Unmanned aerial vehicle and ai-based power distribution network multi-source detection and intelligent identification system and method

Abstract

The application relates to the technical field of intelligent monitoring of power equipment, in particular to a power distribution network multi-source detection and intelligent identification system and method based on a UAV and AI, which adopts a collaborative intelligent architecture of cloud planning, edge execution and cloud closed loop, visible light, infrared, ultraviolet and acoustic data are synchronously collected through a UAV carrying a multi-source sensor array; in the airborne edge computing terminal, real-time processing is sequentially carried out through a feature extraction and space-time alignment module, an environment perception cross-modal adaptive feature fusion module and a defect identification and positioning module, the feature fusion module adopts a double-path attention mechanism and dynamically adjusts the fusion weight according to the environment context; the identification result is fed back to the cloud platform and linked with the power distribution production management system, and operation and maintenance instructions are automatically triggered according to the defect confidence and type to form a business closed loop. The application realizes feature-level deep fusion and all-weather intelligent identification of defects in the power distribution network, and significantly improves the detection precision, environment robustness and operation and maintenance efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for power equipment, specifically to a multi-source detection and intelligent identification system and method for power distribution networks based on drones and AI. Background Technology

[0002] As the final link in power transmission, the power distribution network's equipment is constantly exposed to the natural environment, making it susceptible to defects such as overheating, discharge, and mechanical damage, threatening power supply reliability. Traditional manual inspections are inefficient and risky. In recent years, with the maturity of drone technology, drone-based automated inspections have become an important direction for industry upgrading. By equipping themselves with visible light cameras or infrared thermal imagers, drones can quickly acquire image data of the power line corridor, improving inspection efficiency and safety to some extent. However, this inspection mode, which relies on a single or limited number of sensors, increasingly reveals its inherent limitations when facing the complex and ever-changing operating environment and fault types of the power distribution network. Specifically: 1. Shallow fusion level: Multimodal information fusion often stops at the data level superposition or decision level fusion, failing to achieve interaction and collaboration at the deep feature level, and failing to fully leverage the complementary potential of multi-source data.

[0003] 2. Poor environmental robustness: Visible light imaging fails under extreme conditions such as nighttime, rain, and fog, and the stability of single sensor recognition is insufficient.

[0004] 3. Low detection accuracy: Discharge defect detection is easily affected by environmental interference, has a high false alarm rate, and lacks a multi-source collaborative verification mechanism.

[0005] 4. Insufficient real-time performance: Complex algorithms are difficult to run in real time on airborne edge devices, and reliance on cloud backhaul leads to latency.

[0006] 5. System silos: The detection system is disconnected from business systems such as the power distribution production management system (PMS), failing to form a closed loop for operation and maintenance.

[0007] 6. Insufficient spatiotemporal alignment accuracy: It is difficult to achieve high-precision synchronization and spatial registration of multi-source heterogeneous data, which affects the subsequent fusion effect.

[0008] Therefore, there is an urgent need to develop a new generation of intelligent inspection and fault identification technology for distribution networks that can integrate multi-source sensing, achieve intelligent adaptive fusion at the feature level, have real-time edge processing capabilities, and deeply collaborate with power grid business systems, so as to truly realize the intelligent, efficient and practical nature of inspection operations. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a multi-source detection and intelligent identification system and method for power distribution networks based on drones and AI. It can achieve deep feature-level fusion, has all-weather environmental adaptability, supports real-time edge processing, and can form an intelligent closed loop with business systems.

[0010] This invention is achieved through the following technical solution: This paper provides a multi-source detection and intelligent identification system for power distribution networks based on drones and AI. It adopts a collaborative intelligent architecture of cloud planning, edge execution, and cloud closed-loop, including: The cloud-based collaborative management platform is used to generate the optimal inspection strategy based on task attributes, environmental status, and historical defect distribution information through an optimization model. The optimal inspection strategy includes at least the flight path, sensor working mode, and sampling frequency. The drone platform is equipped with a multi-source sensor array and a high-precision synchronous triggering unit, which is used to synchronously collect visible light, infrared, ultraviolet and acoustic multi-source data of power distribution equipment according to the optimal inspection strategy. The airborne edge computing terminal is used to process multi-source data in real time. Its processing modules include, in sequence: feature extraction and spatiotemporal alignment module, cross-modal adaptive feature fusion module for environmental perception, and defect identification and localization module. The cloud-based collaborative management platform is also used to receive standardized identification results generated by airborne edge computing terminals through standardized interfaces, and to link with the power distribution production management system and fault monitoring system. Based on the confidence level and defect type of the identification results, it automatically triggers operation and maintenance instructions to form a business closed loop.

[0011] Furthermore, the feature extraction and spatiotemporal alignment module includes: Multiple independent lightweight feature extraction branches are used to extract depth features from visible light images, infrared thermal images, ultraviolet imaging, and acoustic data, respectively. A deep learning-based feature alignment network, employing deformable convolutional layers and cross-attention modules, is used to perform high-precision spatial alignment of infrared feature maps, ultraviolet / acoustic joint feature maps, and visible light feature maps. A quality assessment network is used to generate initial confidence scores for each aligned modal feature at each spatial location, and to perform weighted preliminary fusion based on these scores, outputting preliminary fused features with quality perception.

[0012] Furthermore, the cross-modal adaptive feature fusion module for environment perception includes: The environment context encoding unit is used to extract the global environment context vector from multi-source data of the current scene; The dual-path cross-modal attention mechanism includes intra-modal attention path and inter-modal attention path. The intra-modal attention path is used to enhance the salience of defect regions within each modality, while the inter-modal attention path is used to calculate the complementary information weights between features of different modalities. The adaptive weighting unit is used to dynamically adjust the fusion weights of each modality feature based on the environmental context vector, and to perform weighted fusion of the attention-enhanced features.

[0013] Furthermore, the airborne edge computing terminal achieves real-time processing through multi-level optimization techniques, including: A dynamic precision selector is used to adaptively select the computational precision of the model inference based on the complexity of the input data or the urgency of the task. Parallel processing pipelines are used to execute data loading, preprocessing, model inference, and postprocessing stages in parallel to reduce end-to-end processing latency. A power-aware scheduler is used to monitor the status of computing resources and dynamically adjust the hardware operating frequency and voltage to optimize energy consumption while meeting real-time requirements. The hardware acceleration optimization unit utilizes the inference engine to perform graph optimization, layer fusion, and quantization deployment on deep learning models, thereby improving the model's execution efficiency on dedicated computing hardware.

[0014] Furthermore, the defect identification and localization module integrates a multi-source collaborative verification mechanism to cross-verify the spatiotemporally aligned infrared, ultraviolet, and acoustic features, thereby improving the localization accuracy of partial discharge defects and suppressing false alarms.

[0015] Furthermore, the multi-source sensor array includes at least a visible light / multispectral camera, an infrared thermal imager, an ultraviolet imager, an acoustic sensor array, and a separate high-precision digital temperature sensor.

[0016] Furthermore, the preset rules for business linkage are as follows: when the confidence level of the identified defect is higher than the first threshold and belongs to the critical defect type, an emergency work order is automatically generated; when the confidence level is between the first and second thresholds, routine maintenance is scheduled; when the confidence level is lower than the second threshold, it is recorded for trend analysis.

[0017] A method for identifying multiple sources in a power distribution network based on drones and AI includes the following steps: Cloud-based planning steps: The cloud-based collaborative management platform generates the optimal inspection strategy based on task attributes, environmental status, and historical defect distribution information through strategy generation network calculation; Edge execution steps: Based on the optimal inspection strategy, the UAV platform controls multiple source sensors to synchronously collect data through a high-precision synchronous triggering unit, and then performs the following processing sequentially on the airborne edge computing terminal: By extracting features and performing spatiotemporal alignment processing, spatially aligned multimodal deep features are obtained; By performing cross-modal adaptive fusion processing based on environmental perception, multimodal deep features are deeply fused to obtain fused features that are robust to the environment. Defect identification and localization are performed based on fused features to generate standardized identification results; Cloud-based closed-loop process: The standardized identification results are sent back to the cloud-based collaborative management platform, and linked with the power grid business management system according to preset rules to automatically complete operation and maintenance decisions and execution.

[0018] Furthermore, cross-modal adaptive feature fusion processing for environment perception includes: Extract the current scene's context vector; The internal saliency of each modality feature and the complementary weights between modalities are calculated separately using a dual-path attention mechanism. Based on the environmental context vector, the fusion weights of each modality are dynamically generated and weighted fusion is performed.

[0019] Furthermore, the edge execution steps employ dynamic precision selection and power consumption-aware scheduling, adaptively adjusting the operating status of computing resources in real time to meet real-time constraints while optimizing system energy efficiency.

[0020] The beneficial effects of this invention are: This invention achieves adaptive cross-modal deep fusion at the feature level, significantly improving defect recognition accuracy and robustness across all scenarios. It abandons the simple data overlay or decision-level post-fusion in existing technologies. By constructing an environmental perception fusion module that includes a dual-path cross-modal attention mechanism, it realizes the interaction and collaboration of visible light, infrared, ultraviolet and acoustic data in the feature extraction stage of the deep neural network. It can also dynamically adjust the fusion weights of each modality according to the environmental context (such as lighting and weather), enabling the system to adaptively select the most reliable information source. This fundamentally overcomes the information limitations of single-modality or shallow fusion, thereby comprehensively improving the detection rate and recognition accuracy of various defects such as heat, discharge, and mechanical damage.

[0021] By constructing a high-precision spatiotemporal alignment and collaborative verification mechanism, the reliability and positioning accuracy of hidden defects such as discharges are greatly enhanced. Addressing the issues of existing ultraviolet detection being susceptible to interference and having a high false alarm rate, this invention ensures strict spatiotemporal matching of multi-source data through high-precision synchronous trigger control and a feature alignment network based on deep deformable convolution. Furthermore, it innovatively designs a temperature-ultraviolet-acoustic multi-source collaborative verification mechanism, utilizing the correlation of multimodal features after spatiotemporal alignment for cross-verification and false alarm suppression. This achieves precise positioning, intensity quantification, and high-confidence alarm for partial discharge points, solving an industry-wide challenge in the detection of special defects.

[0022] This invention possesses powerful all-weather environmental adaptability and stable perception capabilities under complex conditions. By introducing an adaptive image enhancement preprocessing and dynamic fusion strategy of illumination / environment perception, the system can automatically adjust the processing strategy according to the real-time environmental state vector. Under extreme conditions such as nighttime, rain, fog, and strong backlight, by strengthening the weight of effective modes such as infrared and ultraviolet, it ensures the reliable acquisition and identification of key information, significantly breaking through the environmental limitations of traditional visible light inspection and achieving truly stable operation under all-weather and all-weather conditions.

[0023] This invention achieves low-latency, end-to-end real-time intelligent processing on an airborne edge computing platform. To meet the stringent real-time requirements of UAV inspection, the feature extraction, fusion, and recognition networks have undergone comprehensive lightweight design and hardware acceleration optimization. By employing technologies such as dynamic precision selection, a four-level parallel pipeline, and power-aware scheduling, complex multimodal deep learning models are efficiently deployed on the airborne edge computing unit, achieving millisecond-level end-to-end processing from data acquisition to defect identification and location. This overcomes the latency bottleneck caused by reliance on cloud backhaul and supports the real-time and intelligent operation of inspection tasks.

[0024] This invention establishes a closed-loop business process from intelligent sensing to operation and maintenance execution, effectively improving the overall efficiency of power grid operation and maintenance. Through a cloud-based collaborative management platform and standardized API interfaces, this invention automatically and structurally pushes edge identification results (defect type, location, confidence level, etc.) to the distribution production management system (PMS) and fault monitoring system, and automatically triggers maintenance work orders, resource scheduling, and status updates according to preset linkage rules. This realizes a closed-loop intelligent operation and maintenance process from sensing to diagnosis to decision-making to execution, completely solving the "information silo" problem between advanced detection systems and existing business systems, and significantly improving the speed of operation and maintenance response and the level of automation.

[0025] Furthermore, this invention provides an integrated and implementable system-level solution from task planning to result application. It employs a collaborative intelligent architecture of cloud planning-edge execution-cloud closed loop. The cloud performs globally optimal inspection planning based on digital twins and policy generation networks, while the edge performs high-precision synchronous data collection and real-time intelligent analysis. Finally, the cloud completes business integration and decision support, forming a complete and systematic technical system. This solution not only focuses on breakthroughs in core algorithms but also solves key practical application problems at the engineering level, such as multi-source synchronization, edge deployment, and system integration, ensuring the high efficiency, practicality, and rapid implementation of the technology. Attached Figure Description

[0026] Figure 1 The overall flowchart of the collaborative intelligent architecture of "cloud planning - edge execution - cloud closed loop" provided by the present invention.

[0027] Figure 2This is a schematic diagram of the multimodal feature extraction and high-precision spatiotemporal alignment module in this invention.

[0028] Figure 3 This is a schematic diagram of the cross-modal adaptive feature fusion module for environmental perception in this invention.

[0029] Figure 4 This is a flowchart illustrating the execution process of the edge real-time optimization module in this invention. Detailed Implementation

[0030] To clearly illustrate the technical features of this solution, the following detailed implementation method will be used to explain the solution.

[0031] The hardware deployment of a power distribution network multi-source detection and intelligent identification system based on drones and AI constitutes a complete three-layer collaborative system architecture, adopting a collaborative intelligent architecture of "cloud planning - edge execution - cloud closed loop", including: The cloud-based collaborative management platform is used to generate the optimal inspection strategy based on task attributes, environmental status, and historical defect distribution information through an optimization model. The optimal inspection strategy includes at least the flight path, sensor working mode, and sampling frequency. The drone platform is equipped with a multi-source sensor array and a high-precision synchronous triggering unit, which is used to synchronously collect visible light, infrared, ultraviolet and acoustic multi-source data of power distribution equipment according to the optimal inspection strategy. The airborne edge computing terminal is used to process multi-source data in real time. Its processing modules include, in sequence: feature extraction and spatiotemporal alignment module, cross-modal adaptive feature fusion module for environmental perception, and defect identification and localization module. The cloud-based collaborative management platform is also used to receive standardized identification results generated by airborne edge computing terminals through standardized interfaces, and to link with the power distribution production management system and fault monitoring system. Based on the confidence level and defect type of the identification results, it automatically triggers operation and maintenance instructions to form a business closed loop.

[0032] At the forefront is a drone inspection platform equipped with a core sensor array. A high-stability industrial-grade drone is used, and its high-performance three-axis stabilized gimbal integrates multi-source detection payloads: a visible light / multispectral camera for high-resolution texture and spectral information acquisition, an infrared thermal imager for capturing temperature field distribution, an ultraviolet imager for detecting corona discharge, an acoustic sensor array for picking up discharge sounds, and an independent high-precision digital temperature sensor for key point contact temperature measurement. All sensors are hardware-controlled by a synchronization trigger unit based on a high-precision crystal oscillator, ensuring millisecond-level timestamp synchronization of multi-source data.

[0033] The middle layer is a ruggedized airborne edge computing terminal integrated into the belly of the drone. It has a built-in high-performance AI computing chip (such as GPU / NPU) and is responsible for running all lightweight algorithms to achieve real-time data processing.

[0034] The backend is a cloud-based collaborative management platform deployed in the power data center. It seamlessly connects with the power distribution production management system (PMS), fault monitoring system, and mobile operation and maintenance applications through standardized data interfaces, and is responsible for global task scheduling, data aggregation, intelligent decision-making, and business closed-loop management.

[0035] A method for identifying multiple sources in a power distribution network based on drones and AI is presented. The system follows an intelligent process of "cloud planning - edge execution - cloud closed loop," with specific steps as follows, where formulas describe the implementation details of key algorithms: 1. Cloud-based planning phase: Before the task begins, the cloud-based collaborative management platform makes intelligent decisions based on multi-source information. Decision input vector Includes: Task attribute vector T = [type, priority, region range] T The environmental state vector E = [light intensity, temperature, humidity, wind speed]. T Historical defect distribution vector H Network generation through strategy Calculate the optimal sensing strategy: \MERGEFORMAT (1); in: This is a policy vector that includes the flight path, sensor operating mode, and sampling frequency. For coverage rewards, Rewards for efficiency For energy consumption costs, The weights are adaptive. This strategy, after being verified through digital twin simulation, is then deployed to the edge of the drone.

[0036] 2. Edge execution phase: After the drone takes off, according to the strategy It autonomously flies to the target area. A high-precision synchronous triggering unit controls multiple source sensors to synchronously collect data, forming raw data. Data is transmitted in real time to the airborne edge computing terminal, where the following processes are executed sequentially (corresponding to...). Figure 1 (Flowchart shown) 2.1 Feature Extraction and Spatiotemporal Alignment: Deep features are extracted from each modality of data using a lightweight network: Visible light / multispectral branch: Input image An improved lightweight backbone network is adopted. Extracting rich spatial texture and spectral features: (2); Infrared thermal imaging branch: Input thermal image (Single-channel temperature matrix), through a depthwise separable convolutional network Extracting temperature distribution and thermal gradient characteristics Its network design is particularly enhanced to detect minute temperature differences.

[0037] Ultraviolet / Acoustic branch; for ultraviolet imaging Extracting the morphology and intensity characteristics of the discharge spot for acoustic signals The time-frequency spectrum is obtained after short-time Fourier transform (STFT), and then the frequency domain features of the discharge pulse are extracted through a one-dimensional convolutional network. These two features are then concatenated to form a joint feature. .

[0038] This invention designs a learnable feature alignment network. Its input is the reference modal features ( ) and the modal features to be aligned ( The output is the spatially aligned features: (3); in It uses deformable convolutional layers and a cross-attention module. It predicts a set of 2D offset fields. The sampling position of the convolution kernel on the feature map to be aligned is dynamically adjusted to achieve precise spatial distortion.

[0039] Alignment loss function The approach combines multi-scale feature consistency loss with task-guided loss. (4); in Indicates the first Feature map of the layer To mitigate the losses in downstream identification tasks, ensure alignment is beneficial to the final goal.

[0040] Aligned multimodal features { Instead of direct splicing, it first goes through a lightweight quality assessment network. For each spatial location Each modal feature is assigned an initial confidence score. : (5); The score reflects the reliability of the modal data at that location (e.g., infrared data in areas with thermal reflection interference under strong sunlight scores low). Subsequently, a weighted preliminary fusion is performed based on the scores to obtain preliminary fusion features for quality perception. This will provide high-quality input for the next stage of in-depth integration.

[0041] 2.2 Cross-modal adaptive feature fusion for environmental perception: Modules such as Figure 3 As shown, the system receives aligned multimodal features and performs deep interaction and adaptive integration at the feature level. The environmental context coding system extracts environmental context vectors from the raw data and sensors. : (6); This vector encodes the global visual characteristics, thermal characteristics, and physical environment state of the current scene, serving as global prior information to guide fusion.

[0042] Next, features are enhanced using a dual-path attention mechanism: The dual-path cross-modal attention mechanism is the core of the fusion, which includes two paths: intra-modal attention and inter-modal attention.

[0043] Intramodal attention is applied to each modal feature map. Calculate the correlation between different spatial locations within it. For infrared features... Its location The context-enhanced features at the location are: (7); in: The weight matrix enhances the saliency of defective regions and suppresses background noise.

[0044] Intermodal attention calculates the complementary information weights between features from different modalities. For example, it uses visible light features to query infrared features: (8); This operation enables visible light features to extract thermal anomaly information from infrared features. Simultaneously, a symmetric attention computation process extracts texture structure information from visible light features using infrared features.

[0045] Then, adaptive weighted fusion and optimization are performed, and the modal features are enhanced with attention. This is not equal treatment. A lightweight environment-adaptive gating network. According to the context Generate dynamic fusion weights : (9); At night, reduce, and Increase.

[0046] Weighted fusion of attention-enhanced features: Final fusion features for: (10); Subsequently The information is further refined by using a multilayer perceptron (MLP) with residual connections and then output to the recognition network.

[0047] 2.3 Defect Identification and Location: The fused features and data from independent temperature sensors are input into a lightweight recognition network to complete defect detection. Edge processing employs multi-level optimization techniques to ensure real-time performance. Dynamic precision selection: Use a dynamic precision selector to select the precision based on the complexity of the input data (e.g., image entropy). Based on the urgency of the task, the inference precision is selected in real time. : (11); Parallel processing and power consumption optimization: Minimize cycle time through a four-stage parallel pipeline; After selecting the precision, the data enters a four-stage parallel pipeline: data loading, data preprocessing, model inference, and result post-processing. Through carefully designed pipeline buffering and parallel execution, end-to-end latency is minimized. Pipeline cycle time. Depends on the slowest stage: (12); The optimization goal is to minimize .

[0048] Power-aware scheduling and hardware acceleration systems monitor chip power consumption. ,temperature and battery power The dynamic voltage and frequency adjustment strategy is based on the processing load. Dynamically adjust CPU / GPU / NPU frequency and voltage In order to meet real-time constraints Minimize energy consumption under the premise of: (13); 2.4 Hardware Acceleration and Optimization: To meet the stringent computing power and power consumption constraints of the airborne platform, the above processing steps employed hardware acceleration optimization. This hardware acceleration optimization utilized inference engines such as NVIDIA TensorRT and Qualcomm SNPE to perform graph optimization, layer fusion, and automatic kernel tuning on the deployed deep learning models, and then quantized their deployment to the corresponding GPU / NPU hardware. Specifically, the execution efficiency of convolution and attention operators on high-performance AI chips was optimized to maximize processing speed.

[0049] 2.5 Result Compression and Transmission: Before being transmitted back to the cloud, the standardized results generated by edge processing (including defect type, location, confidence level, etc.) will adaptively select the compression rate and transmission protocol based on defect priority and wireless channel status. For critical defects with high confidence, a low compression rate and high reliability transmission mode is adopted to ensure data integrity and accuracy; for non-critical data or cached data, a high compression rate and best-effort transmission mode is adopted to maximize the use of limited wireless channel bandwidth.

[0050] 3. Cloud-based closed-loop stage: Standardized results generated at the edge (including defect type and confidence level) d Location coordinates timestamp and eigenvectors After compression, the data is transmitted back to the cloud. The cloud platform interfaces with the power distribution production management system (PMS) and fault monitoring system through standardized API interfaces to achieve automatic work order generation, status updates, and decision support, forming a closed business loop of perception-diagnosis-decision-execution.

[0051] Linkage rules can be formally represented as: (14); in As an example of a defect, This is a set of critical defect types.

[0052] The overall optimization objective of this system can be modeled as a multi-constraint optimization problem: MERGEFORMAT(15); in: For all neural network parameters, For inference accuracy.

[0053] A sudden, localized strong wind and heavy rainstorm struck a 10kV power distribution line in the suburbs of a coastal city overnight. Following the rain, the power company's dispatch center used the system described in this invention to conduct emergency, detailed inspections of potentially damaged sections of the line.

[0054] 1. Cloud-based Planning: The cloud-based collaborative management platform receives an emergency inspection task. The task attribute vector T is set to [Type = Special inspection after fault, Priority = High, Area range = Towers A001-A020]. The platform integrates the real-time environmental status vector E (nighttime, high humidity, high wind speed) and the historical defect distribution of the area (previously reported insulator flashover cases), and generates a network through a strategy. (Formula (1)) quickly calculates the optimal inspection strategy The strategy instructs the drone to fly along a preset high-risk path and commands the multi-source sensor array to adopt a high-frequency sampling mode, with particular emphasis on enhancing the sensitivity of the ultraviolet and acoustic sensors to focus on detecting potential discharge hazards.

[0055] 2. Edge execution: The drone flies to the target area (near pole A005).

[0056] Data Acquisition and Feature Alignment: A high-precision synchronous triggering unit controls the sensors to synchronously acquire data. A visible light camera captures a blurred image of the insulator's outline in night mode. I The infrared thermal imager detected a slight temperature rise in the middle of the insulator string, forming a thermal image. I The ultraviolet imager clearly captured intermittent discharge spots on the surface of the insulator. I UV; the acoustic sensor simultaneously recorded a "buzzing" discharge sound. Aaudio At the edge, the Feature Alignment Network (FAN) successfully integrated infrared hotspot features. Fir and ultraviolet discharge characteristics Fuv Location of insulator in visible light image Fvis Perform high-precision spatial alignment.

[0057] Feature fusion for environmental perception: The environmental context encoding unit generates an environmental context vector based on global information (nighttime, high humidity). cenv (Formula (6)). The dual-path attention mechanism begins to work: intramodal attention enhances the salience of the discharge spot in the ultraviolet feature map; intermodal attention allows the visible light features to "sense" the temperature rise region from the infrared features. Subsequently, the environment-adaptive gating network dynamically assigns higher fusion weights to the infrared and ultraviolet modes according to the nighttime environment. βir and βuv Significantly higher than βvis The enhanced features were then weighted and fused to effectively suppress interference from poor-quality visible light images at night.

[0058] Defect identification and localization: Based on the fused robust features, the defect identification module accurately determines that the insulator has a "partial discharge" defect, with a localization accuracy of centimeters, and provides a confidence level of up to 0.96 by integrating multi-source information. The entire edge processing process benefits from dynamic precision selection (automatically enabling FP16 precision in complex scenarios) and hardware acceleration, with an end-to-end latency of only 120 milliseconds.

[0059] 3. Cloud-based closed-loop: The standardized results generated at the edge (defect type: partial discharge; confidence level: 0.96; location: tower A005 insulator) are transmitted back to the cloud in real time. The cloud platform makes a judgment based on preset linkage rules: due to the confidence level... cd =0.96> τhigh And the defect type belongs to the set of critical defects. Υcritical The system automatically generates an emergency maintenance work order and pushes it to the Power Distribution Production Management System (PMS) instantly via a standardized API interface. The PMS system immediately sends an alarm to the nearest maintenance team's mobile terminal and automatically dispatches maintenance resources. Based on the precise location and defect information provided in the work order, maintenance personnel quickly rush to the site to replace the insulator and feed the results back to the system, forming a closed loop.

[0060] From defect detection to work order generation, the entire process requires no manual intervention, reducing response time from hours in traditional methods to minutes. This invention's system successfully overcomes the limitations of single visible light inspection in harsh nighttime environments. Through multi-source fusion and intelligent decision-making, it accurately identifies potential faults, preventing possible line tripping accidents and demonstrating its strong environmental adaptability, detection accuracy, and operational efficiency.

[0061] This invention seeks the optimal solution for detection performance under resource constraints by linking feature-level adaptive fusion, environmental perception decision-making, and real-time edge optimization.

[0062] The system operates following an intelligent process of cloud-based planning, edge execution, and cloud-based closed-loop processing. Before execution, the cloud platform integrates task requirements, real-time environmental data, and a historical defect database. Through strategy generation networks and digital twin simulations, it plans the optimal inspection route, sensor operating modes, and sampling strategies, and distributes these plans to the designated UAVs. After takeoff, the UAVs autonomously fly to the target area. The high-precision synchronization triggering unit, based on strategy instructions, controls visible light, infrared, and ultraviolet sensors to perform strictly synchronized data acquisition of the same target. The raw data packets are then streamed in real-time to the onboard edge computing terminal.

[0063] At the edge, data immediately enters the parallel processing pipeline. First, the feature extraction and high-precision spatiotemporal alignment module is activated: each modality's data is processed by a lightweight network to extract deep features. Then, a deep learning alignment network based on deformable convolution is used to perform pixel-level spatial alignment of infrared, ultraviolet, and other features with visible light features, solving the misalignment problem caused by differences in viewing angle and sensor. Next, the cross-modal adaptive feature fusion module for environmental perception works: the system first encodes the current lighting, weather, and other environmental contexts, and then uses a dual-path (intra-modal and inter-modal) attention mechanism to allow visible light, infrared, ultraviolet, and acoustic features to interact and complement each other at the depth level. The fusion weights of each modality are dynamically adjusted according to the environmental context (e.g., enhancing infrared and ultraviolet weights at night) to generate environmentally robust fusion features. These features, along with accurate temperature measurement data provided by an independent temperature sensor, are input into a lightweight defect recognition network to complete the simultaneous detection, classification, and precise localization of defects such as overheating, discharge, and mechanical damage.

[0064] Standardized results generated by edge processing (including defect type, location, confidence level, temperature, etc.) are compressed and transmitted back to the cloud in real time. The cloud platform automatically parses the results through standardized APIs, associates them with the power grid asset map, and automatically generates maintenance work orders and updates equipment status to the PMS system according to preset linkage rules, while simultaneously pushing alerts to the monitoring system and mobile devices. After on-site handling by maintenance personnel, the results are reported back, and the cloud updates the closed loop, thus achieving full automation from intelligent sensing and accurate diagnosis to maintenance execution.

[0065] This technical solution provides a complete and practical intelligent inspection solution for power distribution networks, from theoretical modeling to engineering implementation, effectively solving the four major bottlenecks of traditional methods in terms of environmental adaptability, information utilization, real-time performance, and system integration.

[0066] Of course, the above description is not limited to the examples above. Technical features not described in this invention can be implemented by or using existing technology, and will not be repeated here. The above embodiments and drawings are only used to illustrate the technical solutions of this invention and are not intended to limit this invention. This invention has been described in detail with reference to preferred embodiments. Those skilled in the art should understand that any changes, modifications, additions or substitutions made by those skilled in the art within the scope of this invention do not depart from the spirit of this invention and should also fall within the scope of protection of the claims of this invention.

Claims

1. A power distribution network multi-source detection and intelligent identification system based on drones and AI, characterized in that: It adopts a collaborative intelligent architecture of "cloud planning - edge execution - cloud closed loop", including: The cloud-based collaborative management platform is used to generate the optimal inspection strategy based on task attributes, environmental status, and historical defect distribution information through an optimization model. The optimal inspection strategy includes at least the flight path, sensor working mode, and sampling frequency. The drone platform is equipped with a multi-source sensor array and a high-precision synchronous triggering unit, which is used to synchronously collect visible light, infrared, ultraviolet and acoustic multi-source data of power distribution equipment according to the optimal inspection strategy. The airborne edge computing terminal is used to process multi-source data in real time. Its processing modules include, in sequence: feature extraction and spatiotemporal alignment module, cross-modal adaptive feature fusion module for environmental perception, and defect identification and localization module. The cloud-based collaborative management platform is also used to receive standardized identification results generated by airborne edge computing terminals through standardized interfaces, and to link with the power distribution production management system and fault monitoring system. Based on the confidence level and defect type of the identification results, it automatically triggers operation and maintenance instructions to form a business closed loop.

2. The power distribution network multi-source detection and intelligent identification system based on UAVs and AI according to claim 1, characterized in that: The feature extraction and spatiotemporal alignment module includes: Multiple independent lightweight feature extraction branches are used to extract depth features from visible light images, infrared thermal images, ultraviolet imaging, and acoustic data, respectively. A deep learning-based feature alignment network, employing deformable convolutional layers and cross-attention modules, is used to perform high-precision spatial alignment of infrared feature maps, ultraviolet / acoustic joint feature maps, and visible light feature maps. A quality assessment network is used to generate initial confidence scores for each aligned modal feature at each spatial location, and to perform weighted preliminary fusion based on these scores, outputting preliminary fused features with quality perception.

3. The power distribution network multi-source detection and intelligent identification system based on UAVs and AI according to claim 1, characterized in that: The cross-modal adaptive feature fusion module for environment perception includes: The environment context encoding unit is used to extract the global environment context vector from multi-source data of the current scene; The dual-path cross-modal attention mechanism includes intra-modal attention path and inter-modal attention path. The intra-modal attention path is used to enhance the salience of defect regions within each modality, while the inter-modal attention path is used to calculate the complementary information weights between features of different modalities. The adaptive weighting unit is used to dynamically adjust the fusion weights of each modality feature based on the environmental context vector, and to perform weighted fusion of the attention-enhanced features.

4. The power distribution network multi-source detection and intelligent identification system based on UAVs and AI according to claim 1, characterized in that: Airborne edge computing terminals achieve real-time processing through multi-level optimization techniques, including: A dynamic precision selector is used to adaptively select the computational precision of the model inference based on the complexity of the input data or the urgency of the task. Parallel processing pipelines are used to execute data loading, preprocessing, model inference, and postprocessing stages in parallel to reduce end-to-end processing latency. A power-aware scheduler is used to monitor the status of computing resources and dynamically adjust the hardware operating frequency and voltage to optimize energy consumption while meeting real-time requirements. The hardware acceleration optimization unit utilizes the inference engine to perform graph optimization, layer fusion, and quantization deployment on deep learning models, thereby improving the model's execution efficiency on dedicated computing hardware.

5. The power distribution network multi-source detection and intelligent identification system based on UAVs and AI according to claim 1, characterized in that: The defect identification and localization module integrates a multi-source collaborative verification mechanism, which is used to cross-verify the spatiotemporally aligned infrared, ultraviolet and acoustic features to improve the localization accuracy of partial discharge defects and suppress false alarms.

6. The power distribution network multi-source detection and intelligent identification system based on UAVs and AI according to claim 1, characterized in that: The multi-source sensor array includes at least a visible light / multispectral camera, an infrared thermal imager, an ultraviolet imager, an acoustic sensor array, and a separate high-precision digital temperature sensor.

7. The power distribution network multi-source detection and intelligent identification system based on UAVs and AI according to claim 1, characterized in that: The preset rules for business linkage are as follows: when the confidence level of the identified defect is higher than the first threshold and belongs to the critical defect type, an emergency work order is automatically generated; when the confidence level is between the first and second thresholds, routine maintenance is scheduled; when the confidence level is lower than the second threshold, it is recorded for trend analysis.

8. A recognition method for a power distribution network multi-source detection and intelligent recognition system based on unmanned aerial vehicles and AI, as described in any one of claims 1 to 7, characterized in that, Includes the following steps: Cloud-based planning steps: The cloud-based collaborative management platform generates the optimal inspection strategy based on task attributes, environmental status, and historical defect distribution information through strategy generation network calculation; Edge execution steps: Based on the optimal inspection strategy, the UAV platform controls multiple source sensors to synchronously collect data through a high-precision synchronous triggering unit, and then performs the following processing sequentially on the airborne edge computing terminal: By extracting features and performing spatiotemporal alignment processing, spatially aligned multimodal deep features are obtained; By performing cross-modal adaptive fusion processing based on environmental perception, multimodal deep features are deeply fused to obtain fused features that are robust to the environment. Defect identification and localization are performed based on fused features to generate standardized identification results; Cloud-based closed-loop process: The standardized identification results are sent back to the cloud-based collaborative management platform, and linked with the power grid business management system according to preset rules to automatically complete operation and maintenance decisions and execution.

9. The identification method of the power distribution network multi-source detection and intelligent identification system based on UAVs and AI as described in claim 8, characterized in that: Cross-modal adaptive feature fusion processing for environment perception includes: Extract the current scene's context vector; The internal saliency of each modality feature and the complementary weights between modalities are calculated separately using a dual-path attention mechanism. Based on the environmental context vector, the fusion weights of each modality are dynamically generated and weighted fusion is performed.

10. The identification method of the power distribution network multi-source detection and intelligent identification system based on UAVs and AI according to claim 8, characterized in that: The edge execution steps employ dynamic precision selection and power consumption-aware scheduling, adaptively adjusting the operating status of computing resources in real time to meet real-time constraints while optimizing system energy efficiency.

Images