A method for soft-sensing dioxin emission concentration

Abstract

The application provides a dioxin emission concentration soft measurement method, comprising the following steps: constructing a concept drift detection model, constructing a soft measurement model, updating the concept drift detection model, updating the soft measurement model, obtaining actual online data of a factory, performing drift detection on the actual online data based on the updated concept drift detection model, if no drift occurs, performing measurement based on the updated soft measurement model, if drift occurs, matching a historical sample most similar to a current working condition, obtaining multiple measurement values, obtaining a soft measurement value by calculating an arithmetic mean, issuing a concept drift warning, and updating the two models according to obtained DXN true value samples after completing online sampling and offline testing under a similar working condition. The dioxin emission concentration soft measurement method provided by the application can realize soft measurement of dioxin emission concentration and is convenient to use.

Description

Technical Field

[0001] This invention relates to the field of urban solid waste treatment technology, and in particular to a soft measurement method for dioxin emission concentration. Background Technology

[0002] While municipal solid waste incineration (MSWI) has solved the "garbage siege" crisis and generated energy, it has also led to environmental pollution. It has become one of the main choices for most MSWI treatments. However, there is currently no detection technology capable of real-time monitoring of dioxin emissions to support the optimized operation of MSWI processes, and real-time estimation of dioxin concentrations emitted by MSWI plants remains challenging. Therefore, it is essential to design a soft measurement method for dioxin emission concentrations. Summary of the Invention

[0003] The purpose of this invention is to provide a soft measurement method for dioxin emission concentration, which can realize soft measurement of dioxin emission concentration and is easy to use.

[0004] To achieve the above objectives, the present invention provides the following solution:

[0005] A soft measurement method for dioxin emission concentration includes the following steps:

[0006] Step 1: Construct a concept drift detection model;

[0007] Step 2: Construct a soft measurement model;

[0008] Step 3: Update the concept drift detection model;

[0009] Step 4: Update the soft measurement model;

[0010] Step 5: Obtain actual online data from the factory and perform drift detection based on the updated concept drift detection model. If no drift occurs, perform measurement based on the updated soft measurement model. If drift occurs, match the historical sample most similar to the current operating condition, obtain multiple measurement values, calculate the soft measurement value by arithmetic mean, issue a concept drift warning, and update both models based on the obtained DXN ground truth samples after completing online sampling and offline testing under similar operating conditions.

[0011] Optionally, in step 1, a concept drift detection model is constructed, specifically as follows:

[0012] Obtain historical training data, where historical training data D Htd From high-dimensional process variables and the corresponding true values ​​of DXN emission concentrations composition;

[0013] Obtain the historical training data matrix X with zero mean and unit variance. Htd Let the noise data matrix be E Htd The actual process data matrix is The original process data matrix X Htd It can be decomposed into:

[0014]

[0015] In the formula, E is a low-rank matrix. Htd It is a sparse matrix;

[0016] Transforming it into a convex optimization problem, we get:

[0017]

[0018]

[0019] In the formula, For matrix The norm of ||E Htd ||1=Σ|E Htd |For matrix E Htd The sum of vector magnitudes;

[0020] The problem is solved iteratively using the alternating direction method. Assume there exists a multiplier matrix Q for the linear constraints. Htd ∈ N×M With the positive penalty parameter β, we obtain the augmented Lagrangian function, which is:

[0021]

[0022] Minimize using an iterative strategy and E Htd ,get:

[0023]

[0024]

[0025]

[0026] for Using the subgradient operator to evaluate the convex function, we obtain:

[0027]

[0028] make get:

[0029]

[0030] In the formula, Let be the Euclidean projection on the set, where:

[0031]

[0032] for Representing this as a minimization problem, we get:

[0033]

[0034] In the formula, Let a matrix be decomposed into singular value decomposition as follows: The solution yields:

[0035]

[0036] use The low-rank matrix is ​​obtained by iterative calculation. Using low-rank matrices Establish T 2 and the control limits of SPE;

[0037] low-rank matrix Correlation matrix R Htd Written as:

[0038]

[0039] Calculate the correlation matrix R based on the SVD method. Htd The eigenvalues ​​of are:

[0040] R Htd =U Htd Σ Htd [V Htd ] T (13)

[0041] In the formula, U Htd and V Htd Represents an orthogonal matrix (V) Htd =[U Htd ] T ), Σ Htd =diag(σ1,σ2,...,σ M ) is a diagonal matrix, σ m It is R Htd The mth eigenvalue;

[0042] Based on contribution rate η Htd The principal components are selected as follows:

[0043]

[0044] In the formula, MPC To select the number of principal components, δ PC The threshold was set based on experience;

[0045] Based on the selected M PC Each principal component will have a principal component loading matrix U. Htd Represented as low-rank matrix Written as:

[0046]

[0047] In the formula, and Represent the principal space and the residual space, respectively;

[0048] Using T 2 And SPE statistical indicators monitor concept drift, where T 2 SPE represents the change in the principal space, while SPE simultaneously represents the change in both the principal space and the residual space in a low-rank matrix. Below, statistics T 2 The control limits of the indicator, i.e. and the control limits of the SPE index, i.e. Represented as:

[0049]

[0050]

[0051] In the formula, f α (M PCA ,NM PCA ) for having M PCA and (NM) PCA F-distribution of degrees of freedom, c α The normal deviation of the percentile with respect to (1-α) is represented by the parameters Θ1, Θ2, and h0, which are solved as follows:

[0052]

[0053]

[0054] Optionally, in step 2, a soft measurement model is constructed, specifically as follows:

[0055] Obtain a historical training dataset consisting of N samples and M-dimensional features. A mapping layer is composed of multiple RDT units, each consisting of non-leaf nodes and leaf nodes. Let the RDT model have T nodes, where (T+1) / 2 represents the number of leaf nodes and (T-1) / 2 represents the number of non-leaf nodes. During the tree growth process of the RDT, the (T-1) / 2 non-leaf nodes are determined through random splitting, expressed as:

[0056]

[0057] In the formula, This represents the t-th (t≤(T-1) / 2) non-leaf node;

[0058] Based on the information of the t-th non-leaf node, the tree structure of the RDT growth strategy is partitioned as follows:

[0059]

[0060] In the formula, and This represents the data of two child nodes. and express and The number of samples in the sample;

[0061] The recursive procedure yields (T-1) / 2 non-leaf nodes. The leaf nodes are then calculated using an arithmetic mean strategy to determine the final predicted value, as follows:

[0062]

[0063] In the formula, Let be the number of samples in the t-th leaf node. This represents the true value of the t-th leaf node;

[0064] The RDT model is represented as:

[0065]

[0066] In the formula, For the nth layer in the mapping layer ML One RDT function, Indicates the nth ML The t-th path of an RDT;

[0067] Let the mapping layer have N ML If there are 1 RDT function, then the output of the mapping layer is represented as:

[0068]

[0069]

[0070] The enhancement layer has a similar structure and units to the mapping layer. Let the enhancement layer consist of N EL It consists of 1 RDT function, denoted as The output of the enhancement layer is then expressed as:

[0071]

[0072]

[0073] By integrating the mapping layer and the enhancement layer, the output is represented as The output of the output layer is:

[0074]

[0075] Among them, W k Estimate the weights of each RDT function in the mapping and enhancement layers:

[0076] W k =([Z k ] T Z k +λ2I)- 1 [Z k ] T y Htd =[Z k ] * y Htd (29)

[0077] In the formula, [Z k ] * Z represents k The pseudo-inverse matrix, where λ² > 0 is the ratio of regularization terms. Represents the identity matrix;

[0078] When both the mapping layer and the enhancement layer perform poorly, the width of the offline incremental layer is adjusted. Each group in the offline incremental layer consists of multiple neurons. When adding the first group of neurons, the corresponding mapping process is as follows:

[0079]

[0080]

[0081] In the formula, This represents the output of the first group of neurons. Indicates the number of neurons. Indicates the first indivual Neuron output, For activation function, and These represent the weights and biases, respectively.

[0082] The outputs of the mapping layer, enhancement layer, and offline incremental layer are represented as follows:

[0083] The Moore-Penrose inverse matrix is ​​solved using a column recursive algorithm, and the new inverse matrix is ​​written as:

[0084]

[0085] In the formula,

[0086]

[0087] Update the weight matrices of the mapping layer, enhancement layer, and offline incremental layer:

[0088]

[0089] After adding the first group of neurons, the prediction results of the offline model are as follows:

[0090]

[0091] Optionally, in step 3, the concept drift detection model is updated, specifically as follows:

[0092] Based on the concept drift detection model, a fixed-size moving sampling window is used during the detection process. When new data is acquired by the sampling window, the zero-mean and unit variance data matrix and the correlation matrix are calculated, based on the high-dimensional historical training data matrix X. Htd and correlation matrix R Htd The mean of each column is expressed as Let X be the online test dataset from the historical data. Otd Combined with online data within window W1 New data matrix and correlation matrix R Htd The recursive calculation is as follows:

[0093]

[0094] In the formula,

[0095]

[0096] Finally, a low-rank matrix is ​​obtained.

[0097] Principal component loading matrix calculated based on principal components. T of samples within the window 2 And SPE are represented as:

[0098]

[0099]

[0100] when Less than a certain control limit And SPE k+1 Less than a certain control limit At that time, it was determined that no drift had occurred;

[0101] when Exceeding control limits or SPE k+1 Exceeding control limits At that time, it is determined that a drift has occurred.

[0102] Optionally, in step 4, the soft measurement model is updated, specifically as follows:

[0103] when and SPE k+1 The values ​​are all less than their control limits. and When this occurs, it indicates that no drift has occurred, meaning that the current operating conditions are similar to historical operating conditions. That is, the current DXN emission concentration is within the distribution range of historical data. The soft sensor model is then rewritten to obtain:

[0104]

[0105] In the formula, F ML (·), F EL (·) and F OffIL (·) are the activation functions for the mapping layer, enhancement layer, and offline incremental layer, respectively;

[0106] when value or SPE k+1 The value exceeds its control limit and When this occurs, the soft sensor model is considered to have drifted, and a drift sample containing process data and DXN ground truth is used to update the model.

[0107] The outputs of the mapping layer, enhancement layer, and offline incremental layer are represented as follows: Add an online incremental layer, and represent the outputs of the mapping layer, enhancement layer, offline incremental layer, and online incremental layer as follows: in, The new online inverse matrix is ​​written as:

[0108]

[0109] in,

[0110]

[0111] The new weight matrix is ​​represented as follows:

[0112]

[0113] After adding the first online neuron group, the updated soft measurement model is represented as follows:

[0114]

[0115] Optionally, in step 5, actual online data from the factory is acquired, and drift detection is performed based on the updated concept drift detection model. If no drift occurs, measurements are performed based on the updated soft measurement model. If drift occurs, the most similar historical sample to the current operating condition is matched, multiple measurement values ​​are obtained, and the soft measurement value is obtained by calculating the arithmetic mean. A concept drift warning is issued. After completing online sampling and offline testing under similar operating conditions, both models are updated based on the obtained DXN ground truth samples. Specifically:

[0116] Acquire actual online data from the factory and perform drift detection based on the updated conceptual drift detection model. If no drift is detected, then measurement is performed based on the updated soft measurement model; if drift occurs, i.e. Value exceeds control limit or SPE k+1 Value exceeds control limit At this point, the similarity between the drift sample and the historical sample is calculated using Euclidean distance:

[0117]

[0118] By manually setting the threshold δ Sim Select the most similar historical samples for:

[0119]

[0120] The most similar historical samples Input the data into the most recently updated soft measurement model to obtain the measurement vector.

[0121] For measurement vector The arithmetic mean is calculated to obtain the final measurement value under the current drift condition.

[0122] Issue an alarm to the expert operating the MSWI process indicating a condition drift, prompting them to record the condition.

[0123] Under similar working conditions, online sampling and offline testing were performed to obtain the true value of DXN;

[0124] This ground truth sample representing the drift condition is then updated to the online test dataset in the historical data, triggering steps 3 and 4 to update the concept drift detection model and the soft measurement model.

[0125] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The present invention provides a soft measurement method for dioxin emission concentration, which includes constructing a concept drift detection model, constructing a soft measurement model, updating the concept drift detection model, updating the soft measurement model, acquiring actual online data from the plant, performing drift detection based on the updated concept drift detection model, if no drift occurs, then performing measurement based on the updated soft measurement model; if drift occurs, then matching the historical sample most similar to the current operating condition, acquiring multiple measurement values, obtaining the soft measurement value by calculating the arithmetic mean, issuing a concept drift warning, and completing online sampling under similar operating conditions. After offline testing, the two models are updated based on the obtained ground truth samples of DXN. This method proposes a hybrid neuron and tree-based width learning system, in which the offline and online incremental learning layers employ neuron groups, improving the model update speed and modeling accuracy. Through historical data experiments and actual MSWI plant applications, it is demonstrated that the proposed method significantly improves the performance and effectiveness of the DXN soft measurement system. By adopting these algorithms and techniques, researchers and practitioners can improve the effectiveness of online soft measurement systems for DXN emission monitoring, which helps to improve environmental management and regulatory compliance, ultimately minimizing the impact of DXN emissions on public health and the environment. Attached Figure Description

[0126] 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.

[0127] Figure 1 This is a schematic diagram of the modeling strategy for the soft measurement method of dioxin emission concentration in an embodiment of the present invention;

[0128] Figure 2 A distribution map of DXN data;

[0129] Figure 3a This is a three-dimensional representation of the original matrix.

[0130] Figure 3b This is a three-dimensional result diagram of the low-rank matrix;

[0131] Figure 3c This is a three-dimensional result diagram of the noise matrix;

[0132] Figure 4This is a schematic diagram illustrating the contribution rate of features in a low-rank matrix.

[0133] Figure 5a For T based on online test dataset 2 Result diagram;

[0134] Figure 5b This is a schematic diagram of SPE results based on an online test dataset;

[0135] Figure 6 This is a schematic diagram illustrating the application of a soft measurement system in an actual MSWI factory. Detailed Implementation

[0136] The purpose of this invention is to provide a soft measurement method for dioxin emission concentration, which can realize soft measurement of dioxin emission concentration and is easy to use.

[0137] 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.

[0138] like Figure 1 As shown, the soft measurement method for dioxin emission concentration provided in this embodiment of the invention includes the following steps:

[0139] Step 1: Construct a concept drift detection model;

[0140] Based on historical training datasets, an RPCA model is constructed to detect concept drift. This model calculates the historical Hotelling statistic (Hotelling T). 2 ,T 2 The squared prediction error (SPE) control limit is used to monitor and detect drift deviations in operating conditions.

[0141] Step 2: Construct a soft measurement model;

[0142] An HNTBLS model is constructed based on historical training data. The HNTBLS model with a mapping layer, an enhancement layer, and an offline incremental layer is trained. This model can accurately estimate DXN emission concentrations based on given input data with the same distribution.

[0143] Step 3: Update the concept drift detection model;

[0144] Using the online test dataset from historical data, calculate the T-values ​​of the samples within a fixed moving window. 2 And SPE values, and then detect concept drift by comparing these values ​​with predefined control limits, if T 2If the SPE value exceeds the control limit, it indicates that the drift is in normal operating condition. At this time, we record the sample representing the drift of this concept and update the RPCA model and control limit. Otherwise, we return to the moving window to get the next sample. It is worth noting that the online test dataset will be updated over time.

[0145] Step 4: Update the soft measurement model;

[0146] When T 2 When the SPE value is within the control range, HNTBLS is reused to perform online measurement of DXN emission concentration using real-time process data. If T 2 If the SPE value exceeds the control limit, it indicates that the operating condition has drifted. The HNTBLS model is then updated online based on the conceptual drift sample to complete the operating conditions that the soft measurement model can adapt to.

[0147] Step 5: Obtain actual online data from the factory and perform drift detection based on the updated concept drift detection model. If no drift occurs, perform measurements based on the updated soft measurement model. If drift occurs, match the historical sample most similar to the current operating condition, obtain multiple measurement values, and calculate the arithmetic mean to obtain the soft measurement value. Even under drift conditions, this method can ensure relatively effective measurement values. Issue a concept drift warning to the operation experts, and after completing online sampling and offline testing under similar operating conditions, update both models based on the obtained DXN ground truth samples.

[0148] In step 1, a concept drift detection model is constructed, specifically as follows:

[0149] The MSWI process consists of multiple stages and operates under high temperature and high pressure. Therefore, the measured values ​​of process variables may deviate from or contain errors compared with the true values. This will result in sparse noise data in the collected process data. The presence of this noise data hinders the correctness and effectiveness of statistical analysis. Therefore, in order to establish an effective concept drift detection model, the influence of noise data must first be eliminated.

[0150] Obtain historical training data, where historical training data D Htd From high-dimensional process variables and the corresponding true values ​​of DXN emission concentrations composition;

[0151] Obtain the historical training data matrix X with zero mean and unit variance. Htd Let the noise data matrix be E Htd The actual process data matrix is The original process data matrix X Htd It can be decomposed into:

[0152]

[0153] In the formula, E is a low-rank matrix. Htd It is a sparse matrix;

[0154] To solve the matrix separation problem, equation (1) can be transformed into a convex optimization problem using the following formula:

[0155]

[0156]

[0157] In the formula, For matrix The norm of ||E Htd ||1=∑|E Htd |For matrix E Htd The sum of vector magnitudes;

[0158] Equation (2) is solved iteratively using the alternating direction method. Let there exist a multiplier matrix Q for the linear constraint. Htd ∈ N×M With the positive penalty parameter β, we obtain the augmented Lagrangian function, which is:

[0159]

[0160] According to formula (2), an iterative strategy is adopted to minimize and E Htd ,get:

[0161]

[0162]

[0163]

[0164] for Using the subgradient operator to evaluate the convex function, we obtain:

[0165]

[0166] make get:

[0167]

[0168] In the formula, Let be the Euclidean projection on the set, where:

[0169]

[0170] for Representing this as a minimization problem, we get:

[0171]

[0172] In the formula, Let a matrix be decomposed using Singular Value Decomposition (SVD) into... The solution yields:

[0173]

[0174] use The low-rank matrix is ​​obtained by iterative calculation. Using low-rank matrices Establish T 2 and the control limits of SPE;

[0175] low-rank matrix Correlation matrix R Htd Written as:

[0176]

[0177] Calculate the correlation matrix R based on the SVD method. Htd The eigenvalues ​​of are:

[0178] R Htd =U Htd Σ Htd [V Htd ] T (13)

[0179] In the formula, U Htd and V Htd Represents an orthogonal matrix (V) Htd =[U Htd ] T ), Σ Htd =diag(σ1,σ2,...,σ M ) is a diagonal matrix, σ m It is R Htd The mth eigenvalue;

[0180] To reduce redundant information, based on the contribution rate η Htd The principal components are selected as follows:

[0181]

[0182] In the formula, M PC To select the number of principal components, δ PC The threshold was set based on experience;

[0183] Based on the selected M PC Each principal component will have a principal component loading matrix U. Htd Represented as low-rank matrix Written as:

[0184]

[0185] In the formula, and Represent the principal space and the residual space, respectively;

[0186] Using T 2 And SPE statistical indicators monitor concept drift, where T 2 SPE represents the change in the principal space, while SPE simultaneously represents the change in both the principal space and the residual space in a low-rank matrix. Below, statistics T 2 The control limits of the indicator, i.e. and the control limits of the SPE index, i.e. Represented as:

[0187]

[0188]

[0189] In the formula, f α (M PCA ,NM PCA ) for having M PCA and (NM) PCA F-distribution of degrees of freedom, c α The normal deviation of the percentile with respect to (1-α) is represented by the parameters Θ1, Θ2, and h0, which are solved as follows:

[0190]

[0191]

[0192] In step 2, the soft measurement model is constructed, specifically as follows:

[0193] A novel HNTBLS method is employed for offline modeling. Key modules in HNTBLS include a mapping layer, an enhancement layer, and an offline incremental layer. The basic unit used in the mapping and enhancement layers is a random DT (RDT). To achieve fast incremental learning performance, the basic unit of the offline incremental layer is a group of neurons, as detailed below:

[0194] Obtain a historical training dataset consisting of N samples and M-dimensional features. A mapping layer is composed of multiple RDT units, each consisting of non-leaf nodes and leaf nodes. Let the RDT model have T nodes, where (T+1) / 2 represents the number of leaf nodes and (T-1) / 2 represents the number of non-leaf nodes. During the tree growth process of the RDT, the (T-1) / 2 non-leaf nodes are determined through random splitting, expressed as:

[0195]

[0196] In the formula, This represents the t-th (t≤(T-1) / 2) non-leaf node;

[0197] Based on the information of the t-th non-leaf node, the tree structure of the RDT growth strategy is partitioned as follows:

[0198]

[0199] In the formula, and This represents the data of two child nodes. and express and The number of samples in the sample;

[0200] The recursive procedure yields (T-1) / 2 non-leaf nodes. The leaf nodes are then calculated using an arithmetic mean strategy to determine the final predicted value, as follows:

[0201]

[0202] In the formula, Let be the number of samples in the t-th leaf node. This represents the true value of the t-th leaf node;

[0203] The RDT model is represented as:

[0204]

[0205] In the formula, For the nth layer in the mapping layer ML One RDT function, Indicates the nth ML The t-th path of an RDT;

[0206] Let the mapping layer have N ML If there are 1 RDT function, then the output of the mapping layer is represented as:

[0207]

[0208]

[0209] The enhancement layer has a similar structure and units to the mapping layer. Let the enhancement layer consist of N EL It consists of 1 RDT function, denoted as The output of the enhancement layer is then expressed as:

[0210]

[0211]

[0212] By integrating the mapping layer and the enhancement layer, the output is represented as The output of the output layer is:

[0213]

[0214] Among them, W k Estimate the weights of each RDT function in the mapping and enhancement layers:

[0215] W k =([Z k ] T Z k +λ2I)- 1 [Z k ] T y Htd =[Z k ] * y Htd (29)

[0216] In the formula, [Z k ] * Z represents k The pseudo-inverse matrix, where λ² > 0 is the ratio of regularization terms. Represents the identity matrix;

[0217] When both the mapping layer and the enhancement layer perform poorly, the width of the offline incremental layer is adjusted. Each group in the offline incremental layer consists of multiple neurons. When adding the first group of neurons, the corresponding mapping process is as follows:

[0218]

[0219]

[0220] In the formula, This represents the output of the first group of neurons. Indicates the number of neurons. Indicates the first indivual Neuron output, For activation function, and These represent the weights and biases, respectively.

[0221] The outputs of the mapping layer, enhancement layer, and offline incremental layer are represented as follows:

[0222] The Moore-Penrose inverse matrix is ​​solved using a column recursive algorithm, and the new inverse matrix is ​​written as:

[0223]

[0224] In the formula,

[0225]

[0226] Update the weight matrices of the mapping layer, enhancement layer, and offline incremental layer:

[0227]

[0228] After adding the first group of neurons, the prediction results of the offline model are as follows:

[0229]

[0230] In step 3, the concept drift detection model is updated, specifically as follows:

[0231] Based on the concept drift detection model, a fixed-size moving sampling window is used during the detection process. When new data is acquired by the sampling window, the zero-mean and unit variance data matrix and the correlation matrix are calculated, based on the high-dimensional historical training data matrix X. Htd and correlation matrix R Htd The mean of each column is expressed as Let X be the online test dataset from the historical data. Otd Combined with online data within window W1 New data matrix and correlation matrix R Htd The recursive calculation is as follows:

[0232]

[0233] In the formula,

[0234]

[0235] Finally, a low-rank matrix is ​​obtained.

[0236] Principal component loading matrix calculated based on principal components. T of samples within the window 2 And SPE are represented as:

[0237]

[0238]

[0239] The concept drift detection strategy is defined as follows:

[0240] when Less than a certain control limit And SPE k+1 Less than a certain control limit At that time, it was determined that no drift had occurred;

[0241] when Exceeding control limits or SPE k+1 Exceeding control limits At that time, it is determined that a drift has occurred.

[0242] In step 4, the soft measurement model is updated, specifically as follows:

[0243] when and SPE k+1 The values ​​are all less than their control limits. and When this occurs, it indicates that no drift has occurred, meaning that the current operating conditions are similar to historical operating conditions. That is, the current DXN emission concentration is within the distribution range of historical data. The soft sensor model is then rewritten to obtain:

[0244]

[0245] In the formula, F ML (·), F EL (·) and F OffIL (·) are the activation functions for the mapping layer, enhancement layer, and offline incremental layer, respectively;

[0246] when value or SPE k+1 The value exceeds its control limit and When this occurs, the soft sensor model is considered to have drifted, and a drift sample containing process data and DXN ground truth is used to update the model.

[0247] The outputs of the mapping layer, enhancement layer, and offline incremental layer are represented as follows: Add an online incremental layer, and represent the outputs of the mapping layer, enhancement layer, offline incremental layer, and online incremental layer as follows: in, The new online inverse matrix is ​​written as:

[0248]

[0249] in,

[0250]

[0251] The new weight matrix is ​​represented as follows:

[0252]

[0253] After adding the first online neuron group, the updated soft measurement model is represented as follows:

[0254]

[0255] In step 5, actual online data from the factory is acquired, and drift detection is performed based on the updated concept drift detection model. If no drift is detected, measurements are taken based on the updated soft measurement model. If drift occurs, the most similar historical sample to the current operating condition is matched, multiple measurement values ​​are obtained, and the soft measurement value is obtained by calculating the arithmetic mean. A concept drift warning is issued. After completing online sampling and offline testing under similar operating conditions, both models are updated based on the obtained DXN ground truth samples. Specifically:

[0256] Currently, the determination of DXN emission concentration mainly consists of two stages: manual on-site sampling, laboratory cleaning, extraction, and high-resolution gas chromatography-high-resolution mass spectrometry (HRGC / HRMS) analysis. In the first stage, at least two hours are required to collect real-time flue gas samples; in the second stage, analysis usually takes more than a day. Therefore, although HRGC / HRMS analysis is an automated procedure for calculating DXN concentration, obtaining the true DXN value for online model updates in a short period of time during actual MSWI is a challenging problem. At the same time, concept drift is an unavoidable phenomenon in the MSWI process.

[0257] Acquire actual online data from the factory and perform drift detection based on the updated conceptual drift detection model. If no drift is detected, then measurement is performed based on the updated soft measurement model; if drift occurs, i.e. Value exceeds control limit or SPE k+1 Value exceeds control limit At this point, the similarity between the drift sample and the historical sample is calculated using Euclidean distance:

[0258]

[0259] By manually setting the threshold δ Sim Select the most similar historical samples for:

[0260]

[0261] The most similar historical samples Input the data into the most recently updated soft measurement model to obtain the measurement vector.

[0262] For measurement vector The arithmetic mean is calculated to obtain the final measurement value under the current drift condition.

[0263] Issue an alarm to the expert operating the MSWI process indicating a condition drift, prompting them to record the condition.

[0264] Under similar working conditions, online sampling and offline testing were performed to obtain the true value of DXN;

[0265] This ground truth sample representing the drift condition is then updated to the online test dataset in the historical data, triggering steps 3 and 4 to update the concept drift detection model and the soft measurement model.

[0266] This invention provides a real-time example to verify the method described in this invention. The DXN data used for modeling comes from an actual MSWI factory, with a data time span from 2009 to 2020. Detailed information is shown in Table 1.

[0267] Table 1 Detailed Information on DXN Data

[0268]

[0269] The historical training dataset consists of ground truth samples from 2009 to 2016, while the online test dataset consists of samples from 2016 to 2020. The true DXN concentration is the average emission concentration over two hours during the current MSWI operation. Furthermore, process variables such as temperature, pressure, and flow rate are generated by the DCS system at a frequency of seconds. To obtain samples corresponding to the true DXN values, an averaging method is used to calculate the process data over the sampling time. The distribution of the true DXN values ​​is as follows: Figure 2 As shown;

[0270] Depend on Figure 2 It can be seen that since the MSWI plant went into operation, the DXN concentration has remained at 0.1 TEQ ng / m³. 3 The following data complies with the pollutant emission control standard (GB18485-2014). It is worth noting that the distribution of historical data is significantly different from that of test data. This difference poses a major challenge to the classic static modeling method, making it difficult to accurately measure DXN emission concentration.

[0271] In the comparison of static modeling methods based on historical training datasets, Random Forest (RF), BPNN, SVR, DFR, BLS, Fuzzy Tree Width Learning System (FTBLS) and the HNTBLS method proposed in this invention were adopted. The hyperparameters of each method are shown below.

[0272] RF: The number of samples in the leaf nodes is set to 10, and the number of features in the random subspace is determined to be the square root of the original feature dimension;

[0273] BPNN: The number of hidden layer neurons is set to (2×M)+1, the learning rate is set to 0.001, the number of iterations is set to 1500, and the convergence error is set to 0.001.

[0274] SVR: The kernel function is a radial basis function, the number of iterations is set to 1500, and the regularization coefficient is set to 1;

[0275] DFR: The number of leaf node samples is set to 10, the number of RF models and full RF models is set to 10, the number of layers is set to 50, the number of random subspace features is the square root of the feature dimension, and the number of DTs is set to 500.

[0276] BLS: The number of feature nodes, mapping layers, and augmentation nodes are set to 9, 5, and 41, respectively;

[0277] FTBLS: The minimum number of samples is set to 10, the number of fuzzy rules is set to (M×10), the learning rate is set to 0.01, the number of iterations is set to 100, the regularization coefficient is set to 0.5, and the number of FDTs in the feature mapping layer, enhancement layer and incremental layer are set to 50, 50 and 5 respectively.

[0278] HNTBLS: The minimum number of samples is set to 3, the number of RDT mapping layers and enhancement layers is set to 20, the regularization coefficient is set to 0.5, and the number of neurons per group in the incremental layer is set to 5.

[0279] The detailed results are shown in Table 2;

[0280] Table 2 Comparison Results of Static Modeling Experiments Based on Historical Training Datasets

[0281]

[0282] As shown in Table 2, all methods demonstrated effectiveness in modeling historical training data. Among them, HNTBLS (RMSE 0, variance 1) and BPNN (RMSE 0.0009, variance 0.999) showed the best learning performance compared to other methods. Furthermore, only DT-based methods (such as RF and DFR) exhibited weaker performance compared to neural network-based methods (BPNN and BLS), indicating that DT-based methods have lower fitting ability relative to the training data than neuron-based methods. For the online test dataset, all methods showed negative variance, indicating that they had difficulty fitting the original data. Relatively speaking, RF and DFR performed relatively well, suggesting that tree-based methods (RF and DFR) have higher stability than neuron-based methods (BP and BLS). In addition, HNTBLS showed better performance on the online test data compared to the FTBLS method.

[0283] In a dynamic modeling experiment based on an online test dataset, the FTBLS and HNTBLS methods were compared. The hyperparameters of these two methods remained the same as before. The results were divided into online concept drift detection and online model update, using the original process data matrix X. Htd Decompose into low-rank matrices and noise matrix E Htd Its distribution in three-dimensional space is as follows Figure 3a , Figure 3b and Figure 3c As shown;

[0284] As shown in Figure 3, the noise information within the original matrix is ​​effectively decomposed. The ranks of the original matrix, the low-rank matrix, and the noise matrix are 70, 37, and 71, respectively. For the low-rank matrix... Principal component analysis was performed, and the contribution rates of its 116 features were as follows: Figure 4 As shown;

[0285] according to Figure 4 The contribution rate threshold δ PC Set to 0.9, nine principal components were selected for concept drift detection. During online drift detection, a moving window of size 5 was used. Table 3 provides statistics on the number of drift samples, the number of online incremental neurons, the updated model RMSE, and the update time of the online incremental neurons. The T value for each sample within the moving window is also shown. 2 and SPE results and T 2 and SPE control limits Figure 5a and Figure 5b As shown;

[0286] Table 3. Drift detection and model update results based on the online test dataset.

[0287]

[0288]

[0289] As shown in Table 3, there are 5 drift samples within each moving window, which means that the online test dataset shows a significant change compared to the historical training dataset. This is consistent with... Figure 4 The distribution of DXN data shown is consistent with the data distribution shown. Furthermore, as shown in Figure 5, T... 2 The fact that SPE statistics consistently exceeded control limits across all windows indicates a substantial change in the operating conditions represented by the process data; therefore, updating the soft measurement model is crucial.

[0290] During online learning, HNTBLS incremental learning had an average of 3 neuron groups per window. Compared to the number of incremental basic units, HNTBLS had fewer than FTBLS (FTBLS had 5 incremental units online). HNTBLS had an average RMSE of 8.2885e-06 per window, indicating that the online learning performance of the HNTBLS method was better than that of FTBLS (FTBLS had an average RMSE of 2.4325e-03). In addition, the average update time of HNTBLS online learning per window was 0.03 seconds, which was about 20 times faster than FTBLS (average update time of 0.63 seconds).

[0291] Compared with FTBLS, the proposed method has significant advantages in learning performance, update time and modeling accuracy for both historical training datasets and online test datasets.

[0292] This invention also provides an embodiment, namely the DXN online soft measurement software, the interface and running results of which are as follows: Figure 6 As shown, the left panel of the interface provides online drift detection and soft measurement result display (including numerical values ​​and curves), DXN sampling date recording, DXN true value statistics and other functions. The right panel displays real-time running process data information, including comprehensive process information and key variable data. For actual MSWI operation, a DCS data security isolation acquisition device has been developed to ensure the safety and integrity of industrial processes. The network connection method adopts the OPC communication protocol.

[0293] This invention provides a soft measurement method for dioxin emission concentration. The method includes constructing a concept drift detection model, constructing a soft measurement model, updating the concept drift detection model, updating the soft measurement model, acquiring actual online data from the plant, performing drift detection based on the updated concept drift detection model, and if no drift occurs, performing measurement based on the updated soft measurement model. If drift occurs, matching the most similar historical sample to the current operating condition, acquiring multiple measurement values, calculating the arithmetic mean to obtain the soft measurement value, issuing a concept drift warning, and updating both models based on the obtained DXN ground truth samples after completing online sampling and offline testing under similar operating conditions. This method proposes a hybrid neuron and tree width learning system, where the offline and online incremental learning layers use neuron groups, improving the model update speed and modeling accuracy. Experiments using historical data and actual MSWI plant applications demonstrate that the proposed method significantly improves the performance and effectiveness of the DXN soft measurement system. By adopting these algorithms and technologies, researchers and practitioners can improve the effectiveness of online soft measurement systems for DXN emission monitoring, which helps improve environmental management and regulatory compliance, ultimately minimizing the impact of DXN emissions on public health and the environment.

[0294] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this 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 this invention.

Claims

1. A soft measurement method for dioxin emission concentration, characterized in that, Includes the following steps: Step 1: Construct a concept drift detection model; Step 2: Construct a soft measurement model; Step 3: Update the concept drift detection model; Step 4: Update the soft measurement model; Step 5: Obtain actual online data from the factory, perform drift detection based on the updated concept drift detection model. If no drift occurs, perform measurement based on the updated soft measurement model. If drift occurs, match the historical sample most similar to the current working condition, obtain multiple measurement values, calculate the soft measurement value by arithmetic mean, issue a concept drift warning, and update the two models based on the obtained DXN ground truth samples after completing online sampling and offline testing under similar working conditions. In step 1, a concept drift detection model is constructed, specifically as follows: Obtain historical training data, including historical training data. From high-dimensional process variables and the corresponding true values ​​of DXN emission concentrations composition; Obtain the historical training data matrix with zero mean and unit variance. Let the noise data matrix be... The actual process data matrix is ,in The original process data matrix It can be decomposed into: （1） In the formula, It is a low-rank matrix. It is a sparse matrix; Transforming it into a convex optimization problem, we get: （2） In the formula, For matrix norm, For matrix The sum of vector magnitudes; The problem is solved iteratively using the alternating direction method. Assume there exists a multiplier matrix for the linear constraints. Sum of positive penalty parameters Then we obtain the augmented Lagrange function, which is: （3） Minimize using an iterative strategy and ,get: （4） （5） （6） for Using the subgradient operator to evaluate the convex function, we obtain: （7） make ,get: （8） In the formula, Let be the Euclidean projection on the set, where: （9） for Expressing it as a minimization problem, we get: （10） In the formula, Let a matrix be decomposed into singular value decomposition as follows: , Solving for the problem yields: （11） use The low-rank matrix is ​​obtained by iterative calculation. ,in Using low-rank matrices Establish T 2 and the control limits of SPE; low-rank matrix Correlation matrix Written as: （12） Calculate the correlation matrix based on the SVD method. The eigenvalues ​​of are: （13） In the formula, and Represents an orthogonal matrix , It is a diagonal matrix. yes The One eigenvalue; Based on contribution rate The principal components are selected as follows: （14） In the formula, To select the number of principal components, The threshold was set based on experience; According to the selection Each principal component will be loaded into a principal component matrix. Represented as , low-rank matrix Written as: （15） In the formula, and Represent the principal space and the residual space, respectively; Using T 2 And SPE statistical indicators monitor concept drift, where T 2 SPE represents the change in the principal space, while SPE simultaneously represents the change in both the principal space and the residual space in a low-rank matrix. Below, statistics T 2 The control limits of the indicator, i.e. and the control limits of the SPE index, i.e. , is represented as: （16） （17） In the formula, For having and F-distribution of degrees of freedom Indicates about The normality deviation of percentiles, parameters , , and The solution is: （18） （19）； In step 2, the soft measurement model is constructed, specifically as follows: Get by One sample and Historical training dataset composed of 3D features The mapping layer is composed of multiple RDT units, each RDT unit consisting of non-leaf nodes and leaf nodes. Let the RDT model have... There are nodes, among which... Indicates the number of leaf nodes. This represents the number of non-leaf nodes during the tree growth process of an RDT. Non-leaf nodes are determined through random splitting, and are represented as follows: （20） In the formula, Indicates the first There are non-leaf nodes, among which ; Based on the The information of each non-leaf node is used to partition the tree structure of the RDT growth strategy, which is represented as follows: （21） In the formula, and This represents the data of two child nodes. and express and The number of samples in the sample; Obtained using a recursive procedure There are non-leaf nodes. The leaf nodes use an arithmetic mean strategy to calculate the final predicted value, represented as: （22） In the formula, For the first The number of samples in each leaf node. For the first The true value of each leaf node; The RDT model is represented as follows: （23） In the formula, For the first in the mapping layer One RDT function, Indicates the first The first RDT One path; Suppose the mapping layer has If there are 1 RDT function, then the output of the mapping layer is represented as: （24） （25） The enhancement layer has a similar structure and units to the mapping layer. Let the enhancement layer consist of... It consists of 1 RDT function, denoted as The output of the enhancement layer is then expressed as: （26） （27） By integrating the mapping layer and the enhancement layer, the output is represented as The output of the output layer is: （28） in, Estimate the weights of each RDT function in the mapping and enhancement layers: （29） In the formula, , express The pseudo-inverse matrix, It is a regular ratio. Represents the identity matrix; When both the mapping layer and the enhancement layer perform poorly, the width of the offline incremental layer is adjusted. Each group in the offline incremental layer consists of multiple neurons. When adding the first group of neurons, the corresponding mapping process is as follows: （30） （31） In the formula, This represents the output of the first group of neurons. Indicates the number of neurons. Indicates the first The output of each neuron, where , For activation function, and These represent the weights and biases, respectively. The outputs of the mapping layer, enhancement layer, and offline incremental layer are represented as follows: ; The Moore-Penrose inverse matrix is ​​solved using a column recursive algorithm, and the new inverse matrix is ​​written as: （32） In the formula, （33） Update the weight matrices of the mapping layer, enhancement layer, and offline incremental layer: （34） After adding the first group of neurons, the prediction results of the offline model are as follows: （35）。 2. The soft measurement method for dioxin emission concentration according to claim 1, characterized in that, In step 3, the concept drift detection model is updated, specifically as follows: Based on the concept drift detection model, a fixed-size moving sampling window is used during the detection process. When new data is acquired by the sampling window, the zero-mean and unit-variance data matrix and the correlation matrix are calculated, based on the high-dimensional historical training data matrix. and correlation matrix The mean of each column is expressed as The online test dataset in the historical data is denoted as Combined with window Online data within New data matrix and correlation matrix The recursive calculation is as follows: （36） In the formula, （37） Finally, a low-rank matrix is ​​obtained. ; Principal component loading matrix calculated based on principal components. T of samples within the window 2 And SPE are represented as: （38） （39） when Less than a certain control limit and Less than a certain control limit At that time, it was determined that no drift had occurred; when Exceeding control limits or Exceeding control limits At that time, it is determined that a drift has occurred.

3. The soft measurement method for dioxin emission concentration according to claim 2, characterized in that, In step 4, the soft measurement model is updated, specifically as follows: when and The values ​​are all less than their control limits. and When this occurs, it indicates that no drift has occurred, meaning that the current operating conditions are similar to historical operating conditions. That is, the current DXN emission concentration is within the distribution range of historical data. The soft sensor model is then rewritten to obtain: （40） In the formula, , and These are the activation functions for the mapping layer, enhancement layer, and offline incremental layer, respectively. when value or The value exceeds its control limit and When this occurs, the soft sensor model is considered to have drifted, and a drift sample containing process data and DXN ground truth is used to update the model. The outputs of the mapping layer, enhancement layer, and offline incremental layer are represented as follows: Add an online incremental layer, and represent the outputs of the mapping layer, enhancement layer, offline incremental layer, and online incremental layer as follows: ,in, The new online inverse matrix is ​​written as: （41） in, （42） The new weight matrix is ​​represented as follows: （43） After adding the first online neuron group, the updated soft measurement model is represented as follows: （44）。 4. The soft measurement method for dioxin emission concentration according to claim 3, characterized in that, In step 5, actual online data from the factory is acquired, and drift detection is performed based on the updated concept drift detection model. If no drift is detected, measurements are taken based on the updated soft measurement model. If drift occurs, the most similar historical sample to the current operating condition is matched, multiple measurement values ​​are obtained, and the soft measurement value is obtained by calculating the arithmetic mean. A concept drift warning is issued. After completing online sampling and offline testing under similar operating conditions, both models are updated based on the obtained DXN ground truth samples. Specifically: Acquire actual online data from the factory and perform drift detection based on the updated conceptual drift detection model. If no drift is detected, then measurement is performed based on the updated soft measurement model; if drift occurs, i.e. Value exceeds control limit or Value exceeds control limit At this point, the similarity between the drifted sample and the historical sample is calculated using Euclidean distance, as follows: （45） By manually setting thresholds Select the most similar historical samples ,for: （46） The most similar historical samples Input the data into the most recently updated soft measurement model to obtain the measurement vector. ; For measurement vector The arithmetic mean is calculated to obtain the final measurement value under the current drift condition. Issue an alarm to the expert operating the MSWI process indicating a condition drift, prompting them to record the condition. Under similar working conditions, online sampling and offline testing were performed to obtain the true value of DXN; This ground truth sample representing the drift condition is then updated to the online test dataset in the historical data, triggering steps 3 and 4 to update the concept drift detection model and the soft measurement model.

Images