A method for monitoring the tail rope status of hoists based on channel-space dual-domain attention.
Patent Information
- Application Number
- CN202511924117.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-19
- Publication Date
- 2026-05-26
Smart Images

Figure CN122090336A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent monitoring technology for mining equipment, and specifically discloses a method for monitoring the status of hoist tail rope based on channel-space dual-domain attention. Background Technology
[0002] As a core piece of equipment in the mine production system, the health of the tail rope of a mine hoist directly affects the safety of the entire transportation system. Traditional manual inspection methods suffer from problems such as slow response and high missed detection rates, especially in deep mines at depths of thousands of meters, where the timeliness and accuracy of manual inspections cannot meet the needs of modern mine safety management. Currently, the automated monitoring technologies commonly used in the industry mainly fall into two categories: 1) Based on the principle of laser scanning, the condition is judged by measuring the surface deformation of the tail rope. Although this approach can avoid optical interference, it has inherent defects such as high equipment costs and functional limitations. Achieving full shaft coverage requires a large investment, and it can only monitor physical deformation, unable to identify complex conditions such as entanglement; 2) Using conventional machine vision solutions, feature extraction is carried out by combining threshold segmentation and morphological processing. This method has a high accuracy rate under ideal lighting conditions; however, its performance drops sharply in actual mine environments. For example, image blurring caused by dust particles weakens the edge features of the target, and motion blur caused by hoist vibration destroys texture information, further exacerbating the difficulty of feature extraction, among other complex interference factors.
[0003] With the development of artificial intelligence technology, real-time monitoring of tail rope status using target detection models has become a new approach, saving significant manpower, material resources, and financial resources. However, general-purpose target detection models have poor adaptability to detecting special tail rope shapes, exhibit prominent detection frame offset issues, and their accuracy and robustness still have considerable room for improvement. Therefore, researching the application of target detection models in tail rope status monitoring has significant theoretical and practical value. Summary of the Invention
[0004] To address the problems in the background technology, this invention discloses a hoist tail rope status monitoring method based on channel-space dual-domain attention. By collaboratively designing channel domain and spatial domain attention mechanisms, a dynamic feature enhancement network is constructed. Combined with simulation training and edge computing deployment technology, accurate perception and rapid response of the tail rope status in complex environments are achieved.
[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The hoist tail rope status monitoring method based on channel-space dual-domain attention includes the following steps: Step 1: Using cameras deployed in the simulated experimental environment and the real field environment respectively, collect multi-dimensional video streams of the tail rope's operating status in the simulated experimental environment and the real field environment to obtain tail rope video data in the simulated experimental environment and the real field environment. Step 2: Perform frame-by-frame acquisition of the tail rope video data collected in Step 1 to obtain the tail rope image dataset; Step 3: Perform multimodal data augmentation on the tail rope image dataset obtained in Step 2, and perform classification and labeling to obtain a self-made dataset of tail ropes in different states; Step 4: Improve the YOLOv8 model by inserting a channel-space dual-domain attention optimization module, Convolutional Block Attention Module (CBAM), at the junction of the Backbone and Neck of the YOLOv8 model to construct the YOLOv8-CBAM model; Step 5: Use the YOLOv8-CBAM model from Step 4 to train the self-made tail rope dataset obtained in Step 3 until the model converges. Step Six: Deploy the model trained in Step Five to the edge intelligent device. Input the tail rope video stream captured by the camera into the edge computing unit in real time. Use the trained model to perform monitoring tasks and push the video stream of the superimposed monitoring results to the designated terminal device for real-time monitoring and analysis by the management personnel.
[0006] Specifically, in step three, gamma correction is set to balance the unevenness of illumination; adaptive histogram equalization is used to enhance the texture features of the tail rope; random salt and pepper noise is added to simulate dust interference in real-world situations; and LabelImg software is used to construct a self-made dataset of tail ropes in both simulated experimental and real-world environments. This dataset covers labeled data in three states: normal, tangled, and detached, and the ratio of labeled data used for training, testing, and validation is 6:2:2.
[0007] Specifically, in step four, a channel-spatial dual-domain attention optimization module (CBAM) is inserted at the junction of the backbone and neck of the YOLOv8 model. This includes the following steps: 4.1) The channel attention module applies global average pooling and global max pooling operations to the input feature map respectively. The results of these two pooling operations are input into a fully connected layer group with shared weights. The compression ratio is set to 16:1. Then, the channel weight matrix is generated by using the Sigmoid activation function to suppress high-frequency channels related to dust. 4.2) The spatial attention module calculates the mean and maximum values of the feature maps along the channel dimension, and concatenates them into a 2-channel feature map. It then uses a vertical preference convolution kernel to calculate the spatial weights and generates a spatial weight matrix through a Sigmoid activation function to enhance the features in the vertical direction of the tail rope. 4.3) Multiply the original feature map element-wise with the channel-space dual-domain attention weight matrix, and then output the optimized feature map to the Neck module.
[0008] Specifically, in step five, a two-stage training strategy is adopted to optimize model performance: the first stage uses simulated experimental data for pre-training, which is generated by a simulated experimental environment; the second stage uses real experimental data for transfer learning and adopts a dynamic class weight adjustment strategy.
[0009] Specifically, in step five, during the network model training phase, the image size input to the model is adjusted to 3×640×640, the batch size is 16, the total number of iterations is 200, the initial learning rate is set to 0.01, and the loss function is: ; in, These are the classification loss coefficient, localization loss coefficient, orientation constraint loss coefficient, and attention loss coefficient, respectively. These are classification loss, localization loss, orientation constraint loss, and attention loss, respectively. The classification loss ensures that the model correctly identifies the different states of the tail rope, and its formula is: ; Where N is the sample size, For real labels, Predict probabilities for the model; Positioning loss can effectively improve the fit between the detection frame and the actual position of the tail rope, and its formula is: ; in, For the improved , is the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; c is the diagonal length of the minimum bounding box; α is the balance coefficient; v is a coefficient used to measure aspect ratio consistency; w, h, and w gt h gt b and b represent the width and height of the predicted bounding box and the width and height of the ground truth bounding box, respectively. gt These are the coordinates of the center points of the predicted bounding box and the ground truth bounding box, respectively. The directional constraint loss enhances the focus on the vertical characteristics of the tail rope, and its formula is as follows: ; in, The weights are for a 1×7 convolution kernel in the vertical direction. The weights are for a 7×1 convolution kernel in the horizontal direction. To enforce the weighting percentage in the vertical direction; The attention loss is used to optimize the attention weight distribution of the CBAM module, and its formula is as follows: ; in, This is the attention weight map predicted by the model. This is the ideal weight map when the tail rope region is 1 and the background region is 0. It serves as a sparsity constraint to suppress responses in irrelevant regions.
[0010] Specifically, step six includes the following sub-steps: 6.1) During the deployment phase of this system, the trained model is first converted into a lightweight format supported by the edge device using a format conversion tool and then embedded into the storage module of the edge smart device. 6.2) Capture the raw video stream of the tail rope monitoring area in real time using camera equipment; 6.3) The edge gateway calls the hardware acceleration unit GPU or NPU to perform frame-by-frame inference on the video frames and performs the following operations: scaling the input image to the resolution specified by the model, completing normalization and channel format conversion; performing forward propagation based on the optimized computation graph, and outputting the bounding box coordinates, category and confidence of the tail rope; using the non-maximum suppression (NMS) algorithm to filter overlapping detection boxes and retain valid results; superimposing the valid results onto the original video frames to generate the superimposed image, and pushing the results to the web interface for real-time display, while also supporting alarm information pop-ups and historical video playback.
[0011] Compared with the prior art, the beneficial effects of the present invention are: (1) In terms of data collection, this invention constructs a dual-scene acquisition system of simulated experimental environment and real field environment to collect multi-dimensional video stream data of tail rope running status. The simulated environment reproduces the actual state of tail rope through parameterized control system. Based on the tail rope motion trajectory characteristics, image screening rules are established to extract typical image samples of tail rope in normal state, entangled state and detached state, providing a standardized training dataset for subsequent state recognition algorithm. (2) At the level of data augmentation and annotation, multimodal data augmentation is performed on the extracted tail rope images. Random affine transformation, Gaussian noise injection and dynamic brightness adjustment are used to improve data diversity. Semi-automatic annotation is performed through the LabelImg platform. The tail rope body category is labeled with a rectangular box, which is normal, tangled or detached. Finally, a standardized dataset containing the augmented images is generated. (3) At the algorithm architecture level, a channel-space dual-domain attention feature extraction network is designed, and the channel domain attention branch and the spatial domain attention branch are embedded into the backbone network. The channel domain attention branch adopts a group convolution strategy to select anti-interference feature channels and suppress the interference of dust occlusion and illumination noise on key features. The spatial domain attention branch captures the geometric features of the tail rope through deformable convolution operation, enhancing the model's ability to locate slender targets. The two generate an optimized feature map through a dynamic weight fusion mechanism, effectively solving the feature degradation problem caused by local occlusion and motion blur. (4) At the model training level, the YOLOv8-CBAM model is used to train the tail rope state monitoring model. The model convergence optimization is achieved through the fusion of multi-dimensional technologies. The CBAM module with integrated channel and spatial attention mechanisms is used to enhance the model's ability to capture tail rope features. A distributed training environment is deployed in combination with the PyTorch framework. The cosine annealing learning rate scheduling strategy and batch training configuration are used to improve convergence efficiency. During the training process, dynamic data augmentation strategy is implemented simultaneously. Hybrid augmentation technology is used to expand the diversity of tail rope morphology. At the same time, the parameters of some layers of the backbone network are frozen to maintain the stability of basic feature extraction. By dynamically adjusting the training cycle until the loss function tends to stabilize, a lightweight model adapted to industrial scenarios is finally generated, which supports model format conversion and edge device deployment, and realizes high-precision real-time monitoring capabilities. (5) At the system integration level, the generated model is deployed to the field equipment to form a complete automatic inspection process. By adjusting the equipment settings and optimizing the processing flow, the tail rope status monitoring system can smoothly process the tail rope images captured by the camera in real time and accurately identify the different states of the tail rope. The monitoring results will be displayed on the equipment monitoring screen in real time, which will help the staff to quickly locate the problem area. This will not only improve the efficiency of equipment maintenance, but also avoid the omissions that may occur during manual inspection. At the same time, the accumulated data can provide a reference for optimizing the equipment maintenance plan in the future. (6) This invention constructs a dynamic feature enhancement network by integrating the collaborative design of the channel-space dual-domain attention mechanism and lightweight edge computing technology. Combined with simulation training and edge computing deployment technology, it realizes accurate perception and rapid response of the tail rope status in complex environments, significantly improving the accuracy and real-time performance of tail rope status monitoring in mine hoists. It provides a high-reliability and low-latency safety guarantee, taking into account monitoring accuracy, environmental adaptability and engineering feasibility. It has important practical significance for promoting the intelligent transformation of mines, improving the inherent safety level of mines and reducing equipment maintenance costs. Attached Figure Description
[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0013] Figure 1 This is a flowchart of the hoist tail rope status monitoring method based on channel-space dual-domain attention of the present invention;
[0014] Figure 2 This is a schematic diagram of the network model structure established in this invention;
[0015] Figure 3 This is a schematic diagram of the CBAM structure in this invention;
[0016] Figure 4 This is a schematic diagram of the channel attention structure in this invention;
[0017] Figure 5 This is a schematic diagram of the spatial attention structure in this invention. Detailed Implementation
[0018] To better understand the present invention, the following embodiments further illustrate the content of the invention, but the scope of protection of the present invention is not limited to the following embodiments. Numerous specific details are set forth in the following description to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that the present invention can be practiced without one or more of these details.
[0019] Combined with appendix Figure 1-5 This paper elaborates on the hoist tail rope status monitoring method based on channel-space dual-domain attention of the present invention, and its detailed steps are as follows: Step 1: Video Data Acquisition Image acquisition systems were deployed in both the simulated experimental environment and the real-world environment to construct a multi-dimensional data source. The simulated experimental environment reproduced the normal, tangled, and detached states of the tail rope, and the video stream was transmitted to the data storage server. In the real-world environment, two mine-use explosion-proof cameras were symmetrically deployed inside the hoist shaft to ensure complementary dual-viewpoints without blind spots. The equipment met the Exd I Mb explosion-proof standard, was installed on a vibration-damping bracket, and connected to a mine-use flame-retardant optical cable to ensure equipment safety. Video streams of the tail rope's operating status were acquired in both the simulated experimental environment and the real-world environment, obtaining tail rope video data from both environments. Step 2: Time-series image acquisition and preprocessing The original video streams acquired in step one are subjected to frame synchronization and quality optimization. Key frames are extracted from both the simulated experimental environment and the real field environment video streams at a frequency of 30fps. Data timing consistency is ensured by timestamp alignment. Adaptive histogram equalization technology is used to eliminate the influence of uneven lighting and preserve the surface texture details of the tail rope. The pre-processed images are classified and stored according to their status, and a dataset containing three categories is established: normal, tangled, and detached. Step 3: Data Augmentation and Intelligent Labeling Gamma correction was used to balance the unevenness of illumination, adaptive histogram equalization was used to enhance the texture features of the tail rope, and random salt-and-pepper noise was added to simulate dust interference in real-world situations. Data enhancement was achieved through gamma correction, adaptive histogram equalization, and random salt-and-pepper noise to improve the model's generalization ability. LabelImg software was used to manually label the enhanced dataset and verify the accuracy of the labels. The labeled dataset was divided into training, validation, and test sets in a 6:2:2 ratio to ensure the objectivity of the model evaluation. Step 4: Model Architecture Optimization Design A channel-space dual-domain attention module is introduced based on the YOLOv8 model. This module consists of a cascaded channel attention module and a spatial attention module, which is inserted at the junction of the backbone and the neck to improve the targeting of feature extraction. The channel attention module adopts a dual-path parallel structure of global average pooling and global maximum response pooling. The channel attention module applies global average pooling and global maximum pooling operations to the input feature map respectively. The results of these two pooling operations are input into a fully connected layer with shared weights, with a compression ratio set to 16:1. Then, the channel weight matrix is generated by using the sigmoid activation function to dynamically suppress high-frequency noise channels related to dust. The spatial attention module calculates the mean and maximum values of the feature map along the channel dimension and concatenates them into a 2-channel feature map. The spatial weights are calculated using a 1×7 vertical bias convolution kernel, and the spatial weight matrix is generated by using the sigmoid activation function to enhance the vertical features of the tail rope. The spatial weight mask is combined to highlight the target region. Step 5: Progressive Transfer Training A two-stage training strategy is used to optimize model performance: (1) In the first stage, simulated data generated in a simulated experimental environment was used for pre-training. The initial learning rate was set to 1e-3, and the Adam optimizer and cosine annealing scheduling strategy were adopted. The batch size was set to 32. In the network model training stage, the image size of the input model was adjusted to 3×640×640, the batch size was 16, the total number of iterations was 200, the initial learning rate was set to 0.01, and the loss function was: ; in, These are the classification loss coefficient, localization loss coefficient, orientation constraint loss coefficient, and attention loss coefficient, respectively. These are classification loss, localization loss, orientation constraint loss, and attention loss, respectively. The formula for classification loss is: ; Where N is the sample size, For real labels, Predict probabilities for the model; The formula for localization loss is: ; in, For the improved , is the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; c is the diagonal length of the minimum bounding box; α is the balance coefficient; v is a coefficient used to measure aspect ratio consistency; w, h, and w gth gt b and b represent the width and height of the predicted bounding box and the width and height of the ground truth bounding box, respectively. gt These are the coordinates of the center points of the predicted bounding box and the ground truth bounding box, respectively. The formula for the orientation constraint loss is: ; in, The weights are for a 1×7 convolution kernel in the vertical direction. The weights are for a 7×1 convolution kernel in the horizontal direction. To enforce the weighting percentage in the vertical direction; The formula for attention loss is: ; in, This is the attention weight map predicted by the model. This is the ideal weight map when the tail rope region is 1 and the background region is 0. This is a sparsity constraint used to suppress responses in irrelevant regions; (2) In the second stage, real data is introduced for fine-tuning. In the initial stage, simulated data and actual data are mixed in a ratio of 7:3. Every 10 training cycles, the weight of simulated data is reduced by 10%, and finally, training is transitioned to all real data. Step Six: Edge Deployment and Application The trained model is deployed to edge intelligent devices. A model conversion tool converts the PyTorch model to ONNX format and further optimizes it into a device-specific inference engine. Video streams captured by cameras are transmitted via fiber optic cable to the edge gateway, where a built-in chip performs real-time inference. The edge gateway uses a hardware acceleration unit to perform frame-by-frame inference, performing the following operations: scaling the input image to the model's specified resolution, performing normalization and channel format conversion; performing forward propagation based on the optimized computation graph, outputting the bounding box coordinates, category, and confidence score of the tail rope; filtering overlapping detection boxes using a non-maximum suppression algorithm, retaining valid results; superimposing the valid results onto the original video frames to generate the superimposed image, which is then pushed to a web interface for real-time display. Monitoring results superimposed onto the video stream are pushed to the monitoring center via the RTMP protocol. Alarm information pop-ups and historical video playback are also supported. Administrators can view monitoring results in real-time through the web interface and receive tiered early warning information.
[0020] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Any other modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention, as long as they do not depart from the spirit and scope of the technical solutions of the present invention, should be covered within the scope of the claims of the present invention.
Claims
1. A hoist tail rope status monitoring method based on channel-space dual-domain attention, characterized in that, Specifically, the following steps are included: Step 1: Using cameras deployed in the simulated experimental environment and the real field environment respectively, collect video streams of the tail rope's running status in the simulated experimental environment and the real field environment to obtain tail rope video data in the simulated experimental environment and the real field environment. Step 2: Perform frame-by-frame acquisition of the tail rope video data collected in Step 1 to obtain the tail rope image dataset; Step 3: Perform data augmentation on the tail rope image dataset obtained in Step 2, and perform classification and labeling to obtain a self-made dataset of tail ropes in different states; Step 4: Improve the YOLOv8 model by inserting a channel-spatial dual-domain attention optimization module at the junction of the Backbone and Neck of the YOLOv8 model to construct the YOLOv8-CBAM model. Step 5: Use the YOLOv8-CBAM model from Step 4 to train the self-made tail rope dataset obtained in Step 3 until the model converges. Step Six: Deploy the model trained in Step Five to the edge intelligent device. Input the tail rope video stream captured by the camera into the edge computing unit in real time. Use the trained model to perform monitoring tasks and push the video stream of the superimposed monitoring results to the designated terminal device for real-time monitoring and analysis by the management personnel.
2. The hoist tail rope status monitoring method based on channel-space dual-domain attention as described in claim 1, characterized in that, In step three, gamma correction is set to balance the unevenness of illumination; adaptive histogram equalization is used to enhance the texture features of the tail rope. Random salt and pepper noise was added to simulate dust interference in real-world conditions. Using LabelImg software, a self-made dataset of tail ropes was constructed to simulate experimental environments and real-world environments. This dataset covers labeled data for three states: normal, tangled, and detached.
3. The hoist tail rope status monitoring method based on channel-space dual-domain attention as described in claim 1, characterized in that, In step four, a channel-space dual-domain attention optimization module is inserted at the junction of the backbone and the neck of the YOLOv8 model. This includes the following steps: (4.1) The channel attention module applies global average pooling and global max pooling operations to the input feature map respectively. The results of these two pooling operations are input into a fully connected layer group with shared weights. The compression ratio is set to 16:
1. Then, the channel weight matrix is generated by using the Sigmoid activation function to suppress high-frequency channels related to dust. (4.2) The spatial attention module calculates the mean and maximum values of the feature maps along the channel dimension, and splices them into a 2-channel feature map. It then uses a vertical preference convolution kernel to calculate the spatial weights and generates a spatial weight matrix through the Sigmoid activation function to enhance the vertical features of the tail rope. (4.3) Multiply the original feature map element-wise with the channel-space dual-domain attention weight matrix, and then output the optimized feature map to the Neck module.
4. The hoist tail rope status monitoring method based on channel-space dual-domain attention as described in claim 1, characterized in that, In step five, a two-stage training strategy is adopted to optimize model performance: the first stage uses simulated experimental data for pre-training, which is generated by a simulated experimental environment; the second stage uses real experimental data for transfer learning and adopts a dynamic class weight adjustment strategy.
5. The hoist tail rope status monitoring method based on channel-space dual-domain attention as described in claim 1, characterized in that, In step five, during the network model training phase, the image size input to the model is adjusted to 3×640×640, the batch size is 16, the total number of iterations is 200, the initial learning rate is set to 0.01, and the loss function is: in, These are the classification loss coefficient, localization loss coefficient, orientation constraint loss coefficient, and attention loss coefficient, respectively. These are classification loss, localization loss, orientation constraint loss, and attention loss, respectively. The formula for classification loss is: ; Where N is the sample size, For real labels, Predict probabilities for the model; The formula for localization loss is: ; in, For the improved , is the Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box; c is the diagonal length of the minimum bounding box; α is the balance coefficient; v is a coefficient used to measure aspect ratio consistency; w, h, and w gt h gt b and b represent the width and height of the predicted bounding box and the width and height of the ground truth bounding box, respectively. gt These are the coordinates of the center points of the predicted bounding box and the ground truth bounding box, respectively. The formula for the orientation constraint loss is: ; in, The weights are for a 1×7 convolution kernel in the vertical direction. The weights are for a 7×1 convolution kernel in the horizontal direction. To enforce the weighting percentage in the vertical direction; The formula for attention loss is: ; in, This is the attention weight map predicted by the model. This is the ideal weight map when the tail rope region is 1 and the background region is 0. It serves as a sparsity constraint to suppress responses in irrelevant regions.
6. The hoist tail rope status monitoring method based on channel-space dual-domain attention as described in claim 1, characterized in that, Step six specifically includes the following sub-steps: (6.1) During the deployment phase of this system, the trained model is first converted into a lightweight format supported by the edge device using a format conversion tool and then embedded into the storage module of the edge smart device. (6.2) Capture the raw video stream of the tail rope monitoring area in real time using camera equipment; (6.3) The edge gateway calls the hardware acceleration unit to perform frame-by-frame inference on the video frames and performs the following operations: scaling the input image to the resolution specified by the model, completing normalization and channel format conversion; performing forward propagation based on the optimized computation graph, and outputting the bounding box coordinates, category and confidence of the tail rope; The non-maximum suppression algorithm is used to filter overlapping detection boxes and retain valid results. The valid results are superimposed onto the original video frames to generate the superimposed image, which is then pushed to the web interface for real-time display. It also supports alarm information pop-ups and historical video playback.
