Soft sensing method of dioxin emission in MSWI process based on width-mixed forest regression
Through soft measurement methods based on width hybrid forest regression, models of feature map layer, potential feature extraction layer, feature enhancement layer and incremental learning layer are constructed, and the real-time online detection problem of dioxin emission concentration detection in MSWI is solved, and the realization of the real-time online detection problem of dioxin emission concentration detection is achieved. High-precision and fast-responsive DXN concentration detection.
Patent Information
- Application Number
- CN202210059984.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-19
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-01-19
AI Technical Summary
The prior art is difficult to realize real-time online detection of dioxin (DXN) emission concentrations during MSWI, and there are problems such as complex equipment, high cost, many interference factors, and unavailability of prediction accuracy.
Using soft measurement methods based on width hybrid forest regression, a BHFR model including feature map layer, latent feature extraction layer, feature enhancement layer and incremental learning layer was constructed. Modeled through non-differential basis learners and hybrid forest groups, using information Fully connected and mutual information measure and process feature information to achieve high-precision DXN concentration detection.
It realizes high-precision online detection of DXN concentration, reduces model complexity and calculation consumption, improves the rapid response and stability of detection, and is suitable for industrial processes with small samples and high-dimensional data.
Smart Images

Figure CN114398836B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of dioxin emission soft measurement, and in particular to a dioxin emission soft measurement method of a MSWI process based on wide mixed forest regression. Background Art
[0002] Municipal Solid Waste Incineration (MSWI) is one of the main ways to solve the "garbage siege" dilemma in cities worldwide. It has significant advantages such as harmlessness, reduction and resource utilization. Dioxin (DXN), as a persistent and highly toxic organic pollutant in the organized waste gas emitted by the MSWI process, is the main reason for the "NIMBY phenomenon" in the construction of incineration plants, and is also one of the important environmental protection indicators that must be minimized in the MSWI process. The offline analysis method based on high-resolution gas chromatography-high-resolution mass spectrometry (HRGC / HRMS) is currently the main means for detecting DXN emission concentrations. It has the disadvantages of high technical difficulty, large time lag, high manpower and economic costs, and has become one of the key factors hindering the real-time optimization control of the MSWI process. Therefore, the online detection of DXN emission concentrations has become the primary challenge of the MSWI process.
[0003] In view of the above problems, the online indirect detection method of using DXN related substances that can be detected online to build a correlation model and then indirectly obtain the DXN concentration has become a hot topic; however, it has problems such as complex equipment, high cost, many interference factors, and the inability to guarantee prediction accuracy. At the same time, it is essentially a detection method combined with data modeling. Compared with offline analysis and online indirect detection methods, soft measurement technology driven by easy-to-detect process data collected by industrial distributed control systems is an effective way to solve the problem that DXN cannot be detected online. It has the characteristics of stability, accuracy and rapid response. Soft measurement technology has been widely used in the detection of difficult-to-measure parameters in complex industrial processes such as petroleum, chemical industry and steelmaking. Summary of the invention
[0004] The purpose of the present invention is to provide a soft sensing method for dioxin emission in MSWI process based on broad hybrid forest regression. With the detection of DXN emission concentration in MSWI process as the goal, a soft sensing modeling algorithm based on broad hybrid forest regression (BHFR) is proposed.
[0005] To achieve the above object, the present invention provides the following solutions:
[0006] A soft-sensing method for dioxin emission in MSWI process based on wide mixed forest regression is proposed. Based on the BLS framework, a BHFR soft-sensing model for small sample high-dimensional data is constructed by replacing neurons with non-differential basis learners. The BHFR soft-sensing model includes the construction of feature mapping layer, potential feature extraction layer, feature enhancement layer and incremental learning layer, and specifically includes the following steps:
[0007] S1, construct a feature mapping layer, construct a mixed forest group consisting of random forest RF and completely random forest CRF to map high-dimensional features;
[0008] S2, constructs a potential feature extraction layer, extracts potential features from the feature space of the fully connected mixed matrix according to the contribution rate, and ensures the maximum transmission of potential valuable information and minimizes redundancy based on the information measurement criterion, thereby reducing model complexity and computational consumption;
[0009] S3, constructing a feature enhancement layer, and training the feature enhancement layer based on the extracted latent features to further enhance the feature representation capability;
[0010] S4, constructing the incremental learning layer, constructing the incremental learning layer through the incremental learning strategy, using Moore-Penrose pseudo-inverse to obtain the weight matrix, and then realizing the high-precision modeling of the BHFR soft sensor model;
[0011] S5, using a high-dimensional benchmark dataset and an industrial process DXN dataset to verify the soft sensor model;
[0012] S6, using the soft measurement model established in steps S1-S5 to perform soft measurement of dioxin emissions from the MSWI process.
[0013] Furthermore, the step S1 constructs a feature mapping layer, constructs a mixed forest group consisting of a random forest RF and a completely random forest CRF to map high-dimensional features, specifically including:
[0014] Assume the original data is {X,y}, where is the original input data, N Raw is the number of raw data, M is the dimension of the raw input data, which comes from the six different stages of the MSWI process, collected and stored in the DCS system in seconds, is the true output value of DXN emission concentration, which comes from the emission DXN detection sample obtained by offline detection method; the modeling process of the feature mapping layer is described by taking the nth mixed forest group of the feature mapping layer as an example:
[0015] Bootstrap and random subspace RSM sampling are performed on {X, y} to obtain J training subsets of the mixed forest group model as follows:
[0016]
[0017] in, and is the input and output of the Jth training subset, and represents the Bootstrap and RSM sampling of the nth mixed forest group in the feature map layer, P Bootstrap Indicates Bootstrap sampling probability;
[0018] based on Train a mixed forest algorithm containing J decision trees, where the jth decision tree of the nth mixed forest group in the feature map layer is expressed as follows:
[0019]
[0020] Where L represents the number of decision tree leaf nodes, I(·) represents the indicator function, and c l The calculation is done by recursive splitting method;
[0021] Split loss function Ω for decision tree in RF i (·) is expressed as:
[0022]
[0023] Among them, Ω i (s,v) represents the loss function value of the sth feature value v as the segmentation criterion, y L represents the true value vector of DXN emission concentration of the left leaf node, E[y L ] means y L The mathematical expectation of R represents the true value vector of DXN emission concentration of the right leaf node, E[y R ] means y R The mathematical expectation of represents the true value of the DXN emission concentration of the i-th left leaf node, represents the true value of the DXN emission concentration of the i-th right leaf node, c L represents the emission concentration prediction output of the left leaf node DXN, c R It represents the emission concentration prediction output of the right leaf node DXN;
[0024] By minimizing Ω i (s,v), the training set Divide into two tree nodes, as follows:
[0025]
[0026] in, and Represents the sample set contained in the left and right tree nodes after segmentation, N L and N R Respectively and The number of samples in ;
[0027] DXN emission concentration prediction output value of the current left and right tree nodes and is the expectation of the true value of the sample, as follows:
[0028]
[0029] Among them, y L and R express and The true value vector of DXN emission concentration in E[y L ] and E[y R ] means y L and R The mathematical expectation of
[0030] Unlike RF, the decision tree splitting in CRF adopts a completely random selection method, which is expressed as,
[0031]
[0032] in, It means that the value v of the sth feature is selected completely randomly as the split point;
[0033] DXN emission concentration prediction output values of randomly split left and right tree nodes and is the expectation of the true value of the sample, as follows:
[0034]
[0035] Through the above process, the nth mixed forest group It can be expressed as,
[0036]
[0037] in, represents the nth random forest, represents the nth completely random forest; furthermore, the nth mapping feature Z n It can be expressed as
[0038]
[0039] in, represents the mapping features of the nth group of mixed forests to the first sample of the original input data from six different stages of the MSWI process, The nth group of mixed forests is derived from the original input data of the six different stages of the MSWI process. Raw The mapping features of the th sample, It represents the nth group of mixed forests for the original input data from six different stages of the MSWI process. Raw The mapping features of the th sample;
[0040] Finally, the output of the feature mapping layer is expressed as:
[0041]
[0042] Among them, Z1 is the first mapping feature, Z2 is the second mapping feature, and Z N is the Nth mapping feature, the mapping feature matrix Z N Contains N Raw samples and 2N-dimensional features.
[0043] Furthermore, the step S2 constructs a potential feature extraction layer, extracts potential features from the feature space of the fully connected mixed matrix according to the contribution rate, ensures the maximum transmission of potentially valuable information and minimizes redundancy based on the information metric criterion, and reduces the model complexity and computational consumption, specifically including:
[0044] First, the original input data X and feature map matrix Z from the six different stages of the MSWI process N The fully connected mixing matrix A is obtained by combination, which is expressed as:
[0045]
[0046] Among them, A contains N Raw samples and (M+2N)-dimensional features;
[0047] Next, considering that the dimension of A is much higher than the original data, PCA is used here to minimize the redundant information in A and calculate the correlation matrix R of A as follows:
[0048]
[0049] Furthermore, singular value decomposition is performed on R to obtain (M+2N) eigenvalues and corresponding eigenvectors, as follows:
[0050] R=U (M+2N) Σ (M+2N) V (M+2N) (13)
[0051] Among them, U (M+2N)represents an (M+2N)-order orthogonal matrix, Σ (M+2N) represents a (M+2N)-order diagonal matrix, V (M+2N) represents an orthogonal matrix of order (M+2N);
[0052]
[0053] Among them, σ1>σ2>…>σ (M+2N) Indicates the eigenvalues arranged from large to small;
[0054] Then, according to the potential feature contribution threshold η, the final number of principal components is determined.
[0055]
[0056] Among them, the number of potential features Q PCA <<(M+2N);
[0057] Based on the above determined Q PCA potential features, and obtain a set of feature values The corresponding eigenvector matrix V QPCA , which is the projection matrix of A; then, feature projection is performed on A to minimize redundant information, and the potential features are recorded as X PCA ,Right now
[0058]
[0059] in, Indicates the front Q PCA The feature vector of potential features;
[0060] Further, calculate the selected latent feature X PCA and truth value The mutual information value I MI ,as follows:
[0061]
[0062] in, represents the qth potential feature The joint probability distribution of the true value y of the DXN emission concentration is, represents the qth potential feature The marginal probability distribution of , p(y) represents the marginal probability distribution of the true value y of DXN emission concentration;
[0063] Next, the information maximization selection mechanism is used to ensure the relevance of the selected potential features to the true value, which is expressed as:
[0064]
[0065] in, Indicates Q PCA Potential features The mutual information value with the true value y, ζ represents the threshold value that maximizes the information, Indicates the highest correlation with the true value y of DXN emission concentration potential characteristics;
[0066] Finally, we obtain A new dataset of potential features And set the extracted dimension
[0067] Furthermore, in step S3, a feature enhancement layer is constructed, and the feature enhancement layer is trained based on the extracted potential features to further enhance the feature representation capability, specifically including:
[0068] First, the new data set {X′, y} is sampled based on Bootstrap and RSM to obtain the Jth training subset of the mixed forest algorithm, as follows:
[0069]
[0070] in, and is the input and output of the Jth training subset, X′ and y are the input and output of the new training set, represents Bootstrap sampling of the kth mixed forest group, represents the RSM sampling of the kth mixed forest group;
[0071] Next, take the construction of the jth RF in the kth mixed forest group as an example, as follows:
[0072]
[0073] in, represents the jth decision tree of the RF in the kth mixed forest group in the feature enhancement layer; L represents the number of decision tree leaf nodes; c l The calculation is performed using a recursive splitting method, and the specific process is shown in formulas (3)-(5);
[0074] Then, the RF model in the kth mixed forest group in the feature enhancement layer can be obtained, which is expressed as,
[0075]
[0076] Then, similarly, taking the construction of the j-th CRF in the k-th mixed forest group as an example, it is as follows:
[0077]
[0078] in, represents the jth decision tree of CRF in the kth mixed forest group in the feature enhancement layer; c l The calculation is performed using a recursive splitting method. The specific process is shown in formulas (6)-(7);
[0079] Then, the CRF model of the kth mixed forest group in the feature enhancement layer can be obtained, which is expressed as:
[0080]
[0081] Through the above process, we get the kth mixed forest group Then, the kth enhanced feature can be expressed as follows:
[0082]
[0083] in, represents the enhanced mapping of the kth mixed forest group to the first sample in the new data, represents the kth mixed forest group for the nth Raw The enhanced map of the th sample, Indicates the kth mixed forest group for the Nth in the new data Raw The enhanced mapping of the th sample;
[0084] Finally, the output H of the feature enhancement layer K It is expressed as follows:
[0085]
[0086] Among them, H1 is the first enhanced feature, H2 is the second enhanced feature, and H K is the Kth enhanced feature;
[0087] When the incremental learning strategy is not considered, the BHFR model is expressed as follows:
[0088]
[0089] Among them, G K Represents the combination of the output of the feature mapping layer and the feature enhancement layer, that is, G K =[Z N |H K ], which contains N Raw samples and (2N+2K)-dimensional features; W K represents the weights between the feature mapping layer, feature enhancement layer and output layer, which is calculated as follows:
[0090] W K =(λΙ+[GK ] T G K ) -1 [G K ] T Y (27)
[0091] Among them, Ι represents the unit matrix, λ represents the regularization term coefficient; accordingly, G K The pseudo-inverse calculation can be expressed as:
[0092]
[0093] Furthermore, the step S4 constructs an incremental learning layer, constructs an incremental learning layer through an incremental learning strategy, and uses Moore-Penrose pseudo-inverse to obtain a weight matrix, thereby realizing high-precision modeling of the BHFR soft sensor model, specifically including:
[0094] First, the new data set {X′, y} is sampled based on Bootstrap and RSM to obtain the mixed forest algorithm training subset. The process is as follows:
[0095]
[0096] in, and is the input and output of the Jth training subset of the mixed forest algorithm, X′ and y are the input and output of the new training set, and Represents Bootstrap sampling and RSM sampling of the pth mixed forest group in the incremental learning layer;
[0097] Next, construct the decision tree in the pth mixed forest group and The process is the same as that of the feature mapping layer and feature increment layer, so it will not be repeated here;
[0098] Furthermore, when one mixed forest group is added, the output G of the feature mapping layer, feature increment layer and incremental learning layer is K+1 It is expressed as follows:
[0099]
[0100] Among them, G k =[Z n |H k ] contains N Raw samples and (2N+2K)-dimensional features, G K+1 Contains N Raw samples and (2N+2K+2J)-dimensional features;
[0101] Then, perform GK+1 The recursive update of the Moore-Penrose inverse matrix is as follows:
[0102]
[0103] Among them, the calculation of matrix C and matrix D is as follows:
[0104]
[0105] Furthermore, G K+1 The recursive formula for the Moore-Penrose inverse matrix is as follows:
[0106]
[0107] Further, the update matrix W of the weights between the feature mapping layer, feature increment layer, incremental learning layer and output layer is calculated K+1 ,as follows:
[0108]
[0109] Among them, W K =(λΙ+[G K ] T G K ) -1 [G K ] T Y;
[0110] Since the above pseudo-inverse update strategy only needs to calculate the pseudo-inverse matrix of the mixed forest group of the incremental learning layer, fast incremental learning can be achieved;
[0111] Furthermore, adaptive incremental learning is achieved according to the convergence degree of the training error;
[0112] Define the convergence threshold of error as θ Con It is used to determine the number p of mixed forest groups in incremental learning; accordingly, the incremental learning training error of the BHFR model is expressed as follows:
[0113]
[0114] Among them, l represents the training error value of the incremental learning of the p+1th and pth mixed forest groups, and represents the training error of the BHFR model containing p and p + 1 mixed forest groups;
[0115] Finally, the prediction output of the proposed BHFR soft sensor model is for,
[0116]
[0117] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects: the soft measurement method of dioxin emission in the MSWI process based on width mixed forest regression provided by the present invention establishes a soft measurement model based on BHFR, which combines algorithms such as width learning modeling, ensemble learning and potential feature extraction. 1) Based on the width learning system framework, a soft measurement model including feature mapping layer, potential feature extraction layer, feature enhancement layer and incremental learning layer is constructed by using a non-differential learner; 2) The internal information of the BHFR model is processed by using information full connection, potential feature extraction and mutual information measurement, which effectively ensures the maximization of the transmission of internal feature information of the BHFR model and the minimization of redundancy; 3) The mixed forest group is used as the mapping unit to realize the incremental learning of the modeling process, and the output layer weight matrix is quickly calculated by the pseudo-inverse strategy, and the incremental learning is adaptively adjusted by using the convergence degree of the training error, so as to realize high-precision soft measurement modeling. The effectiveness and rationality of the proposed method are verified on high-dimensional benchmark datasets and industrial process DXN datasets. BRIEF DESCRIPTION OF THE DRAWINGS
[0118] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0119] Figure 1 This is a flow chart of a soft-sensing method for dioxin emission in a MSWI process based on width mixed forest regression according to an embodiment of the present invention;
[0120] Figure 2 This is a process flow chart of the municipal solid waste incineration process according to an embodiment of the present invention;
[0121] Figure 3 is a training error convergence curve of an embodiment of the present invention;
[0122] Figure 4a is the fitting curve of the training set in the DXN data set of the embodiment of the present invention;
[0123] Figure 4b is the fitting curve of the validation set in the DXN data set of the embodiment of the present invention;
[0124] Figure 4c It is the fitting curve of the test set in the DXN data set of the embodiment of the present invention. DETAILED DESCRIPTION
[0125] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0126] The purpose of the present invention is to provide a soft sensing method for dioxin emission in MSWI process based on broad hybrid forest regression. With the detection of DXN emission concentration in MSWI process as the goal, a soft sensing modeling algorithm based on broad hybrid forest regression (BHFR) is proposed.
[0127] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.
[0128] like Figure 1 As shown, the soft sensing method for dioxin emission in the MSWI process based on width mixed forest regression provided by the present invention comprises the following steps:
[0129] Based on the BLS framework, a BHFR soft-sensing model for small sample high-dimensional data is constructed by replacing neurons with non-differentiable basis learners. The BHFR soft-sensing model includes the construction of feature mapping layer, potential feature extraction layer, feature enhancement layer and incremental learning layer, which specifically includes the following steps:
[0130] S1, construct a feature mapping layer, construct a mixed forest group consisting of random forest RF and completely random forest CRF to map high-dimensional features;
[0131] S2, constructs a potential feature extraction layer, extracts potential features from the feature space of the fully connected mixed matrix according to the contribution rate, and ensures the maximum transmission of potential valuable information and minimizes redundancy based on the information measurement criterion, thereby reducing model complexity and computational consumption;
[0132] S3, constructing a feature enhancement layer, and training the feature enhancement layer based on the extracted latent features to further enhance the feature representation capability;
[0133] S4, constructing the incremental learning layer, constructing the incremental learning layer through the incremental learning strategy, using Moore-Penrose pseudo-inverse to obtain the weight matrix, and then realizing the high-precision modeling of the BHFR soft sensor model;
[0134] S5, using a high-dimensional benchmark dataset and an industrial process DXN dataset to verify the soft sensor model;
[0135] S6, using the soft measurement model established in steps S1-S5 to perform soft measurement of dioxin emissions from the MSWI process.
[0136] The MSWI process includes solid waste storage and transportation, solid waste incineration, waste heat boiler, steam power generation, flue gas purification and flue gas emission. Taking the grate-type MSWI process with a daily processing capacity of 800 tons as an example, its process flow is as follows: Figure 2 shown.
[0137] Combining the whole process of DXN decomposition, generation, adsorption and emission, the main functions of each stage are described as follows:
[0138] 1) Solid waste storage and transportation stage: sanitation vehicles transport MSW from various collection stations in the city to the MSWI power plant, and after weighing and recording, they are dumped from the unloading platform into the unfermented area of the solid waste storage pool. The solid waste grab is then used to mix and stir it, and then it is grabbed to the fermentation area. After 3 to 7 days of fermentation and dehydration, the low calorific value of MSW incineration is guaranteed. Studies have shown that raw MSW contains trace amounts of DXN (about 0.8ng TEQ / Kg) and contains a variety of chlorine-containing compounds required for the DXN production reaction.
[0139] 2) Solid waste incineration stage: The solid waste grab bucket puts the fermented MSW into the feed hopper, and pushes the MSW into the incinerator through the feeder. After drying, combustion 1, combustion 2 and the ember grate, the combustible components in the MSW are completely burned; the required combustion air is injected from the bottom of the grate and the middle of the furnace by the primary fan and the secondary fan, and the ash produced by the final combustion falls from the end of the ember grate to the slag remover, and is sent to the slag pool after water cooling. In order to ensure that the DXN contained in the original MSW and generated during incineration can be completely decomposed under the high-temperature combustion conditions in the furnace, the furnace combustion process must strictly control the flue gas temperature to be above 850°C, the high-temperature flue gas stays in the furnace for more than 2 seconds, and ensures a sufficiently large flue gas turbulence and other process requirements.
[0140] 3) Waste heat boiler stage: The high-temperature flue gas (higher than 850°C) generated in the furnace is sucked into the waste heat boiler system by the induced draft fan, and passes through the superheater, evaporator and economizer equipment. The high-temperature flue gas exchanges heat with the liquid water in the boiler drum to generate high-temperature steam, thereby cooling the high-temperature flue gas, so that the flue gas temperature at the outlet of the waste heat boiler is lower than 200°C (i.e., flue gas G1). From the perspective of the DXN generation mechanism, when the high-temperature flue gas is cooled by the waste heat boiler, the chemical reactions that lead to the generation of DXN include high-temperature gas-phase synthesis reactions (800°C to 500°C), precursor synthesis (450°C to 200°C) and de novo synthesis (350°C to 250°C), but there is no unified conclusion yet.
[0141] 4) Steam power generation stage: The high-temperature steam generated by the waste heat boiler is used to drive the steam turbine generator, converting mechanical energy into electrical energy, achieving self-sufficiency in plant-level electricity consumption and supplying surplus electricity to the grid, realizing resource utilization and obtaining economic benefits.
[0142] 5) Flue gas purification stage: Flue gas purification in the MSWI process mainly includes denitrification (NO x ), desulfurization (HCL, HF, SO2, etc.), removal of heavy metals (Pb, Hg, Cd, etc.), adsorption of dioxins (DXN) and dust removal (particulate matter), so as to achieve the purpose of meeting the emission standards of incineration flue gas pollutants. The use of activated carbon injection system to adsorb DXN in incineration flue gas is currently the most widely used technical means, and the adsorbed DXN is enriched in fly ash.
[0143] 6) Flue gas emission stage: The combustion flue gas (i.e. flue gas G2) containing trace amounts of DXN after cooling and purification is sucked by the induced draft fan and discharged into the atmosphere through the chimney. The uninterrupted and long-term operation characteristics of the MSWI process lead to a large amount of DXN attached to the particles on the inner wall of the chimney (i.e. memory effect). Under what operating conditions is there a possibility of release, which is still a research problem.
[0144] At present, the research on DXN soft measurement detection for MSWI process mainly focuses on the DXN concentration detection in the emission stage (i.e. flue gas G3). The research focus of this application is to construct a soft measurement model at G3 flue gas.
[0145] The BHFR modeling strategy proposed in this application includes four main parts: feature mapping layer, potential feature extraction layer, feature enhancement layer and incremental learning layer.
[0146] like Figure 1 middle, represents the original data, where is the original input data, N Raw is the number of raw data, M is the dimension of the raw input data, which comes from the six different stages of the above MSWI process, collected and stored in the DCS system in seconds, is the true output value of DXN emission concentration, which is derived from the emission dioxin DXN detection sample obtained by offline detection method; {DT1,…,DT J} represents the J decision tree models in the mixed forest algorithm, DT1 is the first decision tree model, DT J is the J-th decision tree model; Bootstrap and RSM represent sample and feature sampling of input data; {RF n ,CRF n} represents the nth mixed forest group model, RF n and CRF n Represents the nth RF and CRF model; Indicates that the feature map layer contains N mixed forest group models; Z N represents the output of the feature mapping layer; H K represents the output of the feature enhancement layer; [X|Z N ] indicates the original data and Z N The fully connected mixing matrix of Represents new training data after potential feature extraction; Represents the K mixed forest group models contained in the feature enhancement layer; represents the P mixed forest group models contained in the incremental learning layer; W K+P Represents the final weight matrix.
[0147] The main functions of each part are as follows:
[0148] 1) Feature Mapping Layer: The original input data from the six different stages of the MSWI process are Through the feature map layer N hybrid forest group Perform feature mapping to obtain the mapping output matrix Z N ;
[0149] 2) Potential feature extraction layer: Use principal component analysis to extract the original input data With the feature map layer output Z N The fully connected mixing matrix [X|Z N ] to extract potential features, remove redundant information in the feature space, and further determine the potential feature dimension through the mutual information between the extracted potential features and the output true value y of the DXN emission concentration to obtain a new training set
[0150] 3) Feature enhancement layer: using new training set As input, it passes through the K hybrid forest groups of the feature enhancement layer The group performs feature mapping to obtain the enhanced layer output matrix H K ;
[0151] 4) Incremental learning layer: using new training sets As input, the weight W is gradually increased and updated with the mixed forest group as the smallest unit. K+P , until the training error converges.
[0152] In essence, BHFR replaces neurons in the original BLS with a mixed forest group consisting of RF and CRF as primitives as the basic mapping unit; the step S1 constructs a feature mapping layer, constructs a mixed forest group consisting of a random forest RF and a completely random forest CRF to map high-dimensional features, specifically including:
[0153] Assume the original data is {X,y}, where is the original input data, N Raw is the number of raw data, M is the dimension of the raw input data, which comes from the six different stages of the MSWI process, collected and stored in the DCS system in seconds, is the true output value of DXN emission concentration, which comes from the emission DXN detection sample obtained by offline detection method; the modeling process of the feature mapping layer is described by taking the nth mixed forest group of the feature mapping layer as an example:
[0154] Bootstrap and random subspace RSM sampling are performed on {X, y} to obtain J training subsets of the mixed forest group model as follows:
[0155]
[0156] in, and is the input and output of the Jth training subset, and represents the Bootstrap and RSM sampling of the nth mixed forest group in the feature map layer, P Bootstrap Indicates Bootstrap sampling probability;
[0157] based on Train a mixed forest algorithm containing J decision trees, where the jth decision tree of the nth mixed forest group in the feature map layer is expressed as follows:
[0158]
[0159] Where L represents the number of decision tree leaf nodes, I(·) represents the indicator function, and c l The calculation is done by recursive splitting method;
[0160] Split loss function Ω for decision tree in RF i (·) is expressed as:
[0161]
[0162] Among them, Ω i (s,v) represents the loss function value of the sth feature value v as the segmentation criterion, y L represents the true value vector of DXN emission concentration of the left leaf node, E[y L ] means y L The mathematical expectation of R represents the true value vector of DXN emission concentration of the right leaf node, E[y R ] means y R The mathematical expectation of represents the true value of the DXN emission concentration of the i-th left leaf node, represents the true value of the DXN emission concentration of the i-th right leaf node, c L represents the emission concentration prediction output of the left leaf node DXN, c R It represents the emission concentration prediction output of the right leaf node DXN;
[0163] By minimizing Ω i (s,v), the training set Divide into two tree nodes, as follows:
[0164]
[0165] in, and Represents the sample set contained in the left and right tree nodes after segmentation, N L and N R Respectively and The number of samples in ;
[0166] DXN emission concentration prediction output value of the current left and right tree nodes and is the expectation of the true value of the sample, as follows:
[0167]
[0168] Among them, y L and R express and The true value vector of DXN emission concentration in E[y L ] and E[y R ] means y L and R The mathematical expectation of
[0169] Unlike RF, the decision tree splitting in CRF adopts a completely random selection method, which is expressed as,
[0170]
[0171] in, It means that the value v of the sth feature is selected completely randomly as the split point;
[0172] DXN emission concentration prediction output values of randomly split left and right tree nodes and is the expectation of the true value of the sample, as follows:
[0173]
[0174] Through the above process, the nth mixed forest group It can be expressed as,
[0175]
[0176] in, represents the nth random forest, represents the nth completely random forest; furthermore, the nth mapping feature Z n It can be expressed as
[0177]
[0178] in, represents the mapping features of the nth group of mixed forests to the first sample of the original input data from six different stages of the MSWI process, The nth group of mixed forests is derived from the original input data of the six different stages of the MSWI process. Raw The mapping features of the th sample, It represents the nth group of mixed forests for the original input data from six different stages of the MSWI process. Raw The mapping features of the th sample;
[0179] Finally, the output of the feature mapping layer is expressed as:
[0180]
[0181] Among them, Z1 is the first mapping feature, Z2 is the second mapping feature, and Z N is the Nth mapping feature, the mapping feature matrix Z N Contains N Raw samples and 2N-dimensional features.
[0182] In order to avoid overfitting caused by information loss during information transmission, the BHFR proposed in this application adopts a fully connected strategy to realize information transmission between the feature mapping layer and the feature enhancement layer and the incremental learning layer. At the same time, in order to minimize information redundancy during model training, principal component analysis (PCA) is used here to extract the potential features of the fully connected mixed matrix feature space, and then mutual information is used to further screen the potential features related to the maximization of true value information, thereby realizing dimensionality reduction processing of high-dimensional data.
[0183] The step S2 constructs a potential feature extraction layer, extracts potential features from the feature space of the fully connected mixed matrix according to the contribution rate, ensures the maximum transmission of potential valuable information and minimizes redundancy based on the information measurement criterion, and reduces the model complexity and computational consumption, specifically including:
[0184] First, the original input data X and feature map matrix Z from the six different stages of the MSWI process N The fully connected mixing matrix A is obtained by combination, which is expressed as:
[0185]
[0186] Among them, A contains N Raw samples and (M+2N)-dimensional features;
[0187] Next, considering that the dimension of A is much higher than the original data, PCA is used here to minimize the redundant information in A and calculate the correlation matrix R of A as follows:
[0188]
[0189] Furthermore, singular value decomposition is performed on R to obtain (M+2N) eigenvalues and corresponding eigenvectors, as follows:
[0190] R=U (M+2N) Σ (M+2N) V (M+2N) (13)
[0191] Among them, U (M+2N) represents an (M+2N)-order orthogonal matrix, Σ (M+2N) represents a (M+2N)-order diagonal matrix, V (M+2N) represents an orthogonal matrix of order (M+2N);
[0192]
[0193] Among them, σ1>σ2>…>σ (M+2N) Indicates the eigenvalues arranged from large to small;
[0194] Then, according to the potential feature contribution threshold η, the final number of principal components is determined.
[0195]
[0196] Among them, the number of potential features Q PCA <<(M+2N);
[0197] Based on the above determined Q PCA potential features, and obtain a set of feature values The corresponding eigenvector matrix V QPCA , which is the projection matrix of A; then, feature projection is performed on A to minimize redundant information, and the potential features are recorded as X PCA ,Right now
[0198]
[0199] in, Indicates the front Q PCA The feature vector of potential features;
[0200] Further, calculate the selected latent feature X PCA and truth value The mutual information value I MI ,as follows:
[0201]
[0202] in, represents the qth potential feature The joint probability distribution of the true value y of the DXN emission concentration is, represents the qth potential feature The marginal probability distribution of , p(y) represents the marginal probability distribution of the true value y of DXN emission concentration;
[0203] Next, the information maximization selection mechanism is used to ensure the relevance of the selected potential features to the true value, which is expressed as:
[0204]
[0205] in, Indicates Q PCA Potential features The mutual information value with the true value y, ζ represents the threshold value that maximizes the information, Indicates the highest correlation with the true value y of DXN emission concentration potential characteristics;
[0206] Finally, we obtain A new dataset of potential features And set the extracted dimension
[0207] In step S3, a feature enhancement layer is constructed, and the feature enhancement layer is trained based on the extracted potential features to further enhance the feature representation capability, specifically including:
[0208] First, the new data set {X′, y} is sampled based on Bootstrap and RSM to obtain the Jth training subset of the mixed forest algorithm, as follows:
[0209]
[0210] in, and is the input and output of the Jth training subset, X′ and y are the input and output of the new training set, represents Bootstrap sampling of the kth mixed forest group, represents the RSM sampling of the kth mixed forest group;
[0211] Next, take the construction of the jth RF in the kth mixed forest group as an example, as follows:
[0212]
[0213] in, represents the jth decision tree of the RF in the kth mixed forest group in the feature enhancement layer; L represents the number of decision tree leaf nodes; c l The calculation is performed using a recursive splitting method, and the specific process is shown in formulas (3)-(5);
[0214] Then, the RF model in the kth mixed forest group in the feature enhancement layer can be obtained, which is expressed as,
[0215]
[0216] Then, similarly, taking the construction of the j-th CRF in the k-th mixed forest group as an example, it is as follows:
[0217]
[0218] in, represents the jth decision tree of CRF in the kth mixed forest group in the feature enhancement layer; c l The calculation is performed using a recursive splitting method. The specific process is shown in formulas (6)-(7);
[0219] Then, the CRF model of the kth mixed forest group in the feature enhancement layer can be obtained, which is expressed as:
[0220]
[0221] Through the above process, we get the kth mixed forest group Then, the kth enhanced feature can be expressed as follows:
[0222]
[0223] in, represents the enhanced mapping of the kth mixed forest group to the first sample in the new data, represents the kth mixed forest group for the nth Raw The enhanced map of the th sample, Indicates the kth mixed forest group for the Nth in the new data Raw The enhanced mapping of the th sample;
[0224] Finally, the output H of the feature enhancement layer K It is expressed as follows:
[0225]
[0226] Among them, H1 is the first enhanced feature, H2 is the second enhanced feature, and H K is the Kth enhanced feature;
[0227] When the incremental learning strategy is not considered, the BHFR model is expressed as follows:
[0228]
[0229] Among them, G K Represents the combination of the output of the feature mapping layer and the feature enhancement layer, that is, G K =[Z N |H K ], which contains N Raw samples and (2N+2K)-dimensional features; W K represents the weights between the feature mapping layer, feature enhancement layer and output layer, which is calculated as follows:
[0230] W K =(λΙ+[G K ] T G K ) -1 [G K ] T Y (27)
[0231] Among them, Ι represents the unit matrix, λ represents the regularization term coefficient; accordingly, G K The pseudo-inverse calculation can be expressed as:
[0232]
[0233] The BHFR proposed in this application uses the mixed forest group as the basic unit to realize incremental learning according to the convergence degree of the training error. The step S4 constructs an incremental learning layer, constructs an incremental learning layer through an incremental learning strategy, and uses Moore-Penrose pseudo-inverse to obtain a weight matrix, thereby realizing high-precision modeling of the BHFR soft sensor model, specifically including:
[0234] First, the new data set {X′, y} is sampled based on Bootstrap and RSM to obtain the mixed forest algorithm training subset. The process is as follows:
[0235]
[0236] in, and is the input and output of the Jth training subset of the mixed forest algorithm, X′ and y are the input and output of the new training set, and Represents Bootstrap sampling and RSM sampling of the pth mixed forest group in the incremental learning layer;
[0237] Next, construct the decision tree in the pth mixed forest group and The process is the same as that of the feature mapping layer and feature increment layer, so it will not be repeated here;
[0238] Furthermore, when one mixed forest group is added, the output G of the feature mapping layer, feature increment layer and incremental learning layer is K+1 It is expressed as follows:
[0239]
[0240] Among them, G k =[Z n |H k ] contains N Raw samples and (2N+2K)-dimensional features, G K+1 Contains N Raw samples and (2N+2K+2J)-dimensional features;
[0241] Then, perform G K+1 The recursive update of the Moore-Penrose inverse matrix is as follows:
[0242]
[0243] Among them, the calculation of matrix C and matrix D is as follows:
[0244] C=H K+1 -G K D (32)
[0245]
[0246] Furthermore, G K+1 The recursive formula for the Moore-Penrose inverse matrix is as follows:
[0247]
[0248] Further, the update matrix W of the weights between the feature mapping layer, feature increment layer, incremental learning layer and output layer is calculated K+1 ,as follows:
[0249]
[0250] Among them, W K=(λΙ+[G K ] T G K ) -1 [G K ] T Y;
[0251] Since the above pseudo-inverse update strategy only needs to calculate the pseudo-inverse matrix of the mixed forest group of the incremental learning layer, fast incremental learning can be achieved;
[0252] Furthermore, adaptive incremental learning is achieved according to the convergence degree of the training error;
[0253] Define the convergence threshold of error as θ Con It is used to determine the number p of mixed forest groups in incremental learning; accordingly, the incremental learning training error of the BHFR model is expressed as follows:
[0254]
[0255] in, Indicates the training error value of the incremental learning of the p+1th and pth mixed forest groups, and represents the training error of the BHFR model containing p and p + 1 mixed forest groups;
[0256] Finally, the prediction output of the proposed BHFR soft sensor model is for,
[0257]
[0258] This application uses actual DXN data from a MSWI power plant for industrial verification. The DXN data comes from a MSWI incineration power plant in Beijing, covering a total of 141 sets of DXN emission concentration modeling data from 2009 to 2020. The true DXN value is the converted concentration after 2-hour sampling and testing. After eliminating missing data and abnormal variables, the input variable is 116 dimensions, and the corresponding value is the mean value within the current DXN true value sampling time period.
[0259] This application selects the root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (Coefficient of Determination, R 2 ) There are three evaluation indicators to compare the performance of different methods, calculated as follows:
[0260]
[0261]
[0262]
[0263] Where N is the number of data, y i is the i-th true value, is the i-th predicted value, is the mean.
[0264] In the DXN dataset, the parameters of the BHFR method are set as follows: the minimum number of samples N for the leaf nodes of the decision tree smples 7, the number of RSM feature selection The number of decision trees N tree is 10, the number of mixed forest groups N in the feature mapping layer and feature enhancement layer Forest are all 10, the potential feature contribution rate threshold η is 0.9, and the regularization parameter λ is 2^-10.
[0265] Similar to the benchmark dataset, the number of potential features used for the feature enhancement layer and the incremental learning layer is first determined based on the fully connected mixing matrix and the feature space A. The feature dimension of A in the DXN dataset is 316. When the potential feature contribution rate threshold η is 0.9, the number of potential features selected in the DXN dataset is 35. Then, the mutual information value between the 35 potential features and the DXN true value is calculated. The mutual information threshold ζ is set to 0.75, and the number of potential features selected in the DXN dataset is 6.
[0266] Furthermore, the number of mixed forest group units in the incremental learning layer is preset to 1000. The relationship between the training error of the BHFR model and the number of mixed forest groups is as follows: Figure 3 shown.
[0267] Depend on Figure 3 From the training error curve shown, it can be seen that the training process of BHFR on the DXN dataset can converge to a certain lower limit value.
[0268] Then, RF, DFR, DFR-clfc and BLS-NN are used to compare with the proposed BHFR, and the parameters are set as follows: (1) RF, the minimum number of samples of decision tree leaf nodes N smples is 3, and the number of RSM feature selections is The number of decision trees N tree is 500; (2) DFR, the minimum number of samples N for the leaf nodes of a decision tree smples is 3, and the number of RSM feature selections is The number of decision trees N tree is 500, the number of RF and CRF models in each layer is N RF and N CRF Both are 2, and the total number of layers is set to 50; (3) DFR-clfc, the minimum number of samples of decision tree leaf nodes N smplesis 3, and the number of RSM feature selections is The number of decision trees N tree is 500, the number of RF and CRF models in each layer is N RF and N CRF are all 2, and the total number of layers is set to 50; (4) BLS-NN, the number of feature nodes is N m 5, the number of enhanced nodes N e is 41, the number of neurons N n The above method was repeated 20 times under the same conditions, and the statistical results and prediction curves are shown in Table 1 and Figure 4a-4c shown.
[0269] Table 1 Experimental results of DXN dataset
[0270]
[0271] From Table 1 and Figure 4a-4c It can be seen that: 1) RMSE, MAE and R of RF in training, validation and testing 2 The statistical results of the mean indicators are better than DFR, but weaker than DFR in terms of stability indicators; 2) DFR and DFR-clfc are close to RF in modeling accuracy, and their modeling stability is better than RF. Among them, DFR-clfc has slightly higher accuracy than DFR in training, validation and test sets, but DFR has better stability; 3) BLS-NN has obvious overfitting of the training data, and its generalization performance and stability in the validation and test sets are the worst, indicating that BLS-NN is difficult to apply to the small sample high-dimensional data of the real industrial process in this application; 4) BHFR has poor RMSE, MAE and R 2 The mean statistical results of the indicators are all the best, and the stability is only weaker than DFR, indicating that BHFR has good generalization performance and stability.
[0272] In summary, the DXN soft sensor modeling experiment shows that the BHFR proposed in this application has better training and learning capabilities than the classic RF, DFR and its improved version DFR-clfc. At the same time, the modeling accuracy and data fitting degree on the test set are also stronger than RF, DFR, DFR-clfc and BLS-NN, reflecting its obvious advantages in constructing the DXN soft sensor model.
[0273] The soft sensor method for dioxin emission in the MSWI process based on wide mixed forest regression provided by the present invention establishes a soft sensor model based on BHFR, which combines algorithms such as wide learning modeling, ensemble learning and potential feature extraction. 1) Based on the wide learning system framework, a soft sensor model including a feature mapping layer, a potential feature extraction layer, a feature enhancement layer and an incremental learning layer is constructed by using a non-differential learner; 2) The internal information of the BHFR model is processed by using full information connection, potential feature extraction and mutual information measurement, which effectively ensures the maximum transmission and minimum redundancy of the internal feature information of the BHFR model; 3) The mixed forest group is used as a mapping unit to realize the incremental learning of the modeling process, the output layer weight matrix is quickly calculated by the pseudo-inverse strategy, and the incremental learning is adaptively adjusted by using the convergence degree of the training error, so as to realize high-precision soft sensor modeling. The effectiveness and rationality of the proposed method are verified on high-dimensional benchmark datasets and industrial process DXN datasets.
[0274] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of the present invention. At the same time, for those skilled in the art, according to the idea of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.
Claims
1. A soft-sensing method for dioxin emissions from MSWI processes based on wide mixed forest regression is proposed. Based on the BLS framework, a BHFR soft-sensing model for small sample high-dimensional data is constructed by replacing neurons with non-differential basis learners. The method is characterized by: The BHFR soft sensor model includes the construction of a feature mapping layer, a potential feature extraction layer, a feature enhancement layer and an incremental learning layer, and specifically includes the following steps: S1, construct a feature mapping layer, construct a mixed forest group consisting of random forest RF and completely random forest CRF to map high-dimensional features; S2, constructs a potential feature extraction layer, extracts potential features from the feature space of the fully connected mixed matrix according to the contribution rate, and ensures the maximum transmission of potential valuable information and minimizes redundancy based on the information measurement criterion, thereby reducing model complexity and computational consumption; S3, constructing a feature enhancement layer, and training the feature enhancement layer based on the extracted latent features to further enhance the feature representation capability; S4, constructing the incremental learning layer, constructing the incremental learning layer through the incremental learning strategy, using Moore-Penrose pseudo-inverse to obtain the weight matrix, and then realizing the high-precision modeling of the BHFR soft sensor model; S5, using a high-dimensional benchmark dataset and an industrial process DXN dataset to verify the soft sensor model; S6, using the soft measurement model established in steps S1-S5 to perform soft measurement of dioxin emissions from the MSWI process; Step S1, constructing a feature mapping layer, constructing a mixed forest group consisting of a random forest RF and a completely random forest CRF to map high-dimensional features, specifically including: Assume the original data is {X,y}, where is the original input data, N Raw is the number of raw data, M is the dimension of the raw input data, which comes from the six different stages of the MSWI process, collected and stored in the DCS system in seconds, is the true output value of DXN emission concentration, which comes from the emission DXN detection sample obtained by offline detection method; the modeling process of the feature mapping layer is described by taking the nth mixed forest group of the feature mapping layer as an example: Bootstrap and random subspace RSM sampling are performed on {X, y} to obtain J training subsets of the mixed forest group model as follows: in, and is the input and output of the Jth training subset, and represents the Bootstrap and RSM sampling of the nth mixed forest group in the feature map layer, P Bootstrap Indicates Bootstrap sampling probability; based on Train a mixed forest algorithm containing J decision trees, where the jth decision tree of the nth mixed forest group in the feature map layer is expressed as follows: Where L represents the number of decision tree leaf nodes, I(·) represents the indicator function, and c l The calculation is done by recursive splitting method; Split loss function Ω for decision tree in RF i (·) is expressed as: Among them, Ω i (s,v) represents the loss function value of the sth feature value v as the segmentation criterion, y L represents the true value vector of DXN emission concentration of the left leaf node, E[y L ] means y L The mathematical expectation of R represents the true value vector of DXN emission concentration of the right leaf node, E[y R ] means y R The mathematical expectation of represents the true value of the DXN emission concentration of the i-th left leaf node, represents the true value of the DXN emission concentration of the i-th right leaf node, c L represents the emission concentration prediction output of the left leaf node DXN, c R It represents the emission concentration prediction output of the right leaf node DXN; By minimizing Ω i (s,v), the training set Divide into two tree nodes, as follows: in, and Represents the sample set contained in the left and right tree nodes after segmentation, N L and N R Respectively and The number of samples in ; DXN emission concentration prediction output value of the current left and right tree nodes and is the expectation of the true value of the sample, as follows: Among them, y L and R express and The true value vector of DXN emission concentration in E[y L ] and E[y R ] means y L and R The mathematical expectation of Unlike RF, the decision tree splitting in CRF adopts a completely random selection method, which is expressed as, in, It means that the value v of the sth feature is selected completely randomly as the split point; DXN emission concentration prediction output values of randomly split left and right tree nodes and is the expectation of the true value of the sample, as follows: Through the above process, the nth mixed forest group f n FML (·) can be expressed as, in, represents the nth random forest, represents the nth completely random forest; Furthermore, the nth mapping feature Z n It can be expressed as in, represents the mapping features of the nth group of mixed forests to the first sample of the original input data from six different stages of the MSWI process, The nth group of mixed forests is derived from the original input data of the six different stages of the MSWI process. Raw The mapping features of the th sample, It represents the nth group of mixed forests for the original input data from six different stages of the MSWI process. Raw The mapping features of the th sample; Finally, the output of the feature mapping layer is expressed as: Among them, Z1 is the first mapping feature, Z2 is the second mapping feature, and Z N is the Nth mapping feature, the mapping feature matrix Z N Contains N Raw samples and 2N-dimensional features.
