A safety rope abnormal behavior recognition and early warning method based on deep learning

Abstract

The application belongs to the technical field of safety monitoring, and discloses a safety rope abnormal behavior identification and early warning method based on deep learning, which can accurately detect various hidden dynamic abnormal behaviors such as not wearing, loosening, breaking and wrong hanging, and optimizes the deep learning model according to the thin strip and easily shielded graphic features of the safety rope, adopts a ShuffleNetV2 backbone network combined with a graphic feature guided attention mechanism to strengthen key feature identification and suppress background interference; a three-level grading early warning mechanism deeply bound with the matching degree of the safety rope graphic features is designed, different early warning disposal strategies are divided according to the danger level, different early warning levels not only link local sound and light, multi-terminal information pushing, but also can realize intelligent emergency linkage such as work equipment power cut-off and camera locking, and the early warning information is attached with abnormal graphic screenshots, coding comparison results and position information, so that the work personnel and the management personnel can quickly judge the abnormality degree.

Description

Technical Field

[0001] This invention belongs to the field of safety monitoring technology, specifically a method for identifying and warning of abnormal behavior of safety ropes based on deep learning. Background Technology

[0002] In high-risk scenarios such as construction, high-altitude maintenance, and cliff work, safety ropes are core protective equipment for ensuring the lives of workers. Their proper wearing, connection reliability, and condition during use directly determine operational safety. With the development of computer vision and deep learning technologies, some image recognition-based safety monitoring solutions have been applied to high-altitude work scenarios. However, existing technologies still have the following technical problems: Most solutions can only identify the single abnormality of not wearing a safety rope, and cannot detect more hidden dynamic abnormal behaviors such as slack, breakage, mishang, low hanging and high use, and empty hanging of the safety rope, resulting in insufficient targeted warnings. Existing image recognition models mostly use general target detection algorithms, which are not optimized for the graphic features of safety ropes, such as thin strips, easy occlusion, and poor scene adaptability. In scenarios with complex lighting, occlusion, and changing postures of workers, the recognition accuracy is low and the false alarm rate is high, making it difficult to meet the actual operation needs. The existing system lacks a closed-loop mechanism for identification, early warning, and response. Early warning information is not delivered in a timely manner, the classification is unclear, and the model cannot be dynamically iterated and optimized, making it difficult to adapt to the differentiated needs of different work scenarios. In addition, existing multimodal fusion solutions are mostly simple data layer splicing, without combining image recognition technology to achieve deep semantic association at the feature layer, and without designing a dedicated encoding and matching mechanism for the graphic features of safety ropes, thus limiting recognition accuracy and scene adaptability. Summary of the Invention

[0003] The purpose of this invention is to provide a method for identifying and warning of abnormal behavior of safety ropes based on deep learning, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying and warning of abnormal behavior of safety ropes based on deep learning, comprising the following specific steps: Preferably, the data acquisition phase establishes a multi-source data acquisition system covering all scenarios of high-altitude operations. The acquisition equipment includes high-definition cameras, infrared thermal imaging cameras, and attitude sensors. The acquisition equipment also includes tension sensors integrated into the safety rope nodes to acquire safety rope tension data. The acquired content includes image recognition-related data, images and video data of workers wearing safety ropes correctly, images and video data of various abnormal behaviors of the safety rope, and simultaneously acquires sample data under different scenarios, different lighting conditions, and different obstruction conditions, as well as fine-grained image samples of key graphic features of the safety rope. The collected image data is graphically labeled and encoded. A graphical feature encoding algorithm is used to encode key graphical features of the safety rope, including its outline, nodes, and connection points. The coordinates, size, texture features, and spatial relationships of each feature are labeled. The posture data is synchronously associated, and the posture parameters are bound to the corresponding graphical feature codes to construct a multimodal graphical encoding dataset for the safety rope. The dataset contains four core dimensions: graphical feature encoding, posture data, tension data, and anomaly type labels.

[0005] Graphical marking uses professional image annotation tools for pixel-level annotation. When annotating, the entire safety rope body and key feature areas are selected, and a unique annotation ID is assigned to each feature area. Annotators must have relevant knowledge of high-altitude work safety ropes. After the annotation is completed, a two-person cross-check is performed. Only annotation data that passes the check can enter the coding stage. Annotation data that fails the check is returned for re-annotation. The annotation check pass rate must reach 100%.

[0006] Preferably, the image preprocessing stage is based on the multimodal image coding dataset obtained in the data acquisition stage. Combined with image feature extraction technology, an adaptive image feature enhancement and coding preprocessing process is designed to improve the extraction accuracy of the thin strip-shaped image features of the safety rope. Subsequently, an adaptive color balance algorithm is used to dynamically adjust the brightness, contrast and color saturation of the image according to the lighting differences in different scenes. An improved Gaussian filtering algorithm is used to remove image noise while retaining key graphic features, including the edges and textures of the safety rope. A graphic recognition preprocessing module is set up to extract the thin strip contour features of the safety rope based on the improved Canny edge detection algorithm. The graphic features of the safety rope nodes and connection points are enhanced through morphological operations, and the subtle graphic features, including wear and breakage of the safety rope, are enhanced. The graphic feature encoding and verification step is incorporated into the preprocessing process. The preprocessed graphic features of the safety rope are re-encoded and compared with the initial encoding in the data acquisition stage to correct feature extraction deviations. The verified graphic feature data is aligned with the attitude sensor data, and standardized multimodal input data consisting of graphic feature encoding and attitude data is output to provide qualified input samples for the model building stage.

[0007] Preferably, in the model building stage, the standardized multimodal input data output from the image preprocessing stage is used to design a lightweight deep learning model by combining the graphic features of the safety rope and the graphic recognition neural network technology. The backbone network adopts the ShuffleNetV2 architecture, an optimized channel shuffling structure is designed, a graphic feature priority encoding module is set, the encoding information of the key graphic features of the safety rope is incorporated into the channel shuffling weight allocation, and depthwise separable convolution is integrated to reduce the computational load of the model and adapt to edge device deployment. An adaptive graphic feature selection module is set up in the neck of the model. Combined with graphic feature classification technology, the fusion weight of safety rope graphic features of different scales is dynamically adjusted to enhance the fusion of subtle graphic features. The head of the model integrates the detection structure of YOLOv7 and sets up a branch for classifying abnormal behavior of the safety rope. A graphic feature matching branch is set up synchronously. Based on graphic recognition technology, the detected graphic features are matched with the preset safety rope standard graphic encoding library, and the feature matching degree is output. The model is embedded with a graphic feature-guided attention mechanism. Based on graphic feature extraction technology, the recognition ability of key graphic features of the safety rope is enhanced, and the interference of background graphic features is suppressed.

[0008] Preferably, the fusion recognition stage is based on the lightweight deep learning model completed in the model building stage, combined with the graphic feature matching technology, to construct a multimodal feature joint embedding space of graphic feature encoding and posture data, and extract the safety rope graphic feature encoding and the three-dimensional posture data collected by the posture sensor from the preprocessed multimodal input data. An improved cosine similarity calculation method is adopted, combined with image coding matching technology, to calculate the semantic similarity between image feature coding and pose data, realizing the deep association between image features and pose data. The fused image feature coding and pose data are input into a lightweight deep learning model to perform abnormal behavior recognition and reasoning, and output the abnormal behavior type, abnormal confidence, abnormal location coordinates, pose parameters and image feature matching degree. The improved formula for calculating cosine similarity is: ,in The image feature encoding vector has 256 dimensions. The feature vectors after mapping the 3D pose data are mapped to 256 dimensions through a fully connected layer; This is the bias term, with a value of 0.1; This is the smoothing term, with a value of 0.01. This represents semantic similarity, with a value ranging from 0 to 1. A similarity score ≥ 0.7 is considered a deep feature association.

[0009] A dynamic threshold adaptive adjustment mechanism is designed to automatically adjust the anomaly recognition threshold and the graphic feature matching threshold based on the environmental complexity and graphic feature clarity of different work scenarios. At the same time, a dedicated graphic feature matching strategy is designed for different scenarios. After the anomaly information is identified, it will be directly transmitted to the early warning triggering stage.

[0010] Preferably, the early warning triggering stage, based on the core information including the abnormal behavior type, confidence level, and graphic feature matching degree output by the fusion recognition stage, and based on the danger level of the abnormal behavior, combines graphic recognition technology to design a three-level graded early warning mechanism to strengthen the correlation between early warning information and abnormal graphic features of the safety rope. Level 1 warning is triggered when the safety rope is slightly slack or slightly worn, and the corresponding graphic feature matching degree is in the high threshold range of Level 1 warning. This triggers a local audible and visual warning and simultaneously pushes a reminder message to the operator's terminal, along with a screenshot of the abnormal graphic feature and the code comparison result, clearly indicating the location of the abnormal graphic. Level 2 warning is triggered when the safety rope is mis-attached, used at a low position, or has moderate wear, and the corresponding graphic feature matching degree is below the lower limit of the Level 1 warning threshold but above the Level 2 warning threshold. This triggers an on-site audible and visual warning and simultaneously pushes information to the management personnel's terminal. The camera in the abnormal work area is locked, and the abnormal image is captured in real time. The abnormal graphic feature code is automatically extracted for the management personnel to quickly determine the degree of abnormality. Level 3 warning is triggered when the safety rope is broken, not worn, or hanging unsupported, and the corresponding graphic feature matching degree is below the Level 2 warning threshold or no valid safety rope graphic feature code is detected. This triggers the highest level warning, immediately cuts off the power to the relevant work equipment, and simultaneously pushes the warning information to the work supervisor, safety officer, and emergency management terminal, initiating the emergency response process. The warning information includes screenshots of abnormal graphics, abnormal graphic feature codes, abnormal location, abnormal type, and handling suggestions. Data generated during the warning process will be synchronously transferred to the data management stage.

[0011] The emergency response process for a Level 3 warning follows standardized core steps. First, the system automatically pushes the anomaly details to the emergency rescue team's terminal, while simultaneously displaying the anomaly location and on-site footage on the large monitoring screen. Second, the safety officer arrives at the scene within 5 minutes of receiving the warning to verify the anomaly, and the workers immediately cease high-altitude operations and evacuate to a safe area. Finally, the on-site supervisor takes appropriate measures based on the anomaly type (breakage / not wearing, etc.), such as replacing the safety rope and rectifying the attachment points. After the measures are completed, the safety officer confirms the cancellation of the warning in the system, and the warning status is lifted.

[0012] Preferably, the data management stage adopts an edge and cloud collaborative architecture for the identification data of the fusion identification stage, the early warning information and handling records of the early warning triggering stage, and combines graphic feature classification and update technology to build a data management and model dynamic evolution mechanism. A lightweight deep learning model is deployed at the edge, and a graphic feature encoding and verification module is built in to perform preliminary encoding and verification of the identified abnormal graphic features. A data management platform is built in the cloud to synchronously receive abnormal data uploaded from the edge. ShardingSphere is used to implement data sharding and storage. Data is managed by partitioning according to the work area ID and graphic feature coding type. An abnormal data full lifecycle traceability mechanism is designed. All abnormal identification, early warning and handling related data are permanently stored in the cloud. All data at the cloud and edge are encrypted and stored using symmetric encryption algorithms, and encrypted transmission protocols are used during data transmission to prevent data leakage. The cloud data management platform sets hierarchical access permissions: operators can only view warning information related to their own work, managers can view all data in their jurisdiction, and system administrators have the highest privileges and can perform data management and permission allocation. All data access operations generate operation logs, which are permanently retained and traceable to ensure data security and standardized usage.

[0013] The dynamic model update mechanism based on graphic features is incorporated into the management system. New abnormal samples accumulated in the cloud are regularly annotated with graphic features and fed back to the model training end. The graphic feature extraction module and feature matching module of the deep learning model are fine-tuned and optimized to improve the model's scene adaptability and recognition accuracy. The standard graphic encoding library of safety ropes in the cloud is updated simultaneously, and graphic feature codes for different scenes and different abnormal types are added to ensure that the model can perform matching and recognition based on the latest standard graphic features. The optimized model and related parameters are pushed to the adaptation and verification stage.

[0014] The standard graphic coding library is stored using a structured database, with a three-level directory structured by safety rope specifications, work scenarios, and anomaly types. The initial construction incorporates standard graphic feature codes for mainstream safety rope specifications in the industry, covering no fewer than 2,000 sets of feature code samples for complete states and various levels of anomaly states. The coding library supports adding, deleting, modifying, and querying feature codes, as well as version management. The coding library is updated with both scheduled and triggered updates. Scheduled updates occur every 30 days, while triggered updates are automatically initiated when the number of new anomaly samples accumulated in the cloud reaches 500. After the update is completed and passes version verification, it can be used for model matching.

[0015] Preferably, the adaptation verification stage is based on the deep learning model and related parameters optimized in the data management stage. For different operation scenarios, including building construction, cliff operation, and indoor maintenance, the performance verification related to graphic feature recognition is performed. Data acquisition equipment and early warning system are deployed in each scenario to simulate various abnormal behaviors of safety ropes and verify the model's recognition accuracy, graphic feature matching degree and early warning response speed of safety rope graphic features in different scenarios. To meet the specific needs of different scenarios, the model's graphic feature extraction parameters were adjusted, the graphic feature matching threshold was calibrated, an emergency response linkage interface was built, and it was connected with the operating equipment and emergency rescue system to verify the reliability of the three-level early warning linkage and to verify the auxiliary decision-making role of abnormal graphic feature coding in emergency response. Through field testing, test data was collected, and indicators including graphic feature recognition accuracy, code matching accuracy, and anomaly detection rate were analyzed. Early warning thresholds and model parameters were calibrated, and corresponding handling procedures were optimized.

[0016] The beneficial effects of this invention are as follows: 1. This invention can accurately detect various hidden dynamic abnormal behaviors such as not wearing, looseness, breakage, and misattachment. It optimizes the deep learning model for the thin strip-shaped and easily obscured graphic features of safety ropes, and adopts the ShuffleNetV2 backbone network combined with graphic feature-guided attention mechanism to enhance the identification of key features and suppress background interference. At the same time, by constructing a multimodal feature joint embedding space, it realizes the deep semantic association between graphic feature encoding and three-dimensional pose data, thereby improving the comprehensiveness and accuracy of anomaly identification.

[0017] 2. This invention designs a three-level graded early warning mechanism that is deeply bound to the matching degree of the safety rope's graphic features. Different early warning and response strategies are divided according to the level of danger. Different early warning levels not only link local sound and light, and multi-terminal information push, but also realize intelligent emergency linkage such as power cut-off of working equipment and camera locking. Moreover, the early warning information is accompanied by abnormal graphic screenshots, code comparison results and location information, allowing operators and managers to quickly judge the degree of abnormality. It realizes the full-link linkage of graphic feature recognition, abnormal classification, accurate early warning and rapid response, and improves the efficiency of emergency response for high-altitude operations.

[0018] 3. This invention constructs a closed-loop data management and dynamic model evolution mechanism that integrates edge and cloud computing. The edge enables millisecond-level real-time inference and local encoding verification, while the cloud achieves full lifecycle data traceability through sharded storage. It can also use accumulated new anomaly samples to fine-tune the model feature extraction and matching modules, and synchronously update the safety rope standard graphic encoding library. Combined with field testing in the adaptation and verification phase, the model parameters and early warning thresholds are calibrated for different work scenarios, realizing a closed loop of monitoring, early warning, data management, model optimization, and scenario verification. This allows the model to continuously evolve and adapt to the differentiated needs of various high-altitude work scenarios. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the method of the present invention; Figure 2 This is a flowchart of the safety rope image preprocessing process of the present invention; Figure 3 This is a flowchart illustrating the construction process of the safety rope model for this invention. Figure 4 This is a flowchart of the three-level early warning triggering and emergency response process of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figures 1 to 4 As shown, this embodiment of the invention provides a method for identifying and warning of abnormal behavior of safety ropes based on deep learning, including the following specific steps: The data acquisition phase involves building a multi-source data acquisition system that covers all scenarios of high-altitude operations. The acquisition equipment includes high-definition cameras, infrared thermal imaging cameras, and attitude sensors. The multi-source data acquisition system also includes a tension sensor integrated into the safety rope node to collect safety rope tension data. The high-definition camera supports 4K@30fps acquisition and is suitable for low-light environments; the infrared thermal imaging camera solves the problem of acquisition in low-light scenarios; the attitude sensor is integrated into the safety rope node to acquire three-dimensional attitude data. The collected data includes image recognition data, images and videos of workers wearing safety ropes correctly, and images and videos of abnormal behaviors such as safety ropes not being worn, being slack, broken, misattached, being used at a high position but not at a low position, being unattached, and being worn. Simultaneously, sample data is collected under different scenarios (tower crane operation, cliff operation, indoor maintenance), different lighting conditions (strong light, weak light, night), and different obstruction conditions (tool obstruction, clothing obstruction, environmental obstruction), as well as fine-grained image samples of key graphic features such as safety rope outline, nodes, connection points, and wear marks. Multi-source data acquisition equipment is deployed differently according to high-altitude operation scenarios. In tower crane operation scenarios, one high-definition camera and one infrared thermal imaging camera are deployed at both ends of the boom and on the operating platform. Each safety rope node is integrated with an attitude sensor and a tension sensor. In cliff operation scenarios, two high-definition cameras (multi-angle without blind spots) are deployed at the top and bottom of the operation area. Infrared thermal imaging cameras are deployed on the side of the operation area. An attitude sensor is integrated every 5 meters on the safety rope. In indoor maintenance scenarios, high-definition cameras are deployed around the operation area. Infrared thermal imaging cameras are added in areas with insufficient light. Attitude sensors are integrated at the safety rope nodes and attachment points.

[0022] The acquired image data is graphically labeled and encoded. A graphical feature encoding algorithm is used to encode key graphical features of the safety rope, such as its outline, nodes, and connection points. The coordinates, size, texture features, and spatial position of each feature are labeled. The attitude data is synchronously associated, and the attitude parameters are bound to the corresponding graphical feature codes. The attitude parameters include angle, displacement, and spatial orientation. The tension data collected by the tension sensor is synchronously associated to construct a multimodal graphical encoding dataset for the safety rope. The dataset contains four core dimensions: graphical feature encoding, attitude data, tension data, and anomaly type labels.

[0023] The tension sensor uses a miniature tension / compression sensor with a range of 0-50kN, a sampling frequency of 10Hz, an accuracy of ±0.5%FS, and outputs a 4-20mA standard analog signal. The analog signal of the tension sensor is converted into a digital signal by an industrial-grade AD converter with an AD conversion accuracy of 16 bits. The sampling frequency is consistent with that of the attitude sensor (10Hz). The converted tension data is synchronized and bound to the image feature code and attitude data at the millisecond level through timestamps. The graphic feature encoding algorithm adopts a shape context-based encoding method. First, the key graphic features of the safety rope are extracted at the pixel level. Then, a unique encoding identifier is assigned according to the feature type. The contour features are encoded according to the pixel coordinate sequence, the node / connection point features are encoded according to the center coordinate and radius size, and the wear mark features are encoded according to the texture gray value distribution range. All encoding information is stored in hexadecimal character form to achieve unique identification of features.

[0024] The image preprocessing stage is based on the multimodal image coding dataset obtained in the data acquisition stage. To address issues such as uneven lighting, noise interference, and blurred image features in the acquired data, an adaptive image feature enhancement and coding preprocessing process is designed in conjunction with image feature extraction technology to improve the extraction accuracy of the thin strip-shaped image features of the safety rope. Subsequently, an adaptive color balance algorithm is used to dynamically adjust the brightness, contrast, and color saturation of the image according to the lighting differences in different scenes, so as to solve the problem of unclear safety rope image features caused by metal reflections and shadows. An improved Gaussian filtering algorithm is used to remove image noise while retaining key graphic features such as the edges and textures of the safety rope, avoiding the loss of graphic features during the filtering process. A graphic recognition preprocessing module is set up to extract the thin strip contour features of the safety rope based on the improved Canny edge detection algorithm. The graphic features of the safety rope nodes and connection points are enhanced through morphological operations, and the subtle graphic features such as wear and breakage of the safety rope are enhanced. The improved Canny edge detection algorithm adjusts the traditional dual threshold to an adaptive multi-threshold, dynamically setting high and low thresholds based on the grayscale difference between the safety rope region and the background region. It also adds an edge connectivity detection step to compensate for the edge breakage problem of thin strip features. The improved Gaussian filtering algorithm adds a feature region mask to the traditional Gaussian filtering, using low variance filtering to preserve details in the safety rope candidate region and high variance filtering to remove noise in the background region. Morphological operations sequentially perform a combination of dilation, erosion, and opening operations. Both the dilation kernel and the erosion kernel use 3×3 rectangular structural elements. The opening operation is used to eliminate the interference of background noise on the safety rope features and enhance the contour highlighting effect of nodes and connection points.

[0025] The graphic feature encoding and verification step is incorporated into the preprocessing process. The preprocessed graphic features of the safety rope are re-encoded and compared with the initial encoding in the data acquisition stage to correct feature extraction deviations. The verified graphic feature data is aligned with the attitude sensor data, and standardized multimodal input data consisting of graphic feature encoding and attitude data is output to provide qualified input samples for the model building stage.

[0026] A dual-dimensional alignment method using timestamps and spatial coordinates is adopted. In the time dimension, the time difference between the image frame acquisition time and the attitude data acquisition time is calibrated to within ±10ms, and interpolation is performed to complete data with time mismatch. In the spatial dimension, a fixed reference point in the work scene is used as the coordinate origin, and the three-dimensional spatial coordinates of the attitude sensor are mapped and transformed with the pixel coordinates of the safety rope graphic features in the image to establish a unified spatial coordinate system, so as to achieve accurate binding of graphic feature encoding and attitude data at the same time and spatial location.

[0027] In the model construction stage, the standardized multimodal input data output from the image preprocessing stage is used to design a lightweight deep learning model by combining graphic features such as the thin strip shape of the safety rope, easy occlusion, and fragmented graphic features, as well as graphic recognition neural network technology. The backbone network adopts the ShuffleNetV2 architecture, and an optimized channel mixing structure is designed. A graphic feature priority encoding module is set up to incorporate the encoding information of key graphic features such as safety rope breakage, connection points, and wear marks into the channel mixing weight allocation, thereby improving the reuse efficiency of safety rope graphic features. The graphic feature priority coding module divides the graphic features of the safety rope into three levels of coding according to their importance. The first level of coding is for features such as breakage and connection points (coding value 10), the second level of coding is for wear marks and node features (coding value 5), and the third level of coding is for contour and texture features (coding value 1). The channel mixing weight is allocated according to the proportion of the coding value. The feature channel weight = the feature coding value / the sum of all feature coding values, so as to achieve high-weight reuse of key feature channels.

[0028] The ShuffleNetV2 backbone network consists of a four-stage feature extraction module, with each stage containing 3, 7, 3, and 1 bottleneck layers respectively. The initial number of input channels is 32, and the number of output channels for each stage is 64, 128, 256, and 512 respectively. Channel shuffling is performed after the pointwise convolution of each bottleneck layer. Depthwise separable convolution is applied to the spatial convolution stage of all bottleneck layers, splitting the standard convolution into depthwise convolution and pointwise convolution. The depthwise convolution kernel size is 3×3, and the pointwise convolution kernel size is 1×1, reducing the computational cost while ensuring feature extraction capabilities.

[0029] Integrating depthwise separable convolutions reduces model computation, adapts to edge device deployment, and enables real-time inference in the field. The model was trained using a stochastic gradient descent optimizer with a batch size of 32 and an initial learning rate of 0.001. The learning rate decayed using a cosine annealing strategy with a weight decay coefficient of 0.0005. The total number of training epochs was 200, with the first 50 epochs serving as warm-up training. The loss function was a weighted fusion of classification loss (cross-entropy loss), detection loss (CIoU loss), and feature matching loss (mean squared error loss), with weight proportions of 0.3, 0.5, and 0.2, respectively. The training dataset was divided into training, validation, and test sets in an 8:1:1 ratio. Data augmentation techniques such as random cropping, flipping, brightness adjustment, and occlusion simulation were used to improve generalization ability. Training was stopped when the validation set accuracy showed no improvement for 10 consecutive epochs, and the optimal model weights were saved.

[0030] The model is lightweighted using the ONNX format, quantized to FP16 precision, and the model size is compressed to less than 8MB. The inference frame rate deployed on edge devices is ≥25fps, and the single-frame inference time is ≤40ms. Supported edge hardware includes embedded GPUs (Jetson Nano) and industrial-grade microcontrollers (STM32H7). The model input is a standardized image of 640×640 pixels and 3D pose data, and the output is 1×N×5 detection box data (N is the number of abnormal targets) and feature matching degree values.

[0031] The model's neck is equipped with an adaptive graphic feature selection module, which, combined with graphic feature classification technology, dynamically adjusts the fusion weights of safety rope graphic features at different scales, such as thin strips at a distance, clear features at close range, and incomplete features under occlusion. This strengthens the fusion of subtle graphic features such as slight wear and minor breaks. The model's head incorporates the detection structure of YOLOv7 and sets up a safety rope abnormal behavior classification branch, which can simultaneously identify abnormal safety rope behaviors. The feature scale is divided according to the pixel proportion and integrity of the safety rope in the image. The long-distance thin strip feature is the feature with a pixel proportion of <5% and a complete outline (weight 0.3). The medium-distance semi-clear feature is the feature with a pixel proportion of 5%-10% and slight occlusion (weight 0.3). The close-distance clear feature is the feature with a pixel proportion of ≥10% and no occlusion (weight 0.2). The occluded incomplete feature is the feature with integrity of <70% (weight 0.2). The total weight is 1, covering all feature ranges.

[0032] A graphic feature matching branch is set up simultaneously. Based on graphic recognition technology, the detected graphic features are matched with the preset safety rope standard graphic encoding library, and the feature matching degree is output to improve the accuracy of anomaly identification. The model embeds a graphic feature-guided attention mechanism, which enhances the recognition ability of key graphic features of the safety rope based on graphic feature extraction technology, suppresses the interference of background graphic features, reduces the false negative rate in occluded scenes, and improves the model's sensitivity to the identification of subtle graphic anomalies.

[0033] The graphic feature-guided attention mechanism is a hybrid spatial and channel attention mechanism, connected between the adaptive graphic feature selection module of the model neck and the head detection structure. First, the channel weights of the key graphic features (contours, nodes, and connection points) of the safety rope are calculated through the channel attention branch. Then, the background area of ​​the work scene is masked and suppressed through the spatial attention branch. The attention weight calculation uses the Sigmoid activation function output, and the weight value ranges from 0 to 1. The threshold for the channel weight of the key graphic features is set to 0.6. Background channel features below this threshold will be weakened.

[0034] The fusion recognition stage is based on a lightweight deep learning model completed in the model building stage. Combined with graphic feature matching technology, it constructs a multimodal feature joint embedding space of graphic feature encoding and posture data to achieve deep semantic fusion of multimodal features. From the preprocessed multimodal input data, it extracts the graphic feature encoding of the safety rope (encoding information of contour, node and texture) and the three-dimensional posture data collected by the posture sensor. The multimodal feature joint embedding space is a 512-dimensional high-dimensional feature space, which is composed of a 256-dimensional graphic feature encoding subspace and a 256-dimensional pose data feature subspace. Feature dimension mapping is achieved through two fully connected layers (1024 and 512 neurons per layer, respectively). The activation function is ReLU, and a Dropout layer (dropout rate of 0.2) is added between layers to prevent overfitting. Features in the embedding space are normalized to the interval [0, 1].

[0035] An improved cosine similarity calculation method, combined with image coding matching technology, is employed to calculate the semantic similarity between image feature codes and attitude data, achieving a deep correlation between image features and attitude data. For example, the slack in a safety rope corresponds to the tension change in attitude data. The levels of wear and slack are defined by quantitative characteristics of graphic features. Slight wear is defined as a change in the gray value of the surface texture of the safety rope of less than 10% and no obvious scratches. Moderate wear is defined as a change in the gray value of the texture of 10%-30% and the appearance of local scratches. Severe wear is defined as a change in the gray value of the surface texture of the safety rope of ≥30% and the appearance of obvious scratches / exposed rope core, which is considered a high-risk anomaly. Slight slack is defined as a change in the outline stretch of the safety rope of less than 15% and no sagging or bending. Moderate slack is defined as a change in the outline stretch of 15%-30% and the appearance of obvious sagging or bending. Both the outline stretch and texture changes are determined based on the comparison results with the standard graphic code.

[0036] By analyzing the relationship between graphic features and the location of the anchor points, we can define "empty anchoring," "misaligned anchoring," and "low-anchoring for high use." Empty anchoring is defined as an image showing no effective connection graphic features at the safety rope anchor point, only a suspended rope, and a tension sensor reading of 0. Misaligned anchoring is defined as an image showing a mismatch between the graphic features of the safety rope connection point and the graphic features of the standard anchor point area, with the rope connection point deviating from the coordinate range of the standard anchor point. Low-anchoring for high use is defined as an image showing the pixel coordinates of the safety rope anchor point graphic features lower than the pixel coordinates of the worker's center of gravity graphic features, with the rope exhibiting an upward-pulled outline.

[0037] Simultaneously, corresponding to the contour stretching changes of the graphic feature encoding, the fused graphic feature encoding and posture data are input into a lightweight deep learning model to perform abnormal behavior recognition and reasoning, and output abnormal behavior type, abnormal confidence, abnormal position coordinates, posture parameters and graphic feature matching degree; among which abnormal behavior types include not wearing, loose, broken, etc.

[0038] The abnormal behavior inference process first compares the fused features with the standard image encoding library through the model's feature matching branch to obtain the feature matching degree. Then, the abnormal behavior classification branch performs feature extraction and classification on the fused features. The matching degree and classification results are combined for double verification. If the confidence of the classification result is ≥0.8 and the matching degree meets the threshold range, it is directly determined as the corresponding abnormal type. If the confidence is between 0.5 and 0.8, it is further verified and determined after considering the trend of the pose data. If the confidence is <0.5, it is marked as a suspected abnormality and no warning is triggered. The data is only stored in the cloud.

[0039] A dedicated review process is established in the cloud for suspected abnormal data. The system pushes suspected abnormal data to the cloud data review terminal according to the work area. Reviewers make manual judgments based on the abnormal graphic features, posture data, and on-site scene information. If the data is determined to be a real anomaly, the corresponding level of warning is manually triggered and the data is marked and added to the model training sample library. If the data is determined to be a false anomaly, it is marked as a normal sample and retained for use in optimizing the model's threshold adjustment strategy. All review results and operation records are synchronously stored in the cloud data management platform.

[0040] A dynamic threshold adaptive adjustment mechanism is designed to automatically adjust the anomaly identification threshold and the graphic feature matching threshold based on the environmental complexity and graphic feature clarity of different operating scenarios. This avoids false alarms and missed alarms caused by fixed thresholds. At the same time, a dedicated graphic feature matching strategy is designed for different scenarios. For example, in the tower crane scenario, the graphic feature matching of the metal reflective area is optimized by the image de-reflection algorithm, and in the cliff scenario, the graphic feature matching of the shadow area is optimized by the shadow enhancement algorithm. The anomaly information after identification will be directly transmitted to the early warning triggering stage.

[0041] Environmental complexity is quantified and calculated using the proportion of image occlusion area and the illumination deviation value. The proportion of occlusion area is the number of pixels occluded in the safety rope area divided by the total number of pixels in the safety rope. The illumination deviation value is the proportion of the difference between the average gray value of the safety rope area in the current image and the gray value under standard illumination. The two are weighted and summed to obtain an environmental complexity value between 0 and 1, with each weighted at 0.5. For indoor maintenance scenarios, a dedicated graphic feature matching strategy of background noise reduction + contour sharpening is adopted. First, Gaussian difference filtering is used to remove background noise such as indoor equipment and pipelines. Then, the Laplacian operator is used to sharpen the thin strip contour of the safety rope to improve the matching accuracy with the standard coding library.

[0042] The warning triggering stage, based on the core information such as the type of abnormal behavior, confidence level, and graphic feature matching degree output by the fusion recognition stage, and based on the danger level of the abnormal behavior, combines graphic recognition technology to design a three-level graded warning mechanism, strengthen the correlation between warning information and abnormal graphic features of the safety rope, and realize the linkage of graphic feature recognition, abnormal classification, accurate warning, and rapid response. Level 1 (Low Risk) Warning: Slight slack or wear of the safety rope, with the corresponding graphic feature matching degree within the Level 1 warning threshold range (high threshold range). This triggers a local audible and visual warning, simultaneously pushing a reminder message to the operator's terminal, along with a screenshot of the abnormal graphic feature and the code comparison result, clearly indicating the location of the abnormal graphic. Level 2 (Medium Risk) Warning: Mishang, low-hanging-high-use, or moderate wear of the safety rope, with the corresponding graphic feature matching degree below the lower limit of the Level 1 warning threshold but above the Level 2 warning threshold. This triggers an on-site audible and visual warning and simultaneously pushes information to the management personnel's terminal, locking the camera in the abnormal work area, capturing abnormal images in real time, and automatically extracting the abnormal graphic feature code for management personnel to quickly determine the degree of abnormality. Level 3 (High Risk) Warning: Broken, not worn, unattached, or severely worn safety rope, with the corresponding graphic feature matching degree below the Level 2 warning threshold or no valid safety rope graphic feature code detected. This triggers the highest level warning, immediately cutting off the power to the relevant work equipment, such as tower cranes and aerial work platforms, simultaneously pushing warning information to the work supervisor, safety officer, and emergency management terminal, and initiating the emergency response process. Initial threshold settings for image feature matching: Level 1 warning threshold range is 0.8-1.0, Level 2 warning threshold range is 0.5-0.8, and Level 3 warning threshold is <0.5; Dynamic adjustment rules are as follows: When the scene environment complexity (occlusion rate, illumination deviation) increases by 10%, the threshold ranges for Level 1 and Level 2 warnings are lowered by 0.05; when the environment complexity decreases by 10%, the threshold ranges are raised by 0.03. The upper and lower limits of threshold adjustment do not exceed the initial range ±0.1. The Level 3 warning threshold (<0.5) remains fixed and does not participate in dynamic adjustment. The environment complexity is quantified and calculated through image blur and occlusion area ratio, with a value of 0-1.

[0043] The three-level early warning system features differentiated audible and visual warning parameters. Level 1 warning is characterized by intermittent flashing of a blue light at a frequency of 1 flash per second, and a low-volume intermittent buzzer at a frequency of 2 flashes per second. Level 2 warning is characterized by rapid flashing of a yellow light at a frequency of 2 flashes per second, and a medium-volume continuous buzzer. Level 3 warning is characterized by alternating constant red light and strobe flashes, and a high-volume rapid continuous alarm. All audible and visual warning devices support manual pausing on-site, and the terminal push notification function is retained even after pausing.

[0044] Camera locking is achieved by calling the PTZ control interface of the video surveillance system, automatically adjusting the camera's focus to the maximum and locating the abnormal position coordinates, while simultaneously activating continuous video capture mode with a capture interval of 0.5 seconds. Power cut-off of the work equipment is achieved through an industrial-grade relay linkage interface, which is hard-connected to the electrical control system of the work equipment. When a level three warning is triggered, a power-off signal is sent directly, while retaining an emergency manual reset function. Multi-terminal push notifications include smart bracelets for workers, mobile apps for managers, industrial control computers in the on-site monitoring room, and large-screen terminals in the emergency management center. Push notifications use the MQTT IoT protocol, with a message push delay of ≤1 second, and smart bracelets simultaneously trigger vibration and light reminders.

[0045] The warning information includes screenshots of abnormal graphics, abnormal graphic feature codes, abnormal location, abnormal type, and handling suggestions. The data generated during the warning process will be synchronously transferred to the data management stage for unified management and iterative optimization.

[0046] The early warning information is managed under a closed-loop system of "whoever handles it, cancels it." After on-site handling is completed, the handling personnel must upload photos of the completed site and handling records to the system terminal, and submit a cancellation application. The safety officer verifies the completion of the handling online. If the verification is successful, the early warning is canceled in the system, and the early warning status is simultaneously lifted. If the verification fails, the cancellation application is rejected and the handling is required to be carried out again. The early warning remains in effect. The cancellation application, verification results, and handling completion materials are all bound to the original early warning data and stored, forming a closed-loop data system for the entire early warning handling process.

[0047] The data management phase, which includes the identification data from the fusion identification phase, the early warning information from the early warning triggering phase, and the handling records, adopts an edge and cloud collaborative architecture. It combines graphic feature classification and update technology to build a data management and model dynamic evolution mechanism. The lightweight deep learning model deployed at the edge can achieve millisecond-level real-time inference. At the same time, it has a built-in graphic feature encoding and verification module to perform preliminary encoding and verification of the identified abnormal graphic features to meet the real-time monitoring needs on site. The edge and cloud adopt an on-demand data exchange rule. Normally identified data without anomalies is cached on the edge's local solid-state drive (SSD) for 7 days. The system automatically cleans up cached data without anomalies older than 7 days every morning. When the edge's SSD storage utilization reaches 90%, an emergency cleanup mechanism is activated, prioritizing the cleanup of the oldest data without anomalies. At the same time, a storage warning message is pushed to the administrator's terminal. Identified abnormal data is immediately uploaded to the cloud after completing local encoding verification, while a copy is retained on the edge. Data transmission uses a compressed binary format to reduce transmission bandwidth consumption. When the edge and cloud networks are disconnected, a local offline storage strategy is activated, and abnormal data is stored on the edge's local SSD. The storage capacity supports no less than 100,000 abnormal data records. After the network is restored, abnormal data stored offline is automatically uploaded in batches to ensure no data loss.

[0048] A cloud-based data management platform is built to synchronously receive abnormal data such as images, graphic feature codes, posture data, and early warning records uploaded from the edge. ShardingSphere is used to implement data sharding and storage, and partition management is carried out according to the work area ID and graphic feature code type to improve data processing efficiency in high-concurrency scenarios. An abnormal data tracing mechanism is designed so that all abnormal identification records, early warning information, handling results, graphic feature codes, and posture data are stored in the cloud. The edge-end graphic feature encoding verification module executes a three-step verification process. First, it verifies whether the encoding format conforms to the hexadecimal identification standard. Second, it verifies whether the dimension of the feature encoding is consistent with the standard encoding library. Finally, it verifies whether the graphic feature position and posture data corresponding to the encoding match. If all three steps pass, the verification is qualified. If any step fails, it is marked as abnormal and uploaded to the cloud for review. The cloud ShardingSphere adopts a composite sharding strategy, using the job area ID as the row sharding key and dividing it into 16 data shards according to the hash algorithm. It also uses the graphic feature encoding type as the column sharding key and divides it into four types of column shards according to the range algorithm: contour, node, wear, and fracture, to achieve efficient data storage and retrieval.

[0049] A dynamic model update mechanism based on graphical features is incorporated into the management system. New anomaly samples, such as unidentified anomaly types, complex scene samples, and newly emerging anomaly graphical features accumulated in the cloud, are periodically annotated with graphical features and fed back to the model training end. The graphical feature extraction and feature matching modules of the deep learning model are then fine-tuned and optimized to improve the model's scene adaptability and recognition accuracy, enabling the model to self-evolve. The cloud-based standard graphical encoding library for safety ropes is updated synchronously, and graphical feature codes for different scenes and anomaly types are added to ensure that the model can perform matching and recognition based on the latest standard graphical features, thereby improving the stability of long-term use. The optimized model and related parameters are then pushed to the adaptation and verification stage to test the actual application effect.

[0050] The model fine-tuning adopts a mini-batch incremental training method with a batch size of 16. The initial learning rate is reduced to 1 / 10 of the original training learning rate. Only the parameters of the last two convolutional layers of the feature extraction module and the fully connected layer of the feature matching module are updated, while the parameters of the remaining layers are fixed. The training rounds are 50, and the stopping condition is that the feature matching degree on the validation set no longer improves. The fine-tuned model is first validated on the test set before being pushed to the edge for deployment.

[0051] After fine-tuning, the model is validated on the test set. The passing criteria are: image feature recognition accuracy ≥95%, encoding matching accuracy ≥98%, and anomaly false negative rate ≤1%. If all three criteria are met, the model is considered qualified and pushed to the edge for deployment. If any criterion is not met, fine-tuning is stopped, the original model parameters are restored, and the sample data of this fine-tuning is marked as abnormal. The model is then handed over to the technical staff to analyze the sample annotation / model structure issues and retrain.

[0052] A model versioning management system is established in the cloud, assigning a unique version number to each optimized model, recording the version update time, optimization content, and applicable scenarios, while also retaining historical model versions for one-click retrieval. When the edge receives a new model deployment, it first conducts a small-scale test locally. After the test passes, the original model is fully replaced. If problems such as decreased recognition accuracy or increased false alarm rate occur after the new model is deployed, the edge can automatically trigger the model rollback mechanism to immediately restore to the previous stable version. At the same time, the abnormal situation is reported to the cloud for technical personnel to investigate and optimize the problem.

[0053] The adaptation verification stage is based on the deep learning model and related parameters optimized in the data management stage. It performs performance verification related to graphic feature recognition for different operation scenarios such as building construction, cliff operation, and indoor maintenance. It deploys acquisition equipment and early warning system in each scenario, simulates various abnormal behaviors of safety ropes, and verifies the model's recognition accuracy, graphic feature matching degree and early warning response speed for safety rope graphic features (especially subtle features and occlusion features) in different scenarios. To meet the specific needs of different scenarios, the model's graphic feature extraction parameters are adjusted. For example, the feature extraction and encoding of shadow areas are optimized for cliff operations, and the feature extraction and encoding of metal reflective areas are optimized for tower crane operations. The graphic feature matching threshold is calibrated to improve scenario adaptability. An emergency response linkage interface is built to connect with operating equipment and emergency rescue systems, verify the reliability of the three-level early warning linkage, and verify the auxiliary decision-making role of abnormal graphic feature encoding in emergency response. Through field testing, test data was collected, and indicators such as graphic feature recognition accuracy, coding matching accuracy, and anomaly detection rate were analyzed. Early warning thresholds and model parameters were calibrated, and corresponding handling procedures were optimized to ensure the practicality and stability of the method in complex scenarios.

[0054] The acceptable thresholds for each performance indicator are: graphic feature recognition accuracy ≥95%, coding matching accuracy ≥98%, anomaly missed detection rate ≤1%, false alarm rate ≤2%, and early warning response speed ≤500ms. The sample size for field testing is divided according to the operation scenario. The standard sample size for each scenario is no less than 5,000 frames, and the abnormal samples such as relaxation, breakage, and mishang are no less than 1,000 frames each. The sample size for complex environments such as occlusion, strong light, and weak light is no less than 40%, ensuring that the test results cover various scenarios of actual operation.

[0055] The specific calculation method for the indicator is as follows: Graphic feature recognition accuracy: Target-level calculation, i.e., the number of targets whose graphic features of the safety rope are correctly identified / the total number of detected safety rope targets × 100%; Coding matching accuracy: (Number of correctly matched graphic feature codes / Total number of graphic feature codes) × 100%; Anomaly false negative rate: (Number of undetected true anomalies / Total number of true anomalies) × 100%, calculated separately for each anomaly type; False alarm rate: Number of falsely detected anomalies / Total number of detected anomalies × 100%, counted on an image frame basis. Multiple false alarms of the same anomaly within a single frame are counted as one.

[0056] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0057] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various 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 method for identifying and warning of abnormal behavior of safety ropes based on deep learning, characterized in that, The specific steps include the following: During the data acquisition phase, a multi-source data acquisition system covering all scenarios of high-altitude operations was built to collect images and attitude data related to safety ropes, as well as various sample data under different operating scenarios. The images were graphically labeled and encoded and associated with attitude parameters to construct a multimodal graphical encoding dataset. In the image preprocessing stage, a dedicated preprocessing workflow is designed based on the multimodal dataset to optimize the clarity of graphic features to improve recognition. After encoding verification and alignment with pose data, standardized multimodal data is output. In the model building phase, standardized multimodal data is used, combined with safety rope graphic features and graphic recognition neural network technology, to design a lightweight deep learning model adapted to high-altitude operation scenarios. The model structure and attention mechanism are optimized to achieve abnormal behavior recognition of safety ropes. At the same time, the model is adapted to edge devices to achieve real-time inference on site. In the fusion recognition stage, a multimodal feature joint embedding space is constructed based on a lightweight deep learning model, which is associated with the graphic feature encoding of the safety rope and the three-dimensional posture data. Through abnormal behavior reasoning and dynamic threshold adaptive adjustment, the core information of abnormal behavior is output, and special graphic feature matching strategies for different work scenarios are designed. During the early warning triggering phase, a three-level graded early warning mechanism is designed based on the core information of abnormal behavior. According to the level of abnormal danger, the corresponding level of early warning and emergency response linkage is triggered, the early warning content is pushed and the relevant data is output to the data management phase simultaneously. During the data management phase, an edge and cloud collaborative architecture is used to build a data management platform to realize data storage, traceability and dynamic model optimization. The safety rope standard graphic coding library is updated regularly to provide optimized models and related parameters for the adaptation and verification phase. During the adaptation and verification phase, the model adaptation and emergency response linkage verification were carried out, the model parameters were adjusted and the thresholds were calibrated, and the practicality and stability of the method were verified through field tests.

2. The method for identifying and warning of abnormal behavior of safety ropes based on deep learning according to claim 1, characterized in that, The multi-source data acquisition system in the data acquisition phase includes a high-definition camera, an infrared thermal imaging camera, and an attitude sensor. The acquired content includes data related to safety rope graphic recognition, standard and abnormal wearing image and video data, as well as sample data under different scenarios, lighting, and occlusion conditions, and fine-grained image samples of key graphic features of the safety rope. The acquisition device also includes a tension sensor integrated into the safety rope node to acquire safety rope tension data. The acquired images are graphically labeled and encoded. Key graphical features of the safety rope are specifically encoded and labeled with relevant parameters. The posture data is associated and the graphical feature codes are bound together to construct a multimodal graphical coding dataset containing four core dimensions: graphical feature codes, posture data, tension data, and anomaly type labels.

3. The method for identifying and warning of abnormal behavior of safety ropes based on deep learning according to claim 2, characterized in that, The image preprocessing stage is based on a multimodal image coding dataset. An adaptive image feature enhancement and coding preprocessing process is designed. Image parameters are dynamically adjusted according to lighting differences. While removing image noise, key image features of the safety rope are preserved, its contour features are extracted, and node, connection point features and various subtle anomaly features are enhanced. The graphic feature encoding verification is incorporated into the preprocessing process. The preprocessed graphic features are then subjected to secondary graphic feature encoding verification to correct feature extraction deviations, align graphic feature data with attitude sensor data, and output standardized multimodal data.

4. The method for identifying and warning of abnormal behavior of safety ropes based on deep learning according to claim 3, characterized in that, The lightweight deep learning model in the model building phase adopts an optimized backbone network and channel shuffling structure, and integrates lightweight convolution technology to reduce the amount of computation to adapt to edge device deployment. The model neck incorporates an adaptive graphic feature selection module to dynamically adjust the fusion weights of graphic features at different scales. The head incorporates a lightweight target detection structure and sets up an abnormal behavior classification branch and a graphic feature matching branch. A graphic feature-guided attention mechanism is embedded to enhance key feature recognition and suppress background interference.

5. The method for identifying and warning of abnormal behavior of safety ropes based on deep learning according to claim 4, characterized in that, The fusion recognition stage constructs a multimodal feature joint embedding space to extract the safety rope graphic feature encoding and three-dimensional posture data, calculates the semantic similarity between the two to achieve deep association, and inputs the deeply associated multimodal features into a lightweight deep learning model for abnormal behavior inference, outputting abnormal behavior type, abnormal behavior classification confidence, abnormal target physical space coordinates, posture parameters and graphic feature matching degree. The dynamic threshold adaptive adjustment mechanism automatically adjusts the anomaly recognition threshold and the graphic feature matching threshold according to the environmental complexity and graphic feature clarity of the work scenario, configures a dedicated graphic feature matching strategy for different work scenarios, and transmits the identified anomaly information to the early warning triggering stage.

6. The method for identifying and warning of abnormal behavior of safety ropes based on deep learning according to claim 5, characterized in that, The three-level early warning mechanism is based on the classification of abnormal danger levels, and strengthens the correlation between early warning information and abnormal graphic features of safety ropes: Level 1 warning corresponds to slight slack or slight wear of the safety rope, triggering a local audible and visual warning and pushing a reminder message, along with the results of the abnormal graphic comparison; Level 2 warning corresponds to mishanging, low hanging and high use of the safety rope, or moderate wear, triggering an on-site audible and visual warning and pushing a message to the management personnel terminal, capturing abnormal images and extracting graphic feature codes; Level 3 warning corresponds to broken, not worn, or unattached safety rope, triggering the highest level Level 3 warning, cutting off the power to the work equipment and initiating the standardized emergency response procedure. The warning information includes screenshots of abnormal graphics, graphic feature codes, physical spatial coordinates of abnormal targets, types of abnormal behavior, and handling suggestions. The warning-related data is simultaneously pushed to the data management stage.

7. The method for identifying and warning of abnormal behavior of safety ropes based on deep learning according to claim 6, characterized in that, In the edge-cloud collaborative architecture of the data management phase, a lightweight model and a graphic feature encoding and verification module are deployed at the edge to perform preliminary real-time verification of the identified abnormal graphic features; a data management platform is built in the cloud to receive abnormal data from the edge and perform sharded storage and partition management, establish an abnormal data tracing mechanism, and retain all relevant data. A dynamic model update mechanism based on graphical features is adopted. The model and feature extraction and matching modules are optimized regularly using new abnormal samples in the cloud. The standard graphical encoding library of safety ropes is updated synchronously, and the optimized model and parameters are pushed to the adaptation and verification stage.

8. The method for identifying and warning of abnormal behavior of safety ropes based on deep learning according to claim 7, characterized in that, The adaptation and verification phase involves deploying data acquisition equipment and early warning systems for different high-altitude operation scenarios, simulating various abnormal behaviors of safety ropes, and verifying the model's graphic feature recognition accuracy, graphic feature matching degree, and early warning response speed. Adjust the model's graphic feature extraction parameters and matching thresholds, connect with the operating equipment and emergency rescue system, verify the reliability of early warning and response linkage, and the role of abnormal graphic feature coding in assisting decision-making in emergency response; collect data through field tests, analyze relevant performance indicators, calibrate graphic feature matching thresholds and model-related parameters, and optimize the response process.

9. The method for identifying and warning of abnormal behavior of safety ropes based on deep learning according to claim 8, characterized in that, The graphical markings are implemented using professional image annotation tools to perform pixel-level annotations, assigning unique annotation IDs to each key feature area of ​​the safety rope.

10. The method for identifying and warning of abnormal behavior of safety ropes based on deep learning according to claim 9, characterized in that, All data at the cloud and edge are encrypted and stored using a symmetric encryption algorithm, and the data transmission process uses an encrypted transmission protocol.

Images