Method for recognizing combustion state of waste incinerator flame based on unsupervised learning

Abstract

The embodiment of the present application relates to the technical field of image processing, and particularly relates to a waste incinerator flame combustion state recognition method based on unsupervised learning. The method effectively solves the problem of lack of labeled data in waste incinerator flame combustion state recognition through an unsupervised learning framework; the mutual information maximization objective function is used to drive the model to autonomously mine the essential feature correlation in the flame image, and the correspondence between the combustion state and the visual feature can be established without relying on manual labeling; through the dual label output of principal clustering and hyper-clustering, hierarchical description of the combustion state is realized, which not only retains the macro working condition classification ability, but also captures the micro dynamic change characteristics. It can be seen that the embodiment of the present application reduces the data labeling cost while enhancing the adaptability of the model to the complex combustion scene, and can realize fine monitoring and recognition of the combustion state of the incinerator.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically, to a method for recognizing the combustion status of a waste incinerator flame based on unsupervised learning. Background Technology

[0002] In the field of waste-to-energy incineration, real-time monitoring and control of flame combustion status is a core element in ensuring incineration efficiency and achieving pollutant emission standards.

[0003] However, traditional manual monitoring methods rely on operators' experience to visually judge flame morphology, brightness, and other characteristics, resulting in significant subjective judgment bias and lag, making it difficult to meet the real-time and accurate status feedback requirements of automatic combustion control systems. Furthermore, existing image analysis-based automated monitoring technologies typically employ fixed threshold segmentation or feature engineering methods to establish combustion state classification models by extracting statistical quantities such as flame area and color distribution. However, this method faces the bottleneck of low classification accuracy during implementation. Moreover, machine learning solutions for identifying the combustion status of waste incinerator flames are limited by the inherent challenges of combustion state labeling. For example, manual labeling struggles to accurately classify discrete categories, and different operators exhibit significant subjective differences in their labeling standards.

[0004] Therefore, how to achieve precise monitoring and identification of the combustion status of incinerators is a technical challenge that needs to be overcome. Summary of the Invention

[0005] In view of this, the present invention provides a method for identifying the flame combustion status of a waste incinerator based on unsupervised learning.

[0006] To achieve the above objectives, embodiments of the present invention provide a method for identifying the flame combustion status of a waste incinerator based on unsupervised learning. The method is applied to a flame combustion status identification system and includes:

[0007] Obtain unlabeled flame combustion sample images of sample waste incinerators;

[0008] The unlabeled flame combustion sample images are input into the original combustion state clustering recognition model to obtain the joint probability matrix of the combustion state of the sample waste incinerator, and the mutual information variables are determined based on the joint probability matrix of the combustion state.

[0009] The original combustion state clustering identification model is trained in the direction of maximizing the mutual information variables until it meets the preset convergence condition, thereby obtaining the target combustion state clustering identification model.

[0010] A current flame combustion image of the target waste incinerator is acquired, and the current flame combustion image is input into the target combustion state clustering identification model to obtain the combustion state clustering identification result of the target waste incinerator; wherein, the combustion state clustering identification result includes a main cluster identification label and a super cluster identification label.

[0011] This invention effectively addresses the problem of insufficient labeled data in identifying the combustion status of waste incinerator flames through an unsupervised learning framework. It utilizes a mutual information maximization objective function to drive the model to autonomously mine essential feature associations in flame images, establishing a correspondence between combustion status and visual features without relying on manual annotation. Through dual-label output of principal clustering and super-clustering, a hierarchical description of the combustion status is achieved, preserving both macroscopic operational condition classification capabilities and capturing microscopic dynamic changes. Therefore, this invention reduces data annotation costs while enhancing the model's adaptability to complex combustion scenarios, enabling refined monitoring and identification of incinerator combustion status.

[0012] Preferably, before inputting the unlabeled flame combustion sample image into the original combustion state clustering recognition model, the method further includes:

[0013] Based on the calibration results of the intrinsic and extrinsic parameters of the target camera, the homography matrix between the image plane of the target camera and the grate surface of the waste incinerator is determined;

[0014] The grate surface diagram of the waste incinerator is divided into grids to obtain the physical space scale grid corresponding to the grate surface diagram of the waste incinerator;

[0015] The physical space scale grid is mapped onto the image plane to obtain the image grid coordinate system;

[0016] The unlabeled flame combustion sample image is segmented using the image grid coordinate system to obtain multiple flame combustion sample image blocks.

[0017] The step of inputting the unlabeled flame combustion sample image into the original combustion state clustering recognition model includes:

[0018] The multiple flame combustion sample images are divided into blocks and input into the original combustion state clustering recognition model.

[0019] By applying the embodiments of this invention, a precise mapping relationship between the image plane and the physical space of the grate is established, ensuring physical scale consistency in the feature extraction process. The grid mapping method based on the homography matrix effectively overcomes the perspective distortion problem caused by differences in camera installation angles, enabling feature comparison of image data acquired from different perspectives according to a unified physical benchmark. The combination of physical space grid partitioning and image block processing decouples the global combustion state into local area monitoring, significantly improving the model's sensitivity to conditions such as fire line deviation and local combustion anomalies, while also enhancing the algorithm's generalization ability to different furnace structures.

[0020] Preferably, the step of inputting the unlabeled flame combustion sample image into the original combustion state clustering recognition model to obtain the joint probability matrix of the combustion state of the sample waste incinerator includes:

[0021] Using the original combustion state clustering recognition model, each flame combustion sample image is divided into blocks and mapped to a multi-channel space to obtain the color transformation result of each flame combustion sample image block;

[0022] Using the original combustion state clustering recognition model, position transformation processing is performed on each of the flame combustion sample image blocks to obtain the position transformation result of each of the flame combustion sample image blocks;

[0023] The mutual information influence factor is determined based on the color transformation result and the position transformation result;

[0024] The multiple flame combustion sample images are divided into blocks according to the target grid segmentation rule, and the image feature code of each flame combustion sample image block in each square corresponding to the target grid segmentation rule is extracted;

[0025] The image feature encoding is processed by the multilayer perceptron branch in the original combustion state clustering identification model to obtain the unique thermal feature vector;

[0026] The joint probability matrix of the combustion state is determined using the unique thermal eigenvector.

[0027] By applying embodiments of this invention, a feature invariance learning mechanism is constructed through color space transformation and geometric deformation enhancement, enabling the model to effectively distinguish the essential characteristics of combustion from interfering noise. A multi-channel feature encoding strategy integrates multi-dimensional visual information such as chroma, saturation, and brightness to comprehensively characterize the dynamic properties of the flame. A gridded feature extraction method transforms irregular image regions into standardized feature vectors, preserving spatial distribution patterns while eliminating the influence of shape differences. This design, through multi-level feature transformation and probability space mapping, significantly improves the model's robustness to interference factors such as changes in illumination and fluctuations in fuel composition.

[0028] Preferably, the unique hot feature vector includes a first unique hot vector and a second unique hot vector. The first unique hot vector is determined by a first multilayer perceptron branch that serves as the main clustering head, and the second unique hot vector is determined by a second multilayer perceptron branch that serves as the superclustering head.

[0029] The step of determining the joint probability matrix of the combustion state using the unique thermal feature vector includes: determining the joint probability matrix of the combustion state based on the transpose of the second unique thermal vector and the first unique thermal vector.

[0030] By applying the embodiments of this invention, a hierarchical feature representation system is constructed through joint probabilistic modeling of principal clustering and superclustering. Principal clustering captures the macroscopic category features of combustion states, while superclustering retains fine-grained differences in operating conditions. Spatial correlation between the two probability distributions is established through transpose multiplication, effectively uncovering the intrinsic connections between features at different levels. This dual-branch structure enables the model to meet both basic state classification requirements and support in-depth analysis of complex operating conditions, providing multidimensional decision-making basis for subsequent control strategy formulation and achieving a balance between accuracy and completeness in combustion state description.

[0031] Preferably, determining the mutual information variables based on the joint probability matrix of the combustion state includes: determining the mutual information variables using the joint probability matrix of the combustion state and the mutual information influence factor.

[0032] This invention introduces a mutual information influence factor to dynamically adjust feature weights, enabling the model to automatically focus on key feature changes during training. This mechanism effectively suppresses interference from irrelevant factors in the clustering process by quantifying the impact of color transformations and positional shifts on feature associations. Combined with marginal probability constraints on the joint probability matrix, it ensures the model maintains a reasonable feature distribution while maximizing mutual information, avoiding getting trapped in local optima. This adaptive optimization strategy significantly improves the model's adaptability to complex feature changes in real-world conditions.

[0033] Preferably, the mutual information variables are determined using the joint probability matrix of the combustion state and the mutual information influence factor, including:

[0034] The mutual information variables are determined based on the joint probability matrix of the combustion state, the first edge probability matrix, the second edge probability matrix, and the mutual information influence factor; wherein, the first edge probability matrix is ​​determined by the first unique heat vector; and the second edge probability matrix is ​​determined by the second unique heat vector.

[0035] By applying embodiments of this invention, a statistical correlation model of the feature space is established using a mutual information calculation method based on marginal probability constraints, accurately quantifying the stable association patterns between original and enhanced features. By weighted correction of the joint probability distribution, the requirements for feature invariance and discriminativeness are effectively balanced. This design enables the model to fully explore common patterns under different transformation conditions while preserving the essential characteristics of the combustion state, thus enhancing the robustness of feature representation. The coupling optimization mechanism of the probability space provides a reliable mathematical foundation for unsupervised learning, ensuring that the clustering results have clear physical interpretability.

[0036] Preferably, the target grid segmentation rule is a 25*25 grid segmentation rule.

[0037] Preferably, determining the mutual information influence factor based on the color transformation result and the position transformation result includes:

[0038] The mutual information influence factor is determined based on the random values ​​added to the chroma channel, saturation channel, and luminance channel in the color transformation result.

[0039] And / or,

[0040] The mutual information influence factor is determined based on the change in longitudinal position offset in the position transformation result.

[0041] By applying the embodiments of this invention, a mutual information influence factor constructed based on multi-channel perturbation quantization is used to achieve fine-grained control over changes in color and spatial features. The combined effect of chromaticity, saturation, and brightness factors enables the model to adaptively adjust its focus on flame color features, while the position offset factor effectively reduces the impact of mechanical motion disturbances on feature learning. This multi-dimensional weight adjustment mechanism allows the model to maintain stable feature extraction capabilities under complex operating conditions, significantly improving its adaptability to real-world scenarios such as changes in fuel composition and grate movement.

[0042] Preferably, the combustion state clustering identification result is achieved by maximizing the unique thermal features of the main cluster and maximizing the unique thermal features of the supercluster.

[0043] By applying this embodiment of the invention, the dual-maximization decision-making mechanism determines the dominant combustion state category through primary clustering and performs fine-grained verification using super-clustering, forming a hierarchical decision-making system. This design ensures the reliability of the basic classification while capturing subtle differences in operating conditions through the probability distribution of the super-clustering space. The synergistic effect of the primary and super-clustering effectively reduces the risk of misjudgment. When faced with combustion states with ambiguous boundaries, it can make more accurate judgments through probability fusion, significantly improving the system's decision confidence under complex operating conditions.

[0044] Preferably, the step of dividing the grate surface diagram of the waste incinerator into a grid includes dividing the grate surface diagram of the waste incinerator into a 30cm*30cm grid.

[0045] In summary, the waste incinerator flame combustion status recognition method based on unsupervised learning provided by the embodiments of the present invention, by constructing an unsupervised feature learning framework and a mutual information-driven optimization mechanism, can achieve end-to-end clustering recognition based on physical space alignment and feature invariance learning, significantly improving the objectivity of combustion status judgment and cross-device generalization ability, compared with traditional methods that rely on manual annotation or fixed feature extraction.

[0046] In detail, by using physical space grid mapping and image block processing, the feature space offset problem caused by differences in camera installation parameters is effectively overcome. In comparative tests of different incinerator types, the cross-device consistency error of fire line height detection is significantly reduced. The introduction of mutual information variables and the maximization training strategy enable the combustion state clustering identification model to automatically focus on the essential combustion characteristics in the fluctuation of waste composition. Even under extreme operating conditions with changes in fuel calorific value, the identification accuracy of the main cluster can still be guaranteed. Furthermore, the hierarchical clustering architecture, through the synergistic effect of the main cluster label and the supercluster label, can achieve the dual functions of deterministic determination of combustion state and quantification of uncertainty.

[0047] In addition, the unsupervised learning mechanism in this embodiment of the invention can eliminate the cost of manual annotation, enabling the flame combustion situation recognition system to directly use historically accumulated video data for model training, thereby improving data utilization. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart illustrating a method for identifying the flame combustion status of a waste incinerator based on unsupervised learning, provided in an embodiment of the present invention.

[0050] Figure 2 This is a network structure diagram of a target combustion state clustering identification model provided in an embodiment of the present invention.

[0051] Figure 3 This is a block diagram of a flame combustion status recognition system provided in an embodiment of the present invention. Detailed Implementation

[0052] 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 with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0053] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0054] like Figure 1 As shown, this embodiment of the invention provides a method for identifying the flame combustion status of a waste incinerator based on unsupervised learning, which is applied to the above-mentioned flame combustion status identification system. The specific steps for implementing this method are as follows: steps 110-140.

[0055] Step 110: Obtain unlabeled flame combustion sample images of the sample waste incinerator.

[0056] In this embodiment of the invention, the process of acquiring unlabeled flame combustion sample images of a waste incinerator needs to strictly adhere to the principles of physical spatial calibration and multi-dimensional data acquisition. For example, it is necessary to calibrate the intrinsic and extrinsic parameters of the industrial cameras deployed at the waste incinerator site, establish a homography mapping relationship between the image plane and the physical space of the grate, and ensure spatial consistency in subsequent feature extraction.

[0057] For example, video acquisition equipment deployed in multiple incineration plant areas can continuously capture video streams of flame combustion under different furnace types, seasonal conditions, and combustion stages, covering various combustion states such as abnormal fire height, flame deviation, and brightness fluctuations. During the acquisition process, a uniform sampling strategy along the time axis should be adopted, extracting keyframe images at fixed intervals to ensure that the sample set covers all dynamic features within the combustion cycle.

[0058] For example, for a mechanical grate incinerator with a daily processing capacity of 800 tons, flame images with a resolution of 1920×1080 need to be captured every 5 seconds during a continuous 72-hour operation cycle. This results in a database of tens of millions of images covering the entire process of grate feeding, ignition, stable combustion, and burnout cooling. All images retain the original sensor data without manual annotation or preprocessing, only adding metadata such as the acquisition timestamp, furnace number, and camera number.

[0059] Step 120: Input the unlabeled flame combustion sample image into the original combustion state clustering recognition model to obtain the joint probability matrix of the combustion state of the sample waste incinerator, and determine the mutual information variables based on the joint probability matrix of the combustion state.

[0060] Please refer to the following: Figure 2 In this embodiment of the invention, the process of inputting the unlabeled flame combustion sample image into the original combustion state clustering recognition model requires multi-level feature transformation and probability space mapping. First, the input image is segmented into a grid based on physical spatial scale. A pre-calibrated homography matrix is ​​used to back-project a 30cm × 30cm physical grid from the grate onto the image plane, generating a set of irregular image blocks corresponding to the combustion area. Each image block is then decomposed in the HSV color space, and a random offset of ±15° is applied to the chroma channel, a linear adjustment of ±30% is applied to the saturation channel, and a normally distributed noise perturbation is superimposed on the luminance channel to generate an enhanced sample with controllable feature changes.

[0061] It is understandable that during processing by the grid flatten layer network, each irregular image patch is divided into a 25×25 uniform sub-grid. The maximum brightness coordinates of each sub-region in the V channel, the mean value in the S channel, and the peak value of the histogram in the H channel are extracted. These are then fused into a 625-dimensional feature vector using learnable weight coefficients. After inputting the above feature vector into the encoder module based on the transformer architecture, the multi-head self-attention mechanism captures the dynamic correlation patterns of flames across regions, outputting a deep feature representation with spatial awareness. The main clustering head and super-clustering head of the original combustion state clustering recognition model generate 9-dimensional and 18-dimensional probability distribution vectors, respectively. By calculating the joint probability matrix of the original samples and the enhanced samples, and combining the brightness adjustment factor and the position offset factor, a weighted mutual information objective function is constructed to quantify the ability of the original combustion state clustering recognition model to capture invariant features.

[0062] Step 130: Train the original combustion state clustering identification model in the direction of maximizing the mutual information variables until the preset convergence condition is met, and obtain the target combustion state clustering identification model.

[0063] In this embodiment, the process of training the original combustion state clustering recognition model in the direction of maximizing the mutual information variables requires an adaptive optimization strategy and a dynamic weight adjustment mechanism. In the initial training phase, network parameters are updated with a high learning rate to encourage the model to quickly establish a rough correlation between flame morphology and cluster centers. As the number of iterations increases, enhanced samples based on brightness transformation and projection transformation are gradually introduced. The mutual information weights of corresponding sample pairs are dynamically reduced through an α factor, forcing the network to ignore changes in non-critical features.

[0064] For example, when the input sample pair contains an affine transformation of 10% in the vertical direction, the position offset factor is set to 0.3, which significantly reduces the contribution of this transformation to the mutual information calculation and guides the network to focus on essential features such as flame distribution patterns. During the optimization process, a joint training strategy of contrastive loss and clustering loss is adopted to simultaneously constrain intra-class compactness and inter-class separation within the feature embedding space.

[0065] For example, when the mutual information gain on the validation set is below 0.1% for five consecutive epochs, an early stop mechanism is triggered to save the current optimal model parameters. After training, the model should be able to accurately distinguish nine main combustion states on the retained test set, such as abnormal fire line height (e.g., exceeding 2 / 3 of the grate height is considered too high) and brightness fluctuations (e.g., a V channel mean value below 40 is considered too dark), and form fine-grained state classification in the super-clustering space.

[0066] Step 140: Obtain the current flame combustion image of the target waste incinerator, input the current flame combustion image into the target combustion state clustering identification model, and obtain the combustion state clustering identification result of the target waste incinerator; wherein, the combustion state clustering identification result includes a main cluster identification label and a super cluster identification label.

[0067] In this embodiment, the process of acquiring the current flame combustion image of the target waste incinerator and performing real-time inference requires the construction of a multi-level decision fusion system. Specifically, the inference module deployed at the industrial control terminal receives real-time video streams from cameras and extracts keyframes at 1-second intervals, inputting them into the target combustion state clustering identification model. The main clustering probability vector output by the target combustion state clustering identification model is used to determine the dominant combustion state category through argmax operation, while simultaneously parsing the auxiliary labels associated with the main category in the super-clustering probability distribution.

[0068] For example, when the main cluster output is "flame skewed to the left," the corresponding set of supercluster numbers is retrieved from a predefined mapping table, and the average probability of these categories is calculated as the overall confidence level. The control system triggers corresponding adjustment commands based on the confidence level threshold: if the confidence level for left-skewed burning exceeds 75%, the primary air volume on the right side is increased by 15%, and the grate movement frequency on the right side is increased by 20%. All inference results and control records are stored in a historical database for iterative optimization of the target combustion state clustering identification model. For newly emerging undefined combustion modes, high-probability categories in the supercluster space will trigger an early warning mechanism, prompting operators to intervene in the analysis and update the classification label library, thus enabling the continuous evolution of the target combustion state clustering identification model's capabilities.

[0069] It is understood that, in implementing the embodiments of the present invention, camera calibration and image segmentation directly affect the accuracy of subsequent feature extraction and clustering recognition. Therefore, as an optional implementation, before inputting the unlabeled flame combustion sample image into the original combustion state clustering recognition model as described in step 120, the method further includes:

[0070] Step 210: Based on the calibration results of the intrinsic and extrinsic parameters of the target camera, determine the homography matrix between the image plane of the target camera and the grate surface diagram of the waste incinerator.

[0071] In this embodiment of the invention, the process of determining the homography matrix between the image plane and the grate surface based on the calibration results of the target camera's intrinsic and extrinsic parameters requires precise optical calibration and spatial geometric modeling. For example, a high-precision checkerboard calibration plate is used to calibrate the intrinsic parameters of the industrial camera, and the lens distortion coefficient, focal length parameter, and principal point offset are calculated by capturing images of the calibration plate at different angles. In the extrinsic parameter calibration stage, the calibration plate is fixed to the surface of the incinerator grate, establishing a spatial correspondence between the grate physical coordinate system and the camera image coordinate system.

[0072] For example, for a reciprocating mechanical grate with dimensions of 6m × 2.5m, five calibration points are placed at the four corners and the center of the grate. Precise physical coordinates (X = 0, Y = 0, Z = 0) and (X = 6000, Y = 0, Z = 0) of each marker are obtained using a three-dimensional spatial coordinate measuring instrument. Based on the pixel coordinates of the marker points captured by the camera, a direct linear transformation algorithm is used to solve for the 3×3 parameter matrix of the homography matrix H, achieving a projection mapping from the physical plane of the grate to the image plane. This homography matrix will serve as the basic spatial transformation tool for subsequent image segmentation processing, ensuring physical scale consistency in flame feature extraction.

[0073] Step 220: Perform grid segmentation on the grate surface diagram of the waste incinerator to obtain the physical spatial scale grid corresponding to the grate surface diagram of the waste incinerator.

[0074] In this embodiment of the invention, the process of physically dividing the grate surface image of a waste incinerator into a grid needs to be combined with the requirements of the combustion process and feature extraction. For example, a Cartesian coordinate system is established based on the actual size of the incinerator grate. Longitudinal grid lines are divided at 30cm intervals along the length of the grate, and transverse grid lines are divided at 30cm intervals along the width, forming a uniform physical grid covering the entire grate surface. For an 8m × 3m inclined reverse-push grate, the physical grid division will generate a grid array of 26 columns (800cm / 30cm ≈ 26.67 rounded) and 10 rows (300cm / 30cm = 10). Each 30cm × 30cm physical grid corresponds to a combustion status monitoring unit in a specific area on the grate, ensuring that subsequent image segmentation can accurately reflect key combustion parameters such as fire line position and flame distribution. The grid division results are stored in the form of a two-dimensional coordinate index. Each grid unit records the physical coordinate values ​​of its lower left and upper right corners, providing a spatial reference for image planar mapping.

[0075] Step 230: Map the physical space scale grid onto the image plane to obtain the image grid coordinate system.

[0076] In this embodiment of the invention, the process of mapping a physical spatial scale grid to an image plane to generate an image grid coordinate system requires perspective projection transformation using a homography matrix. The coordinates of each physical grid corner point established in step 220 are substituted into the homography matrix H, and the corresponding image pixel coordinates are calculated through homogeneous coordinate transformation. For example, a grid corner point with physical coordinates (X = 300, Y = 150) obtains image coordinates (u = 1250, v = 680) after transformation by the H matrix, thereby determining the projection area of ​​the physical grid in the image.

[0077] It is understandable that due to perspective distortion caused by the camera's mounting angle, the mapped image mesh typically exhibits an irregular quadrilateral shape, such as a trapezoid or parallelogram. The image mesh coordinate system stores the pixel coordinates of each mesh vertex in the form of a two-dimensional array, forming an image block index table that corresponds one-to-one with the physical mesh. This coordinate system ensures that images acquired by cameras with different mounting angles and models can be compared in terms of features at a uniform physical scale, improving the model's robustness to device differences.

[0078] Step 240: The unlabeled flame combustion sample image is segmented using the image grid coordinate system to obtain multiple flame combustion sample image segments.

[0079] In this embodiment of the invention, the process of segmenting an unlabeled flame combustion sample image using an image grid coordinate system requires the extraction of image regions based on geometric transformation. For example, based on the vertex coordinate data of the image grid coordinate system, a bilinear interpolation algorithm is used to crop out image blocks corresponding to each grid from the original image.

[0080] For example, for a physical grid mapped to image coordinates (u1=1200, v1=700), (u2=1280, v2=700), (u3=1275, v3=750), and (u4=1195, v4=750), pixel data within that quadrilateral region is extracted to form independent image blocks. Each block retains the original RGB color space information and is appended with a corresponding physical grid number (e.g., G12-05 represents the 12th column, 5th row grid). The image set after block processing constitutes the basic unit for subsequent feature extraction, ensuring that the model can focus on the combustion state analysis of a specific area of ​​the grate and avoid interference noise from global features.

[0081] As an optional implementation, based on steps 210-240, step 120, inputting the unlabeled flame combustion sample image into the original combustion state clustering recognition model to obtain the joint probability matrix of the combustion state of the sample waste incinerator includes:

[0082] Step 121: Using the original combustion state clustering recognition model, map each flame combustion sample image block to a multi-channel space to obtain the color transformation result of each flame combustion sample image block.

[0083] In this embodiment of the invention, the process of mapping flame combustion sample images to a multi-channel space using the original combustion state clustering recognition model requires HSV color space decomposition and controllable feature enhancement. Specifically, each RGB image block is first converted to the HSV color space, where the chroma channel (H) represents the flame color component, the saturation channel (S) reflects the combustion intensity, and the luminance channel (V) characterizes the thermal radiation level. A random color shift of ±15° is applied to the H channel to simulate changes in fuel composition; for example, the H value is adjusted from 30° (orange flame) to 45° (yellow flame) to simulate the combustion state of plastic waste. The S channel undergoes a ±30% linear adjustment to enhance the characteristic differences in combustion intensity, and the V channel is superimposed with Gaussian noise with a mean of 0 and a standard deviation of 20 to simulate fluctuations in illumination conditions. The transformed HSV image block and the original block constitute a sample pair, and their corresponding chroma adjustment ΔH, saturation scaling factor kS, and luminance noise intensity σV are used as parameters for calculating the mutual information influence factor α, which is then used for weighted processing of the subsequent joint probability matrix.

[0084] Step 122: Using the original combustion state clustering recognition model, perform position transformation processing on each of the flame combustion sample image blocks to obtain the position transformation result of each of the flame combustion sample image blocks.

[0085] In this embodiment of the invention, the process of performing position transformation processing on the image blocks of the flame combustion sample needs to simulate the image offset caused by changes in camera viewpoint and mechanical movement of the grate. For example, a projection transformation algorithm can be used to apply random affine transformations to the image blocks, including geometric deformations such as vertical translation (simulating grate lifting and lowering), horizontal shearing (simulating changes in side view angle), and scaling (simulating focal length adjustment).

[0086] For example, a vertical translation transformation is applied to a certain image block, lowering its center point by 5% of the image height to simulate a fire line position drop. Each geometric transformation parameter is recorded as a position transformation factor αpos. When the transformation involves vertical displacement, αpos is set to 0.5 to reduce its contribution to mutual information calculation. The position-transformed image block and the original block form a spatial feature comparison pair, forcing the model to learn the flame distribution pattern rather than absolute position information, thus improving its adaptability to changes in installation position.

[0087] Step 123: Determine the mutual information influence factor based on the color transformation result and the position transformation result.

[0088] In this embodiment of the invention, the process of determining the mutual information influence factor based on the color transformation result and the position transformation result requires the establishment of a multi-dimensional feature weight fusion mechanism. Specifically, the chroma adjustment factor αH = 1 - |ΔH| / 180, saturation factor αS = 1 - |kS-1|, and brightness factor αV = 1 - σV / 255 in step 121 are weighted and multiplied with the position transformation factor αpos in step 122 to obtain the comprehensive mutual information influence factor α = αH × αS × αV × αpos. For example, when a sample pair contains ΔH = 30° (αH = 0.83), kS = 1.2 (αS = 0.8), σV = 30 (αV = 0.88), and vertical translation (αpos = 0.5), the comprehensive mutual information influence factor α = 0.83 × 0.8 × 0.88 × 0.5 ≈ 0.29. This factor will be used as a weighting coefficient in the calculation of the joint probability matrix, effectively suppressing the influence of irrelevant features on the clustering process and guiding the model to focus on the essential features of the combustion state.

[0089] Step 124: Divide the multiple flame combustion sample image blocks according to the target grid segmentation rule, and extract the image feature code of each flame combustion sample image block in each square corresponding to the target grid segmentation rule; the target grid segmentation rule is a 25*25 grid segmentation rule.

[0090] In this embodiment of the invention, the process of feature encoding of flame combustion sample image blocks using a 25×25 grid segmentation rule requires feature standardization of irregular regions. Each image block is first uniformly divided into 25 rows and 25 columns, totaling 625 sub-grids, maintaining a consistent grid density regardless of the original block shape. For each sub-grid, three feature quantities are extracted: the maximum brightness value of the V channel, the mean value of the S channel, and the main peak value of the histogram of the H channel within its coverage area.

[0091] For example, the top-left sub-grid area of ​​a trapezoidal image block might only contain a portion of the effective pixels. In this case, the 95th percentile of the V channel pixel values ​​within that area is calculated as an approximation of the maximum brightness. The three feature values ​​are linearly combined using a learnable weight vector w = [w1, w2, w3] to generate a fused feature value for each sub-grid: f = w1 × V_max + w2 × S_mean + w3 × H_peak. The feature values ​​of all 625 sub-grids are arranged in row-major order to form a fixed-length 625-dimensional feature vector, eliminating the influence of irregular block shapes on the feature dimension.

[0092] Step 125: Process the image feature encoding through the multilayer perceptron branch in the original combustion state clustering identification model to obtain the unique thermal feature vector.

[0093] In this embodiment, the process of image feature encoding through the multilayer perceptron branch in the original combustion state clustering identification model requires mapping the feature space to cluster labels. For example, a 625-dimensional grid flattened feature vector is input into the main clustering branch containing three fully connected layers. The first layer reduces the dimension to 256 and applies the ReLU activation function. The second layer further compresses it to 64 dimensions. Finally, the output layer generates a 9-dimensional main clustering probability distribution through the softmax function.

[0094] Meanwhile, another independent multilayer perceptron branch processes feature vectors with the same structure, outputting an 18-dimensional super-cluster probability distribution. For example, when the input feature represents a left-biased combustion pattern, the main clustering branch might output a high probability of 0.85 in category 1 (left-biased combustion), while the super-clustering branch outputs probability values ​​of 0.6, 0.3, and 0.7 in categories 2, 4, and 5, respectively. The parameters of the two branches are jointly optimized during training through a mutual information maximization objective function, ensuring that the main clustering captures the macroscopic combustion state and the super-clustering retains fine-grained differential features, jointly constructing a hierarchical combustion state description system.

[0095] Step 126: Determine the joint probability matrix of the combustion state using the unique thermal eigenvector.

[0096] In this embodiment of the application, the combustion state joint probability matrix is ​​determined using the unique thermal feature vector, including: step 1260: determining the combustion state joint probability matrix based on the transpose of the second unique thermal vector and the first unique thermal vector.

[0097] In this embodiment of the invention, the process of determining the joint probability matrix of the combustion state based on the transpose of the second unique heat vector and the first unique heat vector requires a probability space alignment operation. For example, the 9-dimensional probability distribution vector P(m) output by the main clustering branch and the 18-dimensional probability distribution vector P(g) output by the transformation branch are multiplied by a matrix to generate a 9×18-dimensional joint probability matrix Pmg = P(m)×P(g)^T. The matrix element Pmg[i,j] represents the joint probability distribution of the original image block belonging to the i-th class of the main cluster and its transformed version belonging to the j-th class of the super cluster. For example, when the probability of the 3rd class of the main cluster (moderate fire line) is 0.8 and the probability of the 7th class of the transformed super cluster (moderate fire line - brightness fluctuation) is 0.6, the corresponding matrix element value is 0.8×0.6 = 0.48. This joint probability matrix provides basic data for subsequent mutual information calculation, driving the model to establish a stable cross-transformation association pattern in the feature space.

[0098] As a preferred design approach, the determination of mutual information variables based on the joint probability matrix of the combustion state in step 120 can be achieved by using the joint probability matrix of the combustion state and the mutual information influence factor to determine the mutual information variables. More specifically, it can be as follows: the mutual information variables are determined according to the joint probability matrix of the combustion state, the first edge probability matrix, the second edge probability matrix and the mutual information influence factor; wherein, the first edge probability matrix is ​​determined by the first one-hot vector; and the second edge probability matrix is ​​determined by the second one-hot vector.

[0099] In this embodiment, the process of determining the mutual information variable based on the joint probability matrix of combustion status and the mutual information influence factor requires the construction of a multidimensional probability space coupling model. Specifically, the first one-hot vector output by the multilayer perceptron branch of the main cluster is normalized by softmax to generate the first marginal probability matrix P(m). This matrix is ​​a 9-dimensional row vector, where each element represents the probability that a sample belongs to the corresponding main cluster category. The second one-hot vector output by the multilayer perceptron branch is transformed and normalized by softmax to generate the second marginal probability matrix P(g). This matrix is ​​an 18-dimensional row vector, representing the distribution probability of the transformed sample in the super-cluster space. The joint probability matrix Pmg = P(m)T × P(g) is subjected to a Hadamard product with the mutual information influence factor α to obtain the corrected joint probability matrix Pmg' = α⊙Pmg. The mutual information variable I(m;g) is calculated as ΣΣPmg'[i,j] × log(Pmg'[i,j] / (P(m)[i] × P(g)[j])), where the logarithmic term quantifies the cross-correlation between the two probability distributions.

[0100] For example, when the probability of the main cluster type 2 (low fire line) is 0.9, and the probability of its corresponding transformed sample in the supercluster type 5 (low fire line - uneven fuel distribution) is 0.7, and the mutual information influence factor α = 0.8, the joint probability contribution of this sample pair is 0.9 × 0.7 × 0.8 = 0.504, significantly affecting the overall mutual information calculation result. This design, by introducing marginal probability constraints and factor weighting mechanisms, ensures that the model accurately distinguishes between intrinsic combustion characteristics and interference transformation characteristics during the optimization process.

[0101] As a preferred design approach, step 123, which describes determining the mutual information influence factor based on the color transformation result and the position transformation result, includes: determining the mutual information influence factor based on the random value addition results corresponding to the chroma channel, saturation channel, and luminance channel in the color transformation result; and / or, determining the mutual information influence factor based on the change in the vertical position offset in the position transformation result.

[0102] In this embodiment of the invention, the process of determining the mutual information influence factor based on the results of color transformation and position transformation requires the establishment of a multi-channel perturbation quantization system. For example, a ±15° random rotation perturbation is applied to the chroma channel in the HSV space, and the chroma influence factor αH = 1 - |ΔH| / 15 is calculated, where ΔH is the actual rotation angle. The saturation channel is linearly scaled by ±30%, and the saturation influence factor αS = 1 - |kS-1| is calculated, where kS is the scaling factor. Gaussian noise with a mean of μ and a standard deviation of σ is added to the luminance channel, and the luminance influence factor αV = 1 - σ / 255 is calculated.

[0103] For example, in terms of positional transformation, a projection transformation with a vertical offset of ΔY pixels is applied to the image block. The positional influence factor αY = 1 - |ΔY| / H, where H is the block height. The comprehensive mutual information influence factor α = αH × αS × αV × αY. For example, when a sample pair contains ΔH = 10° (αH = 0.33), kS = 1.2 (αS = 0.8), σ = 50 (αV = 0.8), and ΔY = 20 pixels (H = 100, αY = 0.8), the comprehensive mutual information influence factor α = 0.33 × 0.8 × 0.8 × 0.8 ≈ 0.17. This factor will significantly reduce the weight of the sample pair in the mutual information calculation, forcing the model to ignore the interference caused by chromatic aberration and positional changes, and focus on stable features such as flame distribution morphology.

[0104] As a preferred design approach, the combustion state clustering identification result is achieved by maximizing the unique thermal features of the main cluster and the unique thermal features of the supercluster.

[0105] In this embodiment of the invention, the process of achieving combustion state clustering identification by maximizing the unique thermal features of the main cluster and the unique thermal features of the supercluster requires the construction of a hierarchical decision fusion mechanism.

[0106] For example, during the online inference phase, after the input image is processed through grid segmentation, the main clustering branch outputs a 9-dimensional probability vector Pm, and the index of the maximum value is used to determine the dominant combustion state category. Simultaneously, the superclustering branch outputs an 18-dimensional probability vector Pg, which is retrieved from a predefined main-supercluster mapping table to obtain the set of supercluster numbers associated with the main category. For instance, when the main cluster is determined to be category 3 (flame right-biased burning), the associated supercluster numbers are 7, 11, and 15, and the average probability of these categories is calculated as the overall confidence level. The control system is equipped with a dual-threshold trigger mechanism: if the main cluster probability exceeds 0.85 and the overall confidence level is greater than 0.7, the control command is executed immediately; if the main cluster probability is between 0.6 and 0.85, the multimodal sensor verification process is initiated.

[0107] For example, a sample of a right-side combustion condition, after model processing, yielded Pm[3] = 0.92, with associated super-cluster probabilities Pg[7] = 0.6, Pg

[11] = 0.8, and Pg

[15] = 0.7. The overall confidence level was (0.6 + 0.8 + 0.7) / 3 = 0.7, triggering a control action that reduced the right-side damper by 10% and increased the left-side secondary air volume by 15%. This design ensures the reliability of the identification results through a dual probability maximization mechanism, while retaining the super-cluster space's ability to represent complex operating conditions.

[0108] As an optional but not limiting embodiment, training the original combustion state clustering identification model in the direction of maximizing the mutual information variables until a preset convergence condition is met to obtain the target combustion state clustering identification model includes:

[0109] The multiple flame combustion sample images are divided into blocks and input into the original combustion state clustering recognition model to generate color transformation results and position transformation results of the flame combustion sample image blocks;

[0110] The mutual information influence factor is calculated based on the chroma channel perturbation amplitude, saturation channel scaling ratio, luminance channel noise intensity, and vertical offset ratio in the position transformation result.

[0111] The flame combustion sample image is divided into grid blocks using the 25*25 grid segmentation rule. The chroma value, saturation value, and luminance value of the pixel with the largest luminance channel in each grid are extracted and weighted and fused to generate the grid feature code of the flame combustion sample image block.

[0112] The grid feature encoding and the physical spatial location encoding of the flame combustion sample image blocks are concatenated to form a joint feature vector;

[0113] The joint feature vector is input into the main clustering multilayer perceptron branch of the original combustion state clustering identification model, and the first unique heat vector is generated by stepwise dimensionality reduction through three fully connected layers.

[0114] The grid feature encoding corresponding to the position transformation result is input into the transformation multilayer perceptron branch of the original combustion state clustering identification model, and the second unique heat vector is generated through a fully connected layer with shared weights.

[0115] The joint probability matrix of combustion status is constructed based on the matrix product of the first independent heat vector and the second independent heat vector, and the mutual information influence factor is used to perform weighted correction on each element of the joint probability matrix of combustion status.

[0116] The mutual information variable is calculated based on the modified joint probability matrix of combustion status, the first marginal probability matrix output by the main clustering multilayer perceptron branch, and the second marginal probability matrix output by the transform multilayer perceptron branch.

[0117] The parameters of the fully connected layers of the main clustering multilayer perceptron branch and the transform multilayer perceptron branch are adjusted by the backpropagation algorithm, so that the mutual information variable continues to increase during the iteration process.

[0118] When the growth rate of the mutual information variable falls below a preset threshold after a preset number of iterations, the original combustion state clustering identification model is determined to meet the preset convergence condition, and the current model parameters are saved as the target combustion state clustering identification model.

[0119] In this embodiment of the invention, during the process of inputting the multiple flame combustion sample images into the original combustion state clustering recognition model to generate color transformation results and position transformation results, HSV color space decomposition is first performed on each image block, and a random angle rotation perturbation is applied to the chroma channel. For example, the original chroma channel peak value of a certain image block is 30° (orange flame), which becomes 42° (orange-red flame) after adding a 12° perturbation, simulating the color shift caused by changes in fuel composition. At the same time, a vertical position offset of 8% is applied to the same block, causing its central area to shift down by 20 pixels, simulating the change in fire line position caused by grate lifting and lowering. The color transformation results retain the perturbed chroma, saturation, and brightness channel data, and the position transformation results record geometric deformation parameters, providing a basis for subsequent mutual information influence factor calculation.

[0120] Secondly, a multi-dimensional perturbation quantization strategy is adopted when calculating the mutual information impact factor based on the chroma channel perturbation amplitude, saturation channel scaling ratio, luminance channel noise intensity, and vertical offset ratio in the position transformation results. For example, when the chroma perturbation amplitude is 12° (maximum allowed 15°), the saturation scaling ratio is 1.2 times (original value 1), the luminance noise intensity is 40 (maximum 255), and the vertical offset ratio is 8% (block height 250 pixels), the chroma impact factor is calculated as 1 - 12 / 15 = 0.2, the saturation impact factor is 1 - |1.2 - 1| = 0.8, the luminance impact factor is 1 - 40 / 255 ≈ 0.84, and the position impact factor is 1 - 8 / 100 = 0.92. The comprehensive mutual information impact factor is 0.2 × 0.8 × 0.84 × 0.92 ≈ 0.124, which significantly reduces the weight of this sample pair in the mutual information calculation, forcing the model to ignore changes in non-critical features.

[0121] Next, when dividing the flame combustion sample image into grids using the 25*25 grid segmentation rule, each irregular image block is uniformly divided into 625 sub-grids. For example, after mapping a trapezoidal block to the image plane, its upper left corner sub-grid only covers one-third of the effective pixel area. The pixel with the largest value in the brightness channel is still searched within this sub-grid. The chromaticity value (45°), saturation value (220), and brightness value (180) of this pixel are extracted and weighted using learnable weight coefficients w1 = 0.6, w2 = 0.3, and w3 = 0.1, resulting in a sub-grid feature encoding value of 0.6 × 45 + 0.3 × 220 + 0.1 × 180 = 27 + 66 + 18 = 111. The feature encodings of all 625 sub-grids are arranged in spatial order, forming a 625-dimensional vector representing the core features of the local flame.

[0122] Furthermore, when the grid feature encoding and the physical spatial location encoding of the flame combustion sample image blocks are concatenated into a joint feature vector, the physical spatial location encoding is derived from the normalized coordinates of the grate grid. For example, a certain block corresponds to the 15th column and 6th row of the physical grid in the grate, and its center point physical coordinates are (X = 450cm, Y = 180cm), which are normalized to (0.5625, 0.6) in an 8m × 3m grate system. This two-dimensional location encoding is concatenated with the 625-dimensional grid feature encoding to form a 627-dimensional joint feature vector, enabling the model to simultaneously perceive its absolute position in the grate when analyzing local flame features, thus enhancing the ability to identify abnormal fire line heights.

[0123] Furthermore, when the joint feature vector is input into the main clustering multilayer perceptron branch of the original combustion state clustering recognition model to generate the first unique heat vector, the 627-dimensional features are first reduced to 256 dimensions through the first fully connected layer, and the ReLU activation function is used to eliminate negative values. The second fully connected layer further compresses the features to 64 dimensions, capturing the high-order spatial-color association patterns across the grid. The final output layer generates a 9-dimensional probability distribution through the softmax function, for example, outputting a probability value of 0.85 in the "excessively high fire line" category, indicating that the model is highly confident that the current block presents a combustion state with an excessively high fire line. This first unique heat vector directly represents the combustion state classification result in the main clustering space.

[0124] Furthermore, when the grid feature encoding corresponding to the position transformation result is input into the transform multilayer perceptron branch to generate the second one-hot vector, the weights of the first two fully connected layers of the main clustering branch are shared. For example, after the position transformation, the block obtains a 625-dimensional vector through grid feature extraction. After passing through the shared 256-dimensional and 64-dimensional fully connected layers, an 18-dimensional super-clustering probability distribution is generated by an independent output layer. If the original block's main cluster is "moderate fireline," its transformed block may output a probability of 0.7 in the "moderate fireline - brightness fluctuation" category in the super-clustering space, reflecting the fine-grained differences caused by brightness changes. This design ensures that the transforming branch focuses on capturing state changes caused by perturbations while sharing the basic feature extraction capabilities.

[0125] In detail, when constructing the joint probability matrix of combustion status based on the matrix product of the first and second unique heat vectors, the 9-dimensional probability vector of the main cluster and the 18-dimensional probability vector of the supercluster are outer-producted. For example, the probability of the second cluster (low fire line) is 0.9, and the probability of the fifth cluster (low fire line - uneven fuel distribution) is 0.6. Then, the initial value of element (2,5) in the joint probability matrix is ​​0.9 × 0.6 = 0.54. This element is then weighted and corrected using a mutual information influence factor of 0.124, resulting in a value of 0.54 × 0.124 ≈ 0.067. This operation significantly reduces the joint probability contribution of sample pairs with strong color or positional perturbations, forcing the model to learn stable combustion characteristics undisturbed by disturbances.

[0126] It is understandable that when calculating the mutual information variable based on the modified joint probability matrix of combustion status, the first marginal probability matrix output by the main cluster multilayer perceptron branch, and the second marginal probability matrix output by the transformed multilayer perceptron branch, a weighted mutual information formula is used. For example, the value of element (2,5) in the modified joint probability matrix is ​​0.067, the corresponding first marginal probability P(m)[2]=0.9, and the second marginal probability P(g)[5]=0.6, then the contribution of this element to the mutual information is 0.067×log(0.067 / (0.9×0.6))≈0.067×(-1.22)≈-0.082. The total mutual information variable is obtained by summing the contribution values ​​of all matrix elements. The larger this value is, the stronger the model's ability to extract invariant features.

[0127] In this embodiment of the invention, when adjusting the parameters of the main clustering multilayer perceptron branch and the fully connected layer of the transformed multilayer perceptron branch using the backpropagation algorithm, the optimization objective is to maximize the mutual information variable. The gradient of the mutual information variable with respect to the network parameters is calculated, and an adaptive moment estimation optimizer is used to update the weights. For example, when a weight of a fully connected layer causes a decrease in the mutual information variable, the gradient direction indicates that the weight needs to be increased to improve feature correlation. After multiple iterations, the model gradually learns to ignore color perturbations and positional shifts, focusing on essential features such as the flame distribution pattern.

[0128] It is understood that when the growth rate of the mutual information variable falls below a preset threshold after a preset number of iterations (e.g., the mutual information gain is less than 0.05% for 10 consecutive training cycles), the model is deemed to have reached convergence. The fully connected layer parameters of the main clustering branch and the transformation branch at this point are saved to form a target combustion state clustering identification model. It is understood that the above target combustion state clustering identification model is robust to color changes and positional shifts, accurately identifies nine main combustion states such as abnormal fire line height and off-center burning direction, and distinguishes 18 fine-grained operating condition variations in the super-clustering space.

[0129] In other words, the step of training the original combustion state clustering identification model in the direction of maximizing the mutual information variables until the preset convergence condition is met, and obtaining the target combustion state clustering identification model, is achieved through a two-branch mutual information maximization training strategy. The training strategy includes: color and position transformation feature extraction (chroma perturbation, saturation scaling, brightness noise, vertical offset), 25×25 grid feature encoding generation (fusion of three-channel values ​​of the maximum brightness pixel), joint feature vector construction (grid encoding and physical spatial position splicing), two-branch probability modeling (dimensionality reduction of the main clustering MLP branch and shared weights of the transformation MLP branch), mutual information weighted correction (adjustment of joint probability matrix elements), edge probability mutual information calculation and backpropagation optimization, and iterative amplification threshold termination condition.

[0130] Figure 3 A block diagram of a flame combustion situation recognition system 10 provided in an embodiment of the present invention is shown. The flame combustion situation recognition system 10 in this embodiment of the present invention has data storage, transmission, and processing functions, such as... Figure 3 As shown, the flame combustion situation recognition system 10 includes: a memory 11, a processor 12, a network module 13, and a flame combustion situation recognition device 20.

[0131] The memory 11, processor 12, and network module 13 are electrically connected directly or indirectly to achieve data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses or signal lines. The memory 11 stores a flame combustion situation identification device 20, which includes at least one software function module that can be stored in the memory 11 in the form of software or firmware. The processor 12 executes various functional applications and data processing by running the software program and module stored in the memory 11, such as the flame combustion situation identification device 20 in this embodiment of the invention, thereby implementing the WEB-based data logic expression processing method in this embodiment of the invention.

[0132] The memory 11 may be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory 11 stores programs, and the processor 12 executes these programs upon receiving execution instructions.

[0133] The processor 12 may be an integrated circuit chip with data processing capabilities. The processor 12 can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.

[0134] The network module 13 is used to establish a communication connection between the flame combustion status identification system 10 and other communication terminal devices via the network, and to realize the transmission and reception of network signals and data. The aforementioned network signals may include wireless signals or wired signals.

[0135] Understandable. Figure 3 The structure shown is for illustrative purposes only; the flame combustion status recognition system 10 may also include a comparison... Figure 3 The more or fewer components shown, or having the same Figure 3 The different configurations shown. Figure 3 The components shown can be implemented using hardware, software, or a combination thereof.

[0136] This invention also provides a computer-readable storage medium, which includes a computer program. When the computer program runs, it controls the flame combustion status recognition system 10, where the storage medium is located, to execute the aforementioned unsupervised learning-based method for recognizing the flame combustion status of a waste incinerator.

[0137] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A waste incinerator flame combustion state recognition method based on unsupervised learning, characterized in that, The method is applied to a flame combustion state identification system, and the method comprises: Obtaining an unannotated flame combustion sample image of a sample waste incinerator; Inputting the unannotated flame combustion sample image into an original combustion state clustering identification model to obtain a combustion state joint probability matrix of the sample waste incinerator, and determining a mutual information variable based on the combustion state joint probability matrix; Training the original combustion state clustering identification model in a direction of maximizing the mutual information variable until a preset convergence condition is met to obtain a target combustion state clustering identification model; Obtaining a current flame combustion image of a target waste incinerator, inputting the current flame combustion image into the target combustion state clustering identification model to obtain a combustion state clustering identification result of the target waste incinerator; wherein the combustion state clustering identification result comprises a main clustering identification label and a super clustering identification label; Before inputting the unannotated flame combustion sample image into the original combustion state clustering identification model, the method further comprises: According to a calibration result of internal and external parameters of a target camera, determining a homography matrix between an image plane of the target camera and a grate plane of a waste incinerator; Grid segmenting the grate plane of the waste incinerator to obtain a physical space scale grid corresponding to the grate plane of the waste incinerator; Mapping the physical space scale grid to the image plane to obtain an image grid coordinate system; Through the image grid coordinate system, block processing is performed on the unannotated flame combustion sample image to obtain a plurality of flame combustion sample image blocks; The inputting of the unannotated flame combustion sample image into the original combustion state clustering identification model comprises: Inputting the plurality of flame combustion sample image blocks into the original combustion state clustering identification model.

2. The method of claim 1, wherein, The inputting of the unannotated flame combustion sample image into the original combustion state clustering identification model to obtain the combustion state joint probability matrix of the sample waste incinerator comprises: Using the original combustion state clustering identification model, mapping each flame combustion sample image block to a multi-channel space to obtain a color transformation result of each flame combustion sample image block; Using the original combustion state clustering identification model, performing position transformation processing on each flame combustion sample image block to obtain a position transformation result of each flame combustion sample image block; Determining a mutual information influence factor based on the color transformation result and the position transformation result; Through a target grid segmentation rule, dividing the plurality of flame combustion sample image blocks to extract an image feature code of each flame combustion sample image block in each square corresponding to the target grid segmentation rule; Through a multi-layer perception branch in the original combustion state clustering identification model, processing the image feature code to obtain a one-hot feature vector; Using the one-hot feature vector to determine the combustion state joint probability matrix.

3. The method of claim 2, wherein, The one-hot feature vector comprises a first one-hot vector and a second one-hot vector, the first one-hot vector is determined by a first multi-layer perception branch as a main cluster head, and the second one-hot vector is determined by a second multi-layer perception branch as a hyper cluster head; The determination of the combustion state joint probability matrix by using the one-hot feature vector comprises: determining the combustion state joint probability matrix according to the transpose of the second one-hot vector and the first one-hot vector.

4. The method of claim 2, wherein, The determination of the mutual information variable based on the combustion state joint probability matrix comprises: determining the mutual information variable by using the combustion state joint probability matrix and the mutual information influence factor.

5. The method of claim 4, wherein, The determination of the mutual information variable by using the combustion state joint probability matrix and the mutual information influence factor comprises: The determination of the mutual information variable according to the combustion state joint probability matrix, a first edge probability matrix, a second edge probability matrix and the mutual information influence factor, wherein the first edge probability matrix is determined by a first one-hot vector, and the second edge probability matrix is determined by a second one-hot vector.

6. The method of claim 2, wherein, The target grid division rule is grid division rule of grid 7. The method of claim 2, wherein, The determination of the mutual information influence factor based on the color transformation result and the position transformation result comprises: The determination of the mutual information influence factor based on the color transformation result and the position transformation result comprises: The determination of the mutual information influence factor based on the color transformation result and the position transformation result comprises:

8. The method of claim 1, wherein, The determination of the mutual information influence factor based on the color transformation result and the position transformation result comprises: The combustion state clustering identification result is achieved by maximizing the main cluster one-hot feature and maximizing the hyper cluster one-hot feature.

9. The method of claim 1, wherein, The grid division on the garbage incinerator grate plan view comprises: dividing the garbage incinerator grate plan view into grids.

Images