A dioxin emission concentration soft measurement method based on an improved generative adversarial network

Abstract

The application provides a dioxin emission concentration soft measurement method based on an improved generative adversarial network, comprising the following steps: inputting MSWI process data into a missing data division module to obtain input features and target features; inputting the input features and the target features into an adversarial generative data filling module to fill data, obtaining filled data; inputting the filled data and dioxin emission concentration experimental data into an input-output data matching module to match data, obtaining modeling data; inputting the modeling data into an improved deep forest regression module to model and predict, obtaining a concentration prediction value. The application can improve the diversity and accuracy of dioxin emission concentration soft measurement.

Description

Technical Field

[0001] This invention relates to the field of urban solid waste incineration technology, and in particular to a soft measurement method for dioxin emission concentration based on an improved generative adversarial network. Background Technology

[0002] Compared to traditional methods such as landfill and biological treatment, municipal solid waste incineration (MSWI) technology has advantages such as harmlessness, volume reduction, and resource recovery. However, this process emits trace amounts of dioxins (DXN), an organic pollutant known as the "poison of the century," and online detection of these pollutants using soft measurement technology is currently a challenge and a hot topic in the industry.

[0003] DXN (Directional Measurement of Networks) detection is characterized by long cycles and high costs. The corresponding process data is often stored on an hourly timescale, and these historical data are frequently incomplete due to sensor malfunctions, uncertain disturbances, or human factors. For example, data from a Beijing incineration plant between 2009 and 2020 contained 2%-3% missing data; clearly, this hinders the discovery of patterns in operational data and the establishment of soft measurement models for DXN emissions. Complete modeling data is clearly the foundation for constructing effective soft measurement models and even optimizing MSWI (Maximum Sampling and Weighing Injection) process operations. Therefore, filling in missing data is crucial. Sufficient complete data and deep learning algorithms suitable for small samples are fundamental to building soft measurement models with high generalization performance. Researchers typically use criteria to identify anomalies and missing data and then delete corresponding samples or features. While this method effectively reduces the impact of data on modeling, it leads to information loss and a reduced sample size. To address these issues, some researchers use mathematical statistics and machine learning to reasonably estimate data values ​​and fill in missing regions to reduce the impact of information loss.

[0004] The DXN emission concentration during urban solid waste incineration is difficult to measure online, and the obtained modeling data suffers from small sample size. Traditional single-model algorithms such as decision trees and neural networks are limited by the number of samples, resulting in problems such as overfitting and poor generalization performance. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a soft measurement method for dioxin emission concentration based on an improved generative adversarial network.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] A soft measurement method for dioxin emission concentration based on an improved generative adversarial network includes:

[0008] The MSWI process data is input into the missing data segmentation module for segmentation to obtain input features and target features;

[0009] The input features and the target features are input into the adversarial generation data filling module for data filling to obtain filled data;

[0010] The filled data and dioxin emission concentration experimental data are input into the input-output data matching module for data matching to obtain modeling data;

[0011] The modeling data is input into the improved deep forest regression module for modeling and prediction to obtain the concentration prediction value.

[0012] Preferably, the step of inputting MSWI process data into the missing data segmentation module for segmentation to obtain input features and target features includes:

[0013] The missing values ​​and process data within a similar time range of the MSWI process data are identified through a sliding detection window and used as replacement data.

[0014] The data to be filled in is divided into complete input data and missing input data based on whether there is missing data.

[0015] Based on whether there is missing data in the features, the missing input data is divided into input features and target features of the missing input data;

[0016] The complete input data is divided into input features and target features based on whether there is missing data in the features.

[0017] Preferably, the step of inputting the input features and the target features into the adversarial generation data filling module for data filling to obtain filled data includes:

[0018] The input features of the complete input data are replaced with random noise, and the replaced data and the target features of the complete input data are input into the generator; the input of the generator is represented as: Where H is the cue matrix, and the H representation of the d-th missing feature is as follows: In the formula, h i,j Let H be the element in the i-th row and j-th column; and Z represents the input features and target features of the complete input data, respectively; Noise The noise is random; the generator consists of N G The generator is composed of a fully connected network of layers, including one input layer, one output layer, and N layers. G -2 hidden layers;

[0019] Will The generator is input to obtain the output of the first hidden layer, which is represented as follows: In the formula, ω G,1 and b G,1 These represent the weights and biases between the generator input layer and the first hidden layer, respectively; relu(·) represents the ReLU activation function; The output of the first hidden layer of the generator; the output of the nth hidden layer of the generator. G The hidden layer output is represented as The padding data output by the generator is calculated as follows: In the formula, This represents the filling result of the generator during training; and They are respectively the Nth G -2. Weights and biases between the hidden layers and the output layer; linear(·) represents the linear activation function;

[0020] A discriminator is used to distinguish between the filled data and the real data. The input to the discriminator is represented as follows:

[0021] The discriminator consists of N D It consists of a fully connected network with one input layer, one output layer, and N layers. D -2 hidden layers, The input discriminator obtains the output of the first hidden layer, which is represented as: In the formula, ω D,1 and b D,1 These are the weights and biases between the discriminator input layer and the first hidden layer, respectively. The output of the first hidden layer of the discriminator; the output of the nth hidden layer of the discriminator. D The hidden layer output is represented as The discrimination result output by the discriminator is represented as follows: In the formula, θ D Indicates the judgment result; and They are respectively the Nth D -2. Weights and biases between the hidden layers and the output layer; sigmoid(·) represents the sigmoid activation function;

[0022] Construct an objective function, which is expressed as:

[0023]

[0024] In the formula, P(ⅠX) represents Ⅰ The distribution of X; Ⅰ θ D P(Z) is the output of the discriminator for the real data; noise ) represents Z noise Distribution; Ⅲ θ DThe output of the discriminator for the padded data; λ MSE For the weights of the MSE constraint terms; f MSE (·) represents the MSE constraint term; λ G and λ D These are the weights of the model complexity penalty terms for the generator and discriminator, respectively.

[0025] The loss function is constructed using a binary cross-entropy function, and the loss function is expressed as follows:

[0026] In the formula, x a and f(x) a ) represent the model's input and output, respectively; y a The labels are binary; N′ represents the amount of data.

[0027] When training the discriminator, the generator parameters are fixed, and the real data is labeled as 1, i.e., y. a =1, and the corresponding first loss is:

[0028] Mark the filled data as 0, i.e., y a =0, and the corresponding second loss is:

[0029] The objective function for discriminator training is obtained based on the first loss, the second loss, and the discriminator complexity penalty term. The objective function for discriminator training is:

[0030]

[0031] When training the generator, the discriminator parameters are fixed, and the padded data is marked as 1, i.e., y. a =1, and the corresponding third loss is:

[0032] The objective function for generator training is obtained by combining the third loss, the MSE constraint term, and the generator complexity penalty term. The objective function for generator training is:

[0033]

[0034] Repeat the above steps, and the generator and discriminator are jointly trained N in a minimax game adversarial exercise. epoch generation;

[0035] Using random noise Z Noise Replace missing input data Input features of missing data and missing input data Combining the inputs to obtain the generator The generator completed through training The missing data is filled in to obtain the filled data.

[0036] Preferably, the filling data and dioxin emission concentration experimental data are input to the input / output data matching module for data matching to obtain modeling data, including:

[0037] Let the filling data be denoted as

[0038] The experimental data on dioxin emission concentrations are denoted as...

[0039] X STS per N scale The average of each sample is used to obtain the input data for a large time scale. The calculation formula is as follows:

[0040]

[0041]

[0042] In the formula, X LTS Input data for large time scales; N large =N / N scale For large-scale data volumes; D represents the dimension of the data to be filled.

[0043] The final modeling data is obtained by matching the large-timescale input data with the experimental data on dioxin emission concentrations, denoted as...

[0044] Preferably, the improved deep forest regression module includes parallel forest units and serial forest units. The enhanced layer regression vector output by each layer of parallel or serial forest units and the input data are fully connected as the input to the next layer of parallel or serial forest units. The output of the last layer of parallel or serial forest units is weighted and summed to obtain the output of the improved deep forest regression module. The parallel forest units consist of random forests and fully random forests. The random forests consist of multiple parallel decision trees. The serial forest units consist of gradient descent trees, extreme gradient boosting trees, and iterative DF algorithms.

[0045] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0046] This invention provides a soft measurement method for dioxin emission concentration based on an improved generative adversarial network (GAN), comprising: inputting MSWI process data into a missing data partitioning module for partitioning to obtain input features and target features; inputting the input features and target features into a generative adversarial network (GAN) data filling module for data filling to obtain filled data; inputting the filled data and dioxin emission concentration experimental data into an input-output data matching module for data matching to obtain modeling data; and inputting the modeling data into an improved deep forest regression module for modeling and prediction to obtain a concentration prediction value. This invention can improve the diversity and accuracy of soft measurement of dioxin emission concentration. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a schematic diagram of the MSWI process flow provided in an embodiment of the present invention;

[0049] Figure 2 A flowchart of the method provided in an embodiment of the present invention;

[0050] Figure 3 This is a schematic diagram of the algorithm provided in an embodiment of the present invention;

[0051] Figure 4 This is a schematic diagram illustrating the implementation process of the adversarial data filling module provided in an embodiment of the present invention.

[0052] Figure 5 This is a schematic diagram of the IDFR module implementation process provided in an embodiment of the present invention;

[0053] Figure 6 This is a schematic diagram illustrating the missing and anomalous data volume of the DXN dataset provided in an embodiment of the present invention; wherein... Figure 6 (a) is a schematic diagram of data without true values; Figure 6 (b) is a schematic diagram showing data with true values;

[0054] Figure 7 A schematic diagram of the MSE data filling for the DXN dataset provided in this embodiment of the invention;

[0055] Figure 8 This is a schematic diagram comparing the distribution of missing data for eight features and the data imputed by two methods, provided in an embodiment of the present invention; wherein... Figure 8 (a) is a schematic diagram of the first feature; Figure 8(b) is a schematic diagram of the second feature;

[0056] Figure 8 (c) is a schematic diagram of the third feature; Figure 8 (d) is a schematic diagram of the fourth feature; Figure 8 (e) is a schematic diagram of the fifth feature; Figure 8 (f) is a schematic diagram of the sixth feature; Figure 8 (g) is a schematic diagram of the seventh feature; Figure 8 (h) is a schematic diagram of the eighth feature;

[0057] Figure 9 This is a schematic diagram of the fitting curve for the soft measurement modeling comparison experiment of the DXN dataset provided in this embodiment of the invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0060] The terms "first," "second," "third," and "fourth," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, including a series of steps, processes, methods, etc., is not limited to the steps listed, but may optionally include steps not listed, or may optionally include other steps inherent to these processes, methods, products, or devices.

[0061] The purpose of this invention is to provide a soft measurement method for dioxin emission concentration based on an improved generative adversarial network, which can improve the diversity and accuracy of soft measurement of dioxin emission concentration.

[0062] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0063] A typical MSWI process mainly consists of five parts: storage and fermentation, solid waste incineration, waste heat exchange, flue gas treatment, and flue gas emission. The process is as follows: Figure 1 As shown.

[0064] Depend on Figure 1 It can be seen that: after being weighed, the transport vehicle enters the unloading platform to unload MSW into the solid waste pool. After a certain period of dehydration, fermentation, and thorough mixing and homogenization by the grab bucket, it is sent to the hopper and quantitatively supplied to the drying grate by the feeder. After primary air drying, it is sent to the combustion grate for incineration. The remaining unburned components are completely burned in the combustion grate. The remaining slag in the incinerator is cooled by the slag remover and then transported to the slag pit. At the same time, the heat generated by incineration is exchanged with the waste heat boiler to generate steam, which drives the steam turbine generator set to generate electricity. The flue gas treatment system, composed of main equipment such as activated carbon tanks, quicklime tanks, and fly ash tanks, performs desulfurization, denitrification (SNCR), dust removal, and DXN removal operations on the discharged flue gas. The final flue gas contains particulate matter, CO, CO2, NOx, SO2, HCl, DXN, and other substances, which are discharged into the atmosphere by the induced draft fan through the chimney.

[0065] MSWI process data is mainly obtained from data acquisition instruments and sensors at each stage. It primarily includes temperature, flow rate, material consumption, pressure, velocity, steam drum water level, and conventional pollutant emission concentrations. Specifically: temperature measurement points are mainly distributed above the furnace and grate to monitor combustion chamber temperature and the inner and outer temperatures of the grates in the drying, combustion, and burnout sections; flow rates mainly include the air flow rates on the left and right sides of the grate, primary and secondary air flow rates, superheater cooling water flow rate, economizer and mixer feedwater flow rate, and urea solvent supply flow rate; material... Material consumption includes the feed values ​​of the lime and activated carbon storage bins; pressure mainly includes the outlet air pressure of the primary and secondary fans, the steam drum pressure, the flue gas pressure at the outlet of the economizer and induced draft fan, and the differential pressure of the bag filter; speed variables refer to the inner and outer speeds of the feeder and grate, which are obtained by the speed detection equipment of the hydraulic drive device; conventional pollutant emission data collection includes four points: soil monitoring, groundwater monitoring, organized waste gas monitoring, and wastewater monitoring, mainly monitoring variables such as flue gas volume, flue gas flow rate, flue gas temperature, and flue gas pressure.

[0066] During the acquisition, transmission, and storage of industrial process data, data gaps are inevitable due to uncontrollable factors such as empty sampling by detection equipment, abnormal data transmission, damaged storage media, or human error. Research shows that missing data can cause biases in the results of training models, making it difficult to extract correlations between data points and hindering subsequent prediction and analysis. Therefore, data imputation is essential.

[0067] Figure 2 The method flowchart provided in the embodiments of the present invention is as follows: Figure 2As shown, the present invention also provides a soft measurement method for dioxin emission concentration based on an improved generative adversarial network, comprising:

[0068] Step 100: Input the MSWI process data into the missing data segmentation module for segmentation to obtain input features and target features;

[0069] Step 200: Input the input features and the target features into the adversarial generation data filling module to fill in the data and obtain the filled data;

[0070] Step 300: Input the filled data and dioxin emission concentration experimental data into the input-output data matching module for data matching to obtain modeling data;

[0071] Step 400: Input the modeling data into the improved deep forest regression module for modeling and prediction to obtain the concentration prediction value.

[0072] The improved deep forest regression soft sensor modeling method based on complete data proposed in this embodiment includes a missing data partitioning module, an adversarial generative data imputation module, and an improved deep forest regression module, such as... Figure 3 As shown.

[0073] Figure 3 In the model, X and Y represent MSWI process data (i.e., model input) and DXN emission concentration (i.e., model output), respectively. Part X represents the process data to be filled; Ⅰ X and Ⅱ X represents complete input data (denoted as Type I data) and missing input data (denoted as Type II data), respectively. and These represent the input features and target features of Class I data, respectively. and Z represents the input features and target features of Class II data, respectively; Noise For random noise; λ G ||ω G ||2 and λ D ||ω D ||2 represents the model complexity penalty terms for the generator and discriminator, respectively, ω G and ω D These are the generator and discriminator weights, λ. G and λ D Weights for complexity penalty terms; H represents the generator's fill-in result during training; H is the cue matrix; θ D For discriminator output; The imputation result for missing data (denoted as Type III data); X STS Input data for the smaller timescale after filling; XLTS D represents the input data over a large time scale; D represents the modeling data after matching the input and output.

[0074] Furthermore, the meanings of the symbols used in this embodiment are as follows:

[0075]

[0076]

[0077]

[0078]

[0079] The functions of the different modules are as follows:

[0080] 1) Missing Data Segmentation Module: Segment process data based on the presence or absence of missing data and features;

[0081] 2) Adversarial Data Imputation Module: The imputed data closely resembles the real data through adversarial interaction between the generator and the discriminator, as well as MSE constraint terms. A model complexity penalty term is used to reduce the impact of overfitting on model performance.

[0082] 3) Input / output data matching module: Compresses process data at a small time scale to match the true output value at a large time scale;

[0083] 4) Improved deep forest regression module: It consists of L layers of forest algorithms. Each layer uses both fully connected parallel and serial forest algorithms to obtain the final predicted value according to the weighting strategy.

[0084] Specifically, the missing data partitioning module has the following functions:

[0085] The original input data X is used to identify missing values ​​and process data within a similar time range through a sliding detection window, which are then used as imputation data. Part X, based on whether there is missing data Part X is divided into complete input data. Ⅰ X (denoted as Type I data) and missing input data Ⅱ X (denoted as Type II data) is represented as follows:

[0086]

[0087]

[0088]

[0089] In the formula, Ⅰ N and Ⅱ N are respectively Ⅰ X andⅡ The number of samples for X; N and D are respectively Part The number of samples and dimensions of X.

[0090] Based on whether there is missing data in the features Ⅱ Input features that classify X into Class II data and target features Accordingly Ⅰ Input features that classify X into Class I data and target features The target feature and input feature are the features to be filled and the features that do not need to be filled, respectively, as shown below:

[0091]

[0092]

[0093]

[0094]

[0095] In the formula, D Fea and D Tar These are the number of input and target features, respectively.

[0096] Furthermore, the specific functions of the adversarial data padding module are as follows:

[0097] and As training data for the adversarial generative data imputation model, the trained generator processes the missing data. To fill in the blanks, the process is as follows: Figure 4 As shown.

[0098] First, random noise Z is used. Noise replace Missing data in As input to the generator, it is represented as follows:

[0099]

[0100] In the formula, H is the cue matrix, and the H of the d-th missing feature is represented as follows:

[0101]

[0102] In the formula, h i,j Let H be the element in the i-th row and j-th column.

[0103] The generator is composed of N G It consists of a fully connected network with one input layer, one output layer, and N layers. G -2 hidden layers, The input generator produces the output of the first hidden layer, as follows:

[0104]

[0105] In the formula, ω G,1 and b G,1 These represent the weights and biases between the generator input layer and the first hidden layer, respectively; relu(·) represents the ReLU activation function; This is the output of the first hidden layer of the generator.

[0106] Correspondingly, the nth generator G The hidden layer output is represented as follows:

[0107]

[0108] The padding data output by the generator is calculated as follows:

[0109]

[0110] In the formula, This represents the filling result of the generator during training; and They are respectively the Nth G -2. Weights and biases between the hidden layers and the output layer; linear(·) represents the linear activation function.

[0111] Then, the discriminator distinguishes between the filled data and the real data. The discriminator input is represented as follows:

[0112]

[0113] The discriminator consists of N D It consists of a fully connected network with one input layer, one output layer, and N layers. D -2 hidden layers, The input discriminator obtains the output of the first hidden layer, as follows:

[0114]

[0115] In the formula, ω D,1 and b D,1 These are the weights and biases between the discriminator input layer and the first hidden layer, respectively. This is the output of the first hidden layer of the discriminator.

[0116] Correspondingly, the discriminator at the nth... D The hidden layer output is represented as follows:

[0117]

[0118] The discrimination result output by the discriminator is represented as follows:

[0119]

[0120] In the formula, θ D Indicates the judgment result; and They are respectively the Nth D -2. Weights and biases between the hidden and output layers; sigmoid(·) represents the sigmoid activation function.

[0121] To improve the quality of the filled data and alleviate the overfitting problem, this embodiment adds an MSE constraint term and a model complexity penalty term to the original GAN ​​objective function. The proposed objective function is expressed as follows:

[0122]

[0123] In the formula, P(ⅠX) represents Ⅰ The distribution of X; Ⅰ θ D P(Z) is the output of the discriminator for the real data; noise ) represents Z noise Distribution; Ⅲ θ D λ is the output of the discriminator for the filled data (denoted as Class III data); MSE For the weights of the MSE constraint terms; f MSE (·) represents the MSE constraint term; λ G and λ D These are the weights of the model complexity penalty terms for the generator and discriminator, respectively.

[0124] The objective function described above is achieved by using a binary cross-entropy function in the GAN's loss function, as follows:

[0125]

[0126] In the formula, x a and f(x) a ) represent the model's input and output, respectively; y a is a binary label; N′ is the amount of data.

[0127] When training the discriminator, the generator parameters are fixed, and the real data is labeled as 1, i.e., y. a =1, and the corresponding losses are as follows:

[0128]

[0129] Mark the filled data as 0, i.e., y a =0, and the corresponding loss is expressed as follows:

[0130]

[0131] Combining formulas (19) and (20) with the discriminator complexity penalty term, the objective function for discriminator training can be obtained as follows:

[0132]

[0133] When training the generator, the discriminator parameters are fixed, and the padded data is marked as 1, i.e., y. a =1, and the corresponding losses are as follows:

[0134]

[0135] Combining formula (22), the MSE constraint term, and the generator complexity penalty term, the objective function for generator training can be obtained as follows:

[0136]

[0137] Repeat the above steps, and the generator and discriminator are jointly trained N in a minimax game adversarial exercise. epoch This process allows the filled data to become increasingly closer to the real data.

[0138] Finally, random noise Z is used. Noise replace Missing data in Combining the inputs to obtain the generator The generator completed through training The missing data is filled in to obtain the filled data.

[0139] Furthermore, the input / output data matching module in this embodiment specifically functions as follows:

[0140] Real-time monitoring data from complex industrial processes is measured on a small timescale, while key operational indicators are measured offline in laboratories on a large timescale. These two sets of data cannot be directly matched to obtain modeling data. Therefore, the average of the small-timescale process data within a certain range is taken as input and matched with the output of the large-timescale operational indicators to obtain the final modeling data.

[0141] The filled small-timescale input data and large-timescale output data are denoted as follows:

[0142]

[0143]

[0144] X STS per N scale The average of each sample is used to obtain the input data for a large time scale, calculated as follows:

[0145]

[0146]

[0147] In the formula, X LTS Input data for large time scales; N large =N / N scale It represents a large amount of data over a long time scale.

[0148] The final modeling data is obtained by matching the input and output data over a large time scale, denoted as .

[0149]

[0150] Furthermore, the improved deep forest regression (IDFR) proposed in this embodiment, based on DFR, employs both serial and parallel forest algorithms at each fully connected layer to increase the diversity of DFR and prevent overfitting. The implementation process of the proposed IDFR is as follows: Figure 5 As shown.

[0151] IDFR includes both serial and parallel forest algorithms. The reinforcement layer regression vector output from each forest layer and the fully connected input data are used as the input to the next forest layer. The output of the last forest layer is weighted and summed to obtain the IDFR output.

[0152] Specifically, parallel forests consist of RF (Radical Randomization) and Completely Random Forest (CRF). RF consists of N... DT It consists of several parallel decision trees (DTs), represented as follows:

[0153]

[0154] In the formula, Let j represent the j-th DT model, and its specific modeling process is shown in reference

[32] .

[0155] CRF and RF belong to the same class of learners. Except for the fact that the DTs partitioning criterion uses completely randomization instead of a loss function, CRF is otherwise identical to RF. Therefore, CRF can be represented as f CRF (x).

[0156] More specifically, the serial forest consists of GBDT and XGBoost, and employs an iterative DF algorithm to reduce residual error. XGBoost uses a loss function... Based on this, a regularization coefficient λ is introduced, which is expressed as follows:

[0157]

[0158] In the formula, c represents the predicted value of the i-th DT. XGBoost The predicted value for the next DT is as follows:

[0159]

[0160] Therefore, the GBDT and XGBoost models are respectively represented as f GBDT (x) and f XGBoost (x). Accordingly, the first level of IDFR is represented as follows:

[0161] F1(·)={f RF (·),f CRF (·),f GBDT (·),f XGBoost (·)} (32)

[0162] The specific steps for weighted output in this embodiment are as follows:

[0163] The predicted values ​​of the four models were calculated, and the regression vector of the first enhancement layer was obtained as follows:

[0164]

[0165] Will The input to layer 2 of IDFR is combined with D. Due to the fully connected approach, the input to layer 1 (l = 3, ..., L) of IDFR is represented as...

[0166]

[0167] Finally, as the number of layers increases, training stops when the prediction accuracy on the validation set decreases, as shown below:

[0168]

[0169] In the formula, F IDFR (·) indicates the constructed IDFR model; and y represents the predicted outputs of the L-1 and L-layer IDFR models on the validation set, respectively; Val To verify the truth value of the set; f RMSE (·) indicates that the RMSE values ​​of both are calculated.

[0170] Final predicted value of IDFR The calculation is as follows:

[0171]

[0172] In the formula, The output weights for the corresponding forest are calculated as follows:

[0173]

[0174] In the formula, The RMSE value between the predicted values ​​and the true values ​​on the validation set of the corresponding forest is expressed as follows:

[0175]

[0176] In the formula, and These are the validation set predictions for IDFR layer L RF, CRF, GBDT, and XGBoost, respectively.

[0177] As an optional implementation method, this embodiment experimentally verifies the above method, specifically as follows:

[0178] The DXN data originated from a MSWI power plant in Beijing, encompassing 109 days of process data without ground truth values ​​and 58 days of modeling data with ground truth values ​​from 2009 to 2020. The input variables were 116-dimensional, and the output variable was DXN emission concentration. For each data set, the process variables were process data from the same time point, sampled hourly, resulting in 24 samples per day. Due to the randomness and uncertainty of missing DXN process data, a sliding detection window was used to fill in missing values ​​using process data from the same year and an improved GAN. After filling in missing and outlier data in the 58 days of modeling data with ground truth values, the input variables were correspondingly set to the mean of the current DXN ground truth sampling period, resulting in 141 modeling samples. These samples were randomly divided into training, validation, and test sets at 1 / 2, 1 / 4, and 1 / 4 ratios to construct the soft sensing model.

[0179] In this embodiment, root mean square error is selected as the evaluation index to compare the performance of different methods. The calculation is as follows:

[0180]

[0181] First, outliers in input variables with and without true values ​​are identified based on the 3σ criterion. The characteristics and number of outliers and missing values ​​are as follows: Figure 6 As shown.

[0182] Next, the structure and experimental setup of the adversarial generative data filling model are shown in Tables 1 and 2.

[0183] Table 1. Structure of the Adversarial Generative Data Filling Model

[0184]

[0185]

[0186] Table 2 Training settings for the adversarial generative data imputation model

[0187]

[0188] Missing input features 9–12 and 17–20 are selected to demonstrate the imputation effect. The specific meanings of the input features are shown in Table 3. The relationship between the MSE of the imputed data and the training algebra number of the real data during the training process for the eight input features is as follows: Figure 7 As shown.

[0189] Table 3 Input Feature Information for the DXN Dataset

[0190]

[0191]

[0192] Depend on Figure 7 It can be seen that the MSE of the data filled with the 8 features and the real data decreases as the number of training generations increases, and the filling effect tends to stabilize when the number of training generations reaches 700-800.

[0193] Adversarial generative imputation and linear interpolation imputation are compared to demonstrate the imputation effect of the model. The distribution of missing data for 8 features and the data imputed by the two methods are compared. Figure 8 As shown.

[0194] Depend on Figure 8 It can be seen that, compared with linear interpolation imputation, the data imputation method based on adversarial generation proposed in this embodiment has a more reasonable data distribution, especially... Figure 8 The distribution of data filled by linear interpolation in (h) is completely different from the distribution of missing data.

[0195] After filling in missing and outlier data with 58 days of modeling data with true values, the input variables were set to the mean of the current DXN true value sampling period, and the statistical information of 141 modeling samples was obtained as shown in Table 4.

[0196] Table 4 Statistical Results of the DXN Dataset

[0197]

[0198] The parameters of the IDFR soft sensor model are set as follows: minimum number of samples θ sample The number of DTs in the forest is 7. DT The number of features selected is 40, and the number of features selected is N. feature for Maximum number of layers L max The number of forest algorithms is 50, and the number of algorithms is N. forest The regularization coefficient λ is 4, and the learning rate α of GBDT and XGBoost is 1. IDFRIt is 0.3. Where N forest =4 refers to the fact that each layer of IDFR contains one RF, one CRF, one GBDT, and one XGBoost. λ only affects XGBoost, while α IDFR This only applies to GBDT and XGBoost. The IDFR output weights are shown in Table 5. Due to randomness, the IDFR soft sensor modeling experiment was repeated 30 times, and the statistical results are shown in Table 6.

[0199] Table 5. Output weight values ​​of IDFR in the DXN dataset.

[0200]

[0201] Table 6 Statistical Results of Soft Measurement Modeling Experiments on the DXN Dataset

[0202]

[0203] To verify the superiority of the proposed IDFR regression soft sensor model based on complete data, it was compared with RF, BPNN, and DFR under unprocessed and linear interpolation conditions. The parameters were set as follows: 1) RF, with a minimum number of samples per leaf node in the decision tree of 7, and the number of features selected was... 1) Decision trees: 50 nodes; 2) BPNN: 116-580-1160-116-1 neurons, sigmoid activation function in hidden layers, linear activation function in output layers, 1000 training generations, learning rate 1E-3; 3) DFR: 50 cascaded forest layers, 2 RF and 2 CRF per layer, minimum number of samples per leaf node in decision trees is 7, and the number of features selected is... The number of decision trees is 50. The specific setup for the comparative experiment is shown in Table 7, and the test set prediction curves and statistical results are as follows: Figure 9 As shown in Table 8.

[0204] Table 7. Experimental Setup for Comparison of DXN Dataset

[0205]

[0206] Table 8. Statistical Results of the Comparative Experiment on Soft Measurement Modeling for the DXN Dataset

[0207]

[0208]

[0209] As shown in Table 8:

[0210] 1) On the training set, BPNN outperforms other methods in RMSE, but the results on the validation and test sets show that BPNN suffers from severe overfitting and is less stable than other methods.

[0211] 2) On the validation set, linear interpolation to fill the data does not significantly improve the performance of RF and DFR modeling;

[0212] 3) On the test set, the RMSE of the IDFR based on complete data proposed in this invention is better than other methods, but its stability is only better than BPNN. Therefore, how to further improve the accuracy and stability of the model is a future research direction.

[0213] This invention addresses the anomalies and missing data issues in MSWI process datasets with multi-scale, small-sample characteristics. It proposes a DXN soft measurement model based on an improved GAN-DFR, with the main contributions being:

[0214] 1) A modeling framework based on GAN to fill missing values ​​in MSWI process data at small time scales and on IDFR to construct a soft measurement model of DXN emission concentration at large time scales;

[0215] 2) Data imputation based on adversarial generation can learn the data distribution in the game between the generator and the discriminator. The MSE constraint term and the model complexity penalty term adopted can effectively guarantee the quality of the imputed data.

[0216] 3) An IDRF algorithm is proposed, which improves the diversity and accuracy of the model by using both fully connected parallel and serial forest algorithms, and adopts an adaptive weight strategy to improve the generalization performance of the model. The effectiveness of the proposed method is verified using actual industrial process data.

[0217] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0218] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A dioxin emission concentration soft-sensing method based on an improved generative adversarial network, characterized in that, The method comprises the following steps: The MSWI process data is input into a missing data division module to obtain input features and target features; the MSWI process data is obtained by data acquisition instruments and sensors in each stage, including temperature, flow, material consumption, pressure, speed, drum water level and conventional pollutant emission concentration; The input features and the target features are input into an adversarial generated data filling module to obtain filling data; The filling data and dioxin emission concentration experimental data are input into an input-output data matching module to obtain modeling data; The modeling data is input into an improved deep forest regression module to obtain concentration prediction values; The method of inputting the MSWI process data into the missing data division module to obtain the input features and the target features comprises the following steps: The MSWI process data is identified by a sliding detection window to obtain missing values and process data in a similar time range as filling data; The filling data is divided into complete input data and missing input data according to whether there is missing data; The missing input data is divided into input features and target features of the missing input data according to whether there is missing data; The complete input data is divided into input features and target features of the complete input data according to whether there is missing data; The method of inputting the input features and the target features into the adversarial generated data filling module to obtain filling data comprises the following steps: input features of the complete input data are replaced by random noise, and the replaced data and target features of the complete input data are input into a generator; the input of the generator is represented as: ; wherein H is a prompt matrix, and H of the dth missing feature is represented as follows: ; in the formula, h i,j is an element of the ith row and jth column of H; and respectively represent the input features and target features of the complete input data; is random noise; the generator is composed of a full connection network of layers, the generator comprises 1 input layer, 1 output layer and hidden layers; Will The generator is input to obtain the output of the first hidden layer, which is represented as follows: In the formula, and These are the weights and biases between the generator input layer and the first hidden layer, respectively. This represents the ReLU activation function; This is the output of the first hidden layer of the generator; the generator's... The hidden layer output is represented as n G =2,...,N G -2; the padding data output by the generator is calculated as follows: In the formula, This represents the filling result of the generator during training; and The first Weights and biases between the hidden layer and the output layer; Represents a linear activation function; A discriminator is used to distinguish between the filled data and the real data. The input to the discriminator is represented as follows: The discriminator is composed of It consists of a fully connected multilayer network, including one input layer, one output layer, and... A hidden layer, The input discriminator obtains the output of the first hidden layer, which is represented as: In the formula, and These are the weights and biases between the discriminator input layer and the first hidden layer, respectively. This is the output of the first hidden layer of the discriminator; the discriminator's... The hidden layer output is represented as The discrimination result output by the discriminator is represented as follows: In the formula, Indicates the judgment result; and The first Weights and biases between the hidden layer and the output layer; This represents the Sigmoid activation function; A target function is constructed, and the target function is represented as: ； wherein represents the distribution of is the output of the discriminator for real data; represents the distribution of is the output of the discriminator for filled data; is the weight of the MSE constraint term; is the MSE constraint term; and are the model complexity penalty term weights for the generator and discriminator, respectively;​​ A binary cross-entropy function is used to construct a loss function, which is represented as: ; wherein, and represent the input and output of the model, respectively; is a binary label; is the data volume; When training the discriminator, the generator parameters are fixed, and the real data is labeled as 1, i.e. , and the corresponding first loss is: ; The padding data is marked as 0, i.e. , and the corresponding second loss is: ; A target function for discriminator training is obtained according to the first loss, the second loss and the discriminator complexity penalty term, and the target function for discriminator training is: ； When training the generator, the discriminator parameters are fixed, and the padding data is labeled as 1, i.e. , and the corresponding third loss is: ; The target function of the generator training is obtained by combining the third loss, the MSE constraint term and the generator complexity penalty term, and the target function of the generator training is: ; The generator and the discriminator are trained together in a minimax game generation of the second generation; Adopting random noise Replacing missing input data with missing data and input features of the missing input data Combining to obtain the input of the generator , filling the missing data of the by the completed generator, to obtain the filled data .​ 2. The dioxin emission concentration soft sensing method based on the improved generative adversarial network according to claim 1, characterized in that, The method of inputting the filling data and the dioxin emission concentration experimental data into the input-output data matching module to obtain modeling data comprises the following steps: The padding data is denoted as ; The dioxin emission concentration experimental data is recorded as ; The Each The average of each sample is taken to obtain large time scale input data, and the calculation formula is: ； ； wherein is the large time scale input data; is the large time scale data volume; D is the dimension of data to be filled; The large time scale input data and the dioxin emission concentration experimental data are matched to obtain final modeling data, denoted as .

3. The dioxin emission concentration soft sensing method based on the improved generative adversarial network according to claim 2, characterized in that, The improved deep forest regression module comprises parallel forest units and serial forest units, an enhanced layer regression vector output by each layer of parallel forest units or serial forest units is fully connected with input data to serve as input of the next layer of parallel forest units or serial forest units, a weighted sum of outputs of the last layer of parallel forest units or serial forest units serves as output of the improved deep forest regression module; the parallel forest units are composed of random forests and complete random forests; the random forests are composed of multiple parallel decision trees; the serial forest units are composed of gradient descent trees, extremely gradient boosting trees and iterative DF algorithms.

Images