Coal quality intelligent analysis and acceptance method based on deep learning

Abstract

The invention discloses a coal quality intelligent analysis and acceptance method based on deep learning, and the method comprises the following steps: collecting coal sample image data through industrial vision equipment, executing the preprocessing operation, and generating a standardized image tensor; inputting the standardized image tensor into an improved DeepLabV3 + network, and generating an enhanced multi-scale semantic feature tensor; inputting the enhanced multi-scale semantic feature tensor into a decoder module to generate a coal segmentation image; performing region identification on the coal segmentation image to generate a structured segmentation region set; generating a coal quality index set based on the structured segmentation region set; comparing the coal quality index set with a coal industry acceptance standard to generate an intelligent acceptance judgment result; and generating a standardized coal quality acceptance report according to an intelligent acceptance judgment result, and connecting the standardized coal quality acceptance report with a business platform. According to the method, deep learning and industry prior fusion is adopted, and automatic analysis and intelligent acceptance of coal quality are realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent detection and analysis technology for coal quality, and in particular to an intelligent analysis and acceptance method for coal quality based on deep learning. Background Technology

[0002] Currently, coal quality testing and acceptance mainly rely on manual sampling and laboratory physicochemical analysis. Common testing procedures include manual collection, weighing, crushing, sieving, and analysis of coal samples. In practice, these traditional methods suffer from long testing cycles, high workload, and heavy reliance on personnel. Furthermore, sample processing and manual interpretation are easily influenced by subjective factors, leading to inconsistencies in the consistency and accuracy of coal quality assessments. While some companies have introduced computer vision-based image processing technology, these systems often only recognize basic features of coal appearance and struggle to accurately separate complex structural areas such as cracks, particles, and impurities. Their ability to automatically extract key coal quality indicators and support intelligent acceptance is limited, resulting in an overall low level of intelligence.

[0003] With the development of deep learning and industrial intelligent inspection technologies, convolutional neural network models have demonstrated strong feature representation and structure recognition capabilities in object segmentation and surface defect detection. However, in coal quality analysis and intelligent acceptance scenarios, most existing methods have failed to be customized for the complex semantic features and multi-structure prior knowledge of the coal industry. The models still have room for improvement in multi-scale structural segmentation, fine-grained semantic enhancement, and automatic index extraction, especially in automatically generating structured coal quality indicators and completing standardized intelligent acceptance processes, lacking efficient and end-to-end closed-loop analysis and business integration capabilities.

[0004] Therefore, how to provide a deep learning-based intelligent analysis and acceptance method for coal quality is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based intelligent analysis and acceptance method for coal quality. This invention fully integrates deep convolutional neural networks with prior knowledge of coal industry structures. Through multi-scale semantic feature extraction and structured segmentation, it achieves automatic identification of cracks, particles, and impurities in coal samples, as well as intelligent extraction of multiple coal quality indicators. Furthermore, it can automatically generate standardized coal quality acceptance reports based on industry acceptance standards. This method boasts advantages such as high detection efficiency, strong judgment accuracy, and outstanding intelligent closed-loop capabilities.

[0006] A method for intelligent analysis and acceptance of coal quality based on deep learning according to an embodiment of the present invention includes the following steps:

[0007] Coal sample image data is acquired using industrial vision equipment and preprocessed to generate standardized image tensors.

[0008] The normalized image tensor is input into the improved DeepLabV3+ network, and feature extraction and multi-scale semantic enhancement are performed on the normalized image tensor in sequence to generate an enhanced multi-scale semantic feature tensor.

[0009] The enhanced multi-scale semantic feature tensor is input into the decoder module, and a coal segmentation image is generated through multi-level upsampling operations;

[0010] Region recognition is performed on coal segmentation images to generate a set of structured segmentation regions;

[0011] Based on the set of structured segmented regions, the attributes of each segmented region are calculated to generate a set of coal quality indicators;

[0012] By comparing the set of coal quality indicators with the coal industry acceptance standards, the quality grade and qualification status of coal samples are automatically determined, and intelligent acceptance judgment results are generated.

[0013] Based on the intelligent acceptance judgment results, a standardized coal quality acceptance report is generated, and the standardized coal quality acceptance report is connected with the business platform to realize closed-loop management of the intelligent coal quality acceptance process.

[0014] Optionally, the preprocessing operation includes size trimming, illumination enhancement, and color channel normalization conversion.

[0015] Optionally, the step of inputting the standardized image tensor into the improved DeepLabV3+ network to generate an enhanced multi-scale semantic feature tensor specifically includes:

[0016] The standardized image tensor is input into the improved DeepLabV3+ network, which includes a backbone feature extraction network, a dynamic hole rate adaptive ASPP module, a coal industry semantic prior enhancement module, and a decoder module.

[0017] The backbone feature extraction network performs multi-layer convolution and nonlinear transformation on the standardized image tensor to extract deep and shallow feature information of the coal image, and outputs the backbone deep feature tensor and the backbone shallow feature tensor.

[0018] The deep feature tensor of the backbone is input into the dynamic dilation rate adaptive ASPP module, and multi-scale dilated convolution processing is performed on the deep feature tensor of the backbone to obtain dilated convolution feature tensors of multiple scales. The dilated convolution feature tensors of multiple scales are concatenated and spliced ​​to form a multi-scale fused feature tensor.

[0019] The multi-scale fused feature tensor is input into the semantic prior enhancement module for the coal industry. Through the channel attention mechanism and the context recalibration mechanism, the enhanced multi-scale semantic feature tensor is output.

[0020] Optionally, the step of inputting the deep feature tensor of the backbone into the dynamic dilation rate adaptive ASPP module to obtain dilated convolutional feature tensors of multiple scales specifically includes:

[0021] The deep feature tensor of the backbone is input to the dynamic hole rate adaptive ASPP module, which includes a dynamic hole rate parameter generation branch, N parallel hole convolution branches and feature concatenation unit.

[0022] The dynamic void ratio parameter generation branch first extracts the global statistical features of the deep feature tensor of the trunk;

[0023] The auxiliary neural network that generates branches by inputting global statistical features into dynamic dilatation rate parameters automatically generates dynamic dilatation rate adjustment values ​​required for multiple parallel dilatation convolution branches, and each dilatation convolution branch is assigned a separate dynamic dilatation rate parameter.

[0024] The deep feature tensor of the backbone is copied into N paths and input into N parallel dilated convolution branches respectively. Each branch is processed by dilated convolution layer, batch normalization layer and activation function layer in sequence, and outputs an activation dilated convolution feature tensor of the corresponding scale.

[0025] The activation function layer performs a nonlinear transformation on each element of the batch normalized dilated convolutional feature tensor. The ReLU activation function is preferred. Each input element is compared with zero and the maximum value is taken to obtain the activated dilated convolutional feature tensor.

[0026] The feature concatenation unit concatenates the activation dilated convolutional feature tensors output from all branches along the channel dimension to obtain a multi-scale fused feature tensor.

[0027] Optionally, the step of inputting the multi-scale fused feature tensor into the coal industry semantic prior enhancement module and outputting the enhanced multi-scale semantic feature tensor specifically includes:

[0028] The multi-scale fusion feature tensor is input into the semantic prior enhancement module for the coal industry. The semantic prior enhancement module for the coal industry includes a channel global pooling branch, a channel attention weight generation branch, a channel weighting branch, a spatial recalibration branch, and a spatial feature weighting branch.

[0029] The channel global pooling branch performs global average pooling and global max pooling on the multi-scale fused feature tensor in each channel dimension to obtain the channel global average value and the channel global maximum value.

[0030] Input the channel global average and channel global maximum values ​​into the channel attention weight generation branch to generate attention weight coefficients;

[0031] The channel-weighted branch uses attention weight coefficients to perform channel-wise scaling on the multi-scale fusion feature tensor to obtain the channel-weighted feature tensor.

[0032] The spatial recalibration branch uses convolution operations on the channel-weighted feature tensor to extract spatial context information, and combines normalization and activation function processing to generate a spatial attention map;

[0033] The spatial feature weighted branch multiplies the spatial attention map and the channel weighted feature tensor point by point at each spatial location, performs weighted adjustments for each spatial location, and combines prior modulation of the structural response of the coal industry to output an enhanced multi-scale semantic feature tensor.

[0034] Optionally, the step of inputting the enhanced multi-scale semantic feature tensor into the decoder module to generate a coal segmentation image specifically includes:

[0035] The enhanced multi-scale semantic feature tensor is input into the decoder module, which includes a multi-level upsampling unit, a feature fusion unit, and a segmentation prediction unit.

[0036] Multi-level upsampling units perform upsampling operations on the enhanced multi-scale semantic feature tensor sequentially, thereby improving the spatial resolution of the feature map step by step.

[0037] After upsampling at each level, the feature fusion unit concatenates or fuses the currently upsampled feature tensor with the corresponding backbone shallow feature tensor in the channel dimension, and then processes it through convolution, normalization and activation functions to obtain a multi-level fused feature tensor.

[0038] The segmentation prediction unit performs convolution and normalization on the last-level fused feature tensor, and generates a coal segmentation image with the same spatial resolution as the input coal sample image through an activation function.

[0039] Optionally, the step of performing region recognition on the coal segmentation image to generate a structured segmentation region set specifically includes:

[0040] Based on the segmentation label of each spatial location in the coal segmentation image, a set of segmentation labels including four categories of coal body, cracks, particles and impurities is set, and each spatial location in the coal segmentation image is traversed one by one.

[0041] For each spatial location, if the segmentation label is a crack category, it is marked as a crack region, and all pixels belonging to the crack category are aggregated to form a crack region mask.

[0042] For each spatial location, if the segmentation label is a particle category, it is marked as a particle region, and all pixels belonging to the particle category are aggregated to form a particle region mask;

[0043] For each spatial location, if the segmentation label is a blemish category, it is marked as a blemish region, and all pixels belonging to the blemish category are aggregated to form a blemish region mask;

[0044] The crack region mask, particle region mask, and impurity region mask are organized and archived to generate a structured segmentation region set.

[0045] Optionally, the step of generating a set of coal quality indicators based on a structured segmentation region set specifically includes:

[0046] For each structural region in the set of structured segmentation regions, count the number of all pixels in the region, and calculate the actual area of ​​the region by combining the image spatial resolution.

[0047] For each structural region in the structured segmentation region set, extract the shape attributes of the region and the color attributes of the pixels within the region;

[0048] Based on the actual area, shape attributes, and color attributes mentioned above, coal quality indicators are automatically extracted for each region to construct a set of coal quality indicators.

[0049] Optionally, the step of comparing the set of coal quality indicators with the coal industry acceptance standards to generate an intelligent acceptance judgment result specifically includes:

[0050] Each indicator in the coal quality indicator set is compared with the superior, qualified, and unqualified ranges set for each indicator in the coal industry acceptance standards.

[0051] When the test values ​​of all coal quality indicators fall within the superior range of the corresponding indicators, the coal sample is judged to be of excellent quality, and the intelligent acceptance judgment result is excellent and qualified.

[0052] When the test values ​​of all coal quality indicators fall within the qualified range of the corresponding indicators, the coal sample is judged to be qualified and the intelligent acceptance judgment result is qualified.

[0053] If the detected value of any coal quality index falls within the non-compliant range of the corresponding index, the coal sample is determined to be non-compliant, and the intelligent acceptance judgment result is non-compliant.

[0054] The judgment results are generated into intelligent acceptance judgment results in the form of structured data.

[0055] Optionally, the step of generating a standardized coal quality acceptance report based on the intelligent acceptance judgment result and connecting the standardized coal quality acceptance report with the business platform specifically includes:

[0056] Based on the quality grade, qualification status and coal quality index set of the coal sample contained in the intelligent acceptance judgment result, a standardized coal quality acceptance report in structured data format is automatically generated.

[0057] The automatic generation process of the standardized coal quality acceptance report specifically includes: pre-establishing a structured data template, automatically filling the corresponding fields of the template with the intelligent acceptance judgment results and the values ​​of the coal quality index set, forming a structured data file, which serves as the standardized coal quality acceptance report;

[0058] Using a structured data interface, the standardized coal quality acceptance report is automatically pushed to the procurement system and warehousing system by calling the data exchange interface of the business platform.

[0059] After receiving the standardized coal quality acceptance report, the procurement system and warehousing system automatically update the quality data information of the coal samples, completing the closed-loop management of the intelligent coal quality analysis and acceptance process.

[0060] The beneficial effects of this invention are:

[0061] This invention combines industrial vision equipment and deep learning models to achieve automatic acquisition and efficient preprocessing of coal sample images, significantly reducing reliance on manual operation. By employing an improved DeepLabV3+ network to extract multi-scale features from standardized image tensors and enhance semantic priors for the coal industry, the accuracy of identifying structural regions such as coal cracks, particles, and impurities is effectively improved, providing a more reliable data foundation for subsequent automatic extraction and intelligent judgment of coal quality indicators.

[0062] After generating the structured segmentation regions, this invention can automatically calculate the area, shape, and color attributes of each region, extract features including crack characteristics, particle distribution, and impurity ratio in batches, and automatically compare them with coal industry acceptance standards. This process, by setting superior, qualified, and unqualified intervals, achieves automatic determination of coal quality grades and qualification status, greatly improving the objectivity and consistency of test results and avoiding subjective differences and efficiency bottlenecks caused by manual interpretation.

[0063] This invention also supports the automatic generation of structured coal quality acceptance reports and can seamlessly integrate with business platforms such as procurement and warehousing systems to form a complete closed-loop management process for intelligent coal quality analysis and acceptance. Through these innovative designs, this invention significantly improves the automation and intelligence level of coal quality inspection and acceptance, and helps promote the digitalization and standardization of related processes in the coal industry. Attached Figure Description

[0064] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0065] Figure 1 This is a flowchart of a deep learning-based intelligent analysis and acceptance method for coal quality proposed in this invention.

[0066] Figure 2 This is a schematic diagram of the improved DeepLabV3+ network structure in the intelligent coal quality analysis and acceptance method based on deep learning proposed in this invention.

[0067] Figure 3 This is a flowchart illustrating the structure of the semantic prior enhancement module for the coal industry in a deep learning-based intelligent analysis and acceptance method for coal quality proposed in this invention. Detailed Implementation

[0068] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0069] refer to Figure 1-3 A deep learning-based intelligent analysis and acceptance method for coal quality includes the following steps:

[0070] Coal sample image data is acquired using industrial vision equipment and preprocessed to generate standardized image tensors.

[0071] The normalized image tensor is input into the improved DeepLabV3+ network, and feature extraction and multi-scale semantic enhancement are performed on the normalized image tensor in sequence to generate an enhanced multi-scale semantic feature tensor.

[0072] The enhanced multi-scale semantic feature tensor is input into the decoder module, and a coal segmentation image is generated through multi-level upsampling operations;

[0073] Region recognition is performed on coal segmentation images to generate a set of structured segmentation regions;

[0074] Based on the set of structured segmented regions, the attributes of each segmented region are calculated to generate a set of coal quality indicators;

[0075] By comparing the set of coal quality indicators with the coal industry acceptance standards, the quality grade and qualification status of coal samples are automatically determined, and intelligent acceptance judgment results are generated.

[0076] Based on the intelligent acceptance judgment results, a standardized coal quality acceptance report is generated, and the standardized coal quality acceptance report is connected with the business platform to realize closed-loop management of the intelligent coal quality acceptance process.

[0077] In this embodiment, the step of acquiring coal sample image data and preprocessing it to generate a standardized image tensor specifically includes:

[0078] An industrial vision acquisition device is set up on the coal sample transport path to acquire images of coal samples after they have passed through belt conveyor or vibration screening. The acquired images are RGB three-channel images.

[0079] Perform a size cropping operation on the RGB image to extract an effective image sub-block containing the complete coal sample area. The size of the cropped area is the fixed input size W×H of the input model, where W and H represent the width and height, respectively.

[0080] The cropped image sub-blocks are subjected to illumination enhancement processing, which includes adaptive histogram equalization and luminance component normalization enhancement operations to improve coal texture details.

[0081] The image after illumination enhancement is subjected to color channel normalization transformation. The R, G, and B values ​​of each pixel are normalized, and the normalized image is converted into a normalized image tensor.

[0082] In this embodiment, the step of inputting the standardized image tensor into the improved DeepLabV3+ network to generate the enhanced multi-scale semantic feature tensor specifically includes:

[0083] The standardized image tensor is input into the improved DeepLabV3+ network, which includes a backbone feature extraction network, a dynamic hole rate adaptive ASPP module, a coal industry semantic prior enhancement module, and a decoder module.

[0084] The backbone feature extraction network performs multi-layer convolution and nonlinear transformation on the standardized image tensor to extract deep and shallow feature information of the coal image, and outputs backbone deep feature tensor and backbone shallow feature tensor. The backbone deep feature tensor includes the deep structural features of the coal image after multi-layer convolution and nonlinear transformation.

[0085] The backbone feature extraction network adopts the ResNet-50 backbone network, which includes an input convolutional layer, a batch normalization layer, an activation function layer, and four sets of residual connection modules. After the normalized image tensor is input into the backbone network, it first obtains a preliminary feature map through convolutional layers and pooling layers. Then, it sequentially extracts deep features and retains shallow features through multiple residual connection modules, and finally outputs the backbone deep feature tensor and the backbone shallow feature tensor respectively.

[0086] The main deep feature tensor is input into the dynamic porosity adaptive ASPP module, which includes multiple parallel porosity convolution branches. The porosity parameters of each branch can be dynamically adjusted according to the coal image features. The main deep feature tensor is subjected to multi-scale porosity convolution processing to obtain porosity convolution feature tensors of multiple scales. The porosity convolution feature tensors of multiple scales are concatenated and spliced ​​to form a multi-scale fused feature tensor.

[0087] The multi-scale fusion feature tensor is input into the semantic prior enhancement module for the coal industry. The semantic prior enhancement module for the coal industry performs channel weighting and spatial feature enhancement on key industry features such as coal cracks, particles and impurities through channel attention mechanism and context recalibration mechanism, and outputs an enhanced multi-scale semantic feature tensor.

[0088] In this embodiment, the step of inputting the deep feature tensor of the backbone into the dynamic dilation rate adaptive ASPP module to obtain dilated convolutional feature tensors of multiple scales specifically includes:

[0089] The deep feature tensor of the backbone is input to the dynamic hole rate adaptive ASPP module, which includes a dynamic hole rate parameter generation branch, N parallel hole convolution branches and feature concatenation unit.

[0090] The dynamic void ratio parameter generation branch first extracts the global statistical features of the deep feature tensor of the backbone. The global statistical features include the channel global average value and the channel global maximum value obtained by performing global average pooling and global max pooling on each channel.

[0091] The auxiliary neural network that generates branches by inputting global statistical features into dynamic dilatation rate parameters has several fully connected layers or lightweight multilayer perceptrons. It automatically generates dynamic dilatation rate adjustment values ​​required for multiple parallel dilatation convolution branches. Each dilatation convolution branch is assigned a separate dynamic dilatation rate parameter, which is the difference between the preset initial dilatation rate and the dynamic dilatation rate adjustment value for each dilatation convolution branch.

[0092] The deep feature tensor of the backbone is copied into N paths and input into N parallel dilated convolution branches. Each branch is processed by a dilated convolution layer, a batch normalization layer and an activation function layer in sequence, and outputs an activated dilated convolution feature tensor of the corresponding scale. The dilated convolution branch includes a dilated convolution layer, a batch normalization layer and an activation function layer.

[0093] The dilated convolutional layer sets the pixel sampling interval within the convolution kernel according to the dynamic dilation rate parameter, and performs dilated convolution operation on the input backbone deep feature tensor. Each time the convolution kernel moves, it only performs multiplication and addition operations with the corresponding pixel points sampled according to the dilation rate in the backbone deep feature tensor, and finally outputs the dilated convolutional feature tensor. The dilated convolutional feature tensor can cover a larger receptive field and fully extract the structural features of coal images at different spatial scales.

[0094] The batch normalization layer performs mean and variance normalization on each channel of the dilated convolution feature tensor in the batch dimension, and performs linear transformation through scaling and offset parameters to obtain the batch normalized dilated convolution feature tensor.

[0095] The activation function layer performs a nonlinear transformation on each element of the batch normalized dilated convolutional feature tensor. The ReLU activation function is preferred. Each input element is compared with zero and the maximum value is taken to obtain the activated dilated convolutional feature tensor.

[0096] The feature concatenation unit concatenates the activation dilated convolutional feature tensors of all branches in the channel dimension. That is, while keeping the spatial size unchanged, the feature tensors are stacked in the channel order to obtain a multi-scale fused feature tensor with the number of channels equal to the sum of the number of channels of all branches.

[0097] In this embodiment, the step of inputting the multi-scale fused feature tensor into the coal industry semantic prior enhancement module and outputting the enhanced multi-scale semantic feature tensor specifically includes:

[0098] The multi-scale fusion feature tensor is input into the semantic prior enhancement module for the coal industry. The semantic prior enhancement module for the coal industry includes a channel global pooling branch, a channel attention weight generation branch, a channel weighting branch, a spatial recalibration branch, and a spatial feature weighting branch.

[0099] The channel global pooling branch performs global average pooling and global max pooling on the multi-scale fused feature tensor in each channel dimension to obtain the channel global average value and the channel global maximum value.

[0100] The channel global average value and channel global maximum value are input into the channel attention weight generation branch. The channel attention weight generation branch assigns higher attention weights to feature channels related to coal cracks, coal particle boundaries, and coal impurity distribution through a fully connected layer and an activation function, generating attention weight coefficients to highlight coal crack feature channels, coal particle boundary feature channels, and coal impurity distribution feature channels.

[0101] The channel-weighted branch uses the aforementioned attention weight coefficients to perform channel-by-channel scaling on the multi-scale fusion feature tensor to obtain a channel-weighted feature tensor, which enhances the response of coal crack regions, coal particle boundary regions, and coal impurity distribution regions.

[0102] The spatial recalibration branch uses convolution operations on the channel-weighted feature tensor to extract spatial context information, and combines normalization and activation function processing to generate a spatial attention map;

[0103] The spatial feature weighted branch multiplies the spatial attention map with the channel-weighted feature tensor point-by-point at each spatial location, performs weighted adjustments on each spatial location of the feature tensor, and combines prior modulation of the coal industry structural response to output an enhanced multi-scale semantic feature tensor:

[0104] F fin (h,w)=F fuse (h,w)·(1+λ crack M crack (h,w)+λ grain M grain (h,w)+λ impurity M impurity (h,w);

[0105] Among them, F fin (h,w) represents the output value of the multi-scale semantic feature tensor at spatial location (h,w), F fuse (h,w) represents the value of the feature tensor after spatial feature weighting branching at spatial location (h,w), λ crack M is the adjustment weight for the spatial response of coal cracks. crack (h,w) represents the value of the coal crack space mask at spatial location (h,w), λ grain M is the adjustment weight for the boundary space response of coal particles. grain (h,w) represents the value of the boundary space mask of the coal particle at the spatial location (h,w), λ impurity M is the adjustment weight for the spatial response of coal impurity distribution. impurity (h,w) represents the value of the spatial mask for the distribution of coal impurities at the spatial location (h,w), where h and w represent the position indices of the height and width dimensions of the image, respectively.

[0106] The initial principle of the above formula originates from the spatial attention mechanism in the field of deep learning. The basic idea is to use a spatial attention map to weight and modulate the feature tensor at each spatial location to highlight important regions. In traditional spatial attention mechanisms, spatial attention weights are usually normalized weights for a single channel. Building on this, and considering the structural characteristics of the coal industry, this invention proposes to decompose the spatial attention weights into a mask weighted sum of cracks, particles, and impurities. By setting corresponding adjustment weights for each, targeted enhancement of coal cracks, coal particle boundaries, and coal impurity distribution areas can be achieved, thereby improving the network's ability to perceive key industry structures.

[0107] The specific derivation process is as follows: First, refer to the general expression for spatial attention weighting, which involves multiplying the spatial attention map and the input feature tensor point by point at each spatial location; second, introduce industry-prior structure masks, and set spatial masks and adjust weights for each type of structure, ultimately forming a weighted expression for multi-structure joint modulation. The above formula enhances structured semantics while ensuring the network's responsiveness to spatially important regions.

[0108] In terms of physical dimensions, all variables in the formula are dimensionless values, λ crack , λ grain , λ impurity M is a dimensionless weighting parameter. crack (h,w), M grain (h,w), M impurity (h,w) are dimensionless spatial mask values, usually binary or probability values. The dimensions on both sides of the formula are consistent, which is consistent with mathematical common sense.

[0109] In this embodiment, the step of inputting the enhanced multi-scale semantic feature tensor into the decoder module to generate a coal segmentation image specifically includes:

[0110] The enhanced multi-scale semantic feature tensor is input into the decoder module, which includes a multi-level upsampling unit, a feature fusion unit, and a segmentation prediction unit.

[0111] The multi-level upsampling unit performs upsampling operations on the enhanced multi-scale semantic feature tensor in turn, gradually improving the spatial resolution of the feature map so that it gradually becomes consistent with the shallow feature tensor of the backbone output by the backbone feature extraction network in terms of spatial size.

[0112] After upsampling at each level, the feature fusion unit concatenates or fuses the currently upsampled feature tensor with the corresponding backbone shallow feature tensor in the channel dimension. After processing with convolution, normalization and activation functions, it achieves full fusion of deep semantic features and shallow spatial detail features, and obtains a multi-level fused feature tensor.

[0113] The segmentation prediction unit performs convolution and normalization on the last-level fused feature tensor, and generates a coal segmentation image with the same spatial resolution as the input coal sample image through an activation function. The coal segmentation image labels the coal body, crack area, particle area and impurity area at each spatial location, providing basic data for the subsequent automatic extraction and analysis of coal quality indicators.

[0114] In this embodiment, the step of performing region recognition on the coal segmentation image to generate a set of structured segmentation regions specifically includes:

[0115] Based on the segmentation label of each spatial location in the coal segmentation image, a set of segmentation labels including four categories of coal body, cracks, particles and impurities is set, and each spatial location in the coal segmentation image is traversed one by one.

[0116] For each spatial location, if the segmentation label is a crack category, then the spatial location is marked as a crack region, and all pixels belonging to the crack category are aggregated to form a crack region mask.

[0117] For each spatial location, if the segmentation label is a particle category, then the spatial location is marked as a particle region, and all pixels belonging to the particle category are aggregated to form a particle region mask.

[0118] For each spatial location, if the segmentation label is a blemish category, the spatial location is marked as a blemish region, and all pixels belonging to the blemish category are aggregated to form a blemish region mask.

[0119] The crack region mask, particle region mask, and impurity region mask are organized and archived to generate a structured segmentation region set.

[0120] In this embodiment, the step of generating a set of coal quality indicators based on a structured segmentation region set specifically includes:

[0121] For each structural region in the set of structured segmentation regions, count the number of all pixels in the region, and calculate the actual area of ​​the region by combining the image spatial resolution.

[0122] For each region, extract the shape attributes of the region and the color attributes of the pixels within the region. The shape attributes include the length, width, perimeter, area, main direction, and shape factor of the region. The color attributes include the average, variance, and extreme values ​​of the red, green, and blue channels.

[0123] Based on the actual area, shape attributes, and color attributes mentioned above, coal quality indicators are automatically extracted for each region to construct a set of coal quality indicators. The set of coal quality indicators includes total crack length, average crack width, structural integrity, number of particles, particle size distribution, average particle size, impurity ratio, and impurity color statistics.

[0124] For the cracked area, calculate the total crack length and average width, and assess the structural integrity of the coal.

[0125] For the granular region, the number of granules is counted, and the granule area, particle size distribution, mean, and variance are analyzed to extract the particle size distribution and particle size statistical indicators of coal.

[0126] For the impurity region, calculate the ratio of the total impurity area to the main coal area, and extract the color attribute of the impurity region.

[0127] In this embodiment, the step of comparing the set of coal quality indicators with the coal industry acceptance standards to determine the quality grade and qualification status of the coal sample and generating an intelligent acceptance judgment result specifically includes:

[0128] The total crack length, average crack width, structural integrity, number of particles, particle size distribution, average particle size, impurity ratio, and impurity color in the coal quality index set are statistically analyzed and compared with the superior, qualified, and unqualified ranges set for each index in the coal industry acceptance standard. The superior range represents the range of values ​​that the index reaches the excellent level, the qualified range represents the range of values ​​that meet the basic acceptance requirements, and the unqualified range represents the range that does not meet the acceptance requirements.

[0129] When the test values ​​of all coal quality indicators fall within the superior range of the corresponding indicators, the coal sample is judged to be of excellent quality, and the intelligent acceptance judgment result is excellent and qualified.

[0130] When the test values ​​of all coal quality indicators fall within the qualified range of the corresponding indicators, the coal sample is judged to be qualified and the intelligent acceptance judgment result is qualified.

[0131] If the detected value of any coal quality index falls within the non-compliant range of the corresponding index, the coal sample is determined to be non-compliant, and the intelligent acceptance judgment result is non-compliant.

[0132] The judgment results are generated into intelligent acceptance judgment results in the form of structured data. The intelligent acceptance judgment results include the quality grade of the coal sample and the corresponding qualification status.

[0133] In this embodiment, the step of generating a standardized coal quality acceptance report based on the intelligent acceptance judgment result and connecting the standardized coal quality acceptance report with the business platform specifically includes:

[0134] Based on the quality grade, qualification status and coal quality index set of the coal sample included in the intelligent acceptance judgment result, a standardized coal quality acceptance report in structured data format is automatically generated.

[0135] The automatic generation process of the standardized coal quality acceptance report specifically includes: pre-establishing a structured data template, which includes the coal sample number, specific values ​​of coal quality indicators and corresponding acceptance intervals, indicator qualification status, and coal sample quality grade; automatically filling the corresponding fields of the template with the intelligent acceptance judgment results and the values ​​of the coal quality indicator set to form a structured data file, which serves as the standardized coal quality acceptance report.

[0136] The standardized coal quality acceptance report includes the coal sample number, the specific values ​​of the coal quality indicators, the acceptance standard range corresponding to each indicator, the qualification status of each indicator, and the final quality grade.

[0137] Using a structured data interface, the standardized coal quality acceptance report is automatically pushed to the procurement system and warehousing system by calling the data exchange interface of the business platform.

[0138] After receiving the standardized coal quality acceptance report, the procurement system and warehousing system automatically update the acceptance status, acceptance level and corresponding quality data of the coal sample, completing the closed-loop management of the intelligent coal quality analysis and acceptance process.

[0139] Example 1:

[0140] To verify the feasibility of this invention in practice, it was applied to the actual coal procurement and acceptance scenario of a large coal enterprise. This enterprise has long faced problems such as low efficiency in coal quality testing, insufficient accuracy of manual testing results, and the inability to effectively integrate the testing process with subsequent business systems. Previously, the enterprise used traditional methods such as manual random sampling and laboratory physicochemical testing, with the quality judgment cycle for each batch of coal samples typically requiring 2-3 working days. Due to the subjectivity of manual testing results, inconsistencies occasionally occurred in the test results of different inspectors for the same sample, affecting the overall efficiency and accuracy of quality control in coal procurement.

[0141] After deploying the deep learning-based intelligent coal quality analysis and acceptance method proposed in this invention, enterprises install industrial vision equipment at the coal conveyor belt to automatically acquire images of batches of coal samples entering the warehouse. Approximately 2000 sample images are acquired daily, with a uniform image resolution of 2048×2048 pixels. The industrial vision equipment operates at a rate of 1500 images per hour. The acquired coal images are preprocessed and then input into the improved DeepLabV3+ network system of this invention. This network automatically identifies and segments crack, particle, and impurity regions in the coal images, achieving an average recognition accuracy of over 96%. The system then extracts indicators based on the segmented regions, automatically calculating the total length of coal cracks, the average particle size of particle regions, and the proportion of impurity regions to automatically generate a set of coal quality indicators.

[0142] Statistical analysis of actual data from 30 consecutive working days of operation revealed that the intelligent acceptance system of this invention automatically analyzed 60,000 coal sample images, with an average processing time of approximately 2.5 seconds per sample image, saving over 80% of the inspection time compared to previous manual inspection methods. Simultaneously, the company conducted manual verification of 3,000 randomly selected samples, finding that the system's accuracy rate was over 98% consistent with the manual verification results, significantly improving the consistency and objectivity of the inspection and eliminating the subjective errors inherent in previous manual inspections. Because coal quality indicators are automatically determined and integrated in real-time with the company's procurement and warehousing business platforms, the timeliness and business continuity of coal quality acceptance data have been greatly improved, increasing the overall efficiency of the coal warehousing process by approximately 75%, and significantly improving the efficiency of the company's procurement decisions and warehousing management.

[0143] To more intuitively demonstrate the beneficial effects of this invention, the table below provides a comparison of actual operational data for coal quality acceptance scenarios before and after implementation:

[0144] Table 1 Comparison of the effects of implementing intelligent coal quality analysis and acceptance methods before and after implementation.

[0145]

[0146] As can be seen from the table above, this invention achieves significant improvements over traditional manual acceptance methods in several core indicators of intelligent coal quality analysis and acceptance. Firstly, in terms of processing efficiency, the intelligent acceptance system can efficiently process approximately 2000 coal samples per day, far exceeding the traditional manual method's capacity of only about 200 samples per day, representing a nearly tenfold increase in sample detection capability. Regarding single-sample processing speed, the system automatically analyzes each image using deep learning in approximately 2.5 seconds, while manual inspection takes an average of about 600 seconds, reducing detection time by 99.6% and greatly improving the timeliness of the inspection process.

[0147] Secondly, regarding the consistency of test results, the test results of this invention achieve a consistency rate of 98% with those of manual verification, far exceeding the 80% of traditional methods, effectively ensuring the objectivity and stability of coal quality assessment. Furthermore, the number of coal quality indicators automatically extracted has been increased from three in traditional manual testing to eight in this invention's system, not only expanding the refined analysis capabilities of coal quality data but also providing more comprehensive data support for enterprise decision-making.

[0148] Furthermore, the intelligent acceptance system can automatically connect the test results with the enterprise's procurement and warehousing business platforms in real time, realizing intelligent, data-driven, and closed-loop management of the coal quality acceptance process, further promoting business collaboration and efficiency improvement. Comprehensive comparative data shows that this invention not only significantly improves the processing capacity and accuracy of coal quality testing and acceptance, but also promotes the development of the coal industry's acceptance process towards efficiency, intelligence, and standardization, possessing broad practical application value and promising prospects for promotion.

[0149] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent analysis and acceptance of coal quality based on deep learning, characterized in that, Includes the following steps: Coal sample image data is acquired using industrial vision equipment and preprocessed to generate standardized image tensors. The normalized image tensor is input into the improved DeepLabV3+ network, and feature extraction and multi-scale semantic enhancement are performed on the normalized image tensor in sequence to generate an enhanced multi-scale semantic feature tensor. The enhanced multi-scale semantic feature tensor is input into the decoder module, and a coal segmentation image is generated through multi-level upsampling operations; Region recognition is performed on coal segmentation images to generate a set of structured segmentation regions; Based on the set of structured segmented regions, the attributes of each segmented region are calculated to generate a set of coal quality indicators; By comparing the set of coal quality indicators with the coal industry acceptance standards, the quality grade and qualification status of coal samples are automatically determined, and intelligent acceptance judgment results are generated. Based on the intelligent acceptance judgment results, a standardized coal quality acceptance report is generated, and the standardized coal quality acceptance report is connected with the business platform to realize closed-loop management of the intelligent coal quality acceptance process.

2. The method for intelligent analysis and acceptance of coal quality based on deep learning according to claim 1, characterized in that, The preprocessing operations include size clipping, illumination enhancement, and color channel normalization.

3. The method for intelligent analysis and acceptance of coal quality based on deep learning according to claim 1, characterized in that, The step of inputting the standardized image tensor into the improved DeepLabV3+ network to generate an enhanced multi-scale semantic feature tensor specifically includes: The standardized image tensor is input into the improved DeepLabV3+ network, which includes a backbone feature extraction network, a dynamic hole rate adaptive ASPP module, a coal industry semantic prior enhancement module, and a decoder module. The backbone feature extraction network performs multi-layer convolution and nonlinear transformation on the standardized image tensor to extract deep and shallow feature information of the coal image, and outputs the backbone deep feature tensor and the backbone shallow feature tensor. The deep feature tensor of the backbone is input into the dynamic dilation rate adaptive ASPP module, and multi-scale dilated convolution processing is performed on the deep feature tensor of the backbone to obtain dilated convolution feature tensors of multiple scales. The dilated convolution feature tensors of multiple scales are concatenated and spliced ​​to form a multi-scale fused feature tensor. The multi-scale fused feature tensor is input into the semantic prior enhancement module for the coal industry. Through the channel attention mechanism and the context recalibration mechanism, the enhanced multi-scale semantic feature tensor is output.

4. The method for intelligent analysis and acceptance of coal quality based on deep learning according to claim 3, characterized in that, The step of inputting the deep feature tensor of the backbone into the dynamic dilation rate adaptive ASPP module to obtain dilated convolutional feature tensors at multiple scales specifically includes: The deep feature tensor of the backbone is input to the dynamic hole rate adaptive ASPP module, which includes a dynamic hole rate parameter generation branch, N parallel hole convolution branches and feature concatenation unit. The dynamic void ratio parameter generation branch first extracts the global statistical features of the deep feature tensor of the trunk; The auxiliary neural network that generates branches by inputting global statistical features into dynamic dilatation rate parameters automatically generates dynamic dilatation rate adjustment values ​​required for multiple parallel dilatation convolution branches, and each dilatation convolution branch is assigned a separate dynamic dilatation rate parameter. The deep feature tensor of the backbone is copied into N paths and input into N parallel dilated convolution branches respectively. Each branch is processed by dilated convolution layer, batch normalization layer and activation function layer in sequence, and outputs an activation dilated convolution feature tensor of the corresponding scale. The activation function layer performs a nonlinear transformation on each element of the batch normalized dilated convolutional feature tensor. The ReLU activation function is preferred. Each input element is compared with zero and the maximum value is taken to obtain the activated dilated convolutional feature tensor. The feature concatenation unit concatenates the activation dilated convolutional feature tensors output from all branches along the channel dimension to obtain a multi-scale fused feature tensor.

5. The method for intelligent analysis and acceptance of coal quality based on deep learning according to claim 3, characterized in that, The process of inputting the multi-scale fused feature tensor into the coal industry semantic prior enhancement module and outputting the enhanced multi-scale semantic feature tensor specifically includes: The multi-scale fusion feature tensor is input into the semantic prior enhancement module for the coal industry. The semantic prior enhancement module for the coal industry includes a channel global pooling branch, a channel attention weight generation branch, a channel weighting branch, a spatial recalibration branch, and a spatial feature weighting branch. The channel global pooling branch performs global average pooling and global max pooling on the multi-scale fused feature tensor in each channel dimension to obtain the channel global average value and the channel global maximum value. Input the channel global average and channel global maximum values ​​into the channel attention weight generation branch to generate attention weight coefficients; The channel-weighted branch uses attention weight coefficients to perform channel-wise scaling on the multi-scale fusion feature tensor to obtain the channel-weighted feature tensor. The spatial recalibration branch uses convolution operations on the channel-weighted feature tensor to extract spatial context information, and combines normalization and activation function processing to generate a spatial attention map; The spatial feature weighted branch multiplies the spatial attention map and the channel weighted feature tensor point by point at each spatial location, performs weighted adjustments for each spatial location, and combines prior modulation of the structural response of the coal industry to output an enhanced multi-scale semantic feature tensor.

6. The method for intelligent analysis and acceptance of coal quality based on deep learning according to claim 1, characterized in that, The step of inputting the enhanced multi-scale semantic feature tensor into the decoder module to generate a coal segmentation image specifically includes: The enhanced multi-scale semantic feature tensor is input into the decoder module, which includes a multi-level upsampling unit, a feature fusion unit, and a segmentation prediction unit. Multi-level upsampling units perform upsampling operations on the enhanced multi-scale semantic feature tensor sequentially, thereby improving the spatial resolution of the feature map step by step. After upsampling at each level, the feature fusion unit concatenates or fuses the currently upsampled feature tensor with the corresponding backbone shallow feature tensor in the channel dimension, and then processes it through convolution, normalization and activation functions to obtain a multi-level fused feature tensor. The segmentation prediction unit performs convolution and normalization on the last-level fused feature tensor, and generates a coal segmentation image with the same spatial resolution as the input coal sample image through an activation function.

7. The method for intelligent analysis and acceptance of coal quality based on deep learning according to claim 1, characterized in that, The process of performing region recognition on the coal segmentation image to generate a set of structured segmented regions specifically includes: Based on the segmentation label of each spatial location in the coal segmentation image, a set of segmentation labels including four categories of coal body, cracks, particles and impurities is set, and each spatial location in the coal segmentation image is traversed one by one. For each spatial location, if the segmentation label is a crack category, it is marked as a crack region, and all pixels belonging to the crack category are aggregated to form a crack region mask. For each spatial location, if the segmentation label is a particle category, it is marked as a particle region, and all pixels belonging to the particle category are aggregated to form a particle region mask; For each spatial location, if the segmentation label is a blemish category, it is marked as a blemish region, and all pixels belonging to the blemish category are aggregated to form a blemish region mask; The crack region mask, particle region mask, and impurity region mask are organized and archived to generate a structured segmentation region set.

8. The intelligent coal quality analysis and acceptance method based on deep learning according to claim 1, characterized in that, The generation of the coal quality index set based on the structured segmentation region set specifically includes: For each structural region in the set of structured segmentation regions, count the number of all pixels in the region, and calculate the actual area of ​​the region by combining the image spatial resolution. For each structural region in the structured segmentation region set, extract the shape attributes of the region and the color attributes of the pixels within the region; Based on the actual area, shape attributes, and color attributes mentioned above, coal quality indicators are automatically extracted for each region to construct a set of coal quality indicators.

9. The method for intelligent analysis and acceptance of coal quality based on deep learning according to claim 1, characterized in that, The process of comparing the set of coal quality indicators with coal industry acceptance standards to generate intelligent acceptance judgment results specifically includes: Each indicator in the coal quality indicator set is compared with the superior, qualified, and unqualified ranges set for each indicator in the coal industry acceptance standards. When the test values ​​of all coal quality indicators fall within the superior range of the corresponding indicators, the coal sample is judged to be of excellent quality, and the intelligent acceptance judgment result is excellent and qualified. When the test values ​​of all coal quality indicators fall within the qualified range of the corresponding indicators, the coal sample is judged to be qualified and the intelligent acceptance judgment result is qualified. If the detected value of any coal quality index falls within the non-compliant range of the corresponding index, the coal sample is determined to be non-compliant, and the intelligent acceptance judgment result is non-compliant. The judgment results are generated into intelligent acceptance judgment results in the form of structured data.

10. The method for intelligent analysis and acceptance of coal quality based on deep learning according to claim 1, characterized in that, The process of generating a standardized coal quality acceptance report based on the intelligent acceptance judgment results and connecting the standardized coal quality acceptance report with the business platform specifically includes: Based on the quality grade, qualification status and coal quality index set of the coal sample contained in the intelligent acceptance judgment result, a standardized coal quality acceptance report in structured data format is automatically generated. The automatic generation process of the standardized coal quality acceptance report specifically includes: pre-establishing a structured data template, automatically filling the corresponding fields of the template with the intelligent acceptance judgment results and the values ​​of the coal quality index set, forming a structured data file, which serves as the standardized coal quality acceptance report; Using a structured data interface, the standardized coal quality acceptance report is automatically pushed to the procurement system and warehousing system by calling the data exchange interface of the business platform. After receiving the standardized coal quality acceptance report, the procurement system and warehousing system automatically update the quality data information of the coal samples, completing the closed-loop management of the intelligent coal quality analysis and acceptance process.

Images