Industrial sewage water quality soft measurement method

Abstract

The invention discloses an industrial sewage water quality soft measurement method. The method comprises the following steps: collecting sewage water quality data of a sewage plant; carrying out feature extraction on the input data by adopting an AHSICLasso method; decomposing the variable signal into a plurality of symplectic geometric components by using symplectic geometric mode decomposition (SGMD), and then performing secondary decomposition on the decomposed high-frequency nonlinear components through singular spectrum analysis (SSA); carrying out complexity quantification on the decomposed multi-mode component by utilizing approximate entropy so as to evaluate the dynamic characteristics of the multi-mode component; according to a quantization result, reconstructing the multi-mode component; using a time step adaptive dynamic selection mechanism and approximate entropy to screen components at past moments, and using CBO to optimize component weights; the AHSICLasso feature extraction data, the screened components at the past moment and the components obtained after secondary decomposition are input into an OfficANet model; the method comprises the following steps: optimizing hyper-parameters of an EfficANet model by using a CBO collider, introducing an adaptive attention weight mechanism into a GTVA module of the EfficANet model for improvement, and learning and predicting a reconstructed multi-modal component to realize soft measurement of total nitrogen in industrial sewage; according to the invention, high-precision and real-time prediction of the total nitrogen concentration of the industrial sewage is realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of nitrogen concentration prediction, and particularly relates to an industrial wastewater quality soft measurement method. BACKGROUND

[0002] In recent years, a soft measurement technology has the advantages of low cost and fast speed by establishing a mathematical model and using easily measured variables to predict difficult-to-directly-measured target variables. However, the existing soft measurement method still has defects in processing industrial wastewater total nitrogen concentration prediction. On the one hand, these methods often fail to fully excavate the complex nonlinear relationship between multiple potential variables in the wastewater and the total nitrogen concentration; on the other hand, for complex industrial wastewater signals, the existing method lacks effective feature extraction and signal reconstruction means, resulting in limited model prediction accuracy and difficulty in meeting the demand of high-precision real-time monitoring. SUMMARY

[0003] The application aims to provide an industrial wastewater quality soft measurement method, and solve the problem of limited model prediction accuracy and difficulty in meeting the demand of high-precision real-time monitoring caused by the lack of effective feature extraction and signal reconstruction means.

[0004] TECHNICAL SOLUTION The application provides an industrial wastewater quality soft measurement method, which comprises the following steps:

[0005] (1) collecting wastewater quality data of a wastewater treatment plant;

[0006] (2) performing feature extraction on input data by using an AHSICLasso method;

[0007] (3) performing decomposition of a variable signal into multiple symplectic geometric components by using a symplectic geometric modal decomposition (SGMD), and then performing secondary decomposition of the decomposed high-frequency nonlinear components by using a singular spectrum analysis (SSA) to extract more fine features;

[0008] (4) performing complexity quantification on the decomposed multi-modal components by using an approximate entropy to evaluate the dynamic characteristics of the multi-modal components; and reconstructing the multi-modal components according to the quantification results;

[0009] (5) screening the components at past time points by using a time step adaptive dynamic selection mechanism and an approximate entropy, and optimizing the component weights by using a CBO; inputting the AHSICLasso feature extraction data, the screened components at past time points and the secondary decomposed components into an EfficANet model;

[0010] (6) optimizing the hyperparameters of the EfficANet model by using a CBO collision body, introducing an adaptive attention weight mechanism into a GTVA module of the EfficANet model for improvement, learning and predicting the reconstructed multi-modal components, and realizing soft measurement of the total nitrogen of the industrial wastewater.

[0011] Furthermore, the formula for step (2) is as follows:

[0012]

[0013] stα1,…,α d ≥0

[0014] Where α=[α1,…α d ] T It is the regression coefficient vector; α k These are the regression coefficients of the k-th feature; ||·||1 and ||·|| Frob These are the L1 and Frobenius norms; λ>0 is the regularization parameter; and It is a centered gram matrix; R d Let represent the d-dimensional real space.

[0015] Furthermore, step (3) includes the following steps:

[0016] (31) Input the multivariate time-domain signal y of wastewater quality, and construct the trajectory matrix X based on the original time series:

[0017]

[0018] Where L (2≤L≤N / 2) is the window length, and the number of columns in the matrix is ​​N-L+1;

[0019] (32) Perform singular value decomposition on matrix X and generate r submatrices X based on the singular values. i :

[0020]

[0021] Where r represents the number of non-zero eigenvalues ​​of matrix X, i.e., the rank of the matrix, and σ i U represents the i-th singular value in matrix factorization. i V represents the left singular vector corresponding to the i-th singular value. i T This represents the right singular vector corresponding to the i-th singular value;

[0022] (33) Submatrix X i Divided into m groups, for X i Diagonal meanization is performed to obtain the subsequence By superimposing subsequences with the same signal, three special sequences representing trends, low-frequency periods, and high-frequency noise signals are obtained:

[0023]

[0024] Among them, L * =min(L,N-L+1),K * =max(L,N-L+1), v∈(1,N).

[0025] (34) Subsequence The high-frequency components in the sequence are subjected to symplectic geometric mode decomposition, and the subsequences are reconstructed in phase space.

[0026]

[0027] Where x is the high-frequency component sequence, d is the embedding dimension, τ is the delay time, m = n - (d - 1)τ, and n is the length of sequence x.

[0028] (35) The formula for constructing the symplectic orthogonal matrix Q is as follows:

[0029]

[0030] Where N = M 2 Let R be a Hammer matrix, R be the transformed submatrix, and B be an upper triangular matrix with eigenvalues ​​λ1, λ2, λ3, ..., λ4. d The covariance matrix A = X T X,

[0031] (36) The sequence for signal reconstruction is as follows:

[0032]

[0033] Among them, z ij (1≤i≤m;1≤j≤d) are elements in the reconstructed signal matrix, where d ★ =min(m,d), m * =max(m,d), v=m+(d-1)τ.

[0034] 4. The method for soft measurement of industrial wastewater quality according to claim 1, characterized in that step (4) includes the following steps:

[0035] (41) Selecting the embedding dimension and tolerance The embedding dimension is selected as 2 or 3, and the tolerance is selected as 0.1 to 0.2 times the standard deviation of the time series.

[0036] (42) The time series X = {x1, x2, ..., x...} is... N} construct dimensional vector

[0037]

[0038] in,

[0039] (43) D ij The maximum distance between components is represented as:

[0040]

[0041] in,

[0042] Count the number of vectors within the tolerance range.

[0043]

[0044] Where θ is the unit step function, the above process is repeated, but the embedding dimension is increased to . calculate

[0045] (44) Calculate the approximate entropy

[0046]

[0047] Furthermore, in step (5), the time step adaptive dynamic selection mechanism automatically adjusts the time step according to the dynamic characteristics of each component.

[0048] Furthermore, step (6) includes the following steps:

[0049] (11) The EfficANet model receives multivariate time series data of wastewater quality X∈R. M×T ; and transform it into a suitable feature space through patching and embedding strategies; reshape it into X through the unsqueeze operation. in ∈R M×1×T , where M is the number of variables and H is the length of the time series;

[0050] (62) Use convolutional dry layers to segment the input sequence into multiple overlapping patches, and process the patches to extract local features.

[0051] (63) The EfficANet model consists of multiple stacked blocks, each containing three key modules: TLDC temporal local convolution, IVGC variable group convolution, and GTVA global temporal-variable attention; an adaptive attention weight mechanism is introduced in the GTVA module to capture the complex relationships in the temporal and variable dimensions.

[0052] (64) After processing by multiple sub-modules, the output is used to generate the final prediction result through the prediction head.

[0053]

[0054] Among them, Z (l) This is the output of the l-th block. This indicates the operation within the l-th block.

[0055] Furthermore, step (63) includes the following steps:

[0056] (631) Temporal Local Convolution (TLDC): Input tensor shape is X emb The input tensor, with a shape of (N,C,T), is reshaped into (N×C,T) groups and divided into N×C groups.

[0057] Applying depthwise convolution (DW) to capture short-term dependencies:

[0058] X local =DW Conv(X emb )

[0059] Applying depthwise dilated convolution (DW-D) to expand the receptive field:

[0060] X dilated =DW-D Conv(X local )

[0061] Add the outputs of the two convolutions:

[0062] X combined =X dilated +X local

[0063] (632) The IVGC module handles the temporal dependency of each variable independently through grouped convolutions: input tensor X combined The shape is (N,C,T), with the time dimension T filled with multiples of the time window size W, and grouped as follows. Groups;

[0064] Calculate the fill length:

[0065]

[0066] Calculate the start and end fill length P left and P right :

[0067]

[0068] Perform 1D grouped convolutions on the tensors of standard padding and head-and-tail padding respectively, and then align and merge the outputs of the two convolutions:

[0069] Y = Conv(Conv(X) padded1 )+Conv(X padded2 ))

[0070] (633) The GTVA module captures the dependencies between time and variables through a global time and variable attention mechanism; it includes two parts: time attention and variable attention. The features on the time axis and variable axis are processed by a fully connected network, and then the two attentions are fused by the Hadamard product.

[0071] (644) An adaptive attention weight mechanism is introduced in the GTVA module to capture the complex relationship between time and variable dimensions: the input tensor has a Y shape and a shape of (N,C,T), the input tensor is reshaped into (N×C,T), and global average pooling is applied:

[0072] T pool =AvgPool(Y temp )

[0073] Temporal attention weights are generated using a two-layer fully connected network:

[0074] T atten =σ(W2·ReLU(W1·T) pool ))

[0075] Reshape the input tensor into (N×T,C) and apply global average pooling:

[0076] V pool =AvgPool(Y var )

[0077] Variable attention weights are generated using a two-layer fully connected network.

[0078] V atten =σ(W4·ReLU(W3·V) pool ))

[0079] Multiply the temporal attention and variable attention weights by the convolution output:

[0080] Y out =σ(T) atten ⊙V atten ⊙Y)

[0081] Multiply the GTVA output by the block input to introduce a feedback mechanism:

[0082] X′ emb =Y out ⊙X emb .

[0083] Furthermore, in step (1), the water quality data variables include total nitrogen (TN), total phosphorus (TP), chemical oxygen demand (COD), biochemical oxygen demand (BOD), pH value, and dissolved oxygen (DO).

[0084] An electronic device according to the present invention includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to implement the steps of any of the methods described herein.

[0085] The present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of any of the methods described herein.

[0086] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: through advanced variable screening, signal decomposition, feature extraction and optimization reconstruction methods, it achieves high-precision and real-time prediction of total nitrogen concentration in industrial wastewater, providing an efficient and reliable solution for industrial wastewater treatment and environmental monitoring. Attached Figure Description

[0087] Fig. 1 This is a flowchart of the method of the present invention;

[0088] Fig. 2 This is a structural diagram of the AHSIClssao invention;

[0089] Fig. 3 This is a structural diagram of the EfficANet model of the present invention. Detailed Implementation

[0090] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0091] like Figs. 1-3 As shown, this embodiment of the invention provides a soft measurement method for industrial wastewater quality, including the following steps:

[0092] Step 1: Collect wastewater quality data from the wastewater treatment plant. The water quality data variables include total nitrogen (TN), total phosphorus (TP), chemical oxygen demand (COD), biochemical oxygen demand (BOD), pH value, and dissolved oxygen (DO), etc.

[0093] Step 2: Use the AHSIClssao method to extract features from the input data. Improve the accuracy of feature extraction by adjusting parameters and using more training data.

[0094] AHSIClssao is a nonlinear Lasso method used for regularized regression analysis in feature selection. Its optimization method is as follows:

[0095]

[0096] stα1,…,α d ≥0

[0097] Where α=[α1,…α d ] TIt is the regression coefficient vector; α k These are the regression coefficients of the k-th feature; ||·||1 and ||·|| Frob These are the L1 and Frobenius norms; λ>0 is the regularization parameter; and It is a centered gram matrix; R d Let represent the d-dimensional real space.

[0098] Step 3: The steps for secondary decomposition using symplectic geometric mode decomposition (SGMD) and singular spectral analysis (SSA) are as follows:

[0099] Step 3.1: Input a multivariate time-domain signal y for wastewater quality, and construct the trajectory matrix X based on the original time series:

[0100]

[0101] Where L (2≤L≤N / 2) is the window length, and the number of columns in the matrix is ​​N-L+1.

[0102] Step 3.2: Perform singular value decomposition on matrix X, and generate r submatrices X based on the singular values. i :

[0103]

[0104] Where r represents the number of non-zero eigenvalues ​​of matrix X, i.e., the rank of the matrix, and σ i U represents the i-th singular value in matrix factorization. i V represents the left singular vector corresponding to the i-th singular value. i T Let represent the right singular vector corresponding to the i-th singular value. 2

[0105] Step 3.3: Submatrix X i Divided into m groups, for X i Diagonal meanization is performed to obtain the subsequence By superimposing subsequences with the same signal, three special sequences representing trends, low-frequency periods, and high-frequency noise signals are obtained:

[0106]

[0107] Among them, L * =min(L,N-L+1),K * =max(L,N-L+1), v∈(1,N).

[0108] Step 3.4: Subsequence The high-frequency components in the sequence are subjected to symplectic geometric mode decomposition, and the subsequences are reconstructed in phase space.

[0109]

[0110] Where x is the high-frequency component sequence, d is the embedding dimension, τ is the delay time, m = n - (d - 1)τ, and n is the length of sequence x.

[0111] Step 3.5: The formula for constructing the symplectic orthogonal matrix Q is as follows:

[0112]

[0113] Where N = M 2 Let R be a Hammer matrix, R be the transformed submatrix, and B be an upper triangular matrix with eigenvalues ​​λ1, λ2, λ3, ..., λ4. d The covariance matrix A = X T X,

[0114] Step 3.6: The signal reconstruction sequence is as follows:

[0115]

[0116] Among them, z ij (1≤i≤m;1≤j≤d) are elements in the reconstructed signal matrix, where d ＊ =min(m,d), m * =max(m,d), v=m+(d-1)τ.

[0117] Step 4: The steps for approximate entropy to perform complex metric quantification on the decomposed multimodal components are as follows:

[0118] Step 4.1: Select the embedding dimension and tolerance Embedding dimension is typically chosen to be 2 or 3, and tolerance is typically chosen to be 0.1 to 0.2 times the standard deviation of the time series.

[0119] Step 4.2: Convert the time series X = {x1, x2, ..., x...} N} construct dimensional vector

[0120]

[0121] in,

[0122] Step 4.3: D ij The maximum distance between components is represented as:

[0123]

[0124] in,

[0125] Count the number of vectors within the tolerance range.

[0126]

[0127] Where θ is the unit step function, the above process is repeated, but the embedding dimension is increased to . calculate

[0128] Step 4.4: Calculate the approximate entropy

[0129]

[0130] Step 5: The adaptive dynamic selection mechanism for time step automatically adjusts the time step based on the dynamic characteristics of each component, aiming to improve computational efficiency while ensuring computational accuracy. By dynamically assigning weight values ​​to the input features, which are SGCs (feature components of mode decomposition) obtained through SGMD mode decomposition, the model explores the relationship between SGCs and the original energy consumption data.

[0131] By assigning adaptive dynamic weights to the time information carried in each historical time step, this mechanism can distinguish the impact of the current signal on the prediction result of the current time step, while autonomously extracting time information from the historical time steps of the data. This allows the key information of the current data to be expressed, ultimately improving the accuracy of the prediction. The time step adaptive dynamic selection mechanism and approximate entropy are used to filter components from past moments, and CBO is used to optimize the component weights. The AHSICLasso feature extraction data, the filtered components from past moments, and the components after secondary decomposition are input into the EfficANet model.

[0132] Step 6: The steps for optimizing the hyperparameters of the EfficANet model using the CBO collider are as follows:

[0133] Step 6.1: The EfficANet model receives a multivariate time series of wastewater quality X∈R. M×T The input is then transformed into a suitable feature space through a patching and embedding strategy. The input is first reshaped into X through an unsqueeze operation. in ∈R M×1×T , where M is the number of variables and H is the length of the time series.

[0134] Step 6.2: Convolutional Dry Layer: The input sequence is segmented into multiple overlapping patches using a convolutional dry layer, which are then processed to extract local features.

[0135] Step 6.3: The core structure of EfficANet consists of multiple stacked blocks, each containing three key modules: TLDC (Temporal Local Convolution), IVGC (Inverted Group Convolution), and GTVA (Global Temporal-Variable Attention). These modules work together to capture the complex relationships in the temporal and variable dimensions.

[0136] Step 6.3.1: Temporal Local Convolution (TLDC)

[0137] The input tensor has a shape of X. emb The input tensor, with a shape of (N,C,T), is reshaped into (N×C,T) groups. A depthwise convolution (DW Conv) is then applied to capture short-term dependencies.

[0138] X local =DW Conv(X emb )

[0139] Applying depthwise dilated convolution (DW-D Conv) to expand the receptive field:

[0140] X dilated =DW-D Conv(X local )

[0141] Add the outputs of the two convolutions:

[0142] X combined =X dilated +X local

[0143] Step 6.3.2: Group Convolution (IVGC): The IVGC module handles the temporal dependencies of each variable independently through group convolution, allowing parallel computation while preserving the unique characteristics of each variable.

[0144] Input tensor X combined The shape is (N,C,T), with the time dimension T filled with multiples of the time window size W, and grouped as follows. Groups.

[0145] Calculate fill length

[0146]

[0147] Calculate the start and end fill length P left and P right

[0148]

[0149] Perform 1D grouped convolutions on the tensors of standard padding and head-and-tail padding respectively, and then align and merge the outputs of the two convolutions.

[0150] Y = Conv(Conv(X) padded1 )+Conv(X padded2 ))

[0151] Step 6.3.3: Global Time-Varying Attention (GTVA): The GTVA module captures the dependencies between time and variables through a global time and variable attention mechanism. It consists of two parts: time attention and variable attention. The time axis and variable axis features are processed separately through a fully connected network. Then, the two attentions are fused through Hadamard product (element-wise multiplication) to simultaneously emphasize time- and variable-related features.

[0152] An adaptive attention weighting mechanism is introduced in the GTVA module, allowing attention weights to be dynamically adjusted based on the features of the input data. The original input features are passed to the attention weight generation module to obtain dynamically generated attention weights. The generated attention weights are multiplied by the original input features to weight the features, thereby enhancing important features and suppressing unimportant features. A global temporal-variable attention mechanism, such as multi-head attention, is applied to the weighted features to capture complex relationships in the temporal and variable dimensions.

[0153] Given an input tensor of shape Y and shape (N,C,T), reshape the input tensor to (N×C,T) and apply global average pooling:

[0154] T pool =AvgPool(Y temp )

[0155] Temporal attention weights are generated using a two-layer fully connected network:

[0156] T atten =σ(W2·ReLU(W1·T) pool ))

[0157] Reshape the input tensor into (N×T,C) and apply global average pooling:

[0158] V pool =AvgPool(Y var )

[0159] Variable attention weights are generated using a two-layer fully connected network.

[0160] V atten =σ(W4·ReLU(W3·V) pool ))

[0161] Multiply the temporal attention and variable attention weights by the convolution output:

[0162] Y out =σ(T) atten ⊙Vatten ⊙Y)

[0163] Multiply the GTVA output by the block input to introduce a feedback mechanism:

[0164] X′ emb =Y out ⊙X emb

[0165] Step 6.4: Prediction Head: After processing by multiple sub-modules, the output is generated by a prediction head to produce the final prediction result.

[0166]

[0167] Among them, Z (l) This is the output of the l-th block. This indicates the operation within the l-th block.

Claims

1. An industrial wastewater water quality soft-sensing method, characterized in that, The method comprises the following steps: (1) collecting sewage quality data of a sewage plant; (2) performing feature extraction on input data by using an AHSICLasso method; (3) decomposing a variable signal into multiple symplectic geometric components by using symplectic geometric modal decomposition (SGMD), and then performing secondary decomposition on the decomposed high-frequency nonlinear components by using singular spectrum analysis (SSA) to extract more fine features; (4) quantifying the complexity of the decomposed multi-modal components by using approximate entropy to evaluate the dynamic characteristics thereof; and reconstructing the multi-modal components according to the quantification results; (5) screening the components at past time points by using a time step adaptive dynamic selection mechanism and approximate entropy, and optimizing the component weights by using CBO; inputting the AHSICLasso feature extraction data, the screened components at past time points and the secondary decomposed components into an EfficANet model; (6) optimizing the hyperparameters of the EfficANet model by using a CBO collision body, introducing an adaptive attention weight mechanism into a GTVA module of the EfficANet model for improvement, learning and predicting the reconstructed multi-modal components, and realizing soft measurement of total nitrogen in industrial sewage.

2. The industrial wastewater quality soft-sensing method according to claim 1, characterized in that, The formula of step (2) is as follows: s.t. α1, …, α d ≥ 0 where a = [a1,... a d ] T is a vector of regression coefficients; a k is the regression coefficient of the kth feature; ||·||1and ||·||F Frob are the L1and Frobenius norms; l > 0 is a regularization parameter; and is a centered gram matrix; R d denotes the d-dimensional real space.

3. The industrial wastewater quality soft-sensing method according to claim 1, characterized in that, Step (3) comprises the following steps: (31) inputting a sewage quality multivariate time domain signal y, and constructing a trajectory matrix X according to an original time sequence: wherein L (2≤L≤N / 2) is a window length, and the column number of the matrix is N-L+1; (32) performing singular value decomposition on the matrix X and generating r sub-matrices X i : where r denotes the number of non-zero eigenvalues of the matrix X, i.e., the rank of the matrix, σ i represents the i-th singular value in the matrix decomposition, U i represents the i-th singular value, V i T represents the i-th singular value, V (33) The sub-matrix X i is divided into m groups, and the diagonal mean value processing is performed on X i to obtain a sub-sequence The sub-sequences with the same signal are superimposed to obtain three special sequences representing the trend, low-frequency period, and high-frequency noise signal: where L * = min(L, N - L + 1), K * = max(L, N - L + 1), v e (1, N). (34) performing a symplectic geometric mode decomposition on high frequency components in the subsequence and phase space reconstruction of the subsequence: wherein x is a high-frequency component sequence, d is an embedding dimension, τ is a delay time, m=n-(d-1)τ, and n is the length of the sequence x. The formula for constructing the symplectic orthogonal matrix Q is as follows: where N = M 2 is a Hamming matrix, R is a transformed sub-matrix, B is an upper triangular matrix, and the eigenvalues thereof are λ1, λ2, λ3,..., λ d , the covariance matrix A = X T X, The sequence for signal reconstruction is as follows: where z ij are the elements in the reconstructed signal matrix, d ★ = min(m, d), m * = max(m, d), v = m + (d - 1)τ.

4. The industrial wastewater quality soft-sensing method according to claim 1, characterized in that, The step (4) comprises the following steps: (41) select embedding dimension and tolerance where embedding dimension is selected as 2 or 3, and tolerance is selected as 0.1 to 0.2 times the standard deviation of the time series; (42) The time series X = {x1, x2,..., x N} is constructed from a vector wherein (43) D ij is expressed as the maximum distance between components wherein Number of vectors within tolerance where θ is the unit step function, and the above process is repeated but with the embedding dimension increased to Computing (44) Computing the approximate entropy 5. The industrial wastewater quality soft-sensing method according to claim 1, characterized in that, In step (5), the time step adaptive dynamic selection mechanism automatically adjusts the time step according to the dynamic characteristics of each component.

6. The industrial wastewater quality soft-sensing method according to claim 1, characterized in that, Step (6) comprises the following steps: (11) The EfficANet model receives the wastewater quality multivariate time series X ∈ R M×T ; and converts it into a suitable feature space through patching and embedding strategies; reshapes through an unsqueeze operation to X in ∈ R M×1×T , where M is the number of variables and H is the length of the time series; (62) dividing the input sequence into multiple overlapping patches by using a convolution dry layer, processing the patches to extract local features. The EfficANet model is composed of multiple stacked blocks, each block containing three key modules: TLDC time local convolution, IVGC variable group convolution and GTVA global time-variable attention; an adaptive attention weight mechanism is introduced into the GTVA module to capture the complex relationship in the time and variable dimensions. (64) After processing by multiple sub-modules, the output passes through a prediction head to generate the final prediction result wherein Z (l) is the output of the first block, denotes an operation within the first block.

7. The industrial wastewater quality soft-sensing method according to claim 6, characterized in that, Step (63) comprises the following steps: (631)Temporal Local Convolution TLDC: input tensor shape is X emb , shape is (N, C, T), reshapes the input tensor to (NxC, T) and split into NxC groups; A deep convolution DW Conv is applied to capture short-term dependencies: X local = DW Conv(X emb ) A deep dilated convolution DW-D Conv is applied to expand the receptive field: X dilated = DW - D Conv(X local ) The outputs of the two convolutions are added: X combined = X dilated + X local (632)The IVGC module processes the time dependence of each variable independently by grouped convolution: the input tensor X combined shape (N, C, T), padding the time dimension T to be a multiple of the window size W, grouped into groups; The padding length is calculated: Calculate the tail padding length P left and P right : The tensors with standard padding and head-tail padding are respectively subjected to 1D group convolution, and the outputs of the two convolutions are aligned and combined: Y = Conv(Conv(X padded1 )+Conv(X padded2 )) (633) The GTVA module captures the dependency between time and variables through a global time and variable attention mechanism; including time attention and variable attention two parts, respectively processing the features on the time axis and the variable axis through a fully connected network, and then fusing the two attentions through Hadamard product. (644)In the GTVA module, an adaptive attention weight mechanism is introduced to capture the complex relationship in time and variable dimensions: the input tensor shape is Y shape (N, C, T), the input tensor is reshaped to (NxC, T), and global average pooling is applied: T pool = AvgPool(Y temp ) Generate time attention weights through two fully connected networks: T atten = σ(W2 · ReLU(W1 · T pool )) Reshape the input tensor to (N*T, C), and apply global average pooling: V pool = AvgPool(Y var ) Generate variable attention weights through two fully connected networks: V atten = σ(W4 · ReLU(W3 · V pool )) Multiply the time attention and variable attention weights with the convolution output: Y out = σ(T atten ⊙V atten ⊙Y) Multiply the GTVA output with the block input to introduce a feedback mechanism: X' emb = Y out O X emb .

8. The industrial wastewater quality soft-sensing method according to claim 1, characterized in that, In step (1), the water quality data variables include total nitrogen TN, total phosphorus TP, chemical oxygen demand COD, biochemical oxygen demand BOD, pH value and dissolved oxygen DO.

9. An electronic device, comprising: A memory and a processor, the memory stores a computer program, the processor executes the program to realize the steps of the method of any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A memory stores a computer program, the program is executed by the processor to realize the steps of the method of any one of claims 1-7.

Images