CN114881111B - Automatic discrimination method of manned spacecraft on-orbit oxygen consumption state based on unsupervised learning - Google Patents

Abstract

The present invention relates to a method for automatically judging the on-orbit oxygen consumption state of a manned spacecraft based on unsupervised deep learning, comprising: 1) sending raw data into an encoder through a data encoding module 1-1 to obtain corresponding hidden features; 2 ) Through the data decoding module 1-2, the hidden features are sent to the decoder to obtain the reconstructed data; 3) Through the pseudo-label generation module 1-3, the K-means algorithm is used to cluster the latent features output by the encoder, and the The clustering result is used as a pseudo-label to automatically determine the oxygen consumption status of the manned spacecraft in orbit; 4) through the parameter update module 1-4, the network parameters of the encoder and decoder are updated by using the minimized loss function; the loss function includes Similarity loss and reconstruction loss. The invention solves the binary classification judgment problem of any pair of data through unsupervised deep learning, thereby realizing the automatic judgment of the on-orbit oxygen consumption state of the manned spacecraft.

Description

Automatic identification method of oxygen consumption state of manned spacecraft on orbit based on unsupervised learning

Technical Field

The present invention relates to the field of intelligent information processing and computer technology, and in particular to a method for automatically distinguishing the oxygen consumption state of a manned spacecraft on orbit based on unsupervised learning.

Background Art

It is of great significance to cluster the oxygen consumption patterns in different states such as unmanned state, light activity state, moderate activity state, and heavy activity state of manned spacecraft in orbit, and automatically learn the oxygen consumption patterns in different states, so as to automatically determine the oxygen consumption state of the manned spacecraft when it is in orbit.

Most traditional time series clustering algorithms divide the clustering task into two independent stages: feature learning and clustering. However, the discriminant features of different data sets are often different, so this type of method has poor generalization ability.

Summary of the invention

The purpose of the present invention is to provide a method for automatically distinguishing the on-orbit oxygen consumption state of a manned spacecraft based on unsupervised learning, and to realize a method for automatically distinguishing the on-orbit oxygen consumption state of a manned spacecraft based on unsupervised deep learning.

The present invention provides a method for automatically distinguishing the oxygen consumption state of a manned spacecraft on orbit based on unsupervised learning, comprising:

1) The original data is sent to the encoder through the data encoding module to obtain the corresponding latent features;

2) The latent features are sent to the decoder through the data decoding module to obtain the reconstructed data;

3) Through the pseudo-label generation module, the K-means algorithm is used to cluster the latent features output by the encoder, and the clustering results are used as pseudo-labels to automatically identify the oxygen consumption state of manned spacecraft on orbit;

4) Through the parameter update module, the network parameters of the encoder and decoder are updated by minimizing the loss function; wherein the loss function includes similarity loss and reconstruction loss; wherein the similarity loss is the cross entropy loss constructed based on the pseudo-label, and the reconstruction loss is the mean square error between the output of the decoder and the original data.

Furthermore, the data encoding module includes two network branches: a fully convolutional network FCN and a residual network ResNet; the fully convolutional network FCN is divided into 4 modules; wherein the first 3 modules are composed of a one-dimensional convolution layer Conv1D, a batch normalization layer BN and a ReLU activation function layer; the number of channels of the convolution layer of each module is 128, 256, and 128 respectively, the convolution kernel size is 8, 5, and 3 respectively, and the step size is 1; the fourth module is a global average pooling layer; each module is connected in sequence, that is, the output of each module will be used as the input of the next module;

The residual network ResNet is divided into 4 modules, wherein each of the first 3 modules contains three layers of one-dimensional convolution, and each layer of convolution is connected to a batch normalization layer and a ReLU activation function; the number of three-layer convolution channels in each module is (64, 64, 64), (192, 192, 192), (192, 192, 192), respectively, the convolution kernel size is (8, 5, 3), and the step size is 1; the first three modules use residual connection, that is, the input and output of each module are added as the input of the next module; the last module in the residual network ResNet is a global average pooling layer;

The data encoding module connects the outputs of the two network branches together, and passes through a fully connected layer to map the high-dimensional features to the sample category space, and finally passes through the softmax layer to obtain the label feature Z of the data; for an original data x _i , the data encoding module is expressed as z _i =f(x _i ; _we ), where _we is the network parameter of the encoder, f(.) represents the encoding process, and z _i is the output of the original data after the encoder.

其中，3个转置卷积层中卷积核的数量分别为256、128、1，卷积核大小分别为3、5、8，步长为1；前两个转置卷积模块中的转置卷积层后添加有BN层和ReLU层，最后一个block中的转置卷积层后只添加有BN层，而不使用激活函数；对于一个隐特征z_i，所述数据解码模块表示为

其中，w_d是解码器的网络参数，g(.)表示解码过程，

是对原始数据x_i的重构，即隐特征z_i经过解码器的输出。Furthermore, the data decoding module reconstructs the features extracted by the encoder, thereby constraining the latent features to retain important information in the original data; the input of the data decoding module is the label feature Z extracted by the encoder; Z first passes through the fully connected layer to restore the time dimension of the data, and then passes through three transposed convolution modules to obtain the reconstructed data

Among them, the number of convolution kernels in the three transposed convolution layers is 256, 128, and 1 respectively, the convolution kernel sizes are 3, 5, and 8 respectively, and the step size is 1; the transposed convolution layers in the first two transposed convolution modules are followed by BN layers and ReLU layers, and the transposed convolution layer in the last block is followed by only BN layers without using activation functions; for a hidden feature z _i , the data decoding module is expressed as

Among them, w _d is the network parameter of the decoder, g(.) represents the decoding process,

It is the reconstruction of the original data x _i , that is, the output of the latent feature z _i after the decoder.

Furthermore, the pseudo-label generation module uses the k-means algorithm to cluster the label feature Z; the distance metric in k-means uses the cosine distance, and the clustering process jointly learns a k×k centroid matrix C and the cluster assignment _yi∈ {0,1} ^k for _zi , where _yi is a 0-1 vector; the objective function of k-means optimization is expressed as follows:

而不使用质心矩阵C；将其转换为样本对标签矩阵L＝{l_ij}；其中，l_ij为一个二元变量，表示x_i和x_j是否属于一类；若l_ij＝1，那么表示两者为一类，l_ij＝0表示两者属于不同的类别；将其转换为矩阵乘法并利用GPU来加速L的计算，即：When the k-means algorithm completes the iteration, it only focuses on the current optimal cluster assignment

Instead of using the centroid matrix C, we convert it into a sample pair label matrix L = {l _ij }, where l _ij is a binary variable indicating whether _xi and _xj belong to the same category. If l _ij = 1, then they belong to the same category, and l _ij = 0 indicates that they belong to different categories. We convert it into matrix multiplication and use the GPU to accelerate the calculation of L, namely:

L = Y ^* Y ^{* T.}

Furthermore, the parameter updating module uses a back propagation algorithm to adjust the network parameters according to the similarity loss function and the reconstruction loss function;

The calculation formula of the similarity loss function is as follows:

s _ij =cos( _zi , _zj )

Where s _ij represents the cosine similarity between the label features z _i and z _j of any two data in the current batch; l _ij is the pseudo label generated by the k-means algorithm; the similarity loss function only updates the network weights w _e of the encoder during back propagation;

The calculation formula of the reconstruction loss function is as follows:

是对原始数据x_i的重构；重构损失则更新编码器w_e和解码器w_d；Where m represents the number of samples in the current batch;

is the reconstruction of the original data x _i ; the reconstruction loss updates the encoder w _e and decoder w _d ;

将其送入编码器得到标签特征z_i，取z_i中最大响应的索引作为最终该数据的类别标签c_i，即：After the model training is completed, a test set of size p is given

Send it to the encoder to get the label feature z _i , and take the index of the maximum response in z _i as the final category label c _i of the data, that is:

c _i =argmax _h (z _ih ), h=1,...,k.

Through the above scheme, the method for automatically distinguishing the oxygen consumption state of manned spacecraft on orbit based on unsupervised learning has the following technical effects:

1) The present invention implements an end-to-end deep clustering algorithm, which jointly learns the feature representation and clustering of data to improve the generalization ability of the clustering algorithm on different data sets.

2) The present invention converts the clustering problem into a binary classification problem for any pair of data, takes the cosine similarity between the latent features of any pair of data in the current batch as the predicted value of the similarity of the data pair, and obtains the pseudo-label of the similarity of any pair of data by performing k-means clustering on the latent features of all data in the current batch, thereby guiding the feature encoding network.

3) The present invention decodes latent features to reconstruct original data, and constrains latent features through reconstruction errors.

The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, the following is a detailed description of the preferred embodiments of the present invention in conjunction with the accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a flow chart of a method for automatically determining the oxygen consumption state of a manned spacecraft on orbit based on unsupervised learning according to the present invention;

FIG2 is a schematic diagram of a network structure of a data encoding module according to the present invention;

FIG3 is a schematic diagram of a network structure of a data decoding module of the present invention;

FIG4 is a typical curve diagram of oxygen consumption state 1 (man-supplied oxygen state) of a manned spacecraft on orbit;

FIG5 is a typical curve diagram of oxygen consumption state 2 (manned oxygen consumption state) of a manned spacecraft on orbit;

FIG6 is a typical curve diagram of oxygen consumption state 3 (unmanned state) of a manned spacecraft on orbit;

FIG. 7 is a label feature distribution diagram of the learning result of the manned spacecraft on-orbit oxygen consumption data according to the present invention.

DETAILED DESCRIPTION

The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

As shown in FIG1 , this embodiment provides a method for automatically determining the oxygen consumption state of a manned spacecraft on orbit based on unsupervised deep learning, including:

1) The original data is sent to the encoder through the data encoding module 1-1 to obtain the corresponding latent features;

2) The latent features are sent to the decoder through the data decoding module 1-2 to obtain the reconstructed data;

3) Through the pseudo-label generation module 1-3, the K-means algorithm is used to cluster the latent features output by the encoder, and the clustering results are used as pseudo-labels to automatically determine the oxygen consumption state of the manned spacecraft on orbit;

4) Through parameter updating modules 1-4, the network parameters of the encoder and decoder are updated by minimizing the loss function; wherein the loss function includes similarity loss and reconstruction loss; wherein the similarity loss is the cross entropy loss constructed based on the pseudo-label, and the reconstruction loss is the mean square error between the output of the decoder and the original data.

The data encoding module contains two branches: the fully convolutional network FCN and the residual network ResNet. The structure of FCN can be divided into 4 modules. The first 3 modules are composed of a one-dimensional convolution layer Conv1D, a batch normalization layer BN (BatchNormalization) and a ReLU activation function layer. The number of channels of each module convolution layer is 128, 256, and 128 respectively, the convolution kernel size is 8, 5, and 3 respectively, and the step size is 1. The fourth module is a global average pooling layer. Each module is connected in sequence, that is, the output of each module will be used as the input of the next module. The structure of ResNet is also divided into 4 modules. Each of the first 3 modules contains three layers of one-dimensional convolution, and each layer of convolution is connected to a batch normalization layer and a ReLU activation function. The number of three-layer convolution channels in each module is (64, 64, 64), (192, 192, 192), (192, 192, 192), respectively, the convolution kernel size is (8, 5, 3), and the step size is 1. The first three modules use residual connection, that is, the input and output of each module are added as the input of the next module. The last module in ResNet is the global average pooling layer. The data encoding module connects the outputs of the two network branches together, passes through a fully connected layer, maps the high-dimensional features to the sample category space, and finally passes through the softmax layer to obtain the label feature Z of the data. For an original data x _i , the data encoding module can be expressed as z _i = f(x _i ; _we ), where _we is the network parameter of the encoder, f(.) represents the encoding process, and z _i is the output of the original data after the encoder.

具体地，3个转置卷积层中卷积核的数量分别为256、128、1，卷积核大小分别为3、5、8，步长为1。前两个模块中的转置卷积层后还添加了BN层和ReLU层，而最后一个block中的转置卷积层后只添加了BN层，而不使用激活函数。对于一个隐特征z_i，数据解码模块可以表示为

其中w_d是解码器的网络参数，g(.)表示解码过程，

是对原始数据x_i的重构，即隐特征z_i经过解码器的输出。The data decoding module reconstructs the features extracted by the encoder, with the goal of constraining the latent features so that they retain the important information in the original data. The input of the data decoding module is the label feature Z extracted by the encoder. Z first passes through the fully connected layer to restore the time dimension of the data, and then passes through three transposed convolution modules to obtain the reconstructed data.

Specifically, the number of convolution kernels in the three transposed convolution layers is 256, 128, and 1, respectively, and the convolution kernel sizes are 3, 5, and 8, respectively, with a stride of 1. The transposed convolution layers in the first two modules are followed by BN layers and ReLU layers, while the transposed convolution layers in the last block are followed by only BN layers without using activation functions. For a hidden feature z _i , the data decoding module can be expressed as

Where w _d is the network parameter of the decoder, g(.) represents the decoding process,

It is the reconstruction of the original data x _i , that is, the output of the latent feature z _i after the decoder.

The pseudo-label generation module uses the k-means algorithm to cluster the label feature Z. The standard k-means clustering algorithm uses the Euclidean distance as the measure of the distance between the sample point and the cluster center. Considering that the metric of the similarity between samples in this method is the cosine distance, in order to avoid the problems caused by different distance metrics, the distance metric in k-means also uses the cosine distance. The clustering process actually jointly learns a k×k centroid matrix C and the cluster assignment _yi ∈{0,1} ^k for _zi , where _yi is a 0-1 vector. Therefore, the objective function of k-means optimization can be expressed as follows:

而不使用质心矩阵C。由于本方法关注的是任意一对数据的相似性，故将其转换为样本对标签矩阵L＝{l_ij}。其中，l_ij为一个二元变量，表示x_i和x_j是否属于一类。若l_ij＝1，那么表示两者为一类，l_ij＝0表示两者属于不同的类别。最简单的获得L的方法是通过二重循环来生成，不过考虑到循环语句会降低模型的训练速度，本文将其转换为矩阵乘法并利用GPU来加速L的计算，即：When the k-means algorithm completes the iteration, this method only focuses on the current optimal clustering assignment

Instead of using the centroid matrix C. Since this method focuses on the similarity of any pair of data, it is converted into a sample pair label matrix L = {l _ij }. Among them, l _ij is a binary variable, indicating whether x _i and x _j belong to the same category. If l _ij = 1, it means that the two are in the same category, and l _ij = 0 means that the two belong to different categories. The simplest way to obtain L is to generate it through a double loop, but considering that the loop statement will reduce the training speed of the model, this paper converts it into matrix multiplication and uses the GPU to accelerate the calculation of L, that is:

L＝Y ^* Y ^{* T}

The parameter update module uses the back propagation algorithm to adjust the network parameters according to the similarity loss function and the reconstruction loss function. The calculation formula of the similarity loss function is as follows:

s _ij =cos( _zi , _zj )

Where s _ij represents the cosine similarity between the label features z _i and z _j of any two data in the current batch; l _ij is the pseudo label generated by the k-means algorithm. The similarity loss function only updates the network weights w _e of the encoder during back propagation.

The calculation formula of the reconstruction loss function is as follows:

是对原始数据x_i的重构。重构损失则更新编码器w_e和解码器w_d。Where m represents the number of samples in the current batch;

is the reconstruction of the original data x _i . The reconstruction loss updates the encoder w _e and decoder w _d .

将其送入编码器得到标签特征z_i，取z_i中最大响应的索引作为最终该数据的类别标签c_i，即：After the model training is completed, a test set of size p is given

Send it to the encoder to get the label feature z _i , and take the index of the maximum response in z _i as the final category label c _i of the data, that is:

c _i =argmax _h (z _ih ),h=1,…,k

Referring to Figure 2, it is a schematic diagram of the network structure of the data encoding module, which includes the following steps: first perform step 2-1, send the original data into a one-dimensional convolution layer with a convolution kernel size of 8, a step size of 1, and a number of channels of 128, and then perform batch normalization and non-linear function ReLU activation operations on the output. Perform step 2-2, send the output of the previous step into a one-dimensional convolution layer with a convolution kernel size of 5, a step size of 1, and a number of channels of 256, and then perform batch normalization and non-linear function ReLU activation operations on the output. Perform step 2-3, send the output of the previous step into a one-dimensional convolution layer with a convolution kernel size of 3, a step size of 1, and a number of channels of 128, and then perform batch normalization and non-linear function ReLU activation operations on the output. Perform step 2-4, and perform a global pooling operation on the output of the previous step. Execute steps 2-5, send the original data into a one-dimensional convolutional layer with a convolution kernel size of 8, a stride of 1, and 64 channels, and perform batch normalization and ReLU activation on the output; then send the output into a one-dimensional convolutional layer with a convolution kernel size of 5, a stride of 1, and 64 channels, and perform batch normalization and ReLU activation on the output; then send the output into a one-dimensional convolutional layer with a convolution kernel size of 3, a stride of 1, and 64 channels, and perform batch normalization and ReLU activation on the output; finally, add the original data and the output as the output of this step. Execute steps 2-6, send the output of the previous step to a one-dimensional convolutional layer with a kernel size of 8, a stride of 1, and 192 channels, and perform batch normalization and ReLU activation on the output; then send the output to a one-dimensional convolutional layer with a kernel size of 5, a stride of 1, and 192 channels, and perform batch normalization and ReLU activation on the output; then send the output to a one-dimensional convolutional layer with a kernel size of 3, a stride of 1, and 192 channels, and perform batch normalization and ReLU activation on the output; finally, add the input and output of this step as the final output. Execute step 2-7, send the output of the previous step to a one-dimensional convolution layer with a convolution kernel size of 8, a stride of 1, and a number of channels of 192, and perform batch normalization and nonlinear function ReLU activation on the output; then send the output to a one-dimensional convolution layer with a convolution kernel size of 5, a stride of 1, and a number of channels of 192, and perform batch normalization and nonlinear function ReLU activation on the output; then send the output to a one-dimensional convolution layer with a convolution kernel size of 3, a stride of 1, and a number of channels of 192, and perform batch normalization and nonlinear function ReLU activation on the output; finally, add the input and output of this step as the final output. Execute step 2-8, perform global average pooling on the output of the previous step. Execute step 2-9, connect the outputs of steps 2-4 and 2-8. Execute step 2-10, send the output of the previous step to a fully connected layer with the number of neurons equal to the number of categories. Execute step 2-11, perform softmax operation on the output of the previous step to obtain the label features of the original data.

Referring to Figure 3, it is a schematic diagram of the network structure of the data decoding module, which includes the following steps: First, perform step 3-1 to send the label features of the data to the fully connected layer, and the number of neural elements in this layer is the length of the original data. Perform step 3-2 to send the output of the previous step to a one-dimensional transposed convolution layer with a convolution kernel size of 3, a step size of 1, and a number of channels of 256, and then perform batch normalization and nonlinear function ReLU activation operations on the output. Perform step 3-3 to send the output of the previous step to a one-dimensional transposed convolution layer with a convolution kernel size of 5, a step size of 1, and a number of channels of 128, and then perform batch normalization and nonlinear function ReLU activation operations on the output. Perform step 3-4 to send the output of the previous step to a one-dimensional transposed convolution layer with a convolution kernel size of 8, a step size of 1, and a number of channels of 1, and then perform batch normalization on the output.

4, 5 and 6, which are typical curve diagrams of several on-orbit oxygen consumption states of manned spacecraft, namely manned oxygen supply state, manned oxygen consumption state and unmanned state.

Referring to Figure 7, it is a label feature distribution diagram for the learning results of the on-orbit oxygen consumption data of manned spacecraft. From the figure, it can be seen that the labels corresponding to the data of different oxygen consumption states are clustered into distinct clusters, with obvious distinctions and accurate judgments.

The present invention extracts features from the original time series data through an encoder, which includes two network branches, a fully convolutional network and a residual network. The features of the two networks are fused as data feature representation; the latent features are then sent to the decoder to reconstruct the original data. Unsupervised deep learning is used to solve the binary classification problem of any pair of data, thereby realizing automatic discrimination of the oxygen consumption state of manned spacecraft on orbit. By calculating the cosine similarity of the latent features of any pair of data in the current batch, and using it as the network's prediction of the similarity of the pair of data, and then using K-means to cluster the latent features to generate pseudo-labels to guide the update of network parameters, finally the oxygen consumption state of the manned spacecraft on orbit is automatically discriminated based on the pseudo-labels. It has the following technical effects:

1) The present invention implements an end-to-end deep clustering algorithm, which jointly learns the feature representation and clustering of data to improve the generalization ability of the clustering algorithm on different data sets.

2) The present invention converts the clustering problem into a binary classification problem for any pair of data, takes the cosine similarity between the latent features of any pair of data in the current batch as the predicted value of the similarity of the data pair, and obtains the pseudo-label of the similarity of any pair of data by performing k-means clustering on the latent features of all data in the current batch, thereby guiding the feature encoding network.

3) The present invention decodes latent features to reconstruct original data, and constrains latent features through reconstruction errors.

The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the technical principles of the present invention, and these improvements and modifications should also be regarded as within the scope of protection of the present invention.

Claims (2)

1. A method for automatically distinguishing the oxygen consumption state of a manned spacecraft on orbit based on unsupervised learning, characterized in that the method automatically distinguishes the oxygen consumption state of a manned spacecraft on orbit through unsupervised deep learning for the oxygen consumption data of the manned spacecraft on orbit, and specifically comprises the following steps: 1) The original time series data is sent to the encoder through the data encoding module to obtain the corresponding latent features; 2) The latent features are sent to the decoder through the data decoding module to obtain the reconstructed data; 3) Through the pseudo-label generation module, the K-means algorithm is used to cluster the latent features output by the encoder, and the clustering results are used as pseudo-labels to automatically identify the on-orbit oxygen consumption state of the manned spacecraft; the on-orbit oxygen consumption state includes manned oxygen supply state, manned oxygen consumption state, and unmanned state; 4) updating the network parameters of the encoder and decoder by minimizing the loss function through the parameter update module; wherein the loss function includes similarity loss and reconstruction loss; wherein the similarity loss is the cross entropy loss constructed based on the pseudo-label, and the reconstruction loss is the mean square error between the output of the decoder and the original data; The data encoding module includes two network branches: a fully convolutional network FCN and a residual network ResNet; the fully convolutional network FCN is divided into 4 modules; the first 3 modules are composed of a one-dimensional convolution layer Conv1D, a batch normalization layer BN and a ReLU activation function layer; the number of channels of the convolution layer of each module is 128, 256, and 128 respectively, the convolution kernel size is 8, 5, and 3 respectively, and the step size is 1; the fourth module is a global average pooling layer; each module is connected in sequence, that is, the output of each module will be used as the input of the next module; The residual network ResNet is divided into 4 modules, wherein each of the first 3 modules contains three layers of one-dimensional convolution, and each layer of convolution is connected to a batch normalization layer and a ReLU activation function; the number of three-layer convolution channels in each module is (64, 64, 64), (192, 192, 192), (192, 192, 192), respectively, the convolution kernel size is (8, 5, 3), and the step size is 1; the first three modules use residual connection, that is, the input and output of each module are added as the input of the next module; the last module in the residual network ResNet is a global average pooling layer; The data encoding module connects the outputs of the two network branches together, passes through a fully connected layer, maps the high-dimensional features to the sample category space, and finally passes through the softmax layer to obtain the label feature Z of the data; for an original data x _i , the data encoding module is expressed as z _i =f(x _i ; _we ), where _we is the network parameter of the encoder, f(.) represents the encoding process, and z _i is the output of the original data after the encoder; The data decoding module reconstructs the features extracted by the encoder, thereby constraining the latent features to retain the important information in the original data; the input of the data decoding module is the label feature Z extracted by the encoder; Z first passes through the fully connected layer to restore the time dimension of the data, and then passes through three transposed convolution modules to obtain the reconstructed data

Among them, the number of convolution kernels in the three transposed convolution layers is 256, 128, and 1 respectively, the convolution kernel sizes are 3, 5, and 8 respectively, and the step size is 1; the transposed convolution layers in the first two transposed convolution modules are followed by BN layers and ReLU layers, and the transposed convolution layer in the last block is followed by only BN layers without using activation functions; for a hidden feature z _i , the data decoding module is expressed as

Among them, w _d is the network parameter of the decoder, g(.) represents the decoding process,

It is the reconstruction of the original data x _i , that is, the output of the latent feature z _i after the decoder; The pseudo-label generation module uses the k-means algorithm to cluster the label feature Z; the distance metric in k-means uses the cosine distance, and the clustering process jointly learns a k×k centroid matrix C and the cluster assignment _yi∈ {0,1} ^k for _zi , where _yi is a 0-1 vector; the objective function of k-means optimization is expressed as follows:

When the k-means algorithm completes the iteration, it only focuses on the current optimal cluster assignment

Instead of using the centroid matrix C, we convert it into a sample pair label matrix L = {l _ij }, where l _ij is a binary variable indicating whether _xi and _xj belong to the same category. If l _ij = 1, then they belong to the same category, and l _ij = 0 indicates that they belong to different categories. We convert it into matrix multiplication and use the GPU to accelerate the calculation of L, namely: L = Y ^* Y ^{* T.} 2. The method for automatically distinguishing the oxygen consumption state of a manned spacecraft on orbit based on unsupervised learning according to claim 1 is characterized in that the parameter updating module uses a back propagation algorithm to adjust the network parameters according to a similarity loss function and a reconstruction loss function; The calculation formula of the similarity loss function is as follows: s _ij =cos(z _i ,z _j )

Where s _ij represents the cosine similarity between the label features z _i and z _j of any two data in the current batch; l _ij is the pseudo label generated by the k-means algorithm; the similarity loss function only updates the network weights w _e of the encoder during back propagation; The calculation formula of the reconstruction loss function is as follows:

Where m represents the number of samples in the current batch;

is the reconstruction of the original data x _i ; the reconstruction loss updates the encoder w _e and decoder w _d ; After the model training is completed, a test set of size p is given

Send it to the encoder to get the label feature z _i , and take the index of the maximum response in z _i as the final category label c _i of the data, that is: c _i =argmax _h (z _ih ), h = 1,...,k.

Images