Urban solid waste incineration process multi-controlled variable prediction method and system based on multi-modal depth feature fusion

Abstract

The invention provides an urban solid waste incineration process multi-controlled variable prediction method and system based on multi-modal depth feature fusion, and the method comprises the steps: collecting a flame image sequence and a process data sequence, and carrying out the preprocessing of a flame image, and obtaining a standard image set; respectively extracting channel features of the standard image set by using eight deep neural networks, and screening the channel features to obtain a flame image depth feature set; performing feature selection and depth feature extraction on the process data sequence to obtain a process data depth feature set; fusing the flame image depth feature set and the process data depth feature set to obtain a final fusion feature; and inputting the final fusion feature into a multi-controlled variable prediction network for prediction to obtain a multi-controlled variable prediction result. The method can effectively extract and integrate the flame image depth features and the important features of the process data, and achieves the precise prediction of multiple controlled variables in the urban solid waste incineration process, so as to improve the operation efficiency and the environmental protection effect.

Description

Technical Field

[0001] This invention relates to the field of urban solid waste incineration process prediction technology, and in particular to a method and system for predicting multiple controlled variables in urban solid waste incineration process based on multimodal deep feature fusion. Background Technology

[0002] The generation of municipal solid waste (MSW) has increased rapidly with economic development and rising living standards. How to effectively reduce and harmlessly treat MSW, as well as its resource utilization, and thus promote sustainable ecological development, has become an urgent problem to be solved worldwide. Due to limitations in land resources, environmental degradation, and technology, traditional landfill and composting methods are no longer sufficient to meet current development needs. MSW incineration (MSWI) has become the preferred method for MSW treatment in developing countries. The MSWI process is a typical complex industrial process involving multiple continuous operating stages, and researching its intelligent optimization control to achieve efficient and sustainable stable operation remains a highly challenging task.

[0003] Developed countries employ Automatic Combustion Control (ACC) systems to ensure the safe and stable operation of MSWI processes. This system automatically adjusts the output values ​​of manipulated variables based on detectable feedback values ​​of controlled variables, fine-tuning rules from domain experts, and advanced PID controllers. However, due to significant differences between developing and developed countries in MSW classification and equipment maintenance levels, the aforementioned ACC system is not suitable for developing countries. In my country, MSWI power plants primarily rely on manual operation by domain experts to ensure stable operation. This involves using multimodal data, including process variables, flame video, and on-site inspection voice feedback, to predict the changing trends of key controlled variables such as furnace temperature (FT), boiler steam flow (BSF), and flue gas oxygen content (FGOC), as well as conventional pollutants like CO. Based on accumulated experience rules, the output values ​​of manipulated variables are then used to control "fuel, air, and water." The operational effectiveness of this model is clearly dependent on the "expert," exhibiting significant randomness and subjectivity. It can even lead to different domain experts interpreting the same combustion conditions differently, thus hindering the long-term stable operation of MSWI power plants. Clearly, effective prediction of multiple controlled variables is the first problem that needs to be solved to achieve intelligent "replication and improvement" of expert predictions. Industrial sites have accumulated a wealth of multimodal data containing the experience of domain experts. Problems remain to be solved, such as how to extract features from flame images, extract features from process data, and construct predictive models for multiple controlled variables.

[0004] Flame images during MSWI processes contain various feature parameters, necessitating feature extraction. Deep neural network-based feature extraction methods can autonomously learn representative features from flame images. Existing correlation analysis methods cover multiple statistical dimensions, including linear, nonlinear, and monotonicity, effectively assessing the correlation strength between features and controlled variables. However, current research has not systematically addressed the following issues: how to select multiple deep neural networks to extract channel features from shallow to deep layers, and how to effectively utilize multidimensional correlation analysis methods to identify and select effective features. Solving this problem is crucial for improving the performance of controlled variable prediction models. Existing methods for constructing controlled variable prediction models mainly rely on single-modal information such as process data or flame images. How to dynamically fuse multimodal data from process data and flame images, and on this basis, achieve prediction of multiple controlled variables during MSWI processes, remains to be solved. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, the purpose of this invention is to provide a method and system for predicting multiple controlled variables in urban solid waste incineration processes based on multimodal deep feature fusion. This method can effectively extract and integrate deep features of flame images and important features of process data, thereby achieving accurate prediction of multiple controlled variables in urban solid waste incineration processes and improving operational efficiency and environmental protection effects.

[0006] To achieve the above objectives, this invention provides the following solution: a method for predicting multiple controlled variables in urban solid waste incineration processes based on multimodal deep feature fusion, comprising: Flame image sequences and process data sequences of solid waste incineration process are collected to obtain raw multimodal datasets, and the flame images are preprocessed to obtain standard image sets; Eight deep neural networks were used to extract the channel features of the standard image set, and the channel features were filtered using multidimensional correlation analysis and empirical thresholds to obtain a set of flame image depth features. The process data sequence is subjected to feature selection and deep feature extraction using Pearson correlation coefficient and a first multilayer feedforward neural network to obtain a set of deep features of process data. By using an iterative cross-attention fusion method, the flame image depth feature set and the process data depth feature set are fused to obtain the final fused features; The final fused features are input into the trained multi-controlled variable prediction network for prediction to obtain the multi-controlled variable prediction results.

[0007] Optionally, flame images and process data of the solid waste incineration process are collected to obtain a raw multimodal dataset, and the flame images are preprocessed to obtain a standard image set, including: A camera is installed on the rear wall of the furnace of solid waste incineration to collect video. The first frame of the video is captured every minute as a static image of the flame. All the static images of the flame are then combined to obtain a flame image sequence. Calculate the time average of each process variable in the video at 1 minute intervals to obtain the process data sequence; Based on the flame image sequence, image scaling and center cropping, tensor transformation and normalization, and notation conventions are performed sequentially to obtain a standard image set.

[0008] Optionally, eight deep neural networks are used to extract hierarchical channel features from the standard image set, and multidimensional correlation analysis and empirical thresholds are used to filter the hierarchical channel features to obtain a set of flame image depth features, including: Eight deep neural networks—Residual Network, Dense Connection Network, EfficientNet, VGG16, MobileNet, RegNet, ConvNeXt, and Vision Transformer—were selected to perform hierarchical feature extraction on the standard image set, resulting in multiple channel features for each network. Using three analytical dimensions—nonparametric rank correlation analysis, analysis of variance, and mutual information analysis—the relationship between the channel characteristics and the three controlled variables was evaluated, yielding correlation measures including correlation coefficient, variance value, and mutual information value; the three controlled variables included furnace temperature, boiler steam flow rate, and flue gas oxygen content. Determine whether the correlation coefficient is greater than or equal to a preset first empirical threshold, whether the variance value is greater than or equal to a preset second empirical threshold, and whether the mutual information value is greater than or equal to a preset third empirical threshold. If yes, retain the channel feature; otherwise, remove the channel feature to complete feature filtering and obtain an initial image feature set. Then, select 12 representative channel features from the initial image feature set to obtain a representative feature pool. By using a concatenation operation, the representative feature pools of eight deep neural networks are connected to obtain a set of deep features for the flame image.

[0009] Optionally, the process data sequence is subjected to feature selection and deep feature extraction using the Pearson correlation coefficient and a first multilayer feedforward neural network to obtain a set of deep features of the process data, including: For each single vector in the process data sequence, Pearson correlation coefficients are calculated with the three controlled variables to obtain the Pearson correlation coefficients for furnace temperature, boiler steam flow rate, and flue gas oxygen content for each single vector. Determine whether the Pearson correlation coefficient of the furnace temperature is greater than or equal to a preset fourth empirical threshold, whether the Pearson correlation coefficient of the boiler steam flow rate is greater than or equal to a preset fifth empirical threshold, and whether the Pearson correlation coefficient of the flue gas oxygen content is greater than or equal to a preset sixth empirical threshold. If yes, retain the single vector; otherwise, extract the single vector to complete the screening of the process data sequence and obtain the initial process data feature set. Using a first multilayer feedforward neural network with two fully connected layers, features are extracted from the initial process data feature set to obtain a process data representative feature set. Then, mutual information analysis is performed between the process data representative feature set and the controlled variable to obtain a process data deep feature set.

[0010] Optionally, the flame image depth feature set and the process data depth feature set are fused using an iterative cross-attention fusion method to obtain the final fused features, including: The flame image depth feature set and the process data depth feature set are each divided into eight non-overlapping sub-blocks, and an attention mechanism is used to perform bidirectional feature enhancement on the sub-blocks; wherein, the output of the first iteration is used as the input of the second iteration, and the deep fusion features are obtained after two iterations; By using a gating mechanism and element-wise multiplication, the flame image and the deep feature set of the process data are concatenated and then weighted and fused with the deep fusion features to obtain the final fused features.

[0011] Optionally, the final fused features are input into a trained multi-controlled variable prediction network for prediction to obtain multi-controlled variable prediction results, including: The original multimodal dataset is divided into multiple training groups in chronological order, and the training rounds, learning rate, batch size, and number of attention heads are set to complete the dataset partitioning and training parameter settings; wherein each training group consists of a first sample, a second sample, a third sample, a fourth sample, and a fifth sample; The first sample, the third sample, and the fifth sample are set as the training set, the second sample is set as the validation set, and the fourth sample is set as the test set. A second multilayer feedforward neural network with three fully connected layers is trained using the partitioned dataset and training parameters, and the corresponding weight matrix and bias vector are set to obtain a multi-controlled variable prediction network. The multi-controlled variable prediction network is then fitted and verified by combining evaluation metrics. The final fused features are input into the multi-controlled variable prediction network to predict the three controlled variables, thus obtaining the multi-controlled variable prediction result.

[0012] This invention also provides a multi-controlled variable prediction system for urban solid waste incineration processes based on multimodal deep feature fusion, comprising: The raw data acquisition module is used to acquire flame image sequences and process data sequences during the solid waste incineration process to obtain a raw multimodal dataset, and to preprocess the flame images to obtain a standard image set. The flame image depth feature construction module is used to extract the channel features of the standard image set using eight deep neural networks, and to filter the channel features using multidimensional correlation analysis and empirical thresholds to obtain a flame image depth feature set. The process data deep feature construction module is used to perform feature selection and deep feature extraction on the process data sequence using Pearson correlation coefficient and first multilayer feedforward neural network to obtain a process data deep feature set. The multimodal deep feature fusion module is used to fuse the flame image deep feature set and the process data deep feature set through an iterative cross-attention fusion method to obtain the final fused features; The multi-controlled variable prediction module is used to input the final fused features into the trained multi-controlled variable prediction network for prediction, and obtain the multi-controlled variable prediction result. The original data acquisition module, the flame image depth feature construction module, the process data depth feature construction module, the multimodal depth feature fusion module, and the multi-controlled variable prediction module are interconnected.

[0013] This invention discloses the following technical effects by providing a method and system for predicting multiple controlled variables in urban solid waste incineration processes based on multimodal deep feature fusion: 1. Achieving "intelligent replication and enhancement" of expert experience: Domain experts essentially form associations and judgments about FT, BSF, and FGOC based on flame image vision, process curve data, and empirical rules; This invention unifies the two sets of information, namely observing flames and observing instruments, into a learnable representation space through multimodal deep feature fusion and cross-modal attention, and then uses MLP for prediction, which is equivalent to mathematizing and modeling the "expert thinking process".

[0014] 2. Multimodal complementarity is fully utilized: Flame images capture spatial distribution and morphological changes, while process data provides precise values ​​for control and feedback quantities. The fusion of the two provides a more comprehensive characterization of the combustion state. Through cross-modal attention, the most important image features in the current process state and the most important process channels in the current flame state can be highlighted.

[0015] 3. The feature selection and fusion mechanism has strong interpretability: By selecting features through explicit indicators such as MI, Spearman, ANOVA, and PCC, some features can be interpreted as physically reasonable measurement points, such as the temperature of the primary combustion chamber and the flue gas temperature of each inlet; the dimensional distribution of the gating values ​​provides the possibility for interpreting the feature fusion weights.

[0016] 4. Significant performance advantages in prediction of multiple controlled variables: Compared with IT2FNN, LRDT, and XGBoost, this method has significantly lower RMSE and MAE scores. 2 Higher values ​​indicate the best fitting ability and overall performance; especially on FT and BSF, this method significantly outperforms traditional machine learning and fuzzy neural network methods.

[0017] 5. It provides a foundation for subsequent intelligent control (prediction and optimization of manipulated variables): Accurate prediction of key controlled variables is a prerequisite for realizing automatic combustion control (ACC) or intelligent control; the current model can already predict FT, BSF, and FGOC under the current operating conditions with relatively high accuracy, and in the future, it can be embedded into the control strategy for solution, thereby realizing automatic recommendation of manipulated variables.

[0018] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0019] 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.

[0020] Figure 1 This is a schematic diagram of the method flow provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the system architecture provided in an embodiment of the present invention; Figure 3 This is a process flow diagram of an MSWI power plant provided in an embodiment of the present invention; Figure 4 A diagram illustrating the multi-controlled variable prediction strategy for the MSWI process of multimodal deep feature fusion provided in this embodiment of the invention. Detailed Implementation

[0021] 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.

[0022] 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.

[0023] Example 1 like Figure 1 As shown, this invention provides a method for predicting multiple controlled variables in urban solid waste incineration processes based on multimodal deep feature fusion, including: Step 1: Collect flame image sequences and process data sequences of the solid waste incineration process to obtain the raw multimodal dataset, and preprocess the flame images to obtain a standard image set; Step 1 includes: 1.1 A camera is installed on the rear wall of the furnace of the solid waste incineration plant to collect video data. The first frame of the video is captured every minute as a static image of the flame. All the static images of the flame are then combined to obtain a flame image sequence. 1.2 Calculate the time average of each process variable in the video at 1 minute intervals to obtain the process data sequence.

[0024] 1.3 Based on the flame image sequence, image scaling and center cropping, tensor transformation and normalization, and notation conventions are performed sequentially to obtain a standard image set.

[0025] Step 2: Extract channel features from the standard image set using eight deep neural networks, and filter the channel features using multidimensional correlation analysis and empirical thresholds to obtain a set of flame image depth features; Step 2 includes: 2.1 Eight deep neural networks were selected: Residual Network, Dense Connection Network, EfficientNet, VGG16, MobileNet, RegNet, ConvNeXt, and Vision Transformer. Hierarchical feature extraction was performed on the standard image set to obtain multiple channel features for each network.

[0026] 2.2 Using three analytical dimensions—nonparametric rank correlation analysis, analysis of variance, and mutual information analysis—the relationship between the channel characteristics and the three controlled variables was evaluated, yielding correlation measures including correlation coefficient, variance value, and mutual information value; the three controlled variables included furnace temperature, boiler steam flow rate, and flue gas oxygen content.

[0027] 2.3 Determine whether the correlation coefficient is greater than or equal to a preset first empirical threshold, whether the variance value is greater than or equal to a preset second empirical threshold, and whether the mutual information value is greater than or equal to a preset third empirical threshold. If yes, retain the channel feature; otherwise, remove the channel feature to complete feature filtering and obtain an initial image feature set. Then, select 12 representative channel features from the initial image feature set to obtain a representative feature pool.

[0028] 2.4 Using a splicing operation, the representative feature pools of the eight deep neural networks are connected to obtain a set of deep features of the flame image.

[0029] Step 3: Using the Pearson correlation coefficient and a first multilayer feedforward neural network, feature selection and deep feature extraction are performed on the process data sequence to obtain a deep feature set of the process data; Step 3 includes: 3.1 For each single vector in the process data sequence, Pearson correlation coefficients are calculated with the three controlled variables to obtain the Pearson correlation coefficients for furnace temperature, boiler steam flow rate, and flue gas oxygen content for each single vector.

[0030] 3.2 Determine whether the Pearson correlation coefficient of the furnace temperature is greater than or equal to a preset fourth empirical threshold, whether the Pearson correlation coefficient of the boiler steam flow rate is greater than or equal to a preset fifth empirical threshold, and whether the Pearson correlation coefficient of the flue gas oxygen content is greater than or equal to a preset sixth empirical threshold. If yes, retain the single vector; otherwise, extract the single vector to complete the screening of the process data sequence and obtain the initial process data feature set.

[0031] 3.3 Using a first multilayer feedforward neural network with two fully connected layers, feature extraction is performed on the initial process data feature set to obtain a process data representative feature set. Then, mutual information analysis is performed between the process data representative feature set and the controlled variable to obtain a process data deep feature set.

[0032] Step 4: The flame image depth feature set and the process data depth feature set are fused using an iterative cross-attention fusion method to obtain the final fused features; Step 4 includes: 4.1 The flame image depth feature set and the process data depth feature set are divided into eight non-overlapping sub-blocks, and an attention mechanism is used to achieve bidirectional feature enhancement.

[0033] 4.2 Iterative cross-modal attention module: the output of the first iteration is used as the input of the second iteration, and the deep fusion features are obtained after two iterations.

[0034] 4.3 Using a gating mechanism and element-wise multiplication, the flame image and process data depth feature set are concatenated and then weighted and fused with the depth fusion feature to obtain the final fused feature.

[0035] Step 5: Input the final fused features into the trained multi-controlled variable prediction network for prediction to obtain the multi-controlled variable prediction result; Step 5 includes: 5.1 Divide the original multimodal dataset into multiple training groups in chronological order, and set the training rounds, learning rate, batch size and number of attention heads to complete the dataset division and training parameter settings; wherein, each training group consists of a first sample, a second sample, a third sample, a fourth sample and a fifth sample.

[0036] 5.2 Set the first sample, the third sample, and the fifth sample as the training set, set the second sample as the validation set, and set the fourth sample as the test set.

[0037] 5.3 A second multilayer feedforward neural network with three fully connected layers is trained using the partitioned dataset and training parameters. The corresponding weight matrix and bias vector are set to obtain a multi-controlled variable prediction network. The multi-controlled variable prediction network is then fitted and verified using evaluation metrics.

[0038] 5.4 Input the final fused features into the multi-controlled variable prediction network to predict the three controlled variables and obtain the multi-controlled variable prediction results.

[0039] like Figure 2 As shown, this invention also provides a multi-controlled variable prediction system for urban solid waste incineration processes based on multimodal deep feature fusion, including: The raw data acquisition module is used to acquire flame image sequences and process data sequences of the solid waste incineration process to obtain raw multimodal datasets, and to preprocess the flame images to obtain a standard image set.

[0040] The flame image depth feature construction module is used to extract channel features of the standard image set using eight deep neural networks, and to filter the channel features using multidimensional correlation analysis and empirical thresholds to obtain a flame image depth feature set.

[0041] The process data deep feature construction module is used to perform feature selection and deep feature extraction on the process data sequence using Pearson correlation coefficient and a first multilayer feedforward neural network to obtain a process data deep feature set.

[0042] The multimodal deep feature fusion module is used to fuse the flame image deep feature set and the process data deep feature set through an iterative cross-attention fusion method to obtain the final fused features.

[0043] The multi-controlled variable prediction module is used to input the final fused features into the trained multi-controlled variable prediction network for prediction, and obtain the multi-controlled variable prediction result.

[0044] The original data acquisition module, the flame image depth feature construction module, the process data depth feature construction module, the multimodal depth feature fusion module, and the multi-controlled variable prediction module are interconnected.

[0045] Example 2 This invention takes a typical grate furnace in China as an example to predict multiple controlled variables in the urban solid waste incineration process.

[0046] I. MSWI Process Description for Predicting Multiple Key Process Parameters Depend on Figure 3 It is known that MSW is first transported to the MSWI power plant by a dedicated collection vehicle, weighed, and then dumped into a solid waste pool from the unloading platform. Further, the MSW in the pool is thoroughly crushed, mixed, and piled using a crane grab bucket. After fermentation for 5-7 days, it is sent to the incinerator. Next, under the supply of air and combustion aids, it is ignited on the grate for combustion. Non-combustible ash is pushed out of the furnace through the burnt-out grate. The generated high-temperature flue gas is then cooled by absorbing heat in a preheating boiler. After purification, the flue gas is discharged into the atmosphere by an induced draft fan, while the generated high-temperature steam is used for power generation. Therefore, the MSWI process involves six stages: storage and fermentation, solid waste incineration, waste heat exchange, steam power generation, flue gas treatment, and flue gas emission. Among these, the solid waste incineration stage is the core of the entire MSWI process.

[0047] The three controlled variables—FT, BSF, and FGOC—form a deeply complementary qualitative relationship with flame images and process data. Flame images, as a direct visual mapping of the combustion state, intuitively reflect the intensity, stability, and uniformity of combustion within the furnace through their morphology, brightness, and color. Combustion intensity is directly related to FT, stability is closely related to FGOC, while the uniformity of heat release indirectly affects BSF output. Process data provides a quantitative basis for system control and feedback. FT and FGOC are directly affected by operating parameters such as combustion air and feed rate; BSF is directly determined by parameters such as boiler feedwater flow rate and steam pressure, and is explicitly coupled with FT through waste heat recovery. Therefore, flame images provide a "visual observation" of the combustion results, while process data reveals its "control causes." Both describe the combustion process from the perspectives of appearance and mechanism, respectively. Integrating this information is a key foundation for achieving accurate prediction, replication, and enhancement of expert intelligence for FT, BSF, and FGOC.

[0048] In the MSWI process, key controlled variables such as FT, BSF, and FGOC play specific roles. The MSWI process involves complex physical and chemical reactions with unclear mechanisms, producing various toxic and harmful gases. To reduce environmental pollution and ensure the complete decomposition and combustion of dioxins, national regulatory authorities require FT to be maintained above 850°C; however, excessively high FT can lead to furnace coking, equipment damage, and increased maintenance costs. MSWI power plants rely on BSF for turbine power generation, making BSF a crucial controlled variable closely related to economic efficiency. Furthermore, there is an explicit direct correlation between FT and BSF; therefore, properly adjusting the FT and feedwater flow rates of the waste heat recovery boiler is key to ensuring BSF output. FGOC is a crucial factor determining whether combustion is complete within the MSWI furnace. Generally, lower FGOC leads to an increase in reducing toxic gases such as CO and increases the risk of furnace coking, even jeopardizing the safe operation of the MSWI power plant. Conversely, higher FGOC reduces the heat transfer efficiency of high-temperature flue gas in the waste heat boiler, thus affecting power generation efficiency. Therefore, from the perspective of expert intelligent models in the field of "replication and enhancement", effectively predicting the three key controlled variables FT, BSF and FGOC is of great significance for ensuring the stable operation of the MSWI process and increasing the economic benefits of the power plant.

[0049] II. Modeling Strategy like Figure 4 As shown, Figure 4 middle, Represents a flame image, in which Representing the Image Represents the number of samples. Represents the coordinates of a pixel in the image; This represents the image after preprocessing. Represents the depth features of a flame image; This represents the process data corresponding to the flame image, where Representing the Process data corresponding to the image; Representative process data reduction characteristics; This represents the output of fully connected layer 1; Represents the depth characteristics of process data; Represents a cross-modal attention-based bidirectional enhancement feature module; The flame image token represents the input to the CMIX module; This represents the process data token input into the CMIX module; Flame image enhancement features representing the CMIX module; Representative process data enhancement features of CMIX modules; Flame image enhancement features representing the ICMIX module; This represents the process data enhancement features of the ICMIX module; This represents the stitching enhancement feature mapped to the original stitching feature dimension; Represents the final fusion characteristic; This represents the output of the fully connected layer 3; This represents the output of fully connected layer 4; , and These represent the predicted values ​​for FT, BSF, and FGOC, respectively.

[0050] Figure 4 The functions of different modules are as follows: The flame image depth feature construction module includes: a flame image preprocessing submodule that performs scaling, center cropping, and normalization on flame images to conform to the image input of CNN and VIT pre-trained models for the ImageNet dataset in transfer learning; and a flame image depth feature extraction and selection submodule that uses eight classic deep neural networks to extract layered channel features of flame images. It combines Spearman correlation coefficient, ANOVA, and mutual information to perform correlation analysis between channel features and FT, BSF, and FGOC from three dimensions: nonlinear dependence, monotonic trend measure, and significance of inter-group differences. After selecting effective channel features and removing redundancy, a CNN feature pool is constructed, which is the flame image depth feature.

[0051] Process data deep feature construction module: The process data feature selection submodule uses PCC to perform initial feature screening of the original process data to remove redundant and irrelevant features. The process data deep feature extraction submodule applies the deep network structure of MLP to perform hierarchical deep feature extraction of the screened effective features.

[0052] The multimodal deep feature fusion module enhances the features of flame images and process data bidirectionally through an iterative cross-attention mechanism, enabling single-modal data to extract information from other modalities. Then, a gating mechanism adaptively aggregates the original stitched features and deep fusion features to obtain the final fused features.

[0053] Multi-controlled variable prediction module: Input the final deep fusion features into the MLP network to achieve prediction of multiple controlled variables.

[0054] III. Algorithm Implementation 1. Flame Image Depth Feature Construction Module This module implements the construction of flame image depth features. First, flame images are preprocessed. Then, seven classic CNNs and VIT models are used to extract flame image depth features. Finally, Spearman correlation coefficient, ANOVA and mutual information are combined to select the features with the highest correlation with FT, BSF and FGOC as flame image depth features from three dimensions: nonlinear dependence, monotonic trend measure and significance of inter-group differences.

[0055] The characteristics of the eight deep neural networks selected in this paper are described as follows: 1) ResNet: Introduces residual connections, solving the gradient vanishing problem in deep networks and making it possible to train extremely deep networks; 2) DenseNet: Connects each layer to all previous layers through dense connections, greatly promoting feature reuse and gradient flow, achieving powerful performance with fewer parameters; 3) EfficientNet: Balances the depth, width, and resolution of the network through a composite model scaling method, achieving an excellent balance between computational efficiency and accuracy; 4) VGG16: A classic deep network whose simple and unified stacked structure laid the foundation for many subsequent networks, and the extracted features have strong generality; 5) MobileNet: Utilizes depth-separable convolutions to greatly reduce computational complexity and the number of parameters; 6) RegNet: A network structure searched through a quantized design space, its design is more regular and performs better than hand-designed networks, while also being faster than models searched through neural architectures; 7) ConvNeXt: Through a modern standard CNN design, it features the use of large convolutional kernels and fewer activation layers, resulting in superior performance; 8) Vision Transformer: The successful Transformer architecture in natural language processing is directly applied to image patches. It is based entirely on self-attention mechanism for feature extraction, representing a different design paradigm from CNN.

[0056] 1.1 Flame Image Preprocessing Flame images are preprocessed to conform to the image input of CNN and VIT pre-trained models on the ImageNet dataset for transfer learning. Flame image The preprocessing process is described using an example.

[0057] First, resize the image and crop it to the center; then, convert the image to a tensor and normalize it.

[0058] (1); in, This represents image scaling and cropping operations; This represents image tensor transformation and normalization processing; This represents the preprocessed flame image. The preprocessed flame image is denoted as... .

[0059] 1.2 Depth Feature Extraction and Selection from Flame Images This invention employs eight classic deep neural networks to extract features from flame images, and combines Spearman correlation coefficient, ANOVA, and mutual information as three correlation analysis methods to select image features.

[0060] Using ResNet as an example, features of flame images are extracted and selected.

[0061] To extract the first Flame image Taking features as an example, first, extract layered image features from shallow to deep layers. (2); in, Representing the Feature maps extracted by layer networks; Represents the ResNet network structure; The first, representing transfer learning Layer network parameters.

[0062] (3); in, Feature maps representing all layers of the ResNet network; Represents the number of network layers.

[0063] Then, global average pooling is performed on the feature map, with the first... The first layer Feature map For example, (4); in, This represents the scalar value obtained after global average pooling. and These represent the height and width of the feature map, respectively. Represents pixel value; Represents the position The activation value.

[0064] (5); in, Representing the Global average pooling characteristics of the layer; Representing the The number of channels in the layer.

[0065] (6); in, This represents the global average pooling feature of ResNet.

[0066] This section uses represent The flame image depth feature set of n samples is used to calculate the correlation between each layer of features and FT, BSF, and FGOC, with the i-th layer as the basis. The first layer Features Take the correlation with the FT as an example.

[0067] 1) Spearman correlation coefficient The calculation is as follows: (7); in, Represents the truth value of the Fourier Transform (FT); and Representing their respective first The ranking of each sample; Represents the number of samples.

[0068] The following criteria were used to select features with high Spearman correlation coefficients with the FT: (8); in, This represents the Spearman correlation coefficient threshold set based on experience; represent The flag value indicates whether the feature is selected. A value of 1 indicates that the feature is selected, while a value of 0 indicates that it is not selected.

[0069] 2) F-value of ANOVA The calculation is as follows: (9); in, The sum of squares between groups; Represents the sum of squares within a group; The number of representative groups.

[0070] (10); in, Representing the The number of samples in the group; Representing the The characteristic mean of the group; The overall mean of the features representing all samples.

[0071] (11); in, Representing the Group 1 The feature values ​​of each sample.

[0072] The calculation is as follows: (12); in, This represents the null hypothesis (that the means of all groups are equal). The cumulative distribution function representing the F-distribution; Degrees of freedom between groups; This represents the degree of freedom within the group.

[0073] The following criteria are used to select the most discriminative features. (13); in, This represents the F-value threshold of ANOVA, set based on experience. represent The flag value indicates whether the feature is selected. A value of 1 indicates that the feature is selected, while a value of 0 indicates that it is not selected.

[0074] 3) Mutual information value The calculation is as follows: (14); in, Represents the joint probability density; and This represents the marginal probability density.

[0075] The following criteria are used to select features with higher mutual information values ​​with the Fourier Transform (FT): (15); in, This represents a threshold value for mutual information set based on experience. represent The flag value indicates whether the feature is selected. A value of 1 indicates that the feature is selected, while a value of 0 indicates that it is not selected.

[0076] Similarly, after performing correlation analysis and initial screening on all network layer channels with BSF and FGOC, the selected features were further refined and screened using Spearman correlation coefficient as the primary factor, supplemented by ANOVA and mutual information. Finally, 12 channel features were selected as the ResNet feature pool, denoted as [missing information]. .

[0077] Feature extraction and selection are performed on all CNN and VIT networks to obtain the feature pool. , (16); in, , , , , , , and These represent the feature pools of eight deep neural networks; This indicates splicing / merging.

[0078] The final deep neural network feature pool As a depth feature of the flame image, it is denoted as .

[0079] 2. Process Data Deep Feature Construction Module This module first uses PCC to perform initial feature screening of the raw process data to remove redundant and irrelevant features. Then, it applies the deep network structure of MLP to perform hierarchical deep feature extraction on the screened effective features. By combining the linear correlation analysis of PCC and the nonlinear feature learning capability of MLP, deep feature extraction of process data is achieved.

[0080] 2.1 Selection of Process Data Features First, calculate separately The individual sub-features and their PCC values ​​for FT, BSF, and FGOC are described, with FT as an example.

[0081] Calculate the first Features PCC value of FT As shown below: (17); in, Represents the truth value of the Fourier Transform (FT). Represents the mean of the FT. represent The mean.

[0082] The following criteria were used to select features with high correlation to the Fourier Transform (FT): (18); in, This represents a PCC value threshold set based on experience. represent The flag value indicates whether the feature is selected. A value of 1 indicates that the feature is selected, while a value of 0 indicates that it is not selected.

[0083] Similarly, after feature selection for BSF and FGOC respectively, the set of process data features after redundancy removal is denoted as the process data reduction features. .

[0084] 2.2 Process Data Deep Feature Extraction Submodule Will Input the MLP to extract deep features from the process data. This network module employs a two-layer structure, containing fully connected layer 1 and fully connected layer 2, with the first layer being the most advanced. Taking feature extraction from a single sample as an example, it is illustrated below. (19); in, This represents the output characteristics of fully connected layer 1; and These represent the weight matrix and bias vector of fully connected layer 1, respectively; This represents matrix multiplication.

[0085] (20); in, This represents the output characteristics of the fully connected layer 2; and These represent the weight matrix and bias vector of the fully connected layer 2, respectively.

[0086] Using the output of fully connected layer 2 as the final extracted features, as shown below, (twenty one); in, This represents the depth characteristics of the process data.

[0087] The final extracted process data depth features are denoted as... .

[0088] 3. Multimodal deep feature fusion module This module employs an iterative cross-attention mechanism to enhance features across modalities in both directions, and then uses gated weighted fusion of the original spliced ​​features and the dynamically fused features to obtain the final fused features.

[0089] With the first Taking the fusion of individual samples as an example, it is represented as follows: First of all, and Evenly divided into and Non-overlapping sub-blocks are linearly projected onto a unified embedding space and learnable positional codes are added to capture spatial relationships between different locations. Then, complementary information is extracted using dual CMIX modules to enhance the features of image and process data respectively; the two CMIX modules do not share parameters. The description uses the CMIX module for the image branch as an example. (twenty two); in, and These represent the image and process data tokens input to the CMIX module, respectively. This represents the image features enhanced by the CMIX module; This is a CMIX module designed for image branching.

[0090] The CMIX module is implemented as follows. First, the process data modality is... Projected onto two matrices and , representing key and value respectively, representing the image modality Projected onto the query matrix , (twenty three); (twenty four); (25); in, These represent the weight matrix, Represents the dimension after uniform projection. This represents the number of attention heads.

[0091] Next, the relevance matrix is ​​calculated by performing a dot product operation between the query and the key, followed by softmax normalization to generate an attention score. The attention score represents the similarity between image and process data features, which we correlate with... Multiplication is used to enhance image features.

[0092] (26); in, Image features representing attention output; represent The transpose of .

[0093] Then, Add to the features of the input image, (27); Finally, a feedforward network consisting of two fully connected layers further refines the global information and outputs the final image enhancement features. , (28); in, It represents the forward feed network.

[0094] Similarly, another CMIX module enhances the process data branch, generating enhanced process data features. .

[0095] To preserve cross-modal complementary information and improve performance, we propose an iterative learning strategy within the CMIX framework. The output of each iteration serves as the input for the next iteration, termed an ICMIX module. Unlike traditional methods that may lead to overfitting by stacking modules, our iterative approach improves cross-modal and intra-modal information without increasing the number of parameters. For m iterations, the process can be simplified to... (29); in, Represents the ICMIX module, which contains two branches: image and process data. and These represent the enhanced image and process data features after m iterations, respectively.

[0096] Then, the flattened and enhanced image and process data features are stitched together and projected onto the dimensions of the original stitched features.

[0097] (30); in, Represents the fusion characteristics after processing; The representative flattens and processes the data; This represents projection processing.

[0098] Finally, by learning the gating weights The final features are obtained by dynamically balancing the original spliced ​​features and the deep fusion features.

[0099] (31); in, It represents the original features of the splicing.

[0100] (32); in, Represents the learned gating weights; Represents the initial gating weights; This represents the Sigmoid function.

[0101] (33); in, Represents the final fusion characteristic; This represents element-wise multiplication.

[0102] 4. Multi-controlled variable prediction module This module enables prediction of multiple controlled variables based on the final deep fusion features, with the first... Taking the prediction of a single sample as an example, it is shown below. Will Input a three-layer MLP network.

[0103] (34); in, This represents the output characteristics of the fully connected layer 3; and These represent the weight matrix and bias vector of the fully connected layer 3, respectively.

[0104] (35); in, This represents the output characteristics of the fully connected layer 4; and These represent the weight matrix and bias vector of the fully connected layer 4, respectively.

[0105] (36); in, This represents the output characteristics of the fully connected layer 5; and These represent the weight matrix and bias vector of the fully connected layer 5, respectively.

[0106] The output of the fully connected 5 network, which represents the prediction results of the three controlled variables, is denoted as follows: , and .

[0107] The final predicted output is denoted as , and .

[0108] 5. Experimental verification 5.1 Data Description The data used in this experiment came from a MSWI power plant. To monitor the flame combustion status, a camera was installed on the rear wall of the furnace. The flame video was transmitted via cable and acquired by a video card. After processing, flame image datasets and process datasets were obtained. The flame image dataset was captured at 1-minute intervals, using the first frame image; the process dataset consisted of the average of the process data within 1 minute. The flame image dataset and process dataset samples corresponded one-to-one.

[0109] 5.2 Evaluation Indicators This article selects R 2 The evaluation is based on three indicators: RMSE and MAE, as shown in the following formula: (37); (38); (39); in, Represents the truth value. Represents the predicted value. represent The average value.

[0110] 5.3 Experimental Results 5.3.1 Results of the Flame Image Depth Feature Construction Module 5.3.1.1 Flame Image Preprocessing Results The preprocessed images achieve a standardized format, ensuring consistent size and numerical properties while preserving key visual features. Maintaining consistency between the preprocessed images and the input requirements of subsequent transfer learning neural network models improves model performance, accelerates convergence, and enhances training stability.

[0111] 5.3.1.2 Results of the Flame Image Depth Feature Extraction and Selection Submodule Using the method described above, the flame image depth features based on different layers of the eight deep networks described above were quantitatively analyzed using three complementary indices: MI, Spearman correlation coefficient, and ANOVA F-value. The correlations of the three controlled variables, FT, BSF, and FGOC, are shown in Figures S1-S23 in the supplementary material. The parameter thresholds used are shown in Table 1.

[0112] Table 1. Statistical table of relevant parameter thresholds used for feature selection of 8 deep networks.

[0113] Based on the above thresholds, the selected features were deredundantd and screened using Spearman correlation coefficient as the primary factor and ANOVA and mutual information as secondary factors. The correlation values ​​of the 12 channel features selected for each neural network are shown in Table 2 below.

[0114] Table 2 Channels based on correlation screening

[0115] Each neural network contains different network layers and the number of channels, and the correlation between the channel features and the controlled variable varies significantly across different network layers. Highly correlated channels are distributed in both shallow and deep networks, forming local clusters in specific network layers. Different statistical indicators exhibit varying consistency patterns across different network layers. Some network layers show synergistic changes across the three evaluation indicators, while others show differences. The synergy and complementarity among statistical indicators ensure the robustness of the conclusions. As shown in Table 2, the features of ResNet and DenseNet, based on residuals and dense connections, are concentrated entirely in the shallowest layers or distributed in the shallow and middle layers, respectively, confirming the effectiveness of skip connections in preserving basic discriminative patterns. The traditional chain-structured VGG16 concentrates on the middle layers to complete the core abstraction. RegNet and ConvNeXt, as representatives of modern CNNs, show a significant bias towards the middle and deep layers, even the final stage, highlighting their orientation towards high-order semantic extraction through normalized module stacking or deep staged design. The lightweight MobileNet and EfficientNet features are distributed across multiple levels, reflecting the discriminative power formation path under efficiency constraints. Vision Transformer has the broadest feature coverage, fully demonstrating the dynamic refinement process of the self-attention mechanism from establishing patch associations to integrating global semantics.

[0116] 5.3.2 Results of the Process Data Deep Feature Construction Module 5.3.2.1 Results of Process Data Feature Selection The process data features showed higher correlations with FT and BSF than with FGOC. Some features were positively correlated with the three controlled variables, while others were negatively correlated. Based on the correlation results of the three variables, 12 features were selected after redundancy removal from the process data. These features are: mean temperature of the primary combustion chamber, temperature on the left side of the primary combustion chamber, flue gas temperature on the middle side of the primary combustion chamber, temperature on the right side of the primary combustion chamber, flue gas temperature at the right inlet of the economizer, flue gas temperature at the right inlet of the evaporator, flue gas temperature at the left inlet of the economizer, flue gas temperature at the left inlet of the tertiary superheater, flue gas dust concentration, air temperature at the top of the grate in the burnout section, flue gas temperature on the left side of the first channel, and cooling air outlet temperature on the left side of the furnace wall.

[0117] 5.3.2.2 Results of Deep Feature Extraction from Process Data The correlation strength between process data depth features and FGOC, FT, and BSF varies. FGOC's MI values ​​are generally low, with a few features reaching relatively high values ​​of 0.4-0.5, indicating that most process data depth features have a relatively weak correlation with flue gas oxygen content, and a few features are key influencing factors. FT's MI values ​​are generally high, with a dense distribution of high MI features and a peak value close to 1.2, reflecting a generally strong correlation between process data depth features and FT, driven by multiple concentrated process data depth features. BSF's MI value range is similar to FT, with a slightly lower density of high MI features, and its correlation strength with process data depth features is also high. The correlation strength between FT and BSF with process data depth features is higher than that of FGOC, and there are differences within FT and BSF.

[0118] 5.3.3 Results of the Multimodal Deep Feature Fusion Module The total feature dimension is 160. A gating value of 1 means that the model uses the fused features completely in that feature dimension; a gating value of 0 means that the original spliced ​​features are used completely.

[0119] The sample gating values ​​show a clear high-low differential distribution, indicating that the model does not indiscriminately fuse and superimpose all feature dimensions, but rather allocates the weights of the fused features in a refined and targeted manner based on the cross-modal correlation value of features in different dimensions. The overall range of the gating values ​​is concentrated between 0.4 and 0.5, with virtually no fused features completely shut off by the gating mechanism, indicating that the fused features contribute weights across all dimensions and contribute to the model's fitting effect.

[0120] 5.3.4 Results of the Multi-Controlled Variable Prediction Module The input consists of deep features from flame images and deep features from process data, with feature dimensions of 96 and 64, respectively. Samples are divided into groups of five sequentially. Within each group, the 1st, 3rd, and 5th samples are used as the training set, the 2nd sample as the validation set, and the 4th sample as the test set. The training set contains 471 samples, and the validation and test sets each contain 157 samples. The parameters of the proposed method are set as follows: 300 training epochs, a learning rate of 5e-4, a batch size of 16, and 4 attention heads.

[0121] FGOC, FT, and BSF test set prediction curves: R-values ​​of experimental data 2 The results of the RMSE and MAE evaluation indicators are shown in Table 3.

[0122] Table 3 Results of Prediction and Evaluation Indicators for Multiple Controlled Variables

[0123] As shown in Table 3, the training set performs very well and fits the training data well; the validation set performs slightly worse than the training set, but still remains at a high level; the test set performs stably, but is lower than both the training and validation sets. This indicates that the model has good generalization ability, but there is some overfitting, and there is still room for improvement.

[0124] 5.4 Method Comparison The parameters for the IT2FNN method are set as follows: upper and lower bound scaling factor of 0.5, number of rules of 20, number of iterations of 500, and learning rate of 0.001; the parameters for the LSDT method are set as follows: regularization factor of 2-5, minimum number of samples of 30, and number of features of 12; the parameters for the XGBoost method are set as follows: minimum number of samples of 30, number of features of 12, number of decision trees of 60, learning rate of 0.3, and regularization factor of 0.5.

[0125] The experimental results of IT2FNN, LSDT, XGBoost and the proposed method are compared using RMSE and MAE evaluation metrics, as shown in Tables 4 and 5.

[0126] Table 4 Comparison of Experimental Results (RMSE)

[0127] Table 5 Comparison of Experimental Results (MAE)

[0128] Tables 4 and 5 show that: 1) FGOC prediction: The proposed method has a small error, indicating strong fitting ability, but some overfitting exists; IT2FNN and LSDT have slightly weaker fitting ability, significantly inferior to the proposed method; XGBoost has poor fitting ability and large error, but its generalization ability is very good; 2) FT prediction: The proposed method performs well and fits the data well; LSDT performs slightly weaker than the proposed method, but its generalization ability is comparable; IT2FNN and XGBoost have significant errors and poor fitting effect; 3) BSF prediction: The proposed method performs well and has strong fitting ability; IT2FNN and LSDT perform well, but their fitting effect is inferior to the proposed method; XGBoost has poor fitting ability, but good generalization ability; 4) Average value: The proposed method performs well on the training set, validation set, and test set, but some overfitting exists; LSDT performs well, but there is a certain gap compared to the proposed method, and the fitting effect is slightly weaker; IT2FNN and XGBoost have poor fitting effect, but good generalization ability.

[0129] In summary, the method presented in this paper performs the best, exhibiting the smallest error in the prediction of multiple controlled variables and the strongest fitting ability.

[0130] 6. Conclusion Accurate prediction of key controlled variables in MSWI processes based on multimodal data is a crucial factor in realizing AI-enabled applications. To predict FT, BSF, and FGOC, this paper constructs a multi-controlled variable prediction strategy for MSWI processes based on multimodal deep feature fusion. The main contributions are: 1) Extracting layered features from flame images based on eight pre-trained deep neural network models, screening effective features through multi-dimensional correlation analysis, and constructing a deep feature pool for flame images, thus solving the problem of how to effectively extract deep features from flame images; 2) Combining Pearson correlation coefficient and multilayer perceptron algorithm to perform initial feature screening of process data and extracting deep features to simulate the perception process of process data by domain experts, thus solving the problems of how to retain important features of process data and how to mine its deep features; 3) Employing an iterative cross-modal attention mechanism to bidirectionally enhance two types of features, and adaptively fusing the original and enhanced features through gating vectors, finally using MLP to predict multiple controlled variables, thus solving the problem of the discrepancy between relying on single-modal information to construct controlled variable prediction models and the static fusion of multimodal data methods and the fact that key process parameters are perceived based on multimodal data in actual industrial settings. Future research in this paper will focus on improving the dynamic fusion mechanism of multimodal data and predicting manipulated variables.

[0131] Therefore, this invention provides a method and system for predicting multiple controlled variables in the urban solid waste incineration process based on multimodal deep feature fusion. This method can effectively extract and integrate the deep features of flame images and important features of process data, thereby achieving accurate prediction of multiple controlled variables in the urban solid waste incineration process and improving operational efficiency and environmental protection.

[0132] 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.

[0133] 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 predicting multiple controlled variables in urban solid waste incineration processes based on multimodal deep feature fusion, characterized in that, include: Flame image sequences and process data sequences of solid waste incineration process are collected to obtain raw multimodal datasets, and the flame images are preprocessed to obtain standard image sets; Eight deep neural networks were used to extract the channel features of the standard image set, and the channel features were filtered using multidimensional correlation analysis and empirical thresholds to obtain a set of flame image depth features. The process data sequence is subjected to feature selection and deep feature extraction using Pearson correlation coefficient and a first multilayer feedforward neural network to obtain a set of deep features of process data. By using an iterative cross-attention fusion method, the flame image depth feature set and the process data depth feature set are fused to obtain the final fused features; The final fused features are input into the trained multi-controlled variable prediction network for prediction to obtain the multi-controlled variable prediction results.

2. The method for predicting multiple controlled variables in urban solid waste incineration process based on multimodal deep feature fusion according to claim 1, characterized in that, Flame images and process data from the solid waste incineration process are collected to obtain a raw multimodal dataset. The flame images are then preprocessed to obtain a standard image set, including: A camera is installed on the rear wall of the furnace of solid waste incineration to collect video. The first frame of the video is captured every minute as a static image of the flame. All the static images of the flame are then combined to obtain a flame image sequence. Calculate the time average of each process variable in the video at 1 minute intervals to obtain the process data sequence; Based on the flame image sequence, image scaling and center cropping, tensor transformation and normalization, and notation conventions are performed sequentially to obtain a standard image set.

3. The method for predicting multiple controlled variables in urban solid waste incineration processes based on multimodal deep feature fusion according to claim 2, characterized in that, Eight deep neural networks were used to extract hierarchical channel features from the standard image set, and multidimensional correlation analysis and empirical thresholds were used to filter the hierarchical channel features to obtain a set of flame image depth features, including: Eight deep neural networks—Residual Network, Dense Connection Network, EfficientNet, VGG16, MobileNet, RegNet, ConvNeXt, and Vision Transformer—were selected to perform hierarchical feature extraction on the standard image set, resulting in multiple channel features for each network. Using three analytical dimensions—nonparametric rank correlation analysis, analysis of variance, and mutual information analysis—the relationship between the channel characteristics and the three controlled variables was evaluated, yielding correlation measures including correlation coefficient, variance value, and mutual information value; the three controlled variables included furnace temperature, boiler steam flow rate, and flue gas oxygen content. Determine whether the correlation coefficient is greater than or equal to a preset first empirical threshold, whether the variance value is greater than or equal to a preset second empirical threshold, and whether the mutual information value is greater than or equal to a preset third empirical threshold. If yes, retain the channel feature; otherwise, remove the channel feature to complete feature filtering and obtain an initial image feature set. Then, select 12 representative channel features from the initial image feature set to obtain a representative feature pool. By using a concatenation operation, the representative feature pools of eight deep neural networks are connected to obtain a set of deep features for the flame image.

4. The method for predicting multiple controlled variables in urban solid waste incineration processes based on multimodal deep feature fusion according to claim 3, characterized in that, The process data sequence is subjected to feature selection and deep feature extraction using Pearson correlation coefficient and a first multilayer feedforward neural network to obtain a set of deep features for the process data, including: For each single vector in the process data sequence, Pearson correlation coefficients are calculated with the three controlled variables to obtain the Pearson correlation coefficients for furnace temperature, boiler steam flow rate, and flue gas oxygen content for each single vector. Determine whether the Pearson correlation coefficient of the furnace temperature is greater than or equal to a preset fourth empirical threshold, whether the Pearson correlation coefficient of the boiler steam flow rate is greater than or equal to a preset fifth empirical threshold, and whether the Pearson correlation coefficient of the flue gas oxygen content is greater than or equal to a preset sixth empirical threshold. If yes, retain the single vector; otherwise, extract the single vector to complete the screening of the process data sequence and obtain the initial process data feature set. Using a first multilayer feedforward neural network with two fully connected layers, features are extracted from the initial process data feature set to obtain a process data representative feature set. Then, mutual information analysis is performed between the process data representative feature set and the controlled variable to obtain a process data deep feature set.

5. The method for predicting multiple controlled variables in urban solid waste incineration process based on multimodal deep feature fusion according to claim 4, characterized in that, By fusing the flame image depth feature set and the process data depth feature set through an iterative cross-attention fusion method, the final fused features are obtained, including: The flame image depth feature set and the process data depth feature set are each divided into eight non-overlapping sub-blocks, and an attention mechanism is used to perform bidirectional feature enhancement on the sub-blocks; wherein, the output of the first iteration is used as the input of the second iteration, and the deep fusion features are obtained after two iterations; By using a gating mechanism and element-wise multiplication, the flame image and the deep feature set of the process data are concatenated and then weighted and fused with the deep fusion features to obtain the final fused features.

6. The method for predicting multiple controlled variables in urban solid waste incineration process based on multimodal deep feature fusion according to claim 5, characterized in that, The final fused features are input into a trained multi-controlled variable prediction network for prediction, resulting in multi-controlled variable prediction results, including: The original multimodal dataset is divided into multiple training groups in chronological order, and the training rounds, learning rate, batch size, and number of attention heads are set to complete the dataset partitioning and training parameter settings; wherein each training group consists of a first sample, a second sample, a third sample, a fourth sample, and a fifth sample; The first sample, the third sample, and the fifth sample are set as the training set, the second sample is set as the validation set, and the fourth sample is set as the test set. A second multilayer feedforward neural network with three fully connected layers is trained using the partitioned dataset and training parameters, and the corresponding weight matrix and bias vector are set to obtain a multi-controlled variable prediction network. The multi-controlled variable prediction network is then fitted and verified by combining evaluation metrics. The final fused features are input into the multi-controlled variable prediction network to predict the three controlled variables, thus obtaining the multi-controlled variable prediction result.

7. A multi-controlled variable prediction system for urban solid waste incineration processes based on multimodal deep feature fusion, characterized in that, include: The raw data acquisition module is used to acquire flame image sequences and process data sequences during the solid waste incineration process to obtain a raw multimodal dataset, and to preprocess the flame images to obtain a standard image set. The flame image depth feature construction module is used to extract the channel features of the standard image set using eight deep neural networks, and to filter the channel features using multidimensional correlation analysis and empirical thresholds to obtain a flame image depth feature set. The process data deep feature construction module is used to perform feature selection and deep feature extraction on the process data sequence using Pearson correlation coefficient and first multilayer feedforward neural network to obtain a process data deep feature set. The multimodal deep feature fusion module is used to fuse the flame image deep feature set and the process data deep feature set through an iterative cross-attention fusion method to obtain the final fused features; The multi-controlled variable prediction module is used to input the final fused features into the trained multi-controlled variable prediction network for prediction, and obtain the multi-controlled variable prediction result. The original data acquisition module, the flame image depth feature construction module, the process data depth feature construction module, the multimodal depth feature fusion module, and the multi-controlled variable prediction module are interconnected.

Images