Solid waste incineration working condition recognition method and system based on hybrid convolution and physical constraints

Abstract

The present application relates to the technical field of computer vision, solid waste incineration process monitoring and industrial intelligent control, and provides a solid waste incineration working condition recognition method and system based on mixed convolution and physical constraints, which comprises: deploying a hearth camera to collect original images, and calibrating to determine the mapping of pixel rows and hearth physical height; preprocessing the original images and dividing pixel bands according to the hearth physical structure to obtain input tensors; using CNN network and ViT network to respectively extract multi-scale feature maps and global feature sequences of the input tensors; taking the divided pixel bands as a priori to guide bidirectional cross attention, fuse the multi-scale feature maps and the global feature sequences, and obtain a fusion feature vector; introducing a physical constraint regularization conforming to the combustion law of solid waste incineration in the training stage; constructing a joint loss function of the model, and outputting the recognition result of the solid waste incineration working condition. Through the method, the accuracy and robustness of solid waste incineration working condition detection are improved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision, solid waste incineration process monitoring and industrial intelligent control, specifically to a method and system for identifying solid waste incineration operating conditions based on hybrid convolution and physical constraints. Background Technology

[0002] Currently, urban solid waste incineration plants widely use furnace cameras to monitor flame status in real time. Existing judgment methods mainly rely on manual observation, combined with operator experience, to determine combustion conditions (normal combustion, smoldering, uneven combustion, cross-contamination, etc.). This method has the following problems: it is highly subjective, human judgment depends on experience, and is easily affected by factors such as visual fatigue and personal habits; the response is not timely, and it cannot provide timely adjustment suggestions when the flame status changes rapidly (such as instantaneous cross-contamination); it lacks stable quantitative indicators, making it difficult to directly provide input references for the control system.

[0003] In recent years, deep learning methods have been introduced into the field of flame image classification. Convolutional Neural Networks (CNNs) can effectively extract local texture and shape features, while Visual Transformers (ViTs) possess global information modeling capabilities. For example, the Chinese invention patent application CN119693719A, "A Method for Recognizing the State of Urban Solid Waste Incineration Flames Based on Multi-Feature Enhancement Res-Transformer," enriches the training data and improves the model's generalization ability by employing image denoising, image dehazing, and geometric transformation preprocessing methods; it achieves efficient automatic feature extraction and state recognition of urban solid waste incineration flame images through deep learning techniques based on multiple feature enhancements. However, these methods rely on a single end-to-end model architecture. These methods share common limitations: pure CNNs struggle to capture the global spatial relationships within the furnace; pure ViTs are insensitive to local flame details; and more importantly, existing methods do not embed the physical structural knowledge of the incineration process (such as the drying section, combustion section, and burnout section), resulting in insufficient generalization ability and robustness when fuel composition changes or operating conditions fluctuate. Therefore, how to effectively integrate local and global visual features in the flame recognition model and introduce the physical laws of the combustion process as constraints is a key challenge to achieving high-precision, robust, and interpretable intelligent incineration monitoring. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of existing methods and provide a working condition identification method that can simultaneously integrate local details and global context of flame images and embed physical constraints of the combustion process, so as to significantly improve the identification accuracy.

[0005] To address the aforementioned technical problems, this invention proposes a physically-prior-guided hybrid deep learning framework. Its core idea is as follows: First, based on the physical structure of the furnace, the image is divided into pixel bands to generate a spatial mask prior. Second, a dual-branch feature extractor combining CNN and ViT is constructed, focusing on local features and global context respectively. Third, a bidirectional cross-attention fusion module guided by segmented information is designed to achieve feature complementarity. Finally, physical constraints of the combustion process are introduced as a regularization term during training, ensuring that the model's predictions conform to industrial mechanisms. Through these methods, physically interpretable, high-precision operational condition identification is achieved.

[0006] This invention provides a method for identifying solid waste incineration conditions based on hybrid convolution and physical constraints, comprising the following steps:

[0007] S1. Deploy furnace cameras to acquire raw images, calibrate and determine the mapping between pixel rows and the physical height of the furnace; preprocess the raw images and divide them into pixel bands according to the physical structure of the furnace to obtain the input tensor. ;

[0008] S2. Use a CNN network to extract the input tensor. Multi-scale local features, output multi-scale feature map ;

[0009] S3. Extracting input tensors using a ViT network. The global context features are used to output the global feature sequence. ;

[0010] S4. Using the segmented pixel bands as priors, guide bidirectional cross-attention and fuse multi-scale feature maps. With global feature sequences Obtain the fused feature vector ;

[0011] S5. Introduce physical constraint regularization that conforms to the combustion law of solid waste incineration during the training phase;

[0012] S6. Construct the joint loss function of the model and output the identification results of solid waste incineration conditions.

[0013] Further, step S1 includes the following steps:

[0014] S11. During the deployment phase of the furnace camera, the mapping between pixel rows and physical height is determined through geometric calibration or manual calibration, and linear mapping is performed. And set the region of interest in the image. ,in: Indicates the physical height of the pixel within the furnace; Indicates the vertical row index of a pixel in the image; Represents the linear mapping coefficients obtained from calibration; The region of interest mask is used to define the effective imaging area of ​​the furnace.

[0015] S12. Perform distortion correction, resolution unification, luminance channel extraction, contrast enhancement and noise suppression on the acquired images to reduce the impact of lighting changes and equipment differences;

[0016] S13. To encode the prior knowledge of the furnace physical structure into an image mask, the ROI region is divided into drying sections using a fixed calibration method or an adaptive combustion line detection method. Combustion section Burning section , thus obtaining the initial segment mask;

[0017] S14. To improve the robustness of the segment mask and ensure time continuity, the initial segment mask is modified... Morphological corrections and connected component filtering are performed to remove small noise regions; frame quality metrics are calculated, and when the quality is below a threshold, interpolation of neighboring frame results is used to replace it, resulting in three segment masks. ;

[0018] S15, will The brightness image is concatenated with the segment mask to obtain the input tensor. As shown in the following formula:

[0019]

[0020] in, , For pre-processed Brightness image, Indicates the direction of the channel splicing; Given the height and width of the input image, This is the number of input channels.

[0021] Further, step S2 includes the following steps:

[0022] S21. To enable multi-level feature abstraction while reducing spatial resolution and expanding the receptive field, a convolutional layer with a 3×3 kernel and a stride of 2 is used to process the input tensor. Downsampling and low-level feature extraction are performed as follows:

[0023]

[0024] in," " indicates convolution operation; This is the output of the first stage convolutional layer; and For convolution weights and biases; For batch normalization; The activation function is used; the spatial size after downsampling is... ;

[0025] To perform deep feature abstraction at this intermediate scale, stacking The residual blocks form residual group 1, and the calculation formula for a single residual block is as follows:

[0026]

[0027] in, For the first One residual block input; Output for this residual block; This represents the weight of the residual block; It consists of convolution, batch normalization, and activation functions; the output of residual group 1 is as follows:

[0028]

[0029] S22. To further expand the receptive field and capture a wider range of semantic information, a convolutional layer with a 3×3 kernel and a stride of 2 is used. Downsampling The space dimensions are To learn more global feature representations at this larger scale, stacking The residual blocks form residual group 2, and its output is as follows:

[0030]

[0031] S23. To obtain a larger receptive field, the global contextual information of the image is aggregated by using a convolutional layer with a 3×3 kernel and a stride of 2. Downsampling The space dimensions are To learn more global feature representations at this large scale, stacking The residual blocks form residual group 3, and its output is ,in , , This represents the number of feature channels in the final output of the CNN branch. The output is as follows:

[0032]

[0033] Further, step S3 includes the following steps:

[0034] S31. To convert the image into a sequence format that Transformer can process, the input tensor... According to size The blocks are divided into There are several patches, and each patch is flattened and linearly projected to obtain the initial patch representation, as shown in the following equation:

[0035]

[0036]

[0037] in, Indicates the first The initial representation of each patch; Indicates the first The flattened vector of each patch; Represents a linear projection matrix; Indicates projection offset; Indicates the side length of the patch; Indicates the dimension of the Transformer model; , This indicates the number of patches obtained from the segmentation; a patch is the input image... The image is divided into fixed-size patches in spatial dimension. Each patch contains... 1 pixel; This means flattening each patch into a one-dimensional vector and projecting it onto a dimension using a linear mapping. The feature space;

[0038] S32. To enable the Transformer to perceive the prior physical location of the furnace, in addition to traditional position coding... Add segment embedding on the basis The segment-aware location code is obtained as follows:

[0039]

[0040] in, Indicates the first The location encoding of each patch; Indicates patch The corresponding rank identification function, =Drying section, = Combustion section, =Ember segment, determined by the segment mask of the center pixel of the patch. Determine its value; Represents a learnable segment embedding vector; Indicates the first The initial token representation of each patch before entering the Transformer encoder;

[0041] S33. To generate a global image representation, insert a learnable class tag before the token sequence. This is used for the final global representation output, as shown in the following formula:

[0042]

[0043] in, Represents the learnable class label initialization vector; This represents the initial token sequence for the Transformer;

[0044] S34, Initial token sequence Input to The layered Transformer encoder contains segment-biased multi-head self-attention and feedforward networks in each layer, and uses residual connections and layer normalization.

[0045] Further, step S34 specifically includes:

[0046] No. The first layer One point of focus, as shown in the following formula:

[0047]

[0048]

[0049] in, This represents the current layer's input token sequence; This represents the linear mapping weight of the k-th head; Represents the rank bias matrix, if If they belong to the same physical stage (drying, burning, or smoldering stage), then ,otherwise 0; Indicate each Dimensions Represents the number of attention heads; This represents the rank weighting coefficient, which can be learned.

[0050] Concatenate and linearly map the outputs of all heads:

[0051]

[0052] in, This represents the output mapping matrix after multiple heads are concatenated;

[0053] The output of each layer is iteratively output using the following formula:

[0054]

[0055]

[0056] in, Represents a two-layer feedforward network ;

[0057] go through After layer iterations, the final output global feature sequence is obtained. :

[0058]

[0059] in, It is divided into two parts: class tag. and patch token sequence .

[0060] Further, step S4 includes the following steps:

[0061] S41. Multi-scale feature maps Flattened into a sequence in the spatial dimension And construct a rank matching matrix This is used to associate the spatial location of the CNN with the ViT patch at different physical segments, ensuring that subsequent attention fusion occurs within the same physical segment. (Sequence) As shown in the following formula:

[0062]

[0063] in, for In spatial location index The eigenvector at that location;

[0064] Construct a rank matching matrix This is used to associate the spatial location of the CNN with the segment of the ViT patch, as shown in the following formula:

[0065]

[0066] S42. To inject the global context information of ViT into the multi-scale feature map of CNN, perform CNN→ViT cross-attention (injecting global context into local features); utilize the patch token sequence obtained by the ViT network. To update multi-scale feature maps The updated feature vector Arranged in the original spatial order, the enhanced multi-scale feature maps are obtained. ;

[0067] S43. To inject local detail information from the CNN into the ViT token, perform ViT→CNN cross-attention (injecting local details into the global token); use each patch token sequence from ViT. Query the multi-scale feature maps aggregated by CNN to obtain the updated patch token sequence. ;

[0068] S44. To enable the class tag to aggregate patch token information with injected local details, perform self-attention aggregation to obtain the enhanced class tag. ;

[0069] S45. To adaptively fuse the enhanced features from the two branches, a gating mechanism is used to fuse the enhanced multi-scale feature maps. and enhanced class tag Obtain the fused feature vector .

[0070] Further, step S42 includes the following steps:

[0071] S421. Define query Q as the feature sequence of the CNN network. Each of them for In spatial location The vector; key values ​​K and V are the patch token sequences of the ViT network. Each of them For the first The patch token output by the layer. ;

[0072] S422, Transfer each CNN feature Projection as query vector As shown in the following formula:

[0073]

[0074] in, To query the projection matrix;

[0075] Each Projection as key vector Sum value vector As shown in the following formula:

[0076]

[0077]

[0078] in, For the key-value projection matrix, ; For each attention head's feature dimension, set ;

[0079] S423. Introduce segment bias and calculate the scalar attention score as follows:

[0080]

[0081] in, This is a learnable rank bias coefficient;

[0082] S424, for each fixed In all Perform softmax normalization to obtain scalar attention weights, as shown in the following formula:

[0083]

[0084] S425. The weighted sum of the ViT value vectors is used to obtain the aggregate vector, which is then projected back to the original dimensions of the CNN network and residual connected, as shown in the following formula:

[0085]

[0086] in To output the projection matrix; ;

[0087] S426. Update the feature vector. Arranged in the original spatial order, the enhanced multi-scale feature maps are obtained. .

[0088] Further, step S43 includes the following steps:

[0089] S431. Define query Q as the first... in the ViT network. Layer original patch token sequence Each of them Output the patch token for layer L, with dimension L. Key values ​​K and V represent the updated CNN network feature sequences. ;

[0090] S432. Project the ViT token into a query vector, as follows:

[0091]

[0092] in, To query the projection matrix; ;

[0093] The updated CNN network features are projected as key-value vectors, as shown in the following equation:

[0094]

[0095]

[0096] in, , The key-value projection matrix; , ; For the feature dimension of the attention head;

[0097] S433. Introduce segment bias and calculate the scalar attention score as follows:

[0098]

[0099] in, This is the learnable rank bias coefficient;

[0100] S434, for each fixed In all Perform softmax normalization to obtain scalar attention weights, as shown in the following formula:

[0101]

[0102] S435. The weighted summation of the CNN network's value vector yields an aggregated vector, which is then projected back to the original dimensions of the ViT network and residual connected, as shown in the following equation:

[0103]

[0104] in To output the projection matrix; ;

[0105] S436, Updated Form a new ViT patch token sequence .

[0106] Further, step S44 includes the following steps:

[0107] S441. Constructing the recombination sequence ;

[0108] S442, will Given a multi-head self-attention module (MHA) and a feedforward network (FFN), perform a standard Transformer encoding computation as follows:

[0109]

[0110]

[0111] The structures of MHA and FFN are consistent with the encoder layer in the ViT network;

[0112] S443, From the output sequence The first vector is extracted and used as an enhanced class label that integrates global semantics and local details. .

[0113] Further, step S45 includes the following steps:

[0114] S451, to Perform global average pooling to obtain the channel statistics vector, and then project it onto the fusion dimension. As shown in the following formula:

[0115]

[0116]

[0117] in, This represents the vector obtained by taking the global average of the updated CNN features; Indicates will Projected to the fusion dimension The result; For hyperparameters; To fuse feature vector dimensions;

[0118] S452, mark the enhanced class. Projecting onto the same fusion dimension, as shown below:

[0119]

[0120] in, Indicates projection onto the fusion dimension The ViT vector; For hyperparameters;

[0121] S453. Based on the above steps, generate the gated weight vector as follows:

[0122]

[0123] in, , For gating layer parameters, For the Sigmoid function; ;

[0124] S454. Adaptive weighted feature fusion using gated weights, as shown in the following formula:

[0125]

[0126] in, This represents element-wise multiplication, where 1 represents a vector containing only 1s. ;

[0127] Further, step S5 includes the following steps:

[0128] S51. Obtain the position of the burning line by inputting the pixel segmentation results, and calculate the normalized height distribution of the burning line in the current frame. , compared with the reference distribution under normal operating conditions By comparison, the KL divergence metric is used to measure the differences, and a distribution consistency penalty term is obtained. As shown in the following formula:

[0129]

[0130] in, Indicates the horizontal position of the current frame's burning line. Normalized height distribution value at; This indicates the horizontal position of the combustion line under normal operating conditions. Reference height distribution value at the location; Indicates the number of horizontal sampling points; This represents a small constant to prevent division by zero or logarithmic singularity;

[0131] S52. Calculate the average brightness of the pixel region of the burning segment by inputting the segmented burning segment and the burnt-out segment pixel regions respectively. Average brightness of the burnt-out pixel area And calculate its brightness ratio. Constrain it to be within a safe range Within, a brightness ratio penalty term is obtained. As shown in the following formula:

[0132]

[0133] in, This indicates the lower limit threshold for brightness. This indicates the upper limit threshold for brightness ratio; The function ensures that only when A non-zero penalty is incurred when the safe zone is exceeded;

[0134] S53, through combustion line height distribution Total variation regularization is used to suppress unreasonable and drastic fluctuations in the combustion line, resulting in a smoothness penalty term. As shown in the following formula:

[0135]

[0136] S54. Weight the penalty terms from the above steps to obtain the comprehensive physical constraint loss, as shown in the following formula:

[0137]

[0138] in, The weighting coefficients representing the losses of the three physical constraints are determined through experiments or experience.

[0139] Further, step S6 includes the following steps:

[0140] S61. To simultaneously optimize classification accuracy, physical plausibility, and model generalization ability, a total loss function for the training phase is constructed as follows:

[0141]

[0142] in, ( , ) is the cross-entropy loss function; The one-hot encoded label vector for the actual working conditions; The class probability distribution vector predicted by the model (output by the Softmax layer); physical constraint loss. This indicates that the prediction results are reasonable in terms of physical laws; This represents the vector of all learnable parameters of the model; It represents the square of the L2 norm of all learnable weight parameters of the model; and These are the weighting coefficients, which are optimized using the validation set.

[0143] S62, Fusion Feature Vector The predicted probability of the working condition is generated by the fully connected layer and the Softmax layer, as shown in the following formula:

[0144]

[0145] in, To output the weight matrix; For output bias terms; Given the predicted class probability distribution, the final output class is:

[0146]

[0147] in: Indicate category The predicted probability, For predicting category indexes.

[0148] This invention also provides a solid waste incineration condition identification system based on hybrid convolution and physical constraints. The system executes the above-described method during operation and includes the following modules:

[0149] The pixel banding module is used to deploy furnace cameras to acquire raw images, calibrate and determine the mapping between pixel rows and the physical height of the furnace; preprocess the raw images and divide them into pixel bands according to the physical structure of the furnace to obtain the input tensor. ;

[0150] The local feature extraction module is used to extract the input tensor using a CNN network. Multi-scale local features, output multi-scale feature map ;

[0151] The global feature extraction module is used to extract input tensors using the ViT network. The global context features are used to output the global feature sequence. ;

[0152] The feature fusion module uses the divided pixel bands as priors to guide bidirectional cross-attention and fuse multi-scale feature maps. With global feature sequences Obtain the fused feature vector ;

[0153] The physical constraint module is used to introduce physical constraint regularization that conforms to the combustion law of solid waste incineration during the training phase;

[0154] The output module is used to construct the joint loss function of the model and output the identification results of solid waste incineration conditions.

[0155] The advantages of this invention are:

[0156] (1) This invention extracts local and global features of solid waste incineration flame images by using CNN network and ViT network respectively, which takes into account both details and global dependencies, and improves the accuracy of working condition classification and recognition.

[0157] (2) By dividing the pixel band, the present invention explicitly introduces the prior physical structure, thereby improving the adaptability and robustness of the model to different working conditions.

[0158] (3) The present invention outputs a stable quantitative classification result, which can be directly used for closed-loop regulation of the furnace control system, thereby improving the stability of the solid waste incineration process. Attached Figure Description

[0159] Figure 1 This is a schematic diagram of the solid waste incineration condition identification method based on hybrid convolution and physical constraints according to an embodiment of the present invention.

[0160] Figure 2 This is a schematic diagram comparing the solid waste incineration operation condition identification of this invention with other model algorithms. Detailed Implementation

[0161] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0162] Example 1

[0163] This embodiment provides a solid waste incineration condition identification method based on hybrid convolution and physical constraints, aiming to solve the problems of low classification accuracy and weak adaptability of solid waste incineration conditions in existing technologies, and to achieve accurate monitoring and intelligent control of the solid waste incineration process. The specific implementation process is as follows: Figure 1 As shown, it includes the following steps:

[0164] S1. Deploy furnace cameras to acquire raw images, calibrate and determine the mapping between pixel rows and the physical height of the furnace; preprocess the raw images and divide them into pixel bands according to the physical structure of the furnace to obtain the input tensor. The specific implementation method includes the following steps:

[0165] S11. During the deployment phase of the furnace camera, the mapping between pixel rows and physical height is determined through geometric calibration or manual calibration, and linear mapping is performed. And set the region of interest in the image. ,in: Indicates the physical height of the pixel within the furnace; Indicates the vertical row index of a pixel in the image; Represents the linear mapping coefficients obtained from calibration; The region of interest mask is used to define the effective imaging area of ​​the furnace.

[0166] S12. Perform distortion correction, resolution unification, luminance channel extraction, contrast enhancement and noise suppression on the acquired images to reduce the impact of lighting changes and equipment differences;

[0167] S13. To encode the prior knowledge of the furnace physical structure into an image mask, the ROI region is divided into drying sections using a fixed calibration method or an adaptive combustion line detection method. Combustion section Burning section The initial segment mask is obtained; the segment division method can adopt either a fixed calibration method or an adaptive detection method, as detailed below:

[0168] (1) Fixed calibration method: Determine the pixel row index during the calibration period. Directly divide the ROI vertically into Suitable for scenarios where the camera is fixed and reliably calibrated. Among them: The pixel row representing the boundary between the drying section and the combustion section; This represents the pixel row that marks the boundary between the burning section and the burnt-out section. Indicates the height of the ROI (in pixels).

[0169] (2) Adaptive detection method: Binarize the brightness channel and calculate the combustion top column by column. The combustion curve was obtained by Gaussian smoothing, and then its median statistical value was used. Adding / subtracting empirical offsets to determine segment boundaries This method is suitable for scenarios where the camera position or furnace conditions change significantly. Specifically: Indicates being listed The top pixel row of the upper flame area; Represents the horizontal column index of the image; Represents all columns Median value; offset represents the pixel row difference determined empirically based on the physical structure of the furnace.

[0170] S14. Initial segment mask Morphological corrections and connected component filtering are performed to remove small noise regions; frame quality metrics (effective activated pixel ratio) are calculated, and when the quality is below a threshold, interpolation of neighboring frame results is used to replace it, resulting in three segment masks. Three segment masks are used to calculate the statistics for physical constraints and classification (segment average brightness). Brightness ratio Combustion line height distribution (etc.), serving as standard inputs for downstream modules. Among them: A binary mask representing the drying section, combustion section, and burnout section; This represents the average value of the bit brightness (unit: gray level). Indicates the brightness ratio, used to characterize the difference in brightness between the combustion stage and the burnout stage; Indicates the burning line is in the column Normalized height.

[0171] S15, will The brightness image is concatenated with the segment mask to obtain the input tensor. As shown in the following formula:

[0172]

[0173] in, , For pre-processed Brightness image, Indicates the direction of the channel splicing; Given the height and width of the input image, The input channel number is 4 in this embodiment, which includes luminance and 3 masks.

[0174] S2. Use a CNN network to extract the input tensor. Multi-scale local features, output multi-scale feature map The specific implementation method includes the following steps:

[0175] S21. To enable multi-level feature abstraction while reducing spatial resolution and expanding the receptive field, a convolutional layer with a 3×3 kernel and a stride of 2 is used to process the input tensor. Downsampling and low-level feature extraction are performed as follows:

[0176]

[0177] in," " indicates convolution operation; This is the output of the first stage convolutional layer; and For convolution weights and biases; For batch normalization; The activation function is used; the spatial size after downsampling is... ;

[0178] To perform deep feature abstraction at this intermediate scale, stacking The residual blocks form residual group 1, and the calculation formula for a single residual block is as follows:

[0179]

[0180] in, For the first One residual block input; Output for this residual block; This represents the weight of the residual block; It consists of convolution, batch normalization, and activation functions; the output of residual group 1 is as follows:

[0181]

[0182] S22. To further expand the receptive field and capture a wider range of semantic information, a convolutional layer with a 3×3 kernel and a stride of 2 is used. Downsampling The space dimensions are To learn more global feature representations at this larger scale, stacking The residual blocks form residual group 2, and its output is as follows:

[0183]

[0184] S23. To obtain a larger receptive field, the global contextual information of the image is aggregated by using a convolutional layer with a 3×3 kernel and a stride of 2. Downsampling The space dimensions are To learn more global feature representations at this large scale, stacking The residual blocks form residual group 3, and its output is ,in , , This refers to the number of feature channels in the final output of the CNN branch. The output is as follows:

[0185] .

[0186] S3. Extracting input tensors using a ViT network. The global context features are used to output the global feature sequence. The specific implementation method includes the following steps:

[0187] S31. To convert the image into a sequence format that Transformer can process, the input tensor... According to size The blocks are divided into There are several patches, and each patch is flattened and linearly projected to obtain the initial patch representation, as shown in the following equation:

[0188]

[0189]

[0190] in, Indicates the first The initial representation of each patch; Indicates the first The flattened vector of each patch; Represents a linear projection matrix; Indicates projection offset; Indicates the side length of the patch; Indicates the dimension of the Transformer model; , This indicates the number of patches obtained from the segmentation; a patch is the input image... The image is divided into fixed-size patches in spatial dimension. Each patch contains... 1 pixel; This means flattening each patch into a one-dimensional vector and projecting it onto a dimension using a linear mapping. The feature space;

[0191] S32. To enable the Transformer to perceive the prior physical location of the furnace, in addition to traditional position coding... Add segment embedding on the basis The segment-aware location code is obtained as follows:

[0192]

[0193] in, Indicates the first The location encoding of each patch; Indicates patch The corresponding rank identification function, =Drying section, = Combustion section, =Ember segment, determined by the segment mask of the center pixel of the patch. Determine its value; Represents a learnable segment embedding vector; Indicates the first The initial token representation of each patch before entering the Transformer encoder;

[0194] S33. To generate a global image representation, insert a learnable class tag before the token sequence. This is used for the final global representation output, as shown in the following formula:

[0195]

[0196] in, Represents the learnable class label initialization vector; This represents the initial token sequence for the Transformer;

[0197] S34, Initial token sequence Input to A layered Transformer encoder, each layer containing segment-biased multi-head self-attention and feedforward networks, and using residual connections and layer normalization. The specific operations are as follows:

[0198] No. The first layer One point of focus, as shown in the following formula:

[0199]

[0200]

[0201] in, This represents the current layer's input token sequence; This represents the linear mapping weight of the k-th head; Represents the rank bias matrix, if If they belong to the same physical tier, then ,otherwise 0; Indicate each Dimensions Represents the number of attention heads; This represents the rank weighting coefficient, which can be learned.

[0202] Concatenate and linearly map the outputs of all heads:

[0203]

[0204] in, This represents the output mapping matrix after multiple heads are concatenated;

[0205] The output of each layer is iteratively output using the following formula:

[0206]

[0207]

[0208] in, Represents a two-layer feedforward network ;

[0209] go through After layer iterations, the final output global feature sequence is obtained. :

[0210]

[0211] in, It is divided into two parts: class tag. and patch token sequence .

[0212] S4. To achieve deep interaction between CNN branches (local details) and ViT branches (global context), this embodiment proposes a Segment-aware Cross-Attention Fusion (SCAF) module. This module uses segment masks generated by image preprocessing and pixel segmentation as priors to guide bidirectional cross-attention (CNN→ViT and ViT→CNN), enabling local and global features to complement each other at the spatial level, thereby improving classification robustness and sensitivity to physical segments.

[0213] Using the segmented pixel bands as priors, bidirectional cross-attention is guided to fuse multi-scale feature maps. With global feature sequences Obtain the fused feature vector The specific implementation method includes the following steps:

[0214] S41. Multi-scale feature maps Flattened into a sequence in the spatial dimension And construct a rank matching matrix This is used to associate the spatial location of the CNN with the ViT patch at different levels, ensuring that subsequent attention fusion occurs within the same physical segment; sequence As shown in the following formula:

[0215]

[0216] in, for In spatial location index The eigenvector at that location;

[0217] Construct a rank matching matrix This is used to associate the spatial location of the CNN with the segment of the ViT patch, as shown in the following formula:

[0218]

[0219] S42. To inject the global context information of ViT into the multi-scale feature map of CNN, perform CNN→ViT cross-attention; utilize the patch token sequence obtained by the ViT network. To update multi-scale feature maps The updated feature vector Arranged in the original spatial order, the enhanced multi-scale feature maps are obtained. Specifically, it includes the following steps:

[0220] S421. Define query Q as the feature sequence of the CNN network. Each of them for In spatial location The vector; key values ​​K and V are the patch token sequences of the ViT network. Each of them For the first The patch token output by the layer. ;

[0221] S422, Transfer each CNN feature Projection as query vector As shown in the following formula:

[0222]

[0223] in, To query the projection matrix;

[0224] Each Projection as key vector Sum value vector As shown in the following formula:

[0225]

[0226]

[0227] in, For the key-value projection matrix, ; For each attention head's feature dimension, set The index of q indicates the identity of the query initiator, and the indexes of k and v indicate the identity of the queryee.

[0228] S423. Introduce segment bias and calculate scalar attention score for the first segment. u The CNN position and the i A ViTtoken, as shown in the following formula:

[0229]

[0230] in, This is a learnable rank bias coefficient;

[0231] S424, for each fixed In all Perform softmax normalization to obtain scalar attention weights, as shown in the following formula:

[0232]

[0233] S425. The weighted sum of the ViT value vectors is used to obtain the aggregate vector, which is then projected back to the original dimensions of the CNN network and residual connected, as shown in the following formula:

[0234]

[0235] in To output the projection matrix; ;

[0236] S426. Update the feature vector. Arranged in the original spatial order, the enhanced multi-scale feature maps are obtained. .

[0237] S43. To inject local detail information from the CNN into the ViT token, perform ViT→CNN cross-attention; use each patch token sequence from ViT. Query the multi-scale feature maps aggregated by CNN to obtain the updated patchtoken sequence. Specifically, it includes the following steps:

[0238] S431. Define query Q as the first... in the ViT network. Layer original patch token sequence Each of them Output the patch token for layer L, with dimension L. Key values ​​K and V represent the updated CNN network feature sequences. ;

[0239] S432. Project the ViT token into a query vector, as follows:

[0240]

[0241] in, To query the projection matrix; ;

[0242] The updated CNN network features are projected as key-value vectors, as shown in the following equation:

[0243]

[0244]

[0245] in, , The key-value projection matrix; , ; For the feature dimension of the attention head;

[0246] S433. Introduce segment bias, calculate scalar attention score, and apply it to the first... i The ViT token and the first u The CNN positions are as follows:

[0247]

[0248] in, This is the learnable rank bias coefficient;

[0249] S434, for each fixed In all Perform softmax normalization to obtain scalar attention weights, as shown in the following formula:

[0250]

[0251] S435. The weighted summation of the CNN network's value vector yields an aggregated vector, which is then projected back to the original dimensions of the ViT network and residual connected, as shown in the following equation:

[0252]

[0253] in To output the projection matrix; ;

[0254] S436, Updated Form a new ViT patch token sequence .

[0255] S44. To enable the class tag to aggregate patch token information with injected local details, perform self-attention aggregation to obtain the enhanced class tag. Specifically, it includes the following steps:

[0256] S441. Constructing the recombination sequence ;

[0257] S442, will Given a multi-head self-attention module (MHA) and a feedforward network (FFN), perform a standard Transformer encoding computation as follows:

[0258]

[0259]

[0260] The structures of MHA and FFN are consistent with the encoder layer in the ViT network;

[0261] S443, From the output sequence The first vector is extracted and used as an enhanced class label that integrates global semantics and local details. .

[0262] S45. To adaptively fuse the enhanced features from the two branches, a gating mechanism is used to fuse the enhanced multi-scale feature maps. and enhanced class tag Obtain the fused feature vector Specifically, it includes the following steps:

[0263] S451, to Perform global average pooling to obtain the channel statistics vector, and then project it onto the fusion dimension. As shown in the following formula:

[0264]

[0265]

[0266] in, This represents the vector obtained by taking the global average of the updated CNN features; Indicates will Projected to the fusion dimension The result; For hyperparameters; To fuse feature vector dimensions;

[0267] S452, mark the enhanced class. Projecting onto the same fusion dimension, as shown below:

[0268]

[0269] in, Indicates projection onto the fusion dimension The ViT vector; For hyperparameters;

[0270] S453. Based on the above steps, generate the gated weight vector as follows:

[0271]

[0272] in, , For gating layer parameters, For the Sigmoid function; ;

[0273] S454. Adaptive weighted feature fusion using gated weights, as shown in the following formula:

[0274]

[0275] in, This represents element-wise multiplication, where 1 represents a vector containing only 1s. .

[0276] Final fused feature vector This vector integrates local details enhanced by global context (from CNN branches) and global semantics enhanced by local details (from ViT branches), and achieves optimal combination through a learnable gating mechanism, providing more discriminative input for subsequent classification.

[0277] S5. To ensure that the classification results conform to the combustion law of solid waste incineration, physical constraint regularization that conforms to the combustion law of solid waste incineration is introduced during the training phase; the specific implementation includes the following steps:

[0278] S51. Obtain the position of the burning line by inputting the pixel segmentation results, and calculate the normalized height distribution of the burning line in the current frame. , compared with the reference distribution under normal operating conditions By comparison, the KL divergence metric is used to measure the differences, and a distribution consistency penalty term is obtained. As shown in the following formula:

[0279]

[0280] in, Indicates the horizontal position of the current frame's burning line. Normalized height distribution value at; This indicates the horizontal position of the combustion line under normal operating conditions. Reference height distribution value at the location; Indicates the number of horizontal sampling points; This represents a small constant to prevent division by zero or logarithmic singularity;

[0281] S52. Calculate the average brightness of the pixel region of the burning segment by inputting the segmented burning segment and the burnt-out segment pixel regions respectively. Average brightness of the burnt-out pixel area And calculate its brightness ratio. Constrain it to be within a safe range Within, a brightness ratio penalty term is obtained. As shown in the following formula:

[0282]

[0283] in, This indicates the lower limit threshold for brightness. This indicates the upper limit threshold for brightness ratio; The function ensures that only when A non-zero penalty is incurred when the safe zone is exceeded;

[0284] S53, through combustion line height distribution Total variation regularization is used to suppress unreasonable and drastic fluctuations in the combustion line, resulting in a smoothness penalty term. As shown in the following formula:

[0285]

[0286] S54. Weight the penalty terms from the above steps to obtain the comprehensive physical constraint loss, as shown in the following formula:

[0287]

[0288] in, The weighting coefficients representing the losses of the three physical constraints are determined through experiments or experience.

[0289] S6. Construct the joint loss function of the model and output the identification results of solid waste incineration conditions. The specific implementation includes the following steps:

[0290] S61. To simultaneously optimize classification accuracy, physical plausibility, and model generalization ability, a total loss function for the training phase is constructed as follows:

[0291]

[0292] in, ( , ) is the cross-entropy loss function; The one-hot encoded label vector for the actual working conditions; The model predicts the class probability distribution vector; physical constraint loss. This indicates that the prediction results are reasonable in terms of physical laws; This represents the vector of all learnable parameters of the model; It represents the square of the L2 norm of all learnable weight parameters of the model; and These are the weighting coefficients, which are optimized using the validation set.

[0293] S62, Fusion Feature Vector The predicted probability of the working condition is generated by the fully connected layer and the Softmax layer, as shown in the following formula:

[0294]

[0295] in, To output the weight matrix; For output bias terms; Given the predicted class probability distribution, the final output class is:

[0296]

[0297] in: Indicate category The predicted probability, For predicting category indexes.

[0298] Through simulation experiments, the method provided in this embodiment is compared with the accuracy of the EfficientNet model, ResNet model, VGG16net model, and AlexNet model in monitoring the classification results of solid waste incineration conditions. Figure 2 As shown, the classification accuracy of the method provided in this embodiment is higher than that of other models.

[0299] Example 2

[0300] It should be further explained that, based on the same inventive concept, this invention also provides a solid waste incineration condition identification system based on hybrid convolution and physical constraints. When the system is running, it executes the method described in Example 1, including the following modules:

[0301] The pixel banding module is used to deploy furnace cameras to acquire raw images, calibrate and determine the mapping between pixel rows and the physical height of the furnace; preprocess the raw images and divide them into pixel bands according to the physical structure of the furnace to obtain the input tensor. ;

[0302] The local feature extraction module is used to extract the input tensor using a CNN network. Multi-scale local features, output multi-scale feature map ;

[0303] The global feature extraction module is used to extract input tensors using the ViT network. The global context features are used to output the global feature sequence. ;

[0304] The feature fusion module uses the divided pixel bands as priors to guide bidirectional cross-attention and fuse multi-scale feature maps. With global feature sequences Obtain the fused feature vector ;

[0305] The physical constraint module is used to introduce physical constraint regularization that conforms to the combustion law of solid waste incineration during the training phase;

[0306] The output module is used to construct the joint loss function of the model and output the identification results of solid waste incineration conditions.

[0307] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying solid waste incineration conditions based on hybrid convolution and physical constraints, characterized in that, Includes the following steps: S1. Deploy furnace cameras to acquire raw images and calibrate the mapping between pixel rows and the physical height of the furnace; preprocess the raw images and divide them into pixel bands according to the physical structure of the furnace; divide the region of interest (ROI) into drying sections using a fixed calibration method or an adaptive combustion line detection method. Combustion section Burning section Obtain the initial segment mask; [The sentence fragment "for the initial segment mask" appears to be incomplete and requires further context.] Perform morphological corrections and connected component filtering to remove small noise regions; Calculate the frame quality index. When the quality is below a threshold, interpolate the results of neighboring frames to obtain three segment masks. ;Will The brightness image is concatenated with the segment mask to obtain the input tensor. ; S2. Use a CNN network to extract the input tensor. Multi-scale local features, output multi-scale feature map ; S3. Extracting input tensors using ViT networks. The global context features are used to output the global feature sequence. ; S4. Using the segmented pixel bands as priors, guide bidirectional cross-attention and fuse multi-scale feature maps. With global feature sequences Obtain the fused feature vector The steps include: S41. Multi-scale feature maps Flattened into a sequence in the spatial dimension And construct a rank matching matrix ; S42. Perform CNN→ViT cross-attention. During this process, segment bias is introduced, and scalar attention score is calculated. S43. Perform ViT→CNN cross-attention. During this process, segment bias is introduced, and scalar attention scores are calculated. S5. Introduce physical constraint regularization that conforms to the combustion law of solid waste incineration during the training phase; S6. Construct the joint loss function of the model and output the identification results of solid waste incineration conditions.

2. The solid waste incineration condition identification method based on hybrid convolution and physical constraints according to claim 1, characterized in that, Step S1 further includes: During the deployment phase of the furnace camera, the mapping between pixel rows and physical height is determined through geometric calibration or manual calibration, and linear mapping is performed. And set the region of interest in the image. ,in: Indicates the physical height of the pixel within the furnace; Indicates the vertical row index of a pixel in the image; Represents the linear mapping coefficients obtained from calibration; The region of interest mask is used to define the effective imaging area of ​​the furnace. The acquired images undergo distortion correction, resolution unification, luminance channel extraction, contrast enhancement, and noise suppression to reduce the impact of lighting variations and equipment differences. The input tensor As shown in the following formula: in, , For pre-processed Brightness image, Indicates the direction of the channel splicing; Given the height and width of the input image, This is the number of input channels.

3. The solid waste incineration condition identification method based on hybrid convolution and physical constraints according to claim 2, characterized in that, Step S2 includes the following steps: S21. Apply a 3×3 convolutional layer with a stride of 2 to the input tensor. Downsampling and low-level feature extraction are performed as follows: in," " indicates convolution operation; This is the output of the first stage convolutional layer; and For convolution weights and biases; For batch normalization; The activation function is used; the spatial size after downsampling is... ; Stacking The residual blocks form residual group 1, and the calculation formula for a single residual block is as follows: in, For the first One residual block input; Output for this residual block; This represents the weight of the residual block; It consists of convolution, batch normalization, and activation functions; the output of residual group 1 is as follows: S22. Using a convolutional layer with a 3×3 kernel and a stride of 2, Downsampling The space dimensions are Stacking The residual blocks form residual group 2, and its output is as follows: S23. Using a convolutional layer with a 3×3 kernel and a stride of 2, Downsampling The space dimensions are Stacking The residual blocks form residual group 3, and its output is ,in , , This refers to the number of feature channels in the final output of the CNN branch. The output is as follows: 。 4. The solid waste incineration condition identification method based on hybrid convolution and physical constraints according to claim 2, characterized in that, Step S3 includes the following steps: S31, Input Tensor According to size The blocks are divided into There are several patches, and each patch is flattened and linearly projected to obtain the initial patch representation, as shown in the following equation: in, Indicates the first The initial representation of each patch; Indicates the first The flattened vector of each patch; Represents a linear projection matrix; Indicates projection offset; Indicates the side length of the patch; Indicates the dimension of the Transformer model; , This indicates the number of patches obtained from the segmentation; a patch is the input image... The image is divided into fixed-size patches in spatial dimension. Each patch contains... 1 pixel; This means flattening each patch into a one-dimensional vector and projecting it onto a dimension using a linear mapping. The feature space; S32, In traditional position encoding Add segment embedding on the basis The positional encoding for rank perception is obtained as follows: in, Indicates the first The location encoding of each patch; Indicates patch The corresponding rank identification function, =Drying section, = Combustion section, =Ember segment, determined by the segment mask of the center pixel of the patch. Determine its value; Represents a learnable segment embedding vector; Indicates the first The initial token representation of each patch before entering the Transformer encoder; S33. Insert a learnable class token before the token sequence. This is used for the final global representation output, as shown in the following formula: in, Represents the learnable class label initialization vector; This represents the initial token sequence for the Transformer; S34, Initial token sequence Input to The layered Transformer encoder contains segment-biased multi-head self-attention and feedforward networks in each layer, and uses residual connections and layer normalization.

5. The solid waste incineration condition identification method based on hybrid convolution and physical constraints according to claim 4, characterized in that, Step S34 specifically involves: No. The first layer One point of focus, as shown in the following formula: in, This represents the current layer's input token sequence; This represents the linear mapping weight of the k-th head; Represents the rank bias matrix, if If they belong to the same physical tier, then ,otherwise 0; Indicate each Dimensions Represents the number of attention heads; This represents the rank weighting coefficient, which can be learned. Concatenate and linearly map the outputs of all heads: in, This represents the output mapping matrix after multiple heads are concatenated; The output of each layer is iteratively output using the following formula: in, Represents a two-layer feedforward network ; go through After layer iterations, the final output global feature sequence is obtained. : in, It is divided into two parts: class tag. and patch token sequence .

6. The solid waste incineration condition identification method based on hybrid convolution and physical constraints according to claim 5, characterized in that, Step S4 further includes: S41, the sequence As shown in the following formula: in, for In spatial location index The eigenvector at that location; The aforementioned rank matching matrix As shown in the following formula: S42. Perform CNN→ViT cross-attention: Utilize the patch token sequence obtained from the ViT network. To update multi-scale feature maps The updated feature vector Arranged in the original spatial order, the enhanced multi-scale feature maps are obtained. ; S43. Perform ViT→CNN cross-attention: Use the sequence of each patch token from ViT. Query the multi-scale feature maps aggregated by CNN to obtain the updated patch token sequence. ; S44. Perform self-attention aggregation to obtain the enhanced class label. ; S45. Multi-scale feature maps enhanced by gating mechanism and enhanced class tag Obtain the fused feature vector .

7. The solid waste incineration condition identification method based on hybrid convolution and physical constraints according to claim 6, characterized in that, Step S42 includes the following steps: S421. Define query Q as the feature sequence of the CNN network. Each of them for In spatial location The vector; key values ​​K and V are the patch token sequences of the ViT network. Each of them For the first The patch token output by the layer. ; S422, Transfer each CNN feature Projection as query vector As shown in the following formula: in, To query the projection matrix; Each Projection as key vector Sum value vector As shown in the following formula: in, For the key-value projection matrix, ; For each attention head's feature dimension, set ; S423. Introduce segment bias and calculate the scalar attention score as follows: in, This is a learnable rank bias coefficient; S424, for each fixed In all Perform softmax normalization to obtain scalar attention weights, as shown in the following formula: S425. The weighted sum of the ViT value vectors is used to obtain the aggregate vector, which is then projected back to the original dimensions of the CNN network and residual connected, as shown in the following formula: in To output the projection matrix; ; S426. Update the feature vector. Arranged in the original spatial order, the enhanced multi-scale feature maps are obtained. .

8. The solid waste incineration condition identification method based on hybrid convolution and physical constraints according to claim 6, characterized in that, Step S43 includes the following steps: S431. Define query Q as the first... in the ViT network. Layer original patch token sequence Each of them Output the patch token for layer L, with dimension L. Key values ​​K and V represent the updated CNN network feature sequences. ; S432. Project the ViT token into a query vector, as follows: in, To query the projection matrix; ; The updated CNN network features are projected as key-value vectors, as shown in the following equation: in, , The key-value projection matrix; , ; For the feature dimension of the attention head; S433. Introduce segment bias and calculate the scalar attention score as follows: in, This is the learnable rank bias coefficient; S434, for each fixed In all Perform softmax normalization to obtain scalar attention weights, as shown in the following formula: S435. The weighted summation of the CNN network's value vector yields an aggregated vector, which is then projected back to the original dimensions of the ViT network and residual connected, as shown in the following equation: in To output the projection matrix; ; S436, Updated Form a new ViT patch token sequence .

9. The solid waste incineration condition identification method based on hybrid convolution and physical constraints according to claim 6, characterized in that, Step S44 includes the following steps: S441. Constructing the recombination sequence ; S442, will Given a multi-head self-attention module (MHA) and a feedforward network (FFN), perform a standard Transformer encoding computation as follows: The structures of MHA and FFN are consistent with the encoder layer in the ViT network; S443, From the output sequence The first vector is extracted and used as an enhanced class label that integrates global semantics and local details. .

10. The solid waste incineration condition identification method based on hybrid convolution and physical constraints according to claim 6, characterized in that, Step S45 includes the following steps: S451, to Perform global average pooling to obtain the channel statistics vector, and then project it onto the fusion dimension. As shown in the following formula: in, This represents the vector obtained by taking the global average of the updated CNN features; Indicates will Projected to the fusion dimension The result; For hyperparameters; To fuse feature vector dimensions; S452, mark the enhanced class. Projecting onto the same fusion dimension, as shown below: in, Indicates projection onto the fusion dimension The ViT vector; For hyperparameters; S453. Based on the above steps, generate the gated weight vector as follows: in, , For gating layer parameters, For the Sigmoid function; ; S454. Adaptive weighted feature fusion using gated weights, as shown in the following formula: in, This represents element-wise multiplication, where 1 represents a vector containing only 1s. .

11. The solid waste incineration condition identification method based on hybrid convolution and physical constraints according to claim 1, characterized in that, Step S5 includes the following steps: S51. Obtain the position of the burning line by inputting the pixel segmentation results, and calculate the normalized height distribution of the burning line in the current frame. , compared with the reference distribution under normal operating conditions By comparison, the KL divergence metric is used to measure the differences, and a distribution consistency penalty term is obtained. As shown in the following formula: in, Indicates the horizontal position of the current frame's burning line. Normalized height distribution value at; This indicates the horizontal position of the combustion line under normal operating conditions. Reference height distribution value at the location; Indicates the number of horizontal sampling points; This represents a small constant to prevent division by zero or logarithmic singularity; S52. Calculate the average brightness of the pixel region of the burning segment by inputting the segmented burning segment and the burnt-out segment pixel regions respectively. Average brightness of the burnt-out pixel area And calculate its brightness ratio. Constrain it to be within a safe range Within, a brightness ratio penalty term is obtained. As shown in the following formula: in, This indicates the lower threshold of the brightness ratio; This indicates the upper limit threshold for brightness ratio; The function ensures that only when A non-zero penalty is incurred when the safe zone is exceeded; S53, through combustion line height distribution Total variation regularization is used to suppress unreasonable and drastic fluctuations in the combustion line, resulting in a smoothness penalty term. As shown in the following formula: S54. Weight the penalty terms from the above steps to obtain the comprehensive physical constraint loss, as shown in the following formula: in, The weighting coefficients representing the losses of the three physical constraints are determined through experiments or experience.

12. The solid waste incineration condition identification method based on hybrid convolution and physical constraints according to claim 11, characterized in that, Step S6 includes the following steps: S61. Construct the total loss function for the training phase, as follows: in, ( , ) is the cross-entropy loss function; The one-hot encoded label vector for the actual working conditions; The model predicts the class probability distribution vector; physical constraint loss. This indicates that the prediction results are reasonable in terms of physical laws; This represents the vector of all learnable parameters of the model; It represents the square of the L2 norm of all learnable weight parameters of the model; and These are the weighting coefficients, which are optimized using the validation set. S62, Fusion Feature Vector The predicted probability of the working condition is generated by the fully connected layer and the Softmax layer, as shown in the following formula: in, To output the weight matrix; For output bias terms; Given the predicted class probability distribution, the final output class is: in: Indicates category The predicted probability, For predicting category indexes.

13. A solid waste incineration condition identification system based on hybrid convolution and physical constraints, characterized in that, The system, when running, executes the method as described in claim 1, comprising the following modules: The pixel banding module is used to deploy furnace cameras to acquire raw images, calibrate and determine the mapping between pixel rows and the physical height of the furnace; preprocess the raw images and divide them into pixel bands according to the physical structure of the furnace to obtain the input tensor. ; The local feature extraction module is used to extract the input tensor using a CNN network. Multi-scale local features, output multi-scale feature map ; The global feature extraction module is used to extract input tensors using the ViT network. The global context features are used to output the global feature sequence. ; The feature fusion module uses the divided pixel bands as priors to guide bidirectional cross-attention and fuse multi-scale feature maps. With global feature sequences Obtain the fused feature vector ; The physical constraint module is used to introduce physical constraint regularization that conforms to the combustion law of solid waste incineration during the training phase; The output module is used to construct the joint loss function of the model and output the identification results of solid waste incineration conditions.

Images