MSWI process flue gas oxygen content prediction method and system based on multi-modal depth characteristics

Abstract

The invention provides an MSWI process flue gas oxygen content prediction method based on multi-modal depth features, and the method comprises the steps: carrying out the preprocessing of a flame image of an MSWI process, and obtaining a preprocessed image; performing multi-level feature extraction on the preprocessed image through a first neural network composed of eight layers of structures to obtain flame depth features; performing feature selection on the process data of the MSWI process through a Pearson correlation coefficient to obtain a reduction feature set; performing depth feature extraction on the reduction feature set through a second neural network composed of two layers of structures to obtain process depth features; and carrying out feature fusion on the flame depth feature and the process depth feature, and predicting the oxygen content of the flue gas through the fused multi-modal depth feature. According to the method, feature depth extraction and fusion operation are carried out on flame images and process data, and correlation coefficients and a multi-layer perceptron model are combined, so that prediction of the FGOC and construction of the model are realized.

Description

Technical Field

[0001] This invention relates to the field of urban solid waste incineration technology, and in particular to a method and system for predicting the oxygen content of flue gas in MSWI processes based on multimodal deep features. Background Technology

[0002] With rapid economic development and rising living standards, the generation of municipal solid waste (MSW) has increased dramatically. How to effectively reduce MSW volume, achieve harmless treatment, and utilize it for resource recovery, thereby promoting sustainable ecological development, has become a critical issue urgently needing to be addressed worldwide. MSW incineration (MSWI) power generation technology has replaced traditional landfill and composting methods and has become the preferred method for MSW disposal in developing countries. Developed countries employ Automatic Combustion Control (ACC) systems to ensure the safe and stable operation of the MSWI process. However, due to differences between developing and developed countries in MSW classification, management measures, and operation and maintenance levels, the aforementioned ACC system is not suitable for developing countries.

[0003] Currently, MSWI power plants primarily rely on the experience of domain experts to ensure stable equipment operation. This model, based on multimodal data such as process variables, flame video, and on-site inspection voice feedback, adjusts the output of manipulated variables by predicting key controlled variables such as flue gas oxygen content (FGOC), thereby achieving stable MSWI operation. However, the randomness and subjectivity of expert experience are detrimental to the long-term stable operation of MSWI power plants. The first problem to be solved in achieving intelligent "replication and improvement" of expert predictions is the prediction of controlled variables. Existing technologies lack methods and models for FGOC prediction in MSWI processes. Therefore, it is essential to design a method and system for predicting flue gas oxygen content in MSWI processes based on multimodal deep feature analysis. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting the oxygen content of flue gas in MSWI processes based on multimodal deep features. By performing feature depth extraction and fusion operations on flame images and process data, and combining correlation coefficients and network models, the prediction of FGOC and the construction of the model can be achieved.

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

[0006] A method for predicting oxygen content in flue gas during MSWI processes based on multimodal deep features includes the following steps:

[0007] The flame image of the MSWI process is preprocessed to obtain a preprocessed image; the preprocessing includes: image scaling, center cropping and normalization.

[0008] The flame depth features are obtained by performing multi-level feature extraction on the preprocessed image through a first neural network consisting of 8 layers.

[0009] Feature selection was performed on the process data of the MSWI process using the Pearson correlation coefficient to obtain a reduced feature set.

[0010] Deep features of the process are obtained by extracting deep features from the reduced feature set through a second neural network consisting of two layers.

[0011] Flame depth features and process depth features are fused, and the oxygen content in flue gas is predicted by the fused multimodal depth features.

[0012] Optionally, the first neural network includes: a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a third pooling layer, and a fourth pooling layer;

[0013] The first convolutional layer performs multi-channel convolution operations on the preprocessed image using the ReLU activation function;

[0014] The first pooling layer performs max pooling on the output of the first convolutional layer using a maximum value function.

[0015] The second convolutional layer performs a convolution operation on the output of the first pooling layer using the second convolutional kernel matrix and the second convolutional bias vector.

[0016] The second pooling layer performs pooling operations on the output of the second convolutional layer using the second pooling kernel matrix.

[0017] The third convolutional layer performs a convolution operation on the output of the second pooling layer using the third convolutional kernel matrix and the third convolutional bias vector.

[0018] The fourth convolutional layer performs a convolution operation on the output of the third convolutional layer using the fourth convolutional kernel matrix and the fourth convolutional bias vector;

[0019] The third pooling layer performs average pooling on the output of the fourth convolutional layer using an averaging function.

[0020] The fourth pooling layer performs adaptive pooling on the output of the third pooling layer using an adaptive function to obtain flame depth features.

[0021] Optionally, feature selection is performed on the process data of the MSWI process using the Pearson correlation coefficient to obtain a reduced feature set, including:

[0022] Through formula Calculate the Pearson correlation coefficient of the process data. ;in, For the p-th feature in the process data, for The mean, Oxygen content in flue gas for The mean;

[0023] When the Pearson correlation coefficient is greater than the preset empirical threshold, the process data is marked as a candidate feature, and the candidate features are integrated into a reduced feature set.

[0024] Optionally, the second neural network consists of a first fully connected layer and a second fully connected layer;

[0025] The first fully connected layer performs matrix multiplication on the reduced feature set using the first fully connected weight matrix and the first fully connected bias vector.

[0026] The second fully connected layer performs a fully connected operation on the output of the first fully connected layer using the second fully connected weight matrix and the second fully connected bias vector to obtain the process depth features.

[0027] Optionally, flame depth features and process depth features are fused, and the oxygen content in flue gas is predicted using the fused multimodal depth features, including:

[0028] The flame depth feature after flattening is spliced ​​with the process depth feature to obtain the spliced ​​feature;

[0029] Fully connected dimensionality reduction is performed on the spliced ​​features to obtain multimodal deep features.

[0030] A multimodal deep feature-based MSWI process flue gas oxygen content prediction system includes:

[0031] The flame image preprocessing module is used to preprocess the flame images from the MSWI process to obtain preprocessed images. The preprocessing includes image scaling, center cropping, and normalization.

[0032] The flame image depth feature extraction module is used to perform multi-level feature extraction on the preprocessed image through a first neural network composed of 8 layers to obtain flame depth features.

[0033] The process data feature selection module is used to select features from the process data of the MSWI process using the Pearson correlation coefficient to obtain a reduced feature set;

[0034] The process data deep feature extraction module is used to extract deep features from the reduced feature set through a second neural network composed of two layers to obtain process deep features.

[0035] The flue gas oxygen content prediction module is used to fuse flame depth features and process depth features, and predict the flue gas oxygen content through the fused multimodal depth features.

[0036] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The present invention provides a method for predicting flue gas oxygen content based on multimodal depth features in MSWI processes. This method includes: preprocessing flame images of MSWI processes to obtain preprocessed images; the preprocessing includes: image scaling, center cropping, and normalization; performing multi-level feature extraction on the preprocessed images using a first neural network composed of 8 layers to obtain flame depth features; performing feature selection on the process data of the MSWI process using Pearson correlation coefficients to obtain a reduced feature set; performing depth feature extraction on the reduced feature set using a second neural network composed of 2 layers to obtain process depth features; fusing the flame depth features and process depth features, and predicting flue gas oxygen content using the fused multimodal depth features. This method, by performing feature depth extraction and fusion operations on flame images and process data, combined with correlation coefficients and a multilayer perceptron model, achieves FGOC prediction and model construction. Attached Figure Description

[0037] 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.

[0038] Figure 1 This is a flowchart of the MSWI process flue gas oxygen content prediction method of the present invention;

[0039] Figure 2 This is a schematic diagram of process data and PCC values ​​of flue gas oxygen content in an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of the activation value distribution of the process data feature dimension in an embodiment of the present invention;

[0041] Figure 4 This is a schematic diagram of the experimental results comparing the prediction of oxygen content in flue gas according to an embodiment of the present invention. Detailed Implementation

[0042] 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.

[0043] 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.

[0044] like Figure 1 As shown, this invention provides a method for predicting the oxygen content in flue gas during MSWI processes based on multimodal deep features, comprising the following steps:

[0045] Step 100: Preprocess the flame image of the MSWI process to obtain a preprocessed image;

[0046] Specifically, first, the flame image is resized and cropped from the center; then, the image is converted to a tensor and normalized, as shown in the expression:

[0047] ;

[0048] in, This represents image scaling and cropping operations. This represents image tensor transformation and normalization processing. This represents the pre-processed flame image, i.e., the pre-processed image.

[0049] Step 200: Multi-level feature extraction is performed on the preprocessed image using a first neural network consisting of 8 layers to obtain flame depth features;

[0050] Specifically, the first neural network includes: a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a third pooling layer, and a fourth pooling layer; the process of multi-level feature extraction through this neural network is as follows:

[0051] The first convolutional layer performs multi-channel convolution operations on the preprocessed image using the ReLU activation function, expressed as follows:

[0052] ;

[0053] in, This represents the output feature map of the first convolutional layer. Represents multi-channel convolution operation. Represents the ReLU activation function. and These represent the kernel matrix and bias vector of the first convolutional layer, respectively.

[0054] The first pooling layer performs max pooling on the output of the first convolutional layer using a max function, expressed as:

[0055] ;

[0056] in, This represents the output feature map of the first pooling layer. The kernel matrix represents the first pooling layer. This represents the maximum value function.

[0057] The second convolutional layer performs a convolution operation on the output of the first pooling layer using the second convolutional kernel matrix and the second convolutional bias vector. The expression is as follows:

[0058] ;

[0059] in, This represents the output feature map of the second convolutional layer. and These represent the kernel matrix and bias vector of the second convolutional layer, respectively.

[0060] The second pooling layer performs a pooling operation on the output of the second convolutional layer using the second pooling kernel matrix, expressed as follows:

[0061] ;

[0062] in, This represents the output feature map of the second pooling layer. The kernel matrix represents the second pooling layer.

[0063] The third convolutional layer performs a convolution operation on the output of the second pooling layer using the third convolutional kernel matrix and the third convolutional bias vector, as expressed in the following expression:

[0064] ;

[0065] in, This represents the output feature map of the third convolutional layer. and These represent the kernel matrix and bias vector of the third convolutional layer, respectively.

[0066] The fourth convolutional layer performs a convolution operation on the output of the third convolutional layer using the fourth convolutional kernel matrix and the fourth convolutional bias vector, as expressed in the following expression:

[0067] ;

[0068] in, This represents the output feature map of the fourth convolutional layer. and These represent the kernel matrix and bias vector of the fourth convolutional layer, respectively.

[0069] The third pooling layer performs average pooling on the output of the fourth convolutional layer using an averaging function, expressed as follows:

[0070] ;

[0071] in, This represents the output feature map of the third pooling layer. The kernel matrix represents the third pooling layer. It represents the average value function.

[0072] The fourth pooling layer performs adaptive pooling on the output of the third pooling layer using an adaptive function, expressed as follows:

[0073] ;

[0074] in, This represents the output feature map of the fourth pooling layer. The kernel matrix representing the fourth pooling layer. This represents the adaptive function. Finally, the output of the fourth pooling layer is used as the final flame image depth feature. , is represented as: .

[0075] Step 300: Perform feature selection on the process data of the MSWI process using the Pearson correlation coefficient to obtain a reduced feature set;

[0076] Specifically, the correlation between each feature in the process data and the FGOC varies. This embodiment calculates the Pearson correlation coefficient (PCC) value between each individual feature and the FGOC to assess feature importance. The process data contains... The first feature, of which the second feature is... Features PCC value The calculation formula is:

[0077] ;

[0078] in, for The mean, Oxygen content in flue gas for The mean; then according to The comparison results with the preset threshold select features with high correlation to FGOC as candidate features, and integrate the candidate features into a reduced feature set, expressed as:

[0079] ;

[0080] in, This represents a threshold set based on experience. represent The flag indicates whether the feature was selected; a value of 1 indicates that the feature was selected as a candidate feature, and a value of 0 indicates that it was not selected as a candidate feature. The number of selected sub-features is... And denote the reduced feature set as .

[0081] Step 400: Deep feature extraction is performed on the reduced feature set using a second neural network consisting of two layers to obtain process deep features;

[0082] Specifically, the second neural network consists of a first fully connected layer and a second fully connected layer;

[0083] Does the first fully connected layer perform matrix multiplication on the reduced feature set using the first fully connected weight matrix and the first fully connected bias vector? The expression is:

[0084] ;

[0085] in, This represents the output features of the first fully connected layer. and These represent the weight matrix and bias vector of the first fully connected layer, respectively. This represents matrix multiplication.

[0086] The second fully connected layer performs a fully connected operation on the output of the first fully connected layer using the second fully connected weight matrix and the second fully connected bias vector. The expression is as follows:

[0087] ;

[0088] in, This represents the output features of the second fully connected layer. and These represent the weight matrix and bias vector of the second fully connected layer, respectively. The output of the second fully connected layer is then used as the process depth feature. , represented as: .

[0089] Step 500: Perform feature fusion between flame depth features and process depth features, and predict the oxygen content of flue gas using the fused multimodal depth features.

[0090] Specifically, this embodiment uses a neural network composed of a third fully connected layer and a fourth fully connected layer to flatten the [processor / processor]. and After concatenation, feature fusion is performed, expressed as follows:

[0091] ;

[0092] in, The representative flattens and processes the data; This represents splicing / joining. This represents the concatenated features. The output formula for the third fully connected layer is:

[0093] ;

[0094] in, This represents the output features of the third fully connected layer. and These represent the weight matrix and bias vector of the third fully connected layer, respectively. The output of the third fully connected layer is then used as the multimodal deep feature. , represented as: Then The input to the fourth fully connected layer is expressed as follows:

[0095] ;

[0096] in, Represents the output of the fourth fully connected layer. and These represent the weights and biases of the fourth fully connected layer, respectively. The output of the fourth fully connected layer is then used as the final output, i.e., the multimodal deep feature, expressed as: ,in, This represents the predicted value of FGOC.

[0097] This invention uses flame image data from a certain MSWI power plant as an example. In this example, the combustion flame video is transmitted via cable and acquired through a video card. After processing, flame image datasets and process datasets are obtained. The flame image dataset is extracted at 1-minute intervals, using the first frame image. The process dataset is the average of the process data within the corresponding 1-minute interval. There is a one-to-one correspondence between the flame image dataset and the process dataset. After flame image preprocessing and depth feature extraction, feature maps with 64 channels are obtained. Then, feature selection is performed based on PCC, with a feature selection threshold set to 0.53. Process data features with a correlation greater than the threshold to FGOC are extracted to form a reduced feature set for the process data. The correlation between process data and FGOC is as follows: Figure 2 As shown, some features are positively correlated with combustion state, while others are negatively correlated, and there are significant differences in correlation between different features. The reduced feature set of process data contains a total of 10 features, such as... As shown.

[0098] Table 1. Set of Characteristics for Process Data Reduction

[0099]

[0100] Simultaneously, deep features were extracted from the process data, and the activation value distribution of the 64 feature dimensions of the process data is as follows: Figure 3As shown, the activation values ​​of the 64 feature dimensions vary significantly, indicating that each feature responds differently to the input data during model processing. Higher activation values ​​indicate that the feature is more important in the network, while lower activation values ​​indicate that the feature has not been effectively activated in the current input data. The 64 feature dimensions do not function uniformly during model processing but rather respond selectively based on the characteristics of different input data. This multi-dimensional response pattern demonstrates that the model can capture different data features during learning, which helps improve its generalization ability.

[0101] After fusing the extracted deep features, the multimodal deep features are input into the network for FGOC prediction. Using an industrial dataset, the samples are divided into 5 equally spaced parts. The first 2 parts and the 5th part are used as training samples, the 4th part as validation samples, and the 3rd part as test samples. The network model parameters are set as follows: 145 training epochs, batch size of 16, and learning rate of 0.001. The R-values ​​of the experimental data are... 2 The results of the MAE and LOSS evaluation indicators are shown in Table 2.

[0102] Table 2 Results of Prediction and Evaluation Indicators for Flue Gas Oxygen Content

[0103]

[0104] Depend on As can be seen, the training set performs very well and fits the training data well; the validation set performance is lower than the training set performance; the test set performance is stable, but lower than the training set performance; indicating that the model has good generalization ability. Figure 4 As shown, this embodiment also compares and analyzes with IT2F-BLS, BPNN, and IT2FNN methods. The IT2F-BLS method has 100 subsystems, 8 subsystem rules, a hybrid weight coefficient of 0.7, 20 initial boosting nodes, a boosting layer scaling factor of 0.9, and a regularization parameter of 0.01. The BPNN method has 43 hidden layer neurons, a maximum convergence iteration count of 3000, a convergence error of 0.00001, and a learning rate of 0.001. The IT2FNN method has upper and lower bound scaling factors of 0.5, 20 rules, 500 iterations, and a learning rate of 0.001. The experimental results of the four methods are presented in terms of RMSE, MAE, and R... 2 The evaluation indicators are compared as shown in Table 3.

[0105] Table 3 Comparison of Experimental Results

[0106]

[0107]

[0108] This invention also provides a flue gas oxygen content prediction system based on multimodal deep features for MSWI processes, comprising:

[0109] The flame image preprocessing module is used to preprocess the flame images from the MSWI process to obtain preprocessed images. The preprocessing includes image scaling, center cropping, and normalization.

[0110] The flame image depth feature extraction module is used to perform multi-level feature extraction on the preprocessed image through a first neural network composed of 8 layers to obtain flame depth features.

[0111] The process data feature selection module is used to select features from the process data of the MSWI process using the Pearson correlation coefficient to obtain a reduced feature set;

[0112] The process data deep feature extraction module is used to extract deep features from the reduced feature set through a second neural network composed of two layers to obtain process deep features.

[0113] The flue gas oxygen content prediction module is used to fuse flame depth features and process depth features, and predict the flue gas oxygen content through the fused multimodal depth features.

[0114] The beneficial effects of this invention are as follows:

[0115] 1) By fusing the deep features of flame images and process data, the complementary information between multimodal data was fully explored, significantly improving the accuracy of flue gas oxygen content prediction;

[0116] 2) The first neural network adopts an 8-layer structure, which combines adaptive pooling layers and multi-level feature extraction processes to improve the image feature extraction capability;

[0117] 3) Redundant features were eliminated through PCC screening, which reduced model complexity and improved generalization ability;

[0118] 4) The process data feature extraction adopts an MLP with only 2 layers. Combined with the previous PCC feature reduction process, the amount of computation is greatly reduced, the prediction efficiency and performance are improved, and it is suitable for industrial real-time deployment.

[0119] 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.

[0120] 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 method for predicting oxygen content in flue gas during MSWI processes based on multimodal deep features, characterized in that, Includes the following steps: The flame image of the MSWI process is preprocessed to obtain a preprocessed image; the preprocessing includes: image scaling, center cropping and normalization. The flame depth features are obtained by performing multi-level feature extraction on the preprocessed image through a first neural network consisting of 8 layers. The process data of the MSWI process are used to select features by Pearson correlation coefficient to obtain a reduced feature set; The process deep features are obtained by performing deep feature extraction on the reduced feature set using a second neural network consisting of two layers. The flame depth features and the process depth features are fused together, and the oxygen content of the flue gas is predicted by the fused multimodal depth features.

2. The method for predicting flue gas oxygen content based on multimodal deep features in MSWI process according to claim 1, characterized in that, The first neural network includes: a first convolutional layer, a first pooling layer, a second convolutional layer, a second pooling layer, a third convolutional layer, a fourth convolutional layer, a third pooling layer, and a fourth pooling layer; The first convolutional layer performs multi-channel convolution operations on the preprocessed image using the ReLU activation function; The first pooling layer performs max pooling on the output of the first convolutional layer using a maximum value function. The second convolutional layer performs a convolution operation on the output of the first pooling layer using the second convolutional kernel matrix and the second convolutional bias vector. The second pooling layer performs a pooling operation on the output of the second convolutional layer using a second pooling kernel matrix; The third convolutional layer performs a convolution operation on the output of the second pooling layer using the third convolutional kernel matrix and the third convolutional bias vector. The fourth convolutional layer performs a convolution operation on the output of the third convolutional layer using the fourth convolutional kernel matrix and the fourth convolutional bias vector; The third pooling layer performs average pooling on the output of the fourth convolutional layer using an average value function. The fourth pooling layer performs adaptive pooling on the output of the third pooling layer using an adaptive function to obtain the flame depth feature.

3. The method for predicting oxygen content in MSWI process flue gas based on multimodal deep features according to claim 1, characterized in that, Feature selection of the process data of the MSWI process is performed using the Pearson correlation coefficient to obtain a reduced feature set, including: Through formula Calculate the Pearson correlation coefficient of the process data. ;in, For the p-th feature in the process data, for The mean, Oxygen content in flue gas for The mean; When the Pearson correlation coefficient is greater than a preset empirical threshold, the process data is marked as a candidate feature, and the candidate features are integrated into the reduced feature set.

4. The method for predicting flue gas oxygen content based on multimodal deep features in MSWI process according to claim 1, characterized in that, The second neural network consists of a first fully connected layer and a second fully connected layer; The first fully connected layer performs matrix multiplication on the reduced feature set using the first fully connected weight matrix and the first fully connected bias vector; The second fully connected layer performs a fully connected operation on the output of the first fully connected layer using a second fully connected weight matrix and a second fully connected bias vector to obtain the process depth features.

5. The method for predicting flue gas oxygen content in MSWI processes based on multimodal deep features according to claim 1, characterized in that, The flame depth features and the process depth features are fused, and the oxygen content of the flue gas is predicted using the fused multimodal depth features, including: The flame depth feature after flattening is spliced ​​with the process depth feature to obtain the spliced ​​feature; The spliced ​​features are subjected to fully connected dimensionality reduction to obtain the multimodal deep features.

6. A flue gas oxygen content prediction system based on multimodal deep features for MSWI processes, characterized in that, include: The flame image preprocessing module is used to preprocess the flame images of the MSWI process to obtain preprocessed images; The preprocessing includes: image scaling, center cropping, and normalization. The flame image depth feature extraction module is used to perform multi-level feature extraction on the preprocessed image through a first neural network composed of 8 layers to obtain flame depth features. The process data feature selection module is used to select features from the process data of the MSWI process using the Pearson correlation coefficient to obtain a reduced feature set; The process data deep feature extraction module is used to extract deep features from the reduced feature set through a second neural network composed of two layers to obtain process deep features. The flue gas oxygen content prediction module is used to fuse the flame depth features and the process depth features, and predict the flue gas oxygen content through the fused multimodal depth features.

Images