Coal ash fusibility determination method based on residual attention and feature pyramid

Abstract

The present invention relates to the field of material analysis technology and discloses a method for measuring the fusibility of coal ash based on residual attention and feature pyramids. The method comprises: extracting morphological features of ash cones formed during coal ash melting to obtain an initial morphological feature map; performing batch normalization on the initial morphological feature map to obtain a main branch feature map; performing dimensionality transformation on the initial morphological feature map to obtain an auxiliary branch feature map; performing element-by-element fusion of the main branch feature map and the auxiliary branch feature map to obtain an initial feature map; performing adaptive calibration on the initial feature map to obtain macroscopic features; performing point convolution on the initial feature map to obtain transverse features; performing upsampling on the initial feature map to obtain longitudinal features; performing cross-scale fusion of the longitudinal and transverse features to obtain microscopic features; and aggregating the macroscopic and microscopic features into morphological features, performing nonlinear transformation and feature reconstruction, and obtaining a fusibility feature map. The present invention can improve the effectiveness of coal ash fusibility measurement.

Description

Technical Field

[0001] The present invention relates to the technical field of material analysis, and in particular to a method for determining the fusibility of coal ash based on residual attention and feature pyramid. Background Art

[0002] Coal ash, the inorganic mineral residue left after coal combustion, is widely used in building materials, agricultural improvements, and environmental protection. The fusibility of coal ash reflects its softening and melting properties at high temperatures and is a key quality indicator for coal used in power generation and gasification. Given my country's dependence on coal as its primary energy source, measuring the fusibility of coal ash is crucial for boiler design, combustion optimization, equipment safety, and clean utilization.

[0003] Currently, the determination of coal ash fusibility is primarily based on GB / T219-2008, "Determination of Coal Ash Fusibility." This method determines the deformation temperature, softening temperature, hemispherical temperature, and flow temperature by manually observing the morphological changes of the coal ash during heating. However, this traditional measurement method has numerous drawbacks. On the one hand, the operation is cumbersome and requires continuous manual observation and recording, which consumes a significant amount of manpower and time, resulting in low measurement efficiency. On the other hand, the measurement results rely heavily on manual experience, and the judgment criteria of different operators may vary. Moreover, in high-temperature environments, the accuracy of manual observation is limited, making it difficult to accurately obtain key data on coal ash morphological changes. This results in poor measurement results and cannot meet the needs for accurate measurement of coal ash fusibility in actual production and research. Summary of the Invention

[0004] The present invention provides a coal ash fusibility measurement method based on residual attention and feature pyramid, the main purpose of which is to solve the problem of poor effect in coal ash fusibility measurement.

[0005] To achieve the above objectives, the present invention provides a method for determining the fusibility of coal ash based on residual attention and feature pyramid, comprising:

[0006] S1. Extracting morphological features of an ash cone formed during coal ash melting to obtain an initial morphological feature map of the ash cone; performing batch normalization on the initial morphological feature map to obtain a main branch feature map of the ash cone; and performing dimensionality transformation on the initial morphological feature map to obtain an auxiliary branch feature map of the ash cone.

[0007] S2. merging the main branch feature map and the auxiliary branch feature map element by element to obtain an initial feature map of the gray cone, and adaptively calibrating the initial feature map based on a channel attention mechanism to obtain macroscopic features of the gray cone;

[0008] S3. Performing point convolution on the initial feature map to obtain horizontal features of the initial feature map, performing upsampling on the initial feature map to obtain vertical features of the initial feature map, and performing cross-scale fusion on the vertical features and the horizontal features to obtain microscopic features of the gray cone;

[0009] S4. Aggregating the macroscopic features and the microscopic features into morphological features of the ash cone, performing nonlinear transformation on the morphological features to obtain target morphological features of the ash cone, and reconstructing the target morphological features to obtain a meltability characteristic map of the ash cone.

[0010] In a preferred embodiment, the morphological feature extraction of the ash cone formed during the coal ash melting process to obtain the initial morphological feature map of the ash cone includes:

[0011] Collecting morphological data of ash cones formed during the fusion process of coal ash;

[0012] performing denoising processing on the morphological data to obtain standard data of the gray cone;

[0013] Performing Fourier transform on the standard data to obtain morphological characteristic data of the gray cone;

[0014] constructing a feature matrix of the gray cone based on the morphological feature data;

[0015] A characteristic map of the gray cone is calculated based on the characteristic matrix to obtain an initial morphological characteristic map of the gray cone. The formula for calculating the characteristic map is as follows:

[0016]

[0017] Where, For the An initial morphological feature map, is the activation function, For the The weight parameters of the convolution kernel, For the offset vectors, For the feature matrix, is the number of the initial feature map, is the convolution operator symbol.

[0018] In a preferred embodiment, performing batch normalization on the initial morphological feature map to obtain the main branch feature map of the gray cone includes:

[0019] Performing ReLU activation on the initial morphological feature map to obtain an activated initial feature map;

[0020] Calculate the mean and variance of feature data in the activated initial feature map;

[0021] The gray cone is normalized based on the mean value and the variance to obtain a standard characteristic graph of the gray cone, wherein the formula for the normalization is as follows:

[0022]

[0023] Where, For the A standard feature map, For the Feature data in the initial feature map after activation, is the average value, is the variance, To prevent the denominator from being zero, a very small constant is set. The number of the feature data in the initial feature map after activation;

[0024] The characteristic scale of the standard characteristic map is adjusted to obtain the main branch characteristic map of the gray cone.

[0025] In a preferred embodiment, the main branch feature map and the auxiliary branch feature map are fused element by element to obtain the initial feature map of the gray cone, wherein the calculation formula of the element-by-element fusion is as follows:

[0026]

[0027] Where, is the initial feature map, is the residual of the main branch feature map, is the auxiliary branch feature map, is the characteristic of the ash cone.

[0028] In a preferred embodiment, the adaptive calibration of the initial feature map based on the channel attention mechanism to obtain the macro features of the gray cone includes:

[0029] Performing global average pooling on the initial feature map to obtain a global feature response value for each feature in the initial feature map;

[0030] Performing two nonlinear transformations on the global feature response value to obtain the key features of the gray cone;

[0031] Generating a channel attention coefficient of the gray cone based on the key features;

[0032] The channel attention coefficient is multiplied by the initial feature map channel by channel to obtain the macro features of the gray cone.

[0033] In a preferred embodiment, performing point convolution on the initial feature map to obtain the lateral features of the initial feature map includes:

[0034] Extracting the middle layer of the initial feature map;

[0035] Perform a linear transformation on the intermediate layer to obtain the lateral features of the initial feature map, wherein the calculation formula of the linear transformation is as follows:

[0036]

[0037] Where, For the the lateral characteristics of the layer, For the said The middle layer of the layer performs point convolution operation, is the middle layer, is the number of layers of the initial feature map.

[0038] In a preferred embodiment, upsampling the initial feature map to obtain the longitudinal features of the initial features includes:

[0039] Extracting a depth feature map of the initial feature map;

[0040] Adjacent pixels in the depth feature map are weighted and refined to obtain the longitudinal features of the initial features.

[0041] In a preferred embodiment, the cross-scale fusion of the longitudinal features and the transverse features to obtain the microscopic features of the gray cone includes:

[0042] Preliminarily fusing the longitudinal features with the transverse features to obtain initial microscopic features of the ash cone;

[0043] The cross-scale aliasing effect is eliminated on the initial microscopic features to obtain the microscopic features of the gray cone, wherein the formula for eliminating the cross-scale aliasing effect is as follows:

[0044]

[0045] Where, For the the lateral characteristics of the layer, is the number of layers of the initial feature map, The lateral characteristics of layer 3, To perform a 3*3 convolution operation on the horizontal features of the third layer, is the microscopic feature of the initial feature map of the third layer, For the The microscopic features of the initial feature map of the layer, For the The initial micro features of the layer are subjected to 3*3 convolution operation. For the The vertical features of the initial feature map of the layer are upsampled.

[0046] In a preferred embodiment, performing nonlinear transformation on the morphological features to obtain target morphological features of the ash cone, and reconstructing the target morphological features to obtain the fusion characteristic map of the ash cone includes:

[0047] Performing high-dimensional mapping on the morphological features to obtain high-dimensional morphological features;

[0048] Performing ReLU activation on the high-dimensional morphological features to obtain activated high-dimensional morphological features;

[0049] performing batch normalization processing on the activated high-dimensional morphological features to obtain standardized activation features of the activated high-dimensional morphological features;

[0050] performing derandomization processing on the standardized activation features to obtain target morphological features of the gray cone;

[0051] The target morphological features are compressed to a dimension of a preset number of categories to obtain a melting characteristic map of the ash cone.

[0052] In a preferred embodiment, after reconstructing the target morphological features to obtain the ash cone melting characteristic map, the method includes:

[0053] Initializing the melting characteristic map to obtain initial parameters of the melting characteristic map;

[0054] Training the initial parameters, and when the training rounds reach a preset round, collecting the highest value, the lowest value, and the average verification accuracy of the initial parameters after training as the measurement result of the melting characteristic map;

[0055] A comparative experiment was conducted on the measurement results to obtain the measurement effect of the melting characteristic diagram.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] 1. This invention uses a dual-branch structure to process ash cone morphology data to generate primary and secondary branch feature maps. This process effectively preserves and enhances key features. Furthermore, after element-by-element fusion of the primary and secondary branch feature maps, adaptive calibration using a channel-attention mechanism precisely focuses on key information related to melting characteristics while suppressing redundant information. This multi-scale, multi-dimensional feature extraction and fusion approach more accurately reflects the coal ash melting process, significantly improving measurement accuracy.

[0058] 2. This method generates a meltability profile by performing nonlinear transformation and feature reconstruction on morphological features. This reduces manual intervention, avoids the drawbacks of traditional methods that rely on manual experience and are cumbersome to perform, and significantly improves measurement efficiency. Furthermore, this method remains stable under varying experimental environments and data conditions, and is widely adaptable to various coal ash samples. This provides reliable technical support for coal ash meltability measurements in various scenarios, effectively improving measurement results. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 A schematic flow chart of a method for determining the fusibility of coal ash based on residual attention and feature pyramid according to an embodiment of the present invention;

[0060] Figure 2 A schematic diagram of an algorithm flow of a method for determining coal ash fusibility based on residual attention and feature pyramid provided in one embodiment of the present invention;

[0061] Figure 3 Flowchart of the channel attention mechanism of the coal ash fusibility determination method based on residual attention and feature pyramid provided by one embodiment of the present invention;

[0062] Figure 4 A graph showing the similarity between the positioning frame and the standard frame image of the coal ash fusibility determination method based on residual attention and feature pyramid provided in one embodiment of the present invention;

[0063] Figure 5 This is a visualization result diagram of the positioning frame and standard frame of the coal ash fusibility determination method based on residual attention and feature pyramid provided by one embodiment of the present invention;

[0064] The purpose, features and advantages of the present invention will be further described with reference to the accompanying drawings and in conjunction with the embodiments. DETAILED DESCRIPTION

[0065] It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0066] The embodiment of the present application provides a method for determining the fusibility of coal ash based on residual attention and feature pyramid. The execution subject of the method for determining the fusibility of coal ash based on residual attention and feature pyramid includes but is not limited to at least one of the electronic devices such as a server, a terminal, etc. that can be configured to execute the method provided by the embodiment of the present application. In other words, the method for determining the fusibility of coal ash based on residual attention and feature pyramid can be executed by software or hardware installed on a terminal device or a server device. The server includes but is not limited to: a single server, a server cluster, a cloud server or a cloud server cluster, etc. The server can be an independent server, or it can be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms.

[0067] Reference Figure 1 FIG. 1 is a flow chart of a method for determining coal ash fusibility based on residual attention and feature pyramid according to an embodiment of the present invention. In this embodiment, the method for determining coal ash fusibility based on residual attention and feature pyramid includes:

[0068] S1. Extracting morphological features of an ash cone formed during coal ash melting to obtain an initial morphological feature map of the ash cone; performing batch normalization on the initial morphological feature map to obtain a main branch feature map of the ash cone; and performing dimensionality transformation on the initial morphological feature map to obtain an auxiliary branch feature map of the ash cone.

[0069] In an embodiment of the present invention, extracting the morphological features of the ash cone formed during the coal ash melting process to obtain an initial morphological feature map of the ash cone includes:

[0070] Collecting morphological data of ash cones formed during the fusion process of coal ash;

[0071] performing denoising processing on the morphological data to obtain standard data of the gray cone;

[0072] Performing Fourier transform on the standard data to obtain morphological characteristic data of the gray cone;

[0073] constructing a feature matrix of the gray cone based on the morphological feature data;

[0074] A characteristic map of the gray cone is calculated based on the characteristic matrix to obtain an initial morphological characteristic map of the gray cone. The formula for calculating the characteristic map is as follows:

[0075]

[0076] Where, For the An initial morphological feature map, is the activation function, For the The weight parameters of the convolution kernel, For the offset vectors, For the feature matrix, is the number of the initial feature map, is the convolution operator symbol.

[0077] The batch normalization process is performed on the initial morphological feature map to obtain the main branch feature map of the gray cone, including:

[0078] Performing ReLU activation on the initial morphological feature map to obtain an activated initial feature map;

[0079] Calculate the mean and variance of feature data in the activated initial feature map;

[0080] The gray cone is normalized based on the mean value and the variance to obtain a standard characteristic graph of the gray cone, wherein the formula for the normalization is as follows:

[0081]

[0082] Where, For the A standard feature map, For the Feature data in the initial feature map after activation, is the average value, is the variance, To prevent the denominator from being zero, a very small constant is set. The number of the feature data in the initial feature map after activation;

[0083] The characteristic scale of the standard characteristic map is adjusted to obtain the main branch characteristic map of the gray cone.

[0084] Specifically, the algorithm flow used in the present invention is as follows: Figure 2 The algorithm flow diagram of the coal ash fusibility determination method based on residual attention and feature pyramid is shown in the figure.

[0085] Specifically, a high-resolution industrial camera, coupled with a ring light source, was used to photograph the ash cone from multiple angles. Images were captured at fixed time intervals during the ash cone melting process to ensure a complete record of the cone's morphological changes.

[0086] Furthermore, the collected images are stored in a lossless format, and metadata such as shooting time and angle are recorded. Each set of gray cone morphological data contains multiple frames of images at different angles, constituting morphological data.

[0087] For example, a complete acquisition is performed from three angles: 0°, 45°, and 90°, with 100 frames of images taken at each angle, for a total of 300 images as morphological data.

[0088] Furthermore, the cv2.cvtColor() function of Python's OpenCV library is used to grayscale the original image, converting the color image into a grayscale image to reduce the data dimension.

[0089] Furthermore, the cv2.medianBlur() function of the OpenCV library is used to remove discrete noise such as salt and pepper noise in the image. The cv2.medianBlur() function replaces the value of each pixel in the image with the median value of its neighboring pixels, which can better preserve edge information while removing noise.

[0090] Furthermore, the cv2.threshold() function of the OpenCV library was used to binarize the denoised grayscale image to highlight the gray cone outline. The Otsu method was used to automatically determine the optimal threshold and convert the image into a black and white binary image.

[0091] Furthermore, after the above processing, a clear and noise-free gray cone binary image is obtained, which is the standard data of the gray cone.

[0092] Furthermore, the cv2.findContours() function of the OpenCV library is used to extract the contours of the binary image and obtain the contour point set of the gray cone.

[0093] Furthermore, the coordinates of the contour point set are considered as a complex number sequence using the np.fft.fft() function in Python's NumPy library.

[0094] Furthermore, the result contains the frequency information of the gray cone contour, where the low-frequency part corresponds to the overall shape of the contour and the high-frequency part corresponds to the details of the contour.

[0095] Furthermore, a certain number of low-frequency coefficients are selected as the morphological characteristic data of the gray cone. These coefficients can effectively characterize the main morphological characteristics of the gray cone.

[0096] Furthermore, for each gray cone data collection, the morphological feature data obtained after Fourier transform is used as a row vector. Assuming that an experiment collects N sets of gray cone data, and each set of data obtains a morphological feature data vector of length M, then an N×M two-dimensional array is constructed as the feature matrix.

[0097] Specifically, in the formula for feature map calculation, the previous feature matrix is convolved with the corresponding convolution kernel.

[0098] Furthermore, each convolution operation extracts a specific feature pattern from the feature matrix, which reflects the characteristics of the gray cone morphology in different local areas.

[0099] Furthermore, the corresponding offset vector is added to the result of the convolution operation. The role of the offset vector is to add a learnable bias to the result of the convolution operation, so that the model can better fit the data.

[0100] Furthermore, by adjusting the value of the offset vector, the numerical distribution of the feature map can be made more reasonable, which helps the model learn more accurate features.

[0101] Furthermore, the result after adding the offset vector is input into the activation function. The role of the activation function is to introduce nonlinear factors, break the limitations of linear operations, and enable the model to learn and express more complex relationships. After being processed by the activation function, the output result is the initial morphological feature map corresponding to the current number.

[0102] Furthermore, in the feature map calculation formula, as the feature map number increases during the calculation process, each calculation is based on the previous feature map. This means that subsequent feature maps continuously integrate and refine the information in the previous feature maps, and the level of feature abstraction becomes increasingly higher. Initially, the extraction of simple basic features such as edges and corners of the gray cone morphology gradually transitions to more advanced and representative complex morphological features.

[0103] Furthermore, during the computational process, by setting different numbers and sizes of convolution kernels, we can control the number of feature maps generated in each calculation. Generally speaking, to extract richer feature information, the number of feature maps increases as the computation progresses. Each feature map focuses on extracting a specific aspect of the gray cone morphology, and together, multiple feature maps provide a more comprehensive description of the gray cone morphology.

[0104] Furthermore, during the convolution operation, the size of the feature map typically decreases due to factors such as the sliding of the convolution kernel and the step size setting. However, this size reduction does not simply result in information loss. Instead, the convolution operation condenses and refines the features, allowing the feature map to express more critical gray cone morphological characteristics at a smaller size.

[0105] Specifically, the activation function can introduce nonlinearity and enhance the network's expressiveness. This article uses the ReLU function, which is expressed as follows:

[0106]

[0107] Where, is the initial morphological feature map, is the initial morphological feature map after activation.

[0108] Furthermore, batch normalization can stabilize the training process, accelerate the training speed, and stabilize the gradient.

[0109] Furthermore, the calculation process is as follows: First, calculate the input data The average value of:

[0110]

[0111] Where, For the Feature data in the initial feature map after activation, is the average value, is the number of feature data in the initial feature map after activation.

[0112] Furthermore, the variance of the batch data is calculated:

[0113]

[0114] Where, For the Feature data in the initial feature map after activation, is the average value, is the number of feature data in the initial feature map after activation, is the variance.

[0115] Furthermore, each data point is normalized to satisfy the distribution with mean 0 and variance 1:

[0116]

[0117] Where, For the A standard feature map, For the Feature data in the initial feature map after activation, is the average value, is the variance, To prevent the denominator from being zero, a very small constant is set. is the number of the feature data in the initial feature map after activation.

[0118] Furthermore, the normalized data is scaled and offset to make it more consistent with the actual data distribution, thereby obtaining the main branch feature map of the gray cone.

[0119] Furthermore, the auxiliary branch implements the dimension transformation of the feature map through 1×1 convolution to ensure that the size of the feature map matches the output of the main branch.

[0120] In general, extracting the morphological features of the ash cone to obtain the initial morphological feature map is the starting step to capture the key information in the ash melting process.

[0121] In general, this step can convert the complex physical morphology of the ash cone into computer-processable image feature data, completely retaining the original information related to meltability, providing a basis for subsequent precise analysis.

[0122] In general, batch normalization processing of the main branch feature map can standardize the data distribution, stabilize the training process, accelerate the model convergence speed, and reduce the gradient vanishing or explosion problem, allowing the model to learn and identify features more accurately, thereby improving the judgment accuracy of the key temperature points of coal ash fusibility.

[0123] In general, the auxiliary branch feature map obtained by dimensional transformation supplements information from another dimension and cooperates with the main branch feature map to enrich the feature expression, enabling the model to more comprehensively capture the morphological changes of gray cones and further improve the accuracy of the measurement.

[0124] In general, processing the gray cone morphological characteristics in a variety of ways can help improve the robustness of the model.

[0125] In general, in the actual coal ash fusibility measurement, the sample data may contain interference factors such as noise and uneven illumination. After feature extraction, batch normalization and dimensionality transformation, the model's sensitivity to these interferences is reduced.

[0126] In general, batch normalization processing can reduce the impact of data distribution differences and enhance the adaptability of the model to different sample data; the introduction of auxiliary branch feature maps provides the model with an additional information dimension, so that it can still stably extract effective features when facing complex and changeable ash cone morphologies, accurately judge the melting characteristics of coal ash, and ensure the reliability and stability of the measurement results.

[0127] S2. merging the main branch feature map and the auxiliary branch feature map element by element to obtain an initial feature map of the gray cone, and adaptively calibrating the initial feature map based on a channel attention mechanism to obtain macroscopic features of the gray cone;

[0128] In an embodiment of the present invention, the main branch feature map and the auxiliary branch feature map are fused element by element to obtain the initial feature map of the gray cone, wherein the calculation formula of the element-by-element fusion is as follows:

[0129]

[0130] Where, is the initial feature map, is the residual of the main branch feature map, is the auxiliary branch feature map, is the characteristic of the ash cone.

[0131] The adaptive calibration of the initial feature map based on the channel attention mechanism to obtain the macro features of the gray cone includes:

[0132] Performing global average pooling on the initial feature map to obtain a global feature response value for each feature in the initial feature map;

[0133] Performing two nonlinear transformations on the global feature response value to obtain the key features of the gray cone;

[0134] Generating a channel attention coefficient of the gray cone based on the key features;

[0135] The channel attention coefficient is multiplied by the initial feature map channel by channel to obtain the macro features of the gray cone.

[0136] Specifically, a 3*3 convolution process is performed on the main branch feature map to obtain main branch feature data of the main branch feature map.

[0137] Furthermore, batch normalization is performed on the main branch feature data to obtain the residual of the main branch feature map, wherein the formula for the batch normalization is as follows:

[0138]

[0139] Where, is the residual of the main branch feature map, is the main branch feature data, is the average value of the main branch feature data, is the variance of the main branch feature data, A very small constant set to prevent the denominator from being zero.

[0140] Furthermore, the key change information carried by the residual of the main branch feature map is combined with the supplementary information of the auxiliary branch feature map, so that the fused initial feature map can describe the characteristics of the gray cone more comprehensively and richly.

[0141] Furthermore, this feature complementation approach can enhance the feature map's ability to express aspects such as the gray cone morphology, which is helpful for subsequent gray cone analysis and recognition tasks.

[0142] Furthermore, the element-by-element fusion process is actually a process of integrating and optimizing feature information from different branches.

[0143] Furthermore, through fusion, some possible redundant information is removed, and useful information is strengthened, making the information in the initial feature map more refined and valuable, which is more conducive to model extraction and utilization of these features for learning and judgment.

[0144] Furthermore, since the initial feature maps obtained by element-by-element fusion contain higher-quality and more comprehensive feature information, subsequent model training and inference operations based on these feature maps are expected to improve the model's performance for gray cone-related tasks, such as improving classification accuracy and reducing errors.

[0145] Specifically, the process of adaptively calibrating the initial feature map based on the channel attention mechanism is as follows: Figure 3 The flow chart of the channel attention mechanism of the coal ash fusibility determination method based on residual attention and feature pyramid is shown.

[0146] Specifically, global context information at the channel level is obtained through global average pooling, and then the importance weight of each channel is calculated through a bottleneck structure consisting of two levels of fully connected layers.

[0147] Furthermore, the first-level fully connected layer compresses the channel dimension to 1 / 16 of the original dimension, the second-level fully connected layer restores the original number of channels, and finally generates a channel attention coefficient between 0 and 1 through the Sigmoid function.

[0148] Furthermore, these coefficients are multiplied channel by channel with the original feature map to achieve adaptive calibration of feature channels.

[0149] In general, the element-by-element fusion of the main branch feature map and the auxiliary branch feature map realizes the organic combination of multi-dimensional features.

[0150] In general, the main branch feature map retains the key morphological features after batch normalization, and the auxiliary branch feature map supplements the information from another perspective after dimension transformation.

[0151] In general, the initial feature map obtained by fusing the two combines the advantages of different branches, contains more comprehensive and rich ash cone morphology information, provides a more sufficient data basis for subsequent analysis, and helps to more accurately reflect the characteristic changes of coal ash during the melting process.

[0152] In general, the adaptive calibration of the initial feature map based on the channel attention mechanism is an important advantage of this method.

[0153] In general, the channel attention mechanism obtains the global feature response value through global average pooling, and then generates the channel attention coefficient through two nonlinear transformations.

[0154] In general, these coefficients are multiplied with the initial feature map channel by channel, enabling the model to automatically focus on key features closely related to ash fusibility and suppress irrelevant or redundant information.

[0155] In general, in actual measurements, the morphological changes of coal ash are complex and diverse. Through this adaptive calibration, the macroscopic features can more accurately highlight the information directly related to the melting characteristics, such as the key contour changes of the ash cone, the morphological characteristics of specific areas, etc., which greatly enhances the specificity of the features and significantly improves the accuracy of the judgment of the fusibility of coal ash.

[0156] S3. Performing point convolution on the initial feature map to obtain horizontal features of the initial feature map, performing upsampling on the initial feature map to obtain vertical features of the initial feature map, and performing cross-scale fusion on the vertical features and the horizontal features to obtain microscopic features of the gray cone;

[0157] In an embodiment of the present invention, performing point convolution on the initial feature map to obtain the lateral features of the initial feature map includes:

[0158] Extracting the middle layer of the initial feature map;

[0159] Perform a linear transformation on the intermediate layer to obtain the lateral features of the initial feature map, wherein the calculation formula of the linear transformation is as follows:

[0160]

[0161] Where, For the the lateral characteristics of the layer, For the said The middle layer of the layer performs point convolution operation, is the middle layer, is the number of layers of the initial feature map.

[0162] The upsampling of the initial feature map to obtain the longitudinal features of the initial features includes:

[0163] Extracting a depth feature map of the initial feature map;

[0164] Adjacent pixels in the depth feature map are weighted and refined to obtain the longitudinal features of the initial features.

[0165] The cross-scale fusion of the longitudinal features and the transverse features to obtain the microscopic features of the gray cone includes:

[0166] Preliminarily fusing the longitudinal features with the transverse features to obtain initial microscopic features of the ash cone;

[0167] The cross-scale aliasing effect is eliminated on the initial microscopic features to obtain the microscopic features of the gray cone, wherein the formula for eliminating the cross-scale aliasing effect is as follows:

[0168]

[0169] Where, For the the lateral characteristics of the layer, is the number of layers of the initial feature map, The lateral characteristics of layer 3, To perform a 3*3 convolution operation on the horizontal features of the third layer, is the microscopic feature of the initial feature map of the third layer, For the The microscopic features of the initial feature map of the layer, For the The initial micro features of the layer are subjected to 3*3 convolution operation. For the The vertical features of the initial feature map of the layer are upsampled.

[0170] Specifically, in PyTorch, the hook() function is used to get the output of the intermediate layer.

[0171] Furthermore, the hook function can obtain the output of a specific layer during the forward propagation of the model without modifying the model structure.

[0172] Furthermore, point convolution uses a 1×1 convolution kernel to perform convolution operations on the intermediate layer of the input.

[0173] Furthermore, the 1×1 convolution kernel slides at each position in the middle layer, multiplies the element on the channel corresponding to the position by the weight in the convolution kernel and sums them to obtain a new value.

[0174] Furthermore, this process is repeated for each position in the intermediate layer to generate a new feature map.

[0175] Furthermore, such a point convolution operation is performed on each intermediate layer to obtain the lateral features of the initial feature map.

[0176] Furthermore, point convolution operations can easily adjust the dimension of features. By setting different numbers of 1×1 convolution kernels, the number of channels in the feature map can be increased or decreased.

[0177] Furthermore, if the number of channels is increased, the lateral features will contain more feature information, which may improve the model's ability to express complex features; if the number of channels is reduced, the dimension of the features can be reduced, the amount of calculation and the complexity of the model can be reduced, and it also helps to alleviate the overfitting problem.

[0178] Furthermore, a simple convolutional neural network model is built using TensorFlow's Keras API to extract deep feature maps.

[0179] Furthermore, the initial feature map data is input into the constructed model, and the forward propagation calculation is performed through the model() function of the model to obtain the depth feature map depth_feature_maps.

[0180] Specifically, the resolution is directly expanded by performing weighted averaging on adjacent pixels in the depth feature map to obtain the longitudinal feature of the initial feature.

[0181] Furthermore, the longitudinal features and transverse features of corresponding scales are fused to obtain the initial microscopic features of the gray cone.

[0182] Furthermore, a 3×3 convolution operation is performed on the horizontal features of the third layer.

[0183] Furthermore, the 3×3 convolution kernel slides over each local region of the third layer’s horizontal features, operates on the elements in the region with the weights of the convolution kernel, and extracts specific local patterns in the features of this layer through weighted summation to obtain the microscopic features of the initial feature map of the third layer.

[0184] Furthermore, for Upsample the vertical features of the initial feature map of the first layer to increase its resolution to the same level as the first layer. The lateral characteristics of the layers are the same or matching in scale.

[0185] Furthermore, upsampling inserts new pixel values between the original pixels through interpolation and other methods to increase the size of the feature map, so that features of different scales can be fused at the same scale.

[0186] Furthermore, the upsampled Layer longitudinal characteristics and The horizontal features of the layers are added to achieve the fusion of features at different scales.

[0187] Furthermore, this step integrates feature information at different levels and dimensions (horizontal and vertical), making up for the limitations of single-scale features.

[0188] Furthermore, a 3×3 convolution operation is performed on the fused features, and the fused features are processed again to further extract and optimize the features to obtain the first The microscopic features of the initial feature map of the layer.

[0189] In general, point convolution obtains lateral features, can focus on the local detail information of the initial feature map, explore the associations and feature differences between different layers, and capture subtle changes in the gray cone in a certain plane dimension, such as the texture of the gray cone surface and local shape changes.

[0190] In general, upsampling obtains longitudinal features, and then the feature map is expanded from the depth dimension, converting the low-resolution feature map into high-resolution, restoring the spatial information lost by the convolution operation, and reflecting the changing trend of the gray cone morphology at different scales.

[0191] In general, these two methods respectively obtain horizontal and vertical characteristics, and carefully depict the gray cone morphology from different dimensions and scales, making the information obtained more comprehensive and providing a rich data foundation for subsequent precise analysis.

[0192] In general, the cross-scale fusion of horizontal and vertical features organically combines information of different scales and dimensions.

[0193] In general, this fusion method avoids the limitations of single-scale analysis, can comprehensively consider the overall structure and local details of the gray cone, and more accurately reflect the true morphology of the gray cone at the microscopic level.

[0194] In general, by eliminating cross-scale aliasing effects and other operations, the fusion effect is further optimized, making the microscopic features more accurate.

[0195] In general, in the determination of coal ash fusibility, microscopic characteristics are crucial for judging the melting properties of coal ash. More precise microscopic characteristics can help more accurately determine key indicators such as the deformation temperature and softening temperature of coal ash, thereby effectively improving the accuracy and reliability of the measurement results and providing more valuable reference for related industrial production and research.

[0196] S4. Aggregating the macroscopic features and the microscopic features into morphological features of the ash cone, performing nonlinear transformation on the morphological features to obtain target morphological features of the ash cone, and reconstructing the target morphological features to obtain a meltability characteristic map of the ash cone.

[0197] In an embodiment of the present invention, performing nonlinear transformation on the morphological features to obtain target morphological features of the ash cone, and reconstructing the target morphological features to obtain the fusion characteristic map of the ash cone include:

[0198] Performing high-dimensional mapping on the morphological features to obtain high-dimensional morphological features;

[0199] Performing ReLU activation on the high-dimensional morphological features to obtain activated high-dimensional morphological features;

[0200] performing batch normalization processing on the activated high-dimensional morphological features to obtain standardized activation features of the activated high-dimensional morphological features;

[0201] performing derandomization processing on the standardized activation features to obtain target morphological features of the gray cone;

[0202] The target morphological features are compressed to a dimension of a preset number of categories to obtain a melting characteristic map of the ash cone.

[0203] After reconstructing the target morphological features to obtain the ash cone melting characteristic map, the method includes:

[0204] Initializing the melting characteristic map to obtain initial parameters of the melting characteristic map;

[0205] Training the initial parameters, and when the training rounds reach a preset round, collecting the highest value, the lowest value, and the average verification accuracy of the initial parameters after training as the measurement result of the melting characteristic map;

[0206] A comparative experiment was conducted on the measurement results to obtain the measurement effect of the melting characteristic diagram.

[0207] Specifically, the 256-dimensional input layer is mapped to a 2048-dimensional latent space and ReLU activation function and batch normalization are applied.

[0208] Furthermore, a 50% Dropout layer is introduced to perform de-randomization processing to prevent overfitting.

[0209] Furthermore, the second layer maintains a 2048-dimensional feature space and repeats activation and normalization operations. Finally, the output layer compresses the high-dimensional features to the dimension of the target category number.

[0210] Furthermore, the entire structure uses two nonlinear transformations and feature reorganization to maintain strong representation capabilities while ensuring the generalization performance and training stability of the model using Dropout and batch normalization, ultimately achieving an effective mapping from original features to classification decisions.

[0211] Specifically, a comparative experiment was conducted on the RAP architecture used in the present invention and the classic network architecture (including AlexNet, VGG16 and ResNet versions) on the task of coal ash fusibility determination.

[0212] Furthermore, the He initialization method was used for parameter initialization, and the cross entropy loss function and Adam optimizer were used for model training. The batch sizes of the training set and test set were set to 32 and 8 respectively, the learning rate was 0.0001, and the number of training rounds was 50.

[0213] Furthermore, the experiment was conducted in a hardware environment equipped with NVIDIA GPU (RTX3080), using 4 worker processes to load data in parallel.

[0214] Furthermore, during the experiment, the training loss and accuracy changes of each round are monitored in real time, and the highest, lowest and average verification accuracy are finally recorded to ensure the reliability and repeatability of the experimental results.

[0215] In detail, the experimental results are shown in Table 1.

[0216]

[0217] In detail, the minimum accuracy, maximum accuracy and average accuracy of the RAP architecture of the present invention are better than other models. The basic architectures AlexNet and VGG16 perform relatively poorly in terms of minimum, maximum and average accuracy. The main reason for this phenomenon is that their network structures lack residual connections and attention mechanisms, which limits the feature extraction capabilities and model performance. In contrast, the ResNet network series shows better performance, among which ResNet-50 outperforms ResNet-34 and ResNet-101 in accuracy. This advantage stems from the moderate network depth and parameter scale of ResNet-50: its depth can effectively learn complex feature representations while avoiding underfitting caused by the network being too shallow or overfitting caused by being too deep, thereby achieving the best performance balance. RAP further enhances the multi-scale feature fusion capability by integrating the residual-attention module and the feature pyramid module, and ultimately achieves the best performance on the test set. These results fully demonstrate the technical advantages of the module combination proposed in this article in the task of coal ash fusibility determination.

[0218] Furthermore, in order to verify the model's ability to identify the ash melting temperature point in dynamic videos, this paper designed a systematic experimental process.

[0219] In detail, the experiment uses comparative analysis method to compare the performance of the proposed RAP model with AlexNet, VGG16 and ResNet series models.

[0220] Specifically, a high-temperature coal ash cone melting video is first input, and each model processes the video frame by frame. When the same temperature label is predicted for five consecutive frames, it is determined to be a valid temperature point and the corresponding first frame image is extracted.

[0221] In detail, to quantify the model performance, this paper designs an evaluation index system from two dimensions: time and morphology.

[0222] In detail, in the temporal dimension, the accuracy of temperature prediction is evaluated by frame error and average frame error.

[0223] In detail, the frame error is defined as the absolute deviation between the predicted temperature value of a single frame and the true value, and the average frame error reflects the overall prediction accuracy.

[0224] In detail, the frame extraction results of each model are shown in Table 2.

[0225]

[0226] Furthermore, in terms of morphological dimension, the structural similarity index (SSIM) was used to evaluate the characterization ability of the melt morphology. SSIM calculates image similarity based on three components: brightness, contrast, and structure. The formula for calculating the structural similarity index is as follows:

[0227]

[0228] Where, is the structural similarity index, 、 and Represent the comparison functions of brightness, contrast and structure respectively, 、 as well as To adjust the weight, Characterized by the molten state.

[0229] In detail, the comprehensive analysis of the above indicators can comprehensively evaluate the performance of the model in temperature point recognition and morphology characterization.

[0230] In detail, the similarity results between the positioning frames of each model and the standard frame images are shown in Figure 4 Image similarity results of the positioning frame and standard frame of the coal ash fusibility determination method based on residual attention and feature pyramid;

[0231] Furthermore, in order to make the model comparison more impressive, the difference between the model extraction frame and the standard frame morphology is shown in the comparison chart. The visualization results are shown in Figure 5 Visualization results of the positioning frame and standard frame of the coal ash fusibility determination method based on residual attention and feature pyramid.

[0232] In detail, the positioning frame is a fusibility image generated by measuring the coal ash image of each model, and the standard frame is real coal ash image data used as a reference.

[0233] In detail, according to the GB / T219-2008 standard, the RAP model outperforms other classic network architectures in terms of frame extraction accuracy and stability at different temperature points.

[0234] In detail, it performs well at temperature points DT, ST, HT, and FT, with the lowest average frame error and higher image similarity, indicating that it has strong adaptability and extraction capabilities for image features at different temperature points.

[0235] In detail, in comparison, although ResNet50 and VGGNet accurately extracted the HT temperature point, the frame error at the temperature point DT was high, which affected their overall performance; although ResNet101 and ResNet34 performed well at some temperature points, the extraction errors at most temperature points were large; AlexNet's overall performance was relatively stable, but the error at the temperature point FT was high, and the average frame error was slightly higher than that of the RAP model.

[0236] In detail, the RAP model has obvious advantages in the task of coal ash fusibility determination and can more accurately extract frames at different temperature points, providing more reliable technical support for subsequent research and applications.

[0237] In the several embodiments provided by the present invention, it should be understood that the disclosed methods can be implemented in other ways.

[0238] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0239] The embodiments of the present application can acquire and process relevant data based on artificial intelligence technology. Artificial intelligence is the theory, method and technology of using digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to achieve optimal results.

[0240] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A coal ash fusibility determination method based on residual attention and feature pyramid, characterized in that: The method comprises: S1. Collecting morphological data of ash cones formed during the fusion process of coal ash; performing denoising processing on the morphological data to obtain standard data of the gray cone; Performing Fourier transform on the standard data to obtain morphological characteristic data of the gray cone; constructing a feature matrix of the gray cone based on the morphological feature data; A characteristic map of the gray cone is calculated based on the characteristic matrix to obtain an initial morphological characteristic map of the gray cone. The formula for calculating the characteristic map is as follows: ， Where, For the An initial morphological feature map, is the activation function, For the The weight parameters of the convolution kernel, For the offset vectors, For the feature matrix, is the number of the initial feature map, is the convolution operator symbol; Performing batch normalization on the initial morphological feature map to obtain a main branch feature map of the gray cone, and performing dimension transformation on the initial morphological feature map to obtain an auxiliary branch feature map of the gray cone; S2. merging the main branch feature map and the auxiliary branch feature map element by element to obtain an initial feature map of the gray cone, and adaptively calibrating the initial feature map based on a channel attention mechanism to obtain macroscopic features of the gray cone; S3. Performing point convolution on the initial feature map to obtain horizontal features of the initial feature map, performing upsampling on the initial feature map to obtain vertical features of the initial feature map, and performing cross-scale fusion on the vertical features and the horizontal features to obtain microscopic features of the gray cone; S4. Aggregating the macroscopic features and the microscopic features into morphological features of the ash cone, performing nonlinear transformation on the morphological features to obtain target morphological features of the ash cone, and performing feature reconstruction on the target morphological features to obtain a meltability characteristic map of the ash cone; When the melting characteristic diagram has the same temperature for 5 consecutive frames, the temperature is determined to be a valid temperature point of the melting characteristic diagram.

2. The method for determining coal ash fusibility based on residual attention and feature pyramid according to claim 1, wherein: The batch normalization process is performed on the initial morphological feature map to obtain the main branch feature map of the gray cone, including: Performing ReLU activation on the initial morphological feature map to obtain an activated initial feature map; Calculate the mean and variance of feature data in the activated initial feature map; The gray cone is normalized based on the mean value and the variance to obtain a standard characteristic graph of the gray cone, wherein the formula for the normalization is as follows: ， Where, For the A standard feature map, For the Feature data in the initial feature map after activation, is the average value, is the variance, To prevent the denominator from being zero, a very small constant is set. The number of the feature data in the initial feature map after activation; The characteristic scale of the standard characteristic map is adjusted to obtain the main branch characteristic map of the gray cone.

3. The method for determining coal ash fusibility based on residual attention and feature pyramid according to claim 1, wherein: The main branch feature map and the auxiliary branch feature map are fused element by element to obtain the initial feature map of the gray cone, wherein the calculation formula of the element-by-element fusion is as follows: , Where, is the initial feature map, is the residual of the main branch feature map, is the auxiliary branch feature map, is the characteristic of the ash cone.

4. The method for determining coal ash fusibility based on residual attention and feature pyramid according to claim 1, wherein: The adaptive calibration of the initial feature map based on the channel attention mechanism to obtain the macro features of the gray cone includes: Performing global average pooling on the initial feature map to obtain a global feature response value for each feature in the initial feature map; Performing two nonlinear transformations on the global feature response value to obtain the key features of the gray cone; Generating a channel attention coefficient of the gray cone based on the key features; The channel attention coefficient is multiplied by the initial feature map channel by channel to obtain the macro features of the gray cone.

5. The method for determining coal ash fusibility based on residual attention and feature pyramid according to claim 1, wherein: The performing point convolution on the initial feature map to obtain the lateral features of the initial feature map includes: Extracting the middle layer of the initial feature map; Perform a linear transformation on the intermediate layer to obtain the lateral features of the initial feature map, wherein the calculation formula of the linear transformation is as follows: ， Where, For the the lateral characteristics of the layer, For the said The middle layer of the layer performs point convolution operation, is the middle layer, is the number of layers of the initial feature map.

6. The method for determining coal ash fusibility based on residual attention and feature pyramid according to claim 1, wherein: The upsampling of the initial feature map to obtain the longitudinal features of the initial features includes: Extracting a depth feature map of the initial feature map; Adjacent pixels in the depth feature map are weighted and refined to obtain the longitudinal features of the initial features.

7. The method for determining coal ash fusibility based on residual attention and feature pyramid according to claim 1, wherein: The cross-scale fusion of the longitudinal features and the transverse features to obtain the microscopic features of the gray cone includes: Preliminarily fusing the longitudinal features with the transverse features to obtain initial microscopic features of the ash cone; The cross-scale aliasing effect is eliminated on the initial microscopic features to obtain the microscopic features of the gray cone, wherein the formula for eliminating the cross-scale aliasing effect is as follows: ， Where, For the the lateral characteristics of the layer, is the number of layers of the initial feature map, The lateral characteristics of layer 3, To perform a 3*3 convolution operation on the horizontal features of the third layer, is the microscopic feature of the initial feature map of the third layer, For the The microscopic features of the initial feature map of the layer, For the The initial micro features of the layer are subjected to 3*3 convolution operation. For the The vertical features of the initial feature map of the layer are upsampled.

8. The method for determining coal ash fusibility based on residual attention and feature pyramid according to claim 1, wherein: The performing nonlinear transformation on the morphological features to obtain target morphological features of the ash cone, and performing feature reconstruction on the target morphological features to obtain a fusibility characteristic map of the ash cone, includes: Performing high-dimensional mapping on the morphological features to obtain high-dimensional morphological features; Performing ReLU activation on the high-dimensional morphological features to obtain activated high-dimensional morphological features; performing batch normalization processing on the activated high-dimensional morphological features to obtain standardized activation features of the activated high-dimensional morphological features; performing derandomization processing on the standardized activation features to obtain target morphological features of the gray cone; The target morphological features are compressed to a dimension of a preset number of categories to obtain a melting characteristic map of the ash cone.

9. The method for determining coal ash fusibility based on residual attention and feature pyramid according to claim 1, wherein: After reconstructing the target morphological features to obtain the ash cone melting characteristic map, the method includes: Initializing the melting characteristic map to obtain initial parameters of the melting characteristic map; Training the initial parameters, and when the training rounds reach a preset round, collecting the highest value, the lowest value, and the average verification accuracy of the initial parameters after training as the measurement result of the melting characteristic map; A comparative experiment was conducted on the measurement results to obtain the measurement effect of the melting characteristic diagram.

Images