Underground abnormity intelligent identification method and system based on image enhancement

Abstract

The invention discloses an intelligent underground anomaly identification method and system based on image enhancement, and relates to the technical field of image enhancement. The method comprises the following steps: acquiring an underground image and extracting image features; calculating a dust shielding coefficient based on the image features and the air sensing data, and performing adaptive image enhancement to obtain an enhanced image; performing preliminary anomaly recognition based on the enhanced image in combination with a same-region historical image set; based on the preliminary identification result, combining with a historical image set to carry out image segmentation and executing secondary enhancement, and obtaining an enhanced segmentation image set and a risk factor; and finally, carrying out specific identification and risk correction in combination with the risk factor, and outputting an underground abnormity intelligent identification result. According to the invention, accurate image enhancement is realized by dynamically sensing environment change, and historical data is comprehensively utilized to carry out anomaly positioning and risk quantification, so that the accuracy and reliability of anomaly recognition and the timeliness of early warning in the underground complex environment are improved.

Description

Technical Field

[0001] This invention relates to the field of image enhancement technology, and more specifically to a method and system for intelligent identification of downhole anomalies based on image enhancement. Background Technology

[0002] The underground environment presents inherent challenges such as insufficient lighting and pervasive dust, resulting in images that are often characterized by low contrast, blurred details, and severe noise interference.

[0003] Existing downhole image anomaly recognition solutions mostly focus on directly applying general image enhancement algorithms or target detection models. They typically perform global, uniform image processing, failing to fully consider the varying impacts of dynamic changes in downhole dust concentration on image degradation. This results in limited enhancement effects and insufficient adaptability. Furthermore, existing methods rely solely on single-frame image information for analysis, lacking the integration and utilization of historical states and anomaly patterns in the monitored area. This leads to weak perception of subtle texture changes, small targets, or positional deviations, resulting in high false alarm and false negative rates, making it difficult to meet the practical requirements for high-reliability safety early warning. Summary of the Invention

[0004] This invention addresses the technical problems of poor adaptability, difficulty in fully utilizing historical information, and insufficient accuracy in identifying complex anomalies in existing downhole image enhancement methods by providing an intelligent downhole anomaly identification method and system based on image enhancement.

[0005] The technical solution of the present invention to solve the above-mentioned technical problems is as follows:

[0006] In a first aspect, the present invention provides an intelligent downhole anomaly identification method based on image enhancement, comprising: Acquire downhole images and extract image features from the downhole images; Based on the image features and air sensor data, the dust obstruction coefficient is obtained, and the downhole image is enhanced to obtain an enhanced image. Based on the enhanced image, combined with a set of historical images of the same area, preliminary anomaly identification is performed to obtain preliminary identification results; Based on the preliminary identification results, image segmentation is performed in conjunction with historical image sets of the same area, followed by secondary enhancement to obtain enhanced segmented image sets and risk factors. Specific identification is then performed to obtain intelligent identification results for downhole anomalies.

[0007] Secondly, the present invention provides an image-enhanced intelligent identification system for downhole anomalies, comprising: An image acquisition module is used to acquire downhole images and extract image features from the downhole images; The image enhancement module is used to obtain a dust obstruction coefficient based on the image features and air sensor data, and to enhance the downhole image to obtain an enhanced image. The preliminary anomaly identification module is used to perform preliminary anomaly identification based on the enhanced image and in combination with a set of historical images of the same area, and to obtain preliminary identification results. The specific identification module is used to perform image segmentation based on the preliminary identification results and the historical image set of the same area, and to perform secondary enhancement to obtain the enhanced segmented image set and risk factors, and to perform specific identification to obtain the intelligent identification results of downhole anomalies.

[0008] The beneficial effects of this invention are: Compared to existing technologies, this invention firstly overcomes the problem of unstable image quality caused by changes in downhole dust concentration by dynamically calculating the dust occlusion coefficient and performing adaptive image enhancement through the fusion of image features and real-time air sensor data. Secondly, based on the initial enhancement, a set of historical images from the same region is introduced as a reference benchmark, and preliminary anomaly localization is achieved through comparative analysis, reducing environmental background interference. Thirdly, based on the preliminary results and historical high-frequency anomaly information, refined image segmentation is performed, and secondary targeted enhancement is applied to the segmented regions, improving the visual expressiveness and feature separability of key anomaly areas. Finally, combined with the calculated hazard factors, a dedicated recognition branch model is guided to perform specific analysis and risk correction, thereby achieving high-precision and high-reliability intelligent identification and early warning of various downhole anomalies, especially small targets and complex texture anomalies. Attached Figure Description

[0009] Figure 1 A flowchart illustrating the intelligent downhole anomaly identification method based on image enhancement provided by this invention; Figure 2 This is a schematic diagram of the structure of the intelligent downhole anomaly identification system based on image enhancement provided by the present invention.

[0010] In the attached diagram, the components represented by each number are as follows: Image acquisition module 11, image enhancement module 12, preliminary anomaly recognition module 13, and specificity recognition module 14. Detailed Implementation

[0011] Example 1, as Figure 1 As shown, this embodiment of the invention provides an intelligent downhole anomaly identification method based on image enhancement, including: S10: Acquire downhole images and extract image features from the downhole images; Specifically, acquiring downhole images and extracting image features from the downhole images includes: Acquire downhole images; The image features of the downhole image are extracted, including color features, texture features, and edge gradient features.

[0012] First, underground images are acquired. "Underground" specifically refers to underground working spaces such as mine roadways, working faces, transport roadways, and electromechanical chambers. Image acquisition equipment includes intrinsically safe visible light cameras, infrared thermal imagers, or multispectral imaging devices. The acquired underground images form the raw data basis for subsequent analysis and processing.

[0013] Simultaneously, image features are extracted from the downhole images. These image features include color features, texture features, and edge gradient features of the downhole images, used to comprehensively characterize the visual attributes of the downhole scene.

[0014] Color features are obtained by analyzing the statistical distribution of each pixel in the color space of the downhole image, reflecting the surface color attributes and lighting conditions of the downhole environment and objects. Texture features are obtained by calculating descriptors such as the gray-level co-occurrence matrix or directional gradient histogram of local regions in the downhole image, used to characterize the roughness, uniformity, and directionality of equipment surfaces, rock walls, or material structures. Edge gradient features are obtained by performing gradient operator convolution calculations on the downhole image, used to describe the local change intensity and direction of object contours, structural boundaries, or areas of abrupt intensity changes in the image, providing a basis for identifying abnormal geometric shapes such as cracks and damage. The extraction of these features together constitutes the image features of the downhole image, used for multi-dimensional quantitative description of the downhole image content.

[0015] S20: Based on the image features and air sensor data, obtain the dust obstruction coefficient, perform image enhancement on the downhole image, and obtain the enhanced image; Secondly, air sensor data is acquired, and combined with the image features extracted in the preceding steps, the dust occlusion coefficient is calculated. The dust occlusion coefficient is a comprehensive quantitative indicator, its value determined by both the real-time dust concentration reflected in the air sensor data and the degree of visual degradation embodied in the image features. It characterizes the scattering and absorption effects caused by the dusty environment on the imaging optical path, and the resulting level of image quality degradation. The dust occlusion coefficient not only describes the loss of image apparent sharpness but also quantifies, from a physical perspective, the actual occlusion impact of environmental dust on the visual information transmission process.

[0016] Because the concentration and distribution of dust in underground environments are dynamically changing, relying solely on the visual features of the image itself to assess the degree of degradation is easily affected by the inherent complexity of the underground scene and the stability of lighting conditions. Conversely, simply relying on airborne dust concentration data makes it difficult to accurately measure the visual impact of that concentration on a specific image at a given moment. Therefore, it is necessary to calculate a dust occlusion coefficient to provide a controllable basis for subsequent image enhancement processing that can adapt to environmental changes.

[0017] Based on the dynamic matching of the dust occlusion coefficient and the implementation of the corresponding intensity of the image enhancement algorithm, it is possible to ensure the restoration and optimization of degraded images under different dust conditions, effectively improve the visibility and contrast of key details in the image, avoid unnecessary over-processing of relatively clear areas, and thus obtain enhanced images with improved quality and high information fidelity as a whole.

[0018] Specifically, based on the image features and air sensor data, a dust obstruction coefficient is obtained, and image enhancement is performed on the downhole image to obtain an enhanced image, including: The image features are input into an image feature recognizer to obtain the image occlusion coefficient; The air sensor data of the area where the downhole image is located is obtained, and the dust occlusion coefficient is calculated by combining the image occlusion coefficient. Based on the dust obscuration coefficient, an image enhancement strategy is matched, and the image enhancement strategy is used to enhance the downhole image to obtain an enhanced image.

[0019] First, the image features are input into an image feature recognizer to obtain the image occlusion coefficient. This image feature recognizer analyzes the color features, texture features, and edge gradient features of the input image to output a quantified image occlusion coefficient. This image occlusion coefficient directly reflects the degree of overall image sharpness reduction and detail loss caused by factors such as dust adhesion and insufficient lighting.

[0020] Specifically, the acquisition of the image feature recognizer includes: Obtain a historical image set, and obtain the image features of each historical image in the historical image set as a sample input set; The historical images are labeled with the degree of image occlusion to obtain a set of sample image occlusion coefficients, wherein the labeling of the degree of image occlusion is obtained based on the degree of image feature degradation; An image feature recognizer is constructed based on machine learning. The image feature recognizer is trained using the sample input set and the sample image occlusion coefficient set until convergence.

[0021] First, a historical image set is acquired, and image features are extracted from each image in the set to form a sample input set. This historical image set covers diverse image samples collected from different areas, times, and working conditions underground. The extracted image features include color features, texture features, and edge gradient features. This step aims to construct a feature set that comprehensively reflects various typical visual patterns and degradation conditions underground, ensuring that the trained image feature recognizer has good generalization ability.

[0022] Secondly, the occlusion levels of historical images in the historical image set are labeled to obtain a set of sample image occlusion coefficients. The labeling of image occlusion levels is based on the degree of degradation of image features. Specifically, the dispersion of color distribution can be evaluated by calculating the histogram variance of each channel in the standard color space; the clarity of texture structure can be measured by calculating the local binary mode variance or the contrast of the gray-level co-occurrence matrix; and the sharpness of edge contours can be evaluated by calculating the average amplitude of the edge detection response map. Next, the three indicators mentioned above—color variance, texture contrast, and edge amplitude—are normalized and weighted summed to generate a scalar value between 0 and 1, which is defined as the image occlusion coefficient.

[0023] Image occlusion coefficient = α×C + β×T + γ×E. Where C represents the color distribution variance, used to quantify the dispersion of the image's color channels; T represents the texture contrast, used to quantify the clarity of the texture structure; and E represents the mean edge gradient magnitude, used to quantify the sharpness of the edge contour. α, β, and γ are the weighting coefficients corresponding to the above three features, and satisfy α + β + γ = 1. Each weighting coefficient is set specifically according to the recognition needs of different monitoring scenarios in the mine. For example, in the equipment surface condition monitoring scenario, the texture feature weight β can be set to a higher value to enhance the sensitivity to texture degradation phenomena such as surface cracks and corrosion; in the personnel activity safety monitoring scenario, the edge gradient feature weight γ can be increased accordingly to strengthen the evaluation ability of personnel contour integrity and movement clarity. This image occlusion coefficient directly quantifies the degradation level of the corresponding image's visual quality; a higher value indicates a more severe overall blurring and loss of detail caused by factors such as dust, water mist, or low illumination.

[0024] Furthermore, by using the sample input set and combining it with the corresponding sample image occlusion coefficient set, a complete training sample pair is formed. Based on machine learning, a model that can automatically infer the degree of occlusion from image features is trained, namely an image feature recognizer.

[0025] For example, considering that the mapping relationship between image features and occlusion degree has strong non-linear characteristics, and that machine learning methods can effectively learn complex patterns from high-dimensional features, machine learning models can be used to build this recognizer, including support vector regression, random forest regression, and fully connected neural networks. Fully connected neural networks are selected as an example for the following detailed explanation.

[0026] Specifically, this image feature recognizer is based on a neural network architecture, mainly consisting of an input layer, several hidden layers, and an output layer. The input layer receives a normalized image feature vector, which is composed of three types of features: color distribution variance, texture contrast, and mean edge gradient magnitude. The hidden layers contain two to four fully connected network layers, with the number of neurons in each layer decreasing sequentially. For example, the number of neurons in the first layer is set to twice the input dimension, and then halved in each subsequent layer. A ReLU activation function is used after each hidden layer to introduce non-linear processing capabilities, followed by a Dropout layer with a dropout rate between 0.1 and 0.3 to suppress overfitting during training and improve the model's generalization performance. The output layer uses a linear activation function to map the network's final abstract representation to a single continuous value, namely the image occlusion coefficient.

[0027] The key parameters for model training are set as follows: the learning rate is set to 0.0005, the maximum number of training epochs is set to 200, and the batch size is set to 32. The learning rate is chosen to balance the stability and convergence speed of gradient descent; the number of training epochs ensures that the model has sufficient opportunity to learn the complex relationship between features and coefficients; and the batch size is chosen to balance training efficiency with hardware memory limitations.

[0028] Specifically, the sample input set obtained above is used, and combined with the corresponding sample image occlusion coefficient set, the sample input set and the corresponding sample image occlusion coefficient set are merged, and randomly divided into training set, validation set and test set in a ratio of 8:1:1.

[0029] Then, using the feature vectors of the training set as model input and their corresponding true coefficient labels as supervision signals, model training begins. Mean squared error (MSE) is used as the loss function to measure the difference between the model's predicted occlusion coefficients and the true labels. The Adam optimizer is used for optimization, iteratively updating all weight parameters in the network through backpropagation. During training, a validation set is used to monitor model performance. Training is terminated early when the loss function value on the validation set no longer decreases over multiple training epochs, or when the model's prediction error on the validation set reaches a preset accuracy requirement (e.g., mean absolute error below 0.05), to prevent overfitting. Finally, the model parameters obtained after training convergence are saved, resulting in a usable image feature recognizer. This image feature recognizer accurately captures the intrinsic correlation between multi-dimensional visual features of an image and its overall occlusion degradation degree, enabling automated and quantitative evaluation of the occlusion coefficient of any input downhole image.

[0030] Finally, the image features are input into the trained image feature recognizer to obtain the image occlusion coefficient.

[0031] Furthermore, air sensor data of the area where the downhole image is located is acquired, and combined with the image occlusion coefficient, a dust occlusion coefficient is calculated. The air sensor data is collected in real time by environmental sensors deployed near the image acquisition point, and includes at least the dust concentration parameter in the air. The real-time dust concentration and the image occlusion coefficient are weighted and fused to obtain a comprehensive dust occlusion coefficient.

[0032] The dust occlusion coefficient is calculated as follows: ω1 × normalized dust concentration + ω2 × image occlusion coefficient. Here, ω1 and ω2 are the weighting coefficients corresponding to dust concentration and image occlusion features, respectively, and satisfy ω1 + ω2 = 1. The weighting coefficients are set based on the reliability of the two data sources and their contribution to the final quality assessment in a specific application scenario. For example, at fixed monitoring points where the dust concentration sensor has been rigorously calibrated and the data confidence is high, a higher ω1 value, such as 0.7, can be set. In areas where the sensor may be subject to occasional interference or where the image content itself is extremely sensitive to occlusion, the weight of ω2 can be appropriately increased, such as setting it to 0.6, to enhance the role of visual features in the comprehensive assessment. This dust occlusion coefficient not only considers the visual degradation of the image itself but also incorporates objective environmental measurement data, thus more accurately characterizing the actual impact of dust on image quality under current environmental conditions.

[0033] Furthermore, based on the calculated dust occlusion coefficient, an image enhancement strategy is matched and applied to enhance the downhole image, thereby obtaining an enhanced image. First, a threshold range for the dust occlusion coefficient, comprising multiple levels, is pre-defined, with each range associated with a specific image enhancement strategy. Specifically, the image enhancement strategy is configured differently in terms of enhancement intensity, contrast adjustment range, noise suppression parameters, and detail enhancement degree, depending on the degree of occlusion. For example, for images with a high dust occlusion coefficient, i.e., when the image is blurry, a stronger enhancement strategy is applied to improve visibility; while for images with a low dust occlusion coefficient, a milder enhancement strategy is applied, aiming to optimize the visual effect while preserving the details and realism of the original image to the maximum extent, avoiding distortion or loss of effective information due to over-processing.

[0034] Adaptive matching of image enhancement strategies enables the most suitable enhancement processing for original input images of varying qualities. By employing the matched image enhancement strategy to enhance the downhole image, an enhanced image can be obtained. This enhanced image is based on visual data dynamically optimized by the dust occlusion coefficient. While preserving key information of the original scene, it improves overall contrast, visibility of local details, and the sharpness of target edges. This enhanced image effectively suppresses noise interference introduced by dust scattering and insufficient illumination, while avoiding overexposure, color cast, or detail loss that might occur with globally uniform enhancement. Therefore, it provides high-quality, high-fidelity visual input for subsequent anomaly identification and analysis.

[0035] S30: Based on the enhanced image, combined with the historical image set of the same area, perform preliminary anomaly identification and obtain preliminary identification results; Specifically, based on the enhanced image and combined with a set of historical images of the same region, preliminary anomaly identification is performed to obtain preliminary identification results, including: Acquire historical image sets for the same area and determine the normal location range of the device; Based on the normal position range of the device, position anomaly identification is performed on the enhanced image to obtain the position anomaly identification result; Based on the image features of the historical image set of the same region and the enhanced image, feature statistical deviation identification is performed to obtain feature anomaly identification results. The preliminary identification result is obtained by combining the location anomaly identification result and the feature anomaly identification result.

[0036] First, a historical image set for the same area is acquired, and the normal position range of the equipment is extracted. The historical image set for the same area refers to a collection of time-series images under normal operating conditions, collected and stored over a long period at the same monitoring point for the same underground scene. By performing target detection and position analysis on all historical images in this set, the spatial distribution range of each key piece of equipment, such as conveyor belt drive rollers, hydraulic supports, or switchgear, in the image pixel coordinate system can be statistically determined. This range is defined as the normal position range of the equipment. This normal position range is typically represented by a bounding box or mask region.

[0037] Secondly, based on the normal position range of the equipment, the enhanced image obtained above is used to identify positional anomalies, and the results are obtained. Specifically, a target detection algorithm is applied to the current enhanced image to identify the current position of key equipment in the enhanced image. Next, this detected position is compared with a predefined normal position range of the equipment. If the detected equipment position is completely or partially outside the normal position range, a positional anomaly is determined, and the category of the abnormal equipment and its degree of deviation are recorded, forming a positional anomaly identification result. Through positional anomaly identification, macroscopic anomalies such as overall equipment displacement, tilting, or significant misalignment can be discovered.

[0038] Furthermore, based on the image features of historical images and enhanced images from the same region, feature statistical deviation identification is performed to obtain feature anomaly identification results. Specifically, firstly, the statistical distribution of all normal images in the historical image set of the same region in terms of features such as color, texture, and edge gradient is calculated, for example, the mean and standard deviation, to establish a multi-dimensional feature benchmark model of the region under normal conditions. Secondly, the same image features of the current enhanced image are extracted, and the extracted feature values ​​are compared with the benchmark model. By calculating Mahalanobis distance, Z-score, or other statistical deviation measures, the degree of deviation between the current image features and the historical normal feature distribution is quantified.

[0039] If the calculated deviation value exceeds the preset statistical significance threshold, the current image is determined to have an anomaly in its visual features, and the dimension of the anomaly and the intensity of the deviation are recorded to form a feature anomaly identification result. The preset statistical significance threshold is set based on the stability of the feature distribution of historical image sets in the same area under normal operating conditions and the allowable reasonable fluctuation range. For example, for a certain texture contrast feature, the threshold can be set to the range of the mean plus or minus three times the standard deviation, based on its historical mean and standard deviation; anything exceeding this range is considered a statistically significant anomaly.

[0040] By identifying feature statistical deviations, it is possible to identify potential microscopic problems such as texture changes on the equipment surface caused by rust or cracks, local color abnormalities caused by leakage or overheating, or fine structural defects caused by parts falling off, thus enabling early warning of early and hidden faults.

[0041] Finally, by combining the results of location anomaly identification and feature anomaly identification, a comprehensive judgment is made to obtain a preliminary identification result. This preliminary identification result is a structured output that integrates all identified location anomaly and feature anomaly information. This preliminary identification result not only indicates whether an anomaly exists, but also preliminarily indicates the type of anomaly, the approximate area where it occurs, and the significance of the anomaly, providing clear guidance for subsequent more refined image segmentation and in-depth analysis.

[0042] S40: Based on the preliminary identification results and the historical image set of the same area, image segmentation is performed, and secondary enhancement is performed to obtain the enhanced segmented image set and risk factors. Specific identification is then performed to obtain the intelligent identification results of downhole anomalies.

[0043] Specifically, image segmentation is performed based on the preliminary identification results and a set of historical images from the same region, including: Based on the preliminary identification results, the preliminary anomaly location is obtained; Image segmentation is performed based on the initial anomaly location to obtain segmented images, which are then added to the segmented image set. Based on the high-frequency anomaly locations in the historical image set of the same region, the enhanced image is segmented according to the high-frequency anomaly locations to obtain segmented images, which are then added to the segmented image set.

[0044] First, based on the preliminary identification results, preliminary anomaly locations are obtained. Specifically, the preliminary identification results include potential problem areas indicated by location anomaly identification and feature anomaly identification. From these, the spatial coordinates or bounding box information of all those marked as anomalies are extracted, which together constitute a preliminary set of anomaly locations that require in-depth analysis.

[0045] Secondly, image segmentation is performed based on the initial anomaly location to obtain segmented images, which are then added to the segmented image set. Specifically, the initial anomaly location is essentially one or more regions with defined spatial extent, typically represented by bounding box coordinates or outline polygons. Based on the spatial extent of this anomaly region, a local image completely covering the region can be directly cropped from the current enhanced image. Alternatively, to provide more comprehensive contextual information, the geometric center of the anomaly region can be used as a reference, and the cropping can be expanded according to a preset fixed size or a scale adaptively determined based on the region size. The preset fixed size refers to a setting based on the physical size of typical underground monitoring targets and their average pixel percentage in the image; for example, it can be set to the minimum rectangular window required to cover key components of common equipment, with a size of 256×256 pixels. Each cropped local image region constitutes a segmented image, its content fully encompassing the specific location initially determined to have an anomaly and its adjacent background. Summarizing all the obtained segmented images forms the first part of the segmented image set, which directly corresponds to newly detected suspicious target regions within the current analysis cycle.

[0046] Furthermore, based on high-frequency anomaly locations in the historical image set of the same region, high-frequency anomaly location segmentation is performed on the enhanced image to obtain segmented images, which are then added to the segmented image set. High-frequency anomaly locations refer to specific image regions identified from the historical image set of the same region where anomalous events have occurred multiple times at different points in the past.

[0047] Specifically, the locations of high-frequency anomalies are obtained, including: Obtain anomalies from the historical image set of the same region; Based on the frequency of occurrence of the aforementioned anomalies in the historical image set, the anomaly frequency is obtained; By combining the historical dust occlusion coefficients of the historical image set, an abnormal frequency threshold is obtained, and the abnormal situations are filtered to obtain high-frequency anomalies; Extract the location of the high-frequency anomaly and use it as the high-frequency anomaly location.

[0048] First, acquire records of anomalies from historical image sets of the same area. These anomalies refer to various abnormal events identified and recorded at the same monitoring point within a historical time period, along with their corresponding detailed information, including the anomaly type, the time of occurrence, and its specific location coordinates in the image.

[0049] Secondly, based on the frequency of occurrence of anomalies in the historical image set, the anomaly frequency is calculated. Specifically, for each anomaly location recorded in the historical image set, the cumulative number of anomaly events occurring at that location within a preset specific time window is counted. For example, the total number of times the same coordinate area was marked as an anomaly in the past three months is counted. This value is defined as the anomaly occurrence frequency at that location. This anomaly occurrence frequency is divided by the total number of image frames acquired during that time period to calculate a standardized anomaly occurrence frequency. This anomaly occurrence frequency quantifies the frequency and probability level of problems occurring in a specific image region historically.

[0050] Furthermore, by combining the historical dust occlusion coefficients of the historical image set, a dynamic anomaly frequency threshold is obtained, and this threshold is used to filter out anomalies, thereby identifying high-frequency anomalies. Considering that the underground dust environment directly affects image quality and the reliability of anomaly recognition, simply using a fixed frequency threshold is not accurate enough. Therefore, the historical dust occlusion coefficient needs to be introduced as an adjustment factor.

[0051] Specifically, for each historical anomaly, the historical dust obstruction coefficient corresponding to the time of its occurrence is associated. Optionally, the anomaly frequency threshold = base frequency threshold × (1 + historical dust obstruction coefficient). The base frequency threshold is set based on the comprehensive requirements of the specific downhole monitoring scenario for the sensitivity and stability of safety early warning, for example, 3 anomalies occurring every 30 days.

[0052] For anomalous events recorded under conditions of high historical dust occlusion coefficients, since the historical dust occlusion coefficients approach 1, the calculated dynamic anomaly frequency threshold will be higher than the base frequency threshold. This indicates that under such poor image quality conditions, a higher anomaly occurrence frequency needs to be observed to classify the relevant area as a high-frequency anomaly. This setting effectively reduces the probability of accidental misjudgments due to brief severe image blurring being included in the high-frequency anomaly set. Conversely, for anomalous events recorded under conditions of low historical dust occlusion coefficients, the historical dust occlusion coefficients approach zero, and the dynamic anomaly frequency threshold is approximately equal to the base frequency threshold. This indicates that under conditions of clear and reliable images, using the basic judgment criteria can maintain the ability to sensitively capture true anomaly patterns.

[0053] By dynamically adjusting the aforementioned abnormal frequency threshold, we can more reasonably distinguish between truly high-frequency problem areas and occasional false alarm areas caused by image quality fluctuations. Finally, we filter out anomalies whose frequencies exceed their corresponding dynamic abnormal frequency thresholds and classify them as high-frequency anomalies.

[0054] Finally, the locations of high-frequency anomalies are extracted and designated as high-frequency anomaly locations. Specifically, from the selected high-frequency anomaly records, their spatial coordinate information in the image is extracted, such as the vertex coordinates of the bounding box or the polygon contour point set, to form the final high-frequency anomaly locations. These locations are used to guide subsequent preventative key region segmentation of the current enhanced image.

[0055] By combining the two parts described above, the resulting segmented image set includes both newly discovered suspicious target regions within the current detection cycle and potential risk areas that require special attention based on historical experience. This segmented image set provides precise, narrowed-range visual analysis targets for subsequent secondary enhancement and specific identification.

[0056] Furthermore, secondary enhancement is performed to obtain an enhanced segmented image set and hazard factors, and specific identification is performed to obtain intelligent identification results for downhole anomalies, including: Based on the preliminary recognition results, the images in the segmented image set are enhanced a second time to obtain an enhanced segmented image set; Based on the image features and preliminary identification results of the enhanced segmented image set, risk factors are obtained, wherein the risk factors include anomaly type and anomaly risk. The anomaly type is obtained based on the preliminary identification results, and the anomaly risk is obtained based on the anomaly frequency of the nearest anomaly location in the enhanced segmented image. By combining the aforementioned risk factors, identification branch matching is completed and specific identification is performed to obtain intelligent identification results for downhole anomalies.

[0057] First, based on the preliminary recognition results, each image in the segmented image set undergoes secondary enhancement to obtain an enhanced segmented image set. Specifically, the secondary enhancement process targets local images focused on specific regions after segmentation. Since the segmented images may originate from different original backgrounds and possess varying degradation characteristics, the secondary enhancement needs to adaptively adjust the enhancement parameters according to the characteristics of each segmented image and the severity of anomalies indicated by the preliminary recognition results. The goal of the secondary enhancement is to improve the visibility and contrast of details in the target region while avoiding edge sharpening artifacts, noise amplification, or loss of true texture information due to overprocessing, thereby achieving a fine balance between enhancement effect and detail fidelity.

[0058] Secondly, based on the image features and preliminary identification results of the enhanced segmented image set, a hazard factor is calculated. This hazard factor is a comprehensive evaluation index comprising two dimensions: anomaly type and anomaly risk. The anomaly type is directly derived from the classification information of the preliminary identification results, such as equipment displacement, surface cracks, or liquid leaks. The quantification of anomaly risk is determined based on the historical frequency of anomalies in the corresponding region of the segmented image. Specifically, for segmented images originating from preliminary anomaly locations, the frequency of similar anomalies occurring at the same or adjacent locations in the recent period is retrieved from historical data based on their spatial location; for segmented images originating from high-frequency anomaly locations, the high-frequency anomaly frequency calculated in the historical analysis is directly used as the base value for its anomaly risk. By combining the anomaly type with the quantified anomaly risk, each enhanced segmented image is assigned a hazard factor characterizing its potential hazard level and urgency.

[0059] Finally, by combining risk factors, the identification branches are matched and specific identification is performed to obtain the final intelligent identification results of downhole anomalies.

[0060] Specifically, by combining the aforementioned risk factors, identification branch matching is completed and specific identification is performed to obtain intelligent identification results for downhole anomalies, including: A downhole anomaly identification model is obtained, wherein the downhole anomaly identification model includes multiple anomaly identification branches, and each anomaly identification branch is trained and obtained based on a set of historical anomaly images and a set of historical anomaly annotations of an anomaly type; The enhanced segmented image is input into the downhole anomaly identification model, and identification branch matching is performed based on the risk factors to perform specific identification and obtain specific identification results. The specific identification results include downhole anomaly type and downhole anomaly risk. Based on the aforementioned risk factors, the specific identification results are risk-corrected, and the downhole anomaly types and corrected downhole anomaly risks are integrated into intelligent downhole anomaly identification results.

[0061] First, a pre-trained downhole anomaly recognition model is acquired. This model is an integrated recognition architecture containing multiple independent and parallel anomaly recognition branches. Each branch is a sub-model specifically optimized for a particular anomaly category, such as longitudinal tearing of conveyor belts, hydraulic support leakage, roadway spalling, or personnel intrusion. The training of each branch is based on a historical set of anomaly images and corresponding historical anomaly annotations for its corresponding anomaly type. The training process allows each branch to deeply learn the unique visual patterns, texture features, and morphological context of the anomaly category it is responsible for, thereby enabling high-precision localization and classification of such anomalies.

[0062] Next, the enhanced segmented image is input into the trained downhole anomaly detection model. Based on the pre-calculated hazard factor of the enhanced segmented image, it is matched to the corresponding dedicated anomaly detection branch. For example, an image initially identified as potentially having a "surface crack" will be matched to the crack detection branch. Within the matched anomaly detection branch, a preliminary specific identification result can be calculated and output. This specific identification result includes two elements: first, a more precise downhole anomaly type confirmed by the model; and second, an initial downhole anomaly risk value calculated by the model based on the current image content. This downhole anomaly risk value typically reflects the salience, size, or severity of the anomaly in the image.

[0063] Furthermore, risk correction is applied to the specific identification results based on hazard factors. These hazard factors include anomalous risk information based on historical frequency. This historical risk information is then fused with the initial downhole anomaly risk value calculated by the model based on a single-frame image. Specifically, the corrected downhole anomaly risk value = initial downhole anomaly risk value + historical anomaly frequency. The historical anomaly frequency is a normalized anomaly occurrence frequency value obtained from historical data of that location. This formula indicates that the final corrected risk value is the sum of the initial risk determined by the model based on the current image and the inherent historical risk tendency of that location. If the historical anomaly frequency is high at a location, its corrected risk value will receive a clear increment from the initial value, reflecting the higher potential risk due to recurring problems at that location; if the historical frequency is low, the corrected risk value mainly depends on the initial determination result. Through the above calculations, a direct and effective fusion of real-time identification results and long-term statistical patterns is achieved, making the output risk value both immediate and trend-based, providing a more comprehensive quantitative basis for risk assessment.

[0064] Finally, the downhole anomaly types confirmed through specific identification are integrated with the revised downhole anomaly risk values, along with anomaly location information, to form a structured, final intelligent downhole anomaly identification result. This intelligent downhole anomaly identification result not only accurately identifies the specific type of the anomaly event and its spatial location, but also outputs a comprehensive risk quantification assessment that integrates real-time image recognition performance with long-term historical statistical trends. This intelligent downhole anomaly identification result can provide a comprehensive, objective, and reliable basis for subsequent implementation of tiered early warning systems, formulation of targeted inspection plans, and safety management decisions.

[0065] In summary, the embodiments of this application have at least the following technical effects: Compared to existing technologies, this invention dynamically calculates the dust occlusion coefficient by fusing real-time air sensor data and image features, and then adaptively matches image enhancement strategies based on this coefficient. This effectively overcomes the differential impact of dynamic changes in the downhole dust environment on image quality, improving the accuracy and adaptability of image enhancement in low-light, high-dust environments. Furthermore, by introducing historical image sets from the same region for comparative analysis, preliminary screening and location of abnormal areas are achieved. Preventive key segmentation is then performed by combining historical high-frequency anomaly information, enhancing the continuous monitoring capability of potential risk areas. In addition, a specific identification and risk correction mechanism based on hazard factors is adopted, combining real-time identification results with historical anomaly frequencies. This improves the classification accuracy of various anomalies, especially small targets and complex texture anomalies, while simultaneously achieving a quantitative assessment of anomaly risk levels. This provides more reliable, comprehensive, and timely intelligent early warning support for safe downhole production.

[0066] Example 2, as Figure 2 As shown, based on the same inventive concept as the image enhancement-based intelligent downhole anomaly identification method provided in Embodiment 1, this embodiment of the invention also provides an image enhancement-based intelligent downhole anomaly identification system, including: Image acquisition module 11 is used to acquire downhole images and extract image features from the downhole images; Image enhancement module 12 is used to obtain a dust obstruction coefficient based on the image features and air sensor data, enhance the downhole image, and obtain an enhanced image. The preliminary anomaly identification module 13 is used to perform preliminary anomaly identification based on the enhanced image and in combination with a set of historical images of the same area, and to obtain preliminary identification results. The specific identification module 14 is used to perform image segmentation based on the preliminary identification results and the historical image set of the same area, and to perform secondary enhancement to obtain the enhanced segmented image set and risk factors, and to perform specific identification to obtain the intelligent identification results of downhole anomalies.

[0067] The image acquisition module 11 is specifically used for: Specifically, acquiring downhole images and extracting image features from the downhole images includes: Acquire downhole images; The image features of the downhole image are extracted, including color features, texture features, and edge gradient features.

[0068] The image enhancement module 12 is specifically used for: Specifically, based on the image features and air sensor data, a dust obstruction coefficient is obtained, and image enhancement is performed on the downhole image to obtain an enhanced image, including: The image features are input into an image feature recognizer to obtain the image occlusion coefficient; The air sensor data of the area where the downhole image is located is obtained, and the dust occlusion coefficient is calculated by combining the image occlusion coefficient. Based on the dust obscuration coefficient, an image enhancement strategy is matched, and the image enhancement strategy is used to enhance the downhole image to obtain an enhanced image.

[0069] Specifically, the acquisition of the image feature recognizer includes: Obtain a historical image set, and obtain the image features of each historical image in the historical image set as a sample input set; The historical images are labeled with the degree of image occlusion to obtain a set of sample image occlusion coefficients, wherein the labeling of the degree of image occlusion is obtained based on the degree of image feature degradation; An image feature recognizer is constructed based on machine learning. The image feature recognizer is trained using the sample input set and the sample image occlusion coefficient set until convergence.

[0070] The preliminary anomaly identification module 13 is specifically used for: Specifically, based on the enhanced image and combined with a set of historical images of the same region, preliminary anomaly identification is performed to obtain preliminary identification results, including: Acquire historical image sets for the same area and determine the normal location range of the device; Based on the normal position range of the device, position anomaly identification is performed on the enhanced image to obtain the position anomaly identification result; Based on the image features of the historical image set of the same region and the enhanced image, feature statistical deviation identification is performed to obtain feature anomaly identification results. The preliminary identification result is obtained by combining the location anomaly identification result and the feature anomaly identification result.

[0071] Specifically, the specific identification module 14 is used for: Specifically, image segmentation is performed based on the preliminary identification results and a set of historical images from the same region, including: Based on the preliminary identification results, the preliminary anomaly location is obtained; Image segmentation is performed based on the initial anomaly location to obtain segmented images, which are then added to the segmented image set. Based on the high-frequency anomaly locations in the historical image set of the same region, the enhanced image is segmented according to the high-frequency anomaly locations to obtain segmented images, which are then added to the segmented image set.

[0072] Specifically, the locations of high-frequency anomalies are obtained, including: Obtain anomalies from the historical image set of the same region; Based on the frequency of occurrence of the aforementioned anomalies in the historical image set, the anomaly frequency is obtained; By combining the historical dust occlusion coefficients of the historical image set, an abnormal frequency threshold is obtained, and the abnormal situations are filtered to obtain high-frequency anomalies; Extract the location of the high-frequency anomaly and use it as the high-frequency anomaly location.

[0073] Furthermore, secondary enhancement is performed to obtain an enhanced segmented image set and hazard factors, and specific identification is performed to obtain intelligent identification results for downhole anomalies, including: Based on the preliminary recognition results, the images in the segmented image set are enhanced a second time to obtain an enhanced segmented image set; Based on the image features and preliminary identification results of the enhanced segmented image set, risk factors are obtained, wherein the risk factors include anomaly type and anomaly risk. The anomaly type is obtained based on the preliminary identification results, and the anomaly risk is obtained based on the anomaly frequency of the nearest anomaly location in the enhanced segmented image. By combining the aforementioned risk factors, identification branch matching is completed and specific identification is performed to obtain intelligent identification results for downhole anomalies.

[0074] Specifically, by combining the aforementioned risk factors, identification branch matching is completed and specific identification is performed to obtain intelligent identification results for downhole anomalies, including: A downhole anomaly identification model is obtained, wherein the downhole anomaly identification model includes multiple anomaly identification branches, and each anomaly identification branch is trained and obtained based on a set of historical anomaly images and a set of historical anomaly annotations of an anomaly type; The enhanced segmented image is input into the downhole anomaly identification model, and identification branch matching is performed based on the risk factors to perform specific identification and obtain specific identification results. The specific identification results include downhole anomaly type and downhole anomaly risk. Based on the aforementioned risk factors, the specific identification results are risk-corrected, and the downhole anomaly types and corrected downhole anomaly risks are integrated into intelligent downhole anomaly identification results.

Claims

1. A method for intelligent identification of downhole anomalies based on image enhancement, characterized in that, include: Acquire downhole images and extract image features from the downhole images; Based on the image features and air sensor data, the dust obstruction coefficient is obtained, and the downhole image is enhanced to obtain an enhanced image. Based on the enhanced image, combined with a set of historical images of the same area, preliminary anomaly identification is performed to obtain preliminary identification results; Based on the preliminary identification results, image segmentation is performed in conjunction with historical image sets of the same area, followed by secondary enhancement to obtain enhanced segmented image sets and risk factors. Specific identification is then performed to obtain intelligent identification results for downhole anomalies.

2. The intelligent downhole anomaly identification method based on image enhancement according to claim 1, characterized in that, Acquire downhole images and extract image features from the downhole images, including: Acquire downhole images; The image features of the downhole image are extracted, including color features, texture features, and edge gradient features.

3. The intelligent downhole anomaly identification method based on image enhancement according to claim 1, characterized in that, Based on the image features and air sensor data, a dust obstruction coefficient is obtained, and image enhancement is performed on the downhole image to obtain an enhanced image, including: The image features are input into an image feature recognizer to obtain the image occlusion coefficient; Obtain air sensor data of the area where the downhole image is located, and calculate the dust occlusion coefficient by combining the image occlusion coefficient; Based on the dust obscuration coefficient, an image enhancement strategy is matched, and the image enhancement strategy is used to enhance the downhole image to obtain an enhanced image.

4. The intelligent downhole anomaly identification method based on image enhancement according to claim 3, characterized in that, The acquisition of the image feature recognizer includes: Obtain a historical image set, and obtain the image features of each historical image in the historical image set as a sample input set; The historical images are labeled with the degree of image occlusion to obtain a set of sample image occlusion coefficients, wherein the labeling of the degree of image occlusion is obtained based on the degree of image feature degradation; An image feature recognizer is constructed based on machine learning. The image feature recognizer is trained using the sample input set and the sample image occlusion coefficient set until convergence.

5. The intelligent downhole anomaly identification method based on image enhancement according to claim 1, characterized in that, Based on the enhanced image, combined with a set of historical images of the same area, preliminary anomaly identification is performed to obtain preliminary identification results, including: Acquire historical image sets for the same area and determine the normal location range of the device; Based on the normal position range of the device, position anomaly identification is performed on the enhanced image to obtain the position anomaly identification result; Based on the image features of the historical image set of the same region and the enhanced image, feature statistical deviation identification is performed to obtain feature anomaly identification results. The preliminary identification result is obtained by combining the location anomaly identification result and the feature anomaly identification result.

6. The intelligent downhole anomaly identification method based on image enhancement according to claim 1, characterized in that, Image segmentation is performed based on the preliminary identification results and a set of historical images of the same region, including: Based on the preliminary identification results, the preliminary anomaly location is obtained; Image segmentation is performed based on the initial anomaly location to obtain segmented images, which are then added to the segmented image set. Based on the high-frequency anomaly locations in the historical image set of the same region, the enhanced image is segmented according to the high-frequency anomaly locations to obtain segmented images, which are then added to the segmented image set.

7. The intelligent downhole anomaly identification method based on image enhancement according to claim 6, characterized in that, Identify high-frequency anomaly locations, including: Obtain anomalies from the historical image set of the same region; Based on the frequency of occurrence of the aforementioned anomalies in the historical image set, the anomaly frequency is obtained; By combining the historical dust occlusion coefficients of the historical image set, an abnormal frequency threshold is obtained, and the abnormal situations are filtered to obtain high-frequency anomalies; Extract the location of the high-frequency anomaly and use it as the high-frequency anomaly location.

8. The intelligent downhole anomaly identification method based on image enhancement according to claim 1, characterized in that, Secondary enhancement is performed to obtain an enhanced segmented image set and hazard factors, and specific identification is performed to obtain intelligent downhole anomaly identification results, including: Based on the preliminary recognition results, the images in the segmented image set are enhanced a second time to obtain an enhanced segmented image set; Based on the image features and preliminary identification results of the enhanced segmented image set, risk factors are obtained, wherein the risk factors include anomaly type and anomaly risk. The anomaly type is obtained based on the preliminary identification results, and the anomaly risk is obtained based on the anomaly frequency of the nearest anomaly location in the enhanced segmented image. By combining the aforementioned risk factors, identification branch matching is completed and specific identification is performed to obtain intelligent identification results for downhole anomalies.

9. The intelligent downhole anomaly identification method based on image enhancement according to claim 1, characterized in that, Combining the aforementioned risk factors, the identification branch matching is completed and specific identification is performed to obtain intelligent identification results for downhole anomalies, including: A downhole anomaly identification model is obtained, wherein the downhole anomaly identification model includes multiple anomaly identification branches, and each anomaly identification branch is trained and obtained based on a set of historical anomaly images and a set of historical anomaly annotations of an anomaly type; The enhanced segmented image is input into the downhole anomaly identification model, and identification branch matching is performed based on the risk factors to perform specific identification and obtain specific identification results. The specific identification results include downhole anomaly type and downhole anomaly risk. Based on the aforementioned risk factors, the specific identification results are risk-corrected, and the downhole anomaly types and corrected downhole anomaly risks are integrated into intelligent downhole anomaly identification results.

10. An intelligent downhole anomaly identification system based on image enhancement, characterized in that: The method for performing the intelligent downhole anomaly identification method based on image enhancement as described in any one of claims 1-9 includes: An image acquisition module is used to acquire downhole images and extract image features from the downhole images; The image enhancement module is used to obtain a dust obstruction coefficient based on the image features and air sensor data, and to enhance the downhole image to obtain an enhanced image. The preliminary anomaly identification module is used to perform preliminary anomaly identification based on the enhanced image and in combination with a set of historical images of the same area, and to obtain preliminary identification results. The specific identification module is used to perform image segmentation based on the preliminary identification results and the historical image set of the same area, and to perform secondary enhancement to obtain the enhanced segmented image set and risk factors, and to perform specific identification to obtain the intelligent identification results of downhole anomalies.

Images