Coke powder moisture online identification method based on multi-class heterogeneous image features

Abstract

The invention discloses a coke powder moisture online identification method based on multi-class heterogeneous image features, and belongs to the field of image identification. In order to solve the problem that a traditional on-line moisture detection method cannot meet the actual sintering production requirement, the coke powder moisture on-line identification method based on the multi-class heterogeneous image features is provided, the sintered coke powder serves as an object, RGB images of the coke powder under different moisture are collected, the images are subjected to preprocessing such as brightness correction and wavelet transformation, and the moisture content of the coke powder is obtained. According to the method, multiple types of heterogeneous coke powder images are formed, then multiple types of image features based on brightness and texture in the multiple types of heterogeneous coke powder images are extracted, an optimal feature subset is searched in combination with an SHAP feature analysis result, the sintering coke powder moisture prediction precision is further improved, and the newly proposed algorithm not only can eliminate useless or redundant features, but also can improve the prediction accuracy of the sintering coke powder moisture. Dependence on computing resources is reduced, the computing speed is greatly improved, the generalization ability and prediction precision of the model can be remarkably improved, and the application effect of the model is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical field of image recognition, and more specifically, to an on-line recognition method for the moisture content of coke fines based on multi-class heterogeneous image features. Background Art

[0002] The sintering process is one of the key links in the metallurgical industry and plays an important role in improving the utilization rate of ore and the smelting efficiency. It mixes, granulates and sinters materials such as iron ore fines, fuel and flux into blocks, and finally produces sinter suitable for blast furnace smelting. As a key raw material for providing heat in the sintering process, the deviation of fuel will not only affect the quality of sinter, but also have an adverse impact on production efficiency and energy consumption. When the raw material structure is determined, too high or too low moisture content of the fuel will cause different degrees of interference to the stability and effect of the sintering process. If the moisture content of coke fines is higher than the detected value, the effective calorific value of the actual fuel during batching will be lower than the calculated value, and it cannot provide enough heat for the sintering process. Due to the high moisture content, a large amount of heat will be used to evaporate the moisture rather than promoting the sintering reaction, resulting in the temperature of the mixed material layer being difficult to reach the ideal level. In this case, the sintering reaction cannot proceed smoothly, resulting in some ores not being fully sintered, affecting the physical strength and metallurgical properties of the finished product, and reducing the quality of sinter. In addition, the uneven distribution of the temperature field will also affect the heat transfer and mass transfer efficiency, increase the instability of the process, force operators to adjust parameters, and thus increase energy consumption and production costs. On the contrary, if the moisture content of coke fines is lower than the detected value, the actual fuel will be higher than the batching value, resulting in waste of heat. At the same time, due to the too low moisture content, the adhesiveness of the mixed material is significantly reduced, the stability of the material layer becomes poor, and it is easy to cause phenomena such as dispersion, pulverization and collapse of the sintering material layer, further affecting the heat transfer and mass transfer effects of the sintering process. The looseness of the material layer structure will reduce the combustion efficiency, make the combustion process of coke fines incomplete, and the heat energy release uneven, resulting in a decrease in the sintering reaction efficiency. In addition, the increase in dust will put pressure on environmental protection equipment and the working environment, and increase the energy consumption burden of the subsequent dust removal system, indirectly increasing the overall production cost. Therefore, accurately detecting the moisture content of sintering fuel is of great significance for optimizing the smelting process, improving production efficiency and ensuring product quality.

[0003] At present, although traditional moisture detection methods such as the drying method and the chemical method are highly accurate, they are complex to operate, time-consuming, and inefficient, and cannot meet the requirements of modern metallurgical enterprises for real-time monitoring and rapid response. The online moisture detection methods mainly include the neutron method, the microwave method, the infrared method, the conductivity method, and the resistance method, etc. However, traditional online moisture detection methods still have significant limitations in the application of coke powder: the neutron method is difficult to comprehensively reflect the moisture distribution due to insufficient actual penetration depth, sensitivity to environmental interference (such as air humidity and composition interference), and radiation safety issues; the microwave method is affected by the dependence on coke powder composition, with a complex inversion model, and the calibration is difficult due to limited penetration; the infrared method can only detect the surface moisture and is easily interfered by the difference in particle reflectivity and temperature fluctuations; the conductivity method has limited accuracy and requires frequent maintenance due to the interference of ash / salt conductivity and poor contact in contact measurement; the resistance method has poor stability and adaptability due to the non-linear superposition of particle contact resistance and the change in dynamic bulk density. Therefore, traditional online detection methods generally have problems such as local sampling deviation, weak anti-interference ability, or safety risks.

[0004] With the development of machine vision technology, people have begun to study image-based moisture recognition. The online moisture detection method based on images has many advantages such as non-contact, non-radiative, low cost, and maintenance-free, and has good application prospects. However, this technology is still in the research and development stage, and currently still faces the situation that the dynamic adaptability of physical supplementary lighting is poor (interference from ambient light fluctuations and surface reflection), and it is overly dependent on color, resulting in less than ideal recognition accuracy, which limits the application and promotion. Summary of the Invention

[0005] 1. Technical problems to be solved by the invention

[0006] Aiming at the situation that traditional online moisture detection methods in the prior art cannot meet the actual production requirements of sintering, the present invention intends to provide an online moisture recognition method for coke powder based on multi-class heterogeneous image features, which can effectively improve the prediction accuracy of sintered coke powder moisture, reduce the dependence on computing resources, and improve the computing speed, and has good application effects.

[0007] 2. Technical solution

[0008] To achieve the above object, the technical solution provided by the present invention is as follows:

[0009] To solve the problem that traditional online moisture detection methods cannot meet the actual production requirements of sintering, the present invention proposes an online recognition method for the moisture of coke powder based on multi-class heterogeneous image features. Taking sintered coke powder as the object, RGB images of coke powder under different moisture contents are collected, and preprocessing such as brightness correction and wavelet transform is performed on the images to form multi-class heterogeneous coke powder images. Subsequently, multi-class image features based on brightness and texture are respectively extracted from the multi-class heterogeneous coke powder images, and the optimal feature subset is found in combination with the SHAP feature analysis results to further improve the prediction accuracy of the moisture of sintered coke powder. The newly proposed algorithm can not only eliminate those useless or redundant features, reduce the dependence on computing resources, and greatly improve the computing speed, but also significantly enhance the generalization ability and prediction accuracy of the model, and enhance the application effect of the model.

[0010] An online recognition method for the moisture of coke powder based on multi-class heterogeneous image features of the present invention includes:

[0011] Step 1, collection of coke powder image data: Cooperate with coke powder samples with different moisture contents, and collect surface image data of coke powder with different moisture contents;

[0012] Specifically, to collect clear images of water-containing coke powder, a CCD industrial camera is used in the image acquisition hardware system, and the height is fixed by a bracket. Both the camera and the coke powder sample are set in a light-shielding hood, and supplementary lights are symmetrically arranged on both sides inside the light-shielding hood to ensure consistent light intensity and avoid the influence of different natural lights on the image brightness. The sample stage is located directly below the camera lens, so that the camera can accurately and clearly capture the coke powder image. And the camera is connected to the computer through an external Ethernet, and the parameters of the camera are adjusted and the shooting is controlled through the computer.

[0013] During image acquisition, coke powder samples with different water contents are respectively configured. For example, eight kinds of coke powder samples are configured. The specific process is as follows: First, the dry coke powder is mixed evenly and different moisture coke powder samples are configured. The configured coke powder is sealed and left standing for 2-3 days to allow the moisture to penetrate evenly. The coke powder is crushed and mixed again and placed in a container, and pressed to ensure the flatness of the coke powder surface, and then the surface image of the coke powder is collected.

[0014] Step 2, construction of an online recognition model for the moisture of coke powder. Through three-stage optimization of brightness self-corrected multi-class heterogeneous image construction - brightness-texture collaborative feature extraction - feature dynamic recursive optimization, high-precision modeling of the moisture of coke powder is realized, specifically including:

[0015] S21, perform brightness correction with light robustness on the image data to unify the brightness of coke powder images under different lighting conditions to a stable interval;

[0016] Specifically, during the analysis of the characteristics of the original coke powder image, the brightness of ore powder images with different water contents is not the same. However, a low brightness will cause the texture features of the coke powder image to become unclear, thereby affecting the extraction of texture features. Aiming at the defects of high maintenance cost, poor dynamic adaptability, and insufficient uniformity of traditional physical supplementary lighting under complex working conditions, the present invention proposes a pixel-level linear scaling algorithm to replace the traditional method of using physical supplementary lighting to control the image brightness. Based on the target brightness, the pixel values of the entire image are dynamically adjusted to unify the brightness of the coke powder images under different lighting conditions into a stable range, eliminating the interference of ambient light fluctuations on texture features.

[0017] The lighting correction method of the present invention first calculates the average brightness of the original image. By summing up the brightness values of all pixels and taking the average, the overall brightness level of the image is obtained. Then, a target brightness value is set as a standard, and by calculating the ratio of the target brightness to the average brightness of the original image, a brightness adjustment coefficient is obtained. Then, by multiplying each pixel value by the brightness adjustment coefficient, the brightness value of each pixel in the image is adjusted to obtain the corresponding final pixel value of the image. The specific formula is as follows:

[0018]

[0019] Among them, I(x, y) is the pixel value at the position (x, y) of the original image, and I′(x, y) represents the pixel value at the position (x, y) of the pixel point in the corrected image. represents the target-set average brightness value. represents the average brightness value of the original image, N = W × H is the total number of pixels of the image, and W and H represent the width and height of the image respectively.

[0020] The present invention adopts a unified standard to ensure the brightness consistency of different images during the processing, thereby avoiding the inconsistency of the feature extraction results caused by the image brightness difference. This brightness correction method is based on pixel-level operations on the image. By calculating the average brightness value of the image and performing linear scaling, it is both intuitive and computationally efficient, avoiding complex mathematical models and calculations. While ensuring the computational efficiency, it can effectively process a large number of images, especially suitable for the image correction requirements under the sintering process. This method can quickly unify the image brightness, reduce the interference caused by light changes, and ensure the consistency and reliability of image processing under high-intensity working conditions.

[0021] S22. Extract multi-class heterogeneous features, break the limitation of only statistically analyzing RGB color features in the traditional way, and construct a brightness-texture collaborative characterization system;

[0022] Specifically, to further extract the texture information of the coke powder image, the present invention utilizes the fact that the spatial specific frequency characteristics in wavelet transform can correspond to the direction of the texture. The image is decomposed into components with different frequencies and directions through two-dimensional wavelet transform, effectively capturing the local texture information in the coke powder image. The two-dimensional wavelet transform basis function can be expressed as follows:

[0023]

[0024] where x and y are the horizontal and vertical pixel coordinates of the coke powder image; is a one-dimensional scaling function that can restore the original signal and is used to extract the low-frequency information of the image; ψ(·) is a one-dimensional wavelet basis function that can represent the difference between the wavelet basis function and the original signal and is used to extract the high-frequency information of the image; is a two-dimensional scaling function; ψ 1 (·, ·), ψ 2 (·, ·), ψ 3 (·, ·) are separable two-dimensional wavelet basis functions, respectively characterizing the numerical changes in the horizontal, vertical, and diagonal directions.

[0025] In view of the advantages of Haar wavelet in image decomposition and reconstruction, such as fast speed, high compression ratio of the reconstructed image, and easy various analysis and processing of the image, therefore, the present invention selects Haar wavelet as the basis function when performing two-dimensional discrete wavelet decomposition on the coke powder image. The specific process is as follows. First, the low-frequency and high-frequency filters of the Haar wavelet basis function are convolved with each row of the image to extract the low-frequency and high-frequency information in the horizontal direction. Then, the same processing steps are used for each column to extract the low-frequency and high-frequency information in the vertical direction. In this way, the image will ultimately be decomposed into four different parts: the low-frequency information image, the high-frequency information image in the vertical direction, the high-frequency information image in the horizontal direction, and the high-frequency information image in the diagonal direction.

[0026] After performing brightness transformation and wavelet transformation on the original RGB image, multiple types of heterogeneous coke powder images are generated. Next, the present invention performs feature extraction on these multiple types of heterogeneous coke powder images, aiming to extract the feature information related to moisture for further analysis and processing. The image feature extraction process is as Figure 2 shown.

[0027] The feature extraction can be specifically divided into two categories. One category is to extract a total of 9 parameters such as mean, peak value, and wavelet mean from the original image and the corresponding wavelet image to describe the brightness characteristics of the coke powder image. The other category is to extract a total of 32 parameters such as standard deviation, entropy, and angular second moment from the image after brightness correction and its corresponding wavelet image to describe the texture characteristics of the coke powder image. Some feature extraction formulas are as follows:

[0028]

[0029] Among them, I(x i , y i ) is the pixel value at the position of the original image (x i , y i ), cA(x i , y i ) is the wavelet coefficient at the position of the original image (x i , y i ), N is the total number of pixel points; μ is the mean value of the image, and σ is the standard deviation of the image.

[0030] S23. Joint feature optimization of SHAP-RFE, screening out the optimal feature subset, and constructing the corresponding moisture identification model.

[0031] In view of the problem that traditional feature screening statistical methods, such as Pearson / chi-square test, ignore feature interaction and model adaptability, the present invention proposes an interpretable dynamic screening strategy. Through the foregoing method, 42 features are extracted from the coke powder image. Next, the relationship between these features and moisture will be fitted through feature screening and model construction. The feature screening process and its model construction process are as Figure 3 shown. It is mainly divided into two parts: feature screening and model construction.

[0032] Feature Screening: Traditional feature analysis methods, such as Person correlation analysis, usually only consider the relationship between features and target values, relying on simple statistical measures such as correlation coefficients and information gain. These methods often ignore the mechanism of action of features in the model, that is, their specific role in the actual model prediction process. This screening method based solely on correlation or simple measures may miss those features that have complex effects in the model but have a weak linear relationship with the target value. Therefore, traditional feature screening methods cannot fully reveal the true contribution of each feature to the model prediction result, especially when dealing with multi-dimensional non-linear models, their effects may be more limited. To solve this problem, this paper introduces the SHAP (Shapley Additive Explanations) algorithm into the image feature screening process. The SHAP algorithm can comprehensively evaluate the importance of each feature by calculating the marginal contribution of each feature to the model prediction result. Different from traditional methods, as a post-hoc interpretability analysis method, the SHAP algorithm does not depend on the distribution form of data. It not only considers the relationship between features and target variables, but also deeply analyzes the mechanism of action of each feature in the model. It evaluates the different combinations of features in a "game" way, calculates the marginal contribution of each feature in all possible combinations, thereby quantifying the specific impact of the feature on the model prediction and measuring the importance of the feature. The specific operation of the feature screening process is as follows. First, the original data set is input into the model. The SHAP algorithm will assign SHAP values to each feature of each data point by calculation, that is, the contribution degree of the corresponding feature to the model prediction. The SHAP value φ j is defined as:

[0033]

[0034] where |·| represents the number of elements in the set; N represents the original feature set; S represents any feature subset in N; Nleft{j} represents the subset of all elements in the sequence before feature j; f(S) represents the output of the machine learning model for the feature subset S; f(S∪{j}) - f(S) represents the cumulative contribution value of feature j; then the features are sorted according to the SHAP values from large to small to obtain the feature importance ranking set; finally, use RFE to sequentially select the features with higher importance in the feature importance ranking set, delete the least important features, and use the XGBoost model to verify the accuracy rate corresponding to the screened feature subset. Through an iterative method, the optimal feature subset is screened out.

[0035] Model Construction: Use the optimal feature subset screened above to construct a corresponding coke breeze moisture prediction model using the XGBoost model.

[0036] 3. Beneficial Effects

[0037] Adopting the technical solution provided by the present invention, compared with the prior art, it has the following beneficial effects:

[0038] (1) An online identification method for the moisture content of coke powder based on multi-class heterogeneous image features of the present invention solves the three major bottlenecks of traditional moisture detection in terms of environmental adaptability, feature completeness, and model generalization by means of algorithm-level calibration to replace hardware supplementary lighting, extraction of multi-class heterogeneous image features to break through color limitations, and interpretable dynamic screening to optimize the model, providing a low-cost, high-precision, and strong-robustness online detection solution for the metallurgical industry, which can be extended to the moisture detection scenarios of bulk materials such as iron ore powder and coal powder.

[0039] (2) For the defects of high maintenance cost, poor dynamic adaptability, uneven illumination, etc. of traditional physical supplementary lighting, the online identification method of the present invention proposes a pixel-level linear brightness correction algorithm. By dynamically scaling the image brightness to the target mean value, the interference of ambient light fluctuations is eliminated without physical supplementary lighting equipment, greatly reducing the maintenance cost and being more conducive to the subsequent image feature analysis process.

[0040] (3) In the online identification method of the present invention, since the traditional method only relies on the feature extraction of RGB images, it often causes the problem of incomplete feature extraction. The present invention performs wavelet transform on the image, constructs a multi-class heterogeneous coke powder image set, extracts the brightness and texture features of the multi-class heterogeneous coke powder images, constructs a brightness-texture collaborative characterization system. In the brightness features, 9-dimensional parameters such as mean and peak value are extracted based on the original and wavelet-transformed images to capture the overall brightness change of the image caused by moisture; in the texture features, 33-dimensional parameters such as standard deviation, entropy, and angular second moment smoothness are extracted based on the brightness-corrected and wavelet-transformed images to capture the global and local texture changes of the image caused by moisture, thereby realizing the in-depth excavation of the differential features of the coke powder image. The 42-dimensional multi-scale features cover the macro and micro responses of moisture, and the MAE of the model is reduced from 0.35 to 0.116, with the accuracy increased by 67%.

[0041] (4) For the problems of ignoring feature interaction and model incompatibility in traditional statistical methods, the online identification method of the present invention designs a SHAP-RFE joint optimization strategy, analyzes the action mechanism of features inside the model, quantifies the global importance of features. For example, the contribution degree of the image brightness mean value is the highest, and based on the SHAP contribution degree ranking, the feature recursive elimination algorithm is used to screen out the optimal feature subset, enhancing the interpretability of the model. The feature dimension is reduced from 42 dimensions to 9 dimensions, the feature dimension is significantly reduced, the prediction accuracy of the model is improved, and the complexity of the model is greatly reduced at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is the flowchart for constructing the online identification model of the moisture content of coke powder of the present invention;

[0043] Figure 2 It is a schematic diagram of the process for extracting coke breeze images in the present invention;

[0044] Figure 3 It is a schematic diagram of the process for feature screening and model construction in the present invention;

[0045] Figure 4 It is a comparison schematic diagram between the original image and the brightness-corrected image in the embodiment, where the upper three figures (a), (b), and (c) are the original images, and the lower three figures (d), (e), and (f) are the corresponding corrected images;

[0046] Figure 5 It is the two-dimensional wavelet transform image in the embodiment, where the left side is the original image and the four small figures on the right side are the wavelet transform images;

[0047] Figure 6 It is the SHAP feature ranking diagram in the embodiment;

[0048] Figure 7 It is the RFE process model index change diagram in the embodiment. Detailed implementation manners

[0049] To further understand the content of the present invention, the present invention will be described in detail with reference to the accompanying drawings.

[0050] In the description of the present invention, it should be noted that the orientation or positional relationship indicated by the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. is based on the orientation or positional relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the present invention. In addition, the terms "first", "second", and "third" are only used for descriptive purposes and cannot be construed as indicating or implying relative importance.

[0051] The present invention will be further described below in conjunction with the embodiments.

[0052] Embodiment

[0053] This embodiment takes the industrial scenario of coke breeze moisture prediction as an example.

[0054] First, in the experiment, eight different moisture ore powder samples are configured within the range of 3.4% - 9.0% with a moisture content gradient of 0.8%, and the coke breeze moisture image data is collected through the coke breeze moisture image acquisition system. The brightness correction operation is performed on the collected images. As Figure 4 shown, through the illumination correction strategy, the texture discrimination between coke breeze images with different moisture contents is obvious, making the subsequent texture extraction more accurate. As Figure 5 shown. The wavelet transform operation is performed on the coke breeze images to extract wavelet features.

[0055] Feature extraction is performed on the obtained multi-class heterogeneous images. The extracted features are divided into two categories. One category is to extract a total of 9 parameters such as mean, peak value, and wavelet mean from the original image and its corresponding wavelet image to describe the brightness features of the coke powder image. The other category is to extract a total of 33 parameters such as standard deviation, entropy, and angular second moment from the image after brightness correction and its corresponding wavelet image to describe the texture features of the coke powder image.

[0056] The extracted features are fitted by XGBoost for SHAP analysis to obtain the feature importance ranking. The feature importance ranking is as Figure 6 shown.

[0057] According to the feature sequence ranked by SHAP importance, using the RFE feature recursive elimination algorithm, variables with relatively low importance in the variable combination are gradually eliminated, and the feature sequence after elimination is used as the input of the XGBoost model to construct a coke powder moisture prediction model. The changes in the model indicators are as Figure 7 shown. Through the feature recursive elimination process, the indicators of the model on the test set generally show a trend of first decreasing and then increasing. In particular, when the number of features is 9, the XGBoost model reaches the optimal effect. On the test set, MAE is 0.116, MSE is 0.048, RMSE is 0.236, and MAPE is 2.12%. Comparing each indicator, it is better than the XGBoost model without feature screening, indicating that the model prediction ability is further improved after feature screening, and the model fits well the relationship between features and moisture.

[0058] The above schematically describes the present invention and its implementation manners. This description is not restrictive and is only one of the implementation manners of the present invention. In fact, it is not limited thereto. Therefore, if those of ordinary skill in the art are inspired by it and design similar structural manners and embodiments to this technical solution without creative efforts without departing from the purpose of the present invention, they shall fall within the protection scope of the present invention.

Claims

1. A method for online identification of moisture in coke powder based on multi-class heterogeneous image features, characterized in that: include: Step 1, coke powder image data collection: prepare coke powder samples with different moisture contents and collect coke powder surface image data with different moisture contents; Step 2: Building an online identification model for coke powder moisture, including: S21, performing illumination robustness brightness correction on the image data to unify the brightness of the coke image under different illumination conditions to a stable range; S22, extract multi-class heterogeneous features and build a brightness-texture collaborative representation system; S23 and SHAP-RFE are combined with feature optimization to screen out the optimal feature subset and build the corresponding moisture recognition model.

2. The method for online identification of moisture in coke powder based on multi-class heterogeneous image features according to claim 1 is characterized in that: Specifically, step S21 uses a pixel-level linear scaling algorithm to dynamically adjust the pixel value of the entire image based on the target brightness, unify the brightness of the coke image under different lighting conditions to a stable range, and eliminate the interference of ambient light fluctuations on texture features. The application formula is: Where I(x, y) is the pixel value at the position (x, y) of the original image, and I′(x, y) is the pixel value at the position (x, y) of the corrected image. Indicates the average brightness value of the target setting, Represents the average brightness value of the original image, N=W×H is the total number of pixels in the image, and W and H represent the width and height of the image respectively.

3. The method for online identification of moisture in coke powder based on multi-class heterogeneous image features according to claim 1 is characterized in that: In step S22, convolution is performed on each row of the image using low-frequency and high-frequency filters of the Haar wavelet basis function to extract low-frequency and high-frequency information in the horizontal direction; then, the same processing steps are used for each column to extract low-frequency and high-frequency information in the vertical direction; the image is finally decomposed into four different parts: a low-frequency information map, a high-frequency information map in the vertical direction, a high-frequency information map in the horizontal direction, and a high-frequency information map in the diagonal direction; After brightness transformation and wavelet transform were performed on the original RGB image, multiple types of heterogeneous coke powder images were generated, and features were extracted from these multiple types of heterogeneous coke powder images.

4. The method for online identification of moisture in coke powder based on multi-class heterogeneous image features according to claim 3 is characterized in that: In step S22, the extracted features are divided into two categories, one is to extract parameters from the original image and the corresponding wavelet image to describe the brightness features of the coke image, and the other is to extract parameters from the brightness-corrected image and the corresponding wavelet image to describe the texture features of the coke image.

5. The method for online identification of moisture in coke powder based on multi-class heterogeneous image features according to claim 4 is characterized in that: In step S23, the 42 features extracted from the coke powder image are screened and a model is constructed to fit the relationship between these features and moisture; The SHAP algorithm is introduced in feature screening to comprehensively evaluate the importance of each feature by calculating the marginal contribution of each feature to the model prediction results. The corresponding coke powder moisture prediction model is constructed using the XGBoost model through the screening of the optimal feature subset.

6. The method for online identification of moisture in coke powder based on multi-class heterogeneous image features according to claim 3 is characterized by: In step S22, the image is decomposed into components of different frequencies and directions by two-dimensional wavelet transform, so as to effectively capture the local texture information in the coke powder image. The basic function of the two-dimensional wavelet transform is: Where x and y are the horizontal and vertical pixel coordinates of the coke image; is a one-dimensional scaling function that can restore the original signal and is used to extract low-frequency information of the image; ψ(·) is a one-dimensional wavelet basis function that can represent the difference between the wavelet basis function and the original signal and is used to extract high-frequency information of the image; is a two-dimensional scaling function; ψ 1 (·,·),ψ 2 (·,·),ψ 3 (·,·) are separable two-dimensional wavelet basis functions, which represent the numerical changes in the horizontal, vertical and diagonal directions respectively.

7. The method for online identification of moisture in coke powder based on multi-class heterogeneous image features according to claim 5 is characterized in that: In step S23, the feature screening process is specifically operated as follows: first, the original data set is input into the model, and the SHAP algorithm assigns a SHAP value to each feature of each data point by calculation, that is, the SHAP value φ corresponding to each feature's contribution to the model prediction is j Defined as: Where |·| represents the number of elements in the set; N represents the original feature set; S represents any feature subset in N; Nleft{j} represents the subset of all elements in the sequence before feature j; f(S) represents the output of the machine learning model of feature subset S; f(S∪{j})-f(S) represents the cumulative contribution value of feature,; then the features are sorted in descending order according to the SHAP value to obtain the feature importance sorting set; finally, RFE is used to select the features with higher importance in the feature importance sorting set in turn, and the least important features are deleted. The XGBoost model is used to verify the accuracy of the selected feature subsets, and the optimal feature subset is screened out through iteration.

8. The method for online identification of moisture in coke powder based on multi-class heterogeneous image features according to any one of claims 1 to 7, characterized in that: When collecting the coke powder image data in step 1, a CCD industrial camera is used, the height is fixed by a bracket, and a light shield and fill lights on both sides are set to ensure consistent light intensity. The sample stage is located directly below the camera lens so that the camera can accurately and clearly capture the coke powder image.

9. The method for online identification of moisture in coke powder based on multi-class heterogeneous image features according to claim 1, characterized in that: When collecting coke powder image data in any of steps 1-7, first mix the dry coke powder and prepare coke powder samples with different moisture contents, seal the prepared coke powder and let it stand to allow moisture to penetrate evenly, crush and mix the coke powder again, press to ensure that the coke powder surface is flat, and then collect the coke powder surface image.

Images