Coal mine underground sensor reading identification method

Abstract

The application relates to an improved YOLO11n and PP-OCR underground coal mine sensor reading identification method in the technical field of intelligent video safety monitoring of coal mines, which comprises the following steps: video data acquisition and preprocessing; constructing an improved YOLO11n target detection model and training; sensor value identification; reading extraction and standardization processing; experiment design and result analysis. The improved YOLO11n replaces DBNet++ in the PP-OCRv5 and removes the direction classifier, a light-weight value identification process is constructed, a value standardization processing rule is designed, the character error and format confusion problems in OCR identification are effectively solved, the accuracy and uniformity of the reading are guaranteed, non-contact automatic reading of the sensor concentration is realized, a video redundancy verification channel independent of a traditional communication link is constructed for coal mine safety monitoring, and the single-channel data cheating and missing report risk is fundamentally prevented.

Description

Technical Field

[0001] This invention belongs to the field of intelligent video safety monitoring technology in coal mines, specifically involving an improved method for recognizing underground sensor readings using YOLO11n and PP-OCR. This method is used for intelligent visual recognition of readings from various underground sensors, constructs an independent verification mechanism for video channels, improves the accuracy and reliability of underground environmental monitoring sensors, and ensures safe production in coal mines. Background Technology

[0002] The underground environment in coal mines is complex and harsh. Various environmental sensors (such as those for methane, carbon monoxide, carbon dioxide, temperature, smoke, and dust) are core equipment for monitoring mine gas concentrations and environmental parameters, preventing disasters and accidents, and are crucial for ensuring the safety of miners and safe production in the mine. Statistics show that most coal mine accidents are related to the failure to detect abnormal accumulations of underground environmental parameters in a timely manner. Accurate and reliable environmental sensors can effectively monitor environmental changes, thereby effectively reducing accident risks. According to mandatory provisions of the "Coal Mine Safety Regulations," various environmental sensors must be used for real-time monitoring in key areas such as mining faces, return airways, and electromechanical chambers, and the equipment must be ensured to be in normal working condition.

[0003] Current safety monitoring systems generally use wired / wireless industrial communication protocols to upload sensor data to a ground monitoring center for dynamic early warning and supervision. This model suffers from a single-channel monitoring vulnerability: the display terminal and transmission link are often unmonitored, making it difficult to detect actions such as obstruction, displacement, disconnection, or injection of false data in real time. This results in discrepancies between the monitoring platform data and the actual on-site concentration, creating a safety blind spot. Video redundancy monitoring, as an important supplementary measure, uses surveillance cameras to capture images from the LED digital tube display of methane sensors, establishing an independent verification mechanism for each video channel, which can significantly improve the reliability of supervision.

[0004] Currently, in order to compensate for the loopholes in traditional single-channel monitoring based on industrial communication protocols, some mines have begun to try to introduce video redundancy monitoring technology as an independent verification method. However, the existing pure vision solution still has the following significant drawbacks in practical applications: First, the underground environment is complex and harsh, and the area faces challenges such as low illumination, high dust levels, and localized overexposure. Under low illumination conditions, the boundary between the sensor target and the background environment is blurred, making it difficult for traditional target detection algorithms to accurately locate the sensor body and its reading area, which easily leads to missed detections and false detections. Secondly, environmental sensors often use LED digital tubes to display values. In low light, reflective conditions, or when obscured by dust, characters are prone to breakage, blurred edges, and extremely low contrast. Conventional OCR character recognition algorithms have poor robustness in such environments, resulting in a significant drop in reading accuracy and making it difficult to achieve accurate comparison and verification with monitoring platform data. Third, the text detection modules (such as DBNet++) of existing conventional OCR tools (such as PP-OCRv5) have high computational complexity and long inference time, making it difficult to meet the high frame rate requirements of real-time processing of video streams in coal mines. Furthermore, the built-in direction classifier causes computational redundancy in the fixed instrument scenarios underground. Summary of the Invention

[0005] Based on the comprehensive considerations of low cost and lightweight design, the technical problem to be solved by this invention is: how to overcome the interference of complex environments such as low illumination and high dust in underground mines, and achieve high-precision and lightweight visual recognition of LED display readings of various environmental sensors under the condition of limited computing power of edge computing devices.

[0006] To address the aforementioned problems, this invention provides an improved method for identifying sensor readings in underground coal mines using YOLO11n and PP-OCR, comprising the following steps: S1 Video Data Acquisition and Preprocessing: Video streams containing images from environmental sensors and LED digital tube displays are acquired through underground coal mine cameras; the video is processed by frame extraction, and rectangular boxes are marked on the sensor and LED digital tube areas in the images; the samples are expanded through data augmentation techniques such as random brightness and contrast adjustment, Gaussian noise, random shadows, and affine transformation to construct training, validation, and test sets; S2 Construct and train an improved YOLO11n object detection model: Construct an improved YOLO11n object detection model, and use the training set, validation set and test set constructed in S1 to iteratively train the improved YOLO11n model and save the optimal weights; S3 Sensor Numerical Recognition: The text detection module in the PP-OCR tool is replaced with a trained and improved YOLO11n, and the built-in orientation classifier module of the PP-OCR tool is removed to build a lightweight recognition process; S4 Reading Extraction and Normalization: The real-time video stream is input into the improved YOLO11n model to locate the LED digital tube area frame by frame, and the target sub-image is cropped according to the coordinates of the output bounding box; then the sub-image is directly sent to the text recognition module of the PP-OCR tool to output the original numerical text; S5 Experimental Design and Results Analysis: The accuracy, recall, average precision, computational load, and inference speed of target recognition were verified through ablation experiments, YOLO series comparative experiments, and overall performance testing of numerical recognition.

[0007] Because YOLO11n faces the following challenges in the sensor LED digital tube detection scenario described in this paper: firstly, the sensor window typically occupies only a tiny pixel in the image, making it a typical small target; secondly, factors such as underground dust cover and miner's lamp reflection make the sensor target susceptible to interference from complex backgrounds. To address these issues, this invention makes three improvements to YOLO11n. The improved YOLO11n target detection model includes: The ODConv-EMA module is introduced into the C3k2 module of the YOLO11n model. It uses full-dimensional dynamic convolution to replace standard convolution, so that the convolution kernel can dynamically adjust the parameters according to the input image, thereby enhancing the adaptability to the changing appearance of the sensor. A global attention mechanism (GAM) is introduced after the feature fusion layer of the neck network. Through sequential channel attention and spatial attention, the feature representation of small targets after multi-scale fusion is enhanced. We replace the default CIoU with the WIoU v3 loss function and use a dynamic non-monotonic focusing mechanism to suppress the harmful gradients of low-quality samples generated by the harsh downhole environment.

[0008] To enable the improved C3k2 module to possess both dynamic adaptive local feature extraction capabilities and multi-scale global perception capabilities, the ODConv-EMA module extends dynamic convolution to four dimensions: the number of convolutional kernels, spatial size, input channels, and output channels. Four types of attention learn in parallel and complement each other, and their mathematical definition is: In the formula, x and y are the input and output features, respectively; , , , These are attention scalars representing the number of kernels, output channels, input channels, and spatial dimensions, respectively. This is element-wise multiplication; Let i be the i-th convolutional kernel; The ODConv-EMA module replaces the second 3×3 convolution of the original Bottleneck module with ODConv. The first 1×1 convolution remains unchanged as a standard convolution, preserving the efficient channel information fusion capability. A dynamic adaptive mechanism is introduced only in the spatial feature extraction layer, followed by multi-scale feature enhancement through the EMA attention mechanism, and finally output through residual connections.

[0009] To achieve the transformation from "single-layer purification" to "fusion enhancement" and ensure that the fused features are fully enhanced in both channel and spatial dimensions, effectively improving the detection accuracy of small targets, the GAM global attention mechanism includes a channel attention module and a spatial attention module. The channel attention module first learns the input features through global average pooling. Channel dependencies are represented by two 1×1 convolutions followed by ReLU and Sigmoid activations to generate a channel attention map. Spatial attention branch for channel attention output Max pooling and average pooling are performed along the channel dimension, and the concatenated arrays are then subjected to a 7×7 convolution and a sigmoid function to generate a spatial attention map. The specific calculations are as follows: In the formula, This indicates element-wise multiplication; F1 is the input feature map for the GAM global attention mechanism; GAP is the global average pooling operation. , These are two 1×1 convolutional layers; ReLU is the ReLU activation function; Sigmoid is the Sigmoid activation function; Mc(F1) is the channel attention weight map; Fc is the feature map after channel attention weighting; MaxPool and AvgPool are the max pooling and average pooling operations along the channel dimension, respectively; [;] indicates tensor concatenation along the channel dimension; The convolution is 7×7; Ms(Fc) is the spatial attention weight map; F2 is the feature map of the fused channel and spatial attention output by the GAM module.

[0010] To avoid the model being interfered with by harmful gradients and to improve the model's generalization ability, the WIoU v3 loss function intelligently assigns gradient gains to samples of different quality using a dynamic non-monotonic focusing mechanism, thereby focusing on the common quality anchor boxes that are most beneficial to the model's improvement. Its definition is: In the formula, For the attention term, amplify the loss of the predicted box that is far from the center point; ==1− for Loss measures the degree of overlap between predicted and ground truth bounding boxes; r is a non-monotonic focusing coefficient, the core of which is based on outlier. The dynamic gradient gain allocation means that small outliers represent high-quality anchor frames, while large outliers correspond to low-quality anchors. In the formula, for The exponential moving average; To prevent small amounts from being divided by zero; and β is a hyperparameter that controls the shape of the non-monotonic curve; r changes non-monotonicly with β: the gradient decreases for high-quality samples, the gradient is suppressed for low-quality samples, and the highest gradient gain is obtained for medium-quality samples.

[0011] To further simplify the inference process, reduce computational overhead, improve inference speed, and meet the real-time requirements of underground coal mines, in step S3, the trained improved YOLO11n is used to replace the DBNet++ text detection module in PP-OCRv5, and the built-in orientation classifier module of PP-OCRv5 is removed; the improved YOLO11n is used to quickly locate the sensor value display area in the video frame, and the target sub-image is cropped according to the coordinates of the output bounding box; then the sub-image is directly sent to the SVTR text recognition module of PP-OCRv5.

[0012] To ensure the accuracy and usability of the readings, in step S4, the original numerical text is normalized: based on the inherent format characteristics of the sensor LED display, a three-in-one mechanism of regular expression matching, rule verification, and anomaly filtering is constructed to automatically clean, correct, and standardize the output for typical misidentification patterns, thereby obtaining standard sensor readings.

[0013] In summary, compared with the prior art, the present invention has the following technical advantages and beneficial effects: (1) To address the challenge of detecting small targets in LED digital tubes of methane sensors in coal mines, YOLO11n was improved in three ways: the C3k2 module introduced ODConv-EMA dynamic convolution and EMA multi-scale attention, the neck network was embedded with GAM global attention, and the WIoU v3 loss function was adopted; the improved model achieved an mAP50 of 94.3%, which is 5.1 percentage points higher than the baseline, and the recall rate was increased to 87.9%, which significantly suppressed the missed detection of small targets; (2) A lightweight numerical recognition process was constructed by replacing DBNet++ in PP-OCRv5 with improved YOLO11n and removing the orientation classifier. The frame rate reached 23 FPS on the mining edge computing device, and the memory and NPU usage were reduced by 50% and 53.2% respectively, meeting the real-time processing requirements in the mine. (3) Numerical normalization processing rules were designed, which effectively solved the problems of character error and format chaos in OCR recognition and ensured the accuracy and uniformity of readings. This method realizes non-contact automatic reading of sensor concentration and builds a video redundancy verification channel independent of traditional communication links for coal mine safety monitoring, fundamentally preventing the risk of single-channel data cheating and underreporting. Attached Figure Description

[0014] Figure 1This is a flowchart of the coal mine underground sensor reading identification method of the present invention; Figure 2 This is a partial dataset image illustration of the present invention; Figure 3 This is a schematic diagram of the structure of the improved YOLO11n model of the present invention; Figure 4 This is a schematic diagram of the C3k2 module of the present invention; Figure 5 This is a schematic diagram of the network structure of the ODConv-EMA module of the present invention; Figure 6 This is a schematic diagram of the structure of the GAM module of the present invention; Figure 7 This is a schematic diagram of the sensor value recognition process of the present invention; Figure 8 This is a visual comparison chart of the detection results of the present invention and existing algorithms. Detailed Implementation

[0015] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0016] like Figure 1 The method for identifying sensor readings in underground coal mines using an improved YOLO11n and PP-OCR method includes the following steps: S1 Video data acquisition and preprocessing; S2 Constructing and training an improved YOLO11n target detection model; S3 Sensor numerical identification; S4 Reading extraction and normalization; S5 Experimental design and result analysis.

[0017] The following is a detailed explanation of each step.

[0018] S1 Video Data Acquisition and Preprocessing Video streams containing images from LED digital tube displays of environmental sensors (such as methane and carbon monoxide sensors) were captured using underground cameras in coal mines. The videos were then processed by frame extraction, and rectangular boxes were used to mark the sensor LED digital tube areas in the images. Data augmentation techniques such as random brightness and contrast adjustment, Gaussian noise, random shadows, and affine transformations were used to expand the sample and construct training and testing sets.

[0019] The raw data comes from monitoring videos of T1 and T2 methane sensors in multiple coal mine longwall faces and tunneling roadways, covering different installation locations, lighting conditions, and dust environments. A total of 15,000 frames were extracted from the collected videos. After filtering and deblurring, 13,000 valid images were retained. Rectangular bounding boxes were used to label the LED digital tube areas of the sensors (labeled "LED"). The training, validation, and test sets were divided in a 7:2:1 ratio. All samples came from real monitoring scenarios in different coal mines, with no duplicate data, ensuring the model's generalization ability. Some dataset images are shown below. Figure 2 As shown.

[0020] S2 constructs and trains an improved YOLO11n object detection model. As a representative algorithm for single-stage object detection, the YOLO series achieves a good balance between real-time performance and accuracy through its end-to-end detection method, and is widely used in industrial inspection, intelligent monitoring, and other fields. YOLO11, launched by the Ultralytics team based on YOLOv5 and YOLOv8, is a new generation object detection framework. It continues the classic three-level structure of the YOLO series: "backbone network - neck network - detection head," and provides model variants with five parameter levels: n, s, m, l, and x, which can flexibly adapt to various deployment environments from edge devices to high-performance servers. Among them, YOLO11n is the smallest and fastest, suitable for the real-time detection needs of edge computing devices in coal mine scenarios. However, YOLO11n faces the following challenges in the sensor LED digital tube detection scenario presented in this paper: first, the sensor window typically occupies only a tiny pixel in the image, representing a typical small target; second, factors such as dust cover and mine lamp reflections make the sensor target susceptible to interference from complex backgrounds.

[0021] To address the aforementioned challenges, this invention makes three improvements to YOLO11n, resulting in the following model structure: Figure 3 As shown: (1) Introduce the ODConv-EMA module into the C3k2 module of the YOLO11n model, and use full-dimensional dynamic convolution to replace the standard convolution, so that the convolution kernel can dynamically adjust the parameters according to the input image, thereby enhancing the adaptive ability to the sensor’s varied appearance.

[0022] (2) A global attention mechanism (GAM) is introduced after the neck network feature fusion layer. Through sequential channel attention and spatial attention, the feature expression of small targets after multi-scale fusion is enhanced.

[0023] (3) The WIoU v3 loss function is used to replace the default CIoU, and the dynamic non-monotonic focusing mechanism is used to suppress the harmful gradient of low-quality samples generated by the harsh downhole environment.

[0024] (1) ODConv-EMA module The C3k2 module, a brand-new module in YOLO11n, is primarily responsible for feature extraction and is characterized by its flexibility and efficiency. The C3k2 module has two switchable operating modes: C3k=false is the simple mode, such as... Figure 4 As shown in (a), the standard bottleneck structure Bottleneck is used; as Figure 4 As shown in (c), this is suitable for lightweight models that require fast inference, characterized by high computation speed and few parameters; C3k=true is a complex mode, such as... Figure 4 As shown in (b), the standard bottleneck structure is replaced with the c3 module; as Figure 4 As shown in (d), the feature extraction capability is strong and can meet high-precision requirements. Although the C3k2 module helps YOLO11n exhibit good performance in many scenarios, the complex environment of LED digital tube sensors in coal mines, characterized by small target size, unclear target edges, unstable lighting, and variable backgrounds, makes it difficult for YOLO11n to efficiently focus on key features, easily leading to false detections and missed detections. Therefore, this invention introduces the ODConv-EMA module into the C3k2 module, as shown in... Figure 4 As shown in (e), it replaces the original Bottleneck module.

[0025] Standard convolution parameters are fixed, and traditional dynamic convolution only introduces attention in the kernel number dimension, resulting in limited adaptive capabilities. ODConv-EMA full-dimensional dynamic convolution extends dynamic convolution to four dimensions: kernel number, spatial size, input channels, and output channels. Four types of attention learn in parallel and complement each other, achieving performance comparable to multi-core traditional dynamic convolution even with a single kernel. Its mathematical definition is: In the formula: x, y are the input and output features, respectively; , , , These are attention scalars representing the number of kernels, output channels, input channels, and spatial dimensions, respectively. This is element-wise multiplication; Let be the i-th convolutional kernel.

[0026] EMA achieves multi-scale attention without dimensionality reduction through channel grouping and cross-spatial learning. This multi-scale perception characteristic gives EMA a significant advantage in scenarios with a large range of target size variations. It can simultaneously enhance the global contour of large targets and the local details of small targets, making it particularly suitable for the small target detection scenario of the sensor in this paper.

[0027] like Figure 5As shown, the ODConv-EMA module replaces the second 3×3 convolution of the original Bottleneck with ODConv, while keeping the first 1×1 convolution unchanged, preserving the efficient channel information fusion capability. A dynamic adaptive mechanism is introduced only in the spatial feature extraction layer, followed by multi-scale feature enhancement through the EMA attention mechanism, and finally output through residual connections. This combination enables the improved C3k2 module to have both dynamic adaptive local feature extraction capability and multi-scale global perception capability.

[0028] (2) Global Attention Mechanism of GAM LED digital tubes typically have a pixel ratio of less than 10%, and their features tend to become sparsified after multiple downsampling by the backbone network. Introducing attention after the neck fusion layer can further refine features and enhance the spatial representation of small targets. Existing research often deploys attention modules at the end of the backbone network, which can only access single-layer feature maps and lacks direct access to multi-scale fusion information. The structure of the GAM module (Global Attention Mechanism) is as follows: Figure 6 As shown, this invention inserts GAM into each feature fusion node of the neck network (after each Concat operation), enabling it to simultaneously receive fusion information from both the top-down path of FPN and the bottom-up path of PAN, thus achieving a transformation from "single-layer purification" to "fusion enhancement." This design ensures that the fused features are sufficiently enhanced in both channel and spatial dimensions, effectively improving the detection accuracy of small targets.

[0029] The structure of the GAM module is as follows: Figure 6 As shown, GAM adopts a sequential architecture of "channel attention → spatial attention", including a channel attention module and a spatial attention module. The channel attention module first learns the input features through global average pooling (GAP). Channel dependencies are represented by two 1×1 convolutions followed by ReLU and Sigmoid activations to generate a channel attention map. Spatial attention branch for channel attention output Max pooling and average pooling are performed along the channel dimension, and the concatenated arrays are then subjected to a 7×7 convolution and a sigmoid function to generate a spatial attention map. The specific calculations are as follows: In the formula, This indicates element-wise multiplication; F1 is the input feature map for the GAM global attention mechanism; GAP is the global average pooling operation. , These are two 1×1 convolutional layers; ReLU is the ReLU activation function; Sigmoid is the Sigmoid activation function; Mc(F1) is the channel attention weight map; Fc is the feature map after channel attention weighting; MaxPool and AvgPool are the max pooling and average pooling operations along the channel dimension, respectively; [;] indicates tensor concatenation along the channel dimension; The convolution is 7×7; Ms(Fc) is the spatial attention weight map; F2 is the feature map of the fused channel and spatial attention output by the GAM module.

[0030] (3) WIoU v3 loss function YOLO11 defaults to using CIoU as the bounding box regression loss, accelerating convergence by introducing centroid distance and aspect ratio penalties. However, underground monitoring videos are affected by dust, reflected light, water mist, etc., resulting in a large number of low-quality frames and blurred correspondences between anchor boxes and target boxes. CIoU applies a fixed geometric metric penalty to all anchor boxes, failing to perceive differences in sample quality. The model is easily affected by harmful gradients, leading to overfitting to difficult samples and weakening generalization ability. To address this, this invention introduces the WIoU v3 (Wise-IoU v3) loss function, which intelligently allocates gradient gains to samples of different quality using a dynamic non-monotonic focusing mechanism, thereby focusing on the ordinary quality anchor boxes that are most beneficial to model improvement. Its definition is: In the formula, For the attention term, amplify the loss of the predicted box that is far from the center point; ==1− for Loss measures the degree of overlap between predicted and ground truth bounding boxes; r is a non-monotonic focusing coefficient, the core of which is based on outlier. The dynamic gradient gain allocation means that small outliers represent high-quality anchor frames, while large outliers correspond to low-quality anchors. In the formula, for The exponential moving average, To prevent small amounts from being divided by zero; and β is a hyperparameter that controls the shape of the non-monotonic curve; r changes non-monotonicly with β: the gradient of high-quality samples is reduced, the gradient of low-quality samples is suppressed, and medium-quality samples obtain the highest gradient gain, thereby improving the generalization ability of the model.

[0031] The improved YOLO11n model is iteratively trained using the training set, validation set, and test set constructed in step S1, and the optimal weights are saved.

[0032] S3 sensor numerical recognition like Figure 7 As shown, the improved YOLO11n after training is used to replace the DBNet++ text detection module in PP-OCRv5; considering the characteristics of the downhole sensor being fixed in the forward direction and the text direction being uniform, the orientation classifier module built into PP-OCRv5 is removed to build a lightweight recognition process.

[0033] PP-OCR is an open-source OCR toolkit built on the Paddle deep learning framework, boasting advantages such as multi-language support, lightweight design, and flexible deployment. This invention selects its latest version, PP-OCRv5, as the core recognition engine. PP-OCRv5 continues the series' two-stage architecture of "text detection + text recognition," further optimizing the balance between accuracy and speed. The text detection module employs a differentiable binarization-based DBNet++ network, effectively segmenting text regions in images. The text recognition module uses the Transformer-based SVTR algorithm, capturing global dependencies in character sequences through a self-attention mechanism to achieve end-to-end image-to-text mapping. Compared to traditional CRNN structures, the SVTR algorithm offers higher recognition efficiency and fewer parameters.

[0034] However, the default DBNet++ text detection algorithm in PP-OCRv5 has high computational complexity and long inference time, making it difficult to meet the high frame rate requirements of real-time video stream processing in coal mines. Therefore, this invention replaces the original DBNet++ detection module with an improved YOLO11n object detection network. Furthermore, considering the sensor reading recognition scenario in coal mines, where instruments and signs are fixed in a positive orientation with uniform text direction and no issues of inversion or rotation, this invention skips the built-in orientation classifier module in PP-OCRv5, further simplifying the inference process and reducing computational overhead. The specific process is as follows: First, the improved YOLO11n is used to quickly locate the sensor value display area in the video frame, and the target sub-image is cropped based on the output bounding box coordinates; then, the sub-image is directly fed into the SVTR text recognition module of PP-OCRv5. This optimization strategy effectively avoids the computational bottleneck of DBNet++, significantly improves inference speed, and meets the real-time requirements of coal mines.

[0035] S4 Reading Extraction and Normalization The real-time video stream is input into the improved YOLO11n model, and the LED digital tube area is located frame by frame. The target sub-image is cropped according to the coordinates of the output bounding box. Then the sub-image is directly sent to the SVTR text recognition module of PP-OCRv5 to output the original numerical text.

[0036] Due to complex working conditions such as underground dust, uneven lighting, missing strokes on the LED screen, and imaging distortion, OCR output readings are prone to errors such as character confusion, missing symbols, and disordered formats, making the raw results unsuitable for direct use in safety monitoring and over-limit assessment. Therefore, numerical standardization processing is performed on the raw recognition results: based on the inherent format characteristics of the sensor's LED display (e.g., X.XX, range 0.00-4.00), a three-pronged mechanism of regular expression matching, rule verification, and anomaly filtering is constructed. This mechanism automatically cleans, corrects, and standardizes typical misidentification patterns such as "digit and letter confusion (e.g., O recognized as 0)," "missing decimal point," and "abnormal numerical range," resulting in standard sensor readings that ensure accuracy and usability. Typical misidentification types and correction rules are shown in Table 1. Table 1 Typical Misidentification Types and Correction Rules S5 Experimental Design and Results Analysis (1) Experimental environment and evaluation indicators The experimental training server for this invention uses Ubuntu 22.04 as its operating system, an NVIDIA A100 graphics card, 40GB of RAM, an Intel(R) Xeon(R) Gold 6248R CPU, 256GB of RAM, and the deep learning framework PyTorch 2.2.2 with CUDA 12.1. The learning rate is 0.01, using the SGD optimizer, momentum 0.937, weight decay 0.005, input size 640×640, 300 training epochs, and a batch size of 32. The inference platform is an RK3588 intrinsically safe edge computing device for mining, with 6 TOPS of computing power.

[0037] This invention uses the precision (%), recall (%), mean precision (mAP50) (mean precision when IoU threshold is 0.5), computational cost (GFLOPs), and inference speed (FPS) (number of image frames processed per second by the model) commonly used in target recognition as evaluation metrics.

[0038] (2) Ablation test Ablation experiments were conducted on a self-made dataset using YOLO11n as the baseline model to verify the effectiveness of each improvement strategy. The results are shown in Table 2. Table 2 Ablation Experiment Results As shown in Table 2, compared to baseline Model 1, Model 2, which incorporates ODConv-EMA, improves mAP50 by 1.8 percentage points, validating the enhancement effect of dynamic adaptive receptive field on feature extraction. Model 3, which incorporates GAM, improves mAP50 by 2.3 percentage points, indicating that global attention can effectively capture the global context of small targets. Model 4, which uses WIoU v3, improves mAP50 by 1.2 percentage points, demonstrating that it enhances regression robustness by focusing on ordinary quality anchor boxes. Model 5, which incorporates both ODConv-EMA and GAM, sees mAP50 jump to 93.1%, reflecting the positive complementarity between dynamic feature extraction and global attention. Model 6 further incorporates WIoU v3, achieving mAP50 of 94.3%, an improvement of 5.1 percentage points compared to the baseline, with P and R increasing to 91.6% and 87.9% respectively, validating the synergistic gain of the three improvement strategies.

[0039] (3) Comparative experiments of YOLO series The algorithm of this invention was compared with the current mainstream YOLO algorithm under the same conditions, and the results are shown in Table 3: Table 3 Comparison of experimental results As shown in Table 3, the algorithm of this invention significantly outperforms the comparative models in terms of precision, recall, and mAP50. The precision reaches 91.6%, which is 1.4 percentage points higher than the second-best YOLOv12n; the recall reaches 87.9%, which is 6.1-7.8 percentage points higher than the other models, indicating that global attention and robust regression loss significantly suppress false negatives; the mAP50 reaches 94.3%, which is 4.5 percentage points higher than YOLOv5n. Although the inference speed is slightly lower than the baseline, it still meets the real-time detection requirements of underground video surveillance in coal mines and has outstanding application value on edge devices.

[0040] To visually compare the performance differences of different object detection algorithms, Figure 8 The visualization detection results of various YOLO series algorithms under four typical interference scenarios, including backlighting and overexposure, blurring and reflection, and dust and water mist, are presented. The improved YOLO11n algorithm proposed in this invention shows significant advantages in both target detection rate and confidence.

[0041] (4) Overall performance test of numerical recognition To verify the effectiveness and end-to-end recognition performance of the improved YOLO11n replacing DBNet++ in PP-OCRv5, 1000 sensor monitoring images were independently collected from the fully mechanized mining face and tunneling roadways in a coal mine as a test set (without overlap with the training / validation set), covering typical working conditions such as normal lighting, low illumination, strong reflection, and dust obstruction. The test was deployed on a mine-use intrinsically safe edge computing device (6 TOPS computing power). Numerical accuracy was defined as the integer value perfect match rate. The results are shown in Table 4. Table 4 Test Results of Numerical Recognition Process As shown in Table 4, the overall time consumption of this invention is reduced by 47.8%, the frame rate is increased by about 1.9 times, the memory and NPU usage are reduced by 50% and 53.2% respectively, and the accuracy is increased by 3.0 percentage points. The performance improvement is due to the improved single-stage efficient detection of small LED targets by YOLO11n and the reduction of computational redundancy brought about by the removal of the orientation classifier.

[0042] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape, principle and application direction of this application should be covered within the scope of protection of this application.

Claims

1. An improved method for identifying sensor readings in underground coal mines using YOLO11n and PP-OCR, characterized in that, Includes the following steps: S1 Video Data Acquisition and Preprocessing: Video streams containing images from environmental sensors and LED digital tube displays are acquired through underground coal mine cameras; the video is processed by frame extraction, and rectangular boxes are marked on the sensor and LED digital tube areas in the images; the samples are expanded through data augmentation techniques such as random brightness and contrast adjustment, Gaussian noise, random shadows, and affine transformation to construct training, validation, and test sets; S2 Construct and train an improved YOLO11n object detection model: Construct an improved YOLO11n object detection model, and use the training set, validation set and test set constructed in S1 to iteratively train the improved YOLO11n model and save the optimal weights; S3 Sensor Numerical Recognition: The text detection module in the PP-OCR tool is replaced with a trained and improved YOLO11n, and the built-in orientation classifier module of the PP-OCR tool is removed to build a lightweight recognition process; S4 Reading Extraction and Normalization: The real-time video stream is input into the improved YOLO11n model to locate the LED digital tube area frame by frame, and the target sub-image is cropped according to the coordinates of the output bounding box; then the sub-image is directly sent to the text recognition module of the PP-OCR tool to output the original numerical text; S5 Experimental Design and Results Analysis: The accuracy, recall, average precision, computational load, and inference speed of target recognition were verified through ablation experiments, YOLO series comparative experiments, and overall performance testing of numerical recognition.

2. The improved YOLO11n and PP-OCR method for identifying sensor readings in underground coal mines according to claim 1, characterized in that, In step S2, the improved YOLO11n target detection model includes: The ODConv-EMA module is introduced into the C3k2 module of the YOLO11n model. It uses full-dimensional dynamic convolution to replace standard convolution, so that the convolution kernel can dynamically adjust the parameters according to the input image, thereby enhancing the adaptability to the changing appearance of the sensor. A global attention mechanism (GAM) is introduced after the feature fusion layer of the neck network. Through sequential channel attention and spatial attention, the feature representation of small targets after multi-scale fusion is enhanced. We replace the default CIoU with the WIoU v3 loss function and use a dynamic non-monotonic focusing mechanism to suppress the harmful gradients of low-quality samples generated by the harsh downhole environment.

3. The improved YOLO11n and PP-OCR method for identifying readings of underground sensors in coal mines according to claim 2, characterized in that, The ODConv-EMA module extends dynamic convolution to four dimensions: number of kernels, spatial size, input channels, and output channels. Four types of attention learn in parallel and complement each other, and their mathematical definition is: In the formula, x and y are the input and output features, respectively; , , , These are attention scalars representing the number of kernels, output channels, input channels, and spatial dimensions, respectively. This is element-wise multiplication; Let i be the i-th convolutional kernel; The ODConv-EMA module replaces the second 3×3 convolution of the original Bottleneck module with ODConv. The first 1×1 convolution remains unchanged as a standard convolution, preserving the efficient channel information fusion capability. A dynamic adaptive mechanism is introduced only in the spatial feature extraction layer, followed by multi-scale feature enhancement through the EMA attention mechanism, and finally output through residual connections.

4. The improved YOLO11n and PP-OCR method for identifying sensor readings in underground coal mines according to claim 2, characterized in that, The GAM global attention mechanism includes a channel attention module and a spatial attention module. The channel attention module first learns the input features through global average pooling. Channel dependencies are represented by two 1×1 convolutions followed by ReLU and Sigmoid activations to generate a channel attention map. Spatial attention branch for channel attention output Max pooling and average pooling are performed along the channel dimension, and the concatenated arrays are then subjected to a 7×7 convolution and a sigmoid function to generate a spatial attention map. The specific calculations are as follows: In the formula, This indicates element-wise multiplication; F1 is the input feature map for the GAM global attention mechanism; GAP is the global average pooling operation. , These are two 1×1 convolutional layers; ReLU is the ReLU activation function; Sigmoid is the Sigmoid activation function; Mc(F1) is the channel attention weight map; Fc is the feature map after channel attention weighting; MaxPool and AvgPool are the max pooling and average pooling operations along the channel dimension, respectively; [;] indicates tensor concatenation along the channel dimension; The convolution is 7×7; Ms(Fc) is the spatial attention weight map; F2 is the feature map of the fused channel and spatial attention output by the GAM module.

5. The improved YOLO11n and PP-OCR method for identifying sensor readings in underground coal mines according to claim 2, characterized in that, The WIoU v3 loss function intelligently assigns gradient gains to samples of different quality using a dynamic non-monotonic focusing mechanism, thereby focusing on the common quality anchor boxes that are most beneficial to model improvement. Its definition is: In the formula, For the attention term, amplify the loss of the predicted box that is far from the center point; ==1− for Loss measures the degree of overlap between predicted and ground truth bounding boxes; r is a non-monotonic focusing coefficient, the core of which is based on outlier. The dynamic gradient gain allocation means that small outliers represent high-quality anchor frames, while large outliers correspond to low-quality anchors. In the formula, for The exponential moving average; To prevent small amounts from being divided by zero; and β is a hyperparameter that controls the shape of the non-monotonic curve; r changes non-monotonicly with β: the gradient decreases for high-quality samples, the gradient is suppressed for low-quality samples, and the highest gradient gain is obtained for medium-quality samples.

6. The improved YOLO11n and PP-OCR method for identifying sensor readings in underground coal mines according to claim 2, characterized in that, In step S3, the improved YOLO11n trained by training is used to replace the DBNet++ text detection module in PP-OCRv5, and the built-in orientation classifier module of PP-OCRv5 is removed. The improved YOLO11n is used to quickly locate the sensor value display area in a video frame, and the target sub-image is cropped based on the output bounding box coordinates. The sub-image is then directly fed into the SVTR text recognition module of PP-OCRv5.

7. The improved YOLO11n and PP-OCR method for identifying sensor readings in underground coal mines according to claim 2, characterized in that, In step S4, the original numerical text is subjected to numerical normalization processing: based on the inherent format characteristics of the sensor LED display, a three-in-one mechanism of regular expression matching, rule verification, and anomaly filtering is constructed to automatically clean, correct, and standardize the output for typical misidentification patterns, thereby obtaining standard sensor readings.

Images