A machine learning-based refractory material quality detection method and system

Abstract

The application provides a kind of machine learning-based refractory quality detection method and system, the method comprises: the multi-angle surface image set of refractory sample is collected, image under different illumination conditions is covered, to capture material surface characteristics comprehensively;Call pre-trained image feature analysis model to extract surface texture features, pore distribution features and micro-crack features from it, form image feature set;Based on the joint classification processing of preset defect classification rule base to image feature set, identify and quantify defect type and its distribution parameter;According to defect type set and distribution parameter, generate quality evaluation parameter, and compare with preset quality standard threshold value, output quality detection result;The detection result is fed back to production line control system, and automatic trigger sorting operation and process parameter adjustment, to realize the intelligent monitoring and optimization of production process, thereby improve the accuracy and efficiency of refractory quality detection, promote the improvement of production quality.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine vision, in particular to a refractory material quality detection method and system based on machine learning. BACKGROUND

[0002] In the field of refractory material manufacturing, quality detection has always been a key link to ensure product performance and safe production. Traditional refractory material quality detection methods mainly rely on manual visual inspection, physical property testing, and limited automated detection equipment.

[0003] Manual visual inspection is highly dependent on the experience and subjective judgment of the inspectors, which is not only inefficient, but also susceptible to factors such as fatigue and changes in lighting conditions, resulting in instability and high error rate of the detection results. Especially when facing complex surface textures, small pores and micro-cracks, manual detection often has difficulty in accurate identification and quantification.

[0004] Secondly, physical property testing methods, such as compressive strength testing and bending strength testing, can provide basic mechanical property indicators of the material, but these tests are usually destructive and cannot fully reflect the internal microstructure and defect distribution of the material, which is not suitable for quality detection scenarios that require the integrity of the material to be maintained.

[0005] Furthermore, existing automated detection equipment is mostly based on single sensors or simple image processing techniques, which have limited detection range and are difficult to capture multi-angle features of the material surface, especially under different lighting conditions. In addition, these devices often lack sufficient intelligence in defect identification and classification, and cannot accurately classify and quantitatively analyze complex defect patterns. SUMMARY

[0006] In view of the above-mentioned problems, in combination with the first aspect of the present application, the embodiments of the present application provide a refractory material quality detection method based on machine learning, which comprises:

[0007] Collecting a multi-angle surface image set of the refractory material sample, the multi-angle surface image set containing material surface images under different lighting conditions;

[0008] Calling a pre-trained image feature analysis model to perform image feature extraction on the multi-angle surface image set to obtain an image feature set of the refractory material sample, the image feature set including surface texture features, pore distribution features and micro-crack features;

[0009] Based on a pre-set defect classification rule library, the image feature set is jointly classified to generate a defect type set of the refractory material sample and corresponding defect distribution parameters;

[0010] generate a quality evaluation parameter of the refractory material sample according to the defect type set and the defect distribution parameter, and compare the quality evaluation parameter with a preset quality standard threshold to output a quality detection result;

[0011] feed back the quality detection result to a production line control system to trigger a sorting operation and a process parameter adjustment operation.

[0012] In still another aspect, the embodiments of the present application also provide a refractory material quality detection system based on machine learning, comprising a processor, a machine readable storage medium, the machine readable storage medium being connected with the processor, the machine readable storage medium being used for storing programs, instructions or codes, and the processor being used for executing the programs, instructions or codes in the machine readable storage medium to realize the above-mentioned method.

[0013] Based on the above aspects, the embodiments of the present application effectively overcome the feature recognition deviation caused by uneven illumination in the traditional detection method by collecting a multi-angle surface image set, especially by incorporating image data under different illumination conditions, and ensure the integrity and accuracy of image feature extraction. The pre-trained image feature analysis model can deeply mine the surface texture features, pore distribution features and micro-crack features of the refractory material sample to form a multi-dimensional image feature set. The preset defect classification rule library is based on a machine learning algorithm to jointly classify the image feature set, not only to identify various defect types, but also to calculate defect distribution parameters, thereby realizing the all-around and fine evaluation of the internal and surface defects of the refractory material. On this basis, the generated quality evaluation parameter is compared with the preset quality standard threshold to output an objective and quantitative quality detection result, effectively avoiding the subjectivity and errors of human judgment. Finally, the quality detection result is fed back to the production line control system in real time to trigger the sorting operation and the process parameter adjustment operation, forming a dynamic linkage mechanism of quality monitoring and process optimization, which not only improves the production efficiency and product qualification rate, but also promotes the continuous optimization and intelligent upgrading of the production process. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1 is an execution flow diagram of the refractory material quality detection method based on machine learning provided by the embodiments of the present application.

[0015] Figure 2 is a schematic diagram of exemplary hardware and software components of the refractory material quality detection system based on machine learning provided by the embodiments of the present application. DETAILED DESCRIPTION

[0016] The present application will be described in detail below with reference to the accompanying drawings, Figure 1is a flowchart of a machine learning-based refractory material quality detection method provided by an embodiment of the present application. The machine learning-based refractory material quality detection method will be described in detail below.

[0017] Step S110: Collect a multi-angle surface image set of the refractory material sample, which contains material surface images under different lighting conditions.

[0018] During the refractory material quality detection process, comprehensive and accurate surface image information is crucial for subsequent analysis. Therefore, a multi-angle surface image set covering different lighting conditions needs to be collected. This step includes the following sub-steps:

[0019] Step S111: Configure the shooting parameters of the industrial camera array, including light source intensity, polarization angle, and focal length, so that the surface images at different angles have consistent brightness distribution and contrast range.

[0020] Before using the industrial camera array to collect images, the shooting parameters need to be configured. The light source intensity directly affects the brightness of the image. If the light source intensity is too high, the image will be overexposed, causing some details to be hidden; if it is too low, the image will be too dark and it will be difficult to see the surface features. For refractory material samples with high surface reflectivity, the light source intensity should be reduced. For example, when the sample surface is similar to a mirror and the light reflection is strong, the power output of the light source can be appropriately adjusted to reduce the interference of reflected light on the image. For samples with dark colors and complex textures, the light source intensity needs to be increased. For example, for samples with blackish colors and complex surface textures, increasing the light source power can make the texture more clearly visible. By continuously adjusting the light source intensity, the brightness of the captured image is moderate, and the brightness of each region is uniform, avoiding local over-brightness or over-darkness.

[0021] For the adjustment of the polarization angle, when light is incident on the surface of the refractory material sample, reflected light is generated. If this reflected light is not controlled, it will cause image flare, affecting the observation of surface features. When adjusting the polarization angle, a polarizer can be installed in front of the camera lens, and the angle of the polarizer can be changed by rotating it. In actual operation, a range of polarization angles can be roughly determined, and then fine-tuned by observing the effect of the captured image. When the reflected light in the image is significantly reduced and the surface features are more clearly visible, it means that the appropriate polarization angle has been found.

[0022] The setting of focal length is directly related to the sharpness of the image. If the focal length is not accurate, the image will be blurred. When setting the focal length, the approximate distance between the industrial camera and the refractory sample can be used for preliminary adjustment. For samples taken at close range, the focal length setting is relatively short; for samples taken at a distance, the focal length needs to be adjusted appropriately. Then, the camera's autofocus function can be used for accurate adjustment, or the focal length can be manually adjusted until the image is the sharpest. At the same time, make sure the camera's focus is accurate on the sample surface, avoid partial area blur caused by focus offset.

[0023] Step S112: Control the industrial camera array to collect material surface images of the refractory sample according to a preset time synchronization sequence, the material surface images including top surface images, side surface images and bottom surface images, wherein the relative position between the industrial camera array and the refractory sample is calibrated by a laser range finder.

[0024] The preset time synchronization sequence is the key to ensure that each industrial camera takes pictures at the same time or at a set time interval. In actual operation, a high-precision clock signal generator can be used to generate a time synchronization signal. This time synchronization signal is transmitted to all industrial cameras at the same time, so that they start shooting at the same time. For example, set all cameras to take pictures at the same time every 0.1 seconds, which can ensure that the multi-angle images collected have consistency in time.

[0025] When collecting material surface images, images of the top, side and bottom need to be obtained. The top surface image can show the overall appearance of the top of the sample, such as whether there are obvious protrusions, depressions or cracks and other defects. The side surface image can reflect the texture and structural characteristics of the side of the sample, which helps to find problems such as delamination and holes on the side. The bottom surface image can provide information about the support condition and possible defects on the bottom of the sample.

[0026] To ensure the accuracy of image collection, a laser range finder can be used to calibrate the relative position between the industrial camera array and the refractory sample. The laser range finder determines the distance by emitting a laser beam and measuring the reflection time. Before collecting images, the laser range finder is installed near the camera array so that it can accurately measure the distance between the camera and the sample. Real-time monitoring of the distance, if the distance changes, adjust the position or focal length of the camera in time to ensure the accuracy of image collection. For example, when the sample moves on the conveyor belt, the laser range finder can provide real-time feedback on distance changes, which can be used to automatically adjust the position of the camera to ensure clear and accurate images.

[0027] Step S113: Perform geometric correction processing on the collected material surface images to generate a multi-angle surface image set.

[0028] In this embodiment, the original material surface images collected may be geometrically deformed due to factors such as camera shooting angle, position, and sample placement. Geometric correction processing is a necessary step to eliminate these deformations and restore the normal geometry of the images.

[0029] The geometric correction processing includes operations such as translation, rotation, and scaling. The translation operation can adjust the position of the image in the plane to align it with the standard coordinate system. In actual operation, the direction and distance of translation can be determined according to specific reference points in the image. For example, if the sample in the image should be located at the center position, but is actually offset to the left, the image needs to be translated to the right by a certain distance.

[0030] The rotation operation is used to correct the tilt of the image caused by the deviation of the shooting angle. The tilt angle can be calculated by detecting obvious straight line features in the image, such as the edges of the sample, and then the image is rotated by the corresponding angle for correction. For example, if the edge of the sample is detected to have a 5-degree angle with the horizontal direction, the image is rotated -5 degrees to restore the horizontal direction.

[0031] The scaling operation is to adjust the size of the image to conform to the unified size standard. Different industrial cameras may have different resolutions, and the sizes of the collected images will also be different. Through the scaling operation, the width and height of all images are unified to the preset standard value. During scaling, attention should be paid to maintaining the proportion of the image unchanged to avoid image deformation. For example, if the standard image size is 800 pixels in width and 600 pixels in height, and the collected image width is 1000 pixels and height is 750 pixels, the image can be scaled down to the standard size by the same proportion.

[0032] Step S114: performing data enhancement processing on the multi-angle surface image set, the data enhancement processing including small-angle rotation, brightness adjustment, and local contrast disturbance, and prohibiting mirror flipping operation to avoid distortion of directional features, to expand the multi-angle surface image set.

[0033] In this embodiment, the data enhancement processing can increase the diversity and richness of image data and improve the generalization ability of the subsequent model. The small-angle rotation operation can simulate different observation angles. In operation, the center of the image is taken as the rotation point, and the image is rotated by a small angle, such as between 10 degrees and 20 degrees positive or negative. The selection of the rotation angle can be adjusted according to the actual situation, and a new image is obtained after each rotation. For example, rotating the image clockwise by 15 degrees obtains a new image with a different viewing angle.

[0034] The brightness adjustment can simulate different lighting conditions. The brightness of the image can be changed by adjusting the pixel values of the image. For example, multiply the brightness value of each pixel in the image by a coefficient. If the coefficient is greater than 1, the image becomes brighter. If the coefficient is less than 1, the image becomes darker. The value range of the coefficient can be determined according to the actual situation, such as between 0.8 and 1.2. By adjusting the brightness coefficient multiple times, images with different brightness can be obtained.

[0035] The local contrast disturbance can enhance the contrast of local regions in the image. The image can be divided into multiple small regions, and the contrast of each small region can be adjusted. For example, the pixel values of the local region are redistributed using histogram equalization or other methods to enhance the contrast of the region. In operation, some small regions can be randomly selected for contrast adjustment, and the degree of adjustment can also be different each time, so that images with different local contrasts are obtained.

[0036] It should be noted that the mirror flip operation is prohibited because the mirror flip will change the directional features of the image, causing distortion of the directional information and affecting the subsequent analysis and judgment of features such as surface defects. Through these data enhancement processing operations, the multi-angle surface image set can be expanded.

[0037] Step S120: calling a pre-trained image feature analysis model to perform image feature extraction on the multi-angle surface image set to obtain an image feature set of the refractory material sample, the image feature set including surface texture features, pore distribution features, and micro-crack features.

[0038] After obtaining the multi-angle surface image set, key image features need to be extracted from it using a pre-trained image feature analysis model. This step includes the following sub-steps:

[0039] Step S121: performing preprocessing operations on each surface image in the multi-angle surface image set, the preprocessing operations including gray scale equalization processing, noise suppression processing, and image size normalization processing.

[0040] Before inputting the surface image into the image feature analysis model, preprocessing operations can improve the subsequent processing effect. Gray scale equalization processing can improve the contrast of the image. In actual images, due to factors such as lighting, the gray scale distribution may be concentrated in a certain area, making it difficult to distinguish some details. Gray scale equalization processing redistributes the gray scale values of the image to make the gray scale range of the image more uniform. For example, the number of pixels for each gray scale value in the image can be counted, and then a transformation function can be calculated according to the statistical results to map the original gray scale values to a new gray scale range. After gray scale equalization processing, the overall contrast of the image is improved, and the surface texture and features are clearer.

[0041] The noise suppression processing is used to remove noise interference in the image. During the image acquisition process, the noise of the camera sensor, environmental interference and other factors will cause various noises in the image, such as salt and pepper noise, Gaussian noise, etc. For salt and pepper noise, a median filter method can be used for processing. The median filter is to replace the gray value of each pixel point in the image with the median value of the gray values of the pixels in its neighborhood. For example, for a 3x3 neighborhood, the gray values of the 9 pixels are sorted, and the middle value is taken as the new gray value of the pixel. For Gaussian noise, a Gaussian filter method can be used. Gaussian filtering is to perform convolution operation on the image according to the Gaussian function, and to smooth the image by weighted average of the pixels in the neighborhood to reduce the influence of noise.

[0042] The image size normalization processing is to unify the surface images of different sizes to the same size standard. Due to the different resolutions of different industrial cameras, the sizes of the acquired images also have differences. Through the image size normalization processing, the width and height of all images are adjusted to the same value. A bilinear interpolation method or the like can be used for image scaling. Bilinear interpolation is to calculate the value of a new pixel according to the values of adjacent pixels during scaling. For example, when scaling down the image, the value of the new pixel is determined by weighted average of the values of the surrounding four pixels, so as to ensure the quality and details of the image.

[0043] Step S122: inputting the preprocessed surface image into a first convolutional layer of the image feature analysis model to perform primary feature extraction, to obtain a primary image feature map set, the primary image feature map set containing material surface edge features and local contrast features.

[0044] The first convolutional layer is the initial processing layer of the image feature analysis model, mainly used for extracting primary features of the image. The convolutional layer extracts features by sliding convolution operation of the convolutional kernel on the image. In actual operation, the convolutional kernel is a small matrix, which slides on the image row by row and column by column, and performs convolution calculation with the local region of the image.

[0045] The convolution calculation process is to multiply each element of the convolution kernel with the corresponding element of the local region of the image, and then add all the products to obtain a new pixel value. For example, a 3x3 convolution kernel is convolved with a 3x3 local region of the image, the 9 elements of the convolution kernel are multiplied with the 9 pixel values of the local region of the image respectively, and then the 9 products are added to obtain a new pixel value. By continuously sliding the convolution kernel, the convolution calculation is performed on each local region of the image, and finally a feature map is obtained.

[0046] In this process, the convolution kernel extracts the edge features and local contrast features of the image. The edge features can reflect the contour and boundary information of the material surface, which is very important for identifying the shape and structure of the surface. For example, when the convolution kernel detects a sudden change in pixel value in the image, it indicates that there may be an edge. The local contrast feature can reflect the brightness difference between different regions in the image, which helps to find the subtle changes and defects of the surface. Through the processing of the first convolution layer, a set of primary image feature maps containing these primary features is obtained.

[0047] Step S123: performing multi-scale feature fusion processing on the set of primary image feature maps through the second convolution layer of the image feature analysis model to generate a set of fused scale feature maps, each feature map in the set of fused scale feature maps corresponding to texture information of different scales.

[0048] The main task of the second convolution layer is to perform multi-scale feature fusion processing on the set of primary image feature maps. The feature maps in the set of primary image feature maps may only contain texture information of a certain scale. In order to more comprehensively capture the texture features of the material surface, multi-scale feature fusion is needed.

[0049] In the second convolution layer, different size convolution kernels are used to process the primary image feature maps. Different size convolution kernels can capture texture information of different scales. Large convolution kernels can capture large-scale texture features, such as macro-texture of the material surface, overall pattern, etc. Small convolution kernels can capture more subtle texture details, such as tiny holes, fine lines, etc.

[0050] In the fusion process, first, different size convolution kernels are used to perform convolution operations on the primary image feature maps to obtain a plurality of feature maps of different scales. Then, these feature maps of different scales are fused. The fusion method can be to perform weighted summation of pixel values at corresponding positions. For example, for two feature maps of different scales, the pixel values at corresponding positions are multiplied by different weight coefficients respectively, and then added to obtain a new pixel value. By continuously adjusting the weight coefficients, the fused feature map can better preserve the texture information of different scales. Finally, a set of fused scale feature maps is generated, each of which corresponds to texture information of different scales, thereby more comprehensively reflecting the texture features of the material surface.

[0051] Step S124: calling the feature pyramid network of the image feature analysis model to perform hierarchical feature aggregation processing on the set of fused scale feature maps to obtain a set of multi-level feature descriptors.

[0052] The role of the feature pyramid network is to perform hierarchical feature aggregation on the fusion scale feature map set. Although the feature maps in the fusion scale feature map set contain texture information of different scales, this information is scattered and needs to be further integrated and aggregated.

[0053] The feature pyramid network performs multi-level processing on the fusion scale feature map set. First, the fusion scale feature map set is taken as input, and a series of convolution and pooling operations are performed to construct feature maps of different levels. In each level, the size and resolution of the feature map gradually decrease, while the abstraction level of the feature gradually increases.

[0054] When performing hierarchical feature aggregation, feature maps of different levels are fused and combined. For example, high-level abstract features are fused with low-level detailed features. The size of the low-level feature map can be adjusted to the same size as the high-level feature map through upsampling operation, and then the pixel values at corresponding positions are added or spliced. In this way, the feature information of different levels is integrated to form a multi-level feature descriptor set. Each descriptor in the multi-level feature descriptor set contains feature information of multiple levels, which can more comprehensively and accurately describe the features of the material surface.

[0055] Step S125: obtaining an image feature set of the refractory material sample according to the multi-level feature descriptor set.

[0056] After obtaining the multi-level feature descriptor set, further processing is needed to obtain the image feature set of the refractory material sample. This step includes the following sub-steps:

[0057] Step S1251: performing channel attention weighting processing on the multi-level feature descriptor set to generate a weighted feature channel set, and the weight of the channel attention weighting processing is determined by the feature response intensity of each channel in the multi-level feature descriptor set.

[0058] The purpose of channel attention weighting processing is to highlight important channel features in the multi-level feature descriptor set. In the multi-level feature descriptor set, the feature response intensity of different channels may be different, and some channel features are more important for describing the features of the material surface, while some channel features are relatively secondary.

[0059] First, the feature response intensity of each channel in the multi-level feature descriptor set is calculated. The feature map of each channel can be compressed into a scalar value by performing global average pooling operation on the feature map of each channel, and the scalar value represents the feature response intensity of the channel. For example, for a channel feature map, add all the pixel values and divide by the total number of pixels to get the average pixel value of the channel as the feature response intensity.

[0060] Then, the weights are determined according to the characteristic response intensity of each channel. For channels with high characteristic response intensity, larger weights are assigned; for channels with low characteristic response intensity, smaller weights are assigned. A nonlinear function such as a sigmoid function can be used to map the characteristic response intensity to a suitable weight range, such as [0, 1].

[0061] Finally, the feature maps of each channel are multiplied by the corresponding weights to obtain a set of weighted feature channels. In this way, the features of important channels are highlighted, and the features of secondary channels are suppressed, making subsequent feature analysis more focused on key information.

[0062] Step S1252: input the set of weighted feature channels into a spatial pyramid pooling layer for multi-resolution feature sampling processing to obtain a set of multi-resolution feature vectors.

[0063] The role of the spatial pyramid pooling layer is to perform multi-resolution feature sampling processing on the set of weighted feature channels. The feature maps in the set of weighted feature channels may have different sizes and resolutions, and in order to unify the representation form of the features, multi-resolution feature sampling needs to be performed.

[0064] The spatial pyramid pooling layer divides the set of weighted feature channels into regions of different sizes. For example, the feature maps can be divided into regions of sizes 1x1, 2x2, 4x4, etc. For each region, a pooling operation is performed, such as max pooling or average pooling. Max pooling selects the pixel value with the maximum value in each region as the feature value of that region; average pooling calculates the average value of all pixel values in each region as the feature value of that region.

[0065] By performing the pooling operation on regions of different sizes, feature vectors of different resolutions are obtained. For example, for a region of size 1x1, a feature vector is obtained; for a region of size 2x2, 4 feature vectors are obtained; for a region of size 4x4, 16 feature vectors are obtained. Combining these feature vectors of different resolutions, we obtain a set of multi-resolution feature vectors. The set of multi-resolution feature vectors can more comprehensively reflect the characteristics of the material surface, as it contains feature information at different scales and resolutions.

[0066] Step S1253: perform dimensionality reduction projection processing on the set of multi-resolution feature vectors to obtain a set of low-dimensional feature projections, each projection vector in the set of low-dimensional feature projections corresponding to a local area feature of the refractory material sample.

[0067] The feature vectors in the multi-resolution feature vector set can have a high dimension, which increases the complexity of subsequent processing. In order to reduce the dimension of the features while preserving important feature information, the multi-resolution feature vector set needs to be processed by dimension reduction projection.

[0068] The dimension reduction projection processing can use methods such as principal component analysis (PCA). The basic idea of the PCA method is to find the principal component directions of the data and project the data onto these principal component directions to achieve dimension reduction. First, the covariance matrix of the multi-resolution feature vector set is calculated, which reflects the correlation between the feature vectors. Then, the covariance matrix is decomposed to obtain the eigenvalues and eigenvectors. The eigenvalues represent the importance of each principal component, and the eigenvectors represent the direction of the principal component.

[0069] The eigenvectors corresponding to the top k largest eigenvalues are selected as the projection matrix. Each feature vector in the multi-resolution feature vector set is multiplied by the projection matrix to obtain the projected low-dimensional feature vectors. These low-dimensional feature vectors form a low-dimensional feature projection set. In the low-dimensional feature projection set, each projection vector corresponds to the local region feature of the refractory material sample, which can more concisely represent the features of the material surface, facilitating subsequent analysis and processing.

[0070] Step S1254: constructing the surface texture feature based on the statistical distribution parameters of the low-dimensional feature projection set, the statistical distribution parameters including variance feature, skewness feature and kurtosis feature.

[0071] In order to construct the surface texture feature, the statistical distribution parameters of the low-dimensional feature projection set need to be extracted. The variance feature reflects the dispersion degree of the data in the low-dimensional feature projection set. In calculating the variance feature, first calculate the difference between each vector element in the low-dimensional feature projection set and the mean value of the set, square these differences, and then average all squared differences. The result is the variance feature. The larger the variance, the more dispersed the data, which means that the texture of the material surface can be more complex and diverse, such as the presence of deep and shallow, large and small grooves or protrusions, etc. On the contrary, the smaller the variance, the more concentrated the data, and the surface texture is relatively more regular and smooth.

[0072] The skewness feature is used to describe the asymmetry of the data distribution. When calculating the skewness feature, the cube of the difference between each vector element in the low-dimensional feature projection set and the mean value is calculated first, the sum of these cubed differences is divided by the number of elements, and then divided by the cube root of the variance to obtain the skewness feature value. If the skewness feature value is positive, it indicates that the data distribution is skewed to the right, i.e. there are some large values deviating from the mean value; if the skewness feature value is negative, it indicates that the data distribution is skewed to the left, i.e. there are some small values deviating from the mean value. Through the skewness feature, it can be understood whether the distribution of the surface texture is skewed to one side, for example, there may be a situation that the texture is relatively dense on one side of the material surface, while the texture is relatively sparse on the other side.

[0073] The kurtosis feature reflects the sharpness of the data distribution. When calculating the kurtosis feature, the fourth power of the difference between each vector element in the low-dimensional feature projection set and the mean value is calculated first, the sum of these fourth power differences is divided by the number of elements, then divided by the square of the variance, and finally subtracted by 3 to obtain the kurtosis feature value. The higher the kurtosis, the more concentrated the data distribution is around the peak value, and the surface texture may have more obvious features, such as obvious periodic texture or prominent local features; the lower the kurtosis, the flatter the data distribution, and the surface texture is relatively uniform and has no obvious prominent features. Combining the variance feature, the skewness feature and the kurtosis feature, the surface texture feature reflecting the texture condition of the material surface is constructed.

[0074] Step S1255: performing a pore region segmentation processing on the low-dimensional feature projection set by a clustering algorithm combined with a region growing strategy, generating a pore connected region set according to the feature similarity and spatial adjacency relationship, and generating the pore distribution feature according to the area proportion of the pore connected region set and the shape complexity parameter generated by the principal component analysis.

[0075] When performing the pore region segmentation processing, the low-dimensional feature projection set is first classified using a clustering algorithm. The clustering algorithm will divide the vectors in the low-dimensional feature projection set into different categories according to the similarity between the vectors. For example, vectors with similar features will be classified into the same category, and in the material surface image, these similar features may correspond to pore regions or non-pore regions. A commonly used clustering algorithm is the K-means clustering algorithm, which first randomly selects K center points, then assigns each vector to the category with the closest center point, and then continuously updates the position of the center point until the category division no longer changes.

[0076] Then, the region growing strategy is combined to further process the preliminary classification results according to the feature similarity and spatial adjacency relationship. The region growing strategy starts from one or more seed points and merges the regions similar in feature and adjacent in space to the seed points into the same region. In the material surface image, a vector preliminarily judged as a pore region is taken as a seed point, and adjacent vectors are searched. If the adjacent vectors are similar in feature to the seed point, they are merged into the pore region. The process is repeated until there is no adjacent vector meeting the condition. In this way, a set of pore connected regions is generated.

[0077] For the generated set of pore connected regions, the area ratio is calculated. The area ratio refers to the ratio of the total area of the pore connected regions to the total area of the material surface image. The area ratio can be obtained by counting the number of vectors included in the pore connected regions and comparing it with the total number of vectors in the low-dimensional feature projection set. The area ratio can reflect the distribution range of the pores on the material surface. The larger the area ratio, the more pores on the material surface.

[0078] At the same time, the principal component analysis is used to analyze the shape of the pore connected regions to generate a shape complexity parameter. The principal component analysis finds the main direction and variation mode of the pore connected region data. The complexity of the shape is evaluated by calculating the variance and contribution rate of the pore connected regions in each principal component direction. For example, if the pore connected regions have large variances in multiple principal component directions, the shape is relatively complex, and there may be irregular shapes, branches, or hole nesting; if there are large variances in only a few principal component directions, the shape is relatively regular. The area ratio and the shape complexity parameter are combined to generate a pore distribution feature that can reflect the distribution of pores on the material surface.

[0079] Step S1256: Under the premise of preserving the consistency of gradient directions in the low-dimensional feature projection set, the direction gradient analysis processing is performed using the histogram of oriented gradients algorithm to extract the direction continuity parameter and the length distribution parameter of the micro-crack feature.

[0080] The features of the micro-cracks on the material surface are crucial for judging the quality of the material, and the key parameters can be effectively extracted by the histogram of oriented gradients algorithm. Before analysis, it is necessary to ensure the consistency of the gradient directions in the low-dimensional feature projection set, which is the basis for accurately extracting the micro-crack features. The consistency of the gradient directions ensures that the direction of the micro-cracks can be truly reflected in the subsequent analysis.

[0081] The histogram of oriented gradients algorithm first calculates the gradient of the low-dimensional feature projection set. Gradient calculation is to calculate the gradient value and gradient direction by calculating the rate of change of each vector element in the horizontal and vertical directions. For example, for a vector in the low-dimensional feature projection set, the difference in the horizontal and vertical directions is calculated by comparing the values of its adjacent elements, and then the gradient value and gradient direction are obtained.

[0082] Next, the image is divided into a plurality of small cells, and the distribution of the gradient direction is counted in each cell to generate the histogram of oriented gradients. In the statistical process, the gradient direction is divided into a plurality of intervals, each interval corresponds to a direction range, and the sum of the gradient values in each interval is counted to obtain the histogram value of the interval. By combining the histograms of oriented gradients of all cells, the histogram of oriented gradients of the entire image is obtained.

[0083] The direction continuity parameter of the microscopic crack feature can be extracted from the histogram of oriented gradients. The direction continuity parameter reflects the continuity of the microscopic crack in the direction. If the histogram value of a certain direction interval in the histogram of oriented gradients changes little in the adjacent cells, it means that the microscopic crack has good continuity in that direction; on the contrary, if the change is large, the continuity is poor. The difference between the histogram values of the same direction interval in adjacent cells can be calculated, and the difference values can be statistically analyzed to obtain the direction continuity parameter.

[0084] At the same time, the length distribution parameter of the microscopic crack feature can also be extracted. The length distribution parameter describes the length distribution of the microscopic crack. According to the distribution of the gradient value in the histogram of oriented gradients, combined with the position information of the pixels in the image, the starting point and the ending point of the microscopic crack can be identified, and the length of each microscopic crack can be calculated. Statistical analysis of the lengths of all microscopic cracks can obtain length distribution parameters such as average length, length variance, etc. These parameters can help judge the severity and distribution of the microscopic crack.

[0085] Step S130: Based on the preset defect classification rule library, the image feature set is jointly classified to generate a set of defect types of the refractory sample and corresponding defect distribution parameters.

[0086] After obtaining the image feature set of the refractory sample, it needs to be jointly classified according to the preset defect classification rule library to determine the defect type and corresponding defect distribution parameter of the sample. This step includes the following sub-steps:

[0087] Step S131: Calculate the difference between the statistical distribution parameters in the surface texture feature and the standard texture feature parameters in the defect classification rule library dimension by dimension to generate a surface texture feature deviation vector.

[0088] The statistical distribution parameters in the surface texture feature contain important information such as variance feature, skewness feature and kurtosis feature. The standard texture feature parameters are stored in the defect classification rule library, which represent the ideal state of the normal refractory material surface texture. The variance feature in the surface texture feature is compared with the standard variance feature in the defect classification rule library dimension by dimension, that is, the numerical difference between the corresponding dimensions is calculated. For example, for each dimension of the variance feature, the value of the dimension of the surface texture feature is subtracted from the value of the corresponding dimension of the standard variance feature to obtain a difference value. The same difference calculation operation is performed on the skewness feature and the kurtosis feature.

[0089] The difference values obtained by the calculation are combined to generate the surface texture feature deviation vector. Each element of the surface texture feature deviation vector corresponds to the difference value of a statistical distribution parameter dimension. The vector can reflect the difference between the surface texture feature and the standard texture feature. If the element value in the deviation vector is large, it means that the surface texture deviates greatly from the standard texture in that dimension, and there may be surface defects such as abnormal surface roughness, uneven texture distribution, etc.

[0090] Step S132: The area ratio and shape complexity parameters in the pore distribution feature are respectively converted into standardized proportions by comparing with the pore distribution threshold in the defect classification rule library, to generate a pore abnormality degree vector.

[0091] The pore distribution feature contains area ratio and shape complexity parameters, which can reflect the distribution of material surface pores. The pore distribution threshold is set in the defect classification rule library, which is the standard for judging whether the pore is abnormal.

[0092] For the area ratio, it is converted into a standardized proportion by comparing with the area ratio threshold in the defect classification rule library. First, the difference between the area ratio and the area ratio threshold is calculated, and then the difference is converted into a standardized value according to the set proportion calculation method. For example, if the area ratio exceeds the threshold, the excess proportion is converted into a value representing the degree of abnormality; if the area ratio is lower than the threshold, the corresponding calculation is also performed.

[0093] For the shape complexity parameter, similar standardized proportion conversion is also performed. The shape complexity parameter is compared with the shape complexity threshold in the defect classification rule library, the difference is calculated and converted into a proportion. The converted area ratio abnormal value and shape complexity abnormal value are combined to generate a pore abnormality degree vector. The pore abnormality degree vector can reflect the degree of abnormality of the pore distribution feature relative to the standard pore distribution. The larger the element value in the vector, the more the pore deviates from the normal standard in terms of area ratio or shape complexity, and there may be defects such as internal pore exceeding the standard.

[0094] Step S133: input the direction continuity parameter and the length distribution parameter of the micro-crack feature into a crack evaluation model in the defect classification rule library, and output a crack abnormality degree vector.

[0095] The direction continuity parameter and the length distribution parameter of the micro-crack feature contain important information of the micro-crack. The crack evaluation model in the defect classification rule library is trained by a large amount of data, which can evaluate the abnormality degree of the micro-crack according to the input direction continuity parameter and length distribution parameter.

[0096] The direction continuity parameter and the length distribution parameter are input into the crack evaluation model, which processes and analyzes these parameters. The crack evaluation model internally calculates and judges whether the direction continuity of the micro-crack meets the normal standard according to the direction continuity parameter. If the continuity is poor, it indicates that there may be more serious crack problems. According to the length distribution parameter, it is judged whether the length of the micro-crack is beyond the normal range. If the length is too long or the length distribution is too dispersed, it also indicates that the crack is abnormal.

[0097] The crack evaluation model outputs a crack abnormality degree vector, each element of which represents the abnormality degree of the micro-crack in different aspects. For example, one element may represent the abnormality degree of the direction continuity, and another element may represent the abnormality degree of the length distribution. The crack abnormality degree vector can comprehensively reflect the abnormality degree of the micro-crack relative to the standard crack condition, providing a basis for judging whether there is a micro-structure fracture defect.

[0098] Step S134: perform non-dimensional normalization processing on the surface texture feature deviation vector, the pore abnormality degree vector and the crack abnormality degree vector to generate a unified dimension defect score vector set.

[0099] The surface texture feature deviation vector, the pore abnormality degree vector and the crack abnormality degree vector may have different dimensions and value ranges. In order to facilitate subsequent joint classification processing, they need to be processed by non-dimensional normalization.

[0100] The purpose of non-dimensional normalization processing is to unify the value range of these vectors to a set interval, such as the interval [0, 1]. For the surface texture feature deviation vector, first find the maximum and minimum values of all elements in the surface texture feature deviation vector, then calculate the difference between each element and the minimum value, and then divide the difference by the difference between the maximum value and the minimum value to get the normalized element value. The pore abnormality degree vector and the crack abnormality degree vector are also normalized in the same way.

[0101] By normalization processing, the dimensional differences between different vectors are eliminated, making them numerically comparable. The normalized surface texture feature deviation vector, the porosity anomaly degree vector and the crack anomaly degree vector are combined to generate a unified dimensional defect score vector set. Each defect score vector in the defect score vector set represents the abnormal degree of a defect type in different aspects.

[0102] Step S135: input the defect score vector set into a multi-label classification network for joint defect probability prediction, and output the probability distribution results of each defect type.

[0103] The multi-label classification network is a model that can simultaneously classify and predict the probability of multiple defect types. The unified dimensional defect score vector set is input into the multi-label classification network, and the network will calculate and analyze the input data.

[0104] The multi-label classification network contains multiple levels of neurons and connection weights. After inputting the data, the data will be transmitted layer by layer in the network. Each neuron will perform weighted summation on the input data according to its connection weight, and then perform nonlinear transformation through an activation function. Through continuous calculation and transmission, the network will predict the probability of each defect type.

[0105] The network output is the probability distribution result of each defect type, i.e. the probability value of each defect type. For example, for different defect types such as surface cracking defect, internal porosity defect and microstructure fracture defect, the network will output their respective probability values. These probability values can help determine whether the refractory material sample has a certain defect and the size of the defect possibility.

[0106] Step S136: generate an initial defect type set according to the defect types in the probability distribution result that exceed the preset activation threshold, and generate a density index in the defect distribution parameter based on the extreme components of the corresponding defect score vector.

[0107] According to the probability distribution results of each defect type output by the multi-label classification network, it is necessary to determine which defect types have probability values exceeding the preset activation threshold. The preset activation threshold is a pre-set standard for determining which defect types are actually present.

[0108] When the probability value of a certain defect type exceeds the preset activation threshold, the defect type is added to the initial defect type set. For example, if the probability value of the surface cracking defect exceeds the set threshold, the surface cracking defect will be included in the initial defect type set.

[0109] Meanwhile, the intensity index in the defect distribution parameter is generated based on the extreme components of the corresponding defect score vector. For each defect type included in the initial defect type set, find the extreme component in the corresponding defect score vector. The extreme component can be the maximum value in the defect score vector, which represents the maximum abnormality of the defect type in a certain aspect.

[0110] According to these extreme components, the intensity index is generated by a set calculation method. For example, the extreme components of all defect types included in the initial defect type set can be summed up, and then the sum is converted into the intensity index according to a set proportional relationship. The intensity index can reflect the distribution intensity of defects on the material surface. The larger the index value is, the more intensive the distribution of defects on the material surface is, and the more serious the defect situation is.

[0111] Step S137: When the probability value of the surface cracking defect exceeds the first activation threshold, the gradient change rate component in the surface texture feature deviation vector is extracted, and the spatial distribution entropy value of the pore abnormality degree vector is weighted and fused to generate the surface defect depth parameter.

[0112] When the probability value of the surface cracking defect output by the multi-label classification network exceeds the first activation threshold, it indicates that the material surface is likely to have a surface cracking defect. At this time, it is necessary to further evaluate the depth of the surface defect.

[0113] First, the gradient change rate component is extracted from the surface texture feature deviation vector. The gradient change rate component reflects the change speed of the surface texture in the gradient direction. In the surface texture feature deviation vector, the gradient change rate component is a dimension value related to the change of the texture gradient. By calculating the gradient of the surface texture and analyzing the change of the gradient value, the gradient change rate component is obtained.

[0114] Meanwhile, the spatial distribution entropy value of the pore abnormality degree vector is calculated. The spatial distribution entropy value is used to measure the uniformity of the spatial distribution of pores. When calculating the spatial distribution entropy value, first divide the pore abnormality degree vector according to the spatial position to obtain the pore abnormality degree values in different regions. Then, according to the calculation method of information entropy, the pore abnormality degree values in these regions are calculated to obtain the spatial distribution entropy value. The larger the spatial distribution entropy value is, the more uneven the spatial distribution of pores is.

[0115] The extracted gradient change rate component and the calculated spatial distribution entropy value are fused by weighting. The weighting fusion refers to assigning different weights to the gradient change rate component and the spatial distribution entropy value, and then combining them. The determination of the weight can be adjusted according to experience or through experiments. For example, it is considered that the gradient change rate component has a greater impact on the depth of surface defects, and a larger weight can be assigned to the gradient change rate component. The result obtained by weighting fusion is a surface defect depth parameter, which can comprehensively reflect the depth of surface cracking defects.

[0116] Step S138: When the probability value of the internal pore over-standard defect exceeds the second activation threshold, the area ratio component in the pore abnormality vector is nonlinearly transformed to generate a pore density over-standard rate, and coupled with the directional derivative of the shape complexity component to generate a pore distribution uniformity parameter.

[0117] When the probability value of the internal pore over-standard defect exceeds the second activation threshold, it indicates that there is a high possibility of pore over-standard problem in the material. In order to further evaluate the distribution of pores, the following processing is needed.

[0118] The area ratio component in the pore abnormality vector is nonlinearly transformed to generate a pore density over-standard rate. Nonlinear transformation refers to using a nonlinear function to process the area ratio component. For example, an exponential function or a logarithmic function can be used to calculate the pore density over-standard rate according to the value of the area ratio component. The purpose of nonlinear transformation is to more accurately reflect the degree of pore density over-standard, because the pore density over-standard situation may not be linearly changed.

[0119] At the same time, the directional derivative of the shape complexity component is calculated. The directional derivative reflects the change rate of the shape complexity in a certain direction. In the pore abnormality vector, the shape complexity component describes the complexity of the pore shape. By taking the derivative operation of the shape complexity component, the change rate of it in different directions, i.e. the directional derivative, is obtained.

[0120] The generated pore density over-standard rate and the directional derivative of the shape complexity component are coupled to operate. Coupling operation refers to combining and correlating two parameters to obtain a comprehensive parameter. For example, the pore density over-standard rate and the directional derivative can be multiplied or added, and the specific operation method can be selected according to the actual situation. The result obtained by coupling operation is a pore distribution uniformity parameter, which can comprehensively reflect the density and distribution uniformity of the internal pore.

[0121] Step S139: When the probability value of the microstructure fracture defect exceeds the third activation threshold, perform a spatial convolution operation on the direction continuity parameter and the length distribution parameter in the crack abnormality vector to generate a crack network density parameter.

[0122] When the probability value of the microstructure fracture defect exceeds the third activation threshold, it indicates that there is a high possibility of microstructure fracture problems inside the material. In order to evaluate the network density of micro cracks, the direction continuity parameter and the length distribution parameter in the crack abnormality vector need to be processed.

[0123] Perform a spatial convolution operation on the direction continuity parameter and the length distribution parameter in the crack abnormality vector. Spatial convolution operation is an operation on two parameters in the spatial domain. First, consider the direction continuity parameter and the length distribution parameter as two functions in space, and then perform convolution operation in space.

[0124] In the convolution process, the direction continuity parameter and the length distribution parameter are multiplied and summed in space. Through this operation, the information of direction continuity and length distribution in space is integrated to obtain a new parameter, which can reflect the distribution and mutual connection of micro cracks in space, i.e. the crack network density parameter. The direction continuity parameter reflects the coherence of micro cracks in the direction, and the length distribution parameter reflects the length of the cracks. Through spatial convolution operation, the two are combined, so that the crack network density parameter can more comprehensively describe the severity of microstructure fracture. If the direction continuity is good and the length distribution is long, the crack network density parameter value obtained after convolution operation will be larger, indicating that the network formed by the mutual connection of micro cracks is relatively dense, and the microstructure fracture is relatively serious; on the contrary, the parameter value is smaller, which indicates that the crack network is relatively sparse, and the integrity of the microstructure is relatively good.

[0125] Step S1310: Input the surface defect depth parameter, pore distribution uniformity parameter and crack network density parameter into the three-dimensional space correlation model. If the area proportion of the continuous overlapping area of the defect regions corresponding to the surface defect depth parameter, pore distribution uniformity parameter and crack network density parameter in the spatial projection exceeds the preset overlapping threshold, add the composite defect to the defect type set and mark the correlation priority of the original independent defect type. For example, this step further includes the following sub-steps:

[0126] Step S1310-1: Perform dimensionless standardization conversion processing on the surface defect depth parameter, pore distribution uniformity parameter and crack network density parameter to generate a standardized defect parameter set with a unified dimension range.

[0127] Since the surface defect depth parameter, the pore distribution uniformity parameter and the crack network density parameter can have different dimensions and value ranges, in order to enable reasonable analysis and comparison in the three-dimensional space correlation model, they need to be non-dimensional standardized conversion processing. First, for the surface defect depth parameter, find its maximum value and minimum value, calculate the difference between each surface defect depth parameter value and the minimum value, and then divide the difference by the difference between the maximum value and the minimum value to obtain the standardized surface defect depth parameter value. For the pore distribution uniformity parameter and the crack network density parameter, the same method is also used for processing. Through such standardized conversion, the value range of the three parameters is unified to a set interval, for example, the [0, 1] interval, eliminating the influence of the dimension, generating a standardized defect parameter set with a unified dimension range, so that they have comparability in subsequent spatial correlation analysis.

[0128] Step S1310-2: mapping each parameter in the standardized defect parameter set to a three-dimensional voxel grid according to the spatial coordinates, wherein the surface defect depth parameter is mapped to a first grid layer, the pore distribution uniformity parameter is mapped to a second grid layer, and the crack network density parameter is mapped to a third grid layer.

[0129] The three-dimensional voxel grid is a three-dimensional spatial structure for spatial representation of defect parameters. The surface defect depth parameter, the pore distribution uniformity parameter and the crack network density parameter in the standardized defect parameter set are respectively mapped to different grid layers of the three-dimensional voxel grid according to their corresponding spatial coordinates. Specifically, the surface defect depth parameter is mapped to the first grid layer, each voxel of this layer corresponds to the surface defect depth condition of a specific position on the material surface; the pore distribution uniformity parameter is mapped to the second grid layer, and the voxels of the second grid layer reflect the pore distribution uniformity of the corresponding position inside the material; the crack network density parameter is mapped to the third grid layer, which reflects the crack network density of the position inside the material. Through this mapping method, the originally abstract defect parameters are intuitively represented in three-dimensional space.

[0130] Step S1310-3: performing spatial projection processing on each grid layer to generate a corresponding two-dimensional defect distribution mask, the two-dimensional defect distribution mask containing coordinate range and intensity information of the defect area.

[0131] The defect parameters mapped to each layer of the three-dimensional voxel grid are subjected to spatial projection processing. The purpose is to convert the three-dimensional defect information into a two-dimensional representation, facilitating subsequent analysis. For the first grid layer (corresponding to the surface defect depth parameter), the voxel information of this layer is projected along the direction perpendicular to the layer, compressing the voxel information in three-dimensional space onto a two-dimensional plane. During the projection process, the defect intensity corresponding to each projection position (i.e., the standardized value of the surface defect depth parameter) is recorded, and the coordinate range of the defect region is determined. The same spatial projection processing is also performed for the second grid layer (corresponding to the pore distribution uniformity parameter) and the third grid layer (corresponding to the crack network density parameter). Through such projection operations, two-dimensional defect distribution masks corresponding to the surface defect depth, pore distribution uniformity, and crack network density are generated. These two-dimensional defect distribution masks not only contain the coordinate range of the defect region in the two-dimensional plane but also record the defect intensity information at each position.

[0132] Step S1310-4: Perform a pixel-by-pixel logical AND operation on the two-dimensional defect distribution masks to generate an overlapping region mask and calculate the area ratio of continuous connected regions in the overlapping region mask.

[0133] The generated three two-dimensional defect distribution masks are subjected to a pixel-by-pixel logical AND operation. Logical AND operation means that for the same position pixels in the three masks, only when the three pixels in each mask represent a defect region (i.e., have a certain defect intensity), the pixel at this position in the overlapping region mask is marked as an overlapping region pixel, otherwise it is marked as a non-overlapping region pixel. Through this pixel-by-pixel logical AND operation, the overlapping region mask is obtained, which clearly represents the overlapping of the surface defect depth, pore distribution uniformity, and crack network density three defect regions in the two-dimensional plane. Then, the overlapping region mask is analyzed to find the continuous connected regions. Continuous connected region refers to a region composed of adjacent overlapping region pixels. The total area of these continuous connected regions is calculated and compared with the area of the entire two-dimensional plane to obtain the area ratio of continuous connected regions. This area ratio reflects the degree of overlap of the three defects in space, and the larger the area ratio, the stronger the correlation of the three defects in space.

[0134] Step S1310-5: When the area ratio exceeds the preset overlap threshold, the spatial position corresponding to the overlapping region mask is marked as a composite defect region, and the spatial coordinates of the composite defect region are associated with the position coordinates of each independent defect type in the defect type set for correlation matching.

[0135] The preset overlap threshold is a standard preset for judging whether the overlap of the three defects in space reaches the degree required to be identified as a composite defect. When the area proportion of the calculated continuous connected region exceeds the preset overlap threshold, it indicates that the overlap of the surface defect depth, pore distribution uniformity and crack network density in space is more serious. At this time, the spatial position corresponding to the overlap region mask is marked as a composite defect region. Then, the spatial coordinates of the composite defect region are associated with the position coordinates of each independent defect type (such as surface crack defect, internal pore defect, and microstructure fracture defect) in the defect type set. The association matching refers to judging the coincidence of the composite defect region and each independent defect type region in space, and determining which independent defect types are associated with the composite defect region. For example, by comparing the coordinate ranges, the independent defect type regions overlapping with the composite defect region are found, and these independent defect types and their overlap degree with the composite defect region are recorded.

[0136] Step S1310-6: According to the coverage rate of the composite defect region covering the independent defect type in the matching result, the priority of the independent defect type is arranged in descending order of coverage rate, and the composite defect is added to the top layer of the defect type set.

[0137] According to the association matching result of the composite defect region and each independent defect type region, the coverage rate of the composite defect region to each independent defect type region is calculated. The coverage rate refers to the proportion of the area of the overlapping part of the composite defect region and the independent defect type region to the total area of the independent defect type region. Each independent defect type is arranged in descending order according to the coverage rate from high to low. The independent defect type with high coverage rate indicates that it is more strongly associated with the composite defect. Then, the composite defect is added to the top layer of the defect type set. This is because the composite defect is formed by the spatial overlap of multiple independent defects, and its impact on the material quality is more serious. Being in the top layer of the defect type set can highlight its importance, facilitating subsequent quality evaluation and processing decision.

[0138] Step S140: According to the defect type set and the defect distribution parameter, a quality evaluation parameter of the refractory material sample is generated, and the quality evaluation parameter is compared with a preset quality standard threshold value to output a quality detection result. This step includes the following sub-steps:

[0139] Step S141: The surface defect depth parameter is normalized and compared with the maximum allowed defect depth in the standard sample library to generate a surface integrity index.

[0140] The standard sample library stores standard parameters of normal refractory material samples, including the maximum allowable defect depth. The surface defect depth parameter is normalized by the maximum allowable defect depth in the standard sample library. First, the ratio of the surface defect depth parameter to the maximum allowable defect depth is calculated, and then the ratio is converted into a value in a set interval (such as [0, 1]) according to a set normalization rule to obtain a surface integrity index. The surface integrity index reflects the perfection of the material surface. The closer the index value is to 1, the closer the surface defect depth is to the maximum allowable value, and the worse the surface integrity. The closer the index value is to 0, the smaller the surface defect depth, and the more complete the surface.

[0141] Step S142: input the pore distribution uniformity parameter into a preset pore tolerance function for inverse proportional mapping, and output an internal structure stability coefficient.

[0142] The preset pore tolerance function is a conventional function determined according to a large number of experiments and experience, which describes the relationship between the pore distribution uniformity and the internal structure stability. The pore distribution uniformity parameter is input into the pore tolerance function for inverse proportional mapping. Inverse proportional mapping means that the larger the pore distribution uniformity parameter value, the smaller the internal structure stability coefficient obtained after function mapping. Conversely, the smaller the pore distribution uniformity parameter value, the larger the internal structure stability coefficient. This is because the more uneven the pore distribution, the worse the internal structure stability of the material. By inverse proportional mapping, the pore distribution uniformity parameter is converted into an internal structure stability coefficient that can reflect the internal structure stability of the material.

[0143] Step S143: convert the crack network density parameter into a micro strength attenuation rate by a piecewise linear interpolation method.

[0144] The piecewise linear interpolation method is a method for data conversion. According to the relationship between the crack network density parameter and the micro strength attenuation rate, a plurality of segmentation points and corresponding micro strength attenuation rate values are preset. When the crack network density parameter is obtained, it is judged which segmentation interval it is in, and then the corresponding micro strength attenuation rate is calculated by linear interpolation method according to the two end point values of the segmentation interval and the value of the crack network density parameter. For example, if the crack network density parameter is between the two end point values of a certain segmentation interval, the micro strength attenuation rate corresponding to the crack network density parameter is calculated according to the linear proportional relationship of the micro strength attenuation rate values corresponding to the two end points. The micro strength attenuation rate reflects the degree of decrease in the micro strength of the material due to the existence of the micro crack network. The larger the crack network density, the higher the micro strength attenuation rate.

[0145] Step S144: performing linear superposition to generate a comprehensive quality score after dynamically assigning weights to the surface integrity index, internal structure stability coefficient and microstructure strength attenuation rate, and determining a quality evaluation parameter according to the interval division result of the comprehensive quality score in a preset quality grade mapping table, wherein the weight value of dynamic weight assignment is determined by the probability distribution result of each defect type in the defect type set. This substep further includes the following substeps:

[0146] Step S1441: determining the weight value of dynamic weight assignment according to the probability distribution result of each defect type in the defect type set.

[0147] The probability distribution result of each defect type in the defect type set reflects the possibility of occurrence of each defect type. The dynamic weight values of the surface integrity index, internal structure stability coefficient and microstructure strength attenuation rate are determined according to these probability distribution results. For example, if the probability of surface cracking defects is high, it indicates that the surface integrity has a greater impact on the material quality, and then a larger weight is assigned to the surface integrity index; if the probability of internal pore defects is high, a larger weight is assigned to the internal structure stability coefficient; if the probability of microstructure fracture defects is high, a larger weight is assigned to the microstructure strength attenuation rate. The specific weight assignment can be realized by a set calculation method, for example, after normalizing the probability value of each defect type, the weight of the corresponding parameter is determined according to the normalized probability value.

[0148] Step S1442: performing linear superposition to generate a comprehensive quality score for the surface integrity index, internal structure stability coefficient and microstructure strength attenuation rate.

[0149] The surface integrity index, internal structure stability coefficient and microstructure strength attenuation rate with the determined weight are linearly superimposed. That is, the surface integrity index is multiplied by its corresponding weight, the internal structure stability coefficient is multiplied by its corresponding weight, and the microstructure strength attenuation rate is multiplied by its corresponding weight, and then the three products are added to obtain a comprehensive quality score. The comprehensive quality score comprehensively considers factors such as material surface integrity, internal structure stability and microstructure strength, and can comprehensively reflect the quality condition of the material.

[0150] Step S1443: determining a quality evaluation parameter according to the interval division result of the comprehensive quality score in a preset quality grade mapping table.

[0151] The preset quality grade mapping table pre-divides different comprehensive quality score intervals, and each interval corresponds to a quality grade. The calculated comprehensive quality score is compared with the intervals in the preset quality grade mapping table to determine which interval the comprehensive quality score falls in, so as to determine the corresponding quality evaluation parameter. For example, the preset quality grade mapping table may divide the comprehensive quality score into three intervals, respectively corresponding to a qualified grade, a degraded use grade and a scrap grade. If the comprehensive quality score falls in the interval corresponding to the qualified grade, the quality evaluation parameter is the qualified grade; if it falls in the interval corresponding to the degraded use grade, the quality evaluation parameter is the degraded use grade; and if it falls in the interval corresponding to the scrap grade, the quality evaluation parameter is the scrap grade.

[0152] Step S1444: When the comprehensive quality score falls into the first numerical interval, it is determined to be a qualified grade and a qualified product sorting instruction is triggered; when the comprehensive quality score falls into the second numerical interval, it is determined to be a degraded use grade and a raw material ratio correction amount is generated according to the difference rate of the surface integrity index and the internal structure stability coefficient; when the comprehensive quality score falls into the third numerical interval, it is determined to be a scrap grade and a smelting process adjustment parameter is generated based on the gradient change of the microcosmic strength attenuation rate; the sorting instruction, the raw material ratio correction amount and the smelting process adjustment parameter are synchronized to the production line control system to trigger corresponding sorting and process adjustment operations.

[0153] When the comprehensive quality score falls into the first numerical interval, it indicates that the quality of the material meets the qualified standard and is determined to be a qualified grade. At this time, a qualified product sorting instruction is triggered, and the production line control system controls the mechanical arm to move the refractory material sample to the qualified product storage area, and records the quality data of the current production batch to the database for subsequent quality tracing and analysis.

[0154] When the comprehensive quality score falls into the second numerical interval, it is determined to be a degraded use grade. In order to improve the quality of subsequent products, a raw material ratio correction amount needs to be generated according to the difference rate of the surface integrity index and the internal structure stability coefficient. First, the difference between the surface integrity index and the internal structure stability coefficient is calculated, and then the difference is divided by the average of the two to obtain the difference rate. According to the size and direction of the difference rate, the raw material ratio correction amount is generated through a set rule. For example, if the difference rate indicates that the surface integrity is poor and the internal structure stability is relatively good, the proportion of some raw materials that can improve the surface performance may need to be increased; on the contrary, if the internal structure stability is poor, the proportion of raw materials related to the internal structure needs to be adjusted.

[0155] When the comprehensive quality score falls into the third numerical interval, it is determined to be a scrap grade. In order to avoid the same problem in subsequent production, it is necessary to generate smelting process adjustment parameters based on the gradient change of the micro-strength attenuation rate. The gradient change of the micro-strength attenuation rate reflects the change of the micro-strength attenuation rate with a certain factor (such as smelting temperature, time, etc.). By analyzing the gradient change of the micro-strength attenuation rate, the smelting process parameters that need to be adjusted are determined, such as increasing or decreasing the smelting temperature, extending or shortening the smelting time, etc.

[0156] Finally, the sorting instructions, raw material ratio correction amount and smelting process adjustment parameters are synchronized to the production line control system. The production line control system will trigger corresponding sorting and process adjustment operations according to these instructions and parameters, sort the qualified products to the storage area, process the products for downgraded use accordingly, and adjust the smelting process parameters to improve the quality of subsequent products.

[0157] Step S150: feeding back the quality detection result to the production line control system to trigger sorting operation and process parameter adjustment operation. This step includes the following sub-steps:

[0158] Step S151: when the quality detection result is a qualified grade, control the mechanical arm to move the refractory material sample to the qualified product storage area, and record the quality data of the current production batch to the database.

[0159] When the quality detection result is determined to be a qualified grade, the production line control system sends instructions to the mechanical arm to accurately grasp the refractory material sample and move it to the qualified product storage area. During the moving process, the mechanical arm will operate according to the preset path and action to ensure the safety and accuracy of the sample to the storage area. At the same time, the system will record the quality data of the current production batch to the database, including surface integrity index, internal structure stability coefficient, micro-strength attenuation rate, comprehensive quality score and other information. Recording quality data helps to trace and analyze the production process to find potential quality problems and optimize production process in time.

[0160] Step S152: when the quality detection result is a downgraded use grade, adjust the raw material delivery rate and molding pressure parameters of the production line control system, and control the mechanical arm to move the refractory material sample to the repair processing station.

[0161] When the quality detection result is a downgraded use grade, the production line control system will adjust the raw material delivery rate based on the previously generated raw material ratio correction amount. The adjustment of the raw material delivery rate is to change the input proportion of the raw materials to improve the quality of the products. For example, if the raw material ratio correction amount indicates that the proportion of a certain raw material needs to be increased, the production line control system will increase the delivery rate of that raw material, and correspondingly adjust the delivery rate of other raw materials.

[0162] At the same time, the system will also adjust the molding pressure parameters. Molding pressure has an important influence on the internal structure and surface quality of refractory materials. According to the quality detection results and the adjustment of raw material ratio, the production line control system will appropriately increase or decrease the molding pressure to optimize the performance of the product. The adjusted raw material conveying rate and molding pressure parameters will take effect in the subsequent production process.

[0163] In addition, the production line control system will control the mechanical arm to move the degraded refractory material samples to the repair processing station. At the repair processing station, workers or automated equipment will further process the samples, such as repairing surface defects, improving internal structure, etc., to improve the quality of the samples so that they can meet certain use requirements.

[0164] Step S153: When the quality detection result is a scrap grade, trigger the audible and light alarm device and suspend the current production line, and send the defect distribution parameters and process parameter abnormality records to the remote monitoring terminal.

[0165] When the quality detection result is determined to be a scrap grade, it means that the product quality is seriously out of requirements, and there may be a big production problem. At this time, the production line control system will trigger the audible and light alarm device to emit loud sound and flashing light signals to attract the attention of the operator. At the same time, the system will suspend the operation of the current production line to avoid continuing to produce unqualified products and reduce resource waste and production cost.

[0166] The system will also send the defect distribution parameters and process parameter abnormality records to the remote monitoring terminal. The defect distribution parameters include surface defect depth parameters, pore distribution uniformity parameters, crack network density parameters, etc., and the process parameter abnormality records include abnormal conditions of raw material conveying rate, molding pressure, smelting temperature, etc. The management personnel of the remote monitoring terminal can understand the problems occurring in the production process in a timely manner, analyze the causes of the problems, and take corresponding measures to adjust and improve.

[0167] Step S154: Generate process parameter optimization strategies based on the set of historical quality detection results, including dynamically adjusting the sintering curve, optimizing the mold design parameters, and updating the raw material screening standards. This step further includes the following sub-steps:

[0168] Step S1541: Perform time series analysis processing on the set of historical quality detection results to extract quality fluctuation period characteristics and defect type evolution trends.

[0169] The historical quality detection result set contains quality detection data of multiple production batches, and the time series analysis is performed on these data. Time series analysis refers to analyzing the quality detection results in chronological order to extract key information. When performing time series analysis, first arrange the historical quality detection results in chronological order according to the production batches. Then, by observing and analyzing the changes of quality evaluation parameters (such as comprehensive quality score, surface integrity index, etc.) over time, the quality fluctuation period characteristics are determined. For example, if it is observed that the quality evaluation parameters show a periodic rising and falling trend over a period of time, the length, amplitude, etc. of this period are the quality fluctuation period characteristics.

[0170] For the extraction of defect type evolution trend, the occurrence of different defect types in different production batches needs to be counted and analyzed. Record the probability distribution of various defect types (such as surface crack defects, internal porosity defects, microstructure fracture defects, etc.) in each production batch. Over time, observe whether the probability distribution of these defect types changes, for example, whether the occurrence probability of a certain defect type gradually increases or decreases, or whether the ratio relationship between different defect types changes. Through the analysis of these changes, the evolution trend of defect types can be understood, which helps to predict the possible defect types in future production, so that appropriate preventive measures can be taken in advance.

[0171] Step S1542: Perform correlation analysis on the quality fluctuation period characteristics and the production environment parameter set to determine the influence weight of temperature fluctuation, humidity change and equipment vibration parameters on the quality evaluation parameters.

[0172] The production environment parameter set contains information such as temperature fluctuation, humidity change and equipment vibration parameters. The quality fluctuation period characteristics extracted earlier are correlated and analyzed with these production environment parameters. The purpose of correlation analysis is to find the relationship between production environment parameters and quality evaluation parameters, and determine the influence of each production environment parameter on quality evaluation parameters.

[0173] First, collect the production environment parameter data corresponding to each production batch, including temperature fluctuation (such as the maximum value, minimum value, fluctuation amplitude of temperature), humidity change (such as the average value, change range of humidity) and equipment vibration parameters (such as the frequency, amplitude of vibration). Then, use statistical analysis methods such as correlation analysis to calculate the correlation between each production environment parameter and the quality evaluation parameter. Correlation analysis will get a correlation coefficient, the larger the absolute value of the correlation coefficient, the closer the relationship between the production environment parameter and the quality evaluation parameter.

[0174] According to the results of the correlation analysis, the influence weights of the temperature fluctuation, humidity change and equipment vibration parameters on the quality evaluation parameter are determined. The influence weight reflects the importance of each production environment parameter in influencing the quality evaluation parameter. For example, if the temperature fluctuation has a strong correlation with the quality evaluation parameter, the influence weight of the temperature fluctuation is relatively large; on the contrary, if the humidity change has a weak correlation with the quality evaluation parameter, the influence weight of the humidity change is relatively small. By determining the influence weights, it can be determined which production environment parameters have a greater impact on the quality, so that these parameters can be controlled and adjusted.

[0175] Step S1543: constructing a multi-objective optimization function according to the influence weights, the multi-objective optimization function being used to minimize the defect occurrence rate and maximize the production efficiency.

[0176] After determining the influence weights of the temperature fluctuation, humidity change and equipment vibration parameters on the quality evaluation parameter, a multi-objective optimization function is constructed according to these influence weights. The goal of the multi-objective optimization function is to simultaneously minimize the defect occurrence rate and maximize the production efficiency.

[0177] The defect occurrence rate refers to the proportion of defective products in the production process, and reducing the defect occurrence rate can improve product quality and reduce production cost. The production efficiency reflects the production efficiency of the production line, and improving the production efficiency can increase the yield and improve the economic benefit.

[0178] When constructing the multi-objective optimization function, each production environment parameter is multiplied by its corresponding influence weight, and then combined according to the target requirements of the defect occurrence rate and the production efficiency. For example, the defect occurrence rate can be expressed as a function related to the production environment parameters, and the production efficiency can also be expressed as a function related to the production environment parameters, and then the two functions are combined by weighting to form a comprehensive multi-objective optimization function. In the combination process, the weights of the defect occurrence rate and the production efficiency need to be determined according to the actual situation to balance the relationship between the two.

[0179] Step S1544: solving the multi-objective optimization function by a particle swarm optimization algorithm with constraints, wherein the constraints include the physical compatible range of sintering temperature and cooling rate, to generate an optimal process parameter combination, the optimal process parameter combination including raw material mixing time, sintering furnace temperature control curve and cooling rate parameters.

[0180] The particle swarm optimization algorithm with constraints is an intelligent algorithm for solving optimization problems. The algorithm is used to solve the multi-objective optimization function constructed in the foregoing to find the optimal process parameter combination.

[0181] The basic idea of the particle swarm optimization algorithm is to simulate the group behavior of bird or fish swarms and find the optimal solution by moving particles in the solution space. In solving multi-objective optimization functions, each particle represents a set of process parameter combinations, including raw material mixing time, sintering furnace temperature control curve, and cooling rate parameters, etc. The particles move constantly in the solution space, adjusting their moving direction and speed according to their own historical optimal position and the historical optimal position of the group.

[0182] At the same time, considering the physical limitations in actual production, constraint conditions are set, including the physical compatible range of sintering temperature and cooling rate. During the movement of particles, it is necessary to ensure that the process parameter combination represented by each particle meets these constraint conditions. If the process parameter combination of a certain particle exceeds the physical compatible range of sintering temperature and cooling rate, the particle needs to be adjusted to meet the constraint conditions.

[0183] Through continuous iteration, the particle swarm optimization algorithm gradually converges to the optimal solution, i.e., generates the optimal process parameter combination. This optimal process parameter combination can make the multi-objective optimization function reach the optimal value under the premise of meeting the constraint conditions, that is, minimize the defect occurrence rate and maximize the production efficiency.

[0184] Step S1545: Update the optimal process parameter combination to the parameter configuration file of the production line control system, and after verifying the quality improvement effect through the trial production batch, trigger the quality detection cycle of the formal production batch.

[0185] The optimal process parameter combination obtained by the particle swarm optimization algorithm is updated to the parameter configuration file of the production line control system. The production line control system will adjust the process parameters in the production process according to the updated parameter configuration file, such as mixing raw materials according to the new raw material mixing time, controlling the temperature of the sintering furnace according to the new sintering furnace temperature control curve, and cooling according to the new cooling rate, etc.

[0186] After updating the parameters, a trial production batch is carried out. The trial production batch is to verify whether the new process parameter combination can effectively improve the product quality. During the trial production process, production is carried out according to the normal production process, and the products of the trial production batch are detected for quality. The detection content includes the characteristics of surface integrity, pore distribution, micro-cracks, and comprehensive quality score, etc.

[0187] The quality detection results of the trial production batch are compared with the previous historical quality detection results to evaluate the quality improvement effect. If the quality of the trial production batch is significantly improved, the defect occurrence rate is reduced, and the production efficiency is improved, it means that the new process parameter combination is effective. At this time, the quality detection cycle of the formal production batch is triggered, that is, the product of the formal production batch is detected according to the quality detection method described above to ensure the stability and reliability of the product quality. If the quality of the trial production batch is not significantly improved, the cause needs to be analyzed again, and the parameters of the multi-objective optimization function or the optimization of the process parameters need to be adjusted.

[0188] Figure 2 A schematic diagram of exemplary hardware and software components of the machine learning based refractory quality detection system 100 that can implement the inventive concept provided by some embodiments of the present application is shown. For example, a processor 120 can be used on the machine learning based refractory quality detection system 100 and used to perform the functions in the present application.

[0189] The machine learning based refractory quality detection system 100 can be a general-purpose server or a special-purpose server, both of which can be used to implement the machine learning based refractory quality detection method of the present application. Although only one server is shown in the present application, for the sake of convenience, the functions described in the present application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0190] For example, the machine learning based refractory quality detection system 100 can include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and different forms of storage media 140, such as a disk, a ROM, or a RAM, or any combination thereof. Exemplarily, the machine learning based refractory quality detection system 100 can also include program instructions stored in a ROM, a RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application can be implemented according to these program instructions. The machine learning based refractory quality detection system 100 also includes an I / O interface 150 between the computer and other input / output devices.

[0191] For the convenience of description, only one processor is described in the machine learning based refractory quality detection system 100. However, it should be noted that the machine learning based refractory quality detection system 100 in the present application can also include multiple processors, and therefore the steps performed by one processor described in the present application can also be jointly performed or separately performed by multiple processors. For example, if the processor of the machine learning based refractory quality detection system 100 performs steps A and B, it should be understood that steps A and B can also be jointly performed by two different processors or separately performed in one processor. For example, a first processor performs step A, a second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0192] In addition, the embodiment of the present application also provides a readable storage medium, wherein computer executable instructions are preset, and when the processor executes the computer executable instructions, the machine learning based refractory quality detection method is realized.

[0193] It should be noted that, in order to simplify the description of the present application and to help the understanding of one or more embodiments of the present application, in the foregoing description of the embodiments of the present application, various features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A machine learning-based refractory quality detection method, characterized by, The method comprises: Collecting a multi-angle surface image set of a refractory material sample, the multi-angle surface image set containing material surface images under different lighting conditions; Calling a pre-trained image feature analysis model to perform image feature extraction on the multi-angle surface image set to obtain an image feature set of the refractory material sample, the image feature set including surface texture features, pore distribution features and microscopic crack features; Based on a preset defect classification rule library, performing joint classification processing on the image feature set to generate a defect type set of the refractory material sample and corresponding defect distribution parameters; Generating a quality evaluation parameter of the refractory material sample according to the defect type set and the defect distribution parameters, comparing the quality evaluation parameter with a preset quality standard threshold, and outputting a quality detection result; Feeding back the quality detection result to a production line control system to trigger sorting operation and process parameter adjustment operation; Based on the preset defect classification rule library, the joint classification processing on the image feature set generates a defect type set of the refractory material sample and corresponding defect distribution parameters, comprising: Performing dimension-by-dimension difference calculation on the statistical distribution parameters in the surface texture features and the standard texture feature parameters in the defect classification rule library to generate a surface texture feature deviation vector; Respectively performing standardized proportion conversion on the area proportion and shape complexity parameters in the pore distribution features and the pore distribution threshold in the defect classification rule library to generate a pore abnormality degree vector; Inputting the direction continuity parameter and length distribution parameter of the microscopic crack feature into a crack evaluation model in the defect classification rule library to output a crack abnormality degree vector; Performing dimensionless normalization processing on the surface texture feature deviation vector, the pore abnormality degree vector and the crack abnormality degree vector to generate a unified dimension defect score vector set; Inputting the defect score vector set into a multi-label classification network to perform joint defect probability prediction and output probability distribution results of each defect type; Generating an initial defect type set according to the defect types in the probability distribution results exceeding a preset activation threshold, and generating a density index in the defect distribution parameters based on the extreme components of the corresponding defect score vectors; The initial defect type set is generated according to the defect types in the probability distribution results exceeding a preset activation threshold, comprising: When the probability value of surface cracking defect exceeds a first activation threshold, the gradient change rate component in the surface texture feature deviation vector is extracted, and the spatial distribution entropy value of the pore abnormality degree vector is weighted and fused to generate a surface defect depth parameter; When the probability value of internal pore over-standard defect exceeds a second activation threshold, the area proportion component in the pore abnormality degree vector is subjected to nonlinear transformation to generate a pore density over-standard rate, and the directional derivative of the shape complexity component is coupled to generate a pore distribution uniformity parameter; When the probability value of microscopic structure fracture defect exceeds a third activation threshold, the direction continuity parameter and the length distribution parameter in the crack abnormality degree vector are subjected to spatial convolution operation to generate a crack network density parameter; The surface defect depth parameter, the pore distribution uniformity parameter and the crack network density parameter are input into a three-dimensional space correlation model, and if the area proportion of the continuous overlapping area of the defect regions corresponding to the surface defect depth parameter, the pore distribution uniformity parameter and the crack network density parameter in the space projection exceeds a preset overlapping threshold, a composite defect is added to the defect type set and the correlation priority of the original independent defect type is marked.

2. The machine learning-based refractory quality detection method according to claim 1, characterized by, The pre-trained image feature analysis model is called to perform image feature extraction on the multi-angle surface image set to obtain an image feature set of the refractory material sample, including: The pre-processing operation on each surface image in the multi-angle surface image set includes gray scale equalization processing, noise suppression processing and image size normalization processing; The pre-processed surface image is input into the first convolutional layer of the image feature analysis model to perform primary feature extraction to obtain a primary image feature map set, and the primary image feature map set contains material surface edge features and local contrast features; The second convolutional layer of the image feature analysis model is used to perform multi-scale feature fusion processing on the primary image feature map set to generate a fused scale feature map set, and each feature map in the fused scale feature map set corresponds to texture information of different scales; The feature pyramid network of the image feature analysis model is called to perform hierarchical feature aggregation processing on the fused scale feature map set to obtain a multi-level feature descriptor set; The image feature set of the refractory material sample is obtained according to the multi-level feature descriptor set.

3. The machine learning-based refractory quality detection method according to claim 2, characterized by, The image feature set of the refractory material sample is obtained according to the multi-level feature descriptor set, including: The channel attention weighting processing is performed on the multi-level feature descriptor set to generate a weighted feature channel set, and the weight of the channel attention weighting processing is determined by the feature response intensity of each channel in the multi-level feature descriptor set; The weighted feature channel set is input into the spatial pyramid pooling layer to perform multi-resolution feature sampling processing to obtain a multi-resolution feature vector set; The multi-resolution feature vector set is subjected to dimension reduction projection processing to obtain a low-dimensional feature projection set, and each projection vector in the low-dimensional feature projection set corresponds to the local area feature of the refractory material sample; The surface texture feature is constructed based on the statistical distribution parameters of the low-dimensional feature projection set, and the statistical distribution parameters include variance features, skewness features and kurtosis features; The low-dimensional feature projection set is subjected to pore region segmentation processing by a clustering algorithm combined with a region growing strategy, a pore connected region set is generated according to feature similarity and spatial adjacency relationship, and the pore distribution feature is generated according to the area proportion of the pore connected region set and the shape complexity parameter generated by principal component analysis; On the premise of preserving the gradient direction consistency in the low-dimensional feature projection set, the direction gradient analysis processing is performed by using the histogram of oriented gradients algorithm to extract the direction continuity parameter and the length distribution parameter of the micro-crack feature.

4. The machine learning-based refractory quality detection method according to claim 1, characterized by, The quality evaluation parameter of the refractory sample is generated according to the defect type set and the defect distribution parameter, and the quality evaluation parameter comprises: The surface defect depth parameter is normalized by a maximum allowable defect depth in a standard sample library to generate a surface integrity index; The pore distribution uniformity parameter is input into a preset pore tolerance function for inverse proportional mapping to output an internal structure stability coefficient; The crack network density parameter is converted into a micro-strength attenuation rate by a piecewise linear interpolation method; The surface integrity index, the internal structure stability coefficient and the micro-strength attenuation rate are subjected to dynamic weight distribution and linear superposition to generate a comprehensive quality score, and a quality evaluation parameter is determined according to an interval division result of the comprehensive quality score in a preset quality grade mapping table, wherein the weight value of the dynamic weight distribution is determined by a probability distribution result of each defect type in the defect type set; The quality evaluation parameter is determined according to the interval division result of the comprehensive quality score in the preset quality grade mapping table, and comprises: When the comprehensive quality score falls into a first numerical interval, it is determined as a qualified grade and a qualified product sorting instruction is triggered; When the comprehensive quality score falls into a second numerical interval, it is determined as a degraded use grade and a raw material ratio correction amount is generated according to a difference rate of the surface integrity index and the internal structure stability coefficient; When the comprehensive quality score falls into a third numerical interval, it is determined as a scrap grade and a smelting process adjustment parameter is generated based on a gradient change of the micro-strength attenuation rate; The sorting instruction, the raw material ratio correction amount and the smelting process adjustment parameter are synchronized to a production line control system to trigger corresponding sorting and process adjustment operations.

5. The machine learning-based refractory quality detection method according to claim 1, characterized by, The quality detection result is fed back to the production line control system to trigger sorting operations and process parameter adjustment operations, and comprises: When the quality detection result is a qualified grade, a mechanical arm is controlled to move the refractory sample to a qualified product storage area, and quality data of a current production batch is recorded to a database; When the quality detection result is a degraded use grade, a raw material delivery rate and a molding pressure parameter of the production line control system are adjusted, and the mechanical arm is controlled to move the refractory sample to a repair processing station; When the quality detection result is a scrap grade, an audible-light alarm device is triggered and the current production line is paused, and defect distribution parameters and process parameter abnormalities are recorded and sent to a remote monitoring terminal; A process parameter optimization strategy is generated based on a set of historical quality detection results, and the process parameter optimization strategy comprises dynamically adjusting a sintering curve, optimizing mold design parameters and updating raw material screening standards.

6. The machine learning-based refractory quality detection method according to claim 5, characterized by, The process parameter optimization strategy is generated based on the set of historical quality detection results, and comprises: The set of historical quality detection results is subjected to time series analysis processing to extract quality fluctuation period characteristics and defect type evolution trends; The quality fluctuation period characteristics are associated with a set of production environment parameters for analysis processing to determine influence weights of temperature fluctuations, humidity changes and equipment vibration parameters on quality evaluation parameters; and The quality fluctuation period characteristics are associated with a set of production environment parameters for analysis processing to determine influence weights of temperature fluctuations, humidity changes and equipment vibration parameters on quality evaluation parameters. According to the influence weight, a multi-objective optimization function is constructed, the multi-objective optimization function is used for minimizing defect occurrence rate and maximizing production efficiency; The multi-objective optimization function is solved by a particle swarm optimization algorithm with constraints, wherein the constraints include a physically compatible range of sintering temperature and cooling rate, to generate an optimal process parameter combination, the optimal process parameter combination including raw material mixing time, sintering furnace temperature control curve and cooling rate parameters; The optimal process parameter combination is updated to a parameter configuration file of the production line control system, and after verifying the quality improvement effect by a trial production batch, a quality detection cycle of a formal production batch is triggered.

7. The machine learning-based refractory quality detection method according to claim 1, characterized by, The multi-angle surface image set of the refractory material sample is collected, including: The shooting parameters of the industrial camera array are configured, including light source intensity, polarization angle and focal length, so that the surface images at different angles have consistent brightness distribution and contrast range; The industrial camera array is controlled to collect the material surface images of the refractory material sample according to a preset time synchronization sequence, the material surface images including top surface images, side surface images and bottom surface images, wherein the relative position between the industrial camera array and the refractory material sample is calibrated by a laser range finder; The collected material surface images are subjected to geometric correction processing to generate a multi-angle surface image set; The multi-angle surface image set is subjected to data enhancement processing, the data enhancement processing including small-angle rotation, brightness adjustment and local contrast disturbance, and mirror flipping operation is prohibited to avoid distortion of directional features, so as to expand the multi-angle surface image set.

8. A machine learning based refractory quality detection system, characterized by, A processor and a memory are included, the memory and the processor are connected, the memory is used for storing programs, instructions or codes, and the processor is used for executing the programs, instructions or codes in the memory to realize the machine learning-based refractory material quality detection method in any one of claims 1-7.

Images