CN115512144A - Automatic XRF spectrogram classification method based on convolution self-encoder - Google Patents

Abstract

The invention discloses an XRF spectrogram automatic classification method based on a convolution self-encoder, and belongs to the field of XRF spectrum analysis processing. The method provided by the invention has the advantages that the acquired XRF spectrum data to be detected is subjected to normalization processing, and the one-dimensional spectrum data vector is converted into a two-dimensional spectrum information matrix form after the normalization processing, so that the XRF spectrum data can be processed by adopting a picture processing mode, and the classification precision is improved. On the basis, a neural network model of a convolution self-encoder is built, and the XRF spectrum spectrogram to be detected is subjected to feature compression to obtain a spectrum data spectrogram after the feature compression; by designing a kmeans classification network, unsupervised classification is realized, and the problems of low classification precision and efficiency caused by complicated XRF spectrum characteristic indexes are effectively solved.

Description

A method for automatic classification of XRF spectra based on convolutional autoencoder

technical field

The invention relates to the field of XRF spectrum analysis and processing, in particular to an automatic classification method for XRF spectrums based on a convolutional autoencoder.

Background technique

XRF (X Ray Fluorescence, X-ray fluorescence spectroscopy) is a method that uses primary X-ray photons to excite atoms in the substance to be measured to generate secondary X-rays for material composition analysis and chemical research analysis. It has many advantages such as simple pretreatment, no pollution in detection, low detection cost, high detection accuracy, wide range of detection elements, fast analysis speed, strong applicability, and good stability. Health and other technical fields have achieved extensive social and economic benefits. The XRF spectrum contains a lot of information. Facing the XRF spectrum of unknown and different samples measured in multiple fields, its spectral classification has always been a research hotspot.

At present, XRF spectrum classification mainly adopts machine learning methods, such as: backpropagation algorithm (BP), support vector machine (SVM), extreme learning machine (ELM), etc., and these algorithms show relatively powerful performance. However, these algorithms also have their own defects: for example, the BP algorithm is easy to fall into the local minimum, the SVM classification performance is unstable, and the calculation complexity is high. Although the ELM can greatly improve the learning speed, the effect is unstable.

In recent years, image classification methods based on deep learning have received extensive attention. The deep convolutional neural network (Convolutional Neural Network, CNN) can directly process the input two-dimensional image, has a powerful feature learning ability, and is widely used in the fields of computer vision and speech recognition of multi-category and large-scale data, and has achieved Great success. However, the application of XRF spectrum is still relatively small, because XRF spectrum is essentially a one-dimensional vector, and the data set is often not large. Although deep learning has a strong learning ability, it is easy to overfit; and the traditional deep learning network Structs are not well suited for working with one-dimensional data.

Therefore, it is necessary to study a new XRF spectrum classification method, so that it can realize automatic classification in the face of a large number of unknown XRF spectrum samples, and improve the classification accuracy and efficiency.

Contents of the invention

The purpose of the present invention is to provide an automatic classification method of XRF spectrum based on convolutional autoencoder, so as to overcome the problems of low classification accuracy and efficiency caused by complex feature indexes of XRF spectrum.

To achieve the above object, the present invention adopts the following technical solutions:

A method for automatic classification of XRF spectrograms based on convolutional autoencoders, comprising the following steps:

Step 1. Use a handheld X-ray fluorescence spectrometer to detect unknown samples as the XRF spectrogram data to be tested;

Step 2. Normalize the XRF spectrum data to be measured obtained in step 1, and convert it into a two-dimensional spectral information matrix form, and each spectral information matrix after conversion into a two-dimensional space contains 512×512 features ; Then create a training sample set according to the converted XRF spectrum information;

Step 3, generate a training set and a test set according to the training sample set created in step 2;

Step 4. Construction and training of convolutional autoencoder neural network model:

Step 4.1, iteratively training the training set based on the convolutional autoencoder neural network, and constructing a convolutional autoencoder neural network model;

Step 4.2, input the test set into the convolutional self-encoder neural network model obtained in step 4.1 to obtain the compressed XRF spectrogram;

Step 5. Build the Kmeans unsupervised classification model, and then input the feature-compressed XRF spectrum data map obtained in step 4.2 into the Kmeans unsupervised classification model, and obtain the XRF spectrum classification result of the unknown sample through training.

Further, the formula used for normalization in step 2 is:

In formula (1), X _max represents the maximum value of the one-dimensional spectral data of the same type of sample input, X _min represents the minimum value of the one-dimensional spectral data of the same type of sample input, and X _norm represents the normalized value of the sample of this type spectral data;

Further, the convolutional self-encoder neural network model constructed in step 4.1 includes an encoder and a decoder; wherein the encoder consists of 1 input layer and m encoding hidden layers; the decoder consists of m-1 decoding hidden layers and 1 output layer; the input data of the input layer is the training set obtained in step 3, which is used to output the input data to m coded hidden layers; m coded hidden layers are used to compress the input data to obtain the input data The core features of the input data are output to m-1 decoding hidden layers for decompression and then passed to the output layer. The output layer reconstructs the received core data and obtains the output of the XRF spectrum after feature compression.

Further, the hidden layers of the encoder and the decoder both use the Tanh nonlinear function as the activation function, and the output layer of the decoder uses the LeakyReLU nonlinear function as the activation function.

Furthermore, the step 4.2 uses the convolutional self-encoder neural network model to obtain the detailed process of the compressed XRF spectrogram as follows:

Step 4.2.1. Take the test set obtained in step 3 as the input data, input it to the input layer of the convolutional autoencoder, and map it to the m hidden layers of the encoder for feature compression and obtain the core features of the input data , to complete the process of convolutional autoencoder encoding, the expression of convolutional autoencoder encoding is as follows:

y=f(w _y x+b _y ) (5)

Among them, x is the loaded input data, y represents the features learned by the middle hidden layer, w _y represents the weight of the hidden layer input, b _y represents the offset coefficient of the hidden unit, and f represents the convolution operation;

Step 4.2.2, transfer the core features obtained in step 4.2.1 to the m-1 hidden layers of the decoder for decompression, and then transfer to an output layer for reconstruction to obtain an output data similar to the input data, namely The compressed XRF spectrogram after feature compression is used to complete the decoding process of the autoencoder. The expression of the decoding process is as follows:

z=f(w _z y+b _z ) (6)

Among them, y represents the feature learned by the middle hidden layer, z is the data reconstructed by the hidden feature y, w _z represents the weight of the hidden layer output, b _z represents the offset coefficient of the output unit, and f represents the convolution operation;

The constraints of the convolutional autoencoder are as follows:

w _y =w' _z =w (7)

Among them, w′ _z represents the transpose of w _z , which means that the convolutional autoencoder has the same binding weight w, which helps to halve the parameter amount of the model;

The goal of convolutional autoencoder training is to continuously reduce the error between input data and reconstructed data. The mathematical expression is as follows:

Here, the parameters that the autoencoder needs to train are: w, b _y , b _z , where x is the loaded sample data, z represents the reconstructed output data, and c(x,z) represents the input data and reconstructed data the error between

Its weight update rule is expressed by the following formula:

表示对权重W求偏导，

表示对隐藏单元的偏移系数b_y求偏导，

表示对输出单元的偏移系数b_z求偏导。Among them, cost(x,z) is the error loss between the input data and the reconstructed data, η is the learning rate,

Represents the partial derivative of the weight W,

Represents the partial derivative of the offset coefficient b _y of the hidden unit,

Represents the partial derivative of the offset coefficient b _z of the output unit.

Further, the construction and training process of the Kmeans unsupervised classification model in the step 5 are as follows:

Step 5.1, Kmeans unsupervised classification model construction

Step 5.1.1. Given a sample data set x={x1,x2,…,xn}, calculate the distance from any data sample in the data set to k (k≤n) centers, and specify The k (k≤n) data sample of k (k≤n) is used as the initial center, and the data sample is assigned to the class with the shortest distance from the data sample to the center;

Step 5.1.2, for the data samples in the class obtained in step 5.1.1, update each class center by methods such as finding the mean value of the class;

Step 5.1.3. Repeat the above two steps to update the class center. If the class center remains unchanged or the change of the class center is less than the preset threshold, the update ends, and the clusters are formed to complete the construction of the Kmeans unsupervised classification model; otherwise, continue;

Iteratively achieve the goal of minimizing the distance between the sample and the center of the class to which it belongs. The objective function is as follows:

Among them, μ _i represents the mean value of the set S _i , and the sum of all elements in the class and the mean distance is VarS _i ;

Choose a distance metric as follows:

where Xi represents the _i -th data sample;

Step 5.2, input the feature compressed XRF spectrogram data into the kmeans unsupervised classification model for training, and obtain the trained kmeans classification network model.

The beneficial effects of the present invention are as follows: the present invention performs normalization processing on the acquired XRF spectrum data to be measured, and after the normalization processing, converts it from a one-dimensional spectral data vector into a two-dimensional spectral information matrix form; It makes it possible to process the XRF spectrogram in the way of image processing, which improves the classification accuracy. On this basis, by building a convolutional autoencoder neural network model, the XRF spectrum spectrum to be measured is compressed to obtain the XRF spectrum data spectrum after feature compression; by designing the kmeans classification network, unsupervised classification is realized. It is more effective to solve the problem that the complexity of its characteristic indicators affects the classification accuracy and efficiency in the presence of a large number of unknown samples. At the same time, it fills in the vacancy in the research direction of the unknown sample classification method using the XRF spectrum of the sample. In addition, in the convolutional autoencoder neural network model, the hidden layers of the encoder and decoder both use the Tanh nonlinear function as the activation function, and the output layer of the decoder uses the LeakyReLU nonlinear function as the activation function to avoid overfitting. . This model combines deep learning neural network with XRF spectrogram, breaks the limitation of quantitative analysis of spectrogram and deep learning, and better integrates XRF spectrum and deep learning for qualitative analysis of samples.

Description of drawings

Fig. 1 is a method sequence diagram of the present invention;

Fig. 2 is the image generated after the data dimension of the one-dimensional spectrogram of the metal sample spectrum in the embodiment changes;

Fig. 3 is the image generated after the dimension change of the one-dimensional spectrogram data of the alloy sample spectrum in the embodiment;

Fig. 4 is the image generated after the embodiment soil sample spectrum one-dimensional spectrogram data dimension changes;

Fig. 5 is a diagram of classification results of the embodiment.

detailed description

In order to make the purpose of the present invention, technical solutions and advantages clearer, in conjunction with the following specific embodiments, and with reference to the accompanying drawings, the present invention is further described, as follows:

As shown in Figure 1, a kind of XRF spectrogram automatic classification method based on convolution autoencoder of the present invention, comprises the following steps:

Step 1. Obtain the XRF spectrum data to be tested:

Use a handheld X-ray fluorescence spectrometer to detect unknown samples as the spectrogram data to be tested;

The second step is the preprocessing of the spectral data to be measured: After normalizing the XRF spectral data to be measured obtained in step 1, the one-dimensional spectral data vector is converted into a two-dimensional spectral information matrix form, and then converted into a two-dimensional dimensional space. The detailed process is as follows:

1.1, adopt the normalization formula to carry out normalization processing to the XRF spectrum data to be measured that step 1 obtains, and its normalization formula is as follows:

Among them, X _max represents the maximum value of the two-dimensional spectral data of the same type of sample input, X _min represents the minimum value of the two-dimensional spectral data of the same type of sample input, and X _norm represents the normalized spectral data of this type of sample.

1.2. Convert the XRF spectrum data normalized in step 1.1 into the form of a two-dimensional feature matrix; each column in the two-dimensional feature matrix represents a spectral dimension, and each row represents all spectral information of each data to be measured. Switching to two-dimensional space, each spectral information matrix contains 512×512 features. This embodiment respectively provides the images generated after the data dimension of the metal sample, alloy sample and soil sample spectrum is changed after the bright XRF spectrum is converted into two-dimensional space; refer to Fig. 2, Fig. 3 and Fig. 4. After the dimension conversion, the spectral data can be processed by image processing.

1.3. Set the number of picture data in each package when loading through the batch_size parameter; then package the spectral image after dimension conversion according to the set loading number, and then load the packaged data to obtain the training sample set.

Step 3. Generate a training set and a test set according to the training sample set obtained in step 2. In this implementation, the training set and test set data generated according to the training sample set are shown in Table 1:

Table 1 XRF spectra of unknown samples collected in the experiment

Step 4. Construction and training of convolutional autoencoder neural network model:

4.1. Perform iterative training on the training set based on the convolutional autoencoder neural network, build a convolutional autoencoder neural network model, and set the activation function and loss function of the model.

The convolutional self-encoder neural network model constructed in step 4.1 includes an encoder and a decoder; wherein the encoder consists of 1 input layer and m encoding hidden layers; the decoder consists of m-1 decoding hidden layers and 1 The output layer is composed of; the input data of the input layer is the training set obtained in step 3, which is used to output the input data to m coded hidden layers; the m coded hidden layers are used to compress the input data to obtain the core features of the input data, And the core features of the input data are output to m-1 decoding hidden layers for decompression and then passed to the output layer, and the output layer reconstructs the received core data to obtain the output of the characteristic compressed XRF spectrum.

Due to the non-linearity of the XRF spectrum data, and too many unknown features. Traditional dimensionality reduction methods cannot solve the existing nonlinear problems well, and different types of unknown samples in the XRF spectrum have too many features in subsequent classification, resulting in a decline in classification efficiency. To solve these problems, both the hidden layers of the encoder and the decoder in this embodiment use the Tanh nonlinear function as the activation function, and the output layer of the decoder uses the LeakyReLU nonlinear function as the activation function. After using the Tanh nonlinear function and the LeakyReLU nonlinear function as the activation function, on the one hand, the performance of the convolutional autoencoder neural network model is improved, so that it can do both linear transformation and nonlinear transformation, and can also handle more complex data. On the other hand, the convolutional autoencoder neural network model is characterized by a reconstruction-oriented training form. After using Tanh and LeakyReLU nonlinear functions as activation functions, the model can restore the original input well, which shows that the intermediate The features stored in the hidden layer retain enough input information.

In this embodiment, the Tanh nonlinear activation function expression is as follows:

Among them, e is the base number of the natural logarithm function, which is a constant; x is the input data of the autoencoder, and f(x) is the output data of the autoencoder after being processed by the Tanh activation function;

The expression of the LeakyReLU nonlinear activation function is as follows:

Among them, x _end is the output data of the last layer of the autoencoder, and f(x _end ) is the output data processed by the LeakyReLU activation function.

It is set to use the root mean square error loss function to measure the deviation between the reconstructed data and the real data, and use the adaptive time estimation method to optimize the algorithm training network parameters. The root mean square error loss function calculation formula is as follows:

为开根号操作，Y_RMSE为均方误差的算术平方根值，

为输入数据真实标签分布，

为自编码器模型的预测值，上标l为样本中类别的总个数，用于指明是哪个真实值与预测值在进行损失计算，N为每一批次的总样本数。Among them, ∑ is the summation operation,

It is the square root operation, Y _RMSE is the arithmetic square root value of the mean square error,

is the true label distribution of the input data,

is the predicted value of the autoencoder model, the superscript l is the total number of categories in the sample, which is used to indicate which real value and predicted value are performing loss calculations, and N is the total number of samples in each batch.

4.2. Input the test set obtained in step 3 into the convolutional autoencoder neural network model to obtain a compressed XRF spectrum through training. The detailed steps are as follows:

4.2.1. Take the test set obtained in step 3 as the input data, input it to the input layer of the convolutional autoencoder neural network, and map it to the m hidden layers in the encoder for feature compression to obtain the core of the input data Features, to complete the encoding process of the convolutional autoencoder neural network model, the encoding expression of the convolutional autoencoder neural network model is as follows:

y=f(w _y x+b _y ) (4)

Among them, x is the loaded input data, y represents the features learned by the middle hidden layer, w _y represents the weight of the hidden layer input, b _y represents the offset coefficient of the hidden unit, and f represents the convolution operation;

4.1.2. Transfer the core features obtained in step 4.1.1 to the m-1 hidden layers of the decoder for decompression, and then transfer them to an output layer for reconstruction, and obtain an output data similar to the input data. That is, the compressed XRF spectrogram after feature compression, thus completing the decoding process of the autoencoder. In this embodiment, the scale of the input layer in the encoder is the same as that of the output layer in the decoder, and the expression of the decoding process is as follows:

z=f(w _z y+b _z ) (5)

Among them, y represents the feature learned by the middle hidden layer, z is the data reconstructed by the hidden feature y, w _z represents the weight of the hidden layer output, b _z represents the offset coefficient of the output unit, and f represents the convolution operation;

w _y =w' _z =w (6)

Among them, w′ _z represents the transpose of w _z , which is the constraint condition of the convolutional autoencoder, which means that the convolutional autoencoder has the same binding weight w, which helps to halve the parameter amount of the model;

The goal of convolutional autoencoder training is to continuously reduce the error between input data and reconstructed data. The mathematical expression is as follows:

Here, the parameters that the autoencoder needs to train are: w, b _y , b _z , where x is the loaded sample data, z represents the reconstructed output data, and c(x,z) represents the input data and reconstructed data the error between

Its weight update rule is expressed by the following formula:

表示对权重W求偏导，

表示对隐藏单元的偏移系数b_y求偏导，

表示对输出单元的偏移系数b_z求偏导。Among them, cost(x,z) is the error loss between the input data and the reconstructed data, η is the learning rate,

Represents the partial derivative of the weight W,

Represents the partial derivative of the offset coefficient b _y of the hidden unit,

Represents the partial derivative of the offset coefficient b _z of the output unit.

Generally speaking, the convolutional autoencoder neural network model of this embodiment uses the training set or test set of unknown samples created in step 3 as input data, and after inputting it to the convolutional autoencoder, the convolution operation is used as the autoencoder. The activation function of the encoder coding layer, and the convolutional autoencoder is trained in an unsupervised manner, and the trained convolutional autoencoder neural network model is obtained, and the training spectral data spectrogram after feature compression is obtained.

Step 5. Build the Kmeans unsupervised classification model, and then input the feature-compressed XRF spectrum data map obtained in step 4.2 into the Kmeans unsupervised classification model, and obtain the XRF spectrum classification result of the unknown sample through training. The detailed process is as follows:

5.1. Given a sample data set x={x1,x2,…,xn}, find the distance from any data sample in the data set to k(k≤n) centers, and specify k( k≤n) The data sample is used as the initial center, and the data sample is assigned to the class with the shortest distance from the data sample to the center;

5.1. Construction of Kmeans unsupervised classification model

5.1.2. For the data samples in the class obtained in step 5.1.1, update each class center by seeking the mean value of the class, etc.;

5.1.3. Repeat the above two steps to update the class center. If the class center remains unchanged or the change of the class center is less than a certain threshold, the update ends, and the clusters are formed to complete the construction of the Kmeans unsupervised classification model; otherwise continue;

Iteratively achieve the goal of minimizing the distance between the sample and the center of the class to which it belongs. The objective function is as follows:

Among them, μ _i represents the mean value of the set S _i , and the sum of all elements in the class and the mean distance is VarS _i ;

Choose a distance metric as follows:

where Xi represents the _i -th data sample;

5.2. Input the feature-compressed XRF spectrum data into the kmeans network for training, and obtain the trained kmeans classification network model.

In order to verify the feasibility and effect of the above method, it is applied in this embodiment.

Step 1. Obtain the XRF spectrum data of the unknown sample to be classified;

Step 2. Perform preprocessing and feature compression on the XRF spectrum data to be classified according to the method in step 2. Specifically: the XRF spectrum to be classified is first normalized and then converted into a two-dimensional spectral information matrix form, and then converted to a two-dimensional space. And create a training sample set according to the converted XRF spectrum information;

Step 3. Input the training sample set created in step 2 into the convolutional self-encoder neural network model for feature compression, and obtain the XRF spectrogram data after the compressed features. The compressed data size is 8×8×3.

Step 4, input the XRF spectrum data graph after feature compression into the trained Kmeans network, and obtain the XRF spectrum graph classification result of the unknown sample.

Through the above steps, the XRF spectrum classification results of unknown samples are finally obtained, as shown in Figure 5. The confusion matrix can intuitively understand the performance of the classification model in each type of sample, often as part of model evaluation, and it is very easy to indicate whether multiple categories are confused. In the mixed line matrix, the diagonal line is used as the dividing line. As shown in Figure 5, the position of the diagonal line indicates that the prediction is correct, and the position outside the diagonal line indicates that the sample is wrongly predicted as other samples. It can be seen that there are only two differences between soil and metal and alloy. Misclassification, the rest of the classification effect is very good, which proves the effectiveness of the embodiment method.

Claims (6)

1. A method for automatic classification of XRF spectrograms based on convolutional self-encoders, characterized in that: comprising the following steps: Step 1. Use a handheld X-ray fluorescence spectrometer to detect unknown samples as the XRF spectrogram data to be tested; Step 2. Normalize the XRF spectrum data to be measured obtained in step 1, and convert it into a two-dimensional spectral information matrix form, and each spectral information matrix after conversion into a two-dimensional space contains 512×512 features ; Then create a training sample set according to the converted XRF spectrum information; Step 3, generate a training set and a test set according to the training sample set created in step 2; Step 4. Construction and training of convolutional autoencoder neural network model: Step 4.1, iteratively training the training set based on the convolutional autoencoder neural network, and constructing a convolutional autoencoder neural network model; Step 4.2, input the test set into the convolutional self-encoder neural network model obtained in step 4.1 to obtain the compressed XRF spectrogram; Step 5. Build the Kmeans unsupervised classification model, and then input the feature-compressed XRF spectrum data map obtained in step 4.2 into the Kmeans unsupervised classification model, and obtain the XRF spectrum classification result of the unknown sample through training. 2. A kind of XRF spectrogram automatic classification method based on convolutional self-encoder according to claim 1, is characterized in that: the formula that carries out normalization processing in described step 2 is:

In formula (1), X _max represents the maximum value of the one-dimensional spectral data of the same type of sample input, X _min represents the minimum value of the one-dimensional spectral data of the same type of sample input, and X _norm represents the normalized value of the sample of this type spectral data. 3. A method for automatic classification of XRF spectrograms based on convolutional self-encoders according to claim 1, characterized in that: the convolutional self-encoder neural network model constructed in the step 4.1 includes an encoder and a decoder ; where the encoder is composed of 1 input layer and m encoding hidden layers; the decoder is composed of m-1 decoding hidden layers and 1 output layer; the input data of the input layer is the training set obtained in step 3, which is used to The input data is output to m encoding hidden layers; the m encoding hidden layers are used to compress the input data to obtain the core features of the input data, and output the core features of the input data to m-1 decoding hidden layers for decompression and transmission To the output layer, the output layer reconstructs the received core data to obtain the output of the characteristic compressed XRF spectrum. 4. a kind of XRF spectrogram automatic classification method based on convolution self-encoder according to claim 3, is characterized in that: the hidden layer of described encoder and decoder all adopts Tanh nonlinear function as activation function, so The output layer of the decoder uses the LeakyReLU non-linear function as the activation function. 5. A method for automatic classification of XRF spectra based on convolutional self-encoders according to claim 4, characterized in that: said step 4.2 uses the convolutional self-encoder neural network model to obtain compressed XRF spectrums The detailed process is: Step 4.2.1. Take the test set obtained in step 3 as the input data, input it to the input layer of the convolutional autoencoder, and map it to the m hidden layers of the encoder for feature compression and obtain the core features of the input data , to complete the process of convolutional autoencoder encoding, the expression of convolutional autoencoder encoding is as follows: y=f(w _y x+b _y ) (5) Among them, x is the loaded input data, y represents the features learned by the middle hidden layer, w _y represents the weight of the hidden layer input, b _y represents the offset coefficient of the hidden unit, and f represents the convolution operation; Step 4.2.2, transfer the core features obtained in step 4.2.1 to the m-1 hidden layers of the decoder for decompression, and then transfer to an output layer for reconstruction to obtain an output data similar to the input data, namely The compressed XRF spectrogram after feature compression is used to complete the decoding process of the autoencoder. The expression of the decoding process is as follows: z=f(w _z y+b _z ) (6) Among them, y represents the feature learned by the middle hidden layer, z is the data reconstructed by the hidden feature y, w _z represents the weight of the hidden layer output, b _z represents the offset coefficient of the output unit, and f represents the convolution operation; The constraints of the convolutional autoencoder are as follows: w _y =w _z '=w (7) Among them, w _z ′ represents the transpose of w _z , which means that the convolutional autoencoder has the same binding weight w, which helps to halve the parameter amount of the model; The goal of convolutional autoencoder training is to continuously reduce the error between input data and reconstructed data. The mathematical expression is as follows:

Here, the parameters that the autoencoder needs to train are: w, b _y , b _z , where x is the loaded sample data, z represents the reconstructed output data, and c(x,z) represents the input data and reconstructed data the error between Its weight update rule is expressed by the following formula:

Among them, cost(x,z) is the error loss between the input data and the reconstructed data, η is the learning rate,

Represents the partial derivative of the weight W,

Represents the partial derivative of the offset coefficient b _y of the hidden unit,

Represents the partial derivative of the offset coefficient b _z of the output unit. 6. A kind of XRF spectrogram automatic classification method based on convolution self-encoder according to claim 1, it is characterized in that: in the described step 5, the construction and training process of Kmeans unsupervised classification model are as follows: Step 5.1, Kmeans unsupervised classification model construction Step 5.1.1. Given a sample data set x={x1,x2,…,xn}, calculate the distance from any data sample in the data set to k (k≤n) centers, and specify The k (k≤n) data sample of k (k≤n) is used as the initial center, and the data sample is assigned to the class with the shortest distance from the data sample to the center; Step 5.1.2, for the data samples in the class obtained in step 5.1.1, update each class center by seeking the mean value of the class; Step 5.1.3. Repeat the above two steps to update the class center. If the class center remains unchanged or the change of the class center is less than the preset threshold, the update ends, and the clusters are formed to complete the construction of the Kmeans unsupervised classification model; otherwise, continue; Iteratively achieve the goal of minimizing the distance between the sample and the center of the class to which it belongs. The objective function is as follows:

Among them, μ _i represents the mean value of the set S _i , and the sum of all elements in the class and the mean distance is VarS _i ; Choose a distance metric as follows:

where Xi represents the _i -th data sample; Step 5.2, input the feature compressed XRF spectrogram data into the kmeans unsupervised classification model for training, and obtain the trained kmeans classification network model.

Images