A method for detecting the concentration of dxn in the MSWI process

Abstract

This invention provides a method for detecting DXN concentration during MSWI (Maximum Saturation Wave Wiring) processes, comprising: acquiring multimodal data; preprocessing the multimodal data to obtain preprocessed multimodal data; constructing a full-process model of DXN concentration; and inputting the preprocessed multimodal data into the full-process model of DXN concentration to obtain DXN emission concentration, DXN generation concentration, and DXN adsorption concentration. This invention solves the problem in existing technologies where the detection cost and efficiency of DXN generated during MSWI processes are high, making it impossible to achieve optimized emission reduction of DXN concentration from its root cause.

Description

Technical Field

[0001] This invention relates to the field of energy conversion technology, and in particular to a method for detecting DXN concentration in MSWI processes. Background Technology

[0002] Under uninterrupted conditions, the global annual production of municipal solid waste (MSW) is projected to reach 3.4 billion tons by 2050. In recent decades, many cities have actively pursued the development of MSW incineration (MSWI) technology to prevent ecological damage and achieve sustainable development goals. MSWI technology is one of the most effective methods for converting MSW into valuable resources, achieved through a series of processes including fermentation, combustion, heat exchange, and gas purification. With the rapid development of MSWI plants, concerns have arisen regarding the secondary environmental pollution caused by their exhaust emissions, particularly dioxins (DXN). DXN, a persistent organic pollutant, is an unintentional byproduct of the MSWI process and can adversely affect the surrounding environment and human health. Advanced air pollution control equipment has successfully and significantly reduced pollutant emission concentrations during MSW treatment. However, the MSWI process remains a major source of DXN emissions into the atmosphere. Due to the significant lag and high cost of existing DXN physical detection technologies, current DXN detection methods are insufficient to support reducing DXN concentrations during the MSWI process at its source through optimized control technologies. Therefore, achieving online real-time measurement of DXN concentration has become a key solution to resolving the inherent contradiction between the sustainable development of the MSWI process and the adverse environmental impacts related to DXN.

[0003] Artificial intelligence (AI)-based soft measurement methods can be used for online measurement of DXN concentration in MSWI processes. However, current inventions all use process numerical data to detect DXN emission concentrations from MSWI process chimneys. These inventions cannot perform full-process soft measurement of DXN concentration across the three stages of DXN generation, absorption, and emission. Consequently, they cannot support optimized emission reductions to reduce DXN concentration at its source. In summary, existing technologies for detecting DXN generated in MSWI processes are costly and inefficient, failing to achieve optimized emission reductions to address the root causes of DXN generation. Summary of the Invention

[0004] To overcome the shortcomings of the prior art, the purpose of this invention is to provide a method for detecting DXN concentration during MSWI. This invention solves the problem that the prior art has high detection costs and low detection efficiency for DXN generated during MSWI, which makes it impossible to achieve optimized emission reduction by reducing DXN concentration from the root cause of DXN generation.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for detecting DXN concentration during MSWI, comprising:

[0007] Acquire multimodal data;

[0008] The multimodal data is preprocessed to obtain preprocessed multimodal data;

[0009] Construct a full-process model of DXN concentration;

[0010] The preprocessed multimodal data is input into the DXN concentration full-process model to obtain DXN emission concentration, DXN generation concentration and DXN adsorption concentration.

[0011] Preferably, the method for constructing a full-process model of DXN concentration includes:

[0012] Construct sub-models for DXN generation concentration, DXN emission concentration, and DXN adsorption concentration;

[0013] The DXN generation concentration sub-model, DXN emission concentration sub-model, and DXN adsorption concentration sub-model are cascaded to obtain the full-process model of DXN concentration.

[0014] Preferably, the multimodal data includes:

[0015] The process numerical data, flame video data, and recording operation data, wherein the flame video data includes: left flame video data and right flame video data.

[0016] Preferably, the recorded operation data includes:

[0017] Feed rate, combustion efficiency, average load, steam production per ton of MSW, and calorific value of MSW.

[0018] Preferably, the multimodal data is preprocessed to obtain preprocessed multimodal data, including:

[0019] The time of the process numerical data, the time of the flame video data, and the time of the recording operation data are synchronized with the time of DXN sampling to obtain the process numerical data, the flame video data, and the recording operation data corresponding to the time.

[0020] Pre-calculate the recording operation data corresponding to the time to obtain a valid recording operation dataset;

[0021] The process numerical data corresponding to the time is subjected to a first average data processing to obtain the average data;

[0022] The average data is sliced ​​to obtain numerical data of DXN generation concentration process, DXN adsorption concentration process, and DXN emission concentration process.

[0023] Extract second-level image information from the flame video data corresponding to the time;

[0024] The second-level image information is subjected to a second averaging process to obtain second-level data;

[0025] The second-level data is processed into grayscale to obtain grayscale image information;

[0026] The grayscale image information is subjected to a third averaging process to obtain the fused video information;

[0027] Preprocessed multimodal data is obtained based on the valid recording operation dataset, DXN generated concentration process numerical data, DXN adsorption concentration process numerical data, DXN emission concentration process numerical data, and fused video information.

[0028] Preferably, the formula for pre-calculating the recording operation data corresponding to the time includes:

[0029]

[0030]

[0031]

[0032]

[0033]

[0034] Where, x FR x IE x AL x SP and x CV These represent the feed rate, incineration efficiency, average load, steam production per ton of municipal solid waste, and MSW calorific value, respectively; Q1 represents the rated capacity of the MSWI process; Q2 represents the actual daily roll processing volume; x 87 H represents steam flow rate; Steam Indicates the enthalpy of vapor; H Water Indicates the enthalpy of water; η Boiler This refers to the boiler's thermal efficiency.

[0035] Preferably, the calculation formula for the first averaging process is:

[0036]

[0037] Where, x i M is the i-th process variable.PND Let T be the dimension of the process numerical data, and T be the time scale.

[0038] Preferably, the method for constructing the DXN concentration sub-model is as follows:

[0039] The construction process numerical data-driven unit and the flame image data-driven unit;

[0040] A DXN concentration generation sub-model is constructed based on the process numerical data driving unit and the flame image data driving model.

[0041] The method for constructing the process numerical data driven unit is as follows:

[0042] Based on the preprocessed multimodal dataset, the process numerical data driving unit is constructed using a cross-layer fully connected adaptive deep forest regression algorithm.

[0043] The numerical data-driven unit described above consists of three parts: a layered forest model, cross-layer fully connected components, and layer adaptive components.

[0044] The method for constructing the flame image data driving unit is as follows:

[0045] A flame image data-driven unit is constructed based on CNN and process numerical data-driven units.

[0046] Preferably, the method for constructing the DXN adsorption concentration sub-model is as follows:

[0047] Virtual data is obtained by generating virtual samples from the DXN-generated concentration and the preprocessed multimodal data;

[0048] Based on the fuzzy forest regression algorithm, a DXN adsorption concentration sub-model is constructed according to the virtual data.

[0049] Preferably, the method for constructing the DXN emission concentration sub-model is as follows:

[0050] The DXN adsorption concentration is obtained according to the DXN adsorption concentration sub-model.

[0051] Based on the width forest regression algorithm, a sub-model of DXN emission concentration is constructed according to the DXN adsorption concentration and the preprocessed multimodal data.

[0052] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0053] This invention provides a method for detecting DXN concentration during MSWI (Maximum Saturation Wave Induction) processes, comprising: acquiring multimodal data; preprocessing the multimodal data to obtain preprocessed multimodal data; constructing a full-process DXN concentration model; and inputting the preprocessed multimodal data into the full-process DXN concentration model to obtain DXN emission concentration, DXN generation concentration, and DXN adsorption concentration. This invention, by constructing a full-process DXN concentration model, visualizes the DXN concentration distribution at multiple stages during MSWI, reducing the cost of DXN concentration monitoring and improving the effectiveness of DXN emission reduction. Attached Figure Description

[0054] 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.

[0055] Figure 1 A flowchart of a method for detecting DXN concentration during MSWI provided in an embodiment of the present invention;

[0056] Figure 2 A schematic diagram of the MSWI process for the generation, adsorption and emission stages of DXN provided in an embodiment of the present invention;

[0057] Figure 3 This is a schematic diagram of the sampling date and time of DXN samples provided in an embodiment of the present invention;

[0058] Figure 4 The diagram shows the distribution of DXN concentration at various stages according to the embodiments of the present invention, wherein (a) is a schematic diagram of the DXN concentration generation and emission stages, and (b) is a schematic diagram of the DXN concentration adsorption stage.

[0059] Figure 5 This is a schematic diagram of the DXN concentration full-process modeling strategy provided in an embodiment of the present invention;

[0060] Figure 6 This is a data distribution diagram provided for an embodiment of the present invention;

[0061] Figure 7 The diagram below shows the training results of the model provided in the embodiment of the present invention, wherein (a) is a schematic diagram of the training error curve of the process numerical data, (b) is a schematic diagram of the training error curve of the flame image data, and (c) is a schematic diagram of the training results of the CNN model.

[0062] Figure 8The diagram shows the training results of the VSG-based absorption model provided in the embodiment of the present invention, wherein (a) is a schematic diagram of the training error curves of the encoder and decoder, and (b) is a schematic diagram of the distribution of the true value and the pseudo-label value;

[0063] Figure 9 A schematic diagram of the training error curve of the BFR-based DXN emission model provided in an embodiment of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] The purpose of this invention is to provide a method for detecting DXN concentration during MSWI. This invention solves the problems of high cost and low efficiency in the detection of DXN generated during MSWI in the prior art, which makes it impossible to achieve optimized emission reduction by reducing DXN concentration from the root cause of DXN generation.

[0066] 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.

[0067] like Figure 1 As shown, the present invention provides a method for detecting DXN concentration during MSWI, comprising:

[0068] Step 100: Acquire multimodal data;

[0069] Step 200: Preprocess the multimodal data to obtain preprocessed multimodal data;

[0070] Step 300: Construct a full-process model of DXN concentration;

[0071] Step 400: Input the preprocessed multimodal data into the DXN concentration full-process model to obtain DXN emission concentration, DXN generation concentration and DXN adsorption concentration.

[0072] Specifically, the MSWI process consists of feeding, incineration, heat exchange, gas purification, and emission stages, such as... Figure 2 As shown in the diagram, the system is a typical MSW grate-type incinerator with a daily processing capacity of 800 tons. The flue gas purification stage is equipped with advanced air pollution control devices (APCDs), including selective non-catalytic reduction, activated carbon injection, semi-dry acid removal, and a bag filter system.

[0073] like Figure 2 As shown, the MSWI process exhibits a clear cascading relationship across multiple stages. Initially, raw MSW containing trace amounts of DXN (approximately 0.8 ng TEQ / kg) is collected and stored for about a week for fermentation to increase its calorific value. Subsequently, the MSW is fed into the furnace via a feeder and undergoes a series of processes, including drying, combustion, and burnout, transforming into high-temperature flue gas (above 850°C). At this high temperature, DXN decomposes. Then, the heat recovery stage begins, where the high-temperature flue gas must be cooled to 200°C. Numerous inventions have shown that there are three pathways for DXN generation at this stage: homogeneous synthesis (500-800°C), precursor synthesis (200-400°C), and de novo synthesis (200-400°C). Therefore, in this invention, the DXN generation point for online sampling and offline analysis is determined to be the economizer outlet (e.g., Figure 2 (As shown). Next, the cooled flue gas is guided to the gas purification stage, where an APCD device is used to remove toxic and harmful substances from the flue gas. The regenerated DXN is adsorbed by activated carbon in the fly ash, and then the fly ash is separated from the flue gas using a bag filter. Therefore, the DXN adsorption point for online sampling and offline analysis in this invention is located at the bottom of the bag filter (e.g., as shown). Figure 2 (As shown). After flue gas purification, the exhaust gas is discharged from the chimney by an induced draft fan. Therefore, the induced draft fan outlet is considered the DXN emission point for online sampling and offline analysis, as well as the conventional DXN emission detection point and the sampling point in this invention.

[0074] In summary, this invention is used to model the DXN concentration distribution at the above three key locations: the economizer outlet (DXN generation), the bottom of the bag filter (DXN adsorption), and the induced draft fan outlet (DXN emission).

[0075] Specifically, in order to obtain full-process data from the MSWI process for DXN concentration modeling, this invention conducted online sampling of flue gas and fly ash at multiple monitoring points of an MSWI plant in Beijing from November 2020 to January 2022, and performed offline analysis in the laboratory in accordance with the national environmental protection standard (HJ 77.2-2008).

[0076] like Figure 2 As shown, the flue gas sampling at the DXN generation and emission detection points adopted a constant velocity tracking mode, with a sampling duration exceeding 2 hours. During the aforementioned 2-hour sampling period, fly ash sampling adopted an equal time interval mode to ensure real-time collection of fly ash samples used for adsorption modeling.

[0077] A total of 47 valid sample sets were collected from the MSWI plant, including 47 generated samples, 47 emission samples, and 16 adsorption samples. The sampling date and time distribution is as follows: Figure 3 As shown.

[0078] Furthermore, in the laboratory, a quartz filter, XAD-2, and sample condensate were mixed for sample extraction and purification. Labeled 13C12 was used as an internal standard for the sample during pretreatment. During instrumental analysis, the temperature program for high-resolution gas chromatography-high-resolution mass spectrometry (HRGC / HRMS) was as follows: 1) the temperature was held at 150°C for 3 minutes; 2) the temperature was increased to 230°C at a rate of 20°C / min and held for 18 minutes; 3) the temperature was increased to 235°C at a rate of 5°C / min and held for 10 minutes; 4) the temperature was increased to 320°C at a rate of 4°C / min and held for 3 minutes.

[0079] The DXN concentration obtained from the HRGC / HRMS device will be labeled as c here. s It is further converted to obtain the TEQ concentration under dry air at 101.325 kPa and 237 K (11% oxygen). The conversion process is described below.

[0080]

[0081] In the formula, c represents the TEQ concentration, with units of ngI-TEQ / Nm³. 3 ; This indicates the oxygen content at the chimney.

[0082] Based on the above sampling strategy, the time scale of DXN concentration data includes: the 2-hour average emission concentration of DXN during the generation and emission stages, and the 4-hour emission concentration during the adsorption stage.

[0083] The TEQ concentration distribution of DXN at the generation, adsorption, and emission points is as follows: Figure 4 As shown, where, Figure 4 (a) is a schematic diagram of the DXN concentration generation and emission stages. Figure 4 (b) is a schematic diagram of the DXN concentration adsorption stage.

[0084] Figure 4 The results show that: 1) Data from the DXN generation stage (sample size 36), DXN adsorption stage (sample size 16), and DXN emission stage (sample size 36) are characterized by small sample sizes; 2) The number of DXN samples in the adsorption stage is extremely small compared to the generation and emission stages; 3) The DXN concentration distribution shows that the distribution of DXN data in the adsorption stage varies considerably, indicating that the data in this stage has a high degree of uncertainty.

[0085] The operational data generated by the MSWI process can be divided into three categories: process numerical data, flame video data, and recorded operational data, such as... Figure 2 As shown. Considering the actual detection points for DXN generation, adsorption, and emission, the above three types of data can be further subdivided according to the following principles:

[0086] 1) In the DXN generation stage, the model input data used includes data from the three modes mentioned above. The process numerical variables in this stage correspond to the feeding, combustion, and heat exchange stages, and Table 1 provides specific information on these variables. Recorded operational data includes feed rate, combustion efficiency, average load, steam production per ton of MSW, and calorific value of MSW. Additionally, the flame video per second contains 25 frames, each image measuring 576 x 720 x 3 pixels.

[0087] 2) In the DXN adsorption stage, the model input data used includes process numerical data and DXN concentration during the generation stage. The process numerical variables mainly focus on activated carbon and lime, and specific variable information is shown in Table 1.

[0088] 3) In the DXN emission stage, the model input data used includes process numerical data, DXN generation concentration, and adsorption concentration. The gas purification and emission stages provide process numerical variables for this stage; Table 1 shows the process numerical variable information for the same stage. Table 1 is shown below:

[0089] Table 1. Information on process numerical variables during the same stage.

[0090]

[0091]

[0092]

[0093]

[0094]

[0095] In Table 1, the dimensions of the process numerical variables at different stages are 119, 5, and 34, respectively.

[0096] Based on the existing data, the input variables in the DXN generation stage exhibit significant high-dimensionality, the variables in the DXN adsorption stage exhibit low-dimensionality, and the variables in the DXN emission stage exhibit medium-dimensionality. In summary, the characteristics of the learning and modeling data at different stages are as follows: the DXN generation stage exhibits multimodal, small sample size, and high-dimensionality features; the DXN adsorption stage exhibits extremely small sample size and low-dimensionality features; and the DXN emission stage exhibits small sample size and medium-sized features. Therefore, different AI-based modeling algorithms should be used to address these characteristics.

[0097] Furthermore, methods for constructing a full-process model of DXN concentration include:

[0098] Construct sub-models for DXN generation concentration, DXN emission concentration, and DXN adsorption concentration;

[0099] The DXN generation concentration sub-model, DXN emission concentration sub-model, and DXN adsorption concentration sub-model are cascaded to obtain the full-process model of DXN concentration.

[0100] Specifically, considering the data characteristics of DXN in its generation, adsorption, and emission stages, a comprehensive modeling strategy, FFFM, is proposed. This strategy includes modules for multimodal data preprocessing, DXN generation concentration model construction, DXN adsorption concentration model construction, and DXN emission concentration model construction. In the FFFM structure, the outputs of soft sensing at each stage are cascaded to achieve full-process modeling of DXN concentration, such as... Figure 5 As shown.

[0101] The functions of the four modules of the FFFM strategy proposed in this invention are described below.

[0102] 1) Multimodal data preprocessing module: The collected multimodal data includes process numerical data, left flame video data, right flame video data, and recording operation data. Based on the time scale of online sampling of DXN data, the multimodal data is sliced ​​and preprocessed according to the generation, adsorption, and emission stages.

[0103] 2) DXN-generated concentration model building module: Two types of data can be used for modeling: process numerical data and flame video data. A new ADFRclfc algorithm is proposed for the DXN-generated concentration model, and a separate modeling strategy is designed for each modality of data. ADFRclfc is used for high-dimensional, small-sample process numerical data, while convolutional neural networks (CNN) are combined with ADFRclfc for flame video data modeling.

[0104] 3) DXN Adsorption Concentration Model Construction Module: Ultra-small sample size and data uncertainty are key challenges that need to be addressed in the adsorption concentration model. To address the ultra-small sample size problem, a novel VSG method, namely latent variable VSG (LVVSG), is proposed to compensate for the lack of modeling samples; the fuzzy forest regression algorithm, namely FFR, is used to build the DXN adsorption concentration model.

[0105] 4) DXN emission concentration model construction module: Using the new BFR method, with DXN generation concentration, DXN adsorption concentration and process numerical variables as input, it can quickly and accurately learn medium-dimensional and small-sample process numerical data.

[0106] Furthermore, the multimodal data includes:

[0107] The process numerical data, flame video data, and recording operation data, wherein the flame video data includes: left flame video data and right flame video data.

[0108] Furthermore, the recorded operation data includes:

[0109] Feed rate, combustion efficiency, average load, steam production per ton of MSW, and calorific value of MSW.

[0110] Furthermore, the multimodal data is preprocessed to obtain preprocessed multimodal data, including:

[0111] The time of the process numerical data, the time of the flame video data, and the time of the recording operation data are synchronized with the time of DXN sampling to obtain the process numerical data, the flame video data, and the recording operation data corresponding to the time.

[0112] Pre-calculate the recording operation data corresponding to the time to obtain a valid recording operation dataset;

[0113] The process numerical data corresponding to the time is subjected to a first average data processing to obtain the average data;

[0114] The average data is sliced ​​to obtain numerical data of DXN generation concentration process, DXN adsorption concentration process, and DXN emission concentration process.

[0115] Extract second-level image information from the flame video data corresponding to the time;

[0116] The second-level image information is subjected to a second averaging process to obtain second-level data;

[0117] The second-level data is processed into grayscale to obtain grayscale image information;

[0118] The grayscale image information is subjected to a third averaging process to obtain the fused video information;

[0119] Preprocessed multimodal data is obtained based on the valid recording operation dataset, DXN generated concentration process numerical data, DXN adsorption concentration process numerical data, DXN emission concentration process numerical data, and fused video information.

[0120] Multimodal data preprocessing module such as Figure 5 As shown, various data types have different time scales, such as hourly time scales for DXN concentration, daily time scales for recording operational data, and second-level time scales for process numerical data and flame video data.

[0121] First, the timing of multimodal data collection is synchronized with the DXN sampling time. That is, the two-hour average concentration of DXN corresponds to two hours of process numerical data, two hours of flame video, and a set of recorded operational data.

[0122] The recorded operational data obtained from domain experts includes feed rate (tons / hour), combustion efficiency (%), average load (tons / hour), steam production per ton of MSW (tons / ton), MSW calorific value (kJ), etc., and its preprocessing is as follows (pre-calculated formula):

[0123]

[0124]

[0125]

[0126]

[0127]

[0128] Where, x FR x IE x AL x SP and x CV These represent the feed rate, incineration efficiency, average load, steam production per ton of municipal solid waste, and MSW calorific value, respectively; Q1 represents the rated capacity of the MSWI process; Q2 represents the actual daily roll processing volume; x 87 H represents steam flow rate; Steam Indicates the enthalpy of vapor; H Water Indicates the enthalpy of water; η Boiler This refers to the boiler's thermal efficiency.

[0129] The valid record operation dataset obtained through the above calculations is represented as follows: Where, N GIt is the sample size, M RD Indicates the dimension of the variable.

[0130] The process of numerical data processing includes two parts: data averaging and variable slicing.

[0131] The specific method for averaging data (the calculation formula for the first averaging process) is as follows:

[0132]

[0133] Where, x i M is the i-th process variable. PND Let T be the dimension of the process numerical data, and T be the time scale.

[0134] The MSWI process numerical dataset is obtained by averaging the data, and is represented as follows: Where M PND The dimension of the numerical variable in the process.

[0135] Subsequently, based on Table 1, the numerical dataset of this process was divided into three subsets: DXN-generated concentration process numerical data. Numerical data on DXN adsorption concentration process Numerical data on DXN emission concentration processes

[0136] The raw flame video data consists of two parts, left and right, each containing different information. The processing flow is as follows:

[0137] 1) In actual combustion, the flame changes relatively slowly. Therefore, to correspond to a second-level timescale, the first frame of the video per second is extracted as image data, with approximately 180,000 flame image data points on each side (left and right), each with a size of 576×720x3 pixels.

[0138] 2) Subsequently, the image information at the second-scale is averaged to obtain approximately 3000 flame images (576x720x3 pixels) on each side. Next, these images are converted to grayscale (from 3 channels to 1 channel) and resized to 32x32 pixels. Then, flame image data is acquired from the left and right sides, with approximately 3000 images (32x32 pixels) on each side.

[0139] 3) Averaging the 3000 images from both sides yields a single flame image of 32x32 pixels. Then, averaging the flame images from both sides again yields the final image data, completing the fusion of video information from both sides at an hourly scale. This process results in a 32x32 pixel flame image dataset.

[0140] Based on the above video preprocessing, the MSWI flame dataset will be obtained, represented as follows:

[0141] Furthermore, the method for constructing the DXN-generated concentration sub-model is as follows:

[0142] The construction process numerical data-driven unit and the flame image data-driven unit;

[0143] A DXN concentration generation sub-model is constructed based on the process numerical data driving unit and the flame image data driving model.

[0144] The method for constructing the process numerical data driven unit is as follows:

[0145] Based on the preprocessed multimodal dataset, the process numerical data driving unit is constructed using a cross-layer fully connected adaptive deep forest regression algorithm.

[0146] The numerical data-driven unit described above consists of three parts: a layered forest model, cross-layer fully connected components, and layer adaptive components.

[0147] The method for constructing the flame image data driving unit is as follows:

[0148] A flame image data-driven unit is constructed based on CNN and process numerical data-driven units.

[0149] Specifically, in the DXN concentration model building module, the input data includes the recorded running dataset. Process numerical dataset Flame Dataset Concentration sets generated by DXN Two models were used in the following two modules: the input is represented as The process numerical data-driven model and input representation are The flame image data-driven model.

[0150] The ADFR clfc model consists of three parts: a layered forest model, cross-layer fully connected components, and layer adaptive components. Each layer of the forest model consists of a random forest (RF) and a complete forest forest (CRF), and each layer of the forest model consists of multiple DT units.

[0151] The training dataset is With the kth G1 Taking layers as an example, the modeling process is described below.

[0152] k G1 The input data for the th layer is:

[0153]

[0154] in, It represents the (k)th G1-1) The output of layer 1 The dimension representing the feature. Indicates k G1 The input representation data of the th layer, J G1,Forest This indicates the forest model number.

[0155] Then, the Bootstrap and random subspace methods are used to represent the input dataset. Preprocessing is performed, as shown below:

[0156] Among them, f Bootstrap (·) is the bootstrap sampling function, J G1,DT f represents the number of bootstraps. RSM (·) represents the feature sampling function. Input data The jth G1,DT A subset, M RSM for The number of features in, where

[0157] Next, using Establish a DT forest model. In RF, each DT uses mean squared error to find the optimal cutoff point to complete tree growth, as shown below:

[0158]

[0159]

[0160] in, The optimal feature sequence number, For optimal eigenvalues, R L and R R The datasets for the left and right child nodes are respectively, c L and c R R respectively L and R R The output average value.

[0161] By using the constraint of the minimum sample in the leaf node, the j-th G1,DT A DT model can be written as follows:

[0162]

[0163] Where I(·) is the sign function.

[0164] Therefore, containing J G1,DT The RF model of a DT model can be represented as:

[0165]

[0166] In CRF, the growth of each complete DT is based on a random split point, and the remaining policies are consistent with DT. Therefore, the CRF model can be represented as... Its description process has been simplified.

[0167] Given the number of RFs and CRFs, the layered forest model is written as:

[0168]

[0169] The structure of ADFR clfc is as follows: Figure 5 As shown, the layer regression vector is the output of the intermediate layer forest model.

[0170] k G1 The layer regression vector is represented as follows:

[0171]

[0172] In the learning process of ADFR-clfc, a cross-layer fully connected strategy is adopted to improve the representation performance of the layer regression vectors. Therefore, it can be written as follows (k G1 Input feature information of layer +1):

[0173]

[0174] in,

[0175] in, It is the kth G1 Output of the forest model.

[0176] Compared to the DFR-clfc method, ADFR-clfc employs a layer-adaptive mechanism to improve learning performance. The learning process automatically stops when the errors converge during layer-by-layer learning. Specifically:

[0177]

[0178] in,

[0179]

[0180] Where, θ G1 The threshold is set manually.

[0181] When the error is not satisfactory, the ADFR clfc model continues to increase the number of layers to obtain better learning results. When the ADFRclfc model learns K... G1 When layering, assuming the error meets the requirements, the ADFR clfc model can be obtained, denoted as F. G1 (·)

[0182] Furthermore, transfer-based CNNs are used to analyze image data. Deep features are extracted from the image dataset, where the transfer CNN is derived from the flame state recognition model. The CNN model employs an 8-layer structure, including an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. The input dataset consists of flame images. The CNN model extracts deep features as follows:

[0183]

[0184]

[0185]

[0186]

[0187]

[0188]

[0189] in, and These are the output feature maps of convolutional layers 1 through 6, respectively; * indicates the convolution process. Indicates multichannel convolution operation; f tanh (·) denotes the Tanh activation function; mean(·) is the matrix averaging function; K i This represents the core matrix.

[0190] Then, the output feature maps of convolutional layers 5 and 6 are used as inputs to ADFR clfc, as shown below:

[0191]

[0192] Among them, F CNN (·) represents the CNN model. Therefore, the depth features of the acquired flame image can be represented as:

[0193] Then, using deep features and DXN concentration value To train a new ADFR-clfc model. Based on the above process, the obtained model is represented as F. G2 (·)

[0194] Prior to this, the DXN generation concentration model was written as follows:

[0195]

[0196] Furthermore, the method for constructing the DXN adsorption concentration sub-model is as follows:

[0197] Virtual data is obtained by generating virtual samples from the DXN-generated concentration and the preprocessed multimodal data;

[0198] Based on the fuzzy forest regression algorithm, a DXN adsorption concentration sub-model is constructed according to the virtual data.

[0199] Specifically, the modeling of the DXN adsorption concentration model involves two steps.

[0200] First, the modeling dataset is expanded based on the original dataset using VSG technology.

[0201] Then, fuzzy modeling is performed by mixing virtual sample data and real data. In the DXN adsorption concentration model building module, the input data includes the DXN generation concentration. and process numerical data The output data is the DXN adsorption concentration. Therefore, the original dataset available to VSG is represented as

[0202] 1) VSG technology for sample expansion

[0203] like Figure 5 As shown, the latent variable space is a two-dimensional coordinate system, and the encoder network has 4 output layers. The encoder network in this paper is designed with multiple hidden layers; therefore, the output of the encoder network is represented as follows:

[0204]

[0205] in, It is a two-dimensional mean vector. f represents a two-dimensional variance vector. Encoder (·) represents the encoder network model, W Encoder and b Encoder These are the weights and biases of the encoder network model, respectively.

[0206] By using the mean and variance information, the latent vector can be written as follows:

[0207]

[0208] in, It is a potential vector. This represents a random number that follows a standard normal distribution.

[0209] By using process numerical data X APND Input to encoder network model f Encoder The latent variable matrix is ​​obtained from (·). Therefore, it can be based on Z EncoderA two-dimensional coordinate system is constructed. Within this system, a VSG method based on clustering and geometric rotation is applied to generate virtual samples. First, the latent variable samples are clustered into K groups using K-means. VSG =3 clusters, as shown below:

[0210]

[0211] Among them, C k Indicates the kth VSG A dataset with multiple clusters. Indicates the kth VSG Cluster C k The number of latent variable samples, It is the kth VSG Cluster C k The center of mass.

[0212] Based on the obtained three clustering datasets and three centroids Three mutually tangent circular regions can be identified in a two-dimensional coordinate system. (like Figure 5 (As shown). The radii of the three circular regions are denoted as r1, r2, and r3, respectively. Based on the distribution of latent variable samples within each circular region, each region can be divided into multiple concentric rings, as shown below:

[0213]

[0214] in, It is the kth VSG The radius difference of each concentric ring in a cluster, Indicates the kth VSG The number of concentric rings in a cluster Indicates the latent variable sample and the kth... VSG The minimum distance between the centroids.

[0215] Therefore, each circle comprises multiple concentric rings with the same radius difference in a two-dimensional coordinate system. Virtual samples are then adaptively generated using a rotation weight matrix method based on rule information. Finally, the final virtual sample dataset of latent variables is represented as follows:

[0216]

[0217] In obtaining virtual samples of latent variables Then, a decoding network, sharing the same structure as the encoder network, virtually samples these latent variables and maps them back to the original, physically meaningful process numerical variables. Therefore, the output of the decoder network is represented as follows:

[0218]

[0219] in, It is a virtual process numerical sample, f Decoder (·) represents the decoder network model, W Decoder and b Decoder These are the weights and biases of the decoder network model, respectively.

[0220] In the decoder network model (f Decoder Based on (·), virtual samples can be input by latent variables. Computing unlabeled virtual process numerical samples The following section proposes a method for unlabeled virtual samples. The labeling and evaluation method for multi-tree collaboration consists of three steps: training a linear tree model, labeling and evaluating unlabeled virtual samples, and collaborative iteration.

[0221] The linear tree model shares the same structure and growth method as the DT (Tree-Dependent Tree) model, differing only in the method for calculating leaf nodes. The leaf node calculation for obtaining the linear tree structure is as follows:

[0222]

[0223] in, It is the kth VSG The output of the q-th leaf node in a linear tree This represents the feature space data of the q-th leaf node. This represents the weight vector (the weight vector is solved using the ridge regression method in this paper). express The number of samples in This represents the target value in the q-th leaf node.

[0224] Therefore, the kth VSG A linear tree model can be represented as follows:

[0225]

[0226] To label and evaluate unlabeled virtual samples, three linear tree models corresponding to the three classes of virtual samples were constructed. Subsequently, these virtual sample data are fed into the corresponding linear tree model and labeled using leaf nodes. The calculation process is shown below:

[0227]

[0228] in, It is the kth VSG The label of the i7th virtual sample in the group, where λ represents the regularization parameter.

[0229] Therefore, the pseudo-labeled virtual sample data can be represented as follows. Furthermore, without modifying the tree structure, the virtual sample data for each pseudo-label is applied to update each linear DT model. The evaluation metric for the pseudo-labeled virtual samples is determined as described below.

[0230]

[0231] in, It is the kth VSG The evaluation index of the i7th pseudo-labeled virtual sample in the group. Indicates the kth VSG The update weights of the q-th leaf node in a linear tree are determined. Evaluation sets for three sets of pseudo-labeled virtual samples are obtained using the equation.

[0232] Then, a collaborative strategy is used to select positive pseudo-labeled virtual samples, including the following steps: 1) Define a threshold θ Percent In each iteration, positive pseudo-labeled virtual samples are selected from the evaluation set. 2) The positive pseudo-labeled virtual samples are added to the training set of other linear tree models, and the leaf nodes in the linear tree models are updated. 3) The remaining pseudo-labeled virtual samples are input into the updated linear tree models to obtain the evaluation metric set. Therefore, the above collaborative strategy is used to obtain positive pseudo-labeled virtual sample data.

[0233] The dataset used for FFR modeling includes real data. Virtual sample data with positive and negative labels The combination of these two datasets is represented as The structure of the FFR model, such as Figure 5 As shown, this includes a collection of several fuzzy tree models in the form of weighted assignments, each fuzzy tree being based on a training dataset. Constructed as follows:

[0234] The structure of the fuzzy tree is determined. Then, a fuzzy system is constructed in the leaf nodes to complete the formation of the fuzzy tree. Assume the input data of the q-th leaf node is represented as... in yes The feature dimensions.

[0235] In the premise of the fuzzy system, each input feature is subjected to the following fuzzy processing:

[0236]

[0237] in, It is the kth Fuzzy The output of a fuzzy rule.

[0238] Given K in the premise sectionFuzzy In the case of fuzzy rules, normalization is used to obtain the k-th... Fuzzy The output weights of the fuzzy rules are as follows:

[0239]

[0240] In the following sections, no activation function is used (the number of neurons equals K). Fuzzy A single-layer hidden neural network is used to map each input feature, represented as...

[0241] Finally, K Fuzzy The leaf node output of the fuzzy rule is shown below:

[0242]

[0243] in, It is the jth FDT The output of the q-th leaf node in a fuzzy tree model.

[0244] In each fuzzy tree, the parameters of the antecedent and consequent are updated using gradient descent and least squares methods, respectively. Therefore, the j-th... FDT A fuzzy tree model can be represented as:

[0245]

[0246] After obtaining J FDT After generating fuzzy trees, the ridge regression method is used to assign weights to each fuzzy tree, as shown below:

[0247]

[0248] Finally, the DXN absorption model based on the FFR algorithm is written as follows:

[0249]

[0250] Furthermore, the method for constructing the DXN emission concentration sub-model is as follows:

[0251] The DXN adsorption concentration is obtained according to the DXN adsorption concentration sub-model.

[0252] Based on the width forest regression algorithm, a sub-model of DXN emission concentration is constructed according to the DXN adsorption concentration and the preprocessed multimodal data.

[0253] In the DXN emission concentration model building module, the input data includes DXN generation concentration. DXN adsorption concentration Process numerical data The output data is the DXN concentration. Therefore, by combining the above dataset to model the DXN emission stages, it is represented as:

[0254] A DXN emission model is established using the BFR algorithm. This model consists of mapping, enhancement, and incremental layers, as follows: Figure 5 As shown. In each layer of the BFR, the basic unit is the DT model, and the number of DTs can be determined experimentally.

[0255] The details of BFR are as follows:

[0256] Based on the preceding modeling process, the DT model in BFR can be written as:

[0257]

[0258] in,

[0259] in, It is the j-th mapping layer Map A DT model.

[0260] Therefore, the j-th layer in the mapping layer Map The output of each DT unit is represented as follows:

[0261]

[0262] Finally, the output of the mapping layer can be represented as At the same time, by combining D E and mapping layer Z Map The output is used to obtain the input for the enhancement layer. After this combination, the input data is represented as... Subsequently, the j-th layer in the enhancement layer Enh The output of each DT unit is represented as follows:

[0263]

[0264] The output of the enhancement layer is represented as The combined output of the mapping and enhancement layers is represented as Therefore, the static part of the BFR model is as follows:

[0265]

[0266] in, This is the weight vector. The weight calculation process is written as follows, connecting the mapping layer, enhancement layer, and output layer:

[0267]

[0268] in, express The pseudo-inverse matrix, where λ represents the regularization coefficient and I represents the identity matrix.

[0269] When the static part of the BFR model encounters challenges in achieving the desired learning performance, BFR achieves rapid learning by adaptively adding DTs. By adding a DT model to the incremental layer, taking the output of the mapped, augmented, and original data as input, the output of the added DT is represented as follows:

[0270]

[0271] Therefore, the new BFR model is written as follows:

[0272]

[0273] Therefore, the update process of the pseudo-inverse matrix is ​​as follows:

[0274]

[0275] in,

[0276]

[0277] When the model error converges, the incremental process automatically stops increasing. The final DXN emission model based on the BFR algorithm is represented as f TBLS (·)

[0278] Furthermore, the input and output data of the FFFM model are as follows: Figure 6 As shown.

[0279] In the modeling experiment, data before October 19, 2021, was used as the training set, while data after that date constituted the test set.

[0280] Specifically, this embodiment discloses the hyperparameter settings and training results for each model of the FFFM method:

[0281] The hyperparameter settings for the ADFR-clfc model are as follows: minimum number of samples in a leaf node is 3, number of DTs in each forest is 5, number of RFs is 2, number of CRFs is 2, and the selected features are... The threshold is 1×10 -4 .

[0282] The hyperparameters of the CNN model are set as follows: The CNN structure consists of three convolutional layers as input, a fully connected layer as output, and the size and number of the three convolutional kernels are set to [5×5×6]-[5×5×16]-[5×5×16]. The number of iterations is 200, and the learning rate is 0.01.

[0283] The results of the model training process, achieved through the above hyperparameter settings, are as follows: Figure 7,in, Figure 7 (a) is a schematic diagram of the training error curve for process numerical data. Figure 7 (b) is a schematic diagram of the training error curve of flame image data. Figure 7 (c) is a schematic diagram of the training results of the CNN model:

[0284] from Figure 7 From this, we can see that: 1) In the process of numerical data modeling, the training error gradually converges as the number of layers increases, such as... Figure 7 (a) The ADFR-clfc model shown in Figure 1 finally converged, reaching a final layer count of 538; (2) The CNN image burning state recognition model demonstrated a classification accuracy of over 90% during training, validation, and testing, such as... Figure 7 (c) shows that this highlights the effectiveness of the CNN model in learning flame images; (3) demonstrates the effective feature extraction capability of the CNN model; (4) Figure 7 (b) shows the training error curves of the ADFR-clfc model driven by flame image data. The results indicate that the ADFR-clfc model maintains effective convergence during training, even when using high-dimensional data.

[0285] The hyperparameters of the VSG algorithm are set as follows: The VAE network structure consists of two hidden layers for input and two hidden layers for output; the number of neurons per layer is set to 6-512-256-4-256-512-6; the number of VAE iterations is 1500; the number of centers is 3; the number of rings is 8; the percentage of co-training is 0.95; the regularization term for co-training is 0.5; and the number of co-training iterations is 5. Under the above hyperparameter settings of the VSG model, the model training results are as follows: Figure 8 As shown. Figure 8 (a) Schematic diagram of the training error curves of the encoder and decoder. Figure 8 (b) is a schematic diagram showing the distribution of true and false labels.

[0286] Figure 8 (a) shows that the errors of the encoder and decoder networks converge effectively during training, indicating that the encoder and decoder networks exhibit strong representational performance for low-dimensional process numerical data in the adsorption phase. The experimental results demonstrate that the employed VSG method can generate a large number of similar virtual samples. After labeling and evaluating the collaborative strategy, the distribution between the obtained pseudo-labeled values ​​and the ground truth values ​​is as follows: Figure 8 As shown in (b), the distribution of pseudo-labels closely matches that of the ground truth labels, demonstrating the effectiveness of the cooperative strategy. Subsequently, virtual samples were added to the original ultra-small sample to evaluate their impact on the fractional flow reserve (FFR) model. The experimental results are shown in the figure. Selecting 20 virtual samples based on the performance curve maximizes the performance of the FFR model.

[0287] The hyperparameters of the BFR model are set as follows: the minimum number of samples in the leaf nodes is set to 3, and the number of selected features is... The number of DTs in the mapping and enhancement layers is 10, and the regularization term is 0.01. Using the above hyperparameter settings for the BFR model, the model training process results are as follows: Figure 9 As shown.

[0288] like Figure 9 As shown, the learning performance of the static part of BFR is limited. After adding the DT model online, the training error of BFR gradually decreases. Therefore, when the number of incremental DTs exceeds 30, the BFR model eventually converges to the emission stage.

[0289] The beneficial effects of this invention are as follows:

[0290] This invention employs different modeling strategies based on the characteristics of different stages. Specifically, ADFR clfc and CNN algorithms are used in the DXN generation stage, VSG and FFR algorithms are used in the DXN adsorption stage, and BFR algorithm is used in the DXN emission stage. The proposed FFFM model was validated using DXN concentration data from different stages at a MSWI plant in Beijing, reducing the cost of DXN concentration monitoring and improving the effectiveness of DXN emission reduction.

[0291] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0292] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for detecting DXN concentration during MSWI, characterized in that, include: Acquire multimodal data; The multimodal data is preprocessed to obtain preprocessed multimodal data; Construct a full-process model of DXN concentration; The preprocessed multimodal data is input into the DXN concentration full-process model to obtain DXN emission concentration, DXN generation concentration and DXN adsorption concentration; Methods for constructing a full-process model of DXN concentration include: Construct sub-models for DXN generation concentration, DXN emission concentration, and DXN adsorption concentration; The DXN generation concentration sub-model, DXN emission concentration sub-model and DXN adsorption concentration sub-model are cascaded to obtain the full process model of DXN concentration. The multimodal data includes: The process numerical data, flame video data, and recording operation data, wherein the flame video data includes: left flame video data and right flame video data; The method for constructing the DXN-generated concentration sub-model is as follows: The construction process numerical data-driven unit and the flame image data-driven unit; A DXN concentration generation sub-model is constructed based on the process numerical data driving unit and the flame image data driving model. The method for constructing the process numerical data driven unit is as follows: Based on the preprocessed multimodal dataset, the process numerical data driving unit is constructed using a cross-layer fully connected adaptive deep forest regression algorithm. The numerical data-driven unit described above consists of three parts: a layered forest model, cross-layer fully connected components, and layer adaptive components. The method for constructing the flame image data driving unit is as follows: A flame image data driving unit is constructed based on CNN and process numerical data driving unit; The method for constructing the DXN adsorption concentration sub-model is as follows: Virtual data is obtained by generating virtual samples from the DXN-generated concentration and the preprocessed multimodal data; Based on the fuzzy forest regression algorithm, a DXN adsorption concentration sub-model is constructed according to the virtual data; The method for constructing the DXN emission concentration sub-model is as follows: The DXN adsorption concentration is obtained according to the DXN adsorption concentration sub-model. Based on the width forest regression algorithm, a sub-model of DXN emission concentration is constructed according to the DXN adsorption concentration and the preprocessed multimodal data.

2. The method for detecting DXN concentration during MSWI according to claim 1, characterized in that, The recorded operation data includes: Feed rate, combustion efficiency, average load, steam production per ton of MSW, and calorific value of MSW.

3. The method for detecting DXN concentration during MSWI according to claim 2, characterized in that, The multimodal data is preprocessed to obtain preprocessed multimodal data, including: The time of the process numerical data, the time of the flame video data, and the time of the recording operation data are synchronized with the time of DXN sampling to obtain the process numerical data, the flame video data, and the recording operation data corresponding to the time. Pre-calculate the recording operation data corresponding to the time to obtain a valid recording operation dataset; The process numerical data corresponding to the time is subjected to a first average data processing to obtain the average data; The average data is sliced ​​to obtain numerical data of DXN generation concentration process, DXN adsorption concentration process, and DXN emission concentration process. Extract second-level image information from the flame video data corresponding to the time; The second-level image information is subjected to a second averaging process to obtain second-level data; The second-level data is processed into grayscale to obtain grayscale image information; The grayscale image information is subjected to a third averaging process to obtain the fused video information; Preprocessed multimodal data is obtained based on the valid recording operation dataset, DXN generated concentration process numerical data, DXN adsorption concentration process numerical data, DXN emission concentration process numerical data, and fused video information.

4. The method for detecting DXN concentration during MSWI according to claim 3, characterized in that, The formula for pre-calculating the recording operation data corresponding to the time includes: ； ； ； ； ； in, , , , and These are feed rate, incineration efficiency, average load, steam production per ton of municipal solid waste, and MSW calorific value. Indicates the rated capacity of the MSWI process; This represents the actual volume processed in one day. Indicates steam flow rate; Indicates the enthalpy of vapor; Indicates the enthalpy of water; This refers to the boiler's thermal efficiency.

5. The method for detecting DXN concentration during MSWI according to claim 4, characterized in that, The calculation formula for the first averaging process is as follows: ； in, For the first One process variable, The dimension of the process numerical data. For time scale.

Images