MSWI Process Combustion State Recognition Method Based on Multi-Feature Fusion and Improved Cascade Forest

Abstract

The present invention relates to a method for identifying the combustion state of the MSWI process based on multi-feature fusion and improved cascade forest. First, preprocessing means such as artificial multi-exposure image fusion, feature normalization, notch filtering, and median filtering are used to de-haze and denoise the MSWI process flame image to obtain a clear image; then, multi-view features such as the brightness, flame, color, and principal components of the image are extracted and fused; to improve the efficiency and accuracy of the recognition model, the multi-view features are reduced based on mutual information (MI) to reduce the input variable dimension; finally, the traditional cascade forest is improved to include three base learners, namely random forest (RF), completely random forest (CRF), and gradient boosting decision tree (GBDT), in each cascade layer, and the above multi-view features are used as the input of the ICF to establish an ICF-based MSWI process combustion state recognition model. Experiments show that the constructed model can achieve a recognition accuracy of more than 96%.

Description

Technical Field

[0001] The present invention relates to the field of urban solid waste incineration process, and in particular to the research on the method for identifying the combustion state during the incineration process. Background Art

[0002] The growth of urbanization level and the improvement of living standards have led to an increasing generation rate of municipal solid waste (MSW), and the types thereof are also constantly increasing. If not treated in time, it will cause serious environmental pollution problems. A series of "acclimatization" phenomena have occurred in the actual operation process of the MSWI technology introduced from abroad, such as unstable combustion state, and there are faults such as furnace coking, ash accumulation, corrosion, and even potential safety hazards such as explosion. The reason is that domestic MSW has characteristics such as complex components, low calorific value, and large moisture change range. Therefore, in order to ensure the sufficiency, stability, safety of the combustion process and reduce the pollutant emission concentration, it is necessary to accurately identify the material combustion state. In view of the complex and changeable operating conditions in the furnace, it is of great practical significance to adopt effective and feasible means to identify the combustion state of the MSWI process. At present, the basis for classifying the combustion state of the MSWI process is mainly the position of the flame combustion line. In the industrial field, there is still no relatively feasible technical means for the automatic identification and detection of the position of the combustion line. There are still great technical breakthrough problems in the combustion state identification based on the position of the combustion line. If the characteristic information contained in the flame image can be effectively extracted, a multi-view feature database is established and fully mined, it will bring new solutions for effectively identifying the combustion state of the MSWI process. There are the following difficulties in the combustion state identification of the MSWI process: 1) The rich information contained in the flame image needs to be fully extracted, which may require the adoption of various feature acquisition strategies; 2) Marking the training samples requires extremely high labor costs and is even infeasible due to severe interference, while a small number of marked sample sets are difficult to meet the training requirements of the neural network; 3) How to fully mine the deep features of the flame image with a small number of samples is a major difficulty. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to invent a method for identifying the combustion state of the MSWI process based on multi-feature fusion of combustion flame and improved cascade forest.

[0004] To solve the above technical problems, the present invention provides a modeling strategy composed of three parts: an image preprocessing module, a multi-view feature extraction and selection module, and a multi-improved cascade forest (ICF) recognition module. The definitions of relevant variables and symbols are as follows: It represents the flame image with the marked combustion state in the MSWI process, where I n (u, v) represents the nth image, N is the number of image samples, and (u, v) represents the pixel point coordinates in each image; Represents the preprocessed image; Denotes the set of extracted reduced flame features; K represents the number of generated underexposed images, γ is the gamma exposure operation coefficient, r represents the guiding filter radius, α, β, σ Dk , σ B Are positive adjustment factors, r notch Is the filtering radius, s a Represents the side length of the median filter window, θ th Denotes the pixel threshold of the flame effective area, ω th Represents the pixel threshold of the high-temperature area; θ MI Denotes the MI threshold, Tn represents the number of decision trees in RF and CRF, η and Respectively represent the GBDT learning rate and the number of iterations, minsamples represents the minimum number of samples per decision tree, Represents the initial input sample Belongs to the function estimate value of the k P Class; Slide represents the sliding window size used when extracting color features; Represents the combustion state recognition result output by the model; y is the manually marked combustion state.

[0005] The method for constructing the combustion state recognition in the MSWI process includes the following steps:

[0006] Step 1: Image preprocessing module: Eliminate the noise introduced in the image due to on-site environment and transmission channels, etc., and separate the flame from the furnace background for subsequent image feature extraction;

[0007] Step 2: Multi-view feature extraction and selection module: Extract multi-view features such as the brightness, flame, color, and principal components of the combustion image, and select features based on MI;

[0008] Step 3: Improved cascade forest (ICF) recognition module: Use the selected multi-view features as the input of the ICF recognition module, and use RF, CRF, and GBDT as the base learners of the cascade forest to recognize the combustion state and obtain the recognition result;

[0009] Due to the above technical solutions adopted by the present invention, the following technical effects are achieved:

[0010] The present invention is used to construct an identification model for the combustion state in the MSWI process. The proposed method combines multi-view features and ICF, and uses 3 base learners in each layer, which improves the diversity and can fully explore deep features. It has an identification accuracy of 96.01%, and has a high precision rate (96.22%), recall rate (96.11%), and F1 score (96.17%), which verifies that the constructed model has good stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 is the modeling strategy diagram of the present invention;

[0012] Figure 2 (a) is the furnace shutdown screen; Figure 2 (b) is the diagram for dividing the burnout line forward movement basis; Figure 2 (c) is the diagram for dividing the normal burnout line basis; Figure 2 (d) is the diagram for dividing the burnout line backward movement basis;

[0013] Figure 3 (a) is the original combustion flame image; Figure 3 (b) is the dehazed image; Figure 3 (c) is the feature normalized image; Figure 3 (d) is the image after notch filtering; Figure 3 (e) is the image after median filtering;

[0014] Figure 4 (a) is the image after preprocessing; Figure 4 (b) is the image after preprocessing; Figure 4 (c) is the V-channel image; Figure 5 is the V-channel equal value diagram;

[0015] Figure 6 (a) is the image of working condition 1; Figure 6 (b) is the image of working condition 2; Figure 6 (c) is the image of working condition 3; Figure 6 (d) is the result of color moment feature extraction under different working conditions;

[0016] Figure 7 (a) is the contribution rate of the first principal component of working condition 1; Figure 7 (b) is the contribution rate of the first principal component of working condition 2; Figure 7 (c) is the contribution rate of the first principal component of working condition 3. DETAILED DESCRIPTION OF THE INVENTION

[0017] Image preprocessing module

[0018] During the MSWI process, combustion in the furnace will generate a large amount of fly ash and smoke, and affected by the high-temperature environment and the use of analog signals for image transmission by the image acquisition device, the obtained color RGB image inevitably introduces signal interference and other physical noises. Therefore, it is necessary to preprocess the image to restore the on-site image as much as possible The preprocessing process of the nth flame image is as follows. 1. Dehazing algorithm based on artificial multi-exposure image fusion

[0019] First, for the original flame image I n(u, v) undergoes a gamma exposure operation to generate K underexposed images K represents the number of underexposed images generated, and K = 6. The k-th image {I n (u, v)} k is generated as follows:

[0020] {I n (u, v)} k = I n (u, v) γ (1)

[0021] where γ is the gamma exposure operation coefficient, and γ = [1, 2, 4, 6, 8, 10]. For simplicity of representation, I k is used to represent {I n (u, v)} k .

[0022] I k is decomposed into a global component B k and a local component D k through a guided filter, as follows:

[0023] B k = G r ∈(L k , L k ) (2)

[0024] D k = I k - B k , k = 1,..., K (3)

[0025]

[0026] where G r represents the guided filter operation, and r = 12 is the filter radius used to control the blur; L k is the grayscale image of the image, which serves as both the input image and the guidance image for the guided filter, and R Channel , G Channel , B Channel represent the three-channel components of the image respectively.

[0027] The exposure weight k of each pixel point in D can be calculated by the following formula:

[0028]

[0029] where is obtained by convolving L k with an average filter of size 7×7, and is the adjustment factor.

[0030] For each global component, obtain the exposure weight using it as the initial exposure feature

[0031]

[0032] Wherein, represents the average grayscale image, m and n respectively represent the length and width of the image, α, σ B are adjustment factors, σ B = 0.5, α = 0.2.

[0033] Perform weighted averaging on the global and local components of the input image to obtain the defogged image Z:

[0034]

[0035] Wherein, β is an adjustment factor, β = 1.1.

[0036] Finally, obtain the original flame image I n The defogged image Z of (u, v) n (u, v).

[0037] 2. Feature normalization

[0038] Feature normalization maps the image pixels from 0 - 255 to 0 - 1, thereby reducing the computational complexity to improve the model operation efficiency, and at the same time separating the flame from the furnace background. The adopted zero-mean normalization is as follows:

[0039]

[0040] Wherein, is the mean of the image, making is the standard deviation of the image, making W n (u, v) is the normalized image.

[0041] 3. Notch filtering

[0042] The image during the transmission process will have frequency-domain interference due to the influence of the high-temperature environment, resulting in stripe noise in the collected flame image. Therefore, it is necessary to eliminate the frequency band where the stripe noise is located through frequency-domain filtering. The notch filter belongs to a band-stop filter, whose stop band has a zero signal frequency point and is 1 at other frequencies. The image obtained after notch filtering is

[0043] V n (u, v) = notch{W n (u, v), rnotch} (9)

[0044] Among them, notch represents the notch filtering operation, and r notch is the filtering radius, and r notch = min(m,n) / 5

[0045] 4. Median filtering

[0046] Isolated noise points in the image caused by fly ash need to be eliminated by median filtering, and a 5×5 window is used for sliding operation. The median value of the window pixels is assigned to the center of the template, and then the filtered image is obtained, denoted as

[0047]

[0048] Among them, median represents the median filtering operation.

[0049] Furthermore, is denoted as the set of finally preprocessed flame images.

[0050] Multi-view feature extraction and selection module

[0051] Multi-view feature extraction sub-module

[0052] During the MSWI process, to ensure that the materials are burned as fully as possible, the grate is always in periodic motion, and the flame images show regular change characteristics. Considering the rich information contained in the flame images, multi-view features are considered for characterization. The nth flame image The feature extraction process is as follows.

[0053] 1. Brightness feature

[0054] For the flame images in the MSWI process, their brightness features can be described from the following multiple perspectives.

[0055] (1) Average gray value: The original image is converted to a grayscale image (Gonzalez, 2007) through the following formula:

[0056]

[0057] Among them, and respectively represent the R Channel , G Channel and B Channel color components at the pixel point (u, v) of the nth image.

[0058] The calculation formula for the average gray value of the image is as follows:

[0059]

[0060] wherein, represents the th pixel point, m and n respectively represent the number of pixel points in the length and width directions of the image.

[0061] (2) Variance of gray value: Calculated by the following formula:

[0062]

[0063] (3) Average brightness value: The above two features mainly describe the brightness feature from the perspective of the computer. Different from the characteristics of the RGB space, the expression of the image in the HSV space is closer to the human perception experience of color. It represents color as a linear combination of three components: color (H component), vividness (S component), and light and dark degree (V component). Therefore, after converting the flame image from the RGB space to the HSV space, extracting the brightness feature of the V channel can better reflect the brightness expression of the flame image from the perspective of human vision.

[0064] Here, is converted to the HSV image space and represented as Select the V channel Calculate the mean of its pixel values as the average brightness value of the image:

[0065]

[0066] (4) Variance of brightness value: The calculation formula is as follows:

[0067]

[0068] 2. Flame features

[0069] The flame features include the area feature of the flame effective area and the area feature of the high-temperature area. Their calculation is based on the V channel because the human eye mainly relies on the brightness feature to distinguish the flame from the background.

[0070] (1) Area of the flame effective area: Defined as the total number of pixel points in the image whose brightness value is greater than the specified threshold θ th , θ th = 0.226 (given method) as follows:

[0071]

[0072] wherein, Θ(·) is the unit step function.

[0073] (2) Area of the high-temperature flame zone: defined as the total number of pixels in the image with a brightness value greater than the specified threshold ω th where ω th = 0.941 (given method) as follows:

[0074]

[0075] 3. Color features

[0076] The principle of flame generation is that atoms in the ground energy state are excited, causing the outer electrons of the nucleus to absorb energy and transition to higher energy levels. When the unstable electrons in the excited state transition to lower energy levels, they emit light with a certain energy and wavelength, and the physical phenomenon corresponding to the wavelength is the color feature of the flame. The mathematical basis of color moments is that any color distribution in an image can be represented by its moments. In this paper, the first-order, second-order, and third-order moments are used to express the color information of the flame image.

[0077] The nth preprocessed flame image Its first-order moment is calculated as follows:

[0078]

[0079] Second-order moment The calculation formulas for each item are as follows:

[0080]

[0081] where represents the first-order moment feature of each color channel of the nth image.

[0082] Third-order moment The calculation formulas for each item are as follows:

[0083]

[0084] A sliding window with a size of 1 / 5 of the original image is used to extract color moment features, and the extracted color features are denoted as:

[0085]

[0086] 4. Principal component features

[0087] The purpose of performing image feature extraction tasks based on principal components analysis (PCA) is to find the basis images so that they contain as much image information as possible.

[0088] To prevent memory shortage during calculation due to the too large size of the original image, the original image is reduced to 0.1 times the original size, and its size is denoted as f×t×3, denoted as Further, it is reconstructed into a matrix χ n (ft×3), then its covariance matrix C n is calculated as follows:

[0089]

[0090] where, is the mean value of C n .

[0091] Further, the eigenvalues λ n of C n and eigenvectors π n can be calculated. The eigenvalues are arranged in descending order, and the eigenvectors corresponding to the eigenvalues with the top 90% of the principal component contribution rate are taken to form a matrix Π n . Correspondingly, the principal component feature PCA n after dimensionality reduction is calculated as follows:

[0092]

[0093] Feature selection sub-module based on mutual information

[0094] The extracted features are fused, then the multi-view features corresponding to the flame image can be expressed as: A n = [Gray_ave n , Gray_var n , Bright_ave n , Bright_var n , A_v n , G_v n , Color_T n , PCA n , and its feature dimension is denoted as Furthermore, the set of flame images corresponding to it can be expressed as

[0095] The redundant features in the multi-view features will affect the model operation efficiency and recognition accuracy. Here, the MI is used to measure the correlation between the extracted features and the combustion state, and this is used as the basis for feature selection.

[0096] First, calculate the MI value as follows: The MI value of the th sub-feature of the multi-view features of the nth flame image

[0097]

[0098] is calculated as follows: where, represents the joint probability density, and k rob (y n ) represents the marginal probability density.

[0099] Furthermore, compare with the threshold θ MI , θ MI = 2.5, screen out the features less than the threshold, and then obtain the multi-view reduced feature set with a higher correlation with the combustion state whose dimension is denoted as Then the flame image set corresponding reduced feature set is denoted as

[0100] ICF recognition module

[0101] Cascade layer sub-module

[0102] In this paper, each layer of CF used contains 2 RFs, 2 CRFs and 2 GBDTs, aiming to further improve the diversity by using different types of base learners.

[0103] 1. Random Forest Algorithm

[0104] RF is a Bagging ensemble model constructed with decision trees (DTs) as base learners proposed by Breiman et al. (1996), and it is an extended variant of the Bagging algorithm.

[0105] First, use Bootstrap to randomly sample the training set D = {(p i , y i ), i = 1 , 2,... b} ∈ P B×R where B represents the number of samples in the training set and R represents the dimension of the training set. The generation process of the RF training subset P can be described as:

[0106]

[0107] In the formula: represents the j-th training subset; f Gini (·) represents the random subspace function; f Bootstrap (·) represents the Bootstrap function; r = 1,..., R j , R j represents the number of features selected for the j-th training subset in the forest, usually R j << R.

[0108] Using the above function J times, the training set of RF can be obtained, and the process is as follows:

[0109]

[0110] Where: J represents the number of Bootstrap times and also represents the number of DTs in RF.

[0111] Using the above J training subsets Construct J DTs in the RF model. For the j-th training subset The construction process is as follows:

[0112] Traverse based on the Gini index criterion to find the optimal splitting feature number and the splitting point s.

[0113]

[0114] where k P represents the k-th P class in the dataset label y, k P ∈ 1, …, K P , represents the proportion of the k-th P class in the total number of labels. From this, the Gini index Gini(·) of the dataset can be calculated; θ Forest represents the sample quantity threshold contained in the leaf node, θ Forest = 8. and respectively represent the label values corresponding to the samples divided into the left and right nodes in the j-th training subset.

[0115] Based on the above criterion, first find the optimal variable number and splitting point value by traversing all input features, and divide the input feature space into left and right regions; then repeat the above process for each region until the number of samples contained in the leaf node is less than the threshold θ Forest , or the Gini index of the samples in the leaf node is 0; finally, divide the input feature space into Q regions. To construct a classification tree model, define the following function:

[0116]

[0117] where

[0118]

[0119] where represents the number of training samples contained in region G q ; represents the label vector corresponding to the sample features in region G q ; represents the final predicted result output by region G q ; I(·) is an indicator function. When When I(·) = 1, otherwise I(·) = 0.

[0120] After constructing J DTs, the RF model is finally obtained.

[0121]

[0122] Completely Random Forest Algorithm

[0123] The difference between CRF and RF is that the former randomly selects the value of a certain feature in the complete feature space as the splitting node, while the latter selects the splitting node through the Gini coefficient in the randomly generated feature subspace after Bootstrap. Correspondingly, the CRF model is represented by F CRF (·).

[0124] Gradient Boosting Decision Tree Algorithm

[0125] GBDT combines the gradient boosting framework with the DT model. By establishing a series of DT models in the gradient direction of residual reduction, the sample estimated value continuously approaches the true value, and finally the decision results of the DTs are superimposed to form the final GBDT model. The construction process of GBDT can be described as follows.

[0126] First, initialize the relevant parameters: the number of iterations The learning rate η = 0.6 and the input samples belonging to the k P class of the initial function estimate

[0127] Performing a Logistic transformation on the function estimate can obtain the probability belonging to the k P class:

[0128]

[0129] Then, define the following loss function

[0130]

[0131] where represents the estimated value of the input sample. When the sample belongs to the k P class, otherwise

[0132] Calculate the gradient direction of residual reduction using the loss function, as shown in the following formula:

[0133]

[0134] For a multi-classification task, the k P -th class samples can fit a regression tree according to the residuals and calculate the gain of the -th leaf node

[0135]

[0136] wherein, represents the P -th leaf node region of the k -th class, and represents the residual of the n

[0137] -th sample. Then, update the value of the -th function estimator by the following formula:

[0138]

[0139] wherein, is the number of iterations, is the indicator function, which is 1 when belongs to the -th node, and 0 otherwise.

[0140] After each class completes the construction process of the decision tree in this round, the next round of iteration process can be carried out until the -th iteration ends. Finally, denote the GBDT classification learning model composed of decision trees as F GBDT (·).(·).

[0141] The weighted average sub-module

[0142] CF adopts 2 F RF (·), 2 F CRF (·) and 2 F GBDT (·) for cascade learning in each layer, and constructs the CF model using the Stack idea. For the input The last layer of CF will output a 6K P -dimensional class distribution vector and use the average and maximum value criteria on it to obtain the combustion state recognition result

[0143]

[0144] Then for the feature set the combustion state recognition result

[0145] Evaluation index

[0146] To verify the effectiveness of the model, the following metrics are used to evaluate the model performance.

[0147] Precision:

[0148]

[0149] Recall:

[0150]

[0151] F1-Score:

[0152]

[0153] Among them, TP (True Positives) represents the number of samples that are actually positive and are classified as positive by the classifier; FP (False Positives) represents the number of samples that are actually negative but are classified as positive by the classifier; FN (False Negatives) represents the number of samples that are actually positive but are classified as negative by the classifier; TN (True Negatives) represents the number of samples that are actually negative and are classified as negative by the classifier.

[0154] For the multi-classification problem, for classification k P There are: and Then the overall evaluation metric can be calculated by the macro-average of each classification evaluation metric:

[0155]

[0156]

[0157]

[0158] The calculation method of the overall accuracy (Accuracy) of the model is as follows:

[0159]

[0160] Summary of formula symbols and their descriptions

[0161]

[0162]

[0163]

[0164]

[0165]

[0166]

[0167] The flame image data used in this experiment is sourced from a MSWI enterprise in Beijing. To monitor the combustion state in real time, an industrial camera is installed on the "rear arch" of the furnace in the MSWI site. After being transmitted through a coaxial cable, it is stored in a monitoring computer using a video capture card. The basis for dividing the combustion state of the flame image is as Figure 2 shown. Among them, Figure 2 (a) shows the in-furnace image taken during the furnace shutdown stage. The front arch, feed inlet, drying section, combustion section, and burnout section of the furnace are all clearly visible, and the combustion state can be divided accordingly; Figure 2 (b) shows the state of the burnout line moving forward, which is mainly distributed in the feeding section and the drying section. At this time, MSW is prone to coking and blocking the feed inlet; Figure 2 (c) shows the normal state of the burnout line, which is mainly distributed in the combustion section. At this time, the flame is bright and the combustion state is good; Figure 2 (d) shows the state of the burnout line moving backward, which is mainly distributed in the burnout section. At this time, it is necessary to adjust the grate speed and air volume in time to ensure the full combustion of MSW.

[0168] Based on the above basis for dividing the combustion state, the flame images collected on site are marked. In this article, the image size is 718×512, the number of samples is 571, and the number of samples corresponding to the three working conditions of the combustion line moving forward, normal, and backward are 184, 213, and 174 respectively.

[0169] Results of image preprocessing

[0170] Figure 3 The results of image preprocessing are shown as follows. The original MSWI flame image is as Figure 3 (a) shows. It contains a large amount of smoke, fly ash, and stripe noise, and the quality is extremely poor. Therefore, first, a defogging algorithm based on artificial multi-exposure image fusion is adopted. The defogging effect is as Figure 3 (b) shows. The amount of smoke is significantly reduced, and the color and shape of the flame are better restored. To further restore the color and shape of the flame, feature normalization is adopted, and the mean and variance of the three channels of the flame image are both set to 0.5 to separate the flame from the background image. The processing effect is as Figure 3 (c) shows. The flame color is bright and vivid, and it is clearly separated from the furnace background and the MSW that has not been fully burned. However, there are still stripe interferences in the image. It is converted to the frequency domain, and notch filtering is used to eliminate the frequency band corresponding to the stripe noise. The effect is as Figure 3 (d) shows. The shape of the flame is well restored. Finally, for the isolated noise points caused by fly ash, median filtering is used to eliminate them. The processing effect is as Figure 3As shown in (e), the fly ash amount is significantly reduced, and the flame morphology is protected.

[0171] Multi-view feature extraction and selection results

[0172] Figure 4 The process of extracting the brightness features of the flame image is shown in detail.

[0173] First, the preprocessed image 4(a) is converted from a color image to a grayscale Figure 4 (b). It can be seen that the area covered by the flame in Figure (b) is slightly smaller than that in Figure (a), indicating that there is a difference in the flame brightness from the computer's perspective and the human perspective. Figure 4 (c) is the V-channel image obtained by converting the flame image from the RGB space to the HSV space, which represents the brightness of the flame image from the human visual perspective. Figure 4 The area covered by the flame in (c) is Figure 4 basically the same as that in (a), which verifies the above conclusion.

[0174] The effective flame region θ selected during the flame feature extraction process th and the high-temperature region threshold ω th are established based on the V-channel image. Subsequently, the areas of the effective flame region and the high-temperature region are calculated and used as the flame features of the flame image. Figure 5 Shows Figure 4 the contour map of the brightness distribution in (c).

[0175] Next, the color moment feature extraction results under different working conditions are as Figure 6 shown. From Figure 6 it can be seen that the color moment features extracted under different combustion states have significantly different distribution characteristics and can better express the color feature differences of the three combustion states.

[0176] When using PCA for feature extraction, it is reduced to 0.1 times the original image. The contribution rate of the first principal component of the flame image under different combustion states reaches more than 95%, as Figure 7 shown, so only the first principal component feature needs to be extracted.

[0177] When performing feature selection based on MI. Most features have similar correlations with the combustion state, and only a small part has relatively low or high correlations.

[0178] ICF recognition module results

[0179] The average value of 20 runs is taken as the final result, as shown in Table 1.

[0180] Table 1 Model recognition accuracy

[0181]

[0182] Method Comparison

[0183] The proposed method is compared with the existing methods in the literature, and the statistical results are shown in Table 2.

[0184] Table 2 Comparison Results of Recognition Methods

[0185]

[0186] The PCA-LSSVM method constructs a combustion state recognition model based on LSSVM using the color moment features of the extracted flame images. Since the performance of LSSVM mainly depends on the selection of the kernel function, there is no relatively mature method for how to select a suitable kernel function according to the actual data model. In addition, the selection of mainstream kernel functions and their parameters are mostly based on experience and manually selected. These factors result in a low recognition performance of the model.

[0187] A combustion state recognition model is constructed based on the convolutional neural network (CNN). It requires a large number of samples during the training process, and the complex combustion state of MSWI leads to a still extremely high labor cost for labeled samples. Therefore, the CNN network has great limitations.

[0188] The method proposed in this paper combines multi-view features and ICF, and uses 3 basic learners in each layer, which improves the diversity and can fully exploit deep features. Therefore, the method in this paper has the highest recognition accuracy, and has a relatively high precision rate (96.22%), recall rate (96.11%) and F1 score (96.17%), which verifies that the established model has good stability.

Claims

1. MSWI Process Combustion State Recognition Method Based on Multi-Feature Fusion and Improved Cascade Forest, Characterized in that: The modeling strategy consists of three parts: an image preprocessing module, a multi-view feature extraction and selection module, and an improved cascade forest recognition module. The definitions of relevant variables and symbols are as follows: represents the flame image with the combustion state labeled during the MSWI process, where I n (u, v) represents the nth image, N is the number of image samples, and (u, v) represents the pixel coordinates in each image; represents the image after preprocessing; represents the reduced set of extracted flame features; K represents the number of underexposed images generated, γ is the gamma exposure operation coefficient, r represents the guiding filter radius, α, β, σ B is a positive adjustment factor, r notch is the filter radius, s a represents the side length of the median filter window, θ th represents the pixel threshold of the flame effective area, ω th represents the pixel threshold of the high-temperature area; θ MI represents the MI threshold, T n represents the number of decision trees in RF and CRF, η and represent the learning rate and the number of iterations of GBDT respectively, minsamples represents the minimum number of samples for each decision tree, represents the initial input sample belongs to the kth P class function estimate; Slide represents the sliding window size used when extracting color features; represents the combustion state recognition result output by the model; y is the manually marked combustion state; The specific steps are as follows: Step 1: Image Preprocessing Module: Eliminate the noise introduced by the on-site environment and transmission channel in the image, and separate the flame from the furnace background for subsequent image feature extraction; Step 2: Multi-View Feature Extraction and Selection Module: Extract multi-view features such as brightness, flame, color, and principal components of the combustion image, and select features based on MI; Step 3: Improved Cascade Forest Recognition Module: Use the selected multi-view features as the input of the ICF recognition module, and use RF, CRF, and GBDT as the base learners of the cascade forest to recognize the combustion state and obtain the recognition result; Specifically as follows 1) Image Preprocessing Module Combustion in the furnace of the MSWI process generates a large amount of fly ash and smoke. Affected by the high-temperature environment and the use of analog signals for image transmission by the image acquisition device, the obtained color RGB images inevitably introduce signal interference and other physical noises; therefore, it is necessary to preprocess the images to restore the on-site images as much as possible The preprocessing process of the nth flame image is as follows; a) Dehazing Algorithm Based on Artificial Multi-Exposure Image Fusion First, perform a gamma exposure operation on the original flame image I n (u, v) to generate K underexposed images K represents the number of underexposed images generated, K = 6; the k-th image {I n (u, v)} k The generation process is as follows: {I n (u,v)} k =I n (u,v) γ (1) where γ is the gamma exposure operation coefficient, γ = [1, 2, 4, 6, 8, 10]; for simplicity of representation, let I k denote {I n (u, v)} k ; Decompose I through a guided filter k into a global component B k and a local component D k as follows: B k = G r ∈ (L k , L k ) (2) D k = I k - B k , k = 1, ..., K (3) Among them, G r represents a guided filtering operation, and r = 12 is the filtering radius used to control the blurriness; L k is the grayscale image of the image, which serves as both the input image and the guidance image for guided filtering, R Channel and G Channel and B Channel represent the three-channel components of the image respectively; D k Exposure weight of each pixel point in is calculated by the following formula: Among them, is obtained by performing a convolution operation on L with an average filter of size 7×7 k and is an adjustment factor, For each global component, an exposure weight is obtained using it as an initial exposure feature Among them, represents the average grayscale image, m and n respectively represent the length and width of the image, and α, σ B are adjustment factors, and σ B = 0.5, α = 0.2; Perform weighted averaging on the global component and local component of the input image to obtain the dehazed image Z: Where β is an adjustment factor, β = 1.1; Finally, the original flame image I is obtained. n The dehazed image Z of (u, v) n (u, v); b) Feature Normalization Feature normalization maps the image pixels from 0-255 to 0-1, thereby reducing the computational complexity to improve the model running efficiency, and at the same time separating the flame from the furnace background; the zero-mean normalization adopted is as follows: Among them, is the mean value of the image, such that is the standard deviation of the image, such that W n (u, v) is the normalized image; c) Notch Filtering The image during the transmission process will have frequency-domain interference due to the influence of the high-temperature environment, resulting in stripe noise in the collected flame image; therefore, it is necessary to eliminate the frequency band where the stripe noise is located through frequency-domain filtering; the notch filter belongs to the band-stop filter, and its stop band has a zero elimination signal frequency point and a value of 1 at other frequencies; the image obtained after notch filtering is V n (u, v) = notch{W n (u, v), r notch}} (9) where, notch represents notch filtering operation, and r notch is the filtering radius, and r notch = min(m,n) / 5 d) Median Filtering Isolated noise points in the image caused by fly ash need to be eliminated by median filtering, using a 5×5 window for sliding operation; assign the median value of the window pixels to the template center, and then obtain the filtered image, denoted as Where median represents the median filtering operation; Denote as the final set of flame images after preprocessing; 2) Multi-View Feature Extraction and Selection Module a) Multi-View Feature Extraction Sub-Module During the MSWI process, to ensure that the materials are burned as fully as possible, the grate is always in periodic motion, and the flame images show regular change characteristics. Considering the rich information contained in the flame images, multi-view features are considered for characterization; the nth flame image The feature extraction process is as follows; (1) Brightness Feature For the flame image of the MSWI process, its brightness feature can be described from the following multiple perspectives; (1)Average gray value: The original image is converted into a grayscale image by the following formula: Conversion to grayscale image: Among them, and respectively represent the R Channel , G Channel and B Channel color components of the three channels at the pixel point (u, v) of the nth image; The calculation formula for the average gray value of the image is as follows: Among them, represents the th pixel point, where m and n respectively represent the number of pixel points in the length and width directions of the image; (2) Variance of Gray Value: Calculated by the following formula: (3) Average Brightness Value: The above two features mainly describe the brightness feature from the computer perspective; different from the characteristics of the RGB space, the HSV space expresses the image closer to the human perception experience of color, and it represents color as a linear combination of three components: color, vividness, and lightness; therefore, after converting the flame image from the RGB space to the HSV space, extracting the brightness feature of the V channel can better reflect the brightness expression of the flame image from the human visual angle; Here, is converted to the HSV image space and represented as The V channel is selected The mean value of its pixel values is calculated as the average brightness value of the image: (4) Variance of Brightness Value: The calculation formula is as follows: (2) Flame Feature The flame characteristics include the area characteristics of the effective flame region and the area characteristics of the high-temperature region, and their calculations are based on the V channel ; (1) Flame effective area: defined as the total number of pixel points in the image with a brightness value greater than the specified threshold θ th , where θ th = 0.226, as follows: Where Θ(·) is the unit step function; (2) Area of the high-temperature flame zone: defined as the total number of pixel points in the image with a brightness value greater than the specified threshold ω th , where ω th = 0.941, as follows: (3) Color Feature Use the first-order, second-order, and third-order moments to express the color information of the flame image; The nth preprocessed flame image Its first moment is calculated as follows: Second moment The calculation formulas for each item are as follows: Among them, represents the first moment feature of each color channel of the nth image; Third moment The calculation formulas for each item are as follows: Use a sliding window with a scale of 1 / 5 of the original image for color moment feature extraction, and record the extracted color feature as: (4) Principal Component Feature The purpose of performing image feature extraction tasks based on principal component analysis is to find the basis image; To prevent memory shortage in calculations caused by the excessively large size of the original image, the original image is reduced to 0.1 times the original size, and its size is denoted as f×t×3, denoted as It is reconstructed into a matrix χ n (ft×3), and then its covariance matrix C n is calculated as follows: Among them, is the mean value of C n ; Calculate to obtain C n The eigenvalue λ n And the eigenvector π n ; Arrange the eigenvalues in descending order, and take the eigenvectors corresponding to the eigenvalues with the top 90% of the principal component contribution rates to form the matrix Π n ; Correspondingly, the principal component feature PCA n After dimensionality reduction is calculated as follows: b) Feature Selection Sub-Module Based on Mutual Information Fuse the extracted features, then the flame image The corresponding multi-view features can be expressed as: A n = [Gray_ave n , Gray_var n , Bright_ave n , Bright_var n , A_v n , G_v n , Color_T n , PCA n , and its feature dimension is denoted as Furthermore,[[]] The set of corresponding flame images can be expressed as The redundant features contained in the multi-view features will affect the model running efficiency and recognition accuracy; here, MI is used to measure the correlation between the extracted features and the combustion state, and this is used as the basis for feature selection; First, calculate the MI value as follows: For the sub-feature of the multi-view features of the nth flame image, the MI value is calculated as follows: Among them, represents the joint probability density, and k rob (y n ) represents the marginal probability density; Compare with the threshold θ MI , where θ MI = 2.

5. Features less than the threshold are filtered out, and then a multi-view reduced feature set with a high correlation with the combustion state is obtained Its dimension is denoted as Then the flame image set The corresponding reduced feature set is denoted as 3) ICF Recognition Module Directly input the reduced multi-view features into the CF to construct a combustion state recognition model, and improve the CF; a) Cascade Layer Sub-Module Each layer of the CF adopted contains 2 RFs, 2 CRFs, and 2 GBDTs; (1) Random Forest Algorithm RF is a Bagging ensemble model constructed with decision tree DT as the base learner and is an extended variant of the Bagging algorithm; First, use Bootstrap to randomly sample the training set D = {(p i , y i ), i = 1, 2, … b} ∈ P B×R where B represents the number of samples in the training set and R represents the dimension of the training set; the generation process of the RF training subset P can be described as follows: In the formula: represents the j-th training subset; f Gini (·) represents the random subspace function; f Bootstrap (·) represents the Bootstrap function; r = 1, …, R j , R j represents the number of features selected for the j-th training subset in the forest, usually R j << R; Use the above function J times to obtain the training set of RF. The process is as follows: Where: J represents the number of Bootstrap times and also represents the number of DTs in RF; The above-mentioned J training subsets Construct J DTs in the RF model; the j-th training subset The construction process is as follows: Traverse and find the best split feature number based on the Gini index criterion and the split point s Among them, k P represents the k P -th class in the dataset label y, where k P ∈ 1, …, K P , represents the proportion of the k P -th class in the total number of labels. From this, the Gini index Gini(·) of the dataset can be calculated; θ Forest represents the sample quantity threshold contained in the leaf node, and θ Forest = 8; and respectively represent the label values corresponding to the samples divided into the left and right nodes in the j-th training subset; Based on the above criteria, first, the optimal variable number and split point value are found by traversing all input features, and the input feature space is divided into two regions, the left and the right; then the above process is repeated for each region until the number of samples contained in a leaf node is less than the threshold θ Forest , or the Gini index of the samples in the leaf node is 0; finally, the input feature space is divided into Q regions; to construct a classification tree model, the following function is defined: Among them, Among them, represents the number of training samples included in region G q ; represents the label vector corresponding to the sample features in region G q ; represents the final predicted result output in region G; I(·) is an indicator function, when q holds, I(·) = 1, otherwise I(·) = 0; ​ After constructing J DTs, the RF model is finally obtained. (2) Completely Random Forest Algorithm The difference between CRF and RF is that the former randomly selects the value of a certain feature in the complete feature space as the splitting node, while the latter selects the splitting node through the Gini coefficient in the randomly generated feature subspace after Bootstrap. Correspondingly, the CRF model is represented by F CRF (·). (3) Gradient Boosting Decision Tree Algorithm GBDT combines the gradient boosting framework with the DT model. By establishing a series of DT models in the gradient direction of residual reduction, the sample estimated value continuously approaches the true value. Finally, the decision results of DTs are superimposed to form the final GBDT model. The construction process of GBDT can be described as follows; First, initialize the relevant parameters: the number of iterations The learning rate η = 0.6 and the input samples belonging to the k P initial estimated value of the function of the class For the function estimated value Performing a Logistic transformation gives The probability of belonging to the k P -th class: Then, define the following loss function. Among them, represents the estimated value of the input sample. When the sample belongs to the k P th class, Otherwise Use the loss function to calculate the gradient direction of residual reduction, as shown in the following formula: For a multi-classification task, the k P -th class samples can fit a regression tree according to the residuals and calculate the gain of the -th leaf node Among them, represents the k P th leaf node region of the category, represents the residual of the nth sample; Next, update the value of the -th function estimate using the following formula: Among them, is the number of iterations, is an indicator function. When belongs to the th node, it is 1, otherwise it is 0; After the construction process of the decision tree for each category is completed, the next round of the iterative process can be carried out until the end of the th iteration; finally, the GBDT classifier learning model composed of GBDT decision trees is denoted as F GBDT (·); b) Weighted Average Sub-module Each layer of CF uses two Fs RF (·), two Fs CRF (·) and two Fs GBDT (·) perform cascade learning and build a CF model using the Stack idea; for the input The last layer of CF will output a 6K P dimensional class distribution vector Apply the average and maximum value criteria to it to obtain the combustion state recognition result Then for the feature set The combustion state recognition result can finally be obtained

Images