光谱与神经网络的变压器油纸绝缘老化诊断方法 |
| CN119903327A (zh) * | 2025-03-31 | 2025-04-29 | 山东大学 | 一种dfb激光器光谱参数的提取方法 |
Also Published As
| Publication number | Publication date |
|---|---|
| US20220051074A1 (en) | 2022-02-17 |
| CN109493287B (zh) | 2022-03-15 |
| CN109493287A (zh) | 2019-03-19 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109493287B (zh) | 一种基于深度学习的定量光谱数据分析处理方法 | |
| Zhang et al. | Noise reduction in the spectral domain of hyperspectral images using denoising autoencoder methods | |
| Jiang et al. | A fully convolutional neural network-based regression approach for effective chemical composition analysis using near-infrared spectroscopy in cloud | |
| Le | Application of deep learning and near infrared spectroscopy in cereal analysis | |
| Li et al. | Quantitative analysis of aflatoxin B1 of peanut by optimized support vector machine models based on near-infrared spectral features | |
| Yang et al. | Detection of starch in minced chicken meat based on hyperspectral imaging technique and transfer learning | |
| CN110991064A (zh) | 土壤重金属含量反演模型生成方法、系统、存储介质及反演方法 | |
| Yuan et al. | Moldy peanuts identification based on hyperspectral images and point-centered convolutional neural network combined with embedded feature selection | |
| Han et al. | Non-destructive detection of freshness in crayfish (Procambarus clarkii) based on near-infrared spectroscopy combined with deep learning | |
| Yang et al. | A recognition method of corn varieties based on spectral technology and deep learning model | |
| Wang et al. | Determination of watermelon soluble solids content based on visible/near infrared spectroscopy with convolutional neural network | |
| CN110736707B (zh) | 一种主仪器向从仪器光谱模型传递的光谱检测优化方法 | |
| CN116883720A (zh) | 基于空谱注意力网络的果蔬农药残留检测方法及系统 | |
| Yang et al. | A quality detection method of corn based on spectral technology and deep learning model | |
| Li et al. | Research on detection of potato varieties based on spectral imaging analytical algorithm | |
| CN113945537A (zh) | 一种高准确度近红外光谱定量模型建立方法 | |
| Yang et al. | A rapid recognition method of Auricularia auricula varieties based on near-infrared spectral characteristics | |
| He et al. | Research on the upgrade and maintenance method of apple soluble solids content models | |
| Zhao et al. | Estimation of soil property content with vis–NIR spectra by multitask deep learning based on attention mechanism and loss-weight balancing | |
| Wang et al. | Optimizing the effects of potato size and shape on near-infrared prediction models of potato quality using a linear-nonlinear algorithm | |
| Fu et al. | A multimodal fusion model for predicting the roasting-induced crispness of walnut kernels using machine vision, hyperspectral imaging, and electronic nose | |
| CN114778457B (zh) | 一种谷物中黄曲霉毒素b1含量检测方法及装置 | |
| CN121564537A (zh) | 基于高光谱的辣椒干燥过程品质指标无损分析方法及系统 | |
| Yang et al. | Corn protein classification and detection method based on near infrared spectral features combined with TCN model | |
| Zhu et al. | Fast and simultaneous detection of wheat kernel adulteration using hyperspectral imaging technology and deep convolutional neural network |
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: 19870634 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: 19870634 Country of ref document: EP Kind code of ref document: A1 |

