CN115758125A - Industrial sewage treatment soft measurement method based on feature structure optimization and deep learning - Google Patents

Abstract

The invention discloses an industrial sewage treatment soft measurement method based on feature structure optimization and deep learning; the invention firstly eliminates the overlapping characteristic of the data collected by the industrial sewage treatment plant. In addition, a full set empirical mode decomposition method is adopted to decompose the input characteristic sequence and historical data. The input sequence and the historical data are decomposed into respective intrinsic mode functions, and then the IMFs are subjected to feature screening by using a feature selection Relief F method. Then, the invention applies a new mixed deep learning model to predict four water-yielding indexes, and the model combines a convolution neural network and a gate control circulation unit network to carry out optimization through an attention mechanism. Compared with a soft measurement algorithm without feature optimization, the method provided by the invention has the advantage that the prediction effect is improved.

Description

Soft-sensing method for industrial wastewater treatment based on feature structure optimization and deep learning

technical field

The invention belongs to the field of water quality soft measurement in sewage treatment system control, and in particular relates to an industrial sewage soft measurement method based on characteristic structure optimization and deep learning.

Background technique

Industrial wastewater refers to the wastewater, sewage and waste liquid produced in the industrial production process. It has various types and complex components, and often contains a variety of toxic substances. With the rapid development of industry, the pollution of industrial wastewater to water bodies is becoming more and more extensive and serious, causing soil and air pollution and threatening human health and safety. It can be seen that the treatment of industrial wastewater is more important than the treatment of urban sewage. How to effectively take corresponding purification measures for comprehensive utilization and treatment according to the composition and concentration of pollutants in wastewater is one of the core issues to be solved to ensure sustainable development. Although industrial wastewater treatment plants (IETP) can effectively treat wastewater through physical, chemical, and biological methods, the fluctuations in the quality and quantity of wastewater and other disturbance factors often weaken the treatment effect of IETP. In addition, the coupling relationship between other process variables in sewage treatment plants and a certain indicator of effluent in time domain and frequency domain is rarely proposed.

Contents of the invention

Aiming at the limitations of the existing soft-sensing methods for industrial sewage treatment, the present invention proposes a soft-sensing method for industrial sewage treatment based on feature structure optimization and deep learning.

In the present invention, firstly, Pearson correlation analysis (PCC) is performed on the collected data features of the industrial sewage treatment plant, and features with overlapping information are eliminated. Furthermore, the input feature sequence and historical data are decomposed using the Complete Ensemble Empirical Mode Decomposition (CEEMD) method. The decomposed IMFs components not only contain the main feature information, but also some IMFs represent the noise components in the original signal, and these pseudo components will become an obstacle for subsequent processing. The Relief F feature selection method is used to select the main process variables in IMFs as input sequences. Then, apply a new hybrid deep learning model, combined with convolutional neural network (CNN) and gated recurrent unit network (GRU), and finally, optimize through attention mechanism (AM), output water quality CODcr, NH ₃ -N, TN and the thick prediction results of TP.

The technical solution adopted by the present invention to solve its technical problems comprises the steps:

Step 1. Eliminate overlapping features for the data collected by industrial sewage treatment plants;

Step 2, performing empirical mode decomposition on the data in the industrial sewage treatment plant with overlapping features eliminated;

Step 3, performing feature screening on the decomposed IMFs;

Step 4, design CNN-GRU-AM hybrid network model;

Step 5. Train and verify the proposed model network, and output prediction indicators.

As preferably, the elimination of overlapping features described in step 1 is as follows:

Before building the model, a Pearson correlation analysis was performed on the data collected by the industrial sewage treatment plant. The strength of the correlation indicates whether there is overlap in the input information; the correlation coefficient ρ of the PCC signals x(t) and y(t) _xy is defined as:

In the formula, σ _xy is the covariance of signals x(t) and y(t), σ _x and σ _y are the standard deviations of signals x(t) and y(t) respectively; finally, the calculated correlation coefficient is Normalization, ρ _xy ∈ [-1, 1], the closer the value of ρ _xy is to 1, the stronger the correlation between the two signals, the threshold is set, and when ρ _xy is greater than the set threshold, it is recognized as the input signal for overlapping signals.

As a preference, the feature described in step 2 is subjected to empirical mode decomposition, specifically as follows:

The complementary integrated empirical mode decomposition method is used to process the original data, combined with the feature screening algorithm, the high-frequency noise interference and false components are eliminated, and the components that can represent effective information are screened out. Complementary integrated empirical mode decomposition not only retains the good adaptability of empirical mode decomposition, but also combines the advantages of integrated empirical mode decomposition to effectively suppress the mode aliasing phenomenon, and effectively overcomes the residual problem of Gaussian white noise. So that the overall model has better noise reduction performance, the specific steps are as follows:

Step ①. Add n pairs of positive and negative white noise ω _i (t) to the original signal s(t):

Step ②, the signal after adding noise is re-decomposed according to the EMD decomposition process, and two sets of corresponding intrinsic mode functions IMFs and residual r _i (t) are obtained, and then the mean value is calculated:

In the formula, C _ij ⁺ (t), C _ij ^- (t) represent the j-th IMF component, r _i ⁺ (t), ri _- ⁽ t) represent the i-th calculation residual.

The data acquisition of industrial sewage treatment plants is often in a complex environment, and the acquired data is usually mixed with a lot of noise. The process of CEEMD decomposition is to decompose the original signal on different time scales to obtain IMFs components with frequencies ranging from high to low. This process is an adaptive and approximately orthogonal filtering process. The IMFs component not only contains the main feature information, but also some IMFs represent the noise components in the original signal, and these pseudo components will become an obstacle for subsequent processing. In the next step, the present invention selects effective IMF components by setting the screening threshold of IMF.

As preferably, the decomposed IMFs described in step 3 are characterized as follows:

The features of the decomposed IMFs are screened by the ReliefF algorithm, specifically: select k nearest neighbor samples from the same and different samples, and average them to obtain the weight of each feature, so as to obtain the weight of each sample instance. Correlation of individual features with classes. Next, the features are sorted according to the weight, and the threshold is set to determine whether the feature is valid or invalid, and the n features with the largest weight are selected, and other features are removed for feature selection;

Proceed as follows:

Let D be the feature set, m be the number of sample resampling, δ be the feature statistic index, k be the number of nearest neighbor samples, and T be the statistic index of outputting each feature.

(1) Set all feature weights to 0

(2) For i＝1 to m do

(3) R is a sample data randomly drawn from the feature set D; H _j (j=1, 2,...k) is the k nearest neighbors H _j (j =1, 2,...k), find k nearest neighbors M _j (C) (j=1, 2,...k) in the feature sets inconsistent with each category.

(4) for i＝1 to N do

The statistical indicators of A for a certain feature are:

where p(C) is the proportion of the category. p(Class(R)) is the proportion of the category of a sample randomly selected.

(6) Sort W

After calculating the feature weight, the larger the weight, the stronger the feature's ability to distinguish samples, and a new feature subset is selected by setting the threshold.

As preferably, the design CNN-GRU-AM network model described in step 4 is as follows:

The CNN structure of the convolutional neural network includes a convolutional layer, a pooling layer, a fully connected layer, and an output layer; the CNN model uses the gradient descent method to train parameters, and the trained model can learn the characteristics of time series data.

The gated recurrent unit GRU neural network includes an update gate and a reset gate; the data propagation process in the GRU unit is described as follows: First, the reset is obtained through the last transmitted evaluation state h _t-1 and the input x _t of the current node The state of gate R _t and update gate Z _t :

r _t ＝σ(x _t W _xr +h _t-1 W _hr +b _r )

z _t ＝σ(x _t W _xr +h _t-1 W _hr +b _z )

集合中所记忆的当前时刻状态表示为：Next, currently

The current state of memory in the collection is expressed as:

Then, the GRU model updates the state by the following formula:

Finally, the output of the forward pass is:

y _t = σ(W _o h _t )

where _ht is the output of the GRU, W is the weight vector, and b is the bias vector of the GRU.

Through the last transmitted state H _t-1 and the input x _t of the current node, σ is a sigmoid function. Through this function, the data is transformed into a value in the range of 0-1 to serve as a gating signal. Among them, H _t-1 contains past information, and ⊙ represents the Hadamard product.

The implementation steps of the attention mechanism AM are shown in the following formula.

(1) Calculate the similarity between the given h _j value and the target state S _t-1 , that is, the weight of the state h _j at each time t:

e _tj = a(S _t-1 , h _j )

(2) Normalize the weight coefficients:

(3) Weighted average of state h _j :

where T is the total number of time steps in the input sequence.

The beneficial effects of the present invention are:

1. The present invention proposes a feature structure optimization method for industrial sewage treatment plant data. Compared with the soft sensor algorithm without feature optimization, the present invention improves the prediction effect.

2. In order to evaluate the prediction accuracy of the proposed CEEMD-ReliefF-CNN-GRU-AM model, taking the effluent CODcr, NH ₃ -N, TN and TP of a real industrial sewage treatment plant as an example, the proposed hybrid model was tested by comparative experiments Efficiency and stability were evaluated on real data sets, and the prediction effect was greatly improved.

Description of drawings

Fig. 1 is the flow chart of the industrial sewage treatment soft-sensing method of feature structure optimization and deep learning of the present invention;

Fig. 2-1 is the evaluation schematic diagram of the result of predicting effluent COD _cr according to the model trained in the embodiment of the present invention;

Fig. 2-2 is a schematic diagram of the evaluation of the result of predicting NH ₃ -N in the effluent according to the trained model in the embodiment of the present invention;

Fig. 2-3 is the evaluation schematic diagram of predicting the effluent TN result of the trained model in the embodiment of the present invention;

2-4 are schematic diagrams of evaluation of the results of predicting effluent TP by the trained model according to the embodiment of the present invention.

Detailed ways

The implementation steps of the present invention will be described in further detail below in conjunction with accompanying drawing 1 .

A soft-sensing method for industrial sewage treatment based on feature structure optimization and deep learning, specifically including the following steps:

Step 1. Eliminate overlapping features, as follows:

Before building the model, this paper conducts Pearson correlation analysis (PCC) on the data collected by industrial sewage treatment plants. The strength of correlation indicates whether there is overlap in the input information; the signals x(t) and y(t) of PCC The correlation coefficient ρ _xy is defined as:

In the formula, σ _xy is the covariance of signals x(t) and y(t), σ _x and σ _y are the standard deviations of signals x(t) and y(t) respectively; finally, the calculated correlation coefficient is Normalization, ρ _xy ∈ [-1, 1], the closer the value of ρ _xy is to 1, the stronger the correlation between the two signals, the threshold is set, and when ρ _xy is greater than the set threshold, it is recognized as the input signal for overlapping signals.

Step 2. Perform empirical mode decomposition on the features, as follows:

Since the data in the industrial sewage treatment plant is highly nonlinear and non-stationary, and mixed with uncertain noise, the present invention uses the Complementary Integrated Empirical Mode Decomposition (CEEMD) method to process the original data, and combines the feature screening algorithm to eliminate High-frequency noise interference and false components are used to screen out components that can represent valid information. Complementary integrated empirical mode decomposition (CEEMD) not only retains the good adaptability of empirical mode decomposition, but also combines the advantages of integrated empirical mode decomposition to effectively suppress the mode aliasing phenomenon, and effectively overcomes the residual problem of Gaussian white noise . So that the overall model has better noise reduction performance, the algorithm is as follows:

Step ①. Add n pairs of positive and negative white noise ω _i (t) to the original signal s(t):

Step ②, the signal after adding noise is re-decomposed according to the EMD decomposition process, and two sets of corresponding intrinsic mode functions IMFs and residual r _i (t) are obtained, and then the mean value is calculated:

In the formula, C _ij ⁺ (t), C _ij ^- (t) represent the j-th IMF component, r _i ⁺ (t), ri _- ⁽ t) represent the i-th calculation residual.

The data acquisition of industrial sewage treatment plants is often in a complex environment, and the acquired data is usually mixed with a lot of noise. The process of CEEMD decomposition is to decompose the original signal on different time scales to obtain IMFs components with frequencies ranging from high to low. This process is an adaptive and approximately orthogonal filtering process. The IMFs component not only contains the main feature information, but also some IMFs represent the noise components in the original signal, and these pseudo components will become an obstacle for subsequent processing. In the next step, the present invention selects effective IMF components by setting the screening threshold of IMF.

Step 3, perform feature screening on the decomposed IMFs, specifically as follows:

The basic idea of the Relief F algorithm is to assign weights to each feature of the sample, and iteratively update the weights, and then sort the corresponding features according to the size of the weights, and select feature subsets accordingly, so that good features are clustered together. samples, while discretizing heterogeneous samples. It can solve multi-category problems as well as regression problems, and can handle datasets with noise, incomplete features, and multi-category attributes. The idea of the algorithm is to select k nearest neighbor samples from samples of the same class and different classes, and calculate the average value of each feature weight to obtain the correlation between each feature and class in each sample instance; then , sort the features according to their weights, and determine whether the features are valid or invalid by setting a threshold, and select n features with the largest weights, and remove other features for feature selection;

Proceed as follows:

Let D be the feature set, m be the number of sample resampling, δ be the feature statistic index, k be the number of nearest neighbor samples, and T be the statistic index of outputting each feature;

(1) Set all feature weights to 0

(2) For i＝1 to m do

(3) R is a sample data randomly selected from the feature set D; H _j is the k nearest neighbors H _j found from the sample set of the same category of R, and k nearest neighbors are found in the feature set inconsistent with each category Nearest neighbor M _j (C), where j = 1, 2, ... k;

(4) for i＝1 to N do

The statistical indicators of A for a certain feature are:

Among them, p(C) is the proportion of the category; p(Class(R)) is the proportion of the category of a sample randomly selected;

(6) Sort W

After calculating the feature weight, the larger the weight, the stronger the feature's ability to distinguish samples, and a new feature subset can be selected by setting the threshold.

Step 4, design the CNN-GRU-AM network model, as follows:

The modules of the CNN-GRU-AM hybrid model are defined as follows: Convolutional neural network (CNN) is a multi-layer feed-forward artificial neural network, which has the characteristics of pooling operation, local connection and weight sharing, and is widely used in different fields. use. CNN has two significant advantages: sparse connection and weight sharing, which can extract deeper features from data and reduce the complexity of network models. The CNN structure is mainly composed of convolutional layers, pooling layers, fully connected layers, and output layers. This structure can reduce the number of weights, thereby making the network model simple, and at the same time, time series data can be directly used as the input of the network, effectively reducing the complexity of feature extraction and data reconstruction. The CNN model uses the gradient descent method to train parameters, and the trained model can learn the features in the time series data.

The gated recurrent unit (GRU) neural network is an optimized and improved variant of LSTM, which reduces training parameters while ensuring prediction accuracy. GRU contains an update gate and a reset gate. Compared with LSTM, the three gate structure parameters are less, the convergence speed is faster, and it has good data feature learning ability. The data propagation process in the GRU unit is described as follows: First, the state of the reset gate R _{t and the update gate Z t is} obtained through the last transmitted evaluation state h _t-1 and the input x _t of the current node:

r _t ＝σ(x _t W _xr +h _t-1 W _hr +b _r )

z _t ＝σ(x _t W _xr +h _t-1 W _hr +b _z )

集合中所记忆的当前时刻状态表示为：Next, currently

The current state of memory in the collection is expressed as:

Then, the GRU model updates the state by the following formula:

Finally, the output of the forward pass is:

y _t = σ(W _o h _t )

where h _t is the output of the GRU, W is the weight vector, and b is the bias vector of the GRU;

Through the last transmitted state H _t-1 and the input x _t of the current node, σ is a sigmoid function, through which the data is transformed into a value in the range of 0-1 to act as a gating signal; where H _{t- 1} contains past information, ⊙ means Hadamard product;

The implementation steps of the attention mechanism AM are shown in the following formula;

(1) Calculate the similarity between the given h _j value and the target state S _t-1 , that is, the weight of the state h _j at each time t:

e _tj = a(S _t-1 , h _j )

(2) Normalize the weight coefficients:

(3) Weighted average of state h _j :

where T is the total number of time steps in the input sequence.

Step 5, train and verify the proposed model network, and then evaluate the soft sensor method, as follows:

The 11532 real data of industrial sewage treatment plants collected in industrial sewage treatment plants are divided into 70% training set and 30% test set. The window length of each sampling point of the prediction model in this paper is set to 10, and the structure of each module includes a GRU layer with 32 nodes and a DNN layer with 16 nodes, and the epochs size is 100. In this section, three experiments are performed to evaluate the proposed model.

Specifically, the present invention will evaluate the prediction model by means of mean absolute error (MAE), mean absolute percentage error (MAPE), R ² _score and mean absolute percentage error (MAPE), and compare it with the comparative experimental model.

表示预测值，y_t表示真实值，

是真实值的平均值。本文选择上述指标对所提出模型进行估计，其中，评估预测值与真实值的基础方法为MAE与RMSE，计算后它们的数值越小，表示预测的误差值越小；MAPE不但考虑了预测值与真实值的误差，还考虑了误差与真实值的比例；R²_score的范围在0～1之间，越接近1，表示模型的预测值越接近真实值；如图2-1、图2-2、图2-3、图2-4所示。Among them, N is the length of the time series,

represents the predicted value, y _t represents the real value,

is the average of the true values. In this paper, the above indicators are selected to estimate the proposed model. Among them, the basic methods for evaluating the predicted value and the real value are MAE and RMSE. The smaller their values after calculation, the smaller the predicted error value; The error of the real value also takes into account the ratio of the error to the real value; the range of R ² _score is between 0 and 1, and the closer to 1, the closer the predicted value of the model is to the real value; as shown in Figure 2-1 and Figure 2- 2. As shown in Figure 2-3 and Figure 2-4.

Claims (5)

1. The industrial sewage treatment soft-sensing method based on feature structure optimization and deep learning is characterized in that, the method comprises the following steps: Step 1. Eliminate overlapping features for the data collected by industrial sewage treatment plants; Step 2, performing empirical mode decomposition on the data in the industrial sewage treatment plant with overlapping features eliminated; Step 3, performing feature screening on the decomposed IMFs; Step 4, design CNN-GRU-AM hybrid network model; Step 5. Train and verify the proposed model network, and output prediction indicators. 2. A kind of industrial sewage treatment soft-sensing method based on feature structure optimization and deep learning according to claim 1, characterized in that the overlapping features described in step 1 are removed, specifically as follows: Before building the model, a Pearson correlation analysis was performed on the data collected by the industrial sewage treatment plant. The strength of the correlation indicates whether there is overlap in the input information; the correlation coefficient ρ of the PCC signals x(t) and y(t) _xy is defined as:

In the formula, σ _xy is the covariance of signals x(t) and y(t), σ _x , σ _y are the standard deviations of signals x(t) and y(t) respectively; finally, the calculated correlation coefficient is Normalization, ρ _xy ∈ [-1,1], the closer the value of ρ _xy is to 1, the stronger the correlation between the two signals. Set the threshold. When ρ _xy is greater than the set threshold, it is considered as the input signal for overlapping signals. 3. the industrial sewage treatment soft-sensing method based on characteristic structure optimization and deep learning according to claim 1, is characterized in that, step 2 described feature is carried out empirical mode decomposition, specifically as follows: The complementary integrated empirical mode decomposition method is used to process the original data, combined with the feature screening algorithm, high-frequency noise interference and false components are eliminated, and the components that can represent effective information are screened out; the specific steps are as follows: Step 1. Add n pairs of positive and negative white noise ω _i (t) to the original signal s(t):

Step ②, the signal after adding noise is re-decomposed according to the EMD decomposition process, and two sets of corresponding intrinsic mode functions IMFs and residual r _i (t) are obtained, and then the mean value is calculated:

In the formula, C _ij ⁺ (t), C _ij ^- (t) represent the j-th IMF component, r _i ⁺ (t), ri _- ⁽ t) represent the i-th calculation residual. 4. a kind of industrial sewage treatment soft-sensing method based on feature structure optimization and deep learning according to claim 3, is characterized in that, the IMFs after decomposing described in step 3 is carried out characteristic screening, specifically as follows: The feature selection of the decomposed IMFs is carried out through the Relief F algorithm, specifically: select k nearest neighbor samples from the same type and different types of samples, and average them to obtain each feature weight, so as to obtain each sample instance The correlation between each feature in the class and the class; then, the features are sorted according to the weight value, and the feature is determined to be valid or invalid by setting the threshold, and the n features with the largest weight are selected, and other features are removed for feature selection. ; Proceed as follows: Let D be the feature set, m be the number of sample resampling, δ be the feature statistic index, k be the number of nearest neighbor samples, and T be the statistic index of outputting each feature; (1) Set all feature weights to 0 (2) For i＝1 to m do (3) R is a sample data randomly selected from the feature set D; H _j is the k nearest neighbors H _j found from the sample set of the same category of R, and k nearest neighbors are found in the feature set inconsistent with each category Nearest neighbor M _j (C), where j=1,2,...k; (4) for i＝1 to N do The statistical indicators of A for a certain feature are:

Among them, p(C) is the proportion of the category; p(Class(R)) is the proportion of the category of a sample randomly selected; (6) Sort W After calculating the feature weight, the larger the weight, the stronger the feature's ability to distinguish samples, and a new feature subset is selected by setting the threshold. 5. a kind of industrial sewage treatment soft-sensing method based on feature structure optimization and deep learning according to claim 4, is characterized in that: the design CNN-GRU-AM network model described in step 4, specifically as follows: The CNN structure of the convolutional neural network includes a convolutional layer, a pooling layer, a fully connected layer, and an output layer; the CNN model uses the gradient descent method to train parameters, and the trained model can learn the characteristics of time series data; The gated recurrent unit GRU neural network includes an update gate and a reset gate; the data propagation process in the GRU unit is described as follows: First, the reset is obtained through the last transmitted evaluation state h _t-1 and the input x _t of the current node The state of gate R _t and update gate Z _t : r _t ＝σ(x _t W _xr +h _t-1 W _hr +b _r ) z _t ＝σ(x _t W _xr +h _t-1 W _hr +b _z ) Next, currently

The current state of memory in the collection is expressed as:

Then, the GRU model updates the state by the following formula:

Finally, the output of the forward pass is: y _t = σ(W _o h _t ) where h _t is the output of the GRU, W is the weight vector, and b is the bias vector of the GRU; Through the last transmitted state H _t-1 and the input x _t of the current node, σ is a sigmoid function, through which the data is transformed into a value in the range of 0-1 to act as a gating signal; where H _{t- 1} contains past information, ⊙ means Hadamard product; The implementation steps of the attention mechanism AM are shown in the following formula; (1) Calculate the similarity between the given h _j value and the target state S _t-1 , that is, the weight of the state h _j at each time t: e _tj = a(S _t-1 , h _j ) (2) Normalize the weight coefficients:

(3) Weighted average of state h _j :

where T is the total number of time steps in the input sequence.

Images