CN118228766B - A method for measuring effluent ammonia nitrogen concentration based on convolutional layer and self-attention mechanism - Google Patents

Abstract

The invention relates to a method for measuring the ammonia nitrogen concentration of effluent water based on a convolution layer and a self-attention mechanism, which comprises the following steps: constructing a sample data set of the variable easy to measure; step 2: constructing a convolution layer, substituting the easily-measured variable into a convolution function to process, and applying a nonlinear activation function to output data of the convolution function by the convolution function; step 3: setting a self-attention mechanism function, and reducing noise on the data; step 4: constructing a long-term and short-term memory neural network model based on a convolution function and a self-attention mechanism function to obtain a measurement model; step 5: adjusting and optimizing model parameters of the measurement model through a self-adaptive Bayesian optimization strategy to obtain an optimal detection model; step 6: and obtaining the ammonia nitrogen concentration of the final effluent. Compared with the prior art, the method has the remarkable advantages in the aspects of cost, noise suppression, generalization capability, real-time response and detection precision, and provides a more efficient and reliable ammonia nitrogen concentration measuring method for the field of urban sewage treatment.

Description

A method for measuring effluent ammonia nitrogen concentration based on convolutional layer and self-attention mechanism

Technical Field

The present invention relates to the technical field of urban sewage treatment detection, and in particular to a method for measuring effluent ammonia nitrogen concentration based on a convolutional layer and a self-attention mechanism.

Background technique

The eutrophication of urban water bodies, mainly caused by nitrogen and phosphorus pollutants, has attracted widespread attention. In order to meet this challenge, my country has increasingly stringent requirements for nitrogen and phosphorus emissions in urban sewage treatment, and regulatory standards are also constantly improving. Urban sewage treatment facilities must not only remove organic pollutants, but also effectively reduce nitrogen and phosphorus emissions, which is of great significance for preventing and controlling water pollution and improving the quality of urban water environment. In this context, accurate measurement of effluent water quality indicators is particularly important, which is the key to ensuring that sewage treatment is continuously effective and that effluent water quality meets standards.

The concentration of ammonia nitrogen in effluent is one of the important indicators for measuring the effect of urban sewage treatment. Real-time and accurate detection of ammonia nitrogen concentration in effluent can help improve the control performance of the sewage treatment process. Traditional methods for measuring ammonia nitrogen concentration in effluent mainly include chemical analysis and online monitoring instruments. The chemical analysis method involves sending water samples to the laboratory, and detecting ammonia nitrogen concentration through chemical reagent treatment and equipment such as spectrophotometers. It can provide accurate results, but there are problems such as time delay, high cost and complex operation. In contrast, online monitoring instruments can monitor ammonia nitrogen concentration in real time, but they have high maintenance costs, are susceptible to interference, and have large initial investments. These traditional measurement methods face many challenges in practical applications, prompting researchers to explore more efficient and economical alternatives.

In recent years, model-based detection methods have been proposed, which aim to detect the effluent ammonia nitrogen concentration through mathematical models. However, the constant changes in influent flow and pollutant composition in the urban sewage treatment process, as well as complex biochemical reactions, make it difficult to establish an accurate mathematical model. At present, detection methods based on fuzzy neural networks (FNN), long short-term memory (LSTM) networks, and extreme learning machines (ELM) have been proposed, but these methods generally have problems with insufficient model generalization ability and poor data quality in practical applications. Especially when dealing with large-scale urban sewage treatment data, these methods still need to be improved in terms of detection accuracy and timeliness. In addition, the input data in the sewage treatment process is easily affected by factors such as climate change and seasonal changes, showing large volatility and uncertainty, which poses a challenge to the generalization ability and stability of the model.

Therefore, it is urgent for technical personnel in this field to solve the problems of high model cost, data noise interference, poor model generalization ability and slow real-time responsiveness in the existing technology for measuring effluent ammonia nitrogen concentration in the field of urban sewage treatment.

Summary of the invention

In view of the problems existing in the prior art, the purpose of the present invention is to provide a measurement method that has improvements in economy, accuracy and timeliness, can process large-scale urban sewage treatment data, improve the generalization ability of the model and the adaptability of data quality, and thus accurately detect the ammonia nitrogen concentration in water in real time.

To achieve the above objectives, the present invention proposes a long short-term memory neural network (CS-LSTM for short) integrating convolutional layers and self-attention mechanism (SE), and uses easily accessible data to build a water ammonia nitrogen concentration detection model, thereby reducing the detection cost. By integrating convolutional layers, SE and LSTM, it aims to solve the problem of difficulty in capturing long-term dependencies in time series data, effectively extract features from complex data, reduce noise, and perform accurate measurements. In addition, the improved Bayesian optimization algorithm solves the problem of model parameter optimization and improves the stability and accuracy of the detection model.

Specifically, the technical solution adopted by the present invention is: a method for measuring effluent ammonia nitrogen concentration based on a convolutional layer and a self-attention mechanism, comprising the following steps:

Step 1: Collect the easily measurable variables that affect the effluent ammonia nitrogen concentration and construct a sample data set of easily measurable variables;

Step 2: Use the convolution function to construct a convolution layer, substitute the easily measurable variables into the convolution function for processing, and obtain data features. The convolution function applies a nonlinear activation function to its output data.

Step 3: Setting a self-attention mechanism function, inputting the data features obtained by the convolution function into the self-attention mechanism function, and the self-attention mechanism function denoises the data by learning the nonlinear relationship between different channels;

Step 4: Construct a long short-term memory neural network model based on convolution function and self-attention mechanism function, model the influencing factors collected in steps 1-3, and obtain the measurement model of effluent ammonia nitrogen concentration ;

Step 5: Use the adaptive Bayesian optimization strategy to optimize the measurement model The model parameters are adjusted and optimized, and the objective function is set , introduce the Gaussian process as the starting prior distribution, calculate the mean and variance of the posterior distribution based on the observed data, and dynamically adjust the existing observed data through the adaptive sampling function to determine the next observation point and obtain the optimal detection model ;

Step 6: Input the variables that affect the effluent ammonia nitrogen concentration into the optimal detection model The final effluent ammonia nitrogen concentration can be obtained .

In the above-mentioned method for measuring effluent ammonia nitrogen concentration based on convolutional layer and self-attention mechanism, in step 1, the easily measurable variables include dissolved oxygen at the end of aerobic phase, total suspended solids at the end of aerobic phase, effluent pH value, effluent redox potential, and effluent nitrate nitrogen. The sample data set is The collected effluent ammonia nitrogen concentration data is expressed as , is the i -th effluent ammonia nitrogen concentration point, and n represents the number of samples of easily measurable variables.

In the above-mentioned method for measuring effluent ammonia nitrogen concentration based on convolutional layer and self-attention mechanism, in step 2, the convolution function is ,in, is the output feature of the lth convolutional layer at position, ( i,j ) is the value of i = 0... H -1, j = 0... W -1, all A set of feature maps with a shape of H × W × c is formed, where H and W represent the height and width of the channel respectively, and c represents the number of channels. represents the activation function, is the weight of position u,v in the convolution kernel of layer l , , M and N are the height and width of the convolution kernel respectively, The corresponding element of the convolution kernel covers the part of the input matrix, and the coverage area is , It is The bias term of the layer.

In the above-mentioned method for measuring effluent ammonia nitrogen concentration based on convolutional layer and self-attention mechanism, step 3 comprises:

Step 3-1: Compression, using global average pooling to generate a channel descriptor , used to emphasize the spatial information of the global distribution: ,in yes The value of the cth channel at position ( i,j ), is the cth element of the compressed feature vector;

Step 3-2: Excitation, using a compression ratio r to reduce the number of parameters and computational complexity, then adding nonlinearity through an activation function and another fully connected layer, and finally using a sigmoid function to generate the weight of each channel;

Step 3-3: Apply the weights obtained after excitation to the original input feature map to perform channel recalibration.

_In the above-mentioned method for measuring effluent ammonia nitrogen concentration based on convolutional layer and self-attention mechanism, step 3-2 includes: setting M1 _and M2 as weights of two fully connected layers respectively, , , the output weight s of the stimulus, for: ,in, is the sigmoid function, z is the output of the compression step;

The step 3-3 includes: the calculation formula of the recalibrated output of the channel is: ,in, is the weight of channel c , is the channel c of the original input feature map, is the output after recalibration.

In the above-mentioned method for measuring effluent ammonia nitrogen concentration based on convolutional layer and self-attention mechanism, in step 4, the measurement model , where W represents the weight matrix connecting the input x to the output gate at time t , h is the hidden state of the previous time step, and b is the bias term, which includes:

Step 4-1: Process the input data through the convolution layer to extract key spatial features, which includes using different convolution kernels to gradually capture the basic features and complex features of the data in consecutive convolution layers;

Step 4-2: Introduce the self-attention mechanism to further process the output of the convolution layer, dynamically adjust the feature channels through global average pooling, excitation, and recalibration steps to highlight important features and suppress unimportant features, effectively reduce noise and enhance the feature expression ability of the model;

Step 4-3: Input the processed features into LSTM and use its gate structure to capture the long-term dependencies in time series data, thereby achieving accurate measurement of effluent ammonia nitrogen concentration.

In the above-mentioned method for measuring effluent ammonia nitrogen concentration based on convolutional layer and self-attention mechanism, in step 5, the model parameters include: the weight matrix W in the LSTM layer, the bias term b , and the number of neurons C.

In the above-mentioned method for measuring effluent ammonia nitrogen concentration based on convolutional layer and self-attention mechanism, in step 5: the adaptive sampling function is: , where i=1, 2, 3, D _n ∈ x .

In the above-mentioned method for measuring effluent ammonia nitrogen concentration based on convolutional layer and self-attention mechanism, the optimization strategy in step 5 includes:

Step 5-1: In the early stages of optimization, there is relatively little information about the objective function. At this time, an upper confidence bound function that focuses more on exploration is used to quickly cover a wide range of parameter space to find potential favorable areas. ,in, It's on point The predicted mean of the objective function at , It's on point The prediction standard deviation at , f is the parameter that controls the balance between exploration and exploitation;

Step 5-2: When partial information about the objective function is available, use the expected improvement function to maintain a balance between exploring unknown areas and exploiting known good areas. ,in, is the input vector corresponding to the currently known optimal parameters, and E represents the expected value;

Step 5-3: Towards the end of the optimization process, use the probability improvement function to carefully search in the promising area that has been discovered to fine-tune the true optimal solution. , where P is the probability measure and z is a non-negative parameter that controls the intensity of exploration.

The beneficial effects of the method for measuring effluent ammonia nitrogen concentration based on a convolutional layer and a self-attention mechanism of the present invention are:

Cost-effectiveness: By using a long short-term memory neural network with integrated convolutional layers and attention mechanism, the cost of data processing and detection of effluent ammonia nitrogen concentration is reduced, and the economic benefit is improved.

Noise suppression: This network structure can effectively extract features from complex data, reduce noise interference in the data, and improve the accuracy and reliability of the measurement results.

Generalization ability: The CS-LSTM model has strong generalization ability and can adapt to large-scale urban sewage treatment data under different conditions, which improves the adaptability and flexibility of the model.

Real-time response: The model can respond to changes in treatment conditions in real time, improving the real-time detection capability of effluent ammonia nitrogen concentration.

Measurement accuracy: The improved Bayesian optimization algorithm improves the stability and accuracy of the model, thereby improving the detection accuracy of effluent ammonia nitrogen concentration.

Compared with the existing technology, the present invention has significant advantages in cost, noise suppression, generalization ability, real-time response and detection accuracy, and provides a more efficient and reliable ammonia nitrogen concentration measurement method for the field of urban sewage treatment.

BRIEF DESCRIPTION OF THE DRAWINGS

Fig. 1 is a flow chart of the measuring method of the present invention;

FIG2 is a network structure diagram of a measurement model of the present invention;

FIG. 3 is a diagram showing the effect of measuring the ammonia nitrogen concentration in effluent water according to the present invention.

Detailed ways

In order to enable those skilled in the art to better understand the technical solution of the present invention, the technical solution of the present invention is described below in conjunction with specific implementation methods and drawings.

Example

As shown in Figure 1-2, a method for measuring effluent ammonia nitrogen concentration based on a convolutional layer and a self-attention mechanism includes the following steps:

Step 1: Collect easily measurable variables that affect the effluent ammonia nitrogen concentration and construct a sample data set of easily measurable variables. The easily measurable variables include dissolved oxygen at the end of aerobic stage, total suspended solids at the end of aerobic stage, effluent pH value, effluent redox potential, and effluent nitrate nitrogen. The sample data set is x = [ x1 _, x2 _, ..., xn ]. The collected effluent ammonia nitrogen concentration data is expressed as Y = [ y1 , ..., _yn ], where yi _is the _i - th effluent ammonia nitrogen concentration point, _and n represents the number of samples of easily measurable variables.

Step 2: Construct a convolution layer by referencing the convolution function. The convolution function is ,in, is the value of the output feature of the lth ( l = 1, 2) convolutional layer at position ( i, j ) ( i = 0... H -1, j = 0... W -1), all A set of feature maps with a shape of H × W × c is formed, where H and W represent the height and width of the channel respectively, and c represents the number of channels. represents the activation function, is the weight of position u,v in the convolution kernel of layer l , , M and N are the height and width of the convolution kernel respectively, The corresponding element of the convolution kernel covers the part of the input matrix, and the coverage area is , It is The bias term of the layer.

Substitute the easily measurable variables into the convolution function for processing to obtain data features, and the convolution function applies a nonlinear activation function to its output data.

Step 3: Setting a self-attention mechanism function, inputting the data features obtained by the convolution function into the self-attention mechanism function, and the self-attention mechanism function denoises the data by learning the nonlinear relationship between different channels. Step 3 includes:

Step 3-1: Compression, use global average pooling to generate a channel descriptor z _c to emphasize the spatial information of global distribution: ,in is the value of the cth channel of x at position ( i,j ) and z _c is the cth element of the compressed feature vector.

Step _3-2 : Excitation, reduce the number of parameters and computational complexity through a compression ratio r , and then increase nonlinearity through an activation function and another fully connected layer. Set M1 _and M2 to be the weights of the _{two fully connected layers, M1∈Rc×c/r, M2∈Rc} / ^r × ^c , _and the weights s, s∈Rc ^×1 of the output of the excitation are: , where σ is the sigmoid function, z is the output of the compression step, and finally the sigmoid function is used to generate the weights of each channel.

Step 3-3: Apply the weights obtained after excitation to the original input feature map to perform channel recalibration. The calculation formula for the channel recalibration output is: , where s _c is the weight of channel c , x _c is the channel c of the original input feature map, is the output after recalibration.

Step 4: Construct a long short-term memory neural network model based on convolution function and self-attention mechanism function, model the influencing factors collected in steps 1-3, and obtain the measurement model F ( x ) of effluent ammonia nitrogen concentration. , where W represents the weight matrix connecting the input x to the output gate at time t , h is the hidden state of the previous time step, and b is the bias term, including:

Step 4-1: Process the input data through the convolution layer to extract key spatial features, which includes using different convolution kernels to gradually capture the basic features and complex features of the data in consecutive convolution layers.

Step 4-2: Introduce the self-attention mechanism to further process the output of the convolutional layer, and dynamically adjust the feature channels through global average pooling, excitation, and recalibration steps to highlight important features and suppress unimportant features, effectively reduce noise and enhance the feature expression ability of the model.

Step 4-3: Input the processed features into LSTM, and use its specially designed gate structure to capture the long-term dependencies in time series data, thereby achieving accurate measurement of effluent ammonia nitrogen concentration.

Step 5: Through the adaptive Bayesian optimization strategy, the model parameters of the measurement model F ( x ) are adjusted and optimized. The model parameters include: the weight matrix W in the LSTM layer, the bias term b , and the number of neurons C.

Set the objective function f ( Θ _p ), introduce the Gaussian process as the starting prior distribution, calculate the mean and variance of the posterior distribution based on the observed data, and dynamically adjust the existing observed data through the adaptive sampling function to determine the next observation point and obtain the optimal detection model .

The adaptive sampling function is: , where i=1, 2, 3, D _n ∈ x .

Optimization strategies include:

Step 5-1: In the early stages of optimization, there is relatively little information about the objective function. At this time, an upper confidence bound function that focuses more on exploration is used to quickly cover a wide range of parameter space to find potential favorable areas. , where μ ( Θ _P ) is the predicted mean of the objective function at point Θ _P , v ( Θ _P ) is the predicted standard deviation at point Θ _P , and f is a parameter that controls the balance between exploration and exploitation;

Step 5-2: When there is already some information about the objective function, use the expected improvement function to maintain a balance between exploring unknown areas and utilizing known good areas. , where Θ _P ⁺ is the input vector corresponding to the currently known optimal parameters, and E represents the expected value;

Step 5-3: Towards the end of the optimization process, use the probability improvement function to carefully search in the promising area that has been discovered to fine-tune the true optimal solution. , where P is the probability measure and z is a non-negative parameter that controls the intensity of exploration.

Step 6: Input the variables that affect the effluent ammonia nitrogen concentration into the optimal detection model The final effluent ammonia nitrogen concentration can be obtained .

Example 2

As shown in Figure 1-3, a method for measuring effluent ammonia nitrogen concentration based on convolutional layers and self-attention mechanism includes the following steps.

Step 1: Collect easily measurable variables that affect the effluent ammonia nitrogen concentration (aerobic terminal dissolved oxygen (DO), aerobic terminal total suspended solids (TSS), effluent (pH), effluent oxidation-reduction potential (ORP), and effluent nitrate nitrogen (NO3-N)), and represent the collected n sample data sets as x = [ x1 _, x2 _, ..., xn ]; at the same time, the collected effluent ammonia nitrogen concentration data is represented as Y = [ y1 _, ..., _{yn ]} , where yi _is the i -th effluent ammonia nitrogen concentration point.

Step 2: After referencing the convolution function to construct the convolution layer, the collected data is first processed by the first convolution layer, which uses multiple convolution kernels to capture the basic features of the data. Subsequently, the data is passed to the second convolution layer, which further uses different convolution kernels to extract more complex data features based on the features extracted by the first layer. In this process, each convolution layer applies a nonlinear activation function to its output data, introducing nonlinearity to learn and represent more complex data features, thereby effectively extracting feature information that is critical to subsequent analysis and detection.

The two-layer convolution function used in the present invention is as follows:

(1);

in, is the value of the output feature of the lth ( l = 1, 2) convolutional layer at position ( i, j ) ( i = 0... H -1, j = 0... W -1), all It forms a set of feature maps with a shape of H × W × c , where H and W represent the height and width of the channel respectively, and c represents the number of channels. δ is the RELU activation function, which is specifically expressed as , which is used to convert the input signal into the output signal in order to introduce nonlinear characteristics into the network so that the network can learn complex features. is the weight of position u,v in the convolution kernel of layer l , , M and N are the height and width of the convolution kernel respectively, The corresponding element of the convolution kernel covers the part of the input matrix, and the coverage area is , It is The bias term of the layer increases the flexibility of the network.

Step 3: After receiving the input of the convolutional layer, the present invention uses the SE layer to reduce noise on the data. The SE layer can effectively enhance the model's ability to express important features by learning the nonlinear relationship between different channels, while suppressing unimportant features, thereby improving the accuracy and generalization ability of the model.

Specifically, compression is first performed, and a channel descriptor z _c is generated using global average pooling to emphasize the spatial information of global distribution:

(2);

in is the value of the cth channel of x at position ( i,j ), and z _c is the cth element of the compressed feature vector.

Then the excitation is performed, and the number of parameters and computational complexity are reduced by a compression ratio r , and then the nonlinearity _is increased by the ReLU activation function and another fully connected layer, and finally the sigmoid function is used to generate the weight of _each channel. Assuming that M1 _and M2 are the weights of the two fully connected layers ( M1∈Rc ^×c/r , M2∈Rc ^/r×c ), the output s ( s∈Rc ^×1 ) _of the excitation is:

(3);

Here, σ is the sigmoid function and z is the output of the compression step.

Finally, the weight s obtained after excitation is applied to the original input feature map to perform channel recalibration. For each channel, its recalibrated output can be calculated by the following formula:

(4);

Among them, s _c is the weight of channel c , x _c is the channel c of the original input feature map, is the output after recalibration.

The data processed by the above steps can effectively reduce noise to improve the accuracy of the model and provide more accurate data for the subsequent LSTM layer detection.

After feature extraction of the input data, the features extracted by the convolutional layer are denoised through the SE layer. This process first applies global average pooling to generate channel descriptors to summarize the global distribution information of each channel. Subsequently, the number of parameters is reduced through a compression ratio r , and nonlinearity is introduced through the ReLU activation function and the fully connected layer. Finally, the sigmoid function is used to generate adjustment weights for each channel. These weights are applied to the original feature map to achieve dynamic adjustment of feature channels and prioritize more important features, thereby improving the model's understanding and processing capabilities of the data, especially when processing complex or noisy data sets, which can effectively improve the accuracy and generalization performance of the model.

Step 4: Use the CS-LSTM method to model the variable data affecting the effluent ammonia nitrogen concentration collected in steps 1-3 to obtain the measurement model of the effluent ammonia nitrogen concentration. , where W represents the weight matrix connecting the input x to the output gate at time t , h is the hidden state of the previous time step, and b is the bias term; specifically:

First, the input data is processed through a convolutional layer to extract key spatial features, which includes using different convolution kernels to gradually capture the basic features and complex features of the data in consecutive convolutional layers. The basic features are represented by the basic correlation of the data, and the complex features are represented by the linear relationship of the data. The basic correlation represents the degree of dependence or interaction between variables, and the linear relationship of the data represents the direct proportional relationship between variables. The basic correlation of the data and the linear relationship of the data are existing technologies.

Next, the SE attention mechanism is introduced to further process the output of the convolutional layer. The feature channels are dynamically adjusted through global average pooling, excitation, and recalibration steps to highlight important features and suppress unimportant features, effectively reducing noise and enhancing the feature expression ability of the model.

Finally, the processed features are input into LSTM, and its specially designed gate structure is used to capture the long-term dependencies in time series data, thereby achieving accurate measurement of effluent ammonia nitrogen concentration.

Step 5: Through the adaptive Bayesian optimization strategy, the model parameters are precisely adjusted and optimized, involving the weight matrix W , the bias term b , and the number of neurons C in the LSTM layer. The initial step includes the setting of the objective function f ( Θ _p ), followed by the introduction of the Gaussian process (GP) as the starting prior distribution. This process covers the definition of the mean function and the kernel function to depict the correlation between different parameter settings. Based on the observed data, the mean and variance of the posterior distribution are calculated, and the following adaptive sampling function is used to dynamically adjust the existing observed data to determine the next observation point. The sampling function can adopt strategies such as upper confidence bound (UCB), expected improvement (EI) or probability improvement (PI) according to the different optimization stages.

The iterative process continuously updates the posterior distribution and determines the parameter point that maximizes the acquisition function value until the termination condition is met.

The adaptive acquisition function is as follows:

(6);

Where i = 1, 2, 3. D _n ∈ x , the specific optimization strategy is as follows:

(1) In the early stages of optimization, there is relatively little information about the objective function. At this time, an upper confidence bound (UCB) function that focuses more on exploration is used to quickly cover a wide range of parameter space to find potential favorable areas.

(7);

Where μ ( ΘP ₎ is _the predicted mean of the objective function at point ΘP , v ( ΘP ) is the predicted _standard deviation at point ΘP , _and f is a parameter (a constant) that controls the balance between exploration and exploitation.

(2) When there is certain objective function information, the expected improvement function is used. The expected improvement function can better balance between exploring unknown areas and utilizing known good areas.

(8);

Among them, Θ _P ⁺ is the input vector corresponding to the currently known optimal parameters.

(3) Towards the end of the optimization process, a probability improvement function is used to conduct a detailed search in the discovered promising regions to fine-tune the true optimal solution.

(9);

Where P is the probability measure and z is a non-negative parameter that controls the intensity of exploration.

In the iterative process, the posterior distribution is continuously updated to find the parameter point that maximizes the sampling function value until the predetermined termination criterion is reached. This algorithm significantly enhances the effectiveness of the model by carefully adjusting the model parameters, thereby maximizing the model performance and obtaining the optimal detection model. .

Specifically, as shown in FIG2, the present invention constructs a CS-LSTM model for detecting the ammonia nitrogen concentration in effluent water. In the initial stage, the convolution layer is used to process the input data to extract the spatial dimension characteristics of the data. Through the convolution kernel, the convolution layer gradually extracts the basic to complex features in the data. Subsequently, the output of the convolution layer is deeply processed by integrating the SE mechanism. This mechanism dynamically adjusts the weights of feature channels through steps such as global average pooling, activation, and recalibration, strengthens key features and suppresses secondary features, thereby effectively reducing noise and improving the model's ability to interpret features. Finally, the screened and optimized features are passed to the LSTM network, which uses its unique gating mechanism to capture long-term dependencies in time series data and achieve accurate detection of effluent ammonia nitrogen concentrations. In FIG2, ① is the LSTM forget gate formula, which is specifically expressed as ,② is the LSTM input gate formula, specifically expressed as , ③ is the lstm updating cell state, which is specifically expressed as , ④ is the LSTM output gate formula, which is specifically expressed as , ⑤ and ⑥ are convolution function formulas, specifically expressed as , Ⅰ is the output of SE layer after folding, which is x _t, Ⅱ is , III is c _t-1, IV is c _t, V is h _t, VI, VII, IX are f _t, i _t, o _{t respectively,} VIII is g _t, where x _t is a two-dimensional vector, ( x _t ∈R ^m×1 ) is the input at time t; is a two-dimensional vector, which is the hidden state of the previous time step; c _t-1 is a two-dimensional vector, which is the cell state at time t -1; c _t is a two-dimensional vector, which is the cell state at time t ; h _t is a two-dimensional vector, ( h _t ∈R ^m×1 ) is the hidden state at time t; f _t, i _t, o _t are the outputs of the forget gate, input gate and output gate at time t respectively; g _t is the candidate value of the cell state at time t ; W _f , W _O , W _i , W _C ( W ∈R ^m×m ) respectively represent the weight matrices connecting the input x _t to different gates at time t , b _f , bi , b _C , b _O ( b ∈R ^m×1 ) are bias terms, _and σ is the sigmoid activation function. The role of the activation function is to introduce nonlinearity.

Step 6: Input the variables that affect the effluent ammonia nitrogen concentration into the determined detection model to obtain the effluent ammonia nitrogen concentration. .

In order to evaluate the performance of the model of the present invention, the root mean square error (RMSE) and the correlation coefficient (R ² ) are used as evaluation indicators, and the calculation formulas are:

(10);

(11);

Where yi _is the true measurement value of the i -th sample in the test set, is the detection value of the test set sample, is the mean of the test samples.

Table 1: Comparison results of different algorithms

.

Table 1 is a table of comparison results of different algorithms. After comparing the performance of multiple models, including FNN, LSTM and ELM models, the present invention shows significant advantages through in-depth analysis of the two key statistical indicators of RMSE and R ^2. Specifically, the RMSE of the present invention is much lower than that of other models, indicating that the difference between our test results and the real data is small and the detection accuracy is high. At the same time, the R ² of the model is close to 1, indicating that the model can accurately mine the intrinsic relationship between the data and has a high degree of fit. These indicators show that our method can not only improve the measurement accuracy, but also grasp the complex dynamics of the data. The present invention is of great significance to improving the accuracy of effluent ammonia nitrogen concentration detection.

The above embodiments are only for illustrating the inventive concept and features of the present invention, and their purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly, and they cannot be used to limit the protection scope of the present invention. Any equivalent changes or modifications made based on the essence of the content of the present invention should be included in the protection scope of the present invention.

Claims (9)

1. The method for measuring the ammonia nitrogen concentration of the effluent based on the convolution layer and the self-attention mechanism is characterized by comprising the following steps of:

Step 1: collecting easily-measured variables influencing the ammonia nitrogen concentration of the effluent, and constructing an easily-measured variable sample data set;

Step 2: constructing a convolution layer by referring to the convolution function, substituting the easily-measured variable into the convolution function for processing to obtain data characteristics, and applying a nonlinear activation function to output data of the convolution function by the convolution function;

Step 3: setting a self-attention mechanism function, inputting the data characteristics obtained by the convolution function into the self-attention mechanism function, and carrying out noise reduction on the data by the self-attention mechanism function through learning nonlinear relations among different channels;

Step 4: constructing a long-term and short-term memory neural network model based on a convolution function and a self-attention mechanism function, modeling the factors of the influence collected in the steps 1-3, and obtaining a measurement model of the ammonia nitrogen concentration of the effluent ；

Step 5: by self-adaptive Bayesian optimization strategy, the measurement model is subjected toModel parameters of (a) are adjusted, optimized and an objective function is setIntroducing Gaussian process as initial prior distribution, calculating average value and variance of posterior distribution according to observed data, dynamically adjusting according to existing observed data by self-adaptive sampling function to determine next observation point, and obtaining optimal detection model；

Step 6: inputting the obtained variable influencing the ammonia nitrogen concentration of the effluent into an optimal detection modelIn the process, the ammonia nitrogen concentration of the final effluent can be obtained。

2. The method for measuring the ammonia nitrogen concentration of effluent based on a convolution layer and a self-attention mechanism according to claim 1, wherein the method comprises the following steps: in the step1, the easily-measured variables comprise aerobic end dissolved oxygen, aerobic end total solid suspended matters, effluent pH value, effluent oxidation-reduction potential and effluent nitrate nitrogen, and the sample data set isThe collected effluent ammonia nitrogen concentration data is expressed as，And n represents the sample number of the easily-measured variable for the ith effluent ammonia nitrogen concentration point.

3. The method for measuring the ammonia nitrogen concentration of effluent based on a convolution layer and a self-attention mechanism according to claim 2, wherein the method comprises the following steps: in step 2, the convolution function isWherein, the method comprises the steps of, wherein,Is the output characteristic of the first convolutional layer at position, (i, j) is a value of i= … … H-1, j= … … W-1, allA set of feature maps of the shape H x W x c is constructed, where H and W represent the height and width of the channel, respectively, c represents the number of channels,The activation function is represented as a function of the activation,Is the weight of the position u, v in the convolution kernel of the first layer,M and N are the height and width of the convolution kernel, respectively,Covering the input matrix part for the corresponding elements of the convolution kernel with a coverage area of，Is the firstBias terms of the layers.

4. The method for measuring ammonia nitrogen concentration in effluent based on a convolution layer and a self-attention mechanism according to claim 3, wherein said step 3 comprises:

Step 3-1: compression, generating a channel descriptor using global averaging pooling Spatial information used to emphasize global distribution: Wherein Is thatThe value of the c-th channel at position (i, j),Is the c-th element of the compressed feature vector;

Step 3-2: excitation, reducing the parameter quantity and the calculation complexity through a compression ratio r, then increasing nonlinearity through an activation function and another full-connection layer, and finally generating the weight of each channel by using a sigmoid function;

Step 3-3: the weights obtained after excitation are applied to the original input feature map to perform channel recalibration.

5. The method for measuring ammonia nitrogen concentration of effluent based on a convolution layer and a self-attention mechanism according to claim 4, wherein said step 3-2 comprises: setting M ₁ and M ₂ as weights of two fully connected layers respectively,，The weight s of the output of the stimulus,The method comprises the following steps: Wherein, the method comprises the steps of, wherein, Is a sigmoid function, z is the output of the compression step;

the step 3-3 comprises the following steps: the calculation formula of the recalibration output of the channel is as follows: Wherein, the method comprises the steps of, wherein, Is the weight of the channel c and,Is the channel c of the original input feature map,Is output after recalibration.

6. The method for measuring ammonia nitrogen concentration in effluent based on a convolution layer and self-attention mechanism as set forth in claim 5, wherein in said step 4, said measuring modelWherein W represents a weight matrix connecting an input x to an output gate at a time t, h is a hidden state of a previous time step, and b is a bias term, and the method specifically includes:

Step 4-1: processing the input data through the convolution layers to extract key spatial features, including progressively capturing basic and complex features of the data in successive convolution layers using different convolution kernels;

Step 4-2: introducing a self-attention mechanism to further process the output of the convolution layer, dynamically adjusting the characteristic channel through the steps of global average pooling, excitation and recalibration to highlight important characteristics and inhibit unimportant characteristics, effectively reducing noise and enhancing the characteristic expression capability of the model;

step 4-3: the processed characteristics are input into the LSTM, and the gate structure is utilized to capture the long-term dependence in the time series data, so that the accurate measurement of the ammonia nitrogen concentration of the effluent is realized.

7. The method for measuring ammonia nitrogen concentration in effluent based on a convolution layer and self-attention mechanism according to claim 6, wherein in step 5, the model parameters include: the weight matrix W, bias term b and neuron number C in LSTM layer.

8. The method for measuring the ammonia nitrogen concentration of effluent based on a convolution layer and a self-attention mechanism according to claim 7, wherein in step 5: the adaptive sampling function is: Where i=1, 2,3, d _n e x.

9. The method for measuring ammonia nitrogen concentration in effluent based on a convolution layer and self-attention mechanism according to claim 8, wherein the optimizing strategy in step 5 comprises:

Step 5-1: at an early stage of the optimization, relatively little information is available about the objective function, at which time the upper confidence function that is more focused on exploration is used, to quickly cover a wide parameter space to find potentially advantageous regions, Wherein, the method comprises the steps of, wherein,Is at the point ofThe predicted mean value of the objective function at that point,Is at the point ofThe prediction standard deviation of the position, f, is a parameter for controlling exploration and utilization balance;

step 5-2: when there is already partial information of the objective function, a balance is maintained between exploring unknown regions and utilizing known good regions using the desired improvement function, Wherein, the method comprises the steps of, wherein,Is the input vector corresponding to the current known optimal parameter, E represents the expected value;

step 5-3: near the end of the optimization process, a probability improvement function is used to search carefully within the found promising areas, to fine tune to find the true optimal solution, Where P is a probability metric and z is a non-negative parameter that controls the strength of the exploration.