A method for feature encoding and extraction of pea protein data

Abstract

This invention discloses a method for feature encoding and extraction of pea protein data, comprising the following steps: encoding the data into a frequency distribution matrix with consistent dimension based on the sequential order of each element in the pea protein sequence data and the frequency of different amino acids; realizing pea protein feature extraction based on the three-dimensional structure by mapping the pea protein three-dimensional model to a voxel grid model; combining the pea protein amino acid frequency matrix with the features extracted by the multilayer combined feature extraction neural network, and performing feature extraction on the pea protein data by fusing the feature extraction neural network. Through discussions on feature extraction and encoding of protein sequence data, sequence data with inconsistent dimensions are encoded into vectors with consistent dimensions, solving the problem that amino acid sequence data cannot be directly input into machine learning algorithms for recognition, and enhancing the accuracy and comprehensiveness of feature extraction by introducing a frequency matrix in feature extraction.

Description

[0001] This invention is a divisional application of the invention patent application number "2021110656587", entitled "A method for extracting features from pea protein data". Technical Field

[0002] This application relates to the field of protein data processing, and in particular to a method for encoding and extracting pea protein data features. Background Technology

[0003] Peas are the world's second largest soybean crop. Pea protein has high solubility, water absorption and emulsifying properties, and high nutritional value. It is a good source of essential amino acids. The extraction of data features directly affects the construction of protein sequence structure and function prediction.

[0004] Protein amino acid sequence data is not only massive in scale but also complex in structure, making traditional data feature extraction methods no longer suitable. Existing pea protein sequence data feature extraction methods mainly use amino acid composition, physicochemical properties of amino acid residues, frequency spectrum, etc., which cannot guarantee the extraction of all feature information. Furthermore, the extraction process is computationally intensive, difficult to extract information, and lacks sufficient extraction accuracy. Summary of the Invention

[0005] The technical problem to be solved by this application is that existing methods for extracting features from pea protein sequence data cannot guarantee the extraction of all feature information, and the extraction process is computationally intensive, difficult to extract information, and lacks precision. Therefore, this application provides a method for feature encoding and extraction of pea protein data.

[0006] The specific technical solution is as follows:

[0007] A method for feature encoding and extraction of pea protein data, the method comprising the following steps:

[0008] S1. Based on the sequential order of each element in the pea protein sequence data and the frequency of different amino acids, the data is encoded into a frequency distribution matrix with consistent dimension;

[0009] S2. By mapping the 3D model of pea protein to a voxel mesh model, feature extraction of pea protein based on the 3D structure is achieved;

[0010] S3. Combine the pea protein amino acid frequency matrix with the features extracted by the multilayer combined feature extraction neural network, and extract features from the pea protein data by fusing the feature extraction neural network.

[0011] Preferably, step S1 includes: a method for calculating the amino acid frequency distribution matrix, wherein the calculation method is as follows:

[0012] Let the set of amino acids contained in pea protein be Ω = {Ω1, Ω2, ..., Ω}. N}, Ω N Let N be the Nth amino acid, and N be the number of amino acid types in pea protein. Map a pea protein amino acid sequence with P amino acid residues to a point in a 20-dimensional Euclidean space, represented by the vector S = {s1, s2, ..., s...}. P}, where P is the number of elements in a given amino acid sequence. For amino acid Ω N Based on the frequency of occurrence in sequence S, the amino acid frequency matrix of pea protein is as follows:

[0013] M = [m ij ] N×N

[0014]

[0015] Where i, j∈N, the sum of all elements in the sequence is Normalize the elements in the sequence:

[0016]

[0017] This yields the normalized amino acid frequency matrix of pea protein.

[0018] Preferably, step S2 includes: constructing a multi-layer combined feature extraction neural network to directly extract pea protein features from the voxel grid model, and extracting features from the pea protein voxel grid through the multi-layer combined feature extraction neural network;

[0019] The detailed steps for feature extraction using the structural feature extraction neural network algorithm are as follows:

[0020] S221. The multilayer combined feature extraction neural network is used to extract deep features from pea protein three-dimensional voxel data. The network input is voxel data X = {x, y, z, 0, μ, τ} with a resolution of τ, where x, y, and z represent the spatial coordinates of the voxel, 0 is the origin coordinate, μ is the mapping factor, and τ is the resolution. The network output is the deep features of the voxel model. The multilayer combined feature extraction neural network contains three sub-neural networks:

[0021] The first layer of the feature extraction neural network uses a set of one-dimensional convolutional networks with 3, 5, 7, 9, and 11 kernels to perform convolution operations on the amino acid sequence in the protein to extract features, with 5 input channels;

[0022] After inputting a set of voxel data X, the convolution kernel sequentially performs convolution operations on each amino acid and its adjacent amino acids, that is, multiplies the matrix elements with the convolution kernel and sums the results. Then, it performs a weighted sum based on the obtained feature data and adds the bias.

[0023]

[0024] in, The output of the convolution is represented by l∈{0,1,...,L1}, ​​where L1 represents the number of output channels, and h 0 b represents the input to the convolution. 0 It is the bias, ω 0 This represents the weights of the input data. This represents the weight of the k-th amino acid residue in the input X. Let X represent the k-th amino acid residue in the input X, K0 be the number of input channels, σ∈G={G1,G2,...}, G be the kernel size, and σ represent the size of any kernel.

[0025] Then activate the output of the convolutional network:

[0026]

[0027] in, θ represents the output of all convolutional kernels in the convolutional network. 20-k This represents the activator of amino acids other than the k-th amino acid residue;

[0028] Finally, a batch normalization function is used to prevent the model from overfitting. This formula is for... Perform batch normalization, where BN represents batch normalization, which is the existing formula;

[0029] Finally, the values ​​of G were set to 3, 5, 7, 9, and 11 respectively, thus obtaining the features extracted by the first layer of the feature extraction neural network.

[0030] S222. The second-layer feature extraction neural network has one more convolution, activation, and normalization operation than the first-layer feature extraction neural network;

[0031] S223. The third-layer feature extraction neural network performs one more convolution, activation, and normalization operation than the second-layer feature extraction neural network; for the input X, three convolution, activation, and normalization operations are performed sequentially, and the features extracted by the third-layer feature extraction neural network are:

[0032] Preferably, step S3 includes: combining the pea protein amino acid frequency matrix M obtained in step S1 with the features extracted in steps 221, 222, and 223 to form the input of the fusion feature extraction neural network. There are m input vectors in total; the fusion feature extraction neural network includes an input layer, a hidden layer 1, a hidden layer 2, a fusion layer, and an output layer; the input layer passes the input vector e to the hidden layer, and the hidden layer activates the input vector using the following activation formula: in, f represents the feature representation extracted by the fusion feature extraction neural network at the t-th position based on the features of the first t-1 data points. t-1 u represents the output at the previous time step. r These are the weights of the input vector, r∈m, e t It is the input at the current moment, b r It is a bias; the input layer passes the input vector e to the hidden second layer, and the hidden second layer activates the input vector using the following activation formula: in, The fusion feature extraction neural network represents the feature representation extracted at position t based on the last mt data features; the hidden layer 1 and hidden layer 2 output the calculation results to the fusion layer, and the fusion layer formula of the fusion feature extraction neural network is:

[0033]

[0034] in, Let ∠ represent the fused data, and ⊙ represent the Hadamard product. In this invention, the number of hidden layer neural units is set to q, then the number of features output by the fused network is 2q. The fusion layer sends the fusion result to the output layer for output, thereby obtaining the data feature f.

[0035] Beneficial effects:

[0036] (1) The protein sequence contains all or part of the 20 common amino acids. The amino acids will appear repeatedly in the pea protein sequence and have their own frequencies. Through the exploration of feature extraction and encoding of protein sequence data, sequence data with inconsistent dimensions are encoded into vectors with consistent dimensions, which solves the problem that amino acid sequence data cannot be directly input into machine learning algorithms for recognition. Furthermore, by introducing a frequency matrix in feature extraction, the accuracy and comprehensiveness of feature extraction are enhanced.

[0037] (2) In the multi-layer combined feature extraction neural network, multiple one-dimensional convolutional kernels of different lengths are used to extract amino acid sequence features, enabling the system to extract features between and within amino acid sequences at the same time, further improving the feature extraction capability of the deep architecture.

[0038] (3) By fusing the features extracted from the amino acid frequency matrix of pea protein with the features extracted from the multi-layer combined feature extraction neural network, the features are extracted by working together in terms of sequence and structure, which enhances the depth of pea protein data feature extraction.

[0039] (4) Extraction of amino acid sequence features of pea protein not only directly determines the quality of prediction methods, but also helps us understand the relationship between pea protein sequence and structure and function. Attached Figure Description

[0040] Figure 1 Flowchart of a method for encoding and extracting pea protein data features according to the present invention;

[0041] Figure 2 A schematic diagram of the pea protein sequence information encoding described in this invention;

[0042] Figure 3 The structure diagram of the multi-layer combined feature extraction neural network and the fusion feature extraction neural network described in this invention. Detailed Implementation

[0043] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, thereby enabling a full understanding of how the present invention uses technical means to solve technical problems and achieve technical effects, and allowing for implementation accordingly. It should be noted that, as long as there is no conflict, the various features in the various embodiments of the present invention can be combined with each other, and the resulting technical solutions are all within the protection scope of the present invention.

[0044] like Figure 1 As shown, the pea protein data feature encoding and extraction method of the present invention includes:

[0045] S1. Based on the sequential order of each element in the pea protein sequence data and the frequency of different amino acids, the data is encoded into a frequency distribution matrix with consistent dimension;

[0046] S11. For any pea protein sequence, the following three layers of information are contained: the protein sequence contains all or some of the 20 common amino acids; the amino acids appear repeatedly in the pea protein sequence and have their own frequencies; the amino acids are located at different positions in the pea protein sequence. An amino acid with a frequency of zero indicates that the protein sequence does not contain that amino acid, and the frequency of the amino acid corresponds to the number of its positions.

[0047] Any pea protein sequence consists of a linear sequence of its amino acid units. Dividing the pea protein sequence starting from the first amino acid yields two subsequences: the preceding amino acid and the following amino acid. Based on the one-to-one correspondence between the amino acid sequence and the following amino acid subsequence, an amino acid frequency distribution vector is obtained. This vector is then used to encode the amino acid positions of all amino acids in the pea protein sequence, representing the frequency of occurrence of all amino acids in the sequence. The table below illustrates this process.

[0048]

[0049] The amino acid frequency distribution vector in the subsequence "LTRALSC" of the sequence "ELTRALSC" is (0,2,1,1,1,1,1), and the amino acid frequency distribution vector in the subsequence "TRALSC" of the sequence "ELTRALSC" is (0,1,1,1,1,1,1,1). By subtracting the amino acid frequency distribution vectors of adjacent subsequences, a difference vector is obtained. This difference vector represents the different amino acids contained in the adjacent subsequences, and the position information of these amino acids in the protein sequence is obtained. After traversal, the encoded amino acid position information is combined by vector summation to merge the codes for the same amino acid at different positions, such as... Figure 2 As shown, this yields the distribution vector of the frequencies of all amino acids in the pea protein sequence, recording the position information of amino acids in the pea protein sequence, thus converting the list of amino acid positions in the pea protein sequence into vector codes.

[0050] S12. The method for calculating the amino acid frequency distribution matrix is ​​as follows:

[0051] Let the set of amino acids contained in pea protein be Ω = {Ω1, Ω2, ..., Ω}. N}, Ω N Let N be the Nth amino acid, and N be the number of amino acid types in pea protein. Mapping a pea protein amino acid sequence with P amino acid residues to a point in a 20-dimensional Euclidean space is represented by the vector S∈{s1, s2, ..., s...} P}, where P is the number of elements in a given amino acid sequence. For amino acid Ω N Based on the frequency of occurrence in sequence S, the amino acid frequency matrix of pea protein is as follows:

[0052] M = [m ij ] N×N

[0053]

[0054] Where i, j ∈ [1, N]. Then the sum of all elements in the sequence is... Normalize the elements in the sequence using iteration:

[0055]

[0056] This yields the normalized amino acid frequency matrix of pea protein.

[0057] The beneficial effects of the amino acid frequency vector encoding method described in step S1 are as follows: Through the exploration of feature extraction and encoding of protein sequence data, sequence data with inconsistent dimensions are encoded into vectors with consistent dimensions, solving the problem that amino acid sequence data cannot be directly input into machine learning algorithms for recognition. Furthermore, by introducing a frequency matrix into feature extraction, the accuracy and comprehensiveness of feature extraction are enhanced.

[0058] S2. By mapping the 3D model of pea protein to a voxel mesh model, feature extraction of pea protein based on the 3D structure is achieved;

[0059] S21. Collect pea protein data and construct a three-dimensional model of pea protein. The method for constructing the three-dimensional model is existing technology and will not be elaborated on here.

[0060] In the 3D model representation of pea protein, each point (x, y, z) constituting the protein is mapped to discrete voxel coordinates. The mapping method is uniform discretization, and the mapping process depends on parameters such as the origin, orientation, and resolution of the voxel grid. After determining the initial parameters such as the origin, orientation, and resolution according to the actual situation, a protein voxel grid model is constructed.

[0061] S22. Extract features from the pea protein voxel grid model to obtain the three-dimensional structural feature matrix of pea protein.

[0062] Construct a multi-layer combined feature extraction neural network to directly extract pea protein features from a voxel grid model, such as... Figure 3 As shown, features are extracted from the above pea protein voxel grid using a multi-layer combined feature extraction neural network.

[0063] The detailed steps for feature extraction using the structural feature extraction neural network algorithm are as follows:

[0064] S221. The multilayer combined feature extraction neural network is used to extract deep features from pea protein three-dimensional voxel data. The network input is voxel data X = {x, y, z, 0, μ, τ} with a resolution of τ, where x, y, and z represent the spatial coordinates of the voxel, 0 is the origin coordinate, μ is the mapping factor, and τ is the resolution. The network output is the deep features of the voxel model. The multilayer combined feature extraction neural network contains three sub-neural networks:

[0065] The first layer of the feature extraction neural network uses a set of one-dimensional convolutional networks with 3, 5, 7, 9, and 11 kernels to perform convolution operations on the amino acid sequence in the protein to extract features, with 5 input channels.

[0066] Taking a one-dimensional convolutional network with 3 kernels as an example, feature extraction is performed on the input data. Each element of the convolutional kernel corresponds to 3 weight coefficients and one bias. After inputting a set of voxel data X, the convolutional kernel sequentially performs convolution operations on each amino acid and its adjacent amino acids, that is, multiplies the matrix elements with the convolutional kernel and sums the results. Then, the resulting feature data is weighted and summed, and the bias is added.

[0067]

[0068] in, The output of the convolution is represented by l∈{0,1,...,L1}, ​​where L1 represents the number of output channels, and h 0 b represents the input to the convolution. 0 It is the bias, ω 0 This represents the weights of the input data. This represents the weight of the k-th amino acid residue in the input X. Let K0 represent the k-th amino acid residue in the input X, K0 be the number of input channels, σ ∈ G = {G1, G2, ...}, G be the kernel size, and σ represent the size of any convolutional kernel. Then, the output of the convolutional network is activated:

[0069]

[0070] Where, θ 20-k This represents the activator of all amino acids except the k-th amino acid residue. Finally, a batch normalization function is used to prevent overfitting of the model. Where BN represents batch normalization. Finally, the values ​​of G are set to 3, 5, 7, 9, and 11 respectively, thus obtaining the features extracted by the first-layer feature extraction neural network.

[0071] S222. The second-layer feature extraction neural network performs one more convolution, activation, and normalization operation than the first-layer feature extraction neural network. Taking a one-dimensional convolutional network with 3 kernels as an example, the input X undergoes two convolution, activation, and normalization operations sequentially:

[0072]

[0073]

[0074]

[0075]

[0076]

[0077]

[0078] in, b represents the output of the second convolution. 1 It is the bias in the second convolution, ω 1 K1 represents the input weights in the second convolution, and L2 represents the number of input channels in the second convolution. In the second convolution, l ∈ {0, 1, ..., L2}, and L2 represents the number of output channels in the second convolution. The output of the second activation, This is the output of the second normalization.

[0079] Finally, the values ​​of G were set to 3, 5, 7, 9, and 11 respectively, resulting in the features extracted by the second-layer feature extraction neural network.

[0080] S223. The third-layer feature extraction neural network performs one more convolution, activation, and normalization operation than the second-layer feature extraction neural network. For the input X, three convolution, activation, and normalization operations are performed sequentially. According to the above calculation process, the features extracted by the third-layer feature extraction neural network are:

[0081] The beneficial effect of the multi-layer combined feature extraction neural network described in step S2 is that multiple one-dimensional convolutional kernels of different lengths are used to extract amino acid sequence features, enabling the system to extract features between and within amino acid sequences simultaneously, further improving the feature extraction capability of deep architecture.

[0082] S3. Combine the pea protein amino acid frequency matrix with the features extracted by the multilayer combined feature extraction neural network, and extract features from the pea protein data by fusing the feature extraction neural network.

[0083] S31. Combine the pea protein amino acid frequency matrix M obtained in step S1 with the features extracted in steps 221, 222, and 223 to form the input of the fusion feature extraction neural network. There are m input vectors in total. The fusion feature extraction neural network consists of an input layer, a hidden layer 1, a hidden layer 2, a fusion layer, and an output layer.

[0084] The input layer passes the input vector e to the hidden layer, which then activates the input vector using the following activation formula:

[0085]

[0086] in, f represents the feature representation extracted by the fusion feature extraction neural network at the t-th position based on the features of the first t-1 data points. t-1 u represents the output at the previous time step. r These are the weights of the input vector, r∈m, e t It is the input at the current moment, b r It is the bias value.

[0087] The input layer passes the input vector e to the hidden layer 2, which then activates the input vector using the following activation formula:

[0088]

[0089] in, This represents the feature representation extracted by the fusion feature extraction neural network at the t-th position based on the last mt data features.

[0090] S32. The hidden layer 1 and hidden layer 2 output the calculation results to the fusion layer. Then the fusion layer formula of the fusion feature extraction neural network is:

[0091]

[0092] in, Let ∠ represent the fused data, and ⊙ represent the Hadamard product. In this invention, the number of hidden layer neural units is set to q, then the number of features output by the fused network is 2q. The fusion layer sends the fusion result to the output layer for output, thereby obtaining the data feature f.

[0093] The training effect of the fusion feature extraction neural network is tested. An error threshold ε is set according to actual needs, and the network error is calculated. There are many methods for calculating the network error; the appropriate method is selected according to actual needs. If error < ε, the actual needs are considered met, and training is complete; otherwise, the network is retrained.

[0094] The beneficial effect of the fusion feature extraction neural network described in step S3 is that by fusing the features extracted from the amino acid frequency matrix of pea protein with those extracted by the multilayer combined feature extraction neural network, the features are extracted in both sequence and structure aspects, thereby enhancing the depth of pea protein data feature extraction.

[0095] The above description is merely a preferred embodiment of the present invention and is not intended to limit this application. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this application should be included within the protection scope of this application. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution, or the direct application of the inventive concept and technical solution to other situations without modification, are all within the protection scope of the present invention.

Claims

1. A method for encoding and extracting pea protein data features, characterized in that, The method includes the following steps: S1. Based on the sequential order of each element in the pea protein sequence data and the frequency of different amino acids, the data is encoded into a frequency distribution matrix with consistent dimension; S2. By mapping the 3D model of pea protein to a voxel grid model, a multi-layer combined feature extraction neural network is constructed to directly extract pea protein features from the voxel grid model. The multi-layer combined feature extraction neural network is used to extract features from the above-mentioned pea protein voxel grid, thereby realizing pea protein feature extraction based on the 3D structure. S3. Combine the amino acid frequency distribution matrix of pea protein with the features extracted by the multilayer combined feature extraction neural network, and extract features from the pea protein data by fusing the feature extraction neural network; Step S1 includes: The pea protein sequence is divided into two subsequences, starting from the first amino acid and ending with it. Based on the one-to-one correspondence between the amino acid sequence and the subsequence, an amino acid frequency distribution vector is obtained. This vector is then used to traverse the subsequence amino acid frequency distribution vectors corresponding to all amino acids in the pea protein sequence, encoding the amino acid positions. The amino acid frequency distribution vector is a vector composed of the frequencies of all amino acids in the pea protein amino acid sequence. By subtracting the amino acid frequency distribution vectors corresponding to adjacent subsequences, a difference vector is obtained, representing the different amino acids contained in the adjacent subsequences and their positions in the protein sequence. After traversal, the encoded amino acid positions are combined using vector summation to obtain the distribution vector of the frequencies of all amino acids in the pea protein sequence, recording the position information of the amino acids in the pea protein sequence. This process converts the list of amino acid positions in the pea protein sequence into a vector code.

2. The method for encoding and extracting pea protein data features according to claim 1, characterized in that, Step S1 includes: a method for calculating the amino acid frequency distribution matrix, wherein the calculation method is as follows: The amino acid set contained in pea protein is defined as follows: , Let N be the Nth amino acid, and N be the number of amino acid types in pea protein. Mapping a pea protein amino acid sequence with P amino acid residues to a point in a 20-dimensional Euclidean space is represented by a vector. P is the number of elements in a given amino acid sequence. amino acids Based on the frequency of occurrence in sequence S, the amino acid frequency distribution matrix of pea protein is as follows: ， ， in, Then the sum of all elements in the sequence is Normalize the elements in the sequence: ， Thus, the normalized amino acid frequency distribution matrix of pea protein was obtained.

3. The method for encoding and extracting pea protein data features according to claim 2, characterized in that, Step S2 includes: constructing a multi-layer combined feature extraction neural network to directly extract pea protein features from the voxel grid model, and extracting features from the above-mentioned pea protein voxel grid through the multi-layer combined feature extraction neural network; The detailed steps of feature extraction are as follows: S221. The multilayer combined feature extraction neural network is used to extract deep features from the three-dimensional voxel data of pea protein. The network input is a resolution of [missing information]. voxel data ,in, The three dimensions represent the spatial coordinates of a voxel, with O being the origin coordinate. As the mapping factor, To achieve the desired resolution, the network output represents the deep features of the voxel model. The multi-layer combined feature extraction neural network comprises three sub-neural networks: The first feature extraction neural network uses a set of one-dimensional convolutional networks with kernel numbers of 3, 5, 7, 9, and 11 to perform convolution operations on the amino acid sequence in the protein to extract features, with 5 input channels; after inputting a set of voxel data X, the convolutional kernel sequentially performs convolution operations on each amino acid and its adjacent amino acids, that is, multiplies the matrix elements with the convolutional kernel and sums the results, then performs a weighted sum based on the obtained feature data and adds a bias: ， in, This represents the output of the convolution. , Indicates the number of output channels. This represents the input to the convolution. It is the bias value. This represents the weights of the input data. This represents the weight of the k-th amino acid residue in the input X. This represents the k-th amino acid residue in input X. This is the number of input channels. , It is the kernel size. This represents the size of any convolutional kernel; Then activate the output of the convolutional network: ， in, This represents the output of all convolutional kernels in the convolutional network. This represents the activators of amino acids other than the k-th amino acid residue; finally, a batch normalization function is used to prevent overfitting of the model. This formula is for Batch normalization is performed; BN represents batch normalization, and this is the existing formula. Finally, the values ​​of G are set to 3, 5, 7, 9, and 11 respectively, thus obtaining the features extracted by the first-layer feature extraction neural network. ; S222. The second-layer feature extraction neural network has one more convolution, activation, and normalization operation than the first-layer feature extraction neural network; S223. The third-layer feature extraction neural network performs one more convolution, activation, and normalization operation than the second-layer feature extraction neural network; for the input X, three convolution, activation, and normalization operations are performed sequentially, and the features extracted by the third-layer feature extraction neural network are: 。 4. The method for encoding and extracting pea protein data features according to claim 3, characterized in that, Step S3 includes: combining the pea protein amino acid frequency distribution matrix M obtained in step S1 with the features extracted in steps 221, 222, and 223 to form the input of the fusion feature extraction neural network. There are a total of m input vectors; The fusion feature extraction neural network consists of an input layer, a hidden layer 1, a hidden layer 2, a fusion layer, and an output layer. The input layer passes the input vector e to the hidden layer, which then activates the input vector using the following activation formula: ; in, This represents the feature representation extracted by the fusion feature extraction neural network at the t-th position based on the features of the first t-1 data points. This represents the output at the previous time step. These are the weight values ​​of the input vector. , It is the input at the current moment. It is the bias value; The input layer passes the input vector e to the hidden layer 2, which then activates the input vector using the following activation formula: ; in, This represents the feature representation extracted by the fusion feature extraction neural network at the t-th position based on the features of the last mt data. The hidden first and hidden second layers output the calculation results to the fusion layer. Therefore, the fusion layer formula for the fusion feature extraction neural network is: ， in, Indicates data fusion. For the Hadamard product, if the number of hidden layer neural units is q, then the number of features output by the network after fusion is 2q. The fusion layer sends the fusion result to the output layer for output, thereby obtaining the data feature f.

Images