An AI recognition-based power operation posture monitoring and early warning method and system

Abstract

The application relates to the technical field of monitoring and early warning, and discloses a power operation posture monitoring and early warning method and system based on AI recognition, which comprises the following steps: determining a target operation posture image, performing posture analysis on the target operation posture image, generating a posture mark according to the result of the posture analysis, extracting skeleton key points of the target operation posture image based on a posture estimation algorithm when a posture suspected mark is generated, constructing a posture recognition model based on a personnel state sample set, the posture recognition model being a comprehensive model of skeleton relationship and global posture features, verifying the posture recognition model based on the personnel state sample set, adjusting the learning rate of the posture recognition model based on cosine annealing if the verification is not up to standard, determining a target posture based on the target posture recognition model and the skeleton key point information, and judging whether to issue an early warning for the target posture according to the result of overlap comparison; and the application ensures the accuracy and reliability of operation posture recognition.

Description

Technical Field

[0001] This invention relates to the field of monitoring and early warning technology, and more specifically, to a method and system for monitoring and early warning of power operation posture based on AI recognition. Background Technology

[0002] In the daily operations of the power industry, whether it's transmission line inspection, substation equipment maintenance, or distribution network maintenance, the standardization of workers' operating postures directly affects personal safety and the stable operation of the power system. Violations, such as haphazard pulling, can lead to serious accidents like electric shock, falls from heights, and equipment damage. Currently, monitoring and early warning of power operation postures relies on on-site personnel. These personnel are responsible for observing every movement of the workers in real time and judging whether it complies with safety regulations. On the one hand, long hours of monitoring can lead to fatigue and distraction, making it difficult to monitor every subtle movement and potentially overlook dangerous postures. On the other hand, differences in experience and judgment standards among different personnel can result in inconsistent assessments of the same operating posture. Once a dangerous posture occurs during operation, a unified and standardized monitoring standard cannot be established. Furthermore, the power operation environment is complex, with some scenarios located at heights, in the field, or in areas with dense high-voltage equipment, posing challenges to human monitoring. For example, in field operations, factors such as weather and terrain can affect the accuracy and timeliness of manual monitoring.

[0003] Therefore, it is necessary to design an AI-based power operation posture monitoring and early warning method and system to solve the problems existing in the current technology. Summary of the Invention

[0004] In view of this, the present invention proposes an AI-based power operation posture monitoring and early warning method and system, which aims to solve the problems that monitoring personnel cannot fully control all the subtle movements of the operators, the judgment results of the same operation posture are inconsistent, the unified and standardized monitoring standards cannot be formed, and the accuracy and timeliness of manual monitoring are insufficient.

[0005] In one aspect, this invention proposes a method for monitoring and early warning of power operation posture based on AI recognition, comprising: Acquire operational posture images, process the operational posture images to determine target operational posture images, perform posture analysis on the target operational posture images, and generate posture imprints based on the results of the posture analysis. The posture imprints include posture matching imprints, posture suspected imprints, and posture deviation imprints. When the suspected posture imprint is generated, the skeletal key points of the target work posture image are extracted based on the posture estimation algorithm, and the skeletal key point information is determined. A personnel state sample set is obtained, and a posture recognition model is constructed based on the personnel state sample set. The posture recognition model is a comprehensive model that fuses skeletal relationships and global posture features. The posture recognition model is validated based on the personnel state sample set. If the validation fails, the learning rate of the posture recognition model is adjusted based on cosine annealing, and the target posture recognition model is determined based on the adjustment result of the learning rate. The target pose is determined based on the target pose recognition model and skeletal key point information, and the target pose is compared with the corresponding standard target pose. Based on the result of the comparison, it is determined whether to issue a warning for the target pose.

[0006] Furthermore, when processing the work posture image to determine the target work posture image, the process includes: The working posture image is divided into several regions to be detected. High-frequency noise points of the working posture image are obtained based on gray-level difference processing and neighborhood pixel correlation. The regions to be detected containing the high-frequency noise points are recorded as environmental interference regions. Based on adaptive weighted median filtering, appropriate weights are set for the environmental interference area and smoothed using a filtering window to determine the smoothed signal-to-noise ratio and edge sharpness. When the signal-to-noise ratio is less than 25dB, the window size of the filtering window is increased; When the edge sharpness is less than 40, the weight of the edge of the environmental interference area is reduced. The target operation posture image is determined based on the adjusted weights and filter window.

[0007] Furthermore, when performing attitude analysis on the target operation posture image and generating attitude imprints based on the results of the attitude analysis, the attitude imprints include attitude matching imprints, attitude suspected imprints, and attitude deviation imprints, the process includes: Determine the standard target operation posture image corresponding to the target operation posture image; Extract all posture pixels from the target operation posture image, and extract the standard posture pixels corresponding to each posture pixel in the standard target operation posture image. Determine the posture pixel values ​​corresponding to all posture pixels and the standard posture pixel values ​​corresponding to all standard posture pixels. If all pose pixel values ​​are equal to the standard pose pixel values, then the pose matching imprint is generated for the target working pose image; If one or more pose pixel values ​​are not equal to the standard pose pixel values, then generate the pose imprint for the target operation pose image; If none of the attitude pixel values ​​are equal to the standard attitude pixel values, then the attitude deviation imprint is generated for the target working attitude image.

[0008] Furthermore, when acquiring the personnel status sample set, it includes: Select at least 50 power workers of different genders and heights, and deploy inertial sensors on each power worker. The deployment positions are the inside of the safety helmet, torso, left / right wrist, and left / right ankle. The sampling frequency of the inertial sensors is set to 200Hz. When a tool is dropped, the sampling frequency is set to 500Hz. Continuous sampling was conducted during high-altitude and ground operations to determine at least two thousand sets of posture samples and construct the personnel state sample set.

[0009] Furthermore, when constructing a pose recognition model based on the aforementioned personnel state sample set, the following steps are included: The personnel status sample set is divided into a model building set and a model validation set; The pose recognition model includes a skeletal relationship feature extraction subnetwork and a global pose feature extraction subnetwork; The skeletal relationship feature extraction subnetwork adopts a graph convolutional neural network structure, based on 17 human skeletal key points as graph nodes, and constructs an adjacency matrix with skeletal connection relationships. The joint angles in the model are input into 3 layers of graph convolutional layers for local relationship feature extraction, and output a 64-dimensional skeletal relationship feature vector. The global attitude feature extraction subnetwork adopts a CNN-LSTM hybrid structure. Based on the sampling points within the continuous time window of the model set, it uses the deployment position of the inertial sensor as the spatial axis to determine the motion features of the time axis and the spatial axis. The motion features include the original components of three-dimensional acceleration, the original components of three-dimensional angular velocity, the first derivative of acceleration amplitude, and the direction angle of angular velocity. A spatiotemporal matrix is ​​constructed based on the motion features. The convolutional layer of the global pose feature extraction subnetwork extracts the local spatiotemporal features of the spatiotemporal matrix and performs dimensionality reduction based on the pooling layer. The fully connected layer outputs a 64-dimensional global pose feature vector based on the dimensionality reduction result.

[0010] Furthermore, when constructing a pose recognition model based on the aforementioned personnel state sample set, the method also includes: The 64-dimensional skeletal relationship feature vector and the 64-dimensional global posture feature vector are fused. The fusion process is based on an attention mechanism algorithm to determine the skeletal feature weight coefficients of the 64-dimensional skeletal relationship feature vector and the posture feature weight coefficients of the 64-dimensional global posture feature vector. The skeletal feature weight coefficients are determined based on the joint motion amplitude, and the posture feature weight coefficients are determined based on the work scenario.

[0011] Furthermore, the pose recognition model also includes a feature fusion layer and a classifier. The feature fusion layer performs weighted summation on the 64-dimensional skeletal relationship feature vector and the 64-dimensional global pose feature vector based on the skeletal feature weight coefficient and the pose feature weight coefficient, respectively, to determine a 128-dimensional fused feature vector. The classifier contains a 3-layer fully connected network and outputs the working pose based on the Softmax activation function and the 128-dimensional fused feature vector.

[0012] Furthermore, when validating the posture recognition model based on the personnel state sample set, and adjusting the learning rate of the posture recognition model based on cosine annealing if the validation fails, and determining the target posture recognition model based on the adjustment result of the learning rate, the process includes: The model validation set is divided into a validation set and a test set, and the performance metrics of the pose recognition model are evaluated based on the validation set. If the performance metric does not reach the performance metric threshold, the learning rate of the pose recognition model is adjusted based on cosine annealing, and the performance metric of the pose recognition model is re-determined based on the validation set until the performance metric reaches the performance metric threshold. If the performance metric reaches the performance metric threshold, the performance metric of the pose recognition model is evaluated again based on the test set. If the performance index reaches the performance index threshold in the re-evaluation, then the posture recognition model is determined as the target posture recognition model. If the performance metrics fail to meet the performance threshold during the re-evaluation, the pose recognition model will be rebuilt.

[0013] Furthermore, when determining the target pose based on the target pose recognition model and skeletal keypoint information, and comparing the target pose with the corresponding standard target pose, and determining whether to issue a warning for the target pose based on the result of the comparison, the process includes: The skeletal key point information is substituted into the target pose recognition model to determine the target pose and obtain the standard target pose corresponding to the target pose. If the overlap area between the target pose and the standard target pose is greater than or equal to the overlap area threshold, then it is determined that no warning will be issued for the target pose. When the overlap area between the target pose and the standard target pose is less than the overlap area threshold, a warning is issued for the target pose.

[0014] Compared with existing technologies, the beneficial effects of this invention are as follows: By constructing a posture recognition model that fuses skeletal relationships and global posture features to replace manual real-time monitoring, it achieves stable monitoring of power operations around the clock, avoids the omission of dangerous postures due to fatigue or distraction of monitoring personnel, improves the ability to control the subtle movements of workers throughout the process, and thus reduces the risk of potential safety hazards being overlooked. The posture recognition model is verified through a personnel state sample set, and the learning rate is adjusted and the model parameters are optimized by combining cosine annealing, so that the target posture recognition model can continuously adapt to diverse work scenarios and continuously improve the accuracy of posture recognition. For complex work scenarios such as high altitude, field, and dense high-voltage equipment, the extraction of skeletal key points and fusion of global features can accurately capture the work posture, avoid interference from environmental factors such as weather and terrain, and improve the accuracy and timeliness of monitoring and early warning. Based on the validated target posture recognition model, the operation posture can be objectively analyzed, eliminating the reliance on the experience level and subjective judgment of the monitoring personnel. This ensures that the judgment results for the same posture remain consistent, thereby forming a standardized and consistent monitoring standard. It reduces the judgment deviation caused by human factors. By comparing the overlap between the target posture and the standard target posture, it can quickly identify the non-compliant posture and issue timely warnings, effectively preventing accidents such as electric shock and falls from heights.

[0015] On the other hand, this application also provides an AI-based power operation posture monitoring and early warning system, used to apply the above-mentioned AI-based power operation posture monitoring and early warning method, including: The acquisition and analysis module is configured to acquire work posture images, process the work posture images, determine target work posture images, perform posture analysis on the target work posture images, and generate posture imprints based on the results of the posture analysis. The posture imprints include posture matching imprints, posture suspected imprints, and posture deviation imprints. The model building module is configured to extract the skeletal key points of the target work posture image based on the posture estimation algorithm when the suspected posture imprint is generated, determine the skeletal key point information, obtain the personnel state sample set, and construct a posture recognition model based on the personnel state sample set. The posture recognition model is a comprehensive model that fuses skeletal relationships and global posture features. The model validation module is configured to validate the posture recognition model based on the personnel state sample set. If the validation fails, the learning rate of the posture recognition model is adjusted based on cosine annealing, and the target posture recognition model is determined based on the adjustment result of the learning rate. The posture warning module is configured to determine the target posture based on the target posture recognition model and skeletal key point information, and to perform an overlap comparison between the target posture and the corresponding standard target posture, and to determine whether to issue a warning for the target posture based on the result of the overlap comparison.

[0016] It is understandable that the above-mentioned AI-based power operation posture monitoring and early warning method and system have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating an AI-based power operation posture monitoring and early warning method provided in this embodiment of the invention; Figure 2 This is a functional block diagram of an AI-based power operation posture monitoring and early warning system provided in an embodiment of the present invention. Detailed Implementation

[0018] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] See Figure 1 As shown in some embodiments of this application, a power operation posture monitoring and early warning method based on AI recognition includes: S100: Acquires and processes the operation posture image, determines the target operation posture image, performs posture analysis on the target operation posture image, and generates posture imprints based on the posture analysis results. The posture imprints include posture matching imprints, posture suspected imprints, and posture deviation imprints. S200: When generating a suspected posture imprint, extract the skeletal key points of the target operation posture image based on the posture estimation algorithm, determine the skeletal key point information, obtain the personnel state sample set, and construct a posture recognition model based on the personnel state sample set. The posture recognition model is a comprehensive model that fuses skeletal relationships and global posture features. S300: The posture recognition model is validated based on the personnel state sample set. If the validation fails, the learning rate of the posture recognition model is adjusted based on cosine annealing, and the target posture recognition model is determined based on the adjustment result of the learning rate. S400: Determines the target pose based on the target pose recognition model and skeletal key point information, and compares the target pose with the corresponding standard target pose. Based on the result of the comparison, it determines whether to issue a warning for the target pose.

[0020] Specifically, infrared cameras and vision devices are used to acquire images of the worker's posture. During image acquisition, due to the unique environment of power operations—which may involve working in the field or at high altitudes—adverse weather conditions such as rain and snow can interfere with the images, making it difficult to accurately determine the worker's posture. For example, in heavy snowfall, falling snowflakes create dynamic noise in the image, causing some skeletal keypoint extraction algorithms to mistakenly identify snowflakes as keypoints, leading to extraction errors. Therefore, after acquiring the work posture images, they are first processed to filter out potential interference factors such as rain and snow, thereby focusing on the details of the worker's posture and improving the accuracy of subsequent analysis. The algorithm dynamically generates posture matching imprints, suspected posture imprints, and posture deviation imprints based on the target work posture images. This classifies the worker's posture for targeted processing, identifies potential violations, and reduces invalid analysis. When a suspected posture imprint is found, the worker is considered to have a potential violation and requires further analysis. The posture estimation algorithm uses visual technology to identify the position of human joints, such as skeletal key points like the shoulder, elbow, wrist, and hip. The algorithm accurately captures the skeletal structure of human movement. Skeletal key point information includes the spatial coordinates and motion angles of these joints, providing quantitative data for subsequent posture recognition. A comprehensive model fusing skeletal relationships and global posture features is constructed based on a personnel state sample set. This is because single skeletal relationships (connections and angles between joints) or global posture features (overall movement shape) have certain limitations. The fused model can comprehensively describe the posture and the worker's posture details, improving the robustness of posture recognition. The personnel state sample set contains a large amount of data on different workers in various postures, providing the foundation for model training. The personnel state sample set was used to validate the posture recognition model. Validation metrics included accuracy, recall, and F1 score. If the metrics were not met, cosine annealing was used to adjust the learning rate. Cosine annealing is a learning rate adjustment strategy that simulates a cosine function, causing the learning rate to initially decrease rapidly and then increase slowly, avoiding the model getting trapped in local optima and finding better parameters to derive the target posture recognition model. Finally, skeletal keypoint information was substituted into the target posture recognition model to determine the target posture. By fusing the spatial topological relationships of skeletal keypoints with global motion features, the target posture recognition model achieved accurate analysis and dynamic tracking of power operation postures. The target posture was overlaid and compared with the corresponding standard target posture. Based on the overlay comparison results, a warning was issued for the target posture, avoiding the subjectivity and blindness of human judgment and improving the accuracy and reliability of monitoring and warning.

[0021] Understandably, by constructing and validating a posture recognition model, and then determining the target posture recognition model, it is possible to provide 24-hour uninterrupted monitoring and early warning for workers based on image acquisition equipment. This avoids missed judgments caused by fatigue or distraction during manual monitoring, thereby capturing subtle dangerous postures and preventing inconsistencies in judgments due to differences in experience and cognition among different monitoring personnel. It ensures consistent evaluation of the same posture. For scenarios where manual monitoring is difficult, such as high altitudes, outdoor areas, and areas with dense high-voltage equipment, image processing and the constructed model can effectively overcome interference from weather and terrain, ensuring the timeliness and reliability of monitoring and providing strong protection for the safety of power operations.

[0022] In some embodiments of this application, when processing the work posture image to determine the target work posture image, the process includes: dividing the work posture image into several detection regions; obtaining high-frequency noise points in the work posture image based on gray-level difference processing and neighborhood pixel correlation; recording the detection regions containing high-frequency noise points as environmental interference regions; setting corresponding weights for the environmental interference regions based on adaptive weighted median filtering and smoothing them using a filtering window; determining the smoothed signal-to-noise ratio and edge sharpness; increasing the window size of the filtering window when the signal-to-noise ratio is less than 25dB; reducing the weight of the edge of the environmental interference region when the edge sharpness is less than 40; and determining the target work posture image based on the adjusted weights and the filtering window.

[0023] Specifically, the work posture image is divided into several detection regions, the number of which can be dynamically adjusted according to the actual size of the work posture image. Gray-level difference processing is used to capture regions with abrupt changes in pixel values ​​(high-frequency noise features). Combined with neighborhood pixel correlation analysis (high correlation in normal areas and low correlation in noisy areas), environmental interference areas are accurately located. These interference areas may contain noise clusters caused by rain or snow. Adaptive weighted median filtering is used to assign differentiated weights to the environmental interference areas. Areas with dense noise are typically given higher weights to enhance filtering, while areas with relatively dispersed noise (usually located at the edges) are given relatively lower weights to reduce damage to image details. Smoothing is achieved through a filtering window. The signal-to-noise ratio (SNR) reflects the quality of the filtered image; rain, snow, and other interferences increase image noise, leading to a decrease in SNR. During light rain and snow, the signal-to-noise ratio (SNR) is typically between 30 and 40 dB. However, during moderate to heavy rain and snow, the SNR drops to 20 to 30 dB. At this point, the noise significantly affects the recognition of pose details. When the SNR is greater than or equal to 25 dB, it indicates that the noise from rain and snow is relatively weak, and pose details in the image, such as human limb movements, are preserved relatively completely. There is no need to excessively increase the filter window to avoid over-filtering and loss of details. When the SNR is less than 25 dB, it indicates that rain and snow noise will significantly interfere with key information, such as hand movements and the limb contours of climbing postures. It is necessary to increase the filter window size to enhance the denoising effect and ensure the accuracy of subsequent pose recognition. Edge sharpness is the average magnitude of the gradient at the edges of an image (such as the gradient value calculated by the Sobel operator). Higher edge sharpness indicates a sharper boundary between the worker and the background environment. Rain and snow can blur edges, causing the gradient magnitude to decrease. In the absence of rain or snow, the average magnitude of the edge gradient is usually between 50 and 80. Under rain and snow interference, it will drop to 30-50. At this time, the worker's limbs are easily confused with the background. When the edge sharpness is greater than or equal to 40, it means that the blurring effect of rain and snow on the edges is low. The limb transitions of the worker's posture, such as the contours of wrist and ankle movements, remain sharp, and there is no need to reduce the weight of the edges. When the edge sharpness is less than 40, it means that rain and snow have blurred the edges. For example, the boundary between the worker's hand and the tool is blurred. In this case, it is necessary to reduce the weight of the edges in the environmental interference area and reduce the excessive smoothing of the edges by the filter to ensure that the positional relationship of the limb joints can be accurately captured during subsequent posture recognition.

[0024] Understandably, the target operation posture image determined by the adjusted weights and filtering window can effectively remove rain and snow noise while preserving the details required for power operation posture recognition to the maximum extent, thereby balancing the requirements of noise reduction effect and detail preservation, and improving the accuracy and reliability of subsequent posture recognition.

[0025] In some embodiments of this application, when performing attitude analysis on a target work posture image and generating attitude imprints based on the results of the attitude analysis, the attitude imprints include attitude matching imprints, attitude suspected imprints, and attitude deviation imprints, the process includes: determining a standard target work posture image corresponding to the target work posture image; extracting all posture pixels from the target work posture image; extracting the standard posture pixels corresponding to each posture pixel in the standard target work posture image; determining the posture pixel values ​​corresponding to all posture pixels; and determining the standard posture pixel values ​​corresponding to all standard posture pixels. If all posture pixel values ​​are equal to the standard posture pixel values, an attitude matching imprint is generated for the target work posture image. If one or more posture pixel values ​​are not equal to the standard posture pixel values, an attitude suspected imprint is generated for the target work posture image. If none of the posture pixel values ​​are equal to the standard posture pixel values, an attitude deviation imprint is generated for the target work posture image.

[0026] Specifically, a standard target working posture image corresponding to the target working posture image is determined, and this standard target working posture image is used as the benchmark for compliance posture. The standard target working posture image is taken in accordance with the mandatory standards of the power industry. For example, in high-altitude operations on transmission lines, the distance between the worker and the live conductor must be greater than or equal to the safe distance (e.g., ≥0.7 meters for 10kV lines). All pixels in the target working posture image that can reflect the working posture, such as limb contours and joint positions, are extracted, i.e., posture pixels. Since there is a one-to-one correspondence between the target working posture image and the standard target working posture image, the posture pixels also correspond one-to-one with the standard posture pixels. A posture imprint is generated by comparing the pixel values ​​of the two one by one. When all corresponding posture pixel values ​​match perfectly, it means that the actual posture is completely consistent with the standard posture required by the specification, and a posture conformity imprint is generated, allowing the power work to continue. When some posture pixel values ​​do not match, it indicates that there may be a local deviation in the posture, but it has not completely deviated from the standard. In this case, a suspected posture imprint is generated for further analysis of the target operation posture image. When all posture pixel values ​​do not match, it means that the posture does not conform to the standard at all. In this case, a posture deviation imprint is generated, and an early warning is issued immediately to remind the operator of the working status. The early warning can be given through sound and light alarms. By comparing pixels, differences in posture can be captured, avoiding simple right or wrong judgments. This ensures the accuracy and reliability of posture judgment, avoids the uncertainty and blindness of human monitoring, and thus improves the efficiency and accuracy of power operation monitoring.

[0027] In some embodiments of this application, the process of obtaining a personnel state sample set includes: selecting at least 50 power workers of different genders and heights, deploying inertial sensors on each power worker, with deployment positions being the inside of the safety helmet, torso, left / right wrist, and left / right ankle. The sampling frequency of the inertial sensors is set to 200Hz, and when a tool is dropped, the sampling frequency is set to 500Hz. Continuous sampling is performed during high-altitude and ground operations to determine at least two thousand sets of posture samples and construct a personnel state sample set.

[0028] Specifically, at least 50 power workers of different genders and heights were selected because there are individual differences in the physiological characteristics of different people, such as limb length and joint range of motion. Height differences will lead to different limb extension ranges during operation. For example, when a 1.9-meter and a 1.6-meter worker operate the same live equipment, the arm raising angle can differ by more than 20°. Gender differences may lead to differences in the strength and coordination of movements. For example, female power workers usually have a higher stride frequency than males when climbing towers. Diverse sample data can cover the postural characteristics of people of different body types and avoid the model from making identification biases for specific groups. Inertial sensors are deployed inside the safety helmet, on the torso, left / right wrists, and left / right ankles. These sensors are determined based on the posture characteristics of electrical workers. The sensors inside the helmet capture head pitch and tilt angles, while the torso sensors record body tilt and center of gravity shift. For example, a torso tilt exceeding 30° forward during high-altitude work may cause imbalance. The wrist sensors capture the operational trajectory of hand movements, and the ankle sensors record the support status of the lower limbs. Combining the sample data from these six locations allows for a complete reconstruction of the electrical worker's three-dimensional motion posture. The 200Hz sampling frequency utilizes the Nyquist criterion to fully preserve the motion. The detailed sampling process avoids signal aliasing, thus meeting the needs of power operations. When a tool falls, it triggers 500Hz high-frequency sampling because tool drop is a sudden dynamic event; for example, the wrist changes abruptly the moment a wrench falls. High-frequency sampling can capture millisecond-level abrupt changes in motion, providing crucial evidence for identifying dangerous postures. Covering both high-altitude and ground-based work scenarios is essential because the posture risks differ between the two. A typical dangerous posture in high-altitude work (such as pole maintenance) is the body being suspended in the air, while in ground-based work (switchgear operation), the dangerous posture often involves the hand accidentally entering a live area. Continuous sampling accumulates at least two thousand sets of samples, ensuring coverage of routine operations (such as normal climbing and tool retrieval), complex conditions (such as high-altitude balance adjustments in strong winds), and dangerous actions. The large number of posture samples effectively reduces posture recognition errors. The diverse characteristics of personnel and the coverage of different scenarios allow the subsequent model to adapt to workers of different body types and complex working environments, thereby reducing generalization errors and laying a reliable data foundation for real-time monitoring and early warning of power operations.

[0029] In some embodiments of this application, when constructing a pose recognition model based on a personnel state sample set, the process includes: dividing the personnel state sample set into a model construction set and a model validation set; the pose recognition model includes a skeletal relationship feature extraction subnetwork and a global pose feature extraction subnetwork; the skeletal relationship feature extraction subnetwork adopts a graph convolutional neural network structure, using 17 human skeletal key points as graph nodes, and constructing an adjacency matrix based on skeletal connection relationships; inputting joint angles from the model construction set into three layers of graph convolutional layers for local relationship feature extraction, and outputting a 64-dimensional skeletal relationship feature vector; and global pose feature extraction... The subnetwork adopts a CNN-LSTM hybrid structure. It uses sampling points within a continuous time window of the model construction set as the time axis and the deployment position of the inertial sensor as the spatial axis to determine the motion features of the time axis and the spatial axis. The motion features include the original components of three-dimensional acceleration, the original components of three-dimensional angular velocity, the first derivative of acceleration amplitude, and the direction angle of angular velocity. Based on the motion features, a spatiotemporal matrix is ​​constructed. The convolutional layer of the global attitude feature extraction subnetwork extracts the local spatiotemporal features of the spatiotemporal matrix and performs dimensionality reduction according to the pooling layer. The fully connected layer outputs a 64-dimensional global attitude feature vector based on the dimensionality reduction result.

[0030] Specifically, the sample set is divided into a model building set and a model validation set, typically in a 7:3 ratio. The model building set is used to build the pose recognition model, while the model validation set is used to evaluate the performance of the built model and ensure its generalization ability on new data. The skeletal relationship feature extraction subnetwork is based on a graph convolutional neural network (GCN). The joints of the human body have certain topological relationships, and GCN is good at processing this type of data. The 17 human skeletal key points, including the left shoulder, right shoulder, left elbow, and right elbow, are used as graph nodes. The adjacency matrix uses 0-1 values ​​to indicate whether there is a skeletal connection between nodes. For example, the left shoulder and right shoulder have no direct connection, which is recorded as 0, and the shoulder and elbow have a connection, which is recorded as 1. After inputting joint angles, such as the elbow joint bending angle and the hip joint rotation angle, the three graph convolutional layers aggregate the neighbor features of each node, such as calculating the feature association between the shoulder node and the elbow node, and gradually extract the spatial constraint relationship of the local skeleton, such as the coordinated change of the shoulder joint and elbow joint angle when raising the arm. Finally, a 64-dimensional skeletal relationship feature vector is output. The global pose feature extraction subnetwork adopts a hybrid CNN-LSTM structure. Since power operation posture is a spatiotemporally continuous dynamic process, such as continuous movements when climbing a tower, a convolutional neural network (CNN) is used to extract spatial features, and a long short-term memory network (LSTM) is used to capture the time series. The time axis represents continuous sampling points at 200Hz, and the spatial axis represents the deployment positions of the six inertial sensors. All inertial sensors uniformly adopt a three-axis configuration with the torso as the origin and the front-back, left-right, and up-down axes to eliminate subtle deviations in individual deployments. Motion features encompass the original components of three-dimensional acceleration, the original components of three-dimensional angular velocity, and the first derivative of the acceleration amplitude. The first derivative of the acceleration amplitude reflects... The rate of change of motion acceleration and the direction angle of angular velocity reflect the rotation direction of the limb. These features can comprehensively characterize the intensity, direction and dynamic trend of the movement of different workers. The spatiotemporal matrix integrates the time series and spatial position. The convolutional layer of the global posture feature extraction subnetwork extracts the local spatiotemporal features of the spatiotemporal matrix, such as dangerous movements where the wrist acceleration surge is accompanied by changes in ankle support force. The features are then reduced in dimensionality by the pooling layer. The fully connected layer compresses the reduced features into a 64-dimensional global posture feature vector. The 64-dimensional skeletal relationship feature vector reflects the static and quasi-static features of how the limbs coordinate locally, while the 64-dimensional global posture feature vector reflects the dynamic features of the overall movement's dynamic trend changing over time.

[0031] Understandably, the combination of local skeletal relationships and global spatiotemporal dynamics captures both joint coordination relationships and identifies continuous movement trends. The complementary features of the two paths ensure the accuracy and reliability of pose recognition, thereby improving the model's robustness to complex pose recognition and avoiding the blindness and uncertainty of human experience judgment.

[0032] In some embodiments of this application, when constructing a posture recognition model based on a set of personnel state samples, the method further includes: fusing the 64-dimensional skeletal relationship feature vector and the 64-dimensional global posture feature vector. The fusion process determines the skeletal feature weight coefficients of the 64-dimensional skeletal relationship feature vector and the posture feature weight coefficients of the 64-dimensional global posture feature vector based on an attention mechanism algorithm. The skeletal feature weight coefficients are determined based on the joint motion amplitude, and the posture feature weight coefficients are determined based on the work scenario.

[0033] Specifically, the fusion process is based on an attention mechanism algorithm, which simulates the selective focusing of human attention and assigns higher weights to features more important to the current task. For the 64-dimensional skeletal relationship feature vector, the skeletal feature weight coefficient is determined by the range of motion of the joints. When a joint, such as the wrist, has a large range of motion, it indicates that it is the core movement part of the posture, and the skeletal feature weight coefficient increases, for example, from 0.3 to 0.6. Conversely, for joints with slow movement, such as the ankle when at rest, the skeletal feature weight coefficient decreases relatively, ensuring that key local movements are identified first. For the 64-dimensional global posture feature vector, the posture feature weight coefficient is... Based on the working scenarios of at least 50 power workers, when working at heights, features such as the overall center of gravity shift and body tilt angle are more critical. Therefore, the weight coefficient of the posture features related to overall balance and safe distance in such scenarios is increased to 0.7. When working on the ground, the relative position of local limbs and live equipment is more important, and the weight coefficient of the posture features can be appropriately reduced. This enables the model to dynamically adjust to adapt to the scenario. The dynamic weight allocation allows local actions to be organically combined with the global scenario, avoiding the limitations of a single feature, focusing on key information, reducing interference from irrelevant features, and improving the accuracy of the model in posture recognition.

[0034] In some embodiments of this application, the pose recognition model further includes a feature fusion layer and a classifier. The feature fusion layer performs weighted summation on the 64-dimensional skeletal relationship feature vector and the 64-dimensional global pose feature vector based on the skeletal feature weight coefficient and the pose feature weight coefficient, respectively, to determine a 128-dimensional fused feature vector. The classifier contains a 3-layer fully connected network and outputs the working pose based on the Softmax activation function and the 128-dimensional fused feature vector.

[0035] Specifically, the core of the feature fusion layer is used for weighted summation and dimensional concatenation. For the 64-dimensional skeletal relationship feature vector, each dimension is weighted according to the skeletal feature weight coefficient, such as 0.6 for the wrist and 0.1 for the ankle, to enhance the local features of key joints. The 64-dimensional global pose feature vector is weighted overall according to the pose feature weight coefficient, highlighting the global dynamics of scene adaptation. The two weighted vectors are concatenated to form a 128-dimensional fused feature vector. This 128-dimensional fused feature vector retains both the 64-dimensional local details of the skeletal features and the 64-dimensional spatiotemporal trends of the global features, achieving complete preservation of both local and global information. The classifier uses a three-layer fully connected network to handle feature mapping and classification decisions. The first layer maps the 128-dimensional fused features to a higher-dimensional space, capturing complex relationships between features, such as the dangerous combination of large wrist movements and torso tilt. The second layer reduces the dimensionality of the features in the higher-dimensional space, thereby compressing redundant information. The third layer uses the Softmax activation function to map the features to the probability distribution of the pose and takes the pose with the highest probability as the output, thus achieving timely and accurate pose recognition and avoiding pose misjudgment or missed judgment caused by human experience bias.

[0036] In some embodiments of this application, when validating a posture recognition model based on a set of personnel state samples, adjusting the learning rate of the posture recognition model based on cosine annealing if the validation fails, and determining the target posture recognition model based on the adjustment result of the learning rate, the process includes: dividing the model validation set into a validation set and a test set; evaluating the performance index of the posture recognition model based on the validation set; if the performance index does not reach the performance index threshold, adjusting the learning rate of the posture recognition model based on cosine annealing, and re-determining the performance index of the posture recognition model based on the validation set until the performance index reaches the performance index threshold; if the performance index reaches the performance index threshold, re-evaluating the performance index of the posture recognition model based on the test set; if the re-evaluated performance index reaches the performance index threshold, determining the posture recognition model as the target posture recognition model; if the re-evaluated performance index does not reach the performance index threshold, reconstructing the posture recognition model.

[0037] Specifically, the model validation set is divided into a validation set and a test set. The validation set is used to fine-tune the model after validation, while the test set is used for the final evaluation of the model. The typical ratio is 4:1 to avoid overfitting due to evaluation on a single dataset. Model performance metrics are evaluated based on the validation set, including accuracy and recall. If the performance metric thresholds are not met, such as accuracy less than 0.85 and recall less than 0.9, cosine annealing is used to adjust the model's learning rate. Cosine annealing simulates the periodic changes of the cosine function, causing the learning rate to first decrease slowly to a minimum and then quickly rebound. This avoids parameter oscillations caused by an excessively high learning rate and prevents getting stuck in local optima due to an excessively low learning rate. Periodic adjustments help the model achieve the required performance. Once the performance metrics validated on the validation set reach the performance threshold, the model is evaluated again using the test set, forming a dual validation mechanism. The test set, as independent data that did not participate in model construction and validation, can truly reflect the model's generalization ability. If the performance metrics on the test set reach the performance threshold, it indicates that the model performs stably on new data and is thus identified as the target pose recognition model. If the performance threshold is not reached, it indicates that the model overfitted the model construction set and validation set during the construction and validation processes, respectively. In this case, the pose recognition model needs to be reconstructed to fundamentally solve the model's generalization defects.

[0038] Understandably, the optimization of the validation set ensures the convergence of the model, while the final verification of the test set guarantees the generalization ability of the model. This allows the target posture recognition model to achieve a balance between convergence and fine-tuning, thereby ensuring that the target posture recognition model can stably identify the working posture in the complex scenario of power operation. Through closed-loop optimization of evaluation, adjustment and re-verification, the error of posture recognition is reduced, ensuring the stability and reliability of posture recognition.

[0039] In some embodiments of this application, when determining the target pose based on the target pose recognition model and skeletal key point information, and comparing the target pose with the corresponding standard target pose, and determining whether to issue a warning for the target pose based on the result of the overlap comparison, the method includes: substituting the skeletal key point information into the target pose recognition model to determine the target pose, obtaining the standard target pose corresponding to the target pose, and determining that no warning is issued for the target pose when the overlap area between the target pose and the standard target pose is greater than or equal to the overlap area threshold, and determining that a warning is issued for the target pose when the overlap area between the target pose and the standard target pose is less than the overlap area threshold.

[0040] Specifically, skeletal keypoint information is input into the target pose recognition model. This keypoint information includes the spatial coordinates and motion angles of joints such as the shoulder, elbow, wrist, and hip. The target pose is determined based on the model's output. Simultaneously, a standard target pose corresponding to the work scenario is invoked. This standard pose is determined based on industry standards and the standardized action postures performed by qualified senior power workers in a simulated work environment, strictly adhering to specifications. A pose imprint is generated based on the standard target pose image for overlay comparison. However, directly comparing the target pose image with the standard target pose image to warn of suspected issues can be problematic. Variations in the worker's movement angle and distance from the image acquisition device can lead to differences in pixel distribution for the same standardized pose from different viewpoints, making it difficult to adapt to such diversity. Furthermore, in power work, tools and equipment may partially obscure parts of the body, causing variations in the pose in the image. Lacking pixel values ​​makes effective judgment impossible. Therefore, a target posture recognition model can better utilize big data to effectively identify work postures. By calculating the intersection of the target posture and the standard target posture (overlap area), an overlap area threshold is dynamically set. This threshold is adjusted based on whether tools or equipment partially obscure the body. For example, an insulating pole (1.5-3 meters long) might obscure the forearm and hand, accounting for about 40% of the upper limb outline. If a conventional overlap area threshold (90%) is used, the obscured upper limb would reduce the overall overlap area, leading to false alarms. Therefore, the overlap area threshold for the upper limb is lowered to 50%, considering only the unobscured parts, while the torso, lower limbs, and other unobscured parts maintain the conventional overlap area threshold. This allows for reduced local overlap due to tool obstruction while using unobscured parts for regular judgment, ensuring the accuracy and stability of overall posture recognition. Warnings can also be issued through audible and visual alarms. Contour overlap calculation based on skeletal key points avoids the subjectivity of human judgment and improves the reliability of recognition in interference scenarios.

[0041] In summary, the beneficial effects of this invention are as follows: By constructing a posture recognition model that fuses skeletal relationships and global posture features to replace manual real-time monitoring, it achieves stable monitoring of power operations around the clock, avoids the omission of dangerous postures due to fatigue or distraction of monitoring personnel, improves the ability to control the subtle movements of workers throughout the process, and thus reduces the risk of potential safety hazards being overlooked. The posture recognition model is validated through a personnel state sample set, and the learning rate is adjusted and model parameters are optimized by combining cosine annealing, so that the target posture recognition model can continuously adapt to diverse work scenarios and continuously improve the accuracy of posture recognition. For complex work scenarios such as high altitude, field, and dense high-voltage equipment, the extraction of skeletal key points and fusion of global features can accurately capture work postures, avoid interference from environmental factors such as weather and terrain, and improve the accuracy and timeliness of monitoring and early warning. Based on the validated target posture recognition model, the operation posture can be objectively analyzed, eliminating the reliance on the experience level and subjective judgment of the monitoring personnel. This ensures that the judgment results for the same posture remain consistent, thereby forming a standardized and consistent monitoring standard. It reduces the judgment deviation caused by human factors. By comparing the overlap between the target posture and the standard target posture, it can quickly identify the non-compliant posture and issue timely warnings, effectively preventing accidents such as electric shock and falls from heights.

[0042] In another preferred embodiment based on the above embodiments, see [reference] Figure 2 As shown, this embodiment provides an AI-based power operation posture monitoring and early warning system, used to apply the above-mentioned AI-based power operation posture monitoring and early warning method, including: The acquisition and analysis module is configured to acquire operation posture images, process the operation posture images, determine the target operation posture image, perform posture analysis on the target operation posture image, and generate posture imprints based on the posture analysis results. The posture imprints include posture matching imprints, posture suspected imprints, and posture deviation imprints. The model building module is configured to extract the skeletal key points of the target work posture image based on the posture estimation algorithm when generating a suspected posture imprint, determine the skeletal key point information, obtain the personnel state sample set, and build a posture recognition model based on the personnel state sample set. The posture recognition model is a comprehensive model that fuses skeletal relationships and global posture features. The model validation module is configured to validate the posture recognition model based on a set of personnel state samples. If the validation fails, the learning rate of the posture recognition model is adjusted based on cosine annealing, and the target posture recognition model is determined based on the adjustment result of the learning rate. The posture warning module is configured to determine the target posture based on the target posture recognition model and skeletal key point information, and to compare the target posture with the corresponding standard target posture. Based on the result of the comparison, it determines whether to issue a warning for the target posture.

[0043] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program goods. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program goods embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0044] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program goods according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0045] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0046] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0047] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for monitoring and early warning of power operation posture based on AI recognition, characterized in that, include: Acquire operational posture images, process the operational posture images to determine target operational posture images, perform posture analysis on the target operational posture images, and generate posture imprints based on the results of the posture analysis. The posture imprints include posture matching imprints, posture suspected imprints, and posture deviation imprints. When the suspected posture imprint is generated, the skeletal key points of the target work posture image are extracted based on the posture estimation algorithm, and the skeletal key point information is determined. A personnel state sample set is obtained, and a posture recognition model is constructed based on the personnel state sample set. The posture recognition model is a comprehensive model that fuses skeletal relationships and global posture features. The posture recognition model is validated based on the personnel state sample set. If the validation fails, the learning rate of the posture recognition model is adjusted based on cosine annealing, and the target posture recognition model is determined based on the adjustment result of the learning rate. The target pose is determined based on the target pose recognition model and skeletal key point information, and the target pose is compared with the corresponding standard target pose. Based on the result of the comparison, it is determined whether to issue a warning for the target pose.

2. The method for monitoring and early warning of power operation posture based on AI recognition according to claim 1, characterized in that, Processing the work posture image to determine the target work posture image includes: The working posture image is divided into several regions to be detected. High-frequency noise points of the working posture image are obtained based on gray-level difference processing and neighborhood pixel correlation. The regions to be detected containing the high-frequency noise points are recorded as environmental interference regions. Based on adaptive weighted median filtering, appropriate weights are set for the environmental interference area and smoothed using a filtering window to determine the smoothed signal-to-noise ratio and edge sharpness. When the signal-to-noise ratio is less than 25dB, the window size of the filtering window is increased; When the edge sharpness is less than 40, the weight of the edge of the environmental interference area is reduced. The target operation posture image is determined based on the adjusted weights and filter window.

3. The method for monitoring and early warning of power operation posture based on AI recognition according to claim 2, characterized in that, When performing attitude analysis on the target operation posture image and generating attitude imprints based on the results of the attitude analysis, the attitude imprints include attitude matching imprints, attitude suspected imprints, and attitude deviation imprints, including: Determine the standard target operation posture image corresponding to the target operation posture image; Extract all posture pixels from the target operation posture image, and extract the standard posture pixels corresponding to each posture pixel in the standard target operation posture image. Determine the posture pixel values ​​corresponding to all posture pixels and the standard posture pixel values ​​corresponding to all standard posture pixels. If all pose pixel values ​​are equal to the standard pose pixel values, then the pose matching imprint is generated for the target working pose image; If one or more pose pixel values ​​are not equal to the standard pose pixel values, then generate the pose imprint for the target operation pose image; If none of the attitude pixel values ​​are equal to the standard attitude pixel values, then the attitude deviation imprint is generated for the target working attitude image.

4. The method for monitoring and early warning of power operation posture based on AI recognition according to claim 3, characterized in that, When obtaining the personnel status sample set, the following are included: Select at least 50 power workers of different genders and heights, and deploy inertial sensors on each power worker. The deployment positions are the inside of the safety helmet, torso, left / right wrist, and left / right ankle. The sampling frequency of the inertial sensors is set to 200Hz. When a tool is dropped, the sampling frequency is set to 500Hz. Continuous sampling was conducted during high-altitude and ground operations to determine at least two thousand sets of posture samples and construct the personnel state sample set.

5. The method for monitoring and early warning of power operation posture based on AI recognition according to claim 4, characterized in that, When constructing a pose recognition model based on the aforementioned personnel state sample set, the following steps are included: The personnel status sample set is divided into a model building set and a model validation set; The pose recognition model includes a skeletal relationship feature extraction subnetwork and a global pose feature extraction subnetwork; The skeletal relationship feature extraction subnetwork adopts a graph convolutional neural network structure, based on 17 human skeletal key points as graph nodes, and constructs an adjacency matrix with skeletal connection relationships. The joint angles in the model are input into 3 layers of graph convolutional layers for local relationship feature extraction, and output a 64-dimensional skeletal relationship feature vector. The global attitude feature extraction subnetwork adopts a CNN-LSTM hybrid structure. Based on the sampling points within the continuous time window of the model set, it uses the deployment position of the inertial sensor as the spatial axis to determine the motion features of the time axis and the spatial axis. The motion features include the original components of three-dimensional acceleration, the original components of three-dimensional angular velocity, the first derivative of acceleration amplitude, and the direction angle of angular velocity. A spatiotemporal matrix is ​​constructed based on the motion features. The convolutional layer of the global pose feature extraction subnetwork extracts the local spatiotemporal features of the spatiotemporal matrix and performs dimensionality reduction based on the pooling layer. The fully connected layer outputs a 64-dimensional global pose feature vector based on the dimensionality reduction result.

6. The method for monitoring and early warning of power operation posture based on AI recognition according to claim 5, characterized in that, When constructing a pose recognition model based on the aforementioned personnel state sample set, the following is also included: The 64-dimensional skeletal relationship feature vector and the 64-dimensional global posture feature vector are fused. The fusion process is based on an attention mechanism algorithm to determine the skeletal feature weight coefficients of the 64-dimensional skeletal relationship feature vector and the posture feature weight coefficients of the 64-dimensional global posture feature vector. The skeletal feature weight coefficients are determined based on the joint motion amplitude, and the posture feature weight coefficients are determined based on the work scenario.

7. The method for monitoring and early warning of power operation posture based on AI recognition according to claim 6, characterized in that, The pose recognition model also includes a feature fusion layer and a classifier. The feature fusion layer performs weighted summation on the 64-dimensional skeletal relationship feature vector and the 64-dimensional global pose feature vector based on the skeletal feature weight coefficient and the pose feature weight coefficient, respectively, to determine a 128-dimensional fused feature vector. The classifier contains a 3-layer fully connected network and outputs the job pose based on the Softmax activation function and the 128-dimensional fused feature vector.

8. The method for monitoring and early warning of power operation posture based on AI recognition according to claim 7, characterized in that, When validating the posture recognition model based on the personnel state sample set, and adjusting the learning rate of the posture recognition model based on cosine annealing if the validation fails, and determining the target posture recognition model based on the adjustment result of the learning rate, the process includes: The model validation set is divided into a validation set and a test set, and the performance metrics of the pose recognition model are evaluated based on the validation set. If the performance metric does not reach the performance metric threshold, the learning rate of the pose recognition model is adjusted based on cosine annealing, and the performance metric of the pose recognition model is re-determined based on the validation set until the performance metric reaches the performance metric threshold. If the performance metric reaches the performance metric threshold, the performance metric of the pose recognition model is evaluated again based on the test set. If the performance index reaches the performance index threshold in the re-evaluation, then the posture recognition model is determined as the target posture recognition model. If the performance metrics fail to meet the performance threshold during the re-evaluation, the pose recognition model will be rebuilt.

9. The method for monitoring and early warning of power operation posture based on AI recognition according to claim 8, characterized in that, When determining the target pose based on the target pose recognition model and skeletal keypoint information, and comparing the target pose with the corresponding standard target pose, and determining whether to issue a warning for the target pose based on the result of the overlap comparison, the process includes: The skeletal key point information is substituted into the target pose recognition model to determine the target pose and obtain the standard target pose corresponding to the target pose. If the overlap area between the target pose and the standard target pose is greater than or equal to the overlap area threshold, then it is determined that no warning will be issued for the target pose. When the overlap area between the target pose and the standard target pose is less than the overlap area threshold, a warning is issued for the target pose.

10. A power operation posture monitoring and early warning system based on AI recognition, used to apply the power operation posture monitoring and early warning method based on AI recognition as described in any one of claims 1-9, characterized in that, include: The acquisition and analysis module is configured to acquire work posture images, process the work posture images, determine target work posture images, perform posture analysis on the target work posture images, and generate posture imprints based on the results of the posture analysis. The posture imprints include posture matching imprints, posture suspected imprints, and posture deviation imprints. The model building module is configured to extract the skeletal key points of the target work posture image based on the posture estimation algorithm when the suspected posture imprint is generated, determine the skeletal key point information, obtain the personnel state sample set, and construct a posture recognition model based on the personnel state sample set. The posture recognition model is a comprehensive model that fuses skeletal relationships and global posture features. The model validation module is configured to validate the posture recognition model based on the personnel state sample set. If the validation fails, the learning rate of the posture recognition model is adjusted based on cosine annealing, and the target posture recognition model is determined based on the adjustment result of the learning rate. The posture warning module is configured to determine the target posture based on the target posture recognition model and skeletal key point information, and to perform an overlap comparison between the target posture and the corresponding standard target posture, and to determine whether to issue a warning for the target posture based on the result of the overlap comparison.

Images