Power transmission line construction disturbance change detection method and device

Abstract

The invention discloses a power transmission line construction disturbance change detection method, and belongs to the technical field of power intelligent monitoring and remote sensing image processing. The method comprises the following steps: acquiring a multi-source image and carrying out geometric correction, registration and ROI (Region of Interest) cutting; denoising the pixel-level sequential sequence by adopting sequential filtering to obtain a high-quality sequential image; selecting a reference image and a current image, inputting the selected reference image and the current image into an image encoder of a general segmentation model SAM, calculating a feature difference to generate a change intensity graph, and automatically generating prompt information; after boundary expansion optimization, an SAM prompt encoder and a mask decoder are input, and a pixel-level change mask is output; and extracting attributes of a change region through morphological operation and edge smoothing, and generating graded early warning according to an area, a type and a frequency threshold value.

Description

Technical Field

[0001] This invention relates to the field of power transmission line monitoring technology, and more particularly to a computer vision-based method for monitoring construction disturbances in power transmission lines. More specifically, it relates to a power transmission line construction disturbance monitoring scheme that integrates temporal filtering technology with a change detection (SAM-CD) method based on the Segment Anything Model (SAM) model. This scheme is applicable to the identification, location, and early warning of construction activities in and around high-voltage power transmission line corridors. It is compatible with multi-source image data such as visible light, infrared, UAV, and satellite remote sensing, and is used to ensure the safe and stable operation of power transmission lines. Background Technology

[0002] As a crucial component of the power grid, the operational status of transmission lines directly impacts the safety, stability, and reliability of the power system. Because transmission lines typically span long distances and cover vast areas, often extending hundreds of kilometers and traversing complex terrains such as mountains, deserts, and cities, they are highly susceptible to various factors, including natural disasters (such as lightning strikes and icing), equipment aging, and human-caused construction disturbances. Among these, construction disturbances (such as mechanical intrusion and earthwork excavation) have become one of the main causes of transmission line failures in recent years.

[0003] Traditional transmission line inspection methods mainly rely on manual inspection and fault alarm equipment. However, manual inspection has problems such as long inspection cycle and strong environmental dependence, while fault alarm methods have low accuracy when facing multi-source disturbances, making it difficult to meet the needs of modern power grids for efficient and intelligent operation and maintenance.

[0004] With the continuous development of power monitoring technology, the prototype of a modern smart grid is gradually taking shape. Existing online monitoring methods, such as power quality monitoring devices and transient waveform recording devices, are capable of monitoring transient voltage and current during and before a fault occurs. However, for external physical disturbances such as construction disturbances, existing methods still have the following shortcomings:

[0005] Limitations of sensor monitoring: Existing sensors such as current and voltage primarily monitor electrical parameters. They lack sufficient sensitivity to non-electrical disturbances such as external construction machinery intrusion and ground excavation, and cannot directly reflect physical intrusion behavior. Initial identification usually requires multi-sensor fusion and complex feature extraction algorithms.

[0006] The shortcomings of traditional image analysis methods: Although visible light or infrared monitoring equipment can provide intuitive image information, traditional image processing algorithms are not robust in complex environments such as changes in lighting, severe weather interference, and vegetation obstruction, and it is difficult to automatically and accurately identify minor changes caused by construction disturbances from massive image data.

[0007] Challenges of Deep Learning Applications: While deep learning-based change detection methods have made some progress in recent years, they still face challenges in power transmission line scenarios, including scarce samples, insufficient model generalization ability, and poor performance in identifying unknown disturbance types. Furthermore, these models typically require substantial computing resources, making real-time detection and analysis difficult to achieve on resource-constrained edge devices.

[0008] Although existing transmission line monitoring technologies and change detection methods have made some progress, there are still many technical pain points and performance bottlenecks in practical applications, mainly including the following aspects:

[0009] Missing temporal correlation:

[0010] Existing change detection methods based on single-frame or dual-frame static images cannot effectively distinguish between "instantaneous interference" (such as birds flying by or flashing light) and "continuous construction disturbances," resulting in a false alarm rate that generally exceeds 15%. The resulting invalid alarms lead to an invalid inspection rate of over 20% for maintenance personnel, wasting significant manpower and reducing overall maintenance efficiency.

[0011] Insufficient sensitivity to minute changes:

[0012] Traditional change detection (CD) models have limited performance in detecting small-scale disturbances. When the tower displacement is less than 5 mm or the excavated area is less than 10 square meters, the recognition rate is usually less than 60%, which is insufficient to meet the engineering safety requirements for high-precision early warning of disturbances during power transmission line construction.

[0013] Poor target segmentation adaptability:

[0014] Existing general segmentation models have low edge extraction accuracy when dealing with specific targets in transmission line scenarios (such as tower angle steel, construction scaffolds, etc.), and the segmentation boundary error often exceeds 3 pixels, which seriously affects the spatial positioning accuracy of the changing area and the accurate assessment of the disturbance range.

[0015] Weak environmental adaptability:

[0016] Under complex environmental conditions such as sandstorms, backlighting, rain, snow, and cloud shadows, the signal-to-noise ratio of monitoring images decreases significantly, resulting in a reduction of more than 30% in change detection accuracy. Existing algorithms are difficult to adapt to the monitoring needs of power transmission line construction disturbances under multiple geographical and climatic conditions.

[0017] In recent years, the Segment Anything Model (SAM) has been widely used in computer vision due to its powerful general image segmentation capabilities and excellent zero-shot generalization performance. Its fine-grained segmentation capabilities provide a new approach for achieving high-precision change detection. However, directly applying the SAM model to the time-series change detection of power transmission lines still faces two challenges:

[0018] On the one hand, the SAM model is not designed specifically for time-series change detection and lacks modeling and utilization of time-dimensional information; on the other hand, transmission line monitoring videos or image sequences usually contain a lot of noise such as illumination changes, camera shake, and cloud shadow interference. If there is a lack of effective time-series filtering and noise suppression mechanisms, the SAM may output false change areas, thereby significantly reducing detection accuracy.

[0019] Therefore, there is an urgent need to propose a change detection method that can integrate temporal information, suppress redundant noise, and give full play to the segmentation performance of the SAM model, so as to achieve high-precision and robust automatic identification and monitoring of construction disturbances in transmission lines. Summary of the Invention

[0020] Technical issues

[0021] To address the problems of high false change detection rate, strong noise sensitivity, poor adaptability to complex environments, and insufficient ability to identify unknown disturbance types in existing technologies, the present invention aims to provide a SAM-CD method for detecting construction disturbance changes in transmission lines that integrates time-series filtering technology. This method fully leverages the high-precision advantage of the SAM (Segment Anything Model) model in target segmentation by introducing time-series information modeling and noise suppression mechanisms, thereby achieving intelligent and high-precision monitoring and early warning of construction disturbances in transmission lines.

[0022] The specific objectives of this invention include:

[0023] Achieve continuous automatic monitoring: Conduct continuous and automatic detection and monitoring of construction disturbances in the transmission line corridor and its surrounding areas, reduce manual intervention, and realize an unmanned and real-time monitoring process.

[0024] Improved anti-interference capability: Through time-series filtering and noise suppression strategies, the effects of factors such as changes in lighting, weather interference, and camera shake are effectively reduced, significantly lowering the false alarm rate and improving the robustness of change detection.

[0025] Enhanced accuracy of change identification and location: Utilizing the high-precision segmentation capability of the SAM model, surface changes caused by construction activities such as mechanical intrusion, earthwork excavation, and new building construction can be accurately identified, located, and segmented into regions, enabling precise tracking of disturbance sources.

[0026] Improve model generalization ability: Build a detection framework with good generalization performance that can adapt to different geographical environments, different seasons and unknown types of construction disturbance scenarios, and ensure the stable performance of the model under multi-source and multi-temporal conditions.

[0027] Provides intelligent early warning support: Generates reliable early warning information based on detection results, providing scientific decision-making basis for transmission line operation and maintenance personnel, and ensuring the safe and stable operation of the power grid.

[0028] Technical solution

[0029] This invention is proposed to solve the above-mentioned problems, and its purpose is to provide a method for detecting changes in construction disturbances in transmission lines, characterized by comprising the following steps:

[0030] S1. Data input and preprocessing: Acquire multi-source image data, and sequentially perform geometric correction, radiometric correction, image registration and region of interest (ROI) cropping on the images to obtain a registered ROI image sequence.

[0031] S2. Temporal filtering: Construct a pixel-level temporal sequence based on the ROI image sequence, and perform temporal filtering on each pixel temporal sequence to obtain the denoised temporal image stack F(TN)…F(T);

[0032] S3. Change Candidate Generation: Select the current image F(T) and the historical reference image F(TK) as keyframe pairs, input them into the image encoder of the general segmentation model SAM to obtain depth features, calculate the difference between the two features to generate a change intensity map, perform threshold segmentation and connected component analysis on the change intensity map to obtain change candidate regions and generate corresponding bounding rectangle prompt information.

[0033] S4. Prompt optimization and fine segmentation: The prompt information is optimized by boundary dilation. The optimized prompt and F(T) are input into the prompt encoder and mask decoder of SAM, and the pixel-level binary transformation mask is output.

[0034] S5. Post-processing and attribute extraction: Perform morphological opening and closing operations and edge smoothing on the binary transformation mask, and extract attributes such as the position, area and type of the transformation area.

[0035] S6. Warning generation and result output: According to preset rules, the changed areas are given graded warnings, and optimized masks, labeled images, attribute reports and warning information for push notifications are output.

[0036] To address the aforementioned problems, a transmission line construction disturbance change detection device according to the present invention is characterized by comprising: a data acquisition module, a temporal filtering module, a change detection module, and an early warning output module; the data acquisition module is used to acquire multi-source image data of the transmission line corridor and its surrounding area, and sequentially perform geometric correction, radiometric correction, image registration, and ROI cropping on the images to generate a registered ROI image sequence; the temporal filtering module is used to perform temporal domain filtering processing on the ROI image sequence to generate a denoised temporal image stack; the change detection module is used to select the reference time image and the current time image and input them into the image encoder of the general segmentation model SAM for feature extraction, and generate a change intensity map and circumscribed rectangle prompt information based on feature differences; the early warning output module is used to perform post-processing and attribute extraction on the change detection results, and generate graded early warning results according to preset rules.

[0037] Beneficial effects

[0038] Compared with existing technologies, the SAM-CD transmission line construction disturbance change detection method with integrated time-series filtering provided by this invention has achieved significant progress in terms of detection accuracy, robustness, automation level, and engineering adaptability, specifically in the following aspects:

[0039] (1) High accuracy and low false alarm rate

[0040] By introducing a temporal filtering module before change detection, transient noise interference such as changes in illumination, weather disturbances, and short-term occlusion is effectively suppressed, thereby significantly reducing false change signals. This preprocessing mechanism enables the feature extraction process of the SAM model to focus more on real surface disturbances, significantly improving detection accuracy and reducing the false alarm rate by more than 60% compared to traditional methods, ensuring the reliability of monitoring results.

[0041] (2) Strong generalization ability

[0042] Leveraging the superior zero-shot segmentation capability of the SAM model, this invention can adapt to diverse geographical environments and unknown construction disturbance types without requiring extensive sample training for specific scenarios or disturbance types. This feature significantly reduces model deployment and maintenance costs, enabling broad adaptation to different regions, climates, and image sources.

[0043] (3) Refined detection and output

[0044] The SAM model can generate pixel-level precision change mask outputs. Compared with traditional detection methods that can only provide rectangular change boxes or coarse areas, this invention can provide accurate shape, area and geographical location information of the change area, providing power transmission line operation and maintenance personnel with more intuitive, quantitative and high-confidence decision-making basis.

[0045] (4) High level of automation and intelligence

[0046] This invention automates the entire process from data preprocessing, time-series filtering, change alert generation to fine segmentation and early warning output, without requiring manual intervention. The system can automatically detect and classify disturbances, enabling intelligent processing of transmission line construction disturbance monitoring and early warning, improving operation and maintenance efficiency by over 80%, and significantly reducing manpower and misjudgment intervention.

[0047] (5) Good engineering adaptability and scalability

[0048] This invention employs a modular design with a flexible system structure, allowing for the selection of suitable data sources (visible light, infrared, UAV, or satellite remote sensing) and filtering algorithms (Kalman filtering, moving average filtering, etc.) based on different application requirements. Some modules (such as temporal filtering and feature difference calculation) can be deployed on edge computing devices to reduce data transmission and latency, while the SAM fine segmentation module can be executed in the cloud or on high-performance servers, forming an "edge + cloud" collaborative computing architecture. This design not only balances real-time performance and accuracy requirements but also effectively reduces hardware deployment costs, demonstrating excellent engineering application prospects.

[0049] In summary, this invention significantly improves the accuracy, robustness, and intelligence of transmission line construction disturbance monitoring by introducing time-series filtering enhancement, adaptive feature difference analysis, and refined SAM segmentation fusion, providing reliable technical support for the safe operation of power systems. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the process of a method for detecting changes in construction disturbances in transmission lines according to an embodiment of the present invention.

[0051] Figure 2 This is a block diagram illustrating a transmission line construction disturbance change detection device according to an embodiment of the present invention.

[0052] Figure 3 This is a schematic diagram illustrating the timing filtering effect according to an embodiment of the present invention.

[0053] Figure 4 This is a schematic diagram illustrating the SAM-CD change detection process according to an embodiment of the present invention.

[0054] Figure 5 This is a logic diagram illustrating the detection logic for changes in construction disturbances of transmission lines according to an embodiment of the present invention. Detailed Implementation

[0055] Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings, so that those skilled in the art can readily understand the present invention.

[0056] This document only describes the parts necessary for understanding the technical content of the present invention, and the description of the remaining parts will be omitted to avoid confusion about the essence of the present invention. This should be noted. Moreover, in this process, for the sake of clarity and convenience of description, the thickness of the lines or the size of the constituent elements shown in the figures may be exaggerated.

[0057] The terminology used herein is for describing embodiments and is not intended to limit or restrict the invention. When describing a component as being "connected," "combined," or "joined" with another component, this includes not only direct connections but also indirect connections where other components are present in between. Furthermore, terms such as "comprising," "including," or "having" indicate the presence of features, numbers, steps, operations, components, or combinations thereof described in the specification, and do not exclude the existence or additional possibility of one or more other features, numbers, steps, operations, components, or combinations thereof. Additionally, terms such as "first" and "second," which may be used herein, are used only to distinguish one component from another and, unless specifically stated otherwise, do not limit the order or importance of the components. Therefore, a first component in one embodiment may be referred to as a second component in another embodiment, and similarly, a second component in one embodiment may be referred to as a first component in another embodiment.

[0058] The embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. It should be understood that those skilled in the art can make various modifications or equivalent substitutions to the following embodiments without departing from the spirit and essence of the present invention, and such modifications or substitutions should all fall within the protection scope of the present invention.

[0059] Figure 1 This is a flowchart illustrating the overall workflow of the SAM-CD transmission line construction disturbance change detection method based on an embodiment of the present invention. Figure 1 As shown, the core idea of ​​the method of the present invention is to construct a collaborative processing flow that combines the noise suppression capability of the time-series filter with the fine variation segmentation capability of the SAM-CD model.

[0060] The SAM-CD transmission line construction disturbance change detection method of the present invention, which integrates time-series filtering, has an overall processing flow comprising four main stages: "data input and preprocessing → time-series filtering → SAM-CD change detection → post-processing and early warning output". The data flow relationship between each module is clear, and the input and output formats are clearly defined, ensuring the integrity and scalability of the system operation.

[0061] (1) Data input and preprocessing

[0062] This stage is responsible for collecting and standardizing multi-source image data.

[0063] Input: Original image file, in JPG, PNG, or TIFF format;

[0064] Processing includes: image geometric correction, radiometric correction, registration, and region of interest (ROI) cropping;

[0065] • Output: Registered ROI image in TIFF format, with an image size of 512×512 or 1024×1024 pixels, used as input for subsequent time-series filtering modules.

[0066] (2) Timing filtering

[0067] This stage involves temporal smoothing and noise suppression of the ROI image sequence to construct a continuous temporal image stack.

[0068] Input: Registered ROI image sequence;

[0069] • Processing content: Kalman filtering or moving average filtering algorithms are used to smooth the reflectivity or brightness values ​​of time-series pixels and remove instantaneous noise;

[0070] Output: A stack of denoised time-series images, saved as an .h5 file, containing continuous image data from F(TN) to F(T).

[0071] The output will be directly used as the input to the SAM-CD change detection module, achieving an efficient connection between "denoising and detection".

[0072] (3) SAM-CD change detection

[0073] This stage utilizes the SAM model for deep feature extraction and fine segmentation of variable regions, which is the core processing step of this invention.

[0074] Input: Two temporal image pairs F(TK) and F(T), in .h5 format;

[0075] Processing content: Extracting depth features, calculating feature differences, generating change intensity maps and SAM cue information, and performing pixel-level segmentation;

[0076] Output:

[0077] o Preliminary change intensity map (PNG format);

[0078] oSAM hint file (JSON format, containing the coordinates of the bounding rectangle of the changing region);

[0079] o pixel-level variation mask (TIFF format, single channel).

[0080] (4) Post-processing and early warning output

[0081] This stage involves morphological optimization of the segmentation results, attribute extraction, and early warning information generation, and then synchronizing the results to the operation and maintenance platform.

[0082] Input: Original segmentation mask (TIFF format);

[0083] Processing content: Perform opening and closing operations and edge smoothing, extract the area, location and type of the changed region, and generate warning level;

[0084] Output:

[0085] o Optimized mask image (TIFF format);

[0086] o Visualized image with change annotations (PNG format);

[0087] o Change Attribute Report (Excel format, including time, location, area, type, and warning level);

[0088] o Early warning push notifications (sent to the operation and maintenance terminal via API interface).

[0089] The data flow relationships between the steps of the method of this invention are as follows:

[0090] • The output of the time-series filtering module is directly input to the SAM-CD change detection module to ensure the continuity and data consistency of the "denoising-detection" processing link;

[0091] • The SAM-CD change detection results are converted into structured output by the post-processing and early warning module, and then imported into coordinates and visualized on the map through the GIS system interface;

[0092] • Maintenance personnel can directly locate the construction disturbance area through the map navigation function, realizing a closed loop of the entire process from detection to on-site response.

[0093] In summary, the process design of this invention realizes fully automated data flow and decision support from raw image acquisition, dynamic noise reduction, depth change detection, and intelligent early warning push, providing an efficient and reliable technical path for the safe operation and maintenance of power transmission lines.

[0094] Figure 2 A block diagram of a transmission line construction disturbance change detection device according to an embodiment of the present invention is shown.

[0095] like Figure 2As shown, the transmission line construction disturbance change detection device of the present invention includes a data acquisition module, a time-series filtering module, a change detection module (SAM-CD), and an early warning output module. Its workflow is as follows: First, the input multi-temporal images are preprocessed and filtered to remove noise; then, keyframe features are extracted using the SAM model, and initial screening of change areas is performed; finally, fine segmentation and result output are completed by combining preceding and following temporal information, thereby achieving high-precision detection and intelligent early warning of transmission line construction disturbances.

[0096] Data acquisition module: performs data input and preprocessing.

[0097] The data acquisition module is used to acquire multi-source time-series image data of the transmission line corridor and surrounding areas, which serves as input for subsequent change detection and analysis.

[0098] The input sources include:

[0099] (1) Continuous image data collected by real-time monitoring equipment such as visible light cameras and infrared thermal imagers deployed on transmission line towers or monitoring points along the line;

[0100] (2) High-resolution image data obtained through regular inspections by drones;

[0101] (3) Satellite remote sensing images serve as a supplementary data source for macroscopic detection and change verification in large-scale construction areas.

[0102] During the data preprocessing stage, the acquired raw images undergo geometric correction, radiometric correction, image registration, and region of interest (ROI) cropping operations sequentially to ensure that images from different time phases have uniform spatial reference coordinates and coverage. Subsequently, the images are uniformly cropped to a fixed size (e.g., 512×512 pixels or 1024×1024 pixels) to adapt to the input requirements of the SAM-CD model and to provide a standardized data foundation for subsequent temporal filtering and feature extraction.

[0103] Timing filtering module: performs timing filtering processing.

[0104] The temporal filtering module is used to smooth the input multi-temporal image sequence in the temporal domain and suppress noise, so as to reduce the error caused by illumination changes, cloud shadow interference and instantaneous occlusion, thereby extracting the real and stable trend of surface change.

[0105] First, such as Figure 1 As shown, preprocessed continuous multi-temporal images (e.g., TN, T-N+1, ..., T-1, T times) are constructed into a temporal image stack in chronological order, and a corresponding temporal reflectance or brightness value sequence is established for each pixel. In this way, it is possible to model the changes of ground features in the time dimension, providing reliable temporal feature input for subsequent change detection.

[0106] During the filtering stage, for each pixel's temporal sequence, a Kalman filter or moving average filter is used for smoothing to suppress instantaneous noise and preserve stable trends. The filtered output is a set of denoised temporal image sequences F(TN), F(T-N+1), ..., F(T-1), F(T).

[0107] This step can effectively reduce noise interference caused by sudden changes in illumination, dynamic changes in cloud shadows, camera shake, or short-term occlusion, thereby improving the feature extraction accuracy and recognition stability of the subsequent change detection module.

[0108] To adapt to the construction monitoring scenario of power transmission lines, the applicability of different filtering algorithms in terms of noise suppression effect and computational complexity is compared as shown in Table 1 below:

[0109] Filtering Algorithm Applicable Scenarios Noise suppression effect (transmission line scenario) computational complexity Kalman filtering Sudden changes in illumination and dynamic interference from cloud shadows It can handle non-stationary noise and improves PSNR by 8.4dB. middle Moving average filtering Stable lighting, no rapid shading When dealing with only stationary noise, the PSNR is improved by 5.2 dB. Low Median filtering Salt and pepper noise (data transmission errors) Ineffective against cloud shadow / lighting interference Low

[0110] Table 1: Comparison of Filtering Algorithm Adaptability

[0111] Regarding parameter settings, the filter window length is set to 5 frames, corresponding to approximately 5 hours of time-series data. This parameter setting is based on the fact that the typical duration of transmission line construction disturbances (such as excavation and mechanical operations) is usually 2 to 4 hours. Therefore, a 5-frame window can completely cover the disturbance cycle while avoiding detection delays caused by an excessively large window. When the window length exceeds 10 frames, the average detection delay will exceed 10 hours, which cannot meet the application requirements for real-time early warning.

[0112] In summary, this temporal filtering module can significantly improve the temporal consistency and signal-to-noise ratio of image data, providing a stable input basis for subsequent SAM-CD change detection.

[0113] Figure 3 This is a schematic diagram illustrating the effect of temporal filtering according to an embodiment of the present invention. The diagram shows the processing flow and effect comparison of the temporal filtering module from the original input image to the output denoised image.

[0114] like Figure 3 As shown in the diagram, the temporal filtering effect of this invention vividly illustrates the working principle and technical value of the temporal filtering module. This flowchart is used to intuitively explain how the temporal filtering module effectively suppresses noise from the original temporal image and outputs high-quality, smooth data for use by the subsequent change detection module.

[0115] Specifically, raw remote sensing time-series image data typically contains high-frequency noise such as clouds, shadows, and changes in illumination. This noise can generate spurious change signals in the time series, severely affecting the accuracy of subsequent change detection. Traditional methods cannot effectively distinguish between real surface changes and illumination disturbances, resulting in a high false alarm rate in detection results.

[0116] To address the aforementioned issues, this invention introduces a temporal filtering module before change detection. This module utilizes smoothing methods in the time or frequency domain to effectively remove high-frequency noise while retaining low-frequency, authentic surface change signals. As the core of the preprocessing in this invention, this module significantly improves the signal-to-noise ratio of the input image, ensuring that subsequent feature extraction processes are more focused on the truly changing areas.

[0117] The filtered time-series image data exhibits a significantly improved signal-to-noise ratio (SNR), with noise signals suppressed and surface change trends smoothly preserved. This result enables subsequent change detection algorithms (such as the SAM-CD module) to operate under higher-quality input conditions, thereby significantly reducing the false positive rate and effectively avoiding misclassification of short-term disturbances such as shadows and clouds as construction disturbances.

[0118] Filtering algorithm selection

[0119] This invention recommends using the Kalman filter as the primary time-series filtering algorithm. This algorithm is an optimal estimation algorithm suitable for linear systems. Its core idea is to use a recursive "prediction-update" process to optimally estimate the true state of each pixel based on the system model and observation data, thereby effectively filtering out observation noise while preserving the true trend of change.

[0120] Its state-space model can be represented as:

[0121] State equation: X(k) = F × X(k-1) + W(k)

[0122] State equation: Z(k) = H × X(k) + V(k)

[0123] Where X(k) is the system state at time k (i.e., the true value of the pixel), Z(k) is the observed value at time k (i.e., the image pixel value), F is the state transition matrix, H is the observation matrix, and W(k) and V(k) are the process noise and observation noise, respectively, both assumed to be white noise following a Gaussian distribution. Using this model, Kalman filtering can achieve dynamic estimation and smooth output of the time series of each pixel.

[0124] Application level and processing method

[0125] Temporal filtering is performed independently at the pixel level. For each location (i,j) in the image, the algorithm reads the observation sequence Z(i,j,1), Z(i,j,2), ..., Z(i,j,T) for that location across all temporal phases, and smooths it using a Kalman filter to obtain the state estimation sequence X̂(i,j,1), X̂(i,j,2), ..., X̂(i,j,T) for that pixel. The filtering results for all pixels are then combined to form a complete "filtered temporal image dataset," enabling denoising and trend preservation across the entire monitoring area.

[0126] Connection with subsequent processes

[0127] The denoised temporal image data output from the filtering module is directly used as input to the subsequent SAM-CD change detection module for keyframe comparison and cue generation. Because the filtering process significantly improves the image's signal-to-noise ratio and temporal stability, the subsequent change detection model can more accurately identify real disturbance areas, avoiding misjudgments caused by unstructured noise such as illumination and cloud shadows. High-quality input data provides a solid foundation for the high-precision segmentation of the SAM model, ensuring the accuracy and robustness of the entire change detection system.

[0128] By introducing and optimizing the aforementioned time-series filtering module, this invention achieves noise suppression, trend smoothing, and signal enhancement in the data preprocessing stage, providing a stable and reliable input environment for subsequent SAM-CD change detection, thereby effectively improving the accuracy and practicality of transmission line construction disturbance monitoring.

[0129] Change Detection Module (SAM-CD): Performs SAM-CD change detection.

[0130] The change detection module is the core component of this invention, used to identify, locate, and segment construction disturbance areas around transmission lines in temporally filtered images. This module combines temporal difference calculation with the general segmentation capabilities of the SAM (Segment Anything Model) model to achieve high-precision change detection from the feature layer to the pixel layer.

[0131] (1) Keyframe selection

[0132] like Figure 1 As shown, from the denoised time-series image sequence output by the time-series filtering module, the filtered image F(T) at the current time T and the filtered image F(TK) at the historical reference time TK are selected as the input image pair, where K represents the time interval frame number.

[0133] In this embodiment, K is set to 30, corresponding to a time interval of approximately one month, to ensure that the baseline image does not contain construction disturbance information, thereby establishing a stable comparative reference. For rainy or frequently changing regions, the K value can be appropriately shortened according to climate and environmental characteristics, for example, set to 15, to reduce the impact of weather changes on the results.

[0134] (2) Feature extraction and initial screening of change regions

[0135] Deep feature extraction: Input F(TK) and F(T) into the image encoder (based on ViT-B / 16 backbone network) of the SAM model respectively to extract high-dimensional deep feature vectors of size 640×640×768, denoted as Feat(TK) and Feat(T) respectively.

[0136] Feature difference calculation: Calculate the cosine distance between Feat(TK) and Feat(T) to generate a change map to reflect the degree of difference between temporal images in the feature space.

[0137] Threshold segmentation: The "OTSU adaptive threshold + empirical correction" strategy is used to binarize the change intensity map. First, the OTSU algorithm automatically distinguishes between changing and non-changing areas. Then, for the vegetation background commonly found in transmission line scenarios, the threshold is lowered by 10% to 15% to avoid misjudgment caused by wind-induced swaying or changes in lighting, thus obtaining a preliminary binary mask for areas with significant changes.

[0138] Connectivity analysis: Connectivity is marked on the binarized results, and the bounding box of each changing region is extracted as a prompt for subsequent SAM segmentation, thus achieving preliminary spatial localization of the changing regions.

[0139] (3) Suggest optimization strategies

[0140] Since the SAM model is highly sensitive to the accuracy of the input prompt bounding box, this invention performs boundary dilation optimization based on the generated bounding box to improve the completeness of the segmentation results.

[0141] In a 1024×1024 pixel image, the expansion is 2 pixels; in a 512×512 pixel image, the expansion is 1 pixel.

[0142] The basis for this parameter setting is that when the expansion range is controlled within 5% of the area, it will not introduce unnecessary background information, but can effectively avoid the problem of missed detection caused by edge disturbances (such as displacement of iron tower angle steel, edge of excavation area, etc.), thereby improving the integrity and robustness of edge change detection.

[0143] (4) Fine segmentation

[0144] The optimized prompt information is input into the prompt encoder and mask decoder of the SAM model. Combined with the original high-resolution image, the initially selected change areas are finely segmented to finally generate a pixel-level precision binary change mask.

[0145] This mask can accurately distinguish newly added construction areas (such as newly piled earthworks and areas intruded by construction machinery), disappeared features (such as areas where vegetation has been cleared), and unchanged areas, achieving high-precision identification and regional positioning of construction disturbances, and providing reliable visual inspection results for the operation and maintenance of power transmission lines.

[0146] Figure 4 A schematic diagram of the SAM-CD change detection process according to an embodiment of the present invention is shown.

[0147] like Figure 4 As shown in the diagram, the SAM-CD change detection process of this invention illustrates the complete data processing logic and decision-making flow of this module. The diagram clearly and structurally reflects the entire process from input keyframes to outputting the final change detection result.

[0148] This diagram aims to visually illustrate how the SAM-CD module, based on filtered keyframe images, achieves automatic detection, accurate segmentation, and result optimization of changing regions. Specifically:

[0149] (a) Input keyframe pair: The input images are two keyframe images from the temporal filtering module, namely the reference image F(TK) (usually a month ago in a state without construction) and the current image F(T) (where construction activities exist, such as excavator operation).

[0150] (b) Change Intensity Map Generation: Calculate the cosine distance between the features of the two images to generate a change intensity map. In this map, the pixel value of the construction area (such as the excavator or excavated soil) is usually ≥0.7 (high change area, marked in red), and the pixel value of the non-change area is ≤0.3 (low change area, marked in blue).

[0151] (c) Automatic prompt generation: The system automatically generates prompt information (Prompt) based on the intensity change map, which is the bounding box of the change area. In actual scenarios, the bounding box coordinates of the excavator area are (320, 450, 480, 620) (x1, y1, x2, y2), which can completely cover the construction area after expansion.

[0152] (d) SAM Fine Mask Output: The current image F(T) and automatically generated prompts are input into the SAM model's prompt encoder and mask decoder to obtain a pixel-level precision binary transformation mask. Construction disturbance targets (such as excavator cab, bucket, and excavation area edges) in the mask are clearly segmented, and the segmentation boundary error is controlled within 1 pixel.

[0153] (e) Post-processing results: Morphological post-processing is performed on the raw mask output by SAM, including opening and closing operations. Opening operations are used to remove noise points with an area smaller than 5 pixels (such as birds, leaves, etc.), and closing operations are used to fill holes in areas of variation (such as soil mound shadows). For example, for an actual excavation area of ​​12㎡, the mask calculation area is 11.8㎡, with an area error of less than 2%.

[0154] Logical Flow Stage Description

[0155] Data input and feature extraction

[0156] The input stage receives registered keyframe image pairs F(TK) and F(T) from the temporal filtering module. Using an image encoder based on the SAM model, the input images are converted into high-dimensional deep features containing semantic information for subsequent change analysis.

[0157] Coarse location of change area and decision judgment

[0158] The system calculates the difference in features between two images (such as cosine distance) to generate a change intensity map. Each pixel value in the change intensity map represents the probability of a change at that location. By analyzing the change intensity map, the system determines whether there are any regions of significant change: if there are no significant changes, it directly outputs a "no change" result; if there are significant changes, it proceeds to the subsequent fine segmentation stage. This decision node can significantly save computational resources and is suitable for real-time deployment at the edge.

[0159] Hint generation and fine segmentation

[0160] For areas of significant change, the system automatically generates a prompt, providing the location of the changed target in the form of a bounding box. Subsequently, this prompt, along with the current image F(T), is input into the prompt encoder and mask decoder of the SAM. Utilizing the zero-sample segmentation capability of the SAM model, pixel-level fine segmentation of the construction disturbance area is achieved, resulting in a high-confidence binary mask result.

[0161] Results Optimization and Output

[0162] Morphological post-processing is performed on the mask output by SAM, including opening / closing operations and edge smoothing, to remove minor noise and complete the target boundaries. The final output includes:

[0163] Image map with marked areas of change;

[0164] Binary transformation mask file;

[0165] Reports that include attributes of the changing area (such as area, location, type).

[0166] The results can be directly used for automated decision-making and early warning of construction disturbances in the operation and maintenance of transmission lines.

[0167] Core logical advantages

[0168] (1) Intelligent decision-making closed loop

[0169] The SAM-CD process introduces a significant change detection node, which dynamically determines whether to initiate computationally intensive segmentation processes based on the change intensity map. If no significant change is detected, the system automatically outputs a "no change" result, thereby saving computing resources and improving response speed, which is especially suitable for resource-constrained edge device deployment scenarios.

[0170] (2) The unity of automation and high precision

[0171] This invention organically combines coarse localization of change regions with fine segmentation using SAM (Signal Analysis and Analysis), achieving a fully automated processing flow from change detection to change recognition. By automatically generating prompts and combining them with the powerful segmentation capabilities of SAM, high-precision disturbance detection and target recognition can be achieved without manual intervention, ensuring the accuracy and consistency of detection results.

[0172] In summary, the SAM-CD module, as the core intelligent detection unit of this invention, achieves high-precision, low-false-report automatic detection of construction disturbances in transmission lines through the coordinated implementation of feature layer change analysis, automatic prompt generation, and refined segmentation. It is a key component in the overall process of this invention.

[0173] Early warning output module: Post-processing and result output

[0174] The early warning output module is used to post-process the fine segmentation mask output by the SAM-CD change detection module, extract change information, and generate intelligent early warnings, so as to realize the automatic identification, classification, and alarm of construction disturbances of transmission lines.

[0175] (1) Post-processing

[0176] Morphological and smoothing operations are performed on the fine segmentation mask Mask_fine to further remove noise, fill holes, and optimize boundaries.

[0177] The specific steps are as follows:

[0178] Opening operation: A morphological opening operation is performed using a 3×3 rectangular structuring element to remove isolated noise points with an area of ​​less than 5 pixels. The actual ground area corresponding to this operation is less than 0.1㎡, which can effectively eliminate transient interference targets such as birds and fallen leaves.

[0179] Closing operation: Morphological closing operations are performed using 5×5 rectangular structuring elements to fill holes in the mask. For example, for local gaps in the excavated area caused by stone shadows or reflections, the closing operation can restore the complete area boundary and avoid omissions in area statistics.

[0180] Edge smoothing: Gaussian filtering (σ=1) is used to smooth the mask edges to reduce jagged edges and improve boundary positioning accuracy, thereby improving the geometric consistency and visualization effect of the final change area.

[0181] (2) Extraction of change information

[0182] After post-processing, key attribute information of construction disturbances, including location, area, and type, is extracted from the change mask.

[0183] Location attributes: Based on image georegistration information, calculate the latitude and longitude coordinates of the center point of the changed area for spatial positioning and geographic visualization.

[0184] Area attribute: Count the number of pixels with a value of 255 in the mask and convert it to the actual ground area based on the image spatial resolution. For example, in an image with a resolution of 1024×1024 pixels and each pixel corresponding to 0.1×0.1㎡, the formula for calculating the area is:

[0185] Area = Number of pixels × 0.01㎡

[0186] Type attribute: Automatically identify the type of disturbance based on mask shape and texture features.

[0187] • Long, narrow areas typically correspond to mechanical intrusion or basket movement;

[0188] Irregular polygonal areas are typically excavation areas;

[0189] • Soil texture areas were identified as foundation pit excavation areas;

[0190] • Areas with metallic reflective characteristics are identified as mechanical equipment or vehicles.

[0191] By analyzing the above combination of features, typical disturbance types such as foundation pit excavation, mechanical intrusion, tower tilting, and material stacking can be automatically classified and identified.

[0192] (3) Early warning generation

[0193] Different levels of early warning events are triggered based on changed attribute information and preset rules. The specific rules are as follows:

[0194] Area threshold rule: When the area of ​​the change is greater than 50㎡ (corresponding to small excavation), or although it is less than 10㎡, it is located within 10 meters of the tower foundation (such as when the risk of excavation around the foundation is high), a level one warning is triggered.

[0195] Type threshold rule: When high-risk disturbance types such as mechanical collisions and tower tilting are detected, a level 1 warning is triggered directly without area conditions.

[0196] Frequency threshold rule: When three or more change events occur consecutively in the same area within 24 hours, a level 2 warning is triggered to indicate continuous construction or repeated intrusion behavior.

[0197] The output formats of the early warning results include:

[0198] Image annotation map: Overlaying the bounding boxes of the changed areas, disturbance type labels, and latitude and longitude information onto the original image;

[0199] Change attribute report: Exported in Excel format, including time, location, area, type and warning level;

[0200] Push notification: Warning information and annotation map links are pushed to the responsible maintenance personnel in real time via the maintenance mobile terminal APP or SMS, enabling remote alarm response.

[0201] Figure 5 This is a logic diagram illustrating the detection logic for changes in construction disturbances of transmission lines according to an embodiment of the present invention.

[0202] like Figure 5 The diagram shown is a logic diagram for detecting changes in construction disturbances in transmission lines according to an embodiment of the present invention. This diagram illustrates the core technical logic of the method in the overall process and the dependencies between various modules.

[0203] Detection of changes caused by construction disturbances in power transmission lines is a comprehensive technical process that integrates remote sensing technology, image processing, and deep learning algorithms. Its core objective is to automatically, accurately, and efficiently identify surface changes caused by construction activities from multi-source image data and generate early warning information in real time to ensure the safe operation of the power grid.

[0204] This logic diagram summarizes the main stages and key technical aspects of the detection process. The overall logic is structured in a closed loop of "noise suppression - change detection - fine segmentation - early warning decision-making", forming a complete disturbance identification and response system.

[0205] (I) Overall System Logic

[0206] The system consists of four stages: data input, change detection and disturbance identification, post-processing and optimization, and result output and application.

[0207] Data Input and Preprocessing

[0208] The system first receives multi-source input image data (including UAV aerial images, visible light camera images, satellite remote sensing images, etc.).

[0209] Image registration, radiometric correction, and denoising are performed to ensure that the input data is consistent in spatial location and lighting conditions, providing a stable input for subsequent temporal filtering and change detection.

[0210] Change detection and disturbance identification

[0211] In this stage, the system comprehensively utilizes multiple detection methods to collaboratively achieve change identification:

[0212] Deep learning-based target recognition: using models such as YOLO to identify dynamic targets such as construction machinery, vehicles and personnel;

[0213] Temporal change detection and analysis: Differential analysis is performed on multi-temporal image sequences to extract potential change regions;

[0214] Binocular or parallax calculation: Estimate the height difference and deformation information of the construction area using binocular images or depth maps;

[0215] Moving target detection: Frame difference method, background modeling or Hough transform is used to identify continuously moving construction equipment and personnel activities.

[0216] Post-processing and optimization

[0217] Morphological operations, contour optimization, and false alarm removal are performed on the change detection results to eliminate minor noise and false targets, ensuring the integrity and accuracy of the output results.

[0218] Results Output and Applications

[0219] The output includes:

[0220] Precisely variable mask: pixel-level boundaries including areas of construction disturbance;

[0221] Disturbance area attributes: including area, location, category, and time information;

[0222] Early warning information: The system automatically generates early warning reports and pushes them to maintenance departments based on the level of disturbance. These reports are used by power dispatch and inspection departments.

[0223] (II) Core Logic Judgment Nodes

[0224] like Figure 5As shown, the process of this invention includes three key decision-making nodes to construct an intelligent decision-making closed loop:

[0225] (1) Registration accuracy judgment node: After temporal filtering, the system automatically calculates the registration error. When the error exceeds 0.5 pixels, the image registration is automatically re-executed to ensure spatial alignment accuracy.

[0226] (2) Change intensity judgment node: After generating the change intensity map, if the maximum pixel value is less than 0.3, the system determines that the current image has no significant change, directly outputs the "no disturbance" result, and skips the subsequent segmentation steps to save computing resources.

[0227] (3) Sensitive Area Judgment Node: After identifying the changing area, the system automatically determines whether the disturbed area is located within 10 meters of the transmission tower foundation. If it is located in a sensitive area, the system automatically raises the warning level and generates a Level 1 alarm.

[0228] Through the above triple judgment logic, this invention achieves dynamic adaptive control of change detection accuracy, algorithm resources and early warning level, forming a stable and efficient automated detection mechanism.

[0229] (III) Technical Linkages and Dependencies

[0230] There are strict data and functional dependencies between the various technical components:

[0231] Image registration provides a spatial consistency basis for temporal filtering;

[0232] • Temporal filtering provides high-quality, denoised image input for SAM-CD change detection;

[0233] The SAM model's prompt generation module provides the target location range for the fine segmentation stage;

[0234] The post-processing module further optimizes the masking results, providing accurate basis for extracting change attributes and issuing early warnings.

[0235] Through the above-mentioned technological integration, this invention establishes a closed-loop detection system from data acquisition to intelligent early warning, ensuring that the entire process is logically rigorous, data transmission is smooth, and processing results are reliable.

[0236] The transmission line construction disturbance change detection system of the present invention, when actually deployed, can include two parts: hardware equipment configuration and software operating environment.

[0237] Hardware deployment

[0238] High-definition network cameras are installed on key towers or monitoring points along transmission lines to ensure their field of view covers the monitoring corridor area. The cameras are equipped with lightning-proof, rainproof, and dustproof housings to ensure the stable operation of power supply and communication lines. Simultaneously, drones can be used to conduct regular inspections to acquire supplementary image data and expand the coverage area.

[0239] Software environment

[0240] The algorithm system of this invention can be deployed on a cloud server or a local high-performance computer. The runtime environment can include components such as the PyTorch deep learning framework, the OpenCV image processing library, and the NumPy scientific computing library to support core operations such as model inference, image registration, and feature calculation.

[0241] Data Acquisition and Time Series Construction

[0242] The system can set the acquisition frequency according to task requirements (e.g., capturing one frame per hour) and accurately register images from different time periods based on feature point matching algorithms (such as SIFT, ORB, or template matching). Subsequently, ROI cropping is performed according to predefined transmission line corridor vector boundaries to reduce unnecessary computational areas. Finally, multiple consecutive image periods are stored sequentially to construct a time-series dataset, providing input for subsequent filtering and change detection.

[0243] System Experiment and Performance Verification

[0244] To verify the effectiveness of the method of the present invention, experimental evaluations were conducted from three dimensions: noise suppression, detection accuracy, and system performance.

[0245] 1. Performance verification of the timing filtering module

[0246] In a real-world transmission line monitoring scenario, the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) were used to evaluate the filtering effect. The experimental results are shown in the table below:

[0247] Noise type PSNR (dB) before noise reduction PSNR (dB) after noise reduction SSIM before noise reduction SSIM after noise reduction Remark Gaussian noise (σ=15) 28.5 36.2 0.78 0.94 Simulated sensor noise Salt and pepper noise (density = 0.05) 24.1 33.8 0.72 0.91 Simulated data transmission errors Cloud shadows 19.8 30.1 0.65 0.88 Simulated weather changes

[0248] Experimental conditions: A total of 1000 sets of continuous time-series images (512×512 pixels) were tested, and the filter window was set to 5 frames.

[0249] The results show that time-series filtering can improve PSNR by about 8.4 dB and SSIM by 0.23, effectively suppressing instantaneous noise such as illumination and cloud shadows, and providing stable input for change detection.

[0250] SAM-CD change detection performance verification

[0251] The method of this invention is compared with traditional change detection algorithms on a self-built transmission line monitoring dataset.

[0252] (1) Comparison of quantitative indicators (test set: 500 pairs of images)

[0253] method Accuracy Precision Recall F1-Score Intersection over Union (IoU) Traditional image difference method 0.75 0.65 0.70 0.67 0.51 CNN-based change detection 0.88 0.82 0.85 0.83 0.71 This method (time-series filtering + SAM-CD) 0.96 0.94 0.93 0.94 0.88

[0254] (2) Detection effect of different construction disturbance types

[0255] Construction disturbance type This method F1-Score CNN-based F1 Score Improved results Mechanical vehicle intrusion 0.96 0.85 +12.9% earthwork excavation 0.93 0.82 +13.4% New building (obstruction) 0.92 0.79 +16.5% Vegetation removal 0.89 0.78 +14.1%

[0256] Experimental results show that the method of the present invention is significantly better than the traditional method in terms of accuracy and IoU, and can effectively suppress spurious changes and generate mask results with accurate boundaries.

[0257] System comprehensive performance test

[0258] (1) Processing efficiency

[0259] Testing was conducted on a server equipped with an NVIDIA Tesla V100 processor. The processing time for a single frame of 1024×1024 video was as follows:

[0260] • Timing filtering (5 frames): approximately 0.5 seconds

[0261] Feature extraction and encoding: approximately 1.8 seconds

[0262] • Prompt generation and segmentation: approximately 0.7 seconds

[0263] • Total processing time: Approximately 3.0 seconds per frame

[0264] This performance meets the real-time accuracy requirements for power transmission line monitoring scenarios.

[0265] (2) Anti-interference performance

[0266] Under simulated conditions of sudden changes in illumination and slight camera shake, the system's detection results remained stable. Compared to the unfiltered baseline method, the false alarm rate was reduced by approximately 60%, and the F1-Score decreased by less than 2%.

[0267] Based on a thorough analysis of the shortcomings of existing transmission line construction disturbance monitoring technologies, this invention proposes a method that integrates time-series filtering technology with SAM-CD (Simultaneous Amplification and Modulation) model-based change detection. This method features innovative designs in system architecture, algorithm fusion, model adaptation, and real-time deployment. The main innovations are as follows:

[0268] (1) Deep fusion of temporal filtering and SAM model

[0269] This invention does not simply superimpose temporal filtering with the SAM model, but constructs the temporal filtering module as a "preprocessing enhancement layer" of the SAM model, performing dynamic denoising and smoothing on the input image before feature extraction, thereby improving the purity and stability of the input data from the source.

[0270] Through this fusion approach, the SAM model can more accurately focus on real disturbance areas during feature extraction, reducing the number of false change areas by more than 60%. In addition, temporal filtering is used to automatically determine the reference time, avoiding the problem of detection error accumulation caused by random selection of reference frames in traditional methods, thereby significantly improving the temporal consistency and reliability of change detection.

[0271] (2) Customized optimization of transmission line scenarios

[0272] In view of the complex environment and strong background interference of transmission line corridors, this invention optimizes the prompt generation and threshold setting of the SAM model in a scenario-based manner.

[0273] For typical backgrounds such as vegetation and towers, a "boundary box expansion + threshold correction" strategy is proposed to reduce false detections caused by background texture fluctuations and improve the edge extraction accuracy of specific targets (such as tower angle steel and construction scaffolding), thereby improving the segmentation accuracy by about 25%.

[0274] It supports multi-source image input, including visible light, infrared, UAV and satellite remote sensing. Through a unified feature extraction and temporal processing workflow, it achieves "one model adapting to multiple scenarios", effectively overcoming the dependence of traditional methods on a single data source.

[0275] (3) Efficient and accurate prompt generation strategy

[0276] This invention proposes an automatic prompt generation mechanism based on the differences in images after temporal filtering, which can automatically generate high-quality bounding boxes of changing regions as input prompt information for the SAM model.

[0277] This strategy enables the SAM model to quickly focus on the real change area, significantly reducing computational redundancy and performance loss caused by misleading prompts. It not only avoids the complexity and subjectivity brought about by manual interaction, but also achieves the unification of automated prompt generation and high-precision detection.

[0278] (4) Design for balancing automation and real-time performance

[0279] This invention achieves fully automated processing, from image input, filtering and noise reduction, change detection and early warning information push, all without manual intervention. Maintenance personnel only need to view the early warning results generated by the system, improving overall work efficiency by about 80%.

[0280] At the system deployment level, through the collaborative computing architecture of "edge + cloud", the time-series filtering and feature difference calculation tasks with smaller computational load are deployed on the edge device (single frame time ≤ 0.5 seconds), while the SAM fine segmentation task with larger computational load is deployed on the cloud (time ≤ 1.8 seconds).

[0281] This architecture not only takes into account the requirements of real-time and high-precision detection, but also effectively reduces hardware costs. Edge devices can run without the support of high-performance GPUs, reducing the overall system cost by about 50%.

[0282] The scope of the claims of this invention is not limited to the specific embodiments described above. Various other embodiments, including modifications or alterations that can be made by those skilled in the art without departing from the spirit and intent of the invention as described in the claims, should also be included within the scope of the claims of this invention.

Claims

1. A method for detecting changes in construction disturbances of transmission lines, characterized in that, Includes the following steps: S1. Data input and preprocessing: Acquire multi-source image data, and sequentially perform geometric correction, radiometric correction, image registration and region of interest (ROI) cropping on the images to obtain a registered ROI image sequence. S2. Temporal filtering: Construct a pixel-level temporal sequence based on the ROI image sequence, and perform temporal filtering on each pixel temporal sequence to obtain the denoised temporal image stack F(TN)…F(T); S3. Change Candidate Generation: Select the current image F(T) and the historical reference image F(TK) as keyframe pairs, input them into the image encoder of the general segmentation model SAM to obtain depth features, calculate the difference between the two features to generate a change intensity map, perform threshold segmentation and connected component analysis on the change intensity map to obtain change candidate regions and generate corresponding bounding rectangle prompt information. S4. Prompt optimization and fine segmentation: The prompt information is optimized by boundary dilation. The optimized prompt and F(T) are input into the prompt encoder and mask decoder of SAM, and the pixel-level binary transformation mask is output. S5. Post-processing and attribute extraction: Perform morphological opening and closing operations and edge smoothing on the binary transformation mask, and extract attributes such as the position, area and type of the transformation area. S6. Warning generation and result output: According to preset rules, the changed areas are given graded warnings, and optimized masks, labeled images, attribute reports and warning information for push notifications are output.

2. The method according to claim 1, characterized in that, The time-domain filtering is a Kalman filter, which is based on the following state-space model: State equation: X(k) = F×X(k−1) + W(k); Observation equation: Z(k) = H×X(k) + V(k); Where X(k) is the system state at time k, Z(k) is the observation value at time k, F is the state transition matrix, H is the observation matrix, and W(k) and V(k) are the process noise and observation noise, respectively.

3. The method according to claim 1, characterized in that, The generation of the intensity change map includes: Input F(T−K) and F(T) into an image encoder based on a ViT-B / 16 backbone network to extract high-dimensional feature vectors Feat(T−K) and Feat(T); Calculate the cosine distance between two features and generate a change intensity map; The intensity change map is binarized using a strategy that combines OTSU adaptive thresholding with empirical correction, where the threshold is lowered by 10% to 15% to reduce the interference from vegetation swaying and light changes.

4. The method according to claim 1, characterized in that, The suggestion optimization includes expanding the boundary of the bounding rectangle of the variable candidate region: When the image size is 1024×1024 pixels, it expands by 2 pixels; When the image size is 512×512 pixels, it expands by 1 pixel; The expansion range shall not exceed 5% of the area of ​​the changed region.

5. The method according to claim 1, characterized in that, The rules for generating the early warning include: A Level 1 warning is triggered when the area of ​​the change is greater than 50㎡, or although it is less than 10㎡, it is located within 10 meters of the foundation of the transmission tower. When a mechanical collision or tower tilting disturbance is detected, a Level 1 warning is triggered directly. A Level 2 alert is triggered when three or more change events occur in the same area within 24 hours.

6. A device for detecting changes in construction disturbances of transmission lines, characterized in that, include: The system includes a data acquisition module, a time-series filtering module, a change detection module, and an early warning output module. The data acquisition module is used to acquire multi-source image data of the transmission line corridor and its surrounding area, and to perform geometric correction, radiometric correction, image registration and ROI cropping on the images in sequence to generate a registered ROI image sequence. The temporal filtering module is used to perform temporal filtering on the ROI image sequence to generate a denoised temporal image stack. The change detection module is used to select the reference time image and the current time image and input them into the image encoder of the general segmentation model SAM for feature extraction, and generate a change intensity map and circumscribed rectangle prompt information based on the feature differences; The early warning output module is used to post-process and extract attributes from the change detection results, and generate graded early warning results according to preset rules.

7. The apparatus according to claim 6, characterized in that, The temporal filtering module employs the Kalman filtering algorithm to achieve pixel-level temporal smoothing through prediction and update processes, thereby suppressing noise interference caused by sudden changes in illumination, cloud shadows, and short-term occlusion.

8. The apparatus according to claim 6, characterized in that, The change detection module includes an image encoder, a cue encoder, and a mask decoder; The image encoder is used to extract the depth features of the input image, the cue encoder is used to receive the bounding rectangle cue information of the change candidate region, and the mask decoder is used to output a pixel-level binary change mask.

9. The apparatus according to claim 6, characterized in that, The early warning output module includes: The post-processing unit is used to perform morphological opening and closing operations and Gaussian edge smoothing; The attribute extraction unit is used to extract the center coordinates, area, and disturbance type of the changed region. The early warning determination unit is used to generate early warning levels and output results based on area thresholds, type thresholds, and frequency thresholds.

10. The apparatus according to claim 6, characterized in that, The device is based on an "edge + cloud" collaborative computing architecture, wherein: Perform temporal filtering and feature difference calculation tasks at the edge. The cloud performs fine segmentation tasks using the general segmentation model SAM; real-time, high-precision construction disturbance detection and intelligent early warning are achieved through collaborative operation between the edge and the cloud.

Images