Differentiation data enhancement strategy generation method and system for fine tuning of coal mine visual model

Abstract

The invention relates to the technical field of image processing, and provides a differential data enhancement strategy generation method and system for fine tuning of a coal mine visual model. The method comprises the following steps: firstly, receiving multi-type fine-tuning data of a coal mine, performing classification preprocessing, unifying formats and resolutions of equipment images, personnel images and gas images, and automatically dividing into a training set and a verification set according to labels; then, fine-grained features are extracted based on the preprocessed data, feature vectors are generated, a coal mine exclusive strategy rule base is matched according to the feature vectors, an initial enhancement strategy is dynamically generated, rotation operation is forbidden when the structural integrity of an equipment image is lower than a threshold value, and overturning operation is forbidden when posture distribution of a personnel image does not meet a preset condition; and when the concentration gradient of the gas image is higher than a threshold value, forbidding the rotation or turnover operation, finally carrying out training verification on the initial enhancement strategy by utilizing the verification set, and dynamically adjusting strategy parameters. The method can effectively protect the key features in the coal mine visual data, and significantly improves the fine adjustment precision of the model.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and system for generating differentiated data enhancement strategies for fine-tuning visual models in coal mines. Background Technology

[0002] As a core sector of the energy industry, the intelligent development of coal mines heavily relies on the accurate application of visual models in scenarios such as equipment fault detection, personnel safety monitoring, and methane concentration identification. Existing technologies often employ a collaborative approach of data augmentation, feature extraction, and model optimization to construct a technical system for fine-tuning visual models. This system encompasses the entire process from data preprocessing to strategy generation, including key stages such as image classification, feature quantification analysis, and constraint rule matching. Data augmentation, as a core means of enhancing model generalization capabilities, requires tailored solutions designed to the specific characteristics of coal mine scenarios. These include constraints such as structural integrity of equipment images (e.g., a fixed 20cm diameter for hydraulic cylinders in support structures), posture distribution characteristics of personnel images (e.g., the proportion of bending / standing individuals), and grayscale mapping relationships in methane images (e.g., grayscale values ​​of 100-150 corresponding to 0.5%-1.0% concentration).

[0003] However, existing fixed data augmentation strategies directly employ general image rotation and cropping operations without establishing a dynamic adaptation mechanism for the fine-grained features of coal mine data. This may lead to the destruction of key features or a decline in model performance. Existing technologies typically use fixed templates preset according to data types, but fail to analyze specific features such as the structural integrity of equipment images, the posture distribution of personnel images, and the concentration gradient of gas images. For example, performing a ±3° rotation operation on a hydraulic support image with a 2° tilt will cause structural deformation in 30% of the samples; similarly, performing ±2% grayscale fine-tuning on a gas image, if its grayscale mean is 120 (corresponding to 0.7% concentration), may lead to 8% concentration misjudgment. Although traditional fixed strategies can achieve basic data expansion, the lack of adaptation to coal mine industry constraints (such as equipment size tolerance ±1%, gas concentration error ≤1%) results in insufficient model reliability in critical scenarios and requires manual debugging (taking 2-3 days per cycle), making it difficult to meet the stringent requirements of coal mines for "second-level response and high-precision recognition." Summary of the Invention

[0004] The present invention aims to at least partially solve one of the technical problems in the related art.

[0005] Therefore, a first aspect of the present invention is to propose a method for generating differentiated data augmentation strategies for fine-tuning a coal mine visual model, comprising: S1, receive multi-type fine-tuning images of coal mines and perform classification preprocessing to unify the image format and resolution. The fine-tuning images are divided into equipment images, personnel images and gas images. S2, perform feature extraction on the preprocessed fine-tuned image to obtain feature vectors; when the fine-tuned image is a device image, the extracted features include contour integrity, angle deviation, size proportion, and gray-level distribution features. The contour integrity is determined by the continuity rate of the device edges, and the angle deviation is determined by the angle between the main axis and the horizontal axis of the device image; when the fine-tuned image is a personnel image, the extracted features include posture distribution, key feature proportion, and brightness distribution features; when the fine-tuned image is a gas image, the extracted features include gray-level gradient, gray-level mean, and Gaussian noise variance. S3, based on different types of fine-tuned images and their feature vectors, match the corresponding policy rule base, and dynamically generate an initial enhancement strategy. The initial enhancement strategy is used to determine the operations that can be performed on the image based on each image type and feature. S4 dynamically adjusts the operating parameters in the initial enhancement strategy based on the fault detection accuracy and gas concentration error index until the index meets the standard or the parameter reaches the lower limit of the constraint, and outputs executable enhancement code and strategy report.

[0006] In one embodiment of the present invention, S2 further includes: S21, when extracting features of the device image, the Canny edge detection method is used to detect the image edge and calculate the contour integrity. The angle between the principal axis and the horizontal axis is detected by Hough transform. The angle between the principal axis and the horizontal axis is used to determine the angular deviation of the device. The proportion of the device area in the image is used as the size proportion. The average gray value of the image is used as the gray value distribution feature to generate the device feature vector. S22, when extracting features from personnel images, MTCNN keypoint detection is used to statistically distribute poses, the size ratio of the safety helmet / reflective clothing area in the image is calculated as the key feature ratio, the average brightness of the personnel area is analyzed as the brightness distribution feature, and a personnel feature vector is generated. S23. When extracting features from the gas image, the Sobel operator is used to calculate the gray-level gradient of the gas image, the mean gray level of the entire gas image is calculated, and the Gaussian noise variance of the image is calculated to generate the gas feature vector.

[0007] In one embodiment of the present invention, S3 further includes: S31, for device images, when the structural integrity of the device image is less than 95%, the rotation operation is disabled; when the size of the device in the device image accounts for less than 60%, the cropping operation is disabled. S32, for the personnel image, when the average brightness exceeds the threshold range, adjust the brightness of the personnel image; when the personnel account for less than 50% of the personnel image, disable the rotation operation. S33, for gas images, when the gas image concentration gradient... When doing so, disable rotation or flipping operations and set grayscale fine-tuning parameters according to the grayscale mean range.

[0008] In one embodiment of the present invention, S32 further includes: when the average brightness of the personnel image is less than 40, performing an operation to increase the brightness of the personnel image; and when the average brightness of the personnel image is greater than 90, performing an operation to decrease the brightness of the personnel image.

[0009] In one embodiment of the present invention, S33 further includes: when the average grayscale value of the gas image is between 80 and 120, performing a grayscale fine-tuning operation of ±1%; when the average grayscale value of the gas image is between 120 and 160, performing a grayscale fine-tuning operation of ±0.5%.

[0010] In one embodiment of the present invention, S4 further includes: S41, perform at least ten rapid training sessions and calculate the accuracy of equipment fault detection, gas concentration error, and misjudgment rate of personnel violation identification as training evaluation indicators. S42, if any training evaluation metrics fail to meet the standards, adjust the parameters of the initial enhancement strategy. S43: Optimization is terminated when the indicator fluctuation is less than 1% after at least two consecutive adjustments, and a strategy report containing feature analysis results, rule matching logic, and optimization process is output.

[0011] To achieve the above objectives, a second aspect of the present invention proposes a differential data augmentation strategy generation system for fine-tuning a coal mine visual model, comprising: The data receiving and classification preprocessing module is used to receive and classify various types of fine-tuned images from the coal mine, unify the image format and resolution, and classify the fine-tuned images into equipment images, personnel images and gas images. The fine-grained feature extraction module is used to extract features from the preprocessed fine-tuned image and obtain feature vectors. When the fine-tuned image is an equipment image, the extracted features include structural integrity, angle deviation, size ratio, and gray-level distribution features. When the fine-tuned image is a personnel image, the extracted features include posture distribution, key feature ratio, and brightness distribution features. When the fine-tuned image is a gas image, the extracted features include concentration gradient, gray-level mean, and noise intensity. The strategy matching and generation module is used to match the corresponding strategy rule library according to different types of fine-tuned images and their feature vectors, and dynamically generate an initial enhancement strategy. The initial enhancement strategy is used to determine the operations that can be performed on the image based on each image type and feature. The training, validation, and parameter optimization module is used to dynamically adjust the operating parameters in the initial enhancement strategy based on the fault detection accuracy and gas concentration error index, until the index meets the standard or the parameters reach the lower limit of the constraint, and outputs executable enhancement code and strategy report.

[0012] In one embodiment of the present invention, the strategy matching and generation module is further configured to: when processing device images, disable rotation operation when the structural integrity of the device image is less than 95%, and disable cropping operation when the size of the device in the device image is less than 60%. When processing images of people, the brightness of the images is adjusted when the average brightness exceeds the threshold range, and the rotation operation is disabled when the proportion of people in the images is less than 50%. When processing gas images, the gas image concentration gradient When doing so, disable rotation or flipping operations and set grayscale fine-tuning parameters according to the grayscale mean range.

[0013] In one embodiment of the present invention, the training verification and parameter optimization module is further configured to: Perform at least ten rapid training sessions and calculate the accuracy of equipment fault detection, gas concentration error, and misjudgment rate of personnel violation identification as training evaluation indicators. If any training evaluation metrics fail to meet the standards, the parameters of the initial augmentation strategy will be adjusted. Optimization will terminate when the indicator fluctuates after at least two consecutive adjustments, and a strategy report containing feature analysis results, rule matching logic, and optimization process will be output.

[0014] To achieve the above objectives, a third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described in the first aspect.

[0015] The methods, systems, and storage media of this invention can effectively protect the fine-grained key features in coal mine visual data, significantly improve the accuracy of model fine-tuning, and greatly reduce manual dependence and improve fine-tuning efficiency through automated strategy generation and optimization mechanisms.

[0016] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0017] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1This is a flowchart of a method for generating a differentiated data augmentation strategy for fine-tuning a coal mine visual model according to an embodiment of the present invention; Figure 2 This is a structural diagram of a differential data augmentation strategy generation system for fine-tuning a coal mine visual model according to an embodiment of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0020] The following description, with reference to the accompanying drawings, describes a method, system, and computer-readable storage medium for generating differentiated data augmentation strategies for fine-tuning a coal mine visual model, according to embodiments of the present invention.

[0021] Example 1 Figure 1 This is a flowchart of a method for generating a differential data enhancement strategy for fine-tuning a coal mine visual model, according to an embodiment of the present invention.

[0022] like Figure 1 As shown, the method for generating differentiated data augmentation strategies for fine-tuning coal mine visual models includes the following steps: S1 receives various types of fine-tuned data from coal mines and performs classification preprocessing, unifying the format and resolution of equipment images, personnel images, and gas images, and automatically dividing the training set and validation set according to the labels.

[0023] In one implementation, this invention unifies equipment images, personnel images, and gas images into JPG format with a resolution of 1024×768. Specifically, in some implementations, the steps of "receiving multi-type fine-tuning data from coal mines and performing classification preprocessing, unifying equipment images, personnel images, and gas images into JPG format with a resolution of 1024×768, and automatically dividing the training set and validation set according to labels" are the basic data processing modules of the entire differentiated data augmentation strategy generation system. Its technical implementation leverages the multi-source heterogeneous data characteristics of coal mine visual model fine-tuning, ensuring the accuracy and stability of subsequent augmentation strategy generation.

[0024] At the technical implementation level, this step first receives various types of image data from underground cameras, mining mobile terminals, and historical fault databases via a coal mine fine-tuning data access module. This includes keyframes from RTSP video streams, JPG / PNG images, and BMP images. The system extracts keyframes from the video stream using frame rate control (e.g., 1fps) and automatically classifies the data based on image tags (e.g., "hydraulic support in fully mechanized mining face - leakage," "personnel - not wearing safety helmets," "gas - 0.5% concentration"). After classification, the system converts all three types of images to JPG format and standardizes the resolution to 1024×768 using image scaling algorithms (e.g., bilinear interpolation) to ensure consistent size and format of the input data during model processing, conforming to the input specifications of coal mine visual models (e.g., ResNet-50, YOLOv8).

[0025] At the parameter level, the following standards must be met during image preprocessing: the grayscale range of equipment images should be controlled between 30 and 80 lux, the brightness range of personnel images should be between 40 and 90 lux, the mean grayscale value of gas images should be between 80 and 160, and the noise intensity variance should not exceed 0.02. Images should be uniformly formatted as JPG, with compression quality parameters set to 85–95 to reduce storage and transmission costs while ensuring image quality. The ratio of training set to validation set should be 8:2, and the validation set should include typical samples from each category to support model evaluation during subsequent strategy optimization.

[0026] At the application level, this step is widely used in the fine-tuning process of models for equipment fault detection, personnel violation identification, and gas concentration identification in coal mines. For example, in the oil seepage detection model for hydraulic supports in fully mechanized mining faces, the system can automatically classify oil seepage images as equipment images and use them in a standardized format for training set construction. In the personnel safety monitoring model, the system can identify images tagged "not wearing a safety helmet" and classify them into the personnel image category, ensuring that the enhancement strategy is adapted to key features (such as posture and brightness).

[0027] From a technical perspective, this step effectively improves data processing efficiency and model input consistency through standardized data formats and automatic classification mechanisms, providing a high-quality, structured data foundation for subsequent generation of differentiated enhancement strategies. Simultaneously, the label-based automatic partitioning mechanism reduces manual intervention, ensuring the representativeness and balance of training and validation data, thereby enhancing the model's generalization ability and stability during fine-tuning.

[0028] S2, perform feature extraction on the preprocessed fine-tuned image to obtain feature vectors; when the fine-tuned image is a device image, the extracted features include contour integrity, angle deviation, size ratio, and grayscale distribution features. The contour integrity is determined by the continuity rate of the device edges, and the angle deviation is determined by the angle between the main axis of the device image and the horizontal axis; when the fine-tuned image is a personnel image, the extracted features include posture distribution, key feature ratio, and brightness distribution features; when the fine-tuned image is a gas image, the extracted features include grayscale gradient, grayscale mean, and noise intensity.

[0029] Specifically, this step, "extracting fine-grained features based on preprocessed data and generating feature vectors," is the core implementation link of the "coal mine data fine-grained feature analysis module" in the technical solution of this invention. Its technical principle lies in extracting fine-grained features with industry-specific semantics from three typical fine-tuning data (equipment, personnel, and gas) in coal mines through image processing and feature engineering methods, and quantifying them into feature vectors of a unified dimension, providing data support for the subsequent generation and optimization of differentiated enhancement strategies.

[0030] At the technical implementation level, this step employs multimodal image analysis methods, combining edge detection, pose estimation, and gradient analysis to extract features from different types of coal mine images. For equipment images, the Canny edge detection algorithm is first used to extract the equipment contour, and the contour completeness is calculated. This completeness is reflected by the continuity rate of the equipment contour (e.g., a cylinder edge continuity rate ≥95% is considered complete). Subsequently, the Hough transform is used to detect the angle between the main axis of the equipment and the horizontal axis of the image. This angle is used to evaluate angular deviation (e.g., a deviation of ≤2° for the main axis of the support is considered normal). Simultaneously, image segmentation techniques are used to calculate the size proportion of the equipment region in the image (e.g., a conveyor belt proportion ≥50% is considered valid), and the average grayscale value of the image is statistically analyzed (normal range 30–80) to reflect underground lighting conditions. For personnel images, the MTCNN keypoint detection algorithm is used to extract pose distribution (e.g., the proportion of standing, bending, and walking), and a target detection model is used to calculate the proportion of key features of safety helmets and reflective clothing in the human body region (e.g., a safety helmet proportion ≥5% is considered valid), while the average brightness of the personnel region is statistically analyzed (normal range 40–90 lux). For gas images, the Sobel operator is used to calculate the gray-level gradient (gradient ≥ 5 indicates effective concentration distribution), the gray-level mean of the entire image is calculated (corresponding to the gray-level range of 80–160 for 0.1%–1.5% concentration), and the image noise intensity is calculated (Gaussian noise variance ≤ 0.02 indicates low noise).

[0031] Regarding parameter settings, the feature extraction process must meet coal mine industry standards and safety requirements. For example, the structural integrity threshold for equipment images is set to ≥95% to avoid misjudgments due to edge breaks; the proportion of safety helmets in personnel images must be ≥5% to ensure that key features are not cropped or occluded during enhancement; and the grayscale gradient of gas images must be ≥5 to preserve the recognizability of concentration distribution. All features are ultimately encoded into 128-dimensional feature vectors, with 128 dimensions each for equipment, personnel, and gas categories, facilitating feature-rule mapping in subsequent strategy matching and optimization modules.

[0032] This step plays a crucial role in the overall technical solution. On one hand, it provides fine-grained feature basis for the generation of differentiated enhancement strategies, ensuring that the enhancement operation does not damage key recognition features. On the other hand, its output feature vector is matched with the constraints in the rule base, thereby achieving automatic adaptation and optimization of enhancement parameters. Through this step, the present invention effectively solves the problems of "high risk of feature destruction" and "poor adaptability" in the prior art, laying a data foundation for high-precision and high-efficiency fine-tuning of coal mine visual models.

[0033] Furthermore, S2 includes: S21, when extracting features from the device image, the Canny edge detection method is used to detect image edges and calculate contour integrity. The angle between the principal axis and the horizontal axis is detected by Hough transform, and the angular deviation of the device is determined by the angle between the principal axis and the horizontal axis. The proportion of the device area in the image is used as the size proportion, and the average gray value of the image is used as the gray value distribution feature to generate a 128-dimensional device feature vector.

[0034] Specifically, in the feature extraction process of coal mine equipment images, performing Canny edge detection and calculating contour integrity is a key step in realizing the quantitative analysis of equipment structural features. Canny edge detection is a multi-stage edge detection algorithm whose core principles include Gaussian filtering for noise reduction, gradient magnitude and direction calculation, non-maximum suppression, and hysteresis thresholding. In this invention, this algorithm is used to extract the structural edges of equipment (such as hydraulic supports in fully mechanized mining faces, conveyor belts, etc.) to assess the integrity of their contours.

[0035] In practice, the input image is first subjected to Gaussian filtering to reduce noise interference with edge detection. Typically, a Gaussian filter is used. A Gaussian kernel is used for smoothing. Then, the gradient magnitude and direction of the image are calculated, and the Sobel operator is used for horizontal and vertical gradient estimation, resulting in gradient magnitude and direction maps. The non-maximum suppression stage preserves local maxima along the gradient direction and suppresses non-edge points, ensuring single-pixel width for edges. Finally, high and low thresholds are set... and Perform hysteresis connections on the edges.

[0036] The contour integrity calculation is based on the edge map output by Canny, statistically analyzing the edge continuity rate of key equipment structures (such as hydraulic cylinders and main support beams). For example, for an image of a hydraulic support in a fully mechanized mining face, if the integrity of its hydraulic cylinder structure is... If the value is within the threshold, the structure is considered intact and subsequent rotation enhancement operations are allowed; if it is below the threshold, rotation is disabled to avoid damage to structural features.

[0037] Specifically, first, the grayscale histogram of the image is obtained. Then, the image gradient components are calculated, and the gradient intensity and gradient direction of each pixel are calculated based on the image gradient components. A high threshold and a low threshold are set, and all pixels with gradient intensities greater than the high threshold are considered strong edge pixels, thus obtaining a contour image. Subsequently, the edge continuity is calculated using the edge closure determination method disclosed in the invention patent with publication number CN 115830378A. If the contour is determined not to be a closed shape, the straight-line distance between adjacent breakpoints is calculated based on the pixel coordinates, and the overall length of the contour is estimated (the pixel coordinates are obtained when calculating edge closure, which is described in detail in patent CN 115830378 A and will not be repeated here). The structural integrity threshold is determined using the ratio of the overall contour length between adjacent breakpoints to the straight-line distance. When this ratio is greater than 95%, the structure is considered complete.

[0038] Furthermore, the angle between the device's principal axis and the image's horizontal axis is detected using the Hough Transform. This transformation maps straight lines in the image to parameter space. The direction of the main axis is determined through an accumulator voting mechanism. In this invention, the main axis angle deviation... If the value exceeds the threshold, it is considered normal. If it exceeds the threshold, it indicates that the image is tilted or the shooting angle is abnormal, and it needs to be restricted in the enhancement strategy.

[0039] In addition, the proportion of the image occupied by the device area and the mean grayscale value are used to assess the salience of targets in the image and lighting conditions. Device Area Proportion For an image to be valid, the grayscale mean must be within [value missing]. Within a certain range, to ensure that the image quality meets the requirements for subsequent model training.

[0040] The aforementioned features are ultimately quantized into a 128-dimensional device feature vector, which serves as the input for generating differentiated data augmentation strategies. This step extracts multi-dimensional features such as structural integrity, angular deviation, size proportion, and grayscale distribution, providing precise feature-driven features for the automated generation and optimization of augmentation strategies. This effectively avoids the destruction of key features by traditional fixed strategies and improves the accuracy and stability of model fine-tuning.

[0041] S22, when extracting features from personnel images, MTCNN keypoint detection is used to statistically distribute poses, the size ratio of the safety helmet / reflective clothing area in the image is calculated as the key feature ratio, the average brightness of the personnel area is analyzed as the brightness distribution feature, and a personnel feature vector is generated.

[0042] Specifically, in the feature analysis of coal mine personnel images, using MTCNN (Multi-task Cascaded Convolutional Neural Networks) for keypoint detection is a fundamental step in realizing personnel posture distribution statistics, calculation of the proportion of safety helmet / reflective clothing areas, and brightness mean analysis. This step uses a deep learning model to locate key points on the personnel's face and body, thereby extracting fine-grained features related to personnel posture, equipment status, and image quality, and ultimately generating a 128-dimensional personnel image feature vector to drive subsequent differentiated data augmentation strategies.

[0043] In its implementation, the MTCNN model first performs face detection and keypoint localization on the input RGB image, outputting five facial keypoints (eyes, nose tip, and corners of the mouth) and auxiliary information for human pose estimation. By statistically analyzing the spatial distribution of these keypoints, the pose distribution of the person in the image can be calculated, such as the proportion of standing, bending, or walking postures. This analysis is based on the relative angles and distances between keypoints, combined with a preset pose classification threshold (e.g., the angle between the shoulder and waist lines is less than a certain value). (It is determined to be bending over), thus realizing the automatic recognition and quantification of posture.

[0044] Furthermore, the system extracts the coverage areas of the safety helmet and reflective clothing using a key-point guided region segmentation method. Specifically, it constructs mask areas for the head and torso using key-point coordinates, and performs region recognition by combining color space (such as HSV or YUV) and semantic segmentation models (such as U-Net), calculating the pixel percentage of the safety helmet and reflective clothing within the personnel area. This percentage must meet the minimum requirements for equipment wearing in coal mine safety standards; for example, the safety helmet area percentage should be... To ensure the effectiveness of identification.

[0045] In addition, the system also calculates the average brightness of the personnel area to assess whether the image lighting conditions meet the acquisition standards for underground safety monitoring. The average brightness is calculated based on the brightness channels (such as the Y channel) of the RGB image, combined with a weighted average using a mask of key point areas to ensure that only the area visible to personnel is evaluated. The average brightness needs to be controlled within a certain range. Within a certain range, to avoid blurring or distortion of equipment features due to excessive darkness or brightness.

[0046] Ultimately, the three types of features (pose distribution, equipment area proportion, and average brightness) were quantized and concatenated into a 128-dimensional feature vector, which served as the input for generating and optimizing differentiated data augmentation strategies. The construction of this feature vector ensures that the augmentation strategy can accurately adapt to the specific needs of coal mine personnel images, avoiding model performance degradation due to pose distortion, equipment feature loss, or abnormal brightness, thereby improving the accuracy and stability of personnel violation identification.

[0047] S23. When extracting features from the gas image, the Sobel operator is used to calculate the gray-level gradient of the gas image, the mean gray level of the entire gas image is calculated, and the Gaussian noise variance of the image is calculated to generate the gas feature vector.

[0048] Specifically, this invention utilizes the method disclosed in invention patent publication number CN106447597A to calculate the image grayscale gradient, specifically: The gradient of each point in the grayscale image is calculated using a 3×3 Sobel filter. The formula is as follows: X direction: g x (x,y)=-f2(x-1,y-1)-2·f2(x-1,y)-f2(x-1,y+1)+f2(x+1,y-1)+2·f2(x+1,y)+f2(x+1,y+1); Y direction: g y (x,y)=f2(x-1,y-1)+2·f2(x,y-1)+f2(x+1,y-1) f2(x-1,y+1)-2·f2(x,y+1)-f2(x+1,y+1); Output the gradient magnitude at each point: g(x,y)=|g x (x,y)+g y (x,y)|; In the formula, x is the horizontal coordinate of the pixel in the image, y is the vertical coordinate of the pixel in the image, and g x (x,y) represents the gradient of this pixel along the x-direction, g y (x,y) represents the gradient of the pixel along the y-direction.

[0049] S3: Based on feature vector matching and a coal mine-specific strategy rule base, dynamically generate initial enhancement strategies. For equipment images, disable rotation when structural integrity is below a threshold; for personnel images, disable flipping when posture distribution does not meet preset conditions; for gas images, disable rotation / flipping when concentration gradient is above a threshold. Specifically, this step, "dynamically generating an initial enhancement strategy based on the feature vector matching coal mine-specific strategy rule library," is one of the core components of the differentiated data enhancement strategy generation method of this invention. Its technology achieves intelligent matching of fine-grained feature analysis and industry-specific rules based on coal mine fine-tuning data, aiming to achieve adaptive generation of enhancement strategies, avoid damage to key features, and improve the accuracy and stability of model fine-tuning.

[0050] At the technical implementation level, this step first relies on the 128-dimensional feature vector extracted by the preceding module. This vector covers the structural integrity, angular deviation, size proportion, and grayscale distribution of equipment images; the posture distribution, key feature proportion, and brightness distribution of personnel images; and the concentration gradient, grayscale mean, and Gaussian noise variance of gas images. The system matches the feature vector with a pre-set coal mine-specific strategy rule base, which stores enhanced operation constraints for different data types and feature combinations. For example, when the structural integrity of an equipment image is below 95%, the system will automatically disable rotation operations; when the proportion of bending over in the posture distribution of personnel images exceeds 50%, the system will disable flipping operations; and when the concentration gradient of a gas image is higher than 5, the system will simultaneously disable rotation and flipping operations to prevent the concentration distribution characteristics from being destroyed.

[0051] At the application level, this step is widely used in the fine-tuning of coal mine visual models, such as fault detection of hydraulic supports in fully mechanized mining faces, identification of personnel violations, and identification of gas concentration. Through dynamic strategy generation, the system can automatically adapt to differences in equipment models, personnel posture distribution, and gas concentration visualization methods in different mines, generating enhancement strategies that meet industry constraints without manual intervention.

[0052] In terms of technical effectiveness, this step effectively solves the problem of existing fixed enhancement strategies damaging key features of coal mines, significantly improving the accuracy and robustness of model fine-tuning. For example, the fault detection accuracy of equipment images can be stabilized at over 95%, the gas concentration identification error is controlled within 1%, and the misjudgment rate of personnel violation identification is reduced to below 1%. Simultaneously, this step provides an initial strategy foundation for subsequent validation set-driven optimization and is a crucial starting point for achieving fully automated enhancement strategy generation.

[0053] Furthermore, S3 includes: S31, for device images, when the structural integrity of the device image is less than 95%, the rotation operation is disabled; when the size of the device in the device image accounts for less than 60%, the cropping operation is disabled.

[0054] Specifically, when the structural integrity of the device image is below 95%, the system will automatically disable the rotation operation and adjust the rotation based on the detected original angular deviation. Dynamic calculation of rotation angle This step is crucial to ensure that the critical structural features of the equipment are not damaged during image enhancement. It is one of the core control logics of the differentiated data enhancement strategy generation and optimization module in this invention, aiming to improve the accuracy and stability of image fine-tuning for coal mine equipment.

[0055] At the technical implementation level, structural integrity is measured by using image edge detection algorithms (such as the Canny operator) to calculate the continuity and integrity of the equipment outline, typically quantified by the missing or broken rate of the main structural edges. When the integrity rate is below 95%, it indicates significant missing or distorted structural features in the image. Continuing to perform rotation operations under these conditions would further disrupt the geometric relationships of critical components such as supports and conveyor belts, thus affecting the model's accurate identification of the equipment's state. Therefore, the system automatically disables rotation operations under this condition to prevent the introduction of additional structural noise due to the enhancement operation.

[0056] Furthermore, when the structural integrity meets the standard (≥95%), the system will determine the original angular deviation based on the detected deviation. The rotation angle is dynamically adjusted. This angular deviation is typically detected by the Hough Transform, which measures the angle between the device's principal axis and the horizontal axis in the image, with an accuracy of ±0.1. ° Rotation angle The calculation formula is: Its design aims to preserve image diversity while avoiding device structural distortion due to rotation. For example, if an image is detected to contain... If the tilt is such that the rotation angle is set to... To align with the main axis of the image, the model's ability to recognize features such as bracket tilt and cylinder position is improved.

[0057] In application scenarios, this step is mainly used for image enhancement of equipment with regular structures such as hydraulic supports and conveyor belts in fully mechanized mining faces. In the low-light, multi-angle shooting environment underground, equipment images often exhibit slight tilt or incomplete edges. Through this dynamic rotation control mechanism, the system can intelligently adjust enhancement parameters according to the actual image state, avoiding feature destruction caused by fixed rotation strategies, thereby improving the robustness and accuracy of the fine-tuned model in equipment fault detection.

[0058] The technical benefits of this step are reflected in a significant reduction in structural noise introduced by the enhancement operation, improving image quality and model training efficiency. Experiments show that after adopting this mechanism, the accuracy of the equipment fault detection model can be improved by 5%-7%, and it has good generalization ability across different mines and equipment models, providing a key guarantee for the automation and high-precision fine-tuning of coal mine visual models.

[0059] S32, for people images, adjust the brightness of the people image when the average brightness exceeds the threshold range, and disable the rotation operation when the proportion of people in the people image is less than 50%.

[0060] Specifically, in some implementations, when people occupy less than 50% of the area in a person image, the system will automatically disable the rotation operation.

[0061] S33, for gas images, when the gas image concentration gradient... When doing so, disable rotation or flipping operations and set grayscale fine-tuning parameters according to the grayscale mean range.

[0062] Specifically, in some implementations, when the grayscale concentration gradient of the gas image... At this time, the system will automatically disable image rotation and flipping operations, and set grayscale fine-tuning parameters according to the range of the image's average grayscale value. This step is technically based on the image's local gradient features and global grayscale statistical characteristics, aiming to avoid disrupting the gas concentration distribution and thus ensuring the model's recognition accuracy and stability during the fine-tuning process.

[0063] From a technical implementation perspective, this step first uses the Sobel operator to perform edge detection on the image and calculates the local gradient magnitude of the grayscale values ​​in the image. If the mean gradient magnitude is greater than or equal to 5, it indicates the existence of a significant concentration change region in the image, with clear directional features. In this case, performing rotation or flipping operations would distort the directional information of the concentration distribution, thus affecting the model's perception and recognition of the concentration distribution. Therefore, the system automatically filters out these operations that might damage key features, ensuring that the physical meaning of the concentration gradient in the image is not interfered with.

[0064] Furthermore, the system dynamically adjusts parameters based on the image's average grayscale value. The formula for calculating the average grayscale value is: ,in Indicates the image at position grayscale value at that location This refers to the image size. If the average grayscale value is within the range of 80–120, then set the grayscale fine-tuning parameter to [value]. If the average grayscale value is within the range of 120–160, then the fine-tuning parameter is: This setting is based on the grayscale-concentration mapping relationship of underground coal mine gas images (e.g., grayscale values ​​of 100–150 correspond to gas concentrations of 0.5%–1.0%), ensuring that the enhancement operation does not exceed the allowable error range. This avoids significant deviations in the model during concentration identification tasks.

[0065] In application scenarios, this step is mainly used in the fine-tuning stage of coal mine underground gas concentration identification models. Since gas images are usually grayscale or pseudo-color images, their key features are grayscale distribution and concentration gradient, rather than texture or color information. Therefore, this enhancement strategy is particularly suitable for monitoring systems in high-gas mines, effectively improving the model's sensitivity to concentration changes and identification stability.

[0066] The technical effect of this step is that by dynamically disabling destructive enhancement operations and finely controlling the grayscale adjustment range, the false positive rate of gas concentration identification is significantly reduced. Experiments show that after adopting this strategy, the gas concentration identification error can be reduced from 20% under the general enhancement strategy to less than 1%, while improving the model's robustness to concentration distribution, providing reliable data augmentation support for coal mine safety monitoring.

[0067] S4 uses the validation set to train and validate the initial strategy. Based on indicators such as fault detection accuracy and gas concentration error, the strategy parameters are dynamically adjusted until the indicators meet the standards or the parameters reach the lower limit of the constraints. Executable enhancement code and strategy report are then output.

[0068] Specifically, in the differentiated strategy generation and optimization module, training and validating the initial strategy using a validation set is a crucial step in ensuring the adaptability of the enhancement strategy and the performance of the model. This step generates an initial enhancement strategy based on the fine-grained feature vectors of coal mine fine-tuning data and the matching of them with the strategy rule base. By training and evaluating the model on the validation set, the strategy parameters are dynamically adjusted until key performance indicators meet the standards or the parameters reach the lower limit of the constraints. Finally, executable enhancement code and a strategy report are output.

[0069] At the technical implementation level, this step first applies the initial strategy to the validation set images, performing operations including rotation, cropping, and brightness fine-tuning. The enhanced images are then used for model fine-tuning training. The training process typically involves 10 rounds, after which the model's performance metrics on the validation set are calculated, such as equipment fault detection accuracy, personnel violation identification rate, and gas concentration error. If the detection accuracy is below 95% or the gas concentration error exceeds 1%, a dynamic adjustment mechanism for the strategy parameters is triggered. The adjustment is based on priority settings in a coal mine-specific constraint rule library; for example, adjusting the rotation angle of the equipment image from... Shrink to Or adjust the cutting ratio from Upgraded to To reduce damage to key structural features.

[0070] At the parameter level, the validation set is divided into training set and validation set. This ensures the strategy optimization process is representative. The rotation angle adjustment range is... to The cutting ratio is to Brightness adjustment range is to The grayscale adjustment range is to The optimization termination condition is that the indicator fluctuation is less than 1% after two consecutive adjustments, or the parameter reaches the lower limit of the constraint set in the rule base.

[0071] At the application level, this step is widely used in the fine-tuning of models for equipment fault detection in fully mechanized coal mining faces, identification of violations by underground personnel, and identification of gas concentration. For example, in the oil seepage detection model of hydraulic supports in fully mechanized coal mining faces, optimizing rotation and trimming parameters through validation set feedback can effectively improve the model's ability to identify subtle oil seepage features.

[0072] The technical benefits of this step are that it significantly improves the accuracy and stability of model fine-tuning while reducing manual debugging time. Through automated verification and parameter adjustment, the model's fault detection accuracy can be stabilized at over 95%, the gas concentration identification error can be controlled within 1%, and the strategy debugging time can be shortened to within 4 hours, achieving an efficient, safe, and traceable coal mine visual model optimization process.

[0073] The differential data enhancement strategy generation method for fine-tuning coal mine visual models in this invention can effectively protect key features in coal mine visual data and improve the accuracy of model fine-tuning. Through automated strategy generation and optimization, it can significantly improve fine-tuning efficiency and reduce reliance on manual intervention.

[0074] Furthermore, S4 includes: S41 performs 10 rounds of rapid training to calculate the accuracy of equipment fault detection. Gas concentration error Misjudgment rate of personnel violation identification Three types of core indicators Specifically, in the differentiated data augmentation strategy generation method of the present invention, 10 rounds of rapid training are performed and the accuracy of equipment fault detection is calculated. Gas concentration error Misjudgment rate of personnel violation identification The three core metrics are crucial steps in validating the effectiveness of the enhancement strategy and the fine-tuning of the model. This step involves rapid iterative training on the validation set, combined with the performance requirements specific to the coal mining industry, to dynamically optimize the enhancement strategy, ensuring that the final output strategy possesses high accuracy and high stability in actual deployment.

[0075] At the technical implementation level, this step employs a lightweight training mechanism, using approximately 100–200 enhanced validation set images per training round, with the training cycle controlled within 10–15 minutes. The training model is based on typical architectures for coal mine fine-tuning tasks (such as ResNet-50 and YOLOv8), and uses cross-entropy loss and mean squared error as optimization objectives. After each training round, the system automatically evaluates the model's performance on the validation set, calculating three core metrics: equipment fault detection accuracy. Gas concentration identification error Misjudgment rate of personnel violation identification .

[0076] At the parameter level, this step sets strict performance thresholds: the accuracy of equipment fault detection must reach [a certain level]. To meet the high reliability requirements of coal mines for identifying the status of key equipment; the gas concentration identification error needs to be controlled within... Ensure that the concentration estimate is within the safe threshold range; the misjudgment rate of personnel violation identification needs to be reduced. To avoid invalid alarms caused by false alarms, if any indicator fails to meet the standard, the system will automatically adjust the enhancement parameters (such as rotation angle, cropping ratio, and brightness variation range) according to the priority rules in the coal mine-specific strategy rule library, and retrain and verify until the indicator fluctuation is less than 1% for two consecutive rounds, ensuring strategy convergence.

[0077] At the application level, this step is widely used in the fine-tuning process of underground visual models in coal mines, especially in scenarios such as equipment status monitoring in fully mechanized mining faces, identification of personnel safety behaviors, and visualization analysis of gas concentration. Through rapid training and index verification, the system can complete strategy optimization within a limited time, adapting to differences in lighting conditions, equipment models, personnel behavior patterns, and other factors in different mines, thereby improving the model's generalization ability and robustness in complex environments.

[0078] At the technical level, this step achieves closed-loop optimization of the enhancement strategy, effectively solving problems such as feature destruction, poor adaptability, and heavy reliance on manual intervention in existing fixed strategies. Through automated training and indicator feedback mechanisms, it not only improves model performance but also significantly reduces debugging time and manual intervention costs, providing the coal mining industry with an efficient, safe, and traceable model fine-tuning solution.

[0079] S42. If the training evaluation index fails to meet the target, the parameters of the initial enhancement strategy are adjusted. When the index fluctuates by 1% after at least two consecutive adjustments, the optimization is terminated, and a strategy report containing feature analysis results, rule matching logic, and optimization process is output.

[0080] Specifically, the present invention will be described below using two consecutive adjustments as an example. When the index fluctuates by 1% after two consecutive adjustments, the optimization is terminated, and a strategy report containing feature analysis results, rule matching logic, and optimization process is output. This is the key termination and output mechanism in the differentiated data augmentation strategy generation method of the present invention. Its technical implementation principle is based on a dynamic optimization process driven by validation set feedback. Combined with the accuracy and stability requirements of fine-tuning the coal mine visual model, it ensures that the strategy optimization process converges within a reasonable range and avoids overfitting or invalid iteration.

[0081] At the technical implementation level, this step determines whether the strategy optimization has reached a stable state by monitoring the continuous changing trend of model performance indicators (such as equipment fault detection accuracy, personnel violation identification rate, and gas concentration error) on the validation set. Specifically, after each strategy parameter adjustment, the system performs a complete augmentation and model training process on the validation set (usually 10 rounds of fine-tuning) and records the magnitude of change in the target indicators. If the indicator fluctuation after two consecutive adjustments is less than 1%, it meets the requirements. If the current strategy has converged to the optimal or suboptimal state, the optimization process terminates. This mechanism is based on the high accuracy and stability requirements of the coal mine model, such as the need for high accuracy in equipment fault detection. The misjudgment rate of personnel violation identification needs to be improved. Gas concentration error needs This ensures that strategy adjustments do not introduce noise or ineffective enhancements due to minor fluctuations.

[0082] At the parameter level, the core parameters involved in this step include: the validation set split ratio (training set:validation set = 8:2), the number of model training epochs (10 epochs), the metric fluctuation threshold (1%), and the minimum step size for strategy adjustment (e.g., the step size for adjusting the rotation angle is...). The cutting ratio step size is These parameters together constitute the control logic of the optimization process, ensuring that strategy adjustments are made within a limited range, while avoiding the omission of better solutions due to premature termination.

[0083] In application scenarios, this step is widely used in the fine-tuning process of visual models in coal mines, such as detecting leakage in hydraulic supports in fully mechanized mining faces, identifying conveyor belt misalignment, and recognizing the wearing of safety helmets by underground personnel. In actual deployment, the system automatically monitors changes in validation set indicators, determining whether the strategy has reached a stable state without manual intervention, and thus outputs the final strategy report.

[0084] The technical benefits of this step are a significant improvement in the efficiency and stability of strategy optimization. By setting reasonable termination conditions, the system can control strategy debugging time to within 4 hours, reducing it by more than 90% compared to existing solutions. Simultaneously, the output strategy report includes complete feature analysis results (such as equipment outline integrity and personnel posture distribution), rule matching logic (such as the association rule between rotation angle and structural integrity), and the optimization process (such as parameter adjustment history and validation set performance curves). This provides traceable technical evidence for subsequent model deployment and regulatory review, ensuring the compliance and reliability of the enhanced strategy in coal mine safety scenarios.

[0085] The differential data enhancement strategy generation method for fine-tuning coal mine visual models in this invention can effectively protect key features in coal mine visual data and improve the accuracy of model fine-tuning. Through automated strategy generation and optimization, it can significantly improve fine-tuning efficiency and reduce reliance on manual intervention.

[0086] Example 2 Figure 2 This is a structural diagram of a differential data augmentation strategy generation system for fine-tuning a coal mine visual model, according to an embodiment of the present invention.

[0087] like Figure 2 As shown, the differential data augmentation strategy generation system for fine-tuning the coal mine visual model includes: The data receiving and classification preprocessing module is used to receive various types of fine-tuned data from coal mines and perform classification preprocessing. It processes equipment images, personnel images, and gas images into a unified format and resolution, and automatically divides the training set and validation set according to the labels. The fine-grained feature extraction module is used to extract fine-grained features based on preprocessed data and generate feature vectors. It extracts structural integrity, angle deviation, size ratio and gray-level distribution features from equipment images, posture distribution, key feature ratio and brightness distribution features from personnel images, and concentration gradient, gray-level mean and noise intensity features from gas images. The strategy matching and generation module is used to match the coal mine-specific strategy rule library based on the feature vector and dynamically generate the initial enhancement strategy. Rotation operation is disabled for equipment images when the structural integrity is below the threshold, and flip operation is disabled for personnel images when the posture distribution does not meet the preset conditions. Rotation / flip operation is disabled for gas images when the concentration gradient is above the threshold. The training, validation, and parameter optimization module is used to train and validate the initial enhancement strategy using the validation set. It dynamically adjusts the strategy parameters based on indicators such as fault detection accuracy and gas concentration error until the indicators meet the standards or the parameters reach the lower limit of the constraints, and outputs executable enhancement code and strategy report.

[0088] Furthermore, the fine-grained feature extraction module is also used to perform Canny edge detection on equipment images to calculate contour integrity, detect the angle between the principal axis and the horizontal axis through Hough transform, and statistically analyze the proportion of the equipment area in the image and the mean gray value to generate a 128-dimensional equipment feature vector; for personnel images, MTCNN keypoint detection is used to statistically analyze pose distribution, calculate the proportion of the safety helmet / reflective clothing area, analyze the mean brightness of the personnel area, and generate a 128-dimensional personnel feature vector.

[0089] Furthermore, the strategy matching and generation module is also used to disable rotation operations when the device image structure integrity is below 95%, and the rotation angle is calculated according to the formula. Dynamic calculation, in which The original angular deviation detected; when the gas image concentration gradient... Disable rotation / flipping operations and set grayscale fine-tuning parameters according to the grayscale mean range: fine-tune ±1% when the grayscale mean is 80-120, and fine-tune ±0.5% when the grayscale mean is 120-160.

[0090] Furthermore, at least ten rapid training sessions are performed, and the accuracy rate of equipment fault detection, gas concentration error, and misjudgment rate of personnel violation identification are calculated as training evaluation indicators. If any training evaluation indicator fails to meet the standard, the parameters of the initial enhancement strategy are adjusted. When the indicator fluctuates after at least two consecutive adjustments, the optimization is terminated, and a strategy report containing feature analysis results, rule matching logic, and optimization process is output.

[0091] The present invention also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, a method for generating the differential data enhancement strategy for fine-tuning the above-mentioned coal mine visual model is implemented.

[0092] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0093] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

Claims

1. A method for generating differentiated data augmentation strategies for fine-tuning a coal mine visual model, characterized in that, include: S1, receive multi-type fine-tuning images of coal mines and perform classification preprocessing to unify the image format and resolution. The fine-tuning images are divided into equipment images, personnel images and gas images. S2, perform feature extraction on the preprocessed fine-tuned image to obtain feature vectors; when the fine-tuned image is a device image, the extracted features include contour integrity, angle deviation, size proportion, and gray-level distribution features. The contour integrity is determined by the continuity rate of the device edges, and the angle deviation is determined by the angle between the main axis and the horizontal axis of the device image; when the fine-tuned image is a personnel image, the extracted features include posture distribution, key feature proportion, and brightness distribution features; when the fine-tuned image is a gas image, the extracted features include gray-level gradient, gray-level mean, and Gaussian noise variance. S3, based on different types of fine-tuned images and their feature vectors, match the corresponding policy rule base, and dynamically generate an initial enhancement strategy. The initial enhancement strategy is used to determine the operations that can be performed on the image based on each image type and feature. S4 dynamically adjusts the operating parameters in the initial enhancement strategy based on the fault detection accuracy and gas concentration error index until the index meets the standard or the parameter reaches the lower limit of the constraint, and outputs executable enhancement code and strategy report.

2. The method as described in claim 1, characterized in that, S2 further includes: S21, when extracting features of the device image, the Canny edge detection method is used to detect the image edge and calculate the contour integrity. The angle between the principal axis and the horizontal axis is detected by Hough transform. The angle between the principal axis and the horizontal axis is used to determine the angular deviation of the device. The proportion of the device area in the image is used as the size proportion. The average gray value of the image is used as the gray value distribution feature to generate the device feature vector. S22, when extracting features from personnel images, MTCNN keypoint detection is used to statistically distribute poses, the size ratio of the safety helmet / reflective clothing area in the image is calculated as the key feature ratio, the average brightness of the personnel area is analyzed as the brightness distribution feature, and a personnel feature vector is generated. S23. When extracting features from the gas image, the Sobel operator is used to calculate the gray-level gradient of the gas image, the mean gray level of the entire gas image is calculated, and the Gaussian noise variance of the image is calculated to generate the gas feature vector.

3. The method as described in claim 1, characterized in that, S3 further includes: S31, for device images, when the structural integrity of the device image is less than 95%, the rotation operation is disabled; when the size of the device in the device image accounts for less than 60%, the cropping operation is disabled. S32, for the personnel image, when the average brightness exceeds the threshold range, adjust the brightness of the personnel image; when the personnel account for less than 50% of the personnel image, disable the rotation operation. S33, for gas images, when the gas image concentration gradient... When doing so, disable rotation or flipping operations and set grayscale fine-tuning parameters according to the grayscale mean range.

4. The method as described in claim 3, characterized in that, S32 further includes: when the average brightness of the personnel image is less than 40, performing an operation to increase the brightness of the personnel image; when the average brightness of the personnel image is greater than 90, performing an operation to decrease the brightness of the personnel image.

5. The method as described in claim 3, characterized in that, S33 further includes: when the average grayscale value of the gas image is between 80 and 120, performing a grayscale fine-tuning operation of ±1%; when the average grayscale value of the gas image is between 120 and 160, performing a grayscale fine-tuning operation of ±0.5%.

6. The method as described in claim 1, characterized in that, S4 further includes: S41, perform at least ten rapid training sessions and calculate the accuracy of equipment fault detection, gas concentration error, and misjudgment rate of personnel violation identification as training evaluation indicators. S42, If any training evaluation metrics fail to meet the standards, adjust the parameters of the initial enhancement strategy. S43: Optimization is terminated when the indicator fluctuation is less than 1% after at least two consecutive adjustments, and a strategy report containing feature analysis results, rule matching logic, and optimization process is output.

7. A system for generating differentiated data augmentation strategies for fine-tuning a visual model in coal mines, characterized in that, include: The data receiving and classification preprocessing module is used to receive and classify various types of fine-tuned images from the coal mine, unify the image format and resolution, and classify the fine-tuned images into equipment images, personnel images and gas images. The fine-grained feature extraction module is used to extract features from the preprocessed fine-tuned image and obtain feature vectors. When the fine-tuned image is an equipment image, the extracted features include structural integrity, angle deviation, size ratio, and gray-level distribution features. When the fine-tuned image is a personnel image, the extracted features include posture distribution, key feature ratio, and brightness distribution features. When the fine-tuned image is a gas image, the extracted features include concentration gradient, gray-level mean, and noise intensity. The strategy matching and generation module is used to match the corresponding strategy rule library according to different types of fine-tuned images and their feature vectors, and dynamically generate an initial enhancement strategy. The initial enhancement strategy is used to determine the operations that can be performed on the image based on each image type and feature. The training, validation, and parameter optimization module is used to dynamically adjust the operating parameters in the initial enhancement strategy based on the fault detection accuracy and gas concentration error index, until the index meets the standard or the parameters reach the lower limit of the constraint, and outputs executable enhancement code and strategy report.

8. The system as described in claim 7, characterized in that, The strategy matching and generation module is also used to: when processing device images, disable rotation operation when the structural integrity of the device image is less than 95%, and disable cropping operation when the size of the device in the device image is less than 60%. When processing images of people, the brightness of the images is adjusted when the average brightness exceeds the threshold range, and the rotation operation is disabled when the proportion of people in the images is less than 50%. When processing gas images, the gas image concentration gradient When doing so, disable rotation or flipping operations and set grayscale fine-tuning parameters according to the grayscale mean range.

9. The system as described in claim 7, characterized in that, The training validation and parameter optimization module is also used for: Perform at least ten rapid training sessions and calculate the accuracy of equipment fault detection, gas concentration error, and misjudgment rate of personnel violation identification as training evaluation indicators. If any training evaluation metrics fail to meet the standards, the parameters of the initial augmentation strategy will be adjusted. Optimization will terminate when the indicator fluctuates after at least two consecutive adjustments, and a strategy report containing feature analysis results, rule matching logic, and optimization process will be output.

10. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the method as claimed in any one of claims 1-6.

Images