Dioxin generation concentration detection method, computer equipment, medium and program product
By constructing a self-supervised interval detection model for dioxin formation concentration, and combining multimodal data augmentation and heterogeneous integrated interval detection, the problem of dioxin formation concentration detection in MSWI process was solved, achieving high-precision detection and reducing the consumption of flue gas purification materials.
Patent Information
- Application Number
- CN202511121734.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-12
- Publication Date
- 2025-11-28
AI Technical Summary
Existing technologies are insufficient to fundamentally reduce dioxin formation concentration during MSWI, and flue gas purification materials are consumed in large quantities. Furthermore, there is a lack of research on the application of multimodal data and multilevel correlation in the detection of dioxin formation concentration.
By constructing a self-supervised interval detection model for dioxin generation concentration, combining multimodal data augmentation modeling and heterogeneous integrated interval detection, and utilizing variational encoder, linear regression decision tree and incremental stochastic weighted neural network, the dioxin generation concentration is detected.
It achieves high-precision detection of dioxin formation concentration, enhances the reliability of the model, and reduces the consumption of flue gas purification materials.
Smart Images
Figure CN121034471A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of environmental protection, and in particular relates to a method for detecting dioxin formation concentration, computer equipment, media, and program products. Background Technology
[0002] Currently, urban solid waste (MSW) continues to increase at an annual growth rate of 8-10%, leading to an increasing number of cities facing the phenomenon of "garbage surrounding the city." Resource-based disposal of MSW is one of the core issues that urgently needs to be addressed in the development of environmentally friendly cities. Urban solid waste incineration (MSWI) is an important link in the process of urban renewable energy recycling and has become an effective way to solve the "garbage surrounding the city" phenomenon.
[0003] MSWI (Maintained Solid Waste Management) is the most scientifically sound method for urban solid waste treatment, but the process generates waste gas, wastewater, and solid waste, which often leads to MSWI companies being listed on the national pollution source emission list, causing them to be plagued by the "NIMBY (Not In My Backyard) effect" for a long time. In particular, the process releases dioxins, a trace organic pollutant known as the "poison of the century," which is non-degradable, has a cumulative effect, and threatens human survival.
[0004] From an operational mechanism perspective, the solid waste incineration stage of the MSWI process generates NO, among other things. x The flue gas purification process primarily focuses on the sources of easily detectable pollutants such as nitrogen oxides, CO, HCl, and SO2, as well as difficult-to-detect pollutants like dioxins. The subsequent flue gas purification stage mainly aims to transfer these pollutants to fly ash through adsorption / catalysis, thereby reducing their atmospheric emission concentration. The mechanisms of dioxin formation, adsorption, and emission during MSWI remain unclear. Existing dioxin detection methods mainly target offline direct detection, online indirect detection, and online soft measurement of dioxin concentrations at flue gas emissions. However, research on how to construct a dioxin formation concentration range detection model using multimodal data and multi-level correlated unlabeled samples, and then detect dioxin formation concentrations, from the perspectives of reducing pollutant formation concentrations at their source and reducing the economic consumption of flue gas purification materials, is rarely reported. Summary of the Invention
[0005] To address the aforementioned technical problems, this invention provides a method for detecting dioxin formation concentration, comprising the following steps:
[0006] S1. Obtain multimodal data samples for enhanced modeling based on dioxin formation concentration;
[0007] S2. Based on multimodal data augmentation modeling samples and unlabeled multimodal data feature samples within a set time period before and after the original ground truth samples, construct and optimize a self-supervised interval detection model for dioxin generation concentration;
[0008] S3, build a dioxin generation concentration isosmotic integrated interval detection model, and detect the dioxin generation concentration.
[0009] Further, step S1 comprises:
[0010] S1.1, synchronize the multi-modal data and the dioxin generation concentration data in the sampling time,
[0011] wherein,
[0012] The multi-modal data comprises the record operation data and the process data, the flame image data in the dioxin generation stage;
[0013] S1.2, perform feature extraction on the multi-modal data;
[0014] S1.3, based on the feature extraction result of the multi-modal data, obtain a virtual sample input / output pair;
[0015] S1.4, combine the virtual sample input / output pair and the multi-modal data to obtain a multi-modal data enhanced modeling sample.
[0016] Further, the dioxin generation concentration data is hourly data, the record operation data is daily data, and the process data and the flame image data are second-level data.
[0017] Further, step S1.2 comprises:
[0018] For the record operation data, based on the actual operation experience and the combustion theory, the record operation data features are obtained by calculation;
[0019] For the process data, the process numerical features are obtained by averaging, based on process flow splitting, etc.
[0020] For the flame image data, an improved denoising diffusion probability model is used, the flame image data is mapped to the latent space for diffusion using a pre-trained encoder, and then a decoder is used for reconstruction to remove the noise in the flame image data; the physical features and the depth features of the flame image data are extracted through frame extraction and time scale average processing.
[0021] Further, step S1.3 comprises:
[0022] S1.3.1, combine the record operation data features, the process numerical features, and the physical features and the depth features of the flame image data to obtain multi-modal data combination features;
[0023] S1.3.2, map the multi-modal data combination features to a two-dimensional vector in the latent space using a variational encoder network, and obtain the latent space virtual sample input in the two-dimensional coordinate system through a virtual sample generation algorithm;
[0024] S1.3.3, input the latent space virtual sample into the dimensionality mapping to map it back to the original feature space using variational decoding, obtain the virtual sample output by training the linear tree model, labeling and evaluation, and then obtain the virtual sample input / output pair.
[0025] Further, step S2 comprises the following steps:
[0026] S2.1, based on the approximate linear dependence, evaluate the quality of the unlabeled sample multi-modal data feature sample, give different trust values to the unlabeled multi-modal data feature samples in different time ranges, obtain multi-layer unlabeled multi-modal data feature samples with differential trust mechanism, select multi-layer unlabeled multi-modal data feature samples with higher trust values,
[0027] After preprocessing the multi-layer unlabeled multi-modal data feature samples, calculate their similarity features with the labeled original true value samples based on domain knowledge as initial trust values, and then calculate the uncertainty of the multiple process input feature samples based on information entropy, combine the information entropy and domain knowledge to update the initial trust values, and obtain multi-layer unlabeled samples with different trust values;
[0028] S2.2, in the pre-training stage, the multi-layer unlabeled samples with higher trust values masked are used as the input of the encoder-decoder model for self-supervised pre-training; in the fine-tuning stage, the weights of the encoder are preserved, the original true value sample with dioxin true value is input into the encoder to extract deep features, and the multilayer perceptron network is fine-tuned through supervised learning, realizing the construction of the dioxin generation concentration self-supervised interval detection model;
[0029] S2.3, combine the multi-modal data enhanced modeling samples, and based on the parallel differential evolution algorithm, the parameters of the dioxin generation concentration self-supervised interval detection model and the number of virtual samples are multi-objective optimized, and the optimized dioxin generation concentration self-supervised interval detection model is obtained.
[0030] Further, step S3 comprises the following steps:
[0031] S3.1, construct a numerical simulation model for dioxin generation concentration, and obtain virtual simulation mechanism data;
[0032] S3.2, construct a hybrid virtual-real data driven multi-input and multi-output linear regression decision tree model with CO2, CO and O2 as output, and solve it using gradient descent method;
[0033] S3.3, based on process data, constructing a multi-input single-output linear regression decision tree model with CO2, CO and O2 as inputs and dioxin generation concentration as output, taking unmarked dioxin combustion state characterization variables as input, obtaining pseudo-label data of dioxin by using a tree-based semi-supervised learning process, and then obtaining a model based on tree structure transfer learning by using the multi-input single-output linear regression decision tree model, and obtaining a mapping tree model based on semi-supervised transfer learning by structure growth learning of the model, that is, a combustion state characterization mechanism model for dioxin interval detection;
[0034] S3.4, using an incremental random weighted neural network with a constraint learning mechanism, fusing the output of the dioxin generation concentration self-supervised interval detection model driven by multi-modal data and the combustion state characterization mechanism model for dioxin generation concentration detection, to obtain a dioxin generation concentration heterogeneous integrated interval detection model;
[0035] S3.5, detecting the dioxin generation concentration by the dioxin generation concentration heterogeneous integrated interval detection model.
[0036] The application also provides a computer device comprising a memory, a processor and a computer program stored in the memory, the computer program being executed by the processor to implement the dioxin generation concentration detection method.
[0037] The application also provides a computer readable storage medium storing a computer program, the computer program being executed by a processor to implement the dioxin generation concentration detection method.
[0038] The application also provides a computer program product comprising a computer program, the computer program being executed by a processor to implement the dioxin generation concentration detection method.
[0039] Compared with the prior art, the dioxin generation concentration detection method provided by the application can obtain the concentration of the detected pollutant, i.e., dioxin, with high accuracy, and enhance the credibility of each model through interval value. BRIEF DESCRIPTION OF DRAWINGS
[0040] Figure 1 A flowchart of the dioxin generation concentration detection method according to an embodiment of the application is shown. DETAILED DESCRIPTION
[0041] The specific embodiments of the present application are described in detail below, it should be understood that the embodiments of the present application are not limited to the examples shown in the drawings, and the protection scope of the present application is not limited by the specific embodiments. The "first", "second" and similar words used in the present application do not represent any order, quantity or importance, but are only used to distinguish different components. Similarly, "one", "a" or "the" and similar words do not represent a quantity limit, but represent the existence of at least one. Unless otherwise explicitly stated, throughout the specification and claims, the term "comprise" or its variants such as "include" or "comprising" will be understood to include the stated elements or components, and not exclude other elements or components. The words "connected" or "connected" and the like are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to represent relative positional relationships, which may change accordingly when the absolute position of the described object changes.
[0042] Unless otherwise defined, all technical and scientific terms used in the present application have the same meaning as commonly understood by a person of ordinary skill in the art. In addition, the meanings of technical and scientific terms used in the present application should be interpreted as consistent with the meanings of corresponding terms defined in common technical manuals, and should not be interpreted as having idealized or excessively formal meanings, unless the present application is explicitly defined.
[0043] Figure 1 A flow chart of the dioxin generation concentration detection method of the embodiment of the present application is shown, see Figure 1 The dioxin generation concentration detection method comprises the following steps:
[0044] S1, obtain multi-modal data enhancement modeling samples for dioxin generation concentration.
[0045] Specifically, step S1 comprises:
[0046] S1.1, synchronize multi-modal data and dioxin generation concentration data sampling time.
[0047] The multi-modal data includes record operation data and process data of the dioxin generation stage, flame image data.
[0048] Specifically, the dioxin generation concentration data is hourly data, the record operation data is daily data, and the process data and the flame image data are second-level data.
[0049] Further, the record operation data includes feed amount, incineration efficiency, average load, ton steam production, and solid waste heat value.
[0050] S1.2, feature extraction is performed on the multi-modal data.
[0051] Specifically, step S1.2 includes:
[0052] For the record operation data, based on the actual operation experience and combustion theory, the record operation data features are obtained by calculation;
[0053] For the process data, process numerical features are obtained by averaging, process stream splitting and other processing;
[0054] For the flame image data, an improved denoising diffusion probability model is used, a pre-trained encoder is used to map the flame image data to a latent space for diffusion, and then a decoder is used for reconstruction to remove noise in the flame image data; physical features and depth features of the flame image data are extracted through frame extraction and time scale average processing.
[0055] S1.3, based on the feature extraction results of the multi-modal data, a virtual sample input / output pair is obtained.
[0056] Specifically, step S1.3 includes:
[0057] S1.3.1, the record operation data features, process numerical features, and physical features and depth features of the flame image data are combined to obtain multi-modal data combination features;
[0058] S1.3.2, the multi-modal data combination features are mapped to a two-dimensional vector in the latent space using a variational encoder network, and a latent space virtual sample input is obtained in the two-dimensional coordinate system through a virtual sample generation algorithm.
[0059] S1.3.3, the latent space virtual sample input is mapped back to the original feature space by dimensionality reduction using a variational decoder, and the virtual sample output is obtained by training a linear tree model, labeling and evaluation, and then the virtual sample input / output pair is obtained.
[0060] S1.4, combining the virtual sample input / output pair and the original true value sample, a multi-modal data augmented modeling sample is obtained.
[0061] Wherein, the original true sample is an input / output pair composed of multi-modal data combination features (including record operation data features, process numerical features, and physical features and depth features of flame image data) and dioxin true value.
[0062] S2, based on the multi-modal data augmented modeling sample and the unmarked multi-modal data feature sample within a certain time period before and after the original true value sample, a dioxin generation concentration self-supervised interval detection model is constructed and optimized.
[0063] Specifically, step S2 includes the following steps:
[0064] S2.1, based on the approximate linear dependence evaluation of unlabeled multi-modal data feature samples, different trust values are given to unlabeled multi-modal data feature samples in different time ranges, and multi-layer unlabeled multi-modal data feature samples with differentiated trust mechanism are obtained, and multi-layer unlabeled multi-modal data feature samples with higher trust values are selected;
[0065] After preprocessing the multi-layer unlabeled multi-modal data feature samples, a plurality of groups of input feature samples are obtained, the similarity features of the input feature samples and the labeled original true value samples are calculated based on the domain knowledge as the initial trust values, and the uncertainty of the plurality of groups of input feature samples is calculated based on the information entropy, and the initial trust values are updated in combination with the information entropy and the domain knowledge, to obtain multi-layer unlabeled samples with different trust values;
[0066] S2.2, in the pre-training stage, the multi-layer unlabeled samples with higher trust values and masks are used as the input of the encoder-decoder model for self-supervised pre-training; in the fine-tuning stage, the weights of the encoder are reserved, the original true value sample with dioxin true value is input into the encoder to extract deep features, and the multilayer perceptron network is fine-tuned through supervised learning, so as to realize the construction of the dioxin generation concentration self-supervised interval detection model;
[0067] S2.3, combined with multi-modal data enhanced modeling samples, the parameters of the dioxin generation concentration self-supervised interval detection model and the number of virtual samples are multi-objective optimized based on the parallel differential evolution algorithm, and the optimized dioxin generation concentration self-supervised interval detection model is obtained.
[0068] S3, a dioxin generation concentration heterogeneous integrated interval detection model is constructed, and the dioxin generation concentration is detected.
[0069] Specifically, step S3 includes the following steps:
[0070] S3.1, a numerical simulation model for dioxin generation concentration is constructed, and virtual simulation mechanism data is obtained;
[0071] S3.2, a hybrid virtual-real data driven multi-input multi-output linear regression decision tree (LRDT) model with CO2, CO and O2 as output is constructed, and gradient descent method is used for solving;
[0072] S3.3 Based on process data, a multi-input single-output LRDT tree model is constructed with CO2, CO, and O2 as inputs and dioxin formation concentration as output. The combustion state characterization variable of unlabeled dioxins is used as input, and a tree-based semi-supervised learning process is used to obtain pseudo-label data of dioxins. Then, the multi-input single-output LRDT tree model is used to obtain a model based on tree structure transfer learning. Through structural growth learning of this model, a mapping tree model based on semi-supervised transfer learning is obtained, which is the combustion state characterization mechanism model for dioxin interval detection.
[0073] S3.4. An incremental stochastic weighted neural network with a constrained learning mechanism is adopted to integrate the outputs of a multimodal data-driven self-supervised interval detection model for dioxin formation concentration and a combustion state characterization mechanism model for dioxin formation concentration detection, thereby obtaining a heterogeneous integrated interval detection model for dioxin formation concentration.
[0074] S3.5. Detect the dioxin formation concentration using a heterogeneous integrated range detection model for dioxin formation concentration.
[0075] The dioxin formation concentration detection method provided by this invention first acquires multimodal data augmentation modeling samples for dioxin formation concentration, then constructs and optimizes a self-supervised interval detection model for dioxin formation concentration based on the multimodal data augmentation modeling samples, and finally constructs a heterogeneous integrated interval detection model for dioxin formation concentration to detect the dioxin formation concentration. This dioxin formation concentration detection method can obtain the concentration of the detected pollutant, i.e., dioxins, with high accuracy, and enhances the reliability of each model through interval values.
[0076] The present invention also provides a computer device, the computer device including a memory, a processor and a computer program stored in the memory, wherein the computer program, when executed by the processor, implements the aforementioned method for detecting dioxin formation concentration.
[0077] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for detecting dioxin formation concentration.
[0078] The present invention also provides a computer program product, the computer program product comprising a computer program, which, when executed by a processor, implements the aforementioned method for detecting dioxin formation concentration.
[0079] The foregoing description of specific exemplary embodiments of the application has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the application to the precise forms disclosed, and obviously many modifications and variations are possible in light of the above teaching. It is intended that the scope of the application be limited not with this detailed description, but rather by the claims appended hereto.
Claims
1. A method for detecting dioxin formation concentration, characterized in that, Includes the following steps: S1. Obtain multimodal data samples for enhanced modeling based on dioxin formation concentration; S2. Based on multimodal data augmentation modeling samples and unlabeled multimodal data feature samples within a set time period before and after the original ground truth samples, construct and optimize a self-supervised interval detection model for dioxin generation concentration; S3. Construct a heterogeneous integrated range detection model for dioxin formation concentration and detect dioxin formation concentration.
2. The method for detecting dioxin formation concentration according to claim 1, characterized in that, Step S1 includes: S1.1 Synchronize the sampling time of multimodal data and dioxin formation concentration data. in, Multimodal data includes recorded operational data, process data of the dioxin formation stage, and flame image data; S1.2, Feature extraction of multimodal data; S1.
3. Based on the feature extraction results of multimodal data, obtain virtual sample input / output pairs; S1.4 Combine virtual sample input / output pairs and multimodal data to obtain multimodal data augmentation modeling samples.
3. The method for detecting dioxin formation concentration according to claim 2, characterized in that, Its features are, Dioxin formation concentration data is available in hourly increments, operational data is available in daily increments, and process data and flame image data are available in second increments.
4. The method for detecting dioxin formation concentration according to claim 2, characterized in that, Its features are, Step S1.2 includes: Based on practical experience and combustion theory, the characteristics of the recorded operation data are calculated. Process data is processed through averaging and process flow-based decomposition to obtain process numerical characteristics; For flame image data, an improved denoising diffusion probability model is adopted. The pre-trained encoder maps the flame image data to the latent space for diffusion and then uses the decoder to reconstruct it to remove noise from the flame image data. Then, the physical features and depth features of the flame image data are extracted through frame extraction and time-scale averaging.
5. The method for detecting dioxin formation concentration according to claim 2, characterized in that, Its features are, Step S1.3 includes: S1.3.1 Combine the recorded operation data features, process numerical features, and physical and depth features of flame image data to obtain multimodal data combination features; S1.3.
2. A variational encoder network is used to map the combined features of multimodal data into a two-dimensional vector in the latent space, and a virtual sample input in the latent space is obtained through a virtual sample generation algorithm in the two-dimensional coordinate system. S1.3.
3. Variational decoding is used to map the virtual sample input in the latent space back to the original feature space by upscaling. The virtual sample output is obtained by training a linear tree model, labeling and evaluating, and then the virtual sample input / output pair is obtained.
6. The method for detecting dioxin formation concentration according to claim 1, characterized in that, Its features are, Step S2 includes the following steps: S2.
1. Evaluate the quality of unlabeled multimodal data feature samples based on approximate linear dependence. Assign different trust values to unlabeled multimodal data feature samples within different time ranges to obtain multilayered unlabeled multimodal data feature samples with differentiated trust mechanisms. Select multilayered unlabeled multimodal data feature samples with higher trust values. After preprocessing the feature samples of multi-layer unlabeled multimodal data, multiple sets of input feature samples are obtained. Based on domain knowledge, the similarity features between these samples and the labeled original ground truth samples are calculated as initial trust values. Then, based on information entropy, the uncertainty of multiple sets of process input feature samples is calculated. The initial trust values are updated by combining information entropy and domain knowledge to obtain multi-layer unlabeled samples with different trust values. S2.2 In the pre-training stage, masked, high-confidence multilayer unlabeled samples are used as input to the encoder-decoder model for self-supervised pre-training. In the fine-tuning stage, the encoder weights are retained, and the original ground truth samples with dioxin ground truth are input to the encoder to extract deep features. The multilayer perceptron network is fine-tuned through supervised learning to realize the construction of a self-supervised interval detection model for dioxin generation concentration. S2.
3. Combine multimodal data to enhance modeling samples. Based on the parallel differential evolution algorithm, perform multi-objective optimization on the parameters and the number of virtual samples of the self-supervised interval detection model of dioxin generation concentration to obtain the optimized self-supervised interval detection model of dioxin generation concentration.
7. The method for detecting dioxin formation concentration according to claim 1, characterized in that, Its features are, Step S3 includes the following steps: S3.1 Construct a numerical simulation model for dioxin formation concentration and obtain virtual simulation mechanism data; S3.2 Construct an interpretable multi-input multi-output linear regression decision tree model driven by hybrid virtual and real data, with CO2, CO, and O2 as outputs, and solve it using the gradient descent method; S3.3 Based on process data, a multi-input single-output linear regression decision tree model is constructed with CO2, CO, and O2 as inputs and dioxin formation concentration as output. The combustion state characterization variable of unlabeled dioxins is used as input, and a tree-based semi-supervised learning process is used to obtain pseudo-label data of dioxins. Then, the multi-input single-output linear regression decision tree model is used to obtain a model based on tree structure transfer learning. Through structural growth learning of this model, a mapping tree model based on semi-supervised transfer learning is obtained, which is the combustion state characterization mechanism model for dioxin interval detection. S3.
4. An incremental stochastic weighted neural network with a constrained learning mechanism is adopted to integrate the outputs of a multimodal data-driven self-supervised interval detection model for dioxin formation concentration and a combustion state characterization mechanism model for dioxin formation concentration detection, thereby obtaining a heterogeneous integrated interval detection model for dioxin formation concentration. S3.
5. Detect the dioxin formation concentration using a heterogeneous integrated range detection model for dioxin formation concentration.
8. A computer device, characterized in that, The computer device includes a memory, a processor, and a computer program stored in the memory, which, when executed by the processor, implements the dioxin generation concentration detection method according to any one of claims 1 to 7.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the dioxin generation concentration detection method according to any one of claims 1 to 7.
10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the dioxin generation concentration detection method according to any one of claims 1 to 7.