Garbage incinerator flame combustion situation identification method based on unsupervised learning

Abstract

The embodiment of the invention relates to the technical field of image processing, in particular to a garbage incinerator flame combustion situation recognition method based on unsupervised learning. According to the method, the problem of lack of labeling data in flame combustion situation recognition of the garbage incinerator is effectively solved through an unsupervised learning framework; a mutual information maximization objective function driving model is utilized to autonomously excavate essential feature association in a flame image, and a corresponding relation between a combustion state and a visual feature can be established without depending on manual labeling; and hierarchical description of the combustion state is realized through dual-label output of main clustering and super clustering, so that the macroscopic working condition classification capability is reserved, and the microscopic dynamic change characteristics are captured. Visibly, according to the embodiment of the invention, the adaptability of the model to a complex combustion scene is enhanced while the data annotation cost is reduced, and refined monitoring and recognition of the combustion state of the incinerator can be realized.

Description

Technical Field

[0001] The embodiments of the present invention relate to the technical field of image processing. Specifically, the embodiments of the present invention relate to a method for identifying the combustion state of a waste incinerator flame based on unsupervised learning. Background Art

[0002] In the field of waste incineration power generation, the real-time monitoring and control of the flame combustion state are the core links to ensure the incineration efficiency and the up-to-standard discharge of pollutants.

[0003] However, traditional manual monitoring methods rely on the experience of operators to visually judge the characteristics of the flame shape, brightness, etc., which have significant subjective judgment biases and lags, and it is difficult to meet the requirements of the automatic combustion control system for real-time and accurate state feedback. In addition, existing automated monitoring technologies based on image analysis usually adopt fixed threshold segmentation or feature engineering methods, and establish a combustion state classification model by extracting statistical quantities such as the flame area and chromaticity distribution. However, this method faces the bottleneck of low classification accuracy in the implementation process. Further, the machine learning solutions for identifying the combustion state of waste incinerator flames are limited by the inherent problems of combustion state annotation. For example, it is difficult for manual annotation to accurately separate discrete categories, and there are significant subjective differences in the annotation standards of different operators.

[0004] Therefore, it can be seen that how to achieve refined monitoring and identification of the combustion state of the incinerator is a technical problem that needs to be overcome at present. Summary of the Invention

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

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

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

[0008] Input the unlabeled flame combustion sample images into the original combustion state clustering and recognition model to obtain the combustion state joint probability matrix of the sample waste incinerator, and determine the mutual information variable based on the combustion state joint probability matrix;

[0009] Train the original combustion state clustering and recognition model in the direction of maximizing the mutual information variable until it meets the preset convergence condition to obtain the target combustion state clustering and recognition model;

[0010] Obtain the current flame combustion image of the target waste incinerator, and input the current flame combustion image into the target combustion situation clustering recognition model to obtain the combustion situation clustering recognition result of the target waste incinerator; wherein, the combustion situation clustering recognition result includes a main clustering recognition label and a super clustering recognition label.

[0011] Applying the embodiment of the present invention effectively solves the problem of lack of labeled data in the recognition of the flame combustion situation of waste incinerators through an unsupervised learning framework; uses the mutual information maximization objective function to drive the model to autonomously mine the essential feature associations in the flame images, and can establish the correspondence between the combustion state and visual features without relying on manual labeling; through the dual-label output of main clustering and super clustering, realizes the hierarchical description of the combustion state, retains both the macroscopic working condition classification ability and captures the microscopic dynamic change features. It can be seen that the embodiment of the present invention reduces the data annotation cost while enhancing the adaptability of the model to complex combustion scenarios, and can realize the refined monitoring and recognition of the combustion state of the incinerator.

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

[0013] According to the calibration results of the internal and external parameters of the target camera, determine the homography matrix between the image plane of the target camera and the grate plane diagram of the waste incinerator;

[0014] Perform grid segmentation on the grate plane diagram of the waste incinerator to obtain the physical space scale grid corresponding to the grate plane diagram of the waste incinerator;

[0015] Map the physical space scale grid to the image plane to obtain an image grid coordinate system;

[0016] Perform block processing on the unlabeled flame combustion sample image through the image grid coordinate system to obtain a plurality of flame combustion sample image blocks;

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

[0018] Input the plurality of flame combustion sample image blocks into the original combustion situation clustering recognition model.

[0019] Applying the embodiments of the present invention, by establishing an accurate mapping relationship between the image plane and the physical space of the grate, it is ensured that the feature extraction process has physical scale consistency. The grid mapping method based on the homography matrix effectively overcomes the perspective distortion problem caused by the difference in the camera installation angle, enabling the image data collected from different perspectives to be feature-compared according to a unified physical benchmark. The combination of physical space grid division and image block processing decouples the global combustion state into local area monitoring, significantly improving the sensitivity of the model to working conditions such as fire line offset and local combustion anomalies, and enhancing the generalization ability of the algorithm to different furnace types.

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

[0021] Using the original combustion situation clustering and recognition model, mapping each block of the flame combustion sample image to a multi-channel space to obtain the color transformation result of each block of the flame combustion sample image;

[0022] Using the original combustion situation clustering and recognition model, performing position transformation processing on each block of the flame combustion sample image to obtain the position transformation result of each block of the flame combustion sample image;

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

[0024] Dividing the multiple blocks of the flame combustion sample image according to the target grid segmentation rule, and extracting the image feature code of each block of the flame combustion sample image in each square corresponding to the target grid segmentation rule;

[0025] Processing the image feature code through the multi-layer perceptron branch in the original combustion situation clustering and recognition model to obtain a one-hot feature vector;

[0026] Determining the combustion situation joint probability matrix using the one-hot feature vector.

[0027] Applying the embodiments of the present invention, by enhancing the construction of a feature invariance learning mechanism through color space transformation and geometric deformation, the model can effectively distinguish the essential features of the combustion state from interference noise. The multi-channel feature coding strategy integrates multi-dimensional visual information such as chromaticity, saturation, and brightness to comprehensively represent the dynamic characteristics of the flame. The grid-based feature extraction method converts irregular image regions into standardized feature vectors, which not only retains the spatial distribution pattern but also eliminates the influence of shape differences. This design significantly improves the robustness of the model to interference factors such as light changes and fuel composition fluctuations through multi-level feature transformation and probability space mapping.

[0028] Preferably, the one-hot feature vector includes a first one-hot vector and a second one-hot vector. The first one-hot vector is determined by a first multi-layer perceptron branch serving as the main clustering head, and the second one-hot vector is determined by a second multi-layer perceptron branch serving as the super-clustering head;

[0029] The step of determining the combustion situation joint probability matrix using the one-hot feature vector includes: determining the combustion situation joint probability matrix based on the transpose of the second one-hot vector and the first one-hot vector.

[0030] Applying the embodiments of the present invention, the joint probability modeling of the main clustering and the super-clustering constructs a hierarchical feature expression system. The main clustering captures the macroscopic category features of the combustion state, and the super-clustering retains the fine-grained difference information of the working conditions. The spatial association of the two types of probability distributions is established through transpose multiplication, effectively mining the internal relationship between different levels of features. This dual-branch structure enables the model to not only meet the basic state classification requirements but also support in-depth analysis of complex working conditions, providing a multi-dimensional decision-making basis for subsequent control strategy formulation and achieving the balance between the accuracy and completeness of the combustion state description.

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

[0032] Applying the embodiments of the present invention, introducing the mutual information influence factor to dynamically adjust the feature weights enables the model to automatically focus on key feature changes during the training process. This mechanism effectively suppresses the interference of irrelevant factors on the clustering process by quantifying the influence degree of color transformation and position offset on feature correlation. Combining with the marginal probability constraint of the joint probability matrix, it ensures the rationality of the feature distribution when the model maximizes the mutual information and avoids falling into local optima. This adaptive optimization strategy significantly improves the model's adaptability to complex feature changes in real working conditions.

[0033] Preferably, the step of determining the mutual information variable using the combustion situation joint probability matrix and the mutual information influence factor includes:

[0034] Determining the mutual information variable according to the combustion situation joint probability matrix, the first marginal probability matrix, the second marginal probability matrix, and the mutual information influence factor; wherein, the first marginal probability matrix is determined by the first one-hot vector; the second marginal probability matrix is determined by the second one-hot vector.

[0035] Applying the embodiments of the present invention, a statistical correlation model of the feature space is established by the mutual information calculation method based on marginal probability constraints, and the stable association pattern between the original features and the enhanced features is accurately quantified. By weighted modification of the joint probability distribution, the requirements of feature invariance and discriminability are effectively balanced. This design enables the model to retain the essential features of the combustion state while fully exploring the common laws under different transformation conditions and enhancing the robustness of feature expression. The coupled optimization mechanism of the probability space provides a reliable mathematical basis for unsupervised learning, ensuring that the clustering results have clear physical interpretability.

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

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

[0038] Determining the mutual information influence factor based on the added results of the random values corresponding to the chrominance channel, saturation channel, and luminance channel in the color transformation result;

[0039] and / or

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

[0041] Applying the embodiments of the present invention, the mutual information influence factor constructed based on multi-channel perturbation quantization realizes the fine control of the changes in color features and spatial features. The combined action of the chrominance, saturation, and luminance factors enables the model to adaptively adjust the degree of attention to the flame color features, and the position offset factor effectively reduces the influence of mechanical motion interference on feature learning. This multi-dimensional weight adjustment mechanism enables the model to maintain a stable feature extraction ability under complex working conditions and significantly improves the adaptability to actual scenarios such as fuel composition changes and grate movement.

[0042] Preferably, the combustion situation clustering and recognition result is realized by maximizing the main cluster one-hot feature and the maximizing super-cluster one-hot feature.

[0043] Applying the embodiments of the present invention, the dual maximization decision mechanism determines the dominant combustion state category through the main cluster and uses the super cluster for fine-grained verification to form a hierarchical decision system. This design not only ensures the reliability of the basic classification but also captures the subtle differences in the working conditions through the probability distribution in the super cluster space. The synergistic effect of the main and super clusters effectively reduces the risk of misjudgment, and in the face of fuzzy combustion states, it can make more accurate judgments through probability fusion, significantly improving the decision confidence of the system under complex working conditions.

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

[0045] In summary, the method for identifying the combustion state of the waste incinerator flame 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, compared with the traditional methods that rely on manual annotation or fixed feature extraction, the embodiments of the present invention can achieve end-to-end clustering recognition based on physical space alignment and feature invariance learning, significantly improving the objectivity of combustion state discrimination and the cross-device generalization ability.

[0046] Specifically, through physical space grid mapping and image block processing, the problem of feature space offset caused by differences in camera installation parameters is effectively overcome. In the comparative tests of waste incinerators with different furnace types, the cross-device consistency error of the fire line height detection is significantly reduced. The introduction of mutual information variables and the maximization training strategy enable the combustion state clustering recognition model to automatically focus on the essential combustion features in the fluctuations of waste components. Even under extreme working conditions with changes in fuel calorific value, the recognition accuracy of the main clusters can still be guaranteed. Further, the hierarchical clustering architecture can achieve the dual functions of determining the certainty of the combustion state and quantifying the uncertainty through the synergistic effect of the main cluster labels and the supercluster labels.

[0047] In addition, the unsupervised learning mechanism in the embodiments of the present invention can eliminate the cost of manual annotation, enabling the flame combustion state recognition system to directly use the historically accumulated video data for model training, thereby improving the data utilization rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, without creative efforts, other relevant drawings can also be obtained based on these drawings.

[0049] Figure 1 It is a flowchart of a method for identifying the combustion state of the waste incinerator flame based on unsupervised learning provided by the embodiments of the present invention.

[0050] Figure 2 It is a network structure diagram of a target combustion state clustering recognition model provided by the embodiments of the present invention.

[0051] Figure 3 It is a module block diagram of a flame combustion state recognition system provided by the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0052] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Usually, the components of the embodiments of the present invention described and illustrated in the accompanying drawings here can be arranged and designed in various different configurations.

[0053] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0054] As Figure 1 shown, the embodiments of the present invention provide a method for identifying the combustion state of the flame of a waste incinerator based on unsupervised learning, which is applied to the above-mentioned flame combustion state recognition system. The specific steps for implementing this method are as shown in Step 110 - Step 140.

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

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

[0057] For another example, video acquisition devices deployed in multiple incineration plants can be used to continuously capture flame combustion video streams in different furnace types, different seasonal operating conditions, and different combustion stages, covering various combustion states such as abnormal fire line height, flame deviation, and brightness fluctuation. During the acquisition process, a uniform sampling strategy on the time axis needs to be adopted, and key frame images are extracted at fixed intervals to ensure that the sample set covers all dynamic features within the combustion cycle.

[0058] For another example, for a mechanical grate incinerator with a daily treatment capacity of 800 tons, it is necessary to intercept flame images with a resolution of 1920×1080 every 5 seconds within a continuous 72-hour operation cycle, and finally form a ten-million-level image database covering the entire process of grate feeding, ignition start, stable combustion, and burnout cooling. All images retain the original sensor data, without manual annotation or preprocessing, and only additional metadata such as acquisition timestamp, furnace number, and camera number are added.

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

[0060] Please refer to Figure 2 , in the embodiment of the present invention, the process of inputting the unlabeled flame combustion sample image into the original combustion situation clustering and recognition model needs to perform multi-level feature transformation and probability space mapping. First, perform grid segmentation based on the physical space scale on the input image, and use the pre-calibrated homography matrix to back-project the physical grid of 30cm×30cm of the grate to the image plane to generate an irregular image block set corresponding to the combustion area. After each image block is decomposed in the HSV color space, a random offset of ±15° is applied to the chrominance channel, a linear adjustment of ±30% is performed on the saturation channel, and a noise perturbation conforming to the normal distribution is superimposed on the luminance channel to generate an enhanced sample with controllable feature changes.

[0061] It can be understood that when processed by the grid flatten layer network layer, each irregular image block is divided into a 25×25 uniform sub-grid, the maximum luminance coordinate in the V channel, the average value in the S channel, and the peak value of the H channel histogram of each sub-region are extracted, and are fused into a 625-dimensional feature vector through learnable weight coefficients. After the above feature vector is input into the encoder module based on the transformer architecture, the multi-head self-attention mechanism is used to capture the cross-region flame dynamic association pattern, and a deep feature representation with spatial perception ability is output. The main clustering head and the super clustering head of the original combustion situation clustering and recognition model respectively generate probability distribution vectors of 9 dimensions and 18 dimensions. By calculating the joint probability matrix of the original sample and the enhanced sample, a weighted mutual information objective function is constructed in combination with the luminance adjustment factor and the position offset factor to quantify the capture ability of the original combustion situation clustering and recognition model for invariant features.

[0062] Step 130: Train the original combustion situation clustering and recognition model in the direction of maximizing the mutual information variable until the preset convergence condition is met to obtain the target combustion situation clustering and recognition model.

[0063] In the embodiment of the present application, the process of training the original combustion situation clustering and recognition model in the direction of maximizing the mutual information variable needs to adopt an adaptive optimization strategy and a dynamic weight adjustment mechanism. At the initial stage of training, the network parameters are updated with a relatively high learning rate to prompt the model to quickly establish a rough association between the flame morphology and the clustering center. As the number of iterations increases, enhanced samples based on luminance transformation and projection transformation are gradually introduced, and the mutual information weight of the corresponding sample pairs is dynamically reduced by the α factor, forcing the network to ignore non-critical feature changes.

[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, significantly reducing the contribution of this transformation to the mutual information calculation and guiding the network to focus on essential features such as the flame distribution pattern. During the optimization process, a joint training strategy of contrast loss and clustering loss is adopted to simultaneously constrain the intra-class compactness and inter-class separation in the feature embedding space.

[0065] Also, for example, when the mutual information gain on the validation set is lower than 0.1% for 5 consecutive epochs, the early stopping mechanism is triggered to save the current optimal model parameters. The trained model should be able to accurately distinguish 9 main combustion states such as abnormal fire line height (e.g., when the fire line exceeds 2 / 3 of the grate height, it is determined as too high) and brightness fluctuation (e.g., when the average value of the V channel is lower than 40, it is determined as too dark) on the holdout test set, and form a fine-grained state division in the hyper-clustering space.

[0066] Step 140: Obtain the current flame combustion image of the target waste incinerator, and input the current flame combustion image into the target combustion situation clustering and recognition model to obtain the combustion situation clustering and recognition result of the target waste incinerator; wherein, the combustion situation clustering and recognition result includes a main clustering recognition label and a hyper-clustering recognition label.

[0067] In the embodiment of the present application, a multi-level decision fusion system needs to be constructed during the process of obtaining the current flame combustion image of the target waste incinerator and performing real-time inference. Specifically, the inference module deployed on the industrial control end receives the real-time video stream from the camera, intercepts key frames at 1-second intervals and inputs them into the target combustion situation clustering and recognition model. The main clustering probability vector output by the target combustion situation clustering and recognition model determines the dominant combustion state category through the argmax operation, and at the same time parses the auxiliary label associated with the main category in the hyper-clustering probability distribution.

[0068] For example, when the main clustering output is "flame burning left", retrieve the predefined mapping table to obtain the corresponding set of hyper-clustering numbers, and calculate the probability mean of these categories as the comprehensive confidence level. The control system triggers the corresponding adjustment instruction according to the confidence threshold: if the confidence level of burning left exceeds 75%, then increase the primary air volume on the right by 15% and increase the grate movement frequency on the right by 20%. All inference results and regulation records are stored in the historical database for subsequent iterative optimization of the target combustion situation clustering and recognition model. For newly emerging undefined combustion modes, the high-probability categories in the hyper-clustering space will trigger the early warning mechanism, prompting the operator to intervene and analyze and update the classification label library to realize the continuous evolution of the capabilities of the target combustion situation clustering and recognition model.

[0069] It can be understood that when implementing the embodiments of the present invention, camera calibration and image block processing directly affect the accuracy of subsequent feature extraction and clustering recognition. Based on this, as an alternative implementation, before inputting the unlabeled flame combustion sample image into the original combustion situation clustering recognition model described in step 120, the method further includes:

[0070] Step 210: Determine the homography matrix between the image plane of the target camera and the grate surface map of the waste incinerator according to the calibration results of the internal and external parameters of the target camera.

[0071] In the embodiments of the present invention, the process of determining the homography matrix between the image plane and the grate surface map according to the calibration results of the internal and external parameters of the target camera requires precise optical calibration and spatial geometry modeling. For example, a high-precision checkerboard calibration plate is used to calibrate the internal parameters of an industrial camera, and the lens distortion coefficient, focal length parameter, and principal point offset are calculated by photographing calibration plate images at different angles. In the external parameter calibration stage, the calibration plate is fixed on the surface of the incinerator grate to establish the spatial correspondence between the physical coordinate system of the grate and the camera image coordinate system.

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

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

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

[0075] Step 230: Map the physically spatially gridded to the image plane to obtain an image grid coordinate system.

[0076] In the embodiments of the present invention, the process of mapping the physically spatially gridded to the image plane to generate an image grid coordinate system needs to apply a homography matrix for perspective projection transformation. Substitute the coordinate of each physical grid corner point established in step 220 into the homography matrix H, and calculate the corresponding image pixel coordinates through homogeneous coordinate transformation. For example, for a grid corner point with physical coordinates (X = 300, Y = 150), after transformation by the H matrix, the image coordinates (u = 1250, v = 680) are obtained, thereby determining the projection area of this physical grid in the image.

[0077] It can be understood that due to the perspective distortion caused by the camera installation angle, the mapped image grid usually presents an irregular quadrilateral shape of a trapezoid or a parallelogram. The image grid coordinate system stores the pixel coordinates of the vertices of each grid in the form of a two-dimensional array, forming an image block index table corresponding one-to-one with the physical grid. This coordinate system ensures that images collected by cameras with different installation angles and different models can be compared for features according to a unified physical scale, improving the robustness of the model to device differences.

[0078] Step 240: Perform block processing on the unannotated flame combustion sample image through the image grid coordinate system to obtain multiple flame combustion sample image blocks.

[0079] In the embodiments of the present invention, the process of performing block processing on the unannotated flame combustion sample image through the image grid coordinate system needs to perform image region extraction based on geometric transformation. For example, according to the vertex coordinate data of the image grid coordinate system, the bilinear interpolation algorithm is used to crop the image blocks corresponding to each grid from the original image.

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

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

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

[0083] In the embodiment of the present invention, the process of mapping the block of the flame combustion sample image to the multi-channel space by using the original combustion situation clustering and recognition model needs to perform HSV color space decomposition and controllable feature enhancement. Specifically, each RGB image block is first converted to the HSV color space, where the hue channel (H) represents the flame color component, the saturation channel (S) reflects the combustion intensity, and the value channel (V) characterizes the thermal radiation level. A random color offset of ±15° is applied to the H channel to simulate the change 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 is linearly adjusted by ±30% to enhance the feature difference of the combustion intensity. The V channel is superimposed with Gaussian noise with a mean of 0 and a standard deviation of 20 to simulate the fluctuation of the illumination condition. The transformed HSV image block and the original block form a sample pair, and the corresponding chromaticity adjustment amount ΔH, saturation scaling factor kS, and luminance noise intensity σV will be used as the calculation parameters of the mutual information influence factor α for the weighted processing of the subsequent joint probability matrix.

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

[0085] In the embodiments of the present invention, the process of performing position transformation processing on the flame combustion sample images in blocks needs to simulate the image offset caused by the change of the camera perspective and the mechanical movement of the grate. For example, the projection transformation algorithm can be used to apply random affine transformations to the image blocks, including vertical translation (simulating the lifting of the grate), horizontal shear (simulating the change of the side view angle), scale scaling (simulating the adjustment of the focal length), and other geometric deformations.

[0086] For another example, a vertical translation transformation is applied to a certain image block to lower the center point by 5% of the image height, simulating the working condition of the flame line position dropping. Each geometric transformation parameter is recorded as the position transformation factor αpos, and when the transformation involves vertical displacement, αpos is set to 0.5 to reduce its contribution to the mutual information calculation. The block image after the position transformation and the original block form a spatial feature comparison pair, forcing the model to learn the flame distribution pattern rather than the absolute position information, and improving the adaptability to the change of the installation position.

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

[0088] In the embodiments of the present invention, the process of determining the mutual information influence factor based on the color transformation result and the position transformation result needs to establish a multi-dimensional feature weight fusion mechanism. Specifically, the chromaticity adjustment factor αH = 1 - |ΔH| / 180, the saturation factor αS = 1 - |kS - 1|, the brightness factor αV = 1 - σV / 255 in Step 121 and the position transformation factor αpos in Step 122 are weighted and multiplied to obtain the comprehensive mutual information influence factor α = αH × αS × αV × αpos. For example, when a certain sample pair includes 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 participate in the joint probability matrix calculation as a weight coefficient, 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 codes 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 the embodiments of the present invention, the process of feature encoding for segmented flame combustion sample images according to the 25×25 grid segmentation rule needs to achieve feature standardization for irregular regions. Each image segment is first evenly divided into 625 sub-grids with 25 rows and 25 columns, maintaining the same grid division density regardless of the shape of the original segment. 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 H channel histogram within the covered area.

[0091] For example, the covered area of the top-left sub-grid of a trapezoidal image segment may only contain some valid pixels. In this case, the 95th percentile of the V-channel pixel values within this area is calculated as an approximation of the maximum brightness. The three feature quantities are linearly combined through 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 the irregular segment shape on the feature dimension.

[0092] Step 125: Process the image feature encoding through the multi-layer perceptron branch in the original combustion situation clustering recognition model to obtain a one-hot feature vector.

[0093] In the embodiments of the present application, the process of processing image feature encoding through the multi-layer perceptron branch in the original combustion situation clustering recognition model needs to complete the mapping from the feature space to the clustering labels. For example, the 625-dimensional grid flatten 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, and the final output layer generates a 9-dimensional main clustering probability distribution through the softmax function.

[0094] Meanwhile, another independent multi-layer perceptron branch processes the feature vector with the same structure and outputs an 18-dimensional super-clustering probability distribution. For example, when the input feature represents a left-biased burning mode, the main clustering branch may output a high probability of 0.85 in class 1 (left-biased burning), while the super-clustering branch outputs probability values of 0.6, 0.3, and 0.7 in classes 2, 4, and 5 respectively. The parameters of the two branches are jointly optimized through the mutual information maximization objective function during the training process to ensure that the main clustering captures the macroscopic combustion state and the super-clustering retains the fine-grained differential features, jointly constructing a hierarchical combustion situation description system.

[0095] Step 126: Use the one-hot feature vector to determine the combustion situation joint probability matrix.

[0096] In the embodiment of the present application, determining the joint probability matrix of the combustion situation by using the one-hot feature vector includes: Step 1260: determining the joint probability matrix of the combustion situation according to the transpose of the second one-hot vector and the first one-hot vector.

[0097] In the embodiment of the present invention, the process of determining the joint probability matrix of the combustion situation according to the transpose of the second one-hot vector and the first one-hot vector needs to perform a probability space alignment operation. For example, the 9-dimensional probability distribution vector P(m) output by the main clustering branch is multiplied by the 18-dimensional probability distribution vector P(g) output by the transformation branch to generate a 9×18 joint probability matrix Pmg = P(m)×P(g)^T. The matrix element Pmg[i, j] represents the joint probability distribution that the original image block belongs to the i-th category of the main clustering and its transformed version belongs to the j-th category of the super clustering. For another example, when the probability of the 3rd category (moderate fire line) of the main clustering is 0.8 and the probability of the 7th category (moderate fire line - brightness fluctuation) of the super clustering after transformation 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 idea, the implementation of determining the mutual information variable based on the joint probability matrix of the combustion situation in step 120 can use the joint probability matrix of the combustion situation and the mutual information influence factor to determine the mutual information variable, and further specifically: determining the mutual information variable according to the joint probability matrix of the combustion situation, the first marginal probability matrix, the second marginal probability matrix, and the mutual information influence factor; wherein, the first marginal probability matrix is determined by the first one-hot vector; the second marginal probability matrix is determined by the second one-hot vector.

[0099] In the embodiment of the present application, the implementation process of determining the mutual information variable based on the joint probability matrix of the combustion situation and the mutual information influence factor needs to construct a multi-dimensional probability space coupling model. Specifically, when implemented, the first one-hot vector output by the main clustering multi-layer perceptron branch is normalized by softmax to generate the first marginal probability matrix P(m), which is a 9-dimensional row vector, and each element represents the probability that the sample belongs to the corresponding main clustering category. The second one-hot vector output by the transformation multi-layer perceptron branch is normalized by softmax to generate the second marginal probability matrix P(g), which is an 18-dimensional row vector, representing the distribution probability of the sample in the super clustering space after transformation. The Hadamard product operation is performed on the joint probability matrix Pmg = P(m)^T×P(g) and the mutual information influence factor α to obtain the corrected joint probability matrix Pmg' = α⊙Pmg. Calculate the mutual information variable I(m; g) = ΣΣPmg'[i, j]×log(Pmg'[i, j] / (P(m)[i]×P(g)[j])), where the logarithmic term quantifies the mutual correlation of the two probability distributions.

[0100] For example, when the probability of the second type of main cluster (too low live wire) is 0.9, the probability of the corresponding transformed sample in the fifth type of super cluster (too low live wire - uneven fuel) is 0.7, and the mutual information influence factor α = 0.8, the joint probability contribution value of this sample pair is 0.9×0.7×0.8 = 0.504, which significantly affects the overall mutual information calculation result. This design ensures that the model can accurately distinguish the essential combustion characteristics and interference transformation characteristics during the optimization process by introducing marginal probability constraints and factor weighting mechanisms.

[0101] As a preferred design idea, determining the mutual information influence factor based on the color transformation result and the position transformation result described in step 123 includes: determining the mutual information influence factor based on the added results of random values corresponding to the hue channel, saturation channel, and brightness channel in the color transformation result; and / or determining the mutual information influence factor based on the change in the longitudinal position offset in the position transformation result.

[0102] In the embodiments of the present invention, a multi-channel perturbation quantization system needs to be established in the process of determining the mutual information influence factor based on the color transformation and position transformation results. For example, a random rotation perturbation of ±15° is applied to the hue channel of the HSV space, and the hue 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 coefficient. Gaussian noise with a mean of μ and a standard deviation of σ is added to the brightness channel, and the brightness influence factor αV = 1 - σ / 255.

[0103] Another example is that in terms of position transformation, a projection transformation with a longitudinal offset of ΔY pixels is applied to the image blocks, and the position 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 includes 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 this sample pair in the mutual information calculation, forcing the model to ignore the interference caused by hue offset and position change and focus on stable features such as the flame distribution pattern.

[0104] As a preferred design idea, the combustion situation clustering and recognition result is achieved by maximizing the main cluster one-hot feature and the super cluster one-hot feature.

[0105] In the embodiments of the present invention, a hierarchical decision fusion mechanism needs to be constructed in the process of realizing combustion situation clustering recognition by maximizing the main clustering one-hot feature and the super-clustering one-hot feature.

[0106] For example, in the online inference stage, after the input image is processed by grid partitioning, the main clustering branch outputs a 9-dimensional probability vector Pm, and the maximum value index is taken to determine the dominant combustion state category. At the same time, the super-clustering branch outputs an 18-dimensional probability vector Pg, retrieves the predefined main-super clustering mapping table, and obtains the set of super-clustering numbers associated with the main category. For example, when the main clustering is determined to be category 3 (flame burning to the right), the associated super-clustering numbers are 7, 11, and 15, and the probability mean of these categories is calculated as the comprehensive confidence. The control system sets a dual-threshold trigger mechanism: if the main clustering probability exceeds 0.85 and the comprehensive confidence is greater than 0.7, the control instruction is immediately executed; if the main clustering probability is between 0.6 and 0.85, the multi-modal sensor review process is started.

[0107] Another example is that a sample of a right-burning condition is processed by the model to obtain Pm[3]=0.92, the associated super-clustering probabilities Pg[7]=0.6, Pg

[11] =0.8, and Pg

[15] =0.7, and the comprehensive confidence is (0.6 + 0.8 + 0.7) / 3 = 0.7, triggering the control actions of reducing the right damper by 10% and increasing the left secondary air volume by 15%. This design ensures the reliability of the recognition result through a dual probability maximization mechanism, while retaining the representation ability of the super-clustering space for complex working conditions.

[0108] As an optional but non-limiting embodiment, training the original combustion situation clustering recognition model in the direction of maximizing the mutual information variable until meeting the preset convergence condition to obtain the target combustion situation clustering recognition model includes:

[0109] Input the segmented multiple flame combustion sample images into the original combustion situation clustering recognition model to generate the color transformation result and position transformation result of the segmented flame combustion sample images;

[0110] Based on the chrominance channel perturbation amplitude, saturation channel scaling ratio, luminance channel noise intensity in the color transformation result, and the longitudinal offset ratio in the position transformation result, calculate the mutual information influence factor;

[0111] Perform grid partitioning on the segmented flame combustion sample images according to the 25*25 grid segmentation rule, extract the chrominance value, saturation value, and luminance value of the pixel point with the largest luminance channel median value in each grid, and generate the grid feature encoding of the segmented flame combustion sample images through weighted fusion;

[0112] Concatenate the grid feature encoding and the physical space position encoding of the segmented flame combustion sample images into a joint feature vector;

[0113] Input the combined feature vector into the main clustering multi-layer perceptron branch of the original combustion situation clustering recognition model, and generate the first one-hot vector by gradually reducing the dimension through three fully connected layers;

[0114] Input the grid feature encoding corresponding to the position transformation result into the transformation multi-layer perceptron branch of the original combustion situation clustering recognition model, and generate the second one-hot vector through a fully connected layer with shared weights;

[0115] Construct the combustion situation joint probability matrix based on the matrix product of the first one-hot vector and the second one-hot vector, and use the mutual information influence factor to weight and correct each element of the combustion situation joint probability matrix;

[0116] Calculate the mutual information variable according to the corrected combustion situation joint probability matrix, the first marginal probability matrix output by the main clustering multi-layer perceptron branch, and the second marginal probability matrix output by the transformation multi-layer perceptron branch;

[0117] Adjust the parameters of the fully connected layers of the main clustering multi-layer perceptron branch and the transformation multi-layer perceptron branch through the backpropagation algorithm, so that the mutual information variable continuously increases during the iteration process;

[0118] When the growth rate of the mutual information variable is lower than the preset threshold after continuously reaching the preset number of iterations, it is determined that the original combustion situation clustering recognition model meets the preset convergence condition, and the current model parameters are saved as the target combustion situation clustering recognition model.

[0119] In the embodiment of the present invention, in the process of inputting the multiple flame combustion sample images in blocks into the original combustion situation clustering recognition model to generate the color transformation result and the position transformation result, first perform HSV color space decomposition on each image block, and apply a random angular rotation perturbation to the chromaticity channel. For example, the original chromaticity channel main peak of a certain image block is 30° (orange flame), and after adding a 12° perturbation, it becomes 42° (orange-red flame), simulating the color shift caused by the change of fuel composition. At the same time, apply a vertical position offset of 8% of the height to the same block, so that the central area moves down 20 pixels, simulating the change of the fire line position caused by the lifting of the grate. The color transformation result retains the chromaticity, saturation, and brightness channel data after perturbation, and the position transformation result records the geometric deformation parameters, providing a basis for the subsequent calculation of the mutual information influence factor.

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

[0121] Next, when dividing the flame combustion sample image into blocks by the 25*25 grid slicing rule, each irregular image block is evenly divided into 625 sub-grids. For example, after a trapezoidal block is mapped in the image plane, the upper left sub-grid only covers one-third of the effective pixel area, and the pixel point with the largest median value in the luminance channel is still searched within this sub-grid. The chromaticity value of 45°, saturation value of 220, and luminance value of 180 of this pixel point are weighted and fused through the learnable weight coefficients w1 = 0.6, w2 = 0.3, and w3 = 0.1, and the feature encoding value of this sub-grid is obtained as 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 to form a 625-dimensional vector representing the local flame core features.

[0122] Furthermore, when splicing the grid feature encoding with the physical space position encoding of the flame combustion sample image block to form a joint feature vector, the physical space position encoding is derived from the normalized coordinates of the grate grid. For example, a certain block corresponds to the 15th column and 6th row physical grid of the grate, and its central point physical coordinates are (X = 450 cm, Y = 180 cm), which are normalized to (0.5625, 0.6) in an 8m × 3m grate system. This two-dimensional position encoding is spliced 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 the local features of the flame and enhancing the ability to identify abnormal fire line heights.

[0123] Furthermore, when the combined feature vector is input into the main clustering multi-layer perceptron branch of the original combustion situation clustering and recognition model to generate the first one-hot vector, the 627-dimensional feature is first reduced to 256 dimensions through the first fully connected layer, and the activation function uses ReLU to eliminate negative values. The second fully connected layer further compresses it to 64 dimensions to capture the high-order spatial-color correlation patterns across grids. The final output layer generates a 9-dimensional probability distribution through the softmax function. For example, for the category of excessive fire line, a probability value of 0.85 is output, indicating that the model is highly confident that the current block presents a combustion situation with an excessive fire line. This first one-hot vector directly represents the classification result of the combustion state in the main clustering space.

[0124] In addition, when the grid feature encoding corresponding to the position transformation result is input into the transformation multi-layer 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, the block after position transformation 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 the independent output layer. If the main clustering of the original block is a moderate fire line, its transformed block may output a probability of 0.7 in the category of moderate fire line - brightness fluctuation in the super-clustering space, reflecting the fine-grained differences brought by brightness changes. This design ensures that the transformation branch focuses on capturing the state changes caused by perturbations while sharing the basic feature extraction ability.

[0125] Specifically, when constructing the combustion situation joint probability matrix based on the matrix product of the first one-hot vector and the second one-hot vector, the outer product operation is performed on the 9-dimensional probability vector of the main clustering and the 18-dimensional probability vector of the super-clustering. For example, if the probability of the second category (low fire line) in the main clustering is 0.9 and the probability of the fifth category (low fire line - uneven fuel) in the super-clustering is 0.6, then the initial value of the element (2,5) in the joint probability matrix is 0.9×0.6 = 0.54. The mutual information influence factor 0.124 is used to weight and correct this element, and the corrected value is 0.54×0.124≈0.067. This operation significantly reduces the contribution of the joint probability of sample pairs with strong color or position perturbations, forcing the model to learn stable combustion features that are not affected by interference.

[0126] It can be understood that when calculating the mutual information variable based on the corrected joint probability matrix of the combustion situation, the first marginal probability matrix output by the main clustering multi-layer perceptron branch, and the second marginal probability matrix output by the transformation multi-layer perceptron branch, the weighted mutual information formula is adopted. For example, the value of the element (2,5) in the corrected 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. Summing up the contribution values of all matrix elements gives the overall mutual information variable. The larger this value is, the stronger the model's ability to extract invariant features.

[0127] In the embodiment of the present invention, when adjusting the parameters of the fully connected layers of the main clustering multi-layer perceptron branch and the transformation multi-layer perceptron branch through the backpropagation algorithm, maximizing the mutual information variable is used as the optimization goal. Calculate the gradient of the mutual information variable with respect to the network parameters, and use the adaptive moment estimation optimizer to update the weights. For example, when the weight of a certain fully connected layer causes the mutual information variable to decrease, the gradient direction indicates that the weight needs to be increased to enhance the feature correlation. After multiple iterations, the model gradually learns to ignore color perturbations and position offsets and focus on essential features such as the flame distribution pattern.

[0128] It can be understood that when the growth rate of the mutual information variable is continuously lower than a preset threshold after reaching a preset number of iterations, for example, the mutual information gain is less than 0.05% within 10 consecutive training cycles, it is determined that the model reaches the convergence condition. Save the parameters of the fully connected layers of the main clustering branch and the transformation branch at this time to form the target combustion situation clustering recognition model. It can be understood that the above target combustion situation clustering recognition model has robustness to color changes and position offsets, can accurately identify 9 main combustion states such as abnormal fire line height and side burning direction, and distinguish 18 fine-grained working condition variants in the super-clustering space.

[0129] In other words, the step of training the original combustion situation clustering recognition model in the direction of maximizing the mutual information variable until it meets the preset convergence condition to obtain the target combustion situation clustering recognition model is realized by the double-branch mutual information maximization training strategy. The training strategy includes: color and position transformation feature extraction (chromaticity perturbation, saturation scaling, brightness noise, longitudinal offset), 25×25 grid feature encoding generation (fusion of the three-channel values of the maximum brightness pixels), joint feature vector construction (concatenation of grid encoding and physical space position), double-branch probability modeling (dimensionality reduction of the main clustering MLP branch and weight sharing of the transformation MLP branch), mutual information weighted correction (adjustment of the elements of the joint probability matrix), marginal probability mutual information calculation and backpropagation optimization, and iteration increase threshold termination condition.

[0130] Figure 3 The block diagram of a flame combustion situation recognition system 10 provided by an embodiment of the present invention is shown. The flame combustion situation recognition system 10 in the embodiment of the present invention has functions of data storage, transmission, and processing. For example, 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, the processor 12, and the network module 13 are directly or indirectly electrically connected 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 recognition device 20. The flame combustion situation recognition device 20 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 software programs and modules stored in the memory 11, such as the flame combustion situation recognition device 20 in the embodiment of the present invention, thereby implementing the WEB-based data logic expression processing method in the embodiment of the present invention.

[0132] Among them, the memory 11 can be, but is not limited to, a random access memory (RAM), a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), etc. Among them, the memory 11 is used to store a program, and the processor 12 executes the program after receiving an execution instruction.

[0133] The processor 12 may be an integrated circuit chip with data processing capabilities. The above-mentioned 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 various methods, steps, and logic block diagrams disclosed in the embodiments of the present invention. The general-purpose processor can be a microprocessor or the processor can also be any conventional processor, etc.

[0134] The network module 13 is used to establish a communication connection between the flame combustion situation recognition system 10 and other communication terminal devices through the network, and realize the transceiver operations of network signals and data. The above network signals may include wireless signals or wired signals.

[0135] It can be understood that Figure 3 the structure shown is only schematic, and the flame combustion situation recognition system 10 may further include Figure 3 more or fewer components than those shown, or have a different configuration from Figure 3 that shown. Figure 3 Each component shown can be implemented by hardware, software or a combination thereof.

[0136] The embodiment of the present invention also provides a computer-readable storage medium, and the readable storage medium includes a computer program. When the computer program runs, it controls the flame combustion situation recognition system 10 where the readable storage medium is located to execute the above-mentioned method for recognizing the flame combustion situation of a waste incinerator based on unsupervised learning.

[0137] The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. A method for identifying the combustion state of the flame in a waste incinerator based on unsupervised learning, characterized in that, The method is applied to a flame combustion situation recognition system, and the method includes: Obtain unlabeled flame combustion sample images of a sample waste incinerator; Input the unlabeled flame combustion sample images into an original combustion situation clustering and recognition model to obtain a combustion situation joint probability matrix of the sample waste incinerator, and determine a mutual information variable based on the combustion situation joint probability matrix; Train the original combustion situation clustering and recognition model in the direction of maximizing the mutual information variable until a preset convergence condition is met to obtain a target combustion situation clustering and recognition model; Obtain the current flame combustion image of a target waste incinerator, and input the current flame combustion image into the target combustion situation clustering and recognition model to obtain a combustion situation clustering and recognition result of the target waste incinerator; wherein, the combustion situation clustering and recognition result includes a main clustering recognition label and a super-clustering recognition label.

2. The method according to claim 1, wherein Before inputting the unlabeled flame combustion sample images into the original combustion situation clustering and recognition model, the method further includes: Determine a homography matrix between the image plane of the target camera and the grate surface map of the waste incinerator according to the calibration result of the internal and external parameters of the target camera; Perform grid segmentation on the grate surface map of the waste incinerator to obtain a physical space scale grid corresponding to the grate surface map of the waste incinerator; Map the physical space scale grid to the image plane to obtain an image grid coordinate system; Perform block processing on the unlabeled flame combustion sample images through the image grid coordinate system to obtain a plurality of flame combustion sample image blocks; The inputting the unlabeled flame combustion sample images into the original combustion situation clustering and recognition model includes: Input the plurality of flame combustion sample image blocks into the original combustion situation clustering and recognition model.

3. The method according to claim 2, wherein The inputting the unlabeled flame combustion sample images into the original combustion situation clustering and recognition model to obtain a combustion situation joint probability matrix of the sample waste incinerator includes: Using the original combustion situation clustering and recognition model, map 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 situation clustering and recognition model, perform position transformation processing on each flame combustion sample image block to obtain a position transformation result of each flame combustion sample image block; Determine a mutual information influence factor based on the color transformation result and the position transformation result; Divide the plurality of flame combustion sample image blocks according to a target grid splitting rule, and extract the image feature encoding of each flame combustion sample image block in each square corresponding to the target grid splitting rule; Process the image feature encoding through a multi-layer perceptron branch in the original combustion situation clustering and recognition model to obtain a one-hot feature vector; Determine the combustion situation joint probability matrix using the one-hot feature vector.

4. The method according to claim 3, characterized in that The one-hot feature vector includes a first one-hot vector and a second one-hot vector. The first one-hot vector is determined by a first multi-layer perceptron branch serving as the main clustering head, and the second one-hot vector is determined by a second multi-layer perceptron branch serving as the super-clustering head; Determining the combustion situation joint probability matrix by using the one-hot feature vector includes: determining the combustion situation joint probability matrix according to the transpose of the second one-hot vector and the first one-hot vector.

5. The method according to claim 3, characterized in that, Determining the mutual information variable based on the combustion situation joint probability matrix includes: using the combustion situation joint probability matrix and the mutual information influence factor to determine the mutual information variable.

6. The method according to claim 5, characterized in that, Using the combustion situation joint probability matrix and the mutual information influence factor to determine the mutual information variable includes: Determining the mutual information variable according to the combustion situation joint probability matrix, the first marginal probability matrix, the second marginal probability matrix, and the mutual information influence factor; wherein, the first marginal probability matrix is determined by the first one-hot vector; the second marginal probability matrix is determined by the second one-hot vector.

7. The method according to claim 3, wherein The target grid segmentation rule is a 25*25 grid segmentation rule.

8. The method according to claim 3, characterized in that, Based on the color transformation result and the position transformation result to determine the mutual information influence factor, including: Determining the mutual information influence factor based on the added results of the random values corresponding to the hue channel, saturation channel, and brightness channel in the color transformation result; And / or Determining the mutual information influence factor based on the change of the longitudinal position offset in the position transformation result.

9. The method according to claim 1, characterized in that The combustion situation clustering and recognition result is realized by maximizing the main clustering one-hot feature and the super-clustering one-hot feature.

10. The method according to claim 2, wherein Performing grid segmentation on the grate surface diagram of the waste incinerator includes: dividing the grate surface diagram of the waste incinerator into grids of 30cm*30cm.

Images