Deep well while-drilling block shedding real-time detection method based on DA-YOLO

Abstract

The application discloses a deep well while drilling block real-time detection method based on DA-YOLO and relates to the technical field of deep well while drilling detection. The method comprises the following steps: collecting solid particle images at the return outlet of a vibrating screen through a height setting camera device and adjusting the size, and then inputting the images into a feature extraction part of a DA-YOLO model; the part adopts a four-stage series module structure, integrates a double-domain strip attention mechanism and a deformable convolution, and realizes cooperative feature extraction and environmental interference suppression. Different levels of extracted features are input into a feature fusion part, multi-scale feature fusion is completed through up-sampling and convolution operation, and then target detection and positioning are performed through three different size decoupling heads to generate preliminary reasoning results. A non-maximum suppression algorithm is used to calculate the overlap degree of the boundary box and combined with a confidence threshold to screen effective detection results. Through modular collaborative design, the method improves the detection accuracy of low contrast and small size block under complex environment, and takes into account the multi-scale adaptability and real-time performance.

Description

Technical Field

[0001] This invention relates to the field of deep well drilling inspection technology, and in particular to a real-time detection method for block dropping in deep wells based on DA-YOLO. Background Technology

[0002] In deep well drilling, real-time monitoring of rock spalling during drilling plays a crucial role in ensuring drilling safety and optimizing construction parameters. During drilling operations, formation rocks are prone to spalling due to drilling fluid erosion and mechanical vibration. The resulting spalled rocks are returned to the surface via a vibrating screen, and their quantity and size changes directly reflect formation stability. Traditional spalling detection relies on manual visual observation, which is affected by factors such as ambient light and personnel experience, resulting in detection lag and insufficient accuracy, making it difficult to meet the real-time monitoring needs of deep wells and complex formations. As drilling technology advances to deeper and ultra-deeper formations, formation conditions become increasingly complex, significantly increasing the risks of stuck pipe and wellbore collapse caused by spalling. There is an urgent need to develop efficient and accurate automated spalling detection technology to achieve timely early warning of formation anomalies during drilling and provide data support for drilling safety decisions.

[0003] Existing automated block detection technologies have two prominent drawbacks: First, insufficient feature extraction capabilities. The vibrating screen area in deep well drilling sites is easily affected by environmental interference such as water mist, dust, and changes in light. Traditional detection models struggle to effectively separate the block target from background noise and have limited feature capture capabilities for low-contrast, small-sized blocks, resulting in high rates of missed and false detections. Second, it is difficult to balance multi-scale adaptability with real-time performance. Block sizes vary widely, from micrometers to centimeters. Existing models either improve multi-scale detection capabilities through complex network structures, but this reduces inference speed and fails to meet the time requirements of real-time monitoring while drilling. Alternatively, they simplify the model structure in pursuit of real-time performance, leading to uneven detection accuracy for blocks of different sizes and making it difficult to adapt to the dynamically changing block detection scenarios during deep well drilling. Summary of the Invention

[0004] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a real-time detection method for block falling during drilling in deep wells based on DA-YOLO.

[0005] The technical solution adopted in this invention is a real-time detection method for block falling during drilling in deep wells based on DA-YOLO, comprising the following steps: S1, acquiring image data of solid particles flowing with drilling fluid at the return outlet of a vibrating screen using a camera device at a set height, adjusting the acquired image data to a fixed size, and then inputting it into the backbone network of the DA-YOLO model; S2, using a four-stage feature extraction structure consisting of a Conv module, a C3k2_DSAM module, a SPPF module, and a C2PSA_DLKA module connected in series in the backbone network to extract features from the input image, wherein the C3k2_DSAM module performs feature extraction and dehazing collaborative processing by embedding a dual-domain strip attention mechanism, and the C2PSA_DLKA module... S3: The sampling position and receptive field are adaptively adjusted by superimposing deformable convolutions; S4: The shallow, middle and deep features output from the backbone network are input into the neck network part, and multi-scale feature fusion is performed through upsampling and convolution operations to construct a multi-dimensional feature representation; S5: The three different sizes of decoupling heads set in the head network part are used to perform target detection and localization on the fused features, generating preliminary inference results including bounding box coordinates, confidence and class probability; S6: The non-maximum suppression algorithm is used to filter redundant bounding boxes in the preliminary inference results, and the effective detection results are selected by calculating the bounding box overlap and combining it with the confidence threshold; S7: Based on the filtered detection results, the missing block target is identified and the number is counted, and the final result of missing block detection is output.

[0006] Furthermore, the execution process of the dual-domain stripe attention mechanism of the C3k2_DSAM module in S2 satisfies: , , in, Indicates the input feature map of the module. Indicates the spatial size of the feature map. This represents the feature map after frequency domain processing. Indicates the size is Frequency strip attention unit, Indicates the size is Spatial strip attention unit, express Convolution operation, Split represents channel splitting operation. This represents the sub-feature maps after splitting. This indicates the final output feature map of the module.

[0007] Furthermore, the feature processing of the frequency strip attention unit in S2 satisfies: , , in, This indicates a strip average pooling operation. Indicates The central strip-shaped region features Indicates low-frequency components, Represents high-frequency components, and These represent the channel attention parameters for low-frequency and high-frequency components, respectively. Note the horizontal operator. This indicates the vertical attention operator. This indicates the output result of the frequency strip attention unit.

[0008] Furthermore, the feature aggregation of the spatial strip attention unit in S2 satisfies: , in, This indicates a global average pooling operation. express In convolution operations, sigmoid represents the activation function. "*" indicates lightweight attention weights, and "*" indicates a convolutional information aggregation method. Note the operators for representing horizontal stripes. Note the operators for representing vertical stripes. This indicates the output result of the spatial stripe attention cell. The input feature map of the unit is represented.

[0009] Furthermore, the deformable big kernel attention mechanism of the C2PSA_DLKA module in S2 satisfies: , Attention , Deformable_LKA Attention , Output _LKA , in, Indicates the input feature map of the module. This represents the activation function. express Convolution operation, This represents deformable depthwise convolution. This represents a deformable depthwise dilated convolution, and Attention represents an attention map. "Indicates element-wise multiplication operation, _ represents the intermediate processing feature map, Output represents the module output feature map, and Deformable_LKA represents the deformable large kernel attention.

[0010] Furthermore, the bounding box selection of the non-maximum suppression algorithm in S5 satisfies: , in, Indicates the predicted bounding box location. Indicates the actual bounding box location. Indicates the area of ​​the overlapping region of the bounding boxes. Indicates the area of ​​the bounding box's joint region. Indicates intersection, union, and ratio. This indicates the probability that a target exists within the predicted bounding box, and Confidence_Score represents the confidence score.

[0011] Further, step S3 includes the following sub-steps: S31, receiving shallow feature maps, mid-level feature maps, and deep feature maps output by the backbone network, each feature map corresponding to different spatial resolutions and channel dimensions; S32, performing an upsampling operation on the deep feature map to make its spatial resolution consistent with that of the mid-level feature map, while adjusting the number of channels to the same dimension through a 1×1 convolution; S33, fusing the adjusted deep feature map and the mid-level feature map element-wise to generate a first fused feature map, and then performing upsampling and channel adjustment operations on the first fused feature map; S34, fusing the adjusted first fused feature map and the shallow feature map element-wise, refining the features through a convolution operation, and outputting the final fused multi-scale feature map.

[0012] Further, step S4 includes the following sub-steps: S41, inputting the fused multi-scale feature maps into the large-size decoupling head, medium-size decoupling head, and small-size decoupling head of the head network, each decoupling head corresponding to a different receptive field range; S42, each decoupling head performs feature mapping on the input feature map through a branch convolutional network, generating bounding box regression branch features, confidence prediction branch features, and category prediction branch features respectively; S43, performing decoding operations on the bounding box regression branch features to obtain the center coordinates, width, and height parameters of the bounding box; S44, performing activation function operations on the confidence prediction branch features and the category prediction branch features respectively to obtain the confidence of the target's existence and the probability distribution of the corresponding category, and integrating the results of each branch to form a preliminary inference result.

[0013] Further, S5 includes the following sub-steps: S51, sorting the preliminary inference results from high to low confidence scores, and selecting the bounding box with the highest confidence score as the initial candidate box; S52, calculating the intersection-union ratio (IU) of the initial candidate box and all remaining bounding boxes, and marking the bounding boxes with IU ratios greater than a set threshold as redundant bounding boxes; S53, removing the marked redundant bounding boxes, and retaining the initial candidate boxes as valid detection results; S54, taking the bounding box with the highest confidence score among the remaining unprocessed bounding boxes as the new initial candidate box, repeating the IU calculation and redundancy removal process until all bounding boxes have been processed, and selecting the final valid detection results.

[0014] A real-time detection method for block falling during drilling in deep wells based on DA-YOLO is implemented through different units, including: an image data acquisition and size adaptation unit, which captures images of solid particles at the return outlet of a vibrating screen using a camera at a preset height, adjusts the image to a fixed size, and then transmits it to a feature extraction unit; a multi-stage feature extraction and enhancement unit, composed of a Conv module, a C3k2_DSAM module, a SPPF module, and a C2PSA_DLKA module connected in series, which receives the adapted image data, performs feature extraction, dehazing, and background interference suppression operations, and outputs a multi-scale feature map; and a feature fusion and dimensionality adjustment unit, which receives the output from the multi-stage feature extraction and enhancement unit. Different levels of feature maps are upsampled, convolutional, and element-wise fusion operations to construct a unified fusion feature map. The target detection and localization unit sets three different sized decoupling heads to perform branch operations on the fusion feature map, generating preliminary detection results including bounding box parameters, confidence scores, and class probabilities. The redundant bounding box filtering and optimization unit uses a non-maximum suppression algorithm to remove redundant bounding boxes through intersection-union ratio calculation and confidence threshold judgment, and outputs optimized detection results. The missing block identification and quantity counting unit receives the optimized detection results, identifies missing block targets, counts their number, and outputs a final detection report. Each unit sequentially interacts with signals and data bidirectionally through a data transmission interface.

[0015] Beneficial Effects: This invention proposes a real-time block detection method for deep well drilling based on DA-YOLO. Through a feature extraction network with a specific structure, it integrates a dual-domain strip attention mechanism and deformable convolution to achieve synergistic feature extraction and environmental interference suppression. This accurately separates the block target from background noise such as water mist and dust, significantly improving the feature capture capability for low-contrast, small-sized blocks and solving the problems of missed detections and false detections caused by insufficient feature extraction in traditional models. By designing multi-scale feature fusion and three different-sized decoupling heads, a multi-dimensional feature representation is constructed to adapt to the detection needs of different-sized blocks. Simultaneously, the computational logic of the core module is simplified, maintaining efficient inference speed while ensuring multi-scale adaptability, breaking the dilemma of balancing real-time performance and detection accuracy in existing technologies. Furthermore, the bounding box filtering process is optimized through a non-maximum suppression algorithm, further improving the accuracy of detection results. Combined with a complete process of image acquisition, feature processing, target localization, result filtering, and statistical output, the entire block detection process is automated and accurate, providing reliable technical support for deep well drilling formation stability monitoring and meeting the real-time monitoring needs in complex drilling scenarios. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a flowchart of method step S3 of the present invention; Figure 3 This is a flowchart of method step S4 of the present invention; Figure 4 This is a flowchart of step S5 of the method of the present invention; Figure 5 This is a diagram showing the unit composition for implementing the method of the present invention. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] like Figure 1 As shown, the real-time detection method for block falling during drilling in deep wells based on DA-YOLO includes the following steps: S1: Acquire image data of solid particles flowing with drilling fluid at the return outlet of the vibrating screen using a camera device at a set height. After adjusting the acquired image data to a fixed size, input it into the backbone network of the DA-YOLO model. Specifically, step S1 serves as the initial data input stage of the detection process, acquiring standardized target image data to support subsequent model calculations. During implementation, an industrial-grade high-definition camera with a resolution of at least 3840×2160 pixels is selected and installed 1.2 to 1.5 meters directly above the return outlet of the vibrating screen. This height range ensures complete coverage of the return outlet width and avoids contamination of the lens by drilling fluid splashes. The camera uses a continuous shooting mode with a frame rate of 30 frames per second, simultaneously enabling automatic exposure and white balance adjustment. The exposure time is set to 1 / 500 second, and the ISO sensitivity is controlled within the range of 100-400 to adapt to changes in lighting at the drilling site and reduce motion blur. The acquired raw image data is transmitted to the backend processing system via an image acquisition card and then adjusted to a fixed size of 640×640 pixels using an image scaling algorithm. This size ensures the integrity of feature information while matching the input requirements of the DA-YOLO model, reducing computational complexity. During the size adjustment process, bilinear interpolation is used to maintain the continuity of image pixel distribution and avoid stretching distortion. The adjusted image data is stored in RGB three-channel format. After the pixel values ​​are normalized to the 0-1 range, they are input to the backbone network of the DA-YOLO model through the data interface to provide a standardized and high-quality input data source for subsequent feature extraction. The parameter settings in this step directly determine the effectiveness of subsequent feature extraction and the accuracy of detection results.

[0019] S2 utilizes a four-stage feature extraction structure consisting of the Conv module, C3k2_DSAM module, SPPF module, and C2PSA_DLKA module connected in sequence in the backbone network to extract features from the input image. Among them, the C3k2_DSAM module performs feature extraction and dehazing collaborative processing by embedding a dual-domain strip attention mechanism, and the C2PSA_DLKA module adaptively adjusts the sampling position and receptive field by superimposing deformable convolutions. Specifically, step S2 is the feature extraction stage, which uses a four-stage cascaded module structure to achieve deep feature mining and optimization of the input image. The first stage of the backbone network is the Conv module, which uses 64 3×3 convolutional kernels with a stride of 2 to perform preliminary convolution operations on the input 640×640 pixel image, outputting a 320×320 pixel, 64-channel feature map, completing image dimensionality compression and basic feature extraction. The second stage, the C3k2_DSAM module, has three parallel feature processing branches. By embedding a dual-domain strip attention mechanism, it enhances the frequency domain and spatial domain features respectively. The frequency domain processing uses strip sizes of 7, 11, and matching the aspect ratio of the feature map, while the spatial domain processing uses strips of the same size. This ensures that the target features are extracted simultaneously. The system first achieves adaptive suppression of water mist interference at the drilling site, outputting a 320×320 pixel, 128-channel feature map. The third stage, the SPPF module, uses a 5×5 pooling kernel and further compresses the feature map dimension to 160×160 pixels through parallel pooling and convolution fusion operations, maintaining the 128 channels and improving the global correlation of features. The fourth stage, the C2PSA_DLKA module, overlays deformable convolutional layers with 256 kernels. By adaptively adjusting the sampling position and receptive field, it accurately captures the morphological features and positional information of the dropped target, outputting a 160×160 pixel, 256-channel deep feature map. This feature extraction structure, formed by the sequential connection of these four modules, achieves a progressive approach from basic texture to advanced semantic features. The synergistic effect of the dual-domain strip attention mechanism and deformable convolution effectively improves the discriminative power of target features in complex environments.

[0020] S3 inputs the shallow, middle and deep features output from the backbone network into the neck network part, and performs multi-scale feature fusion through upsampling and convolution operations to construct a multi-dimensional feature representation; Specifically, step S3 constructs a comprehensive feature representation through multi-scale feature fusion, providing rich feature support for subsequent object detection. The neck network receives three layers of feature maps output from the backbone network. The shallow feature map is 320×320 pixels with 64 channels, focusing on image detail and texture features, such as chipped edges and surface textures. The middle feature map is 320×320 pixels with 128 channels, including target contours and local structural features. The deep feature map is 160×160 pixels with 256 channels, containing target category and global semantic features. The fusion process first performs a 2x upsampling operation on the deep feature map, then uses a transposed convolution algorithm to enlarge its size to 320×320 pixels, while adjusting the number of channels to 128 using a 1×1 convolution kernel to match the channel dimension of the mid-layer feature map. Next, the adjusted deep and mid-layer feature maps are fused element-wise with a fusion weight distribution of 0.6:0.4, generating a first fused feature map. This feature map retains both deep semantic information and incorporates mid-layer structural features. Then, the first fused feature map is upsampled a second time by 2x, adjusting its size to 640×640 pixels, and reducing the number of channels to 64 using a 1×1 convolution to match the shallow feature map. Finally, the adjusted first fused feature map and the shallow feature map are added element-wise with a weight ratio of 0.5:0.5, and then refined using a 3×3 convolution kernel to filter redundant information, outputting a final fused feature map of 640×640 pixels and 64 channels. The entire fusion process achieves complementary fusion of features at different levels through two upsampling operations and two convolutional adjustments, constructing a multi-dimensional and comprehensive feature representation, and providing sufficient feature basis for the detection of blocks of different sizes.

[0021] S4 uses three different sizes of decoupling heads set in the head network to perform target detection and localization on the fused features, generating preliminary inference results including bounding box coordinates, confidence scores and class probabilities; Specifically, step S4 uses multi-size decoupling heads to achieve accurate detection and localization of fused features, generating preliminary inference results. The head network uses three different sizes of decoupling heads: large, medium, and small, corresponding to feature map inputs of 640×640, 320×320, and 160×160 pixels, respectively. The large-size decoupling head is equipped with 128 3×3 convolutional kernels, with a receptive field of 64×64 pixels, specifically designed for detecting larger blocks (diameter greater than 5 cm). The medium-size decoupling head is equipped with 64 3×3 convolutional kernels, with a receptive field of 32×32 pixels, targeting medium-sized blocks (diameter 2-5 cm). The small-size decoupling head is equipped with 32 3×3 convolutional kernels, with a receptive field of 16×16 pixels, focusing on the recognition of small blocks (diameter less than 2 cm). Each decoupling head includes three independent branches: bounding box regression, confidence prediction, and class prediction. The branch convolutional networks use a 3-layer stacked convolutional layer structure, with a linear rectified function as the activation function. The bounding box regression branch outputs four parameters—center x-coordinate, center y-coordinate, width, and height—of the target bounding box through convolution operations on the fused features. The coordinate values ​​are normalized based on the feature map size. The confidence prediction branch outputs the probability of the target's presence using a sigmoid activation function, with a value ranging from 0 to 1. The class prediction branch uses a softmax activation function to output the probability distributions for the missing block and non-missing block classes. The outputs of these three branches are integrated to form a preliminary inference result including bounding box coordinates, confidence scores, and class probabilities. Each feature map pixel corresponds to the detection information of a potential target. The synergistic effect of the three size decoupling heads achieves full coverage detection of missing blocks of different sizes.

[0022] S5. The non-maximum suppression algorithm is used to filter redundant bounding boxes in the preliminary inference results. The effective detection results are filtered by calculating the bounding box overlap and combining it with the confidence threshold. Specifically, step S5 uses a non-maximum suppression algorithm to filter effective detection results, remove redundant bounding boxes, and improve detection accuracy. In implementation, the preliminary inference results are first sorted from highest to lowest confidence score, with a confidence score threshold of 0.5. Detection results below this threshold are discarded, and only preliminary candidate boxes with a confidence score ≥ 0.5 are retained. The first bounding box in the sorted list is selected as the initial candidate box. The intersection-union ratio (IUR) of this candidate box with all remaining candidate boxes is calculated. The IUR is calculated as the ratio of the area of ​​the overlapping region to the area of ​​the joint region of the bounding boxes. The IUR threshold is set to 0.35. Bounding boxes with an IUR greater than this threshold are marked as redundant and removed, retaining the initial candidate boxes as effective detection results. Subsequently, the bounding box with the highest confidence score is selected from the remaining unprocessed bounding boxes as the new initial candidate box, and the above IUR calculation and redundancy removal process is repeated until all candidate boxes have been processed. In this process, the confidence threshold and intersection-over-union (IoU) threshold were verified through multiple sets of experiments. A confidence threshold of 0.5 effectively filters out most false detection results, while an IoU threshold of 0.35 can retain true targets while avoiding duplicate detections. For multiple bounding boxes generated by the same missing block target, the algorithm selects the bounding box with the highest confidence and most accurate location as the final detection result, and removes the remaining redundant boxes with high overlap. This step can control the redundant bounding box removal rate in the initial inference results to over 85%, significantly improving the accuracy and simplicity of the detection results and providing reliable data support for subsequent missing block identification and statistics.

[0023] S6, based on the filtered detection results, performs target identification and quantity statistics for dropped blocks, and outputs the final result of the dropped block detection.

[0024] Specifically, step S6 identifies and counts the number of falling blocks based on the filtered detection results, and outputs the final detection report. First, the filtered valid detection results are categorized. Based on the category probability distribution, the target type corresponding to each bounding box is determined. When the probability value of the falling block category is greater than 0.5, the target is determined to be a falling block; if the probability value of a non-falling block category is greater than 0.5, it is determined to be a non-falling block target and excluded. For targets determined to be falling blocks, their actual size is calculated based on the bounding box coordinate information. By using the calibration ratio between pixel size and actual size (1 pixel corresponds to 0.1 mm), the width and height pixel values ​​of the bounding box are converted into actual physical sizes. Simultaneously, the center position coordinates of each falling block are recorded (based on a coordinate system established on the return outlet plane of the vibrating screen). Then, the falling blocks are classified and counted according to size range, into four categories: micro falling blocks (diameter < 1 cm), small falling blocks (1-3 cm), medium falling blocks (3-5 cm), and large falling blocks (> 5 cm). The number and proportion of each category are then counted. Finally, data such as the number, size distribution, and location of the fallen blocks are integrated to generate a standardized inspection report. The report includes real-time inspection results and historical data trend charts, and is synchronized to the drilling monitoring center via a data transmission interface with a transmission delay controlled within 0.5 seconds. This step transforms inspection results into practical information, providing drilling engineers with intuitive and accurate fallen block monitoring data to support formation stability assessments and drilling parameter adjustment decisions, ensuring the safe operation of deep well drilling.

[0025] Preferably, the execution process of the dual-domain strip attention mechanism of the C3k2_DSAM module in S2 satisfies: , in, Indicates the input feature map of the module. Indicates the spatial size of the feature map. This represents the feature map after frequency domain processing. Indicates the size is Frequency strip attention unit, Indicates the size is Spatial strip attention unit, express Convolution operation, Split represents channel splitting operation. This represents the sub-feature maps after splitting. This indicates the final output feature map of the module.

[0026] Specifically, the execution process of the dual-domain strip attention mechanism in the C3k2_DSAM module in step S2 is specifically defined. This mechanism achieves simultaneous optimization of feature extraction and dehazing through the collaborative processing of the frequency domain and spatial domain. During implementation, the module first receives the feature map output by the preceding Conv module. This feature map includes specific channel number and spatial size information. It then processes the feature map in parallel using three frequency strip attention units of different sizes. The sizes of the three units are set to 7, 11, and a value matching the aspect ratio of the feature map, respectively, enabling comprehensive capture of feature information across different frequency ranges, especially low-frequency noise corresponding to water mist interference at the drilling site and high-frequency signals corresponding to falling blocks. The outputs of the three frequency strip attention units are added element-wise, and then the channel dimensions are integrated through a 1×1 convolution operation to obtain the feature map after frequency domain processing, completing the initial suppression of water mist interference. Subsequently, through channel splitting, the feature map is divided into two sub-feature maps, which are then input into two spatial strip attention units. These units are also set to sizes 7 and 11, corresponding to the frequency strip attention units. This precise enhancement of spatial domain features highlights the spatial location and morphological characteristics of the falling block target. Finally, the outputs of the two spatial strip attention units are concatenated along the channel dimension to form the final feature map output by the module. This feature map retains the key features of the falling block target while effectively filtering out water mist interference, providing high-quality input for subsequent deep feature extraction. This mechanism improves the signal-to-noise ratio of the feature map by more than 30%, significantly enhancing the ability to distinguish target features in complex environments.

[0027] Preferably, the feature processing of the frequency strip attention unit in S2 satisfies: , , , in, This indicates a strip average pooling operation. Indicates The central strip-shaped region features Indicates low-frequency components, Represents high-frequency components, and These represent the channel attention parameters for low-frequency and high-frequency components, respectively. Note the horizontal operator. This indicates the vertical attention operator. This indicates the output result of the frequency strip attention unit.

[0028] Specifically, the feature processing logic of the frequency strip attention unit in step S2 improves the extraction accuracy of target features by separating and enhancing the frequency components of the feature map. During implementation, the unit first receives the input feature map. For each pixel in the feature map, a strip-shaped feature region is defined centered on that pixel. The width and length of the strip-shaped region are adaptively adjusted according to the spatial dimensions of the feature map to ensure coverage of key frequency information surrounding the pixel. The defined strip-shaped feature region is processed using strip average pooling to separate the low-frequency components, which mainly correspond to environmental interference information such as water mist and dust. Then, the high-frequency components are obtained by calculating the difference between the original feature map and the low-frequency components. These high-frequency components concentrate on the detailed features and edge information of the dropped target. Subsequently, two sets of independent channel-based attention parameters are introduced to assign weights to the low-frequency and high-frequency components respectively. The attention parameter weights for the low-frequency components are set to 0.3-0.4, and the attention parameter weights for the high-frequency components are set to 0.6-0.7. This weight adjustment achieves focused enhancement of target features and suppression of interference information. The weighted low-frequency components and high-frequency components are added element-wise to obtain a preliminary feature map. Then, the feature map is processed sequentially by horizontal and vertical attention operators to enhance the texture information in the horizontal and vertical directions, respectively. Finally, the processed feature map is residually connected to the original input feature map, which retains the original feature information while incorporating the optimized frequency features. This process increases the high-frequency feature intensity of the dropped target by more than 40%, effectively solving the problem of target feature blurring in low-contrast environments.

[0029] Preferably, the feature aggregation of the spatial strip attention units in S2 satisfies: , in, This indicates a global average pooling operation. express In convolution operations, sigmoid represents the activation function. "*" indicates lightweight attention weights, and "*" indicates a convolutional information aggregation method. Note the operators for representing horizontal stripes. Note the operators for representing vertical stripes. This indicates the output result of the spatial stripe attention cell. The input feature map of the unit is represented.

[0030] Specifically, the feature aggregation method of the spatial strip attention unit in step S2 improves the spatial correlation of features and the accuracy of target localization through a lightweight attention mechanism and spatial dimension enhancement. In implementation, the unit first receives the input feature map. First, it compresses the spatial dimension of the feature map through global average pooling to obtain global feature information in the channel dimension. This operation integrates the global context of the feature map, avoiding the one-sidedness of local information. Then, the global feature information is input into a 1×1 convolutional layer, which includes convolutional kernels equal to the number of channels in the feature map. Convolutional operations transform the dimensions of the global features and integrate information. After processing with the Sigmoid activation function, lightweight attention weights consistent with the number of channels in the feature map are generated. These weights range from 0 to 1, accurately reflecting the importance of each channel feature. The weights for the feature channels corresponding to the dropped target are concentrated in the range of 0.7-0.9, while the weights for the background interference channels are concentrated in the range of 0.1-0.3. Through convolutional information aggregation, the generated attention weights are fused with the original input feature map channel by channel, strengthening the target feature channels while suppressing interfering feature channels. Subsequently, the fused feature map is processed sequentially using horizontal and vertical stripe attention operators. The horizontal stripe attention operator focuses on the horizontal texture of the feature map, while the vertical stripe attention operator enhances the vertical structure. The processing window size of both operators is dynamically adjusted according to the spatial size of the feature map to ensure complete coverage of the target's spatial range. Finally, the processed feature map is residually concatenated with the original input feature map to maintain the gradient continuity of the feature map. This aggregation method improves the spatial feature response intensity of the feature map by more than 25%, significantly enhancing the spatial localization accuracy of the dropped target.

[0031] Preferably, the deformable big kernel attention mechanism of the C2PSA_DLKA module in S2 satisfies: , Attention Deformable_LKA Attention , Output _LKA , in, Indicates the input feature map of the module. This represents the activation function. express Convolution operation, This represents deformable depthwise convolution. This represents a deformable depthwise dilated convolution, and Attention represents an attention map. "Indicates element-wise multiplication operation, _ represents the intermediate processing feature map, Output represents the module output feature map, and Deformable_LKA represents the deformable large kernel attention.

[0032] Specifically, in step S2, the deformable large kernel attention mechanism of the C2PSA_DLKA module achieves adaptive capture of target features through the combination of deformable convolution and attention map. During implementation, the module first receives the feature map output from the preceding SPPF module. This feature map has a certain number of channels and spatial size. It first adjusts the channel dimension through a 1×1 convolution operation to adapt the number of channels to the subsequent processing requirements. Then, it undergoes a non-linear transformation using the GELU activation function to enhance the expressive power of the features, resulting in an intermediate processed feature map. Subsequently, deformable depthwise convolution is performed on this intermediate feature map. The number of convolution kernels in the deformable depthwise convolution is consistent with the number of channels in the intermediate feature map, and the kernel size is set to 3×3. By adaptively adjusting the sampling position of the convolution kernel, the deformable shape and irregular edges of the dropped target are accurately captured, effectively handling the diverse shape changes of the dropped target. Next, deformable depthwise dilation convolution is performed on the output of the deformable depthwise convolution. The dilation rate is set to 2-3, which expands the receptive field of the convolution while maintaining the spatial size of the feature map, comprehensively covering the global features of the dropped target. The output of the deformable depthwise dilated convolution is channel-integrated through a 1×1 convolution operation to generate an attention map. This map accurately marks the position and importance of the dropped target in the feature map. Element-wise multiplication is then used to weightedly fuse the attention map with the intermediate processed feature map, enhancing the target region features while suppressing background interference. Finally, a 1×1 convolution operation is used to adjust the channels of the fused feature map to match the number of channels in the module's input feature map, and a residual connection is established with the original input feature map. This mechanism improves the module's feature capture accuracy for dropped targets of different shapes by more than 35%, significantly enhancing the model's adaptability to changes in target shape.

[0033] Preferably, the bounding box selection of the non-maximum suppression algorithm in S5 satisfies: , in, Indicates the predicted bounding box location. Indicates the actual bounding box location. Indicates the area of ​​the overlapping region of the bounding boxes. Indicates the area of ​​the bounding box's joint region. Indicates intersection, union, and ratio. This indicates the probability that a target exists within the predicted bounding box, and Confidence_Score represents the confidence score.

[0034] Specifically, in step S5, the bounding box selection criteria of the non-maximum suppression algorithm achieve accurate removal of redundant bounding boxes through intersection-union ratio (IU) calculation and confidence score evaluation. During implementation, the algorithm first obtains the preliminary inference results output by the head network, including the location information of each predicted bounding box, the probability of a target being present, and the class probability. For any two predicted bounding boxes, their respective position coordinates are determined, including the pixel coordinates of the top-left and bottom-right corners. The overlapping area and the joint area of ​​the two bounding boxes are calculated using these coordinates. The overlapping area is the total number of pixels in the intersection of the two bounding boxes, and the joint area is the total number of pixels in the union of the two bounding boxes. The ratio of these two is the IU, which quantifies the degree of overlap between the two bounding boxes. The value ranges from 0 to 1; the closer the ratio is to 1, the higher the degree of overlap. Simultaneously, combining the probability of a target being present in the predicted bounding box with the IU, a confidence score is obtained through multiplication. This score comprehensively reflects the accuracy and uniqueness of the bounding box; a higher score indicates a greater likelihood that the bounding box corresponds to a real target. During the screening process, an intersection-union (IU) threshold of 0.35 is first set. When the IU of two bounding boxes exceeds this threshold, they are considered to have excessive overlap, and the bounding box with the lower confidence score is marked as a redundant bounding box. A confidence score threshold of 0.5 is then set, and bounding boxes with scores below this threshold are directly identified as false detections and removed. This dual-threshold screening method accurately retains the bounding boxes corresponding to the true targets while eliminating redundant and false detections. This screening criterion achieves a redundancy removal rate of over 85% while maintaining a true target retention rate of over 98%, significantly improving the accuracy and reliability of the detection results and providing accurate data support for subsequent block drop identification and statistics.

[0035] Preferred, such as Figure 2 As shown, step S3 includes the following sub-steps: S31, receiving shallow feature maps, mid-level feature maps, and deep feature maps output by the backbone network, each feature map corresponding to different spatial resolutions and channel dimensions; S32, performing an upsampling operation on the deep feature map to make its spatial resolution consistent with that of the mid-level feature map, and simultaneously adjusting the number of channels to the same dimension through a 1×1 convolution; S33, fusing the adjusted deep feature map and the mid-level feature map element-wise to generate a first fused feature map, and then performing upsampling and channel adjustment operations on the first fused feature map; S34, fusing the adjusted first fused feature map and the shallow feature map element-wise, refining the features through a convolution operation, and outputting the final fused multi-scale feature map.

[0036] Specifically, the multi-scale feature fusion process in step S3 is broken down into steps to ensure effective complementarity and integration of features at different levels. S31 first receives three layers of feature maps output from the backbone network. The shallow feature map corresponds to the output of the Conv module, with a spatial resolution of 320×320 and 64 channels, focusing on basic details such as image edges and textures. The middle-layer feature map comes from the C3k2_DSAM module, maintaining a spatial resolution of 320×320 but increasing the number of channels to 128, including local target structure and contour information. The deep feature map is output from the C2PSA_DLKA module, with a spatial resolution compressed to 160×160 and 256 channels, containing target category and global semantic features. The dimensional differences among the three layers of feature maps provide a foundation for multi-scale fusion. S32 performs a 2x upsampling operation on the deep feature map, using a transposed convolution algorithm to enlarge its spatial resolution from 160×160 to 320×320, matching the resolution of the mid-level feature map. Simultaneously, a 1×1 convolutional layer is used to adjust the number of channels from 256 to 128, achieving channel dimension matching and eliminating dimensionality barriers for feature fusion. S33 fuses the adjusted deep and mid-level feature maps element-wise, with the weights of both layers set to 0.5 to ensure a balanced preservation of semantic and structural features, generating a first fused feature map that combines global and local information. This fused feature map is then repeatedly upsampled and adjusted in channel size, increasing the resolution to 640×640 and reducing the number of channels to 64. S34 performs element-wise addition and fusion of the first fused feature map, which has undergone two adjustments, with the shallow feature map. It also adopts the same equal weight allocation strategy. After fusion, the feature is refined through a 3×3 convolutional layer with a convolution stride of 1 and the padding method is "same". Redundant information is filtered and effective features are enhanced. Finally, a multi-scale fused feature map with a resolution of 640×640 and 64 channels is output. This step-by-step process realizes comprehensive feature integration from deep semantics to shallow details, providing sufficient feature support for block detection of different sizes.

[0037] Preferred, such as Figure 3 As shown, step S4 includes the following sub-steps: S41, the fused multi-scale feature maps are input into the large-size decoupling head, medium-size decoupling head, and small-size decoupling head of the head network, respectively, with each decoupling head corresponding to a different receptive field range; S42, each decoupling head performs feature mapping on the input feature map through a branch convolutional network, generating bounding box regression branch features, confidence prediction branch features, and category prediction branch features respectively; S43, the bounding box regression branch features are decoded to obtain the center coordinates, width, and height parameters of the bounding box; S44, activation function operations are performed on the confidence prediction branch features and the category prediction branch features respectively to obtain the confidence of the target's existence and the probability distribution of the corresponding category, and the results of each branch are integrated to form a preliminary inference result.

[0038] Specifically, the target detection and localization logic in step S4 is implemented step by step, achieving accurate detection through multi-size decoupling heads and branch operations. In step S31, the 640×640 resolution multi-scale fused feature map output from S3 is input into three different sized decoupling heads in the head network: large, medium, and small. The large-size decoupling head is adapted to the 640×640 feature map input, equipped with 128 3×3 convolutional kernels, and a receptive field of 64×64, specifically targeting larger blocks with a diameter greater than 5 cm. The medium-size decoupling head corresponds to a 320×320 feature map (after downsampling), configured with 64 3×3 convolutional kernels, and a receptive field of 32×32, focusing on medium-sized blocks with a diameter of 2-5 cm. The small-size decoupling head is adapted to a 160×160 feature map (after secondary downsampling), set with 32 3×3 convolutional kernels, and a receptive field of 16×16, used to capture small blocks with a diameter less than 2 cm. The receptive field and size configuration of the three decoupling heads achieve full coverage of block sizes. S42 Each decoupling head internally constructs three independent branch convolutional networks, corresponding to bounding box regression, confidence prediction, and class prediction tasks, respectively. Each branch network includes three convolutional layers and one activation layer. The activation layer uses a linear rectified function to perform feature mapping on the input feature map through convolution operations, converting high-dimensional features into low-dimensional feature representations for the corresponding task, generating bounding box regression branch features, confidence prediction branch features, and class prediction branch features, respectively. S43 decodes the bounding box regression branch features. Based on the mapping relationship between the feature map pixel coordinates and the original image, it calculates four parameters of the bounding box: the center x-coordinate, the center y-coordinate, the width, and the height. All parameter values ​​are normalized and controlled within the range of 0-1 to ensure positioning accuracy. S44 performs a Sigmoid activation function operation on the confidence prediction branch features, outputting the probability value of the target's existence, ranging from 0 to 1; it uses a Softmax activation function on the category prediction branch features, outputting the probability distributions of the two classes, missing blocks and non-missing blocks, with the sum of the probabilities of the two classes being 1. Subsequently, the bounding box parameters, confidence, and category probabilities are integrated to form the preliminary inference result for each potential target. This step-by-step process achieves accurate localization and category judgment of missing blocks of different sizes through targeted size adaptation and branch design.

[0039] Preferred, such as Figure 4As shown, step S5 includes the following sub-steps: S51, sorting the preliminary inference results from high to low confidence scores, and selecting the bounding box with the highest confidence score as the initial candidate box; S52, calculating the intersection-union ratio (IU) of the initial candidate box and all remaining bounding boxes, and marking the bounding boxes with IU ratios greater than a set threshold as redundant bounding boxes; S53, removing the marked redundant bounding boxes, and retaining the initial candidate boxes as valid detection results; S54, taking the bounding box with the highest confidence score among the remaining unprocessed bounding boxes as the new initial candidate box, repeating the IU calculation and redundancy removal process until all bounding boxes have been processed, and selecting the final valid detection results.

[0040] Specifically, step S5 involves a non-maximum suppression operation, which eliminates redundant bounding boxes through ordered filtering. S51 first preprocesses the preliminary inference results generated in S4, sorting them from highest to lowest confidence score. A confidence threshold of 0.5 is set, discarding detection results with confidence scores below this threshold and retaining only bounding boxes with a confidence score ≥ 0.5 as preliminary candidate boxes. This threshold has been validated through multiple experiments and can effectively filter over 80% of false detection results, reducing subsequent computation. S52 selects the first bounding box from the sorted preliminary candidate boxes as the initial candidate box. This candidate box represents the potential target detection result with the highest confidence. Then, the intersection-union ratio (IUGR) is used to calculate the overlap between the initial candidate box and all remaining preliminary candidate boxes. The IUGR is obtained by the ratio of the overlapping area to the joint area of ​​the bounding boxes. A IUGR threshold of 0.35 is set, marking bounding boxes with a ratio greater than this threshold as redundant bounding boxes. This threshold effectively distinguishes overlapping targets while avoiding missing true targets. S53 directly removes the bounding boxes marked as redundant, retaining only the initial candidate boxes as valid detection results, ensuring that each target corresponds to a unique and accurate bounding box. S54 selects the bounding box with the highest confidence from the remaining unprocessed bounding boxes as new initial candidate boxes, repeating the intersection-union ratio calculation in S52 and the redundancy removal operation in S53. This iterative process continues until all initial candidate boxes have been processed and no remaining bounding boxes can be used as new initial candidate boxes. This step-by-step screening process can control the redundant bounding box removal rate in the initial inference results to above 85%, while ensuring that the retention rate of true targets is not less than 98%, significantly improving the accuracy and simplicity of the detection results and providing a reliable data foundation for subsequent block drop identification and statistics.

[0041] like Figure 5As shown, a real-time detection method for block falling during drilling in deep wells based on DA-YOLO is implemented through different units, including: an image data acquisition and size adaptation unit, used to capture images of solid particles at the return outlet of a vibrating screen using a camera at a preset height, and transmit the images to a feature extraction unit after adjusting them to a fixed size; a multi-stage feature extraction and enhancement unit, composed of a Conv module, a C3k2_DSAM module, a SPPF module, and a C2PSA_DLKA module connected in series, which receives the adapted image data, performs feature extraction, dehazing, and background interference suppression operations, and outputs a multi-scale feature map; and a feature fusion and dimension adjustment unit, which receives the output from the multi-stage feature extraction and enhancement unit. Different levels of feature maps are used to construct a unified fusion feature map through upsampling, convolution, and element-wise fusion operations. The target detection and localization operation unit sets three different sizes of decoupling heads to perform branch operations on the fusion feature map and generate preliminary detection results including bounding box parameters, confidence scores, and class probabilities. The redundant bounding box filtering and optimization unit uses a non-maximum suppression algorithm to remove redundant bounding boxes through intersection-union ratio calculation and confidence threshold judgment, and outputs optimized detection results. The missing block identification and quantity counting unit receives the optimized detection results, identifies missing block targets and counts their number, and outputs a final detection report. Each unit sequentially interacts with signals and data bidirectionally through the data transmission interface.

[0042] The DA-YOLO-based real-time block fall detection method for deep wells employs a four-stage feature extraction structure, incorporating a dual-domain strip attention mechanism and deformable convolution technology. This enhances the ability to capture block target features while simultaneously performing environmental defogging and background interference suppression, effectively adapting to the complex lighting, water mist, and dust environments of deep well drilling sites. By leveraging a multi-scale feature fusion strategy and precise matching with three different sized decoupling heads, it achieves full coverage detection of block targets of varying sizes, accurately identifying and locating both tiny particles and larger blocks. The optimized application of the non-maximum suppression algorithm further improves the accuracy of bounding box selection, reduces redundant detection results, and ensures the reliability and accuracy of the detection output.

[0043] This method addresses the insufficient feature extraction capabilities of traditional models. Its unique module, through the synergistic effect of dual-domain attention and deformable sampling, significantly improves the feature discrimination of low-contrast, small-sized falling blocks, effectively solving the pain points of high false negative and false positive rates. Regarding the contradiction between multi-scale adaptability and real-time performance in existing technologies, this method optimizes the network structure design, simplifying the core computational process while retaining multi-scale detection capabilities. Combined with an efficient feature fusion and inference mechanism, it achieves a balance between detection accuracy and computational speed, meeting the time requirements of real-time monitoring during drilling. The fully automated detection architecture replaces the outdated traditional manual visual observation method, completely eliminating human interference and achieving standardized, accurate, and efficient falling block detection, providing a more reliable technical guarantee for deep well drilling safety.

[0044] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," "link," and "fix" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various equivalent changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A real-time detection method for block falling during drilling in deep wells based on DA-YOLO, characterized in that, Includes the following steps: S1: Image data of solid particles flowing with drilling fluid at the return outlet of the vibrating screen is acquired using a camera device at a set height. The acquired image data is adjusted to a fixed size and then input into the backbone network of the DA-YOLO model. S2: Feature extraction is performed on the input image using a four-stage feature extraction structure consisting of the Conv module, C3k2_DSAM module, SPPF module, and C2PSA_DLKA module connected in series in the backbone network. Specifically, the C3k2_DSAM module performs feature extraction and dehazing collaborative processing by embedding a dual-domain strip attention mechanism, while the C2PSA_DLKA module adaptively adjusts the sampling position and sensing frequency through superimposed deformable convolutions. S3: Input the shallow, middle, and deep features output from the backbone network into the neck network. Perform multi-scale feature fusion through upsampling and convolution operations to construct a multi-dimensional feature representation. S4: Use three different sized decoupling heads set in the head network to perform target detection and localization on the fused features, generating preliminary inference results including bounding box coordinates, confidence scores, and class probabilities. S5: Use a non-maximum suppression algorithm to filter redundant bounding boxes from the preliminary inference results. Calculate the bounding box overlap and combine it with a confidence threshold to complete the selection of effective detection results. S6: Based on the selected detection results, perform target identification and quantity statistics for dropped targets, and output the final result of dropped target detection.

2. The method for real-time detection of block falling during drilling in deep wells based on DA-YOLO as described in claim 1, characterized in that, The execution process of the dual-domain strip attention mechanism in the C3k2_DSAM module of S2 satisfies: ， in, Indicates the input feature map of the module. Indicates the spatial size of the feature map. This represents the feature map after frequency domain processing. Indicates the size is Frequency strip attention unit, Indicates the size is Spatial strip attention unit, express Convolution operation, Split represents channel splitting operation. This represents the sub-feature maps after splitting. This indicates the final output feature map of the module.

3. The method for real-time detection of block falling during drilling in deep wells based on DA-YOLO according to claim 2, characterized in that, The feature processing of the frequency strip attention unit in S2 satisfies: ， ， ， ， in, This indicates a strip average pooling operation. Indicates The central strip-shaped region features Indicates low-frequency components, Represents high-frequency components, and These represent the channel attention parameters for low-frequency and high-frequency components, respectively. Note the horizontal operator. This indicates the vertical attention operator. This indicates the output result of the frequency strip attention unit.

4. The method for real-time detection of block falling during drilling in deep wells based on DA-YOLO according to claim 3, characterized in that, The feature aggregation of the spatial strip attention unit in S2 satisfies: ， ， This indicates a global average pooling operation. express In convolution operations, sigmoid represents the activation function. "*" indicates lightweight attention weights, and "*" indicates a convolutional information aggregation method. Note the operators for representing horizontal stripes. Note the operators for representing vertical stripes. This indicates the output result of the spatial stripe attention cell. The input feature map of the unit is represented.

5. The method for real-time detection of block falling during drilling in deep wells based on DA-YOLO according to claim 4, characterized in that, The deformable large kernel attention mechanism of the C2PSA_DLKA module in S2 satisfies: ， Attention ,Deformable_LKA Attention , Output _LKA , in, Indicates the input feature map of the module. This represents the activation function. express Convolution operation, This represents deformable depthwise convolution. This represents a deformable depthwise dilated convolution, and Attention represents an attention map. "Indicates element-wise multiplication operation, _ represents the intermediate processing feature map, Output represents the module output feature map, and Deformable_LKA represents the deformable large kernel attention.

6. The method for real-time detection of block falling during drilling in deep wells based on DA-YOLO according to claim 5, characterized in that, The bounding box selection in the nonmaximum suppression algorithm in S5 satisfies: ， in, Indicates the predicted bounding box location. Indicates the actual bounding box location. Indicates the area of ​​the overlapping region of the bounding boxes. Indicates the area of ​​the bounding box's joint region. Indicates intersection, union, and ratio. This indicates the probability that a target exists within the predicted bounding box, and Confidence_Score represents the confidence score.

7. The method for real-time detection of block falling during drilling in deep wells based on DA-YOLO according to claim 1, characterized in that, S3 includes the following sub-steps: S31, receiving shallow feature maps, mid-level feature maps, and deep feature maps output by the backbone network, each feature map corresponding to different spatial resolutions and channel dimensions; S32, performing an upsampling operation on the deep feature map to make its spatial resolution consistent with that of the mid-level feature map, and simultaneously adjusting the number of channels to the same dimension through a 1×1 convolution; S33, fusing the adjusted deep feature map and the mid-level feature map element-wise to generate a first fused feature map, and then performing upsampling and channel adjustment operations on the first fused feature map; S34, fusing the adjusted first fused feature map and the shallow feature map element-wise, refining the features through a convolution operation, and outputting the final fused multi-scale feature map.

8. The method for real-time detection of block falling during drilling in deep wells based on DA-YOLO according to claim 1, characterized in that, S4 includes the following sub-steps: S41, the fused multi-scale feature maps are input into the large-size decoupling head, medium-size decoupling head, and small-size decoupling head of the head network, respectively, with each decoupling head corresponding to a different receptive field range; S42, each decoupling head performs feature mapping on the input feature map through a branch convolutional network to generate bounding box regression branch features, confidence prediction branch features, and category prediction branch features, respectively; S43, the bounding box regression branch features are decoded to obtain the center coordinates, width, and height parameters of the bounding box; S44, activation function operations are performed on the confidence prediction branch features and the category prediction branch features, respectively, to obtain the confidence of the target's existence and the probability distribution of the corresponding category, and the results of each branch are integrated to form a preliminary inference result.

9. The method for real-time detection of block falling during drilling in deep wells based on DA-YOLO according to claim 1, characterized in that, S5 includes the following sub-steps: S51, sort the preliminary inference results from high to low confidence scores, and select the bounding box with the highest confidence score as the initial candidate box; S52, calculate the intersection-union ratio (IU) of the initial candidate boxes with all remaining bounding boxes, and mark the bounding boxes with IU greater than a set threshold as redundant bounding boxes; S53, remove the marked redundant bounding boxes and retain the initial candidate boxes as valid detection results; S54, take the bounding box with the highest confidence among the remaining unprocessed bounding boxes as the new initial candidate boxes, and repeat the IU calculation and redundancy removal process until all bounding boxes have been processed and the final valid detection results are selected.

10. The method for real-time detection of block spalling in deep wells based on DA-YOLO according to any one of claims 1-9, characterized in that, This method is implemented through different units, including: an image data acquisition and size adaptation unit, used to capture images of solid particles at the return outlet of the vibrating screen using a camera at a preset height, and transmit the images to the feature extraction unit after adjusting them to a fixed size; a multi-stage feature extraction and enhancement unit, composed of a Conv module, a C3k2_DSAM module, an SPPF module, and a C2PSA_DLKA module connected in series, which receives the adapted image data, performs feature extraction, dehazing, and background interference suppression operations, and outputs multi-scale feature maps; and a feature fusion and dimension adjustment unit, which receives feature maps of different levels output by the multi-stage feature extraction and enhancement unit, and performs feature fusion and dimension adjustment through uploading. The system employs sampling, convolution, and element-wise fusion operations to construct a unified-dimensional fusion feature map. The target detection and localization unit uses three different sized decoupling heads to perform branching operations on the fusion feature map, generating preliminary detection results including bounding box parameters, confidence scores, and class probabilities. The redundant bounding box filtering and optimization unit uses a non-maximum suppression algorithm, calculating the intersection-union ratio and determining the confidence threshold to remove redundant bounding boxes and output optimized detection results. The missing block identification and counting unit receives the optimized detection results, identifies missing block targets, counts their number, and outputs a final detection report. Each unit sequentially interacts with the other through a data transmission interface, exchanging signals and data bidirectionally.

Images