An intelligent monitoring method and system for electric power construction

Abstract

This application discloses an intelligent supervision method and system for power construction, relating to the field of intelligent supervision of power construction. The method includes: acquiring image sequences of the construction area, extracting these sequences from a video stream using intelligent frame extraction technology; constructing a power construction risk detection model, which includes a feature extraction backbone network, a temporal awareness alignment module, a multi-scale spatial modulation fusion module, a feature pyramid network, and a detection head; inputting the image sequences into the power construction risk detection model to obtain safety hazard identification results; and generating a dynamic risk level score based on the safety hazard identification results through three-factor linear weighting, triggering tiered early warnings, and pushing early warning information and rectification instructions to on-site supervisors and construction personnel. By constructing a dedicated deep learning model integrating temporal awareness alignment and knowledge-guided spatial modulation fusion mechanisms, the method fundamentally improves the intelligent identification and assessment capabilities for complex dynamic risks.

Description

Technical Field

[0001] This application relates to the field of intelligent monitoring technology for power construction, and in particular to an intelligent monitoring method and system for power construction. Background Technology

[0002] With the continuous expansion of power engineering construction scale and the increasing complexity of construction environments, traditional safety supervision methods mainly rely on manual inspections and fixed camera monitoring, which suffer from problems such as insufficient real-time performance, high missed detection rates, and delayed response. Especially in dynamic construction scenarios, personnel behavior, equipment status, and environmental risks often intertwine. Simply relying on static image recognition or rule engines is difficult to effectively capture continuous risk behaviors in a temporal sequence. Features are easily drifted due to camera shake, target movement, or occlusion, affecting detection stability and tracking continuity. Existing visual supervision systems mostly focus on general target detection and lack dedicated risk feature extraction and temporal modeling mechanisms for power construction scenarios. They are unable to accurately locate risk targets under complex background interference, and cannot perform dynamic fusion assessment and graded early warning of multiple risk factors.

[0003] Furthermore, existing safety risk assessment methods often rely on simple accumulation or threshold judgment based on single-frame recognition results, neglecting the coupling relationship between the spatial clustering effect, temporal persistence, and category heterogeneity of risk targets. This results in overly static and one-sided risk assessment results, failing to accurately reflect the real-time risk situation at the construction site. Therefore, how to construct an intelligent monitoring method that integrates temporal perception, spatial modulation, and multi-dimensional risk assessment to achieve proactive perception, accurate identification, and dynamic early warning of all aspects of power construction has become an urgent technical challenge to improve construction safety management. Summary of the Invention

[0004] To address the aforementioned challenges, this application provides an intelligent monitoring method and system for power construction. By constructing a dedicated deep learning model that integrates a temporal awareness alignment and knowledge-guided spatial modulation fusion mechanism, it fundamentally enhances the intelligent identification and assessment capabilities for complex dynamic risks.

[0005] To achieve the above objectives, this application provides an intelligent monitoring method for power construction, comprising the following steps: S1: Real-time acquisition of image sequences of the construction area. The image sequences are extracted from the video stream using intelligent frame extraction technology. Each frame of the image is accompanied by a timestamp and camera location information. S2: Construct a power construction risk detection model, which includes a feature extraction backbone network, a time-aware alignment module, a multi-scale spatial modulation fusion module, a feature pyramid network, and a detection head; S3: Input the image sequence into the power construction risk detection model to obtain the safety hazard identification result of power construction. The identification result includes the location bounding box of the risk target, the risk category label and the confidence score. S4: Based on the safety hazard identification results, a dynamic risk level score is generated through three-factor linear weighting, and a graded early warning is triggered to push early warning information and rectification instructions to on-site supervisors and construction workers.

[0006] Preferably, the feature extraction backbone network is a cascaded multi-stage structure, wherein the shallow stage outputs high-resolution, detailed primary feature maps, and the deep stage outputs low-resolution, semantically rich high-level feature maps.

[0007] Preferably, the temporal awareness alignment module adopts an inter-frame self-attention mechanism, receives a multi-scale feature map sequence from a single-modal feature extraction backbone network, generates motion offset by calculating the spatial correlation between the current frame and adjacent reference frames, performs alignment correction on the current frame features based on differentiable spatial transformation, and fuses the original features through residual connections to output enhanced features with temporal consistency for association analysis of continuous risk behaviors.

[0008] Preferably, the inter-frame self-attention mechanism specifically includes the following steps: A five-frame continuous feature map sequence is constructed as a time-series-aware window, and the feature map sequence is extracted by a single-modal feature extraction backbone network; Using the current frame feature map as the query and the feature maps of its neighboring frames as the key and value, the spatial correlation between frames is calculated through three-dimensional convolution operation, and spatial motion offset is generated accordingly. Using the motion offset, spatial transformation correction is performed on the current frame feature map through a differentiable spatial sampling operation; The corrected feature map and the original current frame feature map are fused by residual connection using a dynamically predicted fusion weight, resulting in a feature map with enhanced temporal consistency.

[0009] Preferably, the multi-scale spatial modulation fusion module receives enhanced features from the temporal awareness alignment module and performs adaptive weighted fusion of multi-scale features at different levels through a channel-space dual attention mechanism. The channel attention generates channel weights through global average pooling and a multilayer perceptron, while the spatial attention generates spatial position weights through convolution operations. The two weights are multiplied element-wise and applied to the input features to dynamically suppress background region responses and enhance the expression of construction risk target features. The fused multi-scale feature pyramid is then output to the detection head.

[0010] Preferably, the channel-space dual attention mechanism specifically includes: For the input P3, P4 and P5 three-layer enhanced feature maps, each feature map is first compressed into a 1×1×C channel description vector through global average pooling, and then a channel weight vector is generated through two fully connected layers, where C is the number of channels, and the number of neurons in the two fully connected layers are 64 and 256, respectively. The spatial attention branch uses a 7×7 convolution kernel with a stride of 1 and padding of 3 to generate a spatial attention map for each feature map layer. The convolution output is normalized by the Sigmoid function, so that the model focuses on the high-risk construction area and suppresses static background interference such as scaffolding and fences. Channel weights and spatial weights are multiplied element-wise through a broadcast mechanism to generate dual modulation weights of the same dimension as the input feature map. The modulated multi-scale features are then fused from top to bottom in the feature pyramid network through upsampling and lateral connections, and finally output as a three-layer fused feature pyramid (P3-P5) to the detection head.

[0011] Preferably, the detection head is a multi-task parallel detection head targeting power construction risks, comprising: The equipment status anomaly detection branch is used to identify unsafe operating conditions of construction equipment; The personnel behavior compliance detection branch is used to identify unsafe work behaviors of construction workers; The general construction hazard detection branch is used to identify environmental and material storage risks; Each detection branch generates detection results with location, category, and confidence level in parallel based on the multi-scale feature pyramid output by the multi-scale spatial modulation fusion module, which are then summarized for subsequent risk assessment steps.

[0012] Preferably, the three-factor linear weighting specifically includes: a category hazard coefficient factor, which pre-sets differentiated weights based on the severity of the accident consequences corresponding to the risk type, with 4, 3, 2, and 1 level coefficients for fatal risk, severe risk, general risk, and low risk, respectively; a target quantity factor, which assigns a stepwise value based on the number of risk targets detected in a single frame, with a higher coefficient for a larger number; and a duration factor, which uses a cross-frame tracking algorithm to statistically analyze the continuous detection duration of risk targets and classifies them into three levels: instantaneous, short-term, and continuous. The dynamic risk level score is calculated by linearly weighting and summing the category coefficient, quantity factor, and duration factor, and is mapped to a four-level warning level according to a preset scoring threshold, triggering a differentiated response mechanism.

[0013] An intelligent monitoring system for power construction, employing an intelligent monitoring method for power construction, includes: The data acquisition module is used to acquire image sequences of the construction area in real time. The image sequences are extracted from the video stream using intelligent frame extraction technology, and each frame image is accompanied by a timestamp and camera location information. The model processing module, which embeds a power construction risk detection model, is used to receive the image sequence and output the safety hazard identification results. The identification results include the location bounding box of the risk target, the risk category label, and the confidence score. The risk assessment and early warning decision-making module is used to generate a dynamic risk level score based on the safety hazard identification results through three-factor linear weighting, and to trigger graded early warning instructions; The early warning execution and feedback module is used to push early warning information and rectification instructions to designated on-site monitoring terminals and construction personnel terminals according to the early warning instructions.

[0014] Therefore, this application adopts the above-mentioned intelligent monitoring method and system for power construction. First, intelligent frame extraction is performed on the video stream to obtain image sequences, and a time-series-aware alignment module is used to correct feature drift caused by equipment vibration and target movement, so as to achieve stable tracking of the violation process. Then, a multi-scale spatial modulation fusion module is used to focus on high-risk areas and suppress background interference to accurately extract multi-scale risk target features. Finally, the system uses a product-based fusion architecture to conduct dynamic risk assessment by comprehensively considering the category, spatial density and duration of risk targets, so as to achieve hierarchical early warning, thereby completing the leap from passive monitoring to active perception, and from static identification to dynamic assessment. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating an intelligent monitoring method for power construction according to this application. Figure 2 This is a framework diagram of an intelligent monitoring system for power construction according to this application. Detailed Implementation

[0016] The following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0017] Unless otherwise defined, the technical or scientific terms used in this application shall have the ordinary meaning as understood by a person of ordinary skill in the art to which this application pertains.

[0018] The terms "comprising" or "including," as used in this application, mean that the element preceding the term encompasses the element listed after the term, and do not exclude the possibility of encompassing other elements as well. The terms "inner," "outer," "upper," and "lower," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. When the absolute position of the described object changes, the relative positional relationship may also change accordingly. In this application, unless otherwise expressly specified and limited, the term "attached," etc., should be interpreted broadly. For example, it can refer to a fixed connection, a detachable connection, or an integral part; it can refer to a direct connection or an indirect connection through an intermediate medium; it can refer to the internal communication of two elements or the interaction relationship between two elements. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances.

[0019] Example 1: A smart monitoring method for power construction, such as Figure 1 As shown, it includes the following steps: S1: Real-time acquisition of image sequences of the construction area. The image sequences are extracted from the video stream using intelligent frame extraction technology. Each frame of the image is accompanied by a timestamp and camera location information. Specifically, high-definition network cameras are deployed in key areas of power construction sites to capture construction video streams in real time via wired or wireless transmission. Intelligent frame extraction technology is used to dynamically extract image sequences from the continuous video stream. This frame extraction strategy can adaptively adjust the frame rate according to the complexity of the scene, increasing the sampling frequency when personnel and equipment are dense or when potential moving targets are detected, and reducing the frame rate in static scenes to save storage and computing resources. At the same time, each extracted image automatically embeds unique identification information, including a timestamp accurate to milliseconds (for time sequence alignment and behavior persistence analysis) and the spatial location coordinates and device number of the camera (for risk localization and multi-view correlation), thereby providing a high-quality, timely, and spatiotemporally correlated data foundation for subsequent deep learning risk detection models.

[0020] S2: Construct a power construction risk detection model, which includes a feature extraction backbone network, a time-aware alignment module, a multi-scale spatial modulation fusion module, a feature pyramid network, and a detection head; The feature extraction backbone network is a cascaded multi-stage structure, in which the shallow stage outputs high-resolution, detailed primary feature maps, and the deep stage outputs low-resolution, semantically rich high-level feature maps.

[0021] Typically based on an improved ResNet-50 or CSPDarkNet architecture, this network is divided into five stages (Stage 1 to Stage 5). The shallow stages (Stages 1-2) maintain a high spatial resolution (1 / 4 to 1 / 8 scale of the input image), extracting primary feature maps rich in texture, edge, and color details by stacking residual blocks. These details are crucial for recognizing small protective equipment such as helmets and seat belts. The deep stages (Stages 3-5) reduce the spatial dimension to 1 / 16 to 1 / 32 of the input through progressive downsampling (pooling or strided convolution), while expanding the channel dimension (from 64 channels to 512 / 1024 channels). Utilizing the expanded receptive field of deep convolutions, they extract high-level feature maps rich in semantic information, enabling accurate understanding of the semantic context of complex scenes such as people standing under crane arms or crossing safety fences. The outputs of each stage are labeled as C3, C4, and C5 feature maps, corresponding to different spatial scales, forming the basis for multi-scale feature representation.

[0022] The temporal-aware alignment module adopts an inter-frame self-attention mechanism, receives multi-scale feature map sequences from a single-modal feature extraction backbone network, generates motion offsets by calculating the spatial correlation between the current frame and adjacent reference frames, performs alignment correction on the current frame features based on differentiable spatial transformation, and fuses the original features through residual connections to output enhanced features with temporal consistency for the correlation analysis of continuous risk behaviors.

[0023] The inter-frame self-attention mechanism specifically includes the following steps: A five-frame continuous feature map sequence is constructed as a time-series-aware window, and the feature map sequence is extracted by a single-modal feature extraction backbone network; Using the current frame feature map as the query and the feature maps of its neighboring frames as the key and value, the spatial correlation between frames is calculated through three-dimensional convolution operation, and spatial motion offset is generated accordingly. Using the motion offset, spatial transformation correction is performed on the current frame feature map through a differentiable spatial sampling operation; The corrected feature map and the original current frame feature map are fused by residual connection using a dynamically predicted fusion weight, resulting in a feature map with enhanced temporal consistency.

[0024] Specifically, the temporal awareness alignment module employs an inter-frame self-attention mechanism. This module receives a multi-scale feature map sequence (i.e., the {C3, C4, C5} sequence corresponding to five consecutive frames) from the feature extraction backbone network, aiming to solve the problem of feature space position drift caused by camera mechanical vibration, windy environment shaking, or rapid movement of dynamic targets. Specifically, for each feature map layer (taking C5 as an example), the module first constructs a five-frame consecutive feature map sequence as a temporal awareness window (denoted as Ft-2, Ft-1, Ft, Ft+1, Ft+2). This sequence is extracted by the single-modal feature extraction backbone network through forward propagation of the five consecutive frames. Subsequently, using the current frame feature map Ft as the query and the feature maps {Ft-2, Ft-1, Ft+1, Ft+2} of its four adjacent frames as keys and values, the module calculates the inter-frame spatial correlation through three-dimensional convolution operations (kernel size 3×3×3, joint convolution along the spatiotemporal dimension), capturing pixel-level motion between adjacent frames. The pattern is used to generate a spatial motion offset (denoted as Δp, with dimensions H×W×2, representing the offset of each spatial position in the x and y directions). Using this motion offset, the current frame feature map Ft is spatially transformed and corrected through a differentiable spatial sampling operation (usually implemented using bilinear interpolation), aligning its feature spatial position with the reference frame. Finally, the corrected feature map Ft' and the original current frame feature map Ft are fused using a residual connection with dynamically predicted fusion weights (adaptively generated by a lightweight convolutional network based on alignment quality, denoted as α and β, satisfying α+β=1), outputting a temporally consistent enhanced feature Fout=α·Ft'+β·Ft. This enhanced feature effectively suppresses feature jitter caused by motion blur or positional drift, enabling subsequent detection heads to stably track continuous risky behaviors (such as the complete process of a person climbing from the ground to the crane operator's cab).

[0025] The multi-scale spatial modulation fusion module receives enhanced features from the temporal awareness alignment module and performs adaptive weighted fusion of multi-scale features at different levels through a channel-space dual attention mechanism. Channel attention generates channel weights through global average pooling and multilayer perceptron, while spatial attention generates spatial position weights through convolution operations. The two weights are multiplied element-wise and applied to the input features to dynamically suppress background region responses and enhance the expression of construction risk target features. The fused multi-scale feature pyramid is then output to the detection head.

[0026] The channel-space dual attention mechanism specifically includes: For the input P3, P4 and P5 three-layer enhanced feature maps, each feature map is first compressed into a 1×1×C channel description vector through global average pooling, and then a channel weight vector is generated through two fully connected layers, where C is the number of channels, and the number of neurons in the two fully connected layers are 64 and 256, respectively. The spatial attention branch uses a 7×7 convolution kernel with a stride of 1 and padding of 3 to generate a spatial attention map for each feature map layer. The convolution output is normalized by the Sigmoid function, so that the model focuses on the high-risk construction area and suppresses static background interference such as scaffolding and fences. Channel weights and spatial weights are multiplied element-wise through a broadcast mechanism to generate dual modulation weights of the same dimension as the input feature map. The modulated multi-scale features are then fused from top to bottom in the feature pyramid network through upsampling and lateral connections, and finally output as a three-layer fused feature pyramid (P3-P5) to the detection head.

[0027] The multi-scale spatial modulation fusion module receives temporal consistency enhancement features (denoted as P3_in, P4_in, and P5_in, corresponding to feature maps with resolutions of 80×80, 40×40, and 20×20 respectively, calculated based on a 640×640 input) from the temporal awareness alignment module. It then adaptively weights and fuses these multi-scale features at different levels using a channel-spatial dual attention mechanism. This aims to address the negative impact of complex background interference (such as dense scaffolding, green protective netting, and temporary fencing) on ​​risk target detection at power construction sites. The specific implementation includes two parallel branches: Channel Attention Branch: For the input P3, P4, and P5 enhanced feature maps, each feature map first undergoes global average pooling to compress its spatial dimension, generating a 1×1×C channel description vector (where C is the number of channels, typically 256). This vector encodes the global statistical characteristics of each channel feature map. Subsequently, this vector is passed through a bottleneck structure consisting of two fully connected layers to generate channel weight vectors. The first layer compresses the dimension from 256 to 64 (a dimensionality reduction ratio of 16) and introduces non-linearity using ReLU activation. The second layer restores the dimension from 64 to 256 and normalizes the weights to the (0,1) interval using the Sigmoid function. This mechanism enables the model to adaptively enhance channels critical to the risk detection task (such as edge detection channels and texture analysis channels) while suppressing the responses of irrelevant channels (such as sky background and uniform color regions).

[0028] Spatial Attention Branch: A large-size convolutional kernel (7×7 kernel, stride 1, padding 3 to maintain size) is used to encode the spatial dimension of each feature map layer. This large kernel captures a wider range of contextual relationships, making it suitable for identifying spatial layout risks between personnel, equipment, and the environment. The convolutional output is normalized pixel-by-pixel using the Sigmoid function to generate a spatial attention map of the same size as the input feature map. High-response regions correspond to areas with high construction risk (such as near openings, around high-voltage equipment), while low-response regions correspond to static backgrounds (scaffolding poles, fencing materials). This design allows the model to focus on the true risk sources, avoiding false alarms caused by cluttered backgrounds.

[0029] Channel weights (1×1×C) and spatial weights (H×W×1) are multiplied element-wise through a broadcasting mechanism to generate dual modulation weights of the same dimension (H×W×C) as the input feature map. The modulated multi-scale features are then fused top-down in the Feature Pyramid Network (FPN) through upsampling (nearest neighbor interpolation) and lateral connections (1×1 convolution adjusting channels and then element-wise addition): first, the semantically rich P5 layer is upsampled and fused with P4 to generate enhanced P4; then, enhanced P4 is upsampled and fused with P3 to generate enhanced P3; finally, the three-layer fused feature pyramid (P3-P5) is output to the detection head. This multi-scale fusion enables the detection head to simultaneously detect small targets (such as a head without a helmet at a distance, handled by P3) and large targets (such as a large piece of equipment tilting nearby, handled by P5).

[0030] The detection head is a multi-task parallel detection head designed for power construction risk targets. Unlike the single-head structure of general target detection, this detection head adopts a three-branch architecture to adapt to the refined management requirements of power construction safety regulations, including: The equipment status anomaly detection branch is used to identify unsafe operating conditions of construction equipment, including crane outriggers not fully extended, people within the swing radius of excavators, open electrical distribution box doors, and crane hooks without anti-detachment devices. This branch sets anchors in layers P4 and P5 of the feature pyramid (suitable for detecting medium to large-scale equipment) and uses classification-regression joint loss training. The personnel behavior compliance detection branch is used to identify unsafe work behaviors of construction workers, including not wearing a safety helmet, not wearing a safety belt, illegally crossing fences, climbing or sitting in unsafe positions, and using mobile phones while operating. This branch conducts intensive sampling at layers P3 and P4 (suitable for detecting human-scale targets) and combines posture estimation to assist in the visibility judgment of key points. The general construction hazard detection branch is used to identify environmental and material storage risks, including unprotected edges of foundation pits, incomplete scaffolding planks, excessively high material stacking, water-damaged cables, and blocked fire lanes. This branch utilizes the full-level features of P3-P5 and adopts deformable convolution to adapt to irregularly shaped hazards.

[0031] Each detection branch generates detection results in parallel, incorporating location, category, and confidence score, based on the multi-scale feature pyramid output by the multi-scale spatial modulation fusion module. These results are then aggregated for subsequent risk assessment steps. The outputs of all branches are then aggregated for multi-dimensional fusion calculations in the subsequent risk assessment steps. The advantages of the three-branch structure are: it allows for the configuration of differentiated detection thresholds and NMS parameters for different risk types (e.g., higher recall for personnel behavior, higher accuracy for equipment status), and facilitates rapid model iteration for specific risk types (e.g., adding a new equipment type does not require retraining the personnel branch).

[0032] S3: Input the image sequence into the power construction risk detection model to obtain the safety hazard identification result of power construction. The identification result includes the location bounding box of the risk target, the risk category label and the confidence score. S4: Based on the safety hazard identification results, a dynamic risk level score is generated through three-factor linear weighting, and a graded early warning is triggered to push early warning information and rectification instructions to on-site supervisors and construction workers.

[0033] The three-factor linear weighting specifically includes: a category hazard coefficient factor, which pre-sets differentiated weights based on the severity of the accident consequences corresponding to the risk type, with 4, 3, 2, and 1 level coefficients for fatal risk, severe risk, general risk, and low risk, respectively; a target quantity factor, which assigns a stepwise value based on the number of risk targets detected in a single frame, with a higher coefficient for a larger number; and a duration factor, which uses a cross-frame tracking algorithm to statistically analyze the continuous detection duration of risk targets, classifying them into three levels: instantaneous, short-term, and continuous. Detection of the same target for more than 3 consecutive frames is considered continuous. When the duration is less than 10 seconds (instantaneous), the duration is set to 1; when the duration is 10 seconds to 5 minutes, the duration is set to 2; and when the duration is greater than 5 minutes, the duration is set to 3. The dynamic risk level score is calculated by linearly weighting and summing the category coefficient, quantity factor, and duration factor, and is mapped to a four-level warning level based on a preset scoring threshold, triggering a differentiated response mechanism.

[0034] Specifically, the preset category risk coefficients can be shown in Table 1 below: Table 1

[0035] The dynamic risk level score is specifically expressed as follows: ; in, For category risk factor, For the number of risk targets, For duration, , , , where is the weighting coefficient, and the values ​​are 0.5, 0.3, and 0.2.

[0036] The rating mapping and early warning triggering are shown in Table 2 below. Table 2

[0037] Example 2: An intelligent monitoring system for power construction, such as Figure 2 As shown, an intelligent monitoring method for power construction is adopted, including: The data acquisition module is used to acquire image sequences of the construction area in real time. The image sequences are extracted from the video stream using intelligent frame extraction technology, and each frame image is accompanied by a timestamp and camera location information. The model processing module, which embeds a power construction risk detection model, is used to receive the image sequence and output the safety hazard identification results. The identification results include the location bounding box of the risk target, the risk category label, and the confidence score. The risk assessment and early warning decision-making module is used to generate a dynamic risk level score based on the safety hazard identification results through three-factor linear weighting, and to trigger graded early warning instructions; The early warning execution and feedback module is used to push early warning information and rectification instructions to designated on-site monitoring terminals and construction personnel terminals according to the early warning instructions.

[0038] Therefore, this application adopts the aforementioned intelligent monitoring method and system for power construction. It achieves efficient acquisition and accurate positioning of video streams in the construction area through intelligent frame extraction and spatiotemporal information embedding technology; solves the problems of feature drift and tracking loss in dynamic scenes through a time-series perception alignment module; suppresses complex background interference and enhances the expression of risk target features through a multi-scale spatial modulation fusion module; achieves refined identification of equipment status, personnel behavior, and construction hazards through a multi-task parallel detection head; and finally, generates a dynamic risk level score and triggers graded early warnings by fusing category hazard coefficients, target cluster density, and risk duration through a three-factor linear weighted dynamic risk assessment model. This technical solution forms a complete closed loop of acquisition-detection-assessment-early warning, realizing a technological leap from passive monitoring to active perception, from static identification to dynamic assessment, and from single alarms to graded responses. It significantly improves the accuracy of safety risk identification, the stability of time-series tracking, and the timeliness of early warning response at power construction sites, effectively solving the technical problems of insufficient real-time performance, high missed detection rate, delayed response, and one-sided risk assessment in traditional safety supervision methods. It provides reliable technical support for intelligent safety management of the entire process and all elements of power engineering construction.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and not to limit them. Although this application has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of this application, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of this application.

Claims

1. A smart monitoring method for power construction, characterized in that, Includes the following steps: S1: Real-time acquisition of image sequences of the construction area. The image sequences are extracted from the video stream using intelligent frame extraction technology. Each frame of the image is accompanied by a timestamp and camera location information. S2: Construct a power construction risk detection model, which includes a feature extraction backbone network, a time-aware alignment module, a multi-scale spatial modulation fusion module, a feature pyramid network, and a detection head; S3: Input the image sequence into the power construction risk detection model to obtain the safety hazard identification result of power construction. The identification result includes the location bounding box of the risk target, the risk category label and the confidence score. S4: Based on the safety hazard identification results, a dynamic risk level score is generated through three-factor linear weighting, and a graded early warning is triggered to push early warning information and rectification instructions to on-site supervisors and construction workers.

2. The intelligent monitoring method for power construction according to claim 1, characterized in that, The feature extraction backbone network is a cascaded multi-stage structure, in which the shallow stage outputs high-resolution, detailed primary feature maps, and the deep stage outputs low-resolution, semantically rich high-level feature maps.

3. The intelligent monitoring method for power construction according to claim 2, characterized in that, The temporal awareness alignment module adopts an inter-frame self-attention mechanism, receives a multi-scale feature map sequence from a single-modal feature extraction backbone network, generates motion offset by calculating the spatial correlation between the current frame and adjacent reference frames, performs alignment correction on the current frame features based on differentiable spatial transformation, and fuses the original features through residual connections to output enhanced features with temporal consistency for the correlation analysis of continuous risk behaviors.

4. The intelligent monitoring method for power construction according to claim 3, characterized in that, The inter-frame self-attention mechanism specifically includes the following steps: A five-frame continuous feature map sequence is constructed as a time-series-aware window, and the feature map sequence is extracted by a single-modal feature extraction backbone network; Using the current frame feature map as the query and the feature maps of its neighboring frames as the key and value, the spatial correlation between frames is calculated through three-dimensional convolution operation, and spatial motion offset is generated accordingly. Using the motion offset, spatial transformation correction is performed on the current frame feature map through a differentiable spatial sampling operation; The corrected feature map and the original current frame feature map are fused by residual connection using a dynamically predicted fusion weight, resulting in a feature map with enhanced temporal consistency.

5. The intelligent monitoring method for power construction according to claim 4, characterized in that, The multi-scale spatial modulation fusion module receives enhanced features from the temporal awareness alignment module and performs adaptive weighted fusion of multi-scale features at different levels through a channel-space dual attention mechanism. Channel attention generates channel weights through global average pooling and multilayer perceptron, while spatial attention generates spatial position weights through convolution operations. The two weights are multiplied element-wise and applied to the input features to dynamically suppress background region responses and enhance the expression of construction risk target features. The fused multi-scale feature pyramid is then output to the detection head.

6. The intelligent monitoring method for power construction according to claim 5, characterized in that, The channel-space dual attention mechanism specifically includes: For the input P3, P4 and P5 three-layer enhanced feature maps, each feature map is first compressed into a 1×1×C channel description vector through global average pooling, and then a channel weight vector is generated through two fully connected layers, where C is the number of channels, and the number of neurons in the two fully connected layers are 64 and 256, respectively. The spatial attention branch uses a 7×7 convolution kernel with a stride of 1 and padding of 3 to generate a spatial attention map for each feature map layer. The convolution output is normalized by the Sigmoid function, so that the model focuses on the high-risk construction area and suppresses static background interference such as scaffolding and fences. Channel weights and spatial weights are multiplied element-wise through a broadcast mechanism to generate dual modulation weights of the same dimension as the input feature map. The modulated multi-scale features are then fused from top to bottom in the feature pyramid network through upsampling and lateral connections, and finally output as a three-layer fused feature pyramid (P3-P5) to the detection head.

7. The intelligent monitoring method for power construction according to claim 6, characterized in that, The detection head is a multi-task parallel detection head designed for power construction risk targets, including: The equipment status anomaly detection branch is used to identify unsafe operating conditions of construction equipment; The personnel behavior compliance detection branch is used to identify unsafe work behaviors of construction workers; The general construction hazard detection branch is used to identify environmental and material storage risks; Each detection branch generates detection results with location, category, and confidence level in parallel based on the multi-scale feature pyramid output by the multi-scale spatial modulation fusion module, which are then summarized for subsequent risk assessment steps.

8. The intelligent monitoring method for power construction according to claim 7, characterized in that, The three-factor linear weighting specifically includes: a category hazard coefficient factor, which pre-sets differentiated weights based on the severity of the accident consequences corresponding to the risk type, with 4, 3, 2, and 1 level coefficients for fatal risk, severe risk, general risk, and low risk, respectively; a target quantity factor, which assigns a stepwise value based on the number of risk targets detected in a single frame, with a higher coefficient for a larger number; and a duration factor, which uses a cross-frame tracking algorithm to statistically analyze the continuous detection duration of risk targets and classifies them into three levels: instantaneous, short-term, and continuous. The dynamic risk level score is calculated by linearly weighting and summing the category coefficient, quantity factor, and duration factor, and is mapped to a four-level warning level according to a preset scoring threshold, triggering a differentiated response mechanism.

9. An intelligent monitoring system for power construction, employing an intelligent monitoring method for power construction as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to acquire image sequences of the construction area in real time. The image sequences are extracted from the video stream using intelligent frame extraction technology, and each frame image is accompanied by a timestamp and camera location information. The model processing module, which embeds a power construction risk detection model, is used to receive the image sequence and output the safety hazard identification results. The identification results include the location bounding box of the risk target, the risk category label, and the confidence score. The risk assessment and early warning decision module is used to generate a dynamic risk level score based on the safety hazard identification results through three-factor linear weighting, and trigger graded early warning instructions. The early warning execution and feedback module is used to push early warning information and rectification instructions to designated on-site monitoring terminals and construction personnel terminals according to the early warning instructions.

Images