Power transmission line hidden danger detection method based on model lightweight

Abstract

The invention relates to the technical field of power transmission line intelligent monitoring, in particular to a power transmission line hidden danger detection method based on model lightweight, which comprises the following steps: acquiring an image and a sensing data stream of a power transmission line, performing mode deconstruction on the data stream according to a predefined rule, separating an abnormal mode and a background mode, and forming a heterogeneous data pair. And inputting the heterogeneous data pair into a lightweight detection network, performing feature extraction on abnormal data by adopting a dynamic convolution kernel through parallel branches of the lightweight detection network to generate an abnormal feature map, and performing environment modeling on background data by adopting a static convolution kernel to generate an environment reference map. And the fusion judgment branch calculates the relevance of the two maps and outputs a structured hidden danger report according to the relevance. According to the method, through data mode decoupling and heterogeneous parallel processing, while the detection precision is maintained, the model calculation complexity and resource consumption are reduced, and real-time and accurate hidden danger identification on the edge side with limited resources is realized.

Description

Technical Field

[0001] This invention relates to the field of intelligent monitoring technology for power transmission lines, and in particular to a model-based lightweight method for detecting potential hazards in power transmission lines. Background Technology

[0002] Current intelligent monitoring of power transmission lines relies on analyzing continuous data streams composed of image data and time-series sensor data. Existing technologies generally employ end-to-end deep learning models to process this type of multi-source monitoring data. These models take the raw data stream as a whole input, learn directly through complex network structures, and output hazard identification results. The goal of this method is to achieve automated detection of various hazards, and its effectiveness depends on the model possessing sufficiently powerful feature extraction and fusion capabilities.

[0003] This uniform processing approach has its flaws. Monitoring data streams are essentially a mixture of sudden abnormal signals and steady-state background information, and existing methods fail to effectively distinguish between these two components during the processing flow. Inputting both background and abnormal information into a single model forces the network parameters to simultaneously fit significantly different data patterns, increasing the complexity of model training and the difficulty of convergence. To achieve acceptable detection accuracy, the system often has to rely on models with deeper structures and larger parameter sets, resulting in high computational complexity and high resource consumption.

[0004] Existing solutions struggle to achieve efficient deployment and real-time analysis on edge monitoring terminals with limited computing power and storage resources. There is an urgent need for a detection method that optimizes data processing from the source and matches a lightweight model architecture, thereby reducing hardware computing power requirements while ensuring detection accuracy, and meeting the practical requirements of real-time performance and low power consumption in distributed monitoring scenarios for power transmission lines. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a model-based lightweight method for detecting potential hazards in power transmission lines.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a model-based lightweight method for detecting hidden dangers in transmission lines, comprising: Acquire raw monitoring data streams of power transmission lines that include image data and time-series sensor data; Based on the predefined hidden danger characterization rules, the original monitoring data stream of the transmission line is deconstructed to separate the abnormal mode data stream and the background mode data stream. The abnormal mode data stream and the background mode data stream constitute a heterogeneous data pair. The heterogeneous data pairs are input into a preloaded lightweight detection network, which includes parallel processing branches and fusion decision branches. In the parallel processing branch, a dynamic convolution kernel is used to extract features from the abnormal pattern data stream, while a static convolution kernel is used to model the environment of the background pattern data stream, generating an abnormal feature map and an environmental baseline map respectively. The abnormal feature map and the environmental baseline map are input into the fusion decision branch to perform cross-map correlation calculation; Based on the correlation calculation results, a structured hazard report containing the location and type of the hazard is generated.

[0007] As a further aspect of the present invention, the step of performing pattern deconstruction on the original monitoring data stream of the transmission line according to predefined hidden danger characterization rules specifically includes: For the image data in the original monitoring data stream of the transmission line, the edge saliency detection operator is invoked to extract high gradient change regions and low texture smooth regions in the image; For the time-series sensor data in the original monitoring data stream of the transmission line, a sliding window differential algorithm is applied to identify abrupt change points and periodic fluctuation segments in the data sequence. The high gradient change region and the mutation point are spatiotemporally aligned and logically bound to form an initial candidate set for the abnormal mode data stream; The low-texture smooth region and the periodic fluctuation segment are fused to form a stable reference set for the background pattern data stream; A confidence level filter is applied to the initial candidate set of the abnormal pattern data stream to remove pseudo-abnormal data with a confidence level lower than a preset threshold, thus forming the final abnormal pattern data stream.

[0008] As a further aspect of the present invention, before the step of applying confidence filtering to the initial candidate set of the abnormal pattern data stream, the method further includes: A multi-level confidence model is constructed to evaluate the pseudo-anomaly data. The multi-level confidence model integrates historical false alarm statistics and real-time environmental interference factors. From the initial candidate set of the abnormal mode data stream, the spatiotemporal attributes and physical quantity attributes are extracted item by item; The spatiotemporal attributes and physical quantity attributes of each item are input into the multi-level confidence model to calculate the individual confidence score; Based on all individual confidence scores, generate a confidence distribution histogram of the initial candidate set of the abnormal pattern data stream; Based on the confidence distribution histogram, the value of the preset threshold is dynamically adjusted.

[0009] As a further aspect of the present invention, in the step of inputting the heterogeneous data pair into the preloaded lightweight detection network, the loading process of the lightweight detection network includes: Download the basic network topology that matches the current transmission line tower model from the cloud model repository; Read the hardware performance configuration file deployed on the local edge device, which contains memory capacity and processor computing power information; Based on the hardware performance configuration file, an online pruning operation is performed on the basic network topology to remove redundant connection layers and activation layers; Using the network topology generated after pruning, the corresponding model weight parameter files are incrementally pulled from the cloud model repository; The model weight parameter file is loaded into memory to complete the instantiation and warm-up of the lightweight detection network.

[0010] As a further aspect of the present invention, the step of performing online pruning operation on the basic network topology specifically includes: Analyze the contribution weight of each layer of the basic network topology to the historical hidden danger detection task; Based on the contribution weight, the network layers are sorted, and the set of network layers whose contribution is lower than the pruning threshold is marked. Check the data dependencies between the network layers in the set of network layers to be removed and their adjacent network layers; For network layers with strong data dependencies, their structure is retained but their weights are reset to zero; for network layers without strong data dependencies, they are directly removed from the basic network topology. Traverse and update the data flow paths of the entire network to generate a simplified network topology description file.

[0011] As a further aspect of the present invention, the step of using dynamic convolutional kernels to extract features from the abnormal pattern data stream specifically includes: Based on the statistical characteristics of data segments in the abnormal mode data stream, the size parameters and stride parameters of the convolution kernel are calculated in real time. Using the calculated size and stride parameters, convolutional computation units are dynamically assembled in memory; The abnormal mode data stream is divided into multiple data blocks, which are then sequentially input into the dynamically assembled convolutional computation unit for convolutional operations. Collect the intermediate feature maps output after each data block undergoes convolution operations; All intermediate feature maps are subjected to multi-scale pooling and normalization processing, and then integrated to form the abnormal feature map.

[0012] As a further aspect of the present invention, the step of performing cross-graph correlation calculation specifically includes: The anomaly feature map and the environmental baseline map are aligned by gridding to ensure that the anomaly feature map and the environmental baseline map have a one-to-one correspondence of grid cells in spatial coordinates; For each aligned grid cell, feature vectors are extracted from the anomaly feature map and the environmental baseline map, respectively. Calculate the cosine similarity and Euclidean distance between the feature vectors from the anomaly feature map and the feature vectors from the environmental baseline map; The cosine similarity and the Euclidean distance are combined into a correlation strength index; Traverse all grid cells to generate a correlation intensity distribution map with the same spatial resolution as the original map.

[0013] As a further aspect of the present invention, after generating a correlation intensity distribution map with the same spatial resolution as the original map, the method further includes: The correlation strength distribution map is segmented by region growing to identify connected regions whose correlation strength exceeds the judgment threshold; Calculate the geometric center coordinates and the bounding rectangle of each connected region; Based on the geometric center coordinates, backtrack to the corresponding position in the original anomaly feature map and extract the original fragment of multimodal sensing data corresponding to the position; By combining the circumscribed rectangle boundary with the extracted original fragments of multimodal sensor data, a descriptive field for the location of the hazard and a classification label for the hazard type are generated in the structured hazard report.

[0014] As a further aspect of the present invention, the step of acquiring the original monitoring data stream of the transmission line containing image data and time-series sensing data specifically includes: Poll the data acquisition proxy service deployed at each monitoring point on the transmission line; When the data acquisition agent service returns to the ready state, a data retrieval command with a timestamp is sent to it; Receive the packaged data block returned by the data acquisition agent service according to the data retrieval instruction; The packaged data blocks are unpacked and verified to separate independent image data files and time-series sensor data files; The image data file and the time-series sensor data file are synchronized and spliced ​​according to a unified time reference to form the original monitoring data stream of the power transmission line.

[0015] As a further aspect of the present invention, after the step of unpacking and verifying the packaged data block, the method further includes: If the verification fails, the device identifier and failure time of the data acquisition agent service corresponding to the packaged data block are recorded. Send a data retransmission request to the data acquisition agent service and start the retransmission timer; If the correct data packet is received before the retransmission timer expires, the processing continues; If the retransmission timer times out and no correct data packet is received, the device identifier is marked as abnormal and temporarily removed from the current polling list. At the same time, the backup data acquisition agent service in the same area is activated to collect supplementary data.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Based on predefined hazard characterization rules, the raw monitoring data stream is deconstructed to separate abnormal pattern data streams from background pattern data streams. This step completes preliminary information filtering and focusing during the data preprocessing stage, transforming the unstructured raw data stream into structured heterogeneous data pairs. Subsequent detection networks no longer need to search for anomaly clues from massive amounts of redundant data; their processing objectives are more clearly defined, reducing the complexity of feature learning. This makes it possible to build lighter and more efficient dedicated detection networks, providing a direct technical path for deploying high-performance analysis models in field equipment with limited computing and storage resources.

[0017] A lightweight detection network incorporating parallel processing branches is constructed, employing differentiated processing strategies for the deconstructed heterogeneous data pairs. Dynamic convolutional kernels are used specifically to handle anomalous pattern data streams; their parameters can be dynamically adjusted based on local features of the input data, enhancing the network's adaptive capture capability for multi-scale and multi-morphological anomalous features. Simultaneously, static convolutional kernels are used to process the background pattern data stream, extracting environmental baseline features in a stable and efficient manner. This parallel heterogeneous processing architecture enables the network to maintain stable modeling of the background environment and high sensitivity to sudden anomalies, achieving effective decoupling and enhancement of anomalous and background information during the feature extraction stage. Finally, through cross-graph correlation calculations, potential hazards can be located and their types identified more accurately, improving the robustness and accuracy of hazard detection in complex natural scenes while significantly reducing the model's computational burden. Attached Figure Description

[0018] Figure 1 This is a flowchart of the model-based lightweight transmission line hazard detection method described in this invention; Figure 2 A flowchart for pattern deconstruction; Figure 3 A flowchart for loading a lightweight detection network; Figure 4 A bar chart for evaluating the correlation strength of grid cells in the detection of hidden dangers in power transmission lines; Figure 5 Performance evaluation diagrams for each level of a lightweight model for detecting hidden dangers in power transmission lines. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] See Figure 1 The overall implementation scheme of the model-based lightweight transmission line hazard detection method is as follows: This method acquires the original monitoring data stream of the transmission line, which includes image data and time-series sensor data. Based on predefined hazard characterization rules, the original monitoring data stream is deconstructed to separate abnormal pattern data streams and background pattern data streams, forming a heterogeneous data pair. This heterogeneous data pair is input into a pre-loaded lightweight detection network, which includes a parallel processing branch and a fusion decision branch. In the parallel processing branch, dynamic convolutional kernels are used to extract features from the abnormal pattern data stream, while static convolutional kernels are used to model the environment of the background pattern data stream, generating anomaly feature maps and environmental baseline maps, respectively. The anomaly feature maps and environmental baseline maps are input into the fusion decision branch to perform cross-map correlation calculations. Based on the correlation calculation results, a structured hazard report containing hazard location and hazard type is generated.

[0022] In one embodiment of the present invention, see [reference] Figure 2In practical implementation, for image data in the original monitoring data stream of transmission lines, an edge saliency detection operator is invoked to extract high gradient change regions and low-texture smooth regions in the image. In the example scenario, the transmission line monitoring image includes conductors, insulators, and the background environment. The edge saliency detection operator calculates the gradient magnitude of each pixel. Regions with gradient magnitudes exceeding a set upper limit are marked as high gradient change regions, such as abrupt edge changes caused by conductor surface corrosion or hanging foreign objects. Regions with gradient magnitudes below a set lower limit are marked as low-texture smooth regions, such as the sky or distant mountain background. Data comparison shows that the gradient magnitude distribution of high gradient change regions is concentrated in a higher range, while the gradient magnitude distribution of low-texture smooth regions is concentrated in a lower range. In some embodiments, the edge saliency detection operator employs multi-scale gradient calculation to adapt to images of different resolutions.

[0023] For time-series sensor data in the raw monitoring data stream of transmission lines, a sliding window differential algorithm is applied to identify abrupt changes and periodic fluctuation segments in the data sequence. In the example scenario, the time-series sensor data includes conductor temperature and vibration frequency. The sliding window differential algorithm sets a fixed time window and calculates the difference between adjacent data points within the window. Points with differences exceeding a dynamic threshold are marked as abrupt changes, such as a sudden temperature rise possibly caused by a short circuit. Segments with regularly changing differences are marked as periodic fluctuation segments, such as the periodic swaying of the conductor caused by wind load. Data comparison shows that the difference value sequence of abrupt changes exhibits spikes, while the difference value sequence of periodic fluctuation segments exhibits a sinusoidal pattern.

[0024] High-gradient change regions and abrupt change points are spatiotemporally aligned and logically bound to form the initial candidate set for the anomaly pattern data stream. Spatiotemporal alignment is based on the synchronization of image acquisition time and sensor timestamps, while logical binding uses Boolean operations to confirm that the spatial coordinates of the high-gradient change region and the timestamp of the abrupt change point are within the same event window. Low-texture smooth regions and periodic fluctuation segments are fused to form a stable reference set for the background pattern data stream. Data fusion uses a feature concatenation method, connecting the image texture feature vectors of the low-texture smooth regions with the sensor data feature vectors of the periodic fluctuation segments. In specific implementation, a multi-level confidence model is constructed to evaluate false anomaly data. This model integrates historical false alarm statistics and real-time environmental interference factors. Historical false alarm statistics are derived from the frequency distribution of false alarm events in historical detection logs, while real-time environmental interference factors include current wind speed and rainfall sensor readings. From the initial candidate set of the anomaly pattern data stream, spatiotemporal and physical attributes are extracted item by item. Spatiotemporal attributes include the GPS coordinates and timestamps of the events, while physical attributes include the mean image gradient and the peak value of the sensor data. The spatiotemporal attributes and physical quantity attributes of each item are input into the multi-level confidence model to calculate the individual confidence score. The multi-level confidence model uses the following formula: in: This indicates the confidence score for a single item. This indicates the statistical characteristics of historical false alarms. Indicates the real-time environmental interference factor. This is a balance coefficient with a value range of [0,1]. Optionally, the balance coefficient can be adjusted based on seasonal factors. Based on all individual confidence scores, a confidence distribution histogram of the initial candidate set for the anomaly pattern data stream is generated, with the score value on the horizontal axis and the frequency of candidate events on the vertical axis. Based on the confidence distribution histogram, the value of a preset threshold is dynamically adjusted. In specific implementations, the preset threshold is initially set to the median of the histogram scores. When the histogram shows a high density of candidate events in low-score areas, the preset threshold is increased. This dynamic adjustment process avoids insufficient or excessive filtering caused by a fixed threshold. After completing the above steps, confidence filtering is applied to the initial candidate set of the anomaly pattern data stream to remove pseudo-anomaly data with confidence scores below the preset threshold, forming the final anomaly pattern data stream. In some embodiments, the filtered candidate events are used for subsequent analysis. This confidence filtering improves the reliability of the anomaly pattern data stream.

[0025] In one embodiment of the present invention, see [reference] Figure 3 In the specific implementation, a basic network topology matching the current transmission line tower model is downloaded from the cloud model repository. In the example scenario, the transmission line tower model is "ZBC-220". The cloud model repository stores pre-trained network architecture configuration files for different tower models. Data comparison shows that the basic network topology corresponding to the "ZBC-220" model contains a specific sequence of convolutional layers, pooling layers, and fully connected layers. The hardware performance configuration file deployed on the local edge device is read. The hardware performance configuration file contains memory capacity and processor computing power information. In the example scenario, the local edge device has 4GB of memory and a processor computing power of 1 trillion floating-point operations per second. The hardware performance configuration file is stored in JSON format. Based on the hardware performance configuration file, an online pruning operation is performed on the basic network topology, removing redundant connection layers and activation layers. Data comparison shows that before the online pruning operation, the basic network topology contained 120 network layers, and after the online pruning operation, the number of network layers was reduced to 80 layers. The online pruning operation specifically includes analyzing the contribution weight of each layer of the basic network topology to historical hazard detection tasks. The contribution weight is calculated using historical task data, which includes hazard detection records for the same transmission line segment within the past year. In some embodiments, the contribution weight calculation formula is: in: The weights represent the contribution of the r-th layer of the network. This indicates the total number of samples from historical hazard detection tasks. This represents the output feature map of the r-th layer network when processing the q-th historical sample. Let represent the mean of the output feature maps of the r-th layer network on all historical samples. Represents the L2 norm. This is a very small constant used to prevent division by zero errors. Network layers are sorted according to their contribution weights, and a set of network layers with contributions below the pruning threshold is marked as to be removed. The pruning threshold is dynamically set based on the memory capacity in the hardware performance configuration file; in the example scenario, the pruning threshold is 0.05. The data dependencies between the network layers in the set to be removed and their adjacent layers are examined. These dependencies are determined by analyzing the inter-layer connection directions and dimension matching in the network data flow graph.

[0026] For network layers with strong data dependencies, their structure is preserved but their weights are reset to zero. Strong data dependencies are defined as strict consistency in feature map dimensions and skip connections between adjacent layers. For network layers without strong data dependencies, they are directly removed from the basic network topology. The data flow paths of the entire network are traversed and updated to generate a simplified network topology description file, which is represented as a directed graph. In specific implementations, the corresponding model weight parameter files are incrementally pulled from the cloud model repository using the pruned network topology. Incremental pulling only downloads weight parameters that match the simplified network topology. Data comparison shows that the full download of the model weight parameter file is 500MB, while the incremental pulling reduces the file size to 300MB. Optionally, differential compression technology is used in the incremental pulling process. The model weight parameter files are loaded into memory to complete the instantiation and warm-up of the lightweight detection network. The warm-up process includes running several simulation inferences to initialize the network cache. In some embodiments, the instantiated lightweight detection network occupies 2.5GB of memory. It is understandable that online pruning adaptively adjusts the network size based on hardware constraints. Optionally, the pruning threshold can be further calibrated based on processor computing power information. It is also understandable that incrementally pulling model weight parameter files reduces network transmission latency.

[0027] In practice, the calculation of contribution weights relies on historical hazard detection task data. The construction of historical hazard detection task data is a multi-source data fusion process. This process specifically serves the evaluation of network layer effectiveness during online pruning operations. Historical hazard detection task data is not simply a log of historical detection results, but a structured collection of task instances containing multi-dimensional information. Each task instance corresponds to a monitoring cycle of a transmission line that has been detected in the past, usually with one hour as a cycle unit. Each task instance encapsulates the original heterogeneous data pairs flowing into the lightweight detection network within that cycle, all intermediate feature map snapshots generated by the network intermediate layer when processing the data pairs, and the hazard labeling results finally confirmed for that cycle. The original heterogeneous data pairs come from fragments of abnormal pattern data streams and background pattern data streams stored in history. The network intermediate layer feature map snapshots are obtained by real-time capture and persistent storage through probe layers implanted during the execution of historical detection tasks. The finally confirmed hazard labeling results come from the closed-loop records of historical work orders of maintenance personnel. The three are associated and bound through a unified task identifier and timestamp.

[0028] After constructing a complete historical database of hazard detection tasks, when calculating the contribution weight of a certain network layer, the system retrieves all task instances containing valid hazard labels, extracts the intermediate feature map snapshots corresponding to that network layer in each task instance, compares the correlation differences between the output feature maps of different network layers in the same task instance and the final hazard area, and analyzes the distribution discrimination of the feature map of that layer in positive and negative sample task instances, comprehensively calculates the quantitative contribution value of that layer to driving the network to make correct decisions. This calculation process fully integrates the internal signal propagation characteristics of the network and external verification feedback, so that the contribution weight can accurately reflect the functional importance of the network layer in real scenarios.

[0029] The dynamic pruning threshold setting and adjustment mechanism constitutes another core support for online pruning operations. The pruning threshold is not a fixed global parameter; its value directly determines how many low-contribution network layers will be removed. Therefore, it needs to be tightly coupled with the real-time computing status of the edge device. While hardware performance profiles provide static benchmarks for memory capacity and processor computing power, the actual load on the device fluctuates dynamically. The dynamic pruning threshold adjustment mechanism continuously monitors the current available memory and instantaneous processor occupancy of the edge device. When available memory is scarce or the processor is under sustained high load, the mechanism tends to increase the pruning threshold, triggering more aggressive network pruning to remove more network layers and quickly reduce the model's memory usage and computational load. Conversely, when device resources are scarce... When resources are plentiful, the threshold value can be appropriately lowered to retain more network layer capacity and maintain the model's detection capability. This adjustment process is achieved through a lightweight resource monitoring agent and feedback controller. The resource monitoring agent collects device performance indicators at a fixed frequency, and the feedback controller maps the corresponding pruning threshold value adjustment amount according to a preset resource threshold strategy. For example, when available memory is less than 20% of the total capacity, the pruning threshold value is increased by 0.02 on the base value. The adjusted threshold value immediately takes effect in the marking process of the current set of network layers to be removed, forming a closed-loop control that dynamically optimizes model complexity based on real-time hardware load. This mechanism ensures that the lightweight detection network can obtain an adapted topology on terminal devices with different resource constraints, achieving the best balance between model performance and resource consumption.

[0030] In one embodiment of the present invention, in a specific implementation, the size parameter and stride parameter of the convolution kernel are calculated in real time based on the statistical characteristics of data segments in the abnormal mode data stream. In an example scenario, the abnormal mode data stream contains infrared image sequence segments of local overheated areas on the surface of a conductor. The statistical characteristics include the standard deviation of pixel intensity values ​​and the mean of gradient magnitude within the segments. Data comparison shows that the standard deviation of pixel intensity of image segments containing obvious foreign objects is about 60% higher than that of normal background segments. The size parameter is calculated based on the statistical characteristics. When, a reference intensity mean is introduced. Pixel intensity standard deviation Normalized to the relative rate of change, the formula is expressed as: in: Indicates dimensional parameters, This represents the base value of the dimensional parameter. The standard deviation of pixel intensity in the current data segment. This represents a predefined reference intensity mean. It is a dimensionless proportional coefficient that controls the sensitivity of size adjustment; step size parameter. Automatic adjustment is made based on the spatial continuity of the data segments, which is obtained by calculating the matching rate of feature points between adjacent frames. The calculated size parameters are then used. With step size parameter The convolutional computation units are dynamically assembled in memory. The dynamic assembly process includes assembling them according to size parameters. Allocate convolutional kernel weight matrices of the appropriate dimensions in memory, and determine the weights based on the stride parameter. Configure the stride during convolution operations. In some embodiments, the convolution kernel weight matrix is ​​initialized by selecting the closest set from a predefined weight pool based on the size parameter. The abnormal mode data stream is divided into multiple data blocks, which are sequentially input into dynamically assembled convolutional computation units for convolution operations. The data block division is based on time windows and spatial regions. In the example scenario, each data block corresponds to a sub-region containing 5 consecutive frames of infrared images with a spatial range of 32x32 pixels. Data comparison shows that data blocks divided according to this specification have a more balanced memory usage in subsequent convolution operations.

[0031] The intermediate feature maps output after convolution operations on each data block are collected. These intermediate feature maps are three-dimensional tensors, and their number of channels equals the number of convolutional kernels in the dynamically assembled convolutional computation unit. Optionally, the spatiotemporal index information of the corresponding data block is included with the intermediate feature maps during collection. Multi-scale pooling and normalization are performed on all intermediate feature maps, which are then integrated to form an anomaly feature map. Multi-scale pooling uses pooling windows of different sizes to downsample the same intermediate feature map, generating multi-resolution feature representations. Normalization uses layer normalization to standardize the feature values ​​of each channel. The integration process concatenates all feature tensors after multi-scale pooling and normalization along the channel dimension. In some embodiments, the multi-scale pooling window size includes 2x2 and 4x4. It can be understood that the dynamic convolutional kernel adaptively adjusts its parameters for data segments with different statistical characteristics. Optionally, a scaling factor... Different values ​​can be pre-configured based on different types of potential hazards. It's understandable that multi-scale pooling helps capture abnormal features at different granularities.

[0032] In practice, multi-scale pooling involves parallel downsampling of the same intermediate feature map at different spatial granularities to construct a multi-resolution feature representation. The intermediate feature map originates from the output of the dynamic convolutional computation unit and is a three-dimensional tensor with a specific height, width, and number of channels. Multi-scale pooling does not pre-specify a fixed pooling window size but dynamically calculates a set of pooling window sizes based on the spatial dimensions of the current intermediate feature map. For example, for a feature map of size H×W, the system calculates two sets of pooling window sizes, one set being (H / a, W / a) and the other being (H / b, W / b), where a and b are preset scaling factors. The pooling operation itself uses max pooling, which selects the maximum value of all feature values ​​within each pooling window as the maximum value. The output of the window region, after being processed by pooling windows of different sizes, transforms the same intermediate feature map into multiple smaller feature maps. These feature maps have the same number of channels but different spatial resolutions. In order to integrate these feature information at different scales, they need to be upsampled or spatially transformed to unify them to the same size. The system uses bilinear interpolation to enlarge the smaller-scale feature maps to the same spatial size as the original feature map. After size alignment, the feature data of all channels are stitched together according to the channel dimension to form an integrated feature map with a multiplied number of channels. The integrated feature map carries information at multiple levels, from fine local details to relative global context. This multi-scale pooling processing method enables subsequent analysis to simultaneously capture the subtle structure of anomalous features and their distribution patterns over a larger range.

[0033] The layer normalization method aims to address the inconsistency in the numerical distribution of intermediate feature maps caused by dynamic convolution kernels. Since the convolution kernel parameters are dynamically generated based on the input data segments, the intermediate feature maps obtained after convolution operations on different data blocks may differ in numerical range and distribution. The layer normalization method operates on each independent data sample of each intermediate feature map. For an input 3D feature tensor, this method calculates the mean and variance of the tensor's feature values ​​across all channels and spatial locations. The mean is calculated by summing all elements in the tensor and dividing by the total number of elements. The variance is calculated based on the sum of squared deviations of the same set of elements from the mean. Using the calculated mean and variance, each element in the original feature tensor is standardized. The standardization formula is: subtract the mean, divide by the variance, and add the square root of a very small constant. This operation transforms the feature values ​​into a standard distribution with a mean of 0 and a variance of 1. To preserve the network's representational power, the layer normalization method further introduces learnable scaling and translation parameters. These scaling and translation parameters are one-dimensional vectors with the same number of channels. The normalized feature tensor is multiplied by the scaling parameter channel by channel and then added to the translation parameter channel by channel, allowing the network to adaptively recover a numerical distribution suitable for subsequent processing. The feature map after layer normalization has improved numerical stability, and the feature maps generated by different data blocks are normalized to similar distribution intervals. This is beneficial for subsequent multi-scale pooling integration and correlation calculation of fusion decision branches. The layer normalization calculation process is performed online immediately after each data block is processed, and its parameters are learned together with the network weights during the model training phase and are fixed for use during the inference phase.

[0034] In one embodiment of the present invention, the anomaly feature map and the environmental reference map are aligned using a grid to ensure that they have a one-to-one correspondence of grid cells in spatial coordinates. In the example scenario, the anomaly feature map is a thermal anomaly feature map generated by processing an infrared image sequence with a dynamic convolutional kernel, and the environmental reference map is a texture reference map generated by processing a visible light background image with a static convolutional kernel. The two maps have the same spatial resolution. Data comparison shows that there is a slight offset in the coordinate system of the two maps before alignment, but the coordinates of each pixel point are matched after alignment. For each aligned grid cell, feature vectors are extracted from the anomaly feature map and the environmental reference map respectively. The grid cell size is 8x8 pixels. The feature vector extracted from the anomaly feature map includes the mean and variance of the multi-channel feature values ​​within the grid, and the feature vector extracted from the environmental reference map includes the texture descriptor and color histogram statistics within the grid. The dimension of the feature vectors is 64 dimensions. The cosine similarity and Euclidean distance between feature vectors from the anomaly feature map and feature vectors from the environmental baseline map are calculated. Cosine similarity measures the proximity of two vectors in a direction, while Euclidean distance measures the absolute difference between two vectors in space. The cosine similarity and Euclidean distance are combined into a correlation strength index. The calculation of the correlation strength index requires a dimensionless result; therefore, the Euclidean distance is normalized, as expressed by the formula: in: Indicators representing the strength of correlation, Represents the cosine similarity value. Represents the Euclidean distance value. This represents the reference distance value used for normalization, which is taken from the statistical median of the Euclidean distances of all grid cells. It is a dimensionless scaling factor. This is a very small constant used to prevent division by zero errors. Iterate through all grid cells to generate a correlation intensity distribution map with the same spatial resolution as the original map. The gray value of each pixel in the correlation intensity distribution map is related to the correlation intensity index of its corresponding grid cell. Proportional. In some embodiments, the scaling factor... The value is set to 1.0. This can be understood as normalizing the Euclidean distance to ensure the dimensionless nature of the correlation strength index.

[0035] Region growing segmentation is performed on the correlation intensity distribution map to identify connected regions where the correlation intensity exceeds a judgment threshold. The seed point for region growing segmentation is selected from the pixel with the highest gray value in the correlation intensity distribution map. The growth criterion is that the difference in gray value between adjacent pixels is less than a fixed tolerance and all gray values ​​are higher than the judgment threshold. The judgment threshold is dynamically determined based on the overall gray-level histogram of the correlation intensity distribution map. The geometric center coordinates and the bounding rectangle boundary of each connected region are calculated. The geometric center coordinates are obtained by calculating the mean of the coordinates of all pixels in the connected region, and the bounding rectangle boundary is determined by finding the minimum and maximum x and y coordinates of the pixels in the connected region. Based on the geometric center coordinates, the corresponding position in the original anomaly feature map is traced back to extract the original multimodal sensor data fragments corresponding to the position. The original multimodal sensor data fragments include the original infrared image data block and the original waveform fragment of the vibration sensor near the timestamp corresponding to the geometric center coordinates. By combining the circumscribed rectangle boundary with the extracted original multimodal sensor data fragments, a structured hazard report is generated, containing a description field for the hazard location and a classification label for the hazard type. The description field records the corner coordinates of the circumscribed rectangle boundary in text form, and the classification label is determined based on the pattern recognition results of the original multimodal sensor data fragments. For an example of calculating the correlation strength index in some embodiments, please refer to Table 1.

[0036] Table 1: Calculation Table of Mesh Cell Correlation Strength It is understandable that region growing and segmentation aggregates discrete high-intensity grid cells into continuous regions. Optionally, the dynamic determination of the judgment threshold adopts the Otsu method. It is understandable that backtracking and extracting original fragments of multimodal sensor data provides the original basis for hazard classification.

[0037] In its implementation, the Otsu method is used to dynamically determine the threshold of the correlation intensity distribution map to achieve adaptive region segmentation. The correlation intensity distribution map is a single-channel grayscale image where the grayscale value of each pixel represents the correlation intensity index of the corresponding grid cell. The core idea of ​​the Otsu method is to find an optimal grayscale threshold such that the inter-class variance of the foreground and background regions segmented according to this threshold is maximized. Maximizing the inter-class variance means that the difference between the foreground and background regions in the segmentation result is the most significant, thus obtaining the optimal binary segmentation effect. The specific calculation process involves statistically analyzing the grayscale histogram of the entire correlation intensity distribution map, i.e., counting the number of pixels at each grayscale level (e.g., 0-255), normalizing the grayscale histogram to convert the number of pixels at each grayscale level into a probability value, and iterating through all possible grayscale threshold assumptions. If the current threshold is k, the pixels are divided into two categories based on the threshold: pixels below or equal to k are classified as background pixels, and pixels above k are classified as foreground pixels. The probability of background pixels and their average gray value are calculated, as are the probability of foreground pixels and their average gray value. Based on this, the inter-class variance under the current threshold is calculated. The formula is: background probability multiplied by foreground probability multiplied by the square of the difference between the average gray values ​​of the two classes. After traversal, the gray value that maximizes the inter-class variance is selected as the final threshold. This threshold will be directly used in subsequent region growing and segmentation to determine whether a pixel belongs to a connected region with high correlation strength. Since the Otsu method automatically calculates the threshold based on the gray value distribution of the image itself, it can adapt to changes in the overall brightness and contrast of the correlation strength distribution map under different monitoring scenarios, thus achieving dynamic adjustment of the judgment.

[0038] The process of determining classification labels from pattern recognition results of raw multimodal sensor data fragments relies on a pre-built hazard feature template library. This library is derived from historically confirmed hazard cases, with each template corresponding to a known hazard type, such as bird nests, insulator damage, or foreign objects in the conductor. Each template contains typical feature data for that type of hazard across multiple modes, such as the hotspot distribution characteristics of infrared image fragments, the spectral characteristics of vibration sensor waveform fragments, and their temporal statistical characteristics. When it is necessary to backtrack and extract raw multimodal sensor data fragments based on geometric center coordinates, the system extracts features from both the raw infrared image data blocks and the raw vibration sensor waveform fragments. Infrared image feature extraction includes calculating the shape factor of the hotspot region, temperature gradient distribution, and statistical information relative to the conductor's position. Vibration waveform feature extraction includes calculating specific frequency bands (e.g., 1-5). The power spectral density (0Hz), waveform attenuation coefficient, and duration of the impact event are extracted and combined into a multimodal feature vector. The system performs a one-to-one matching process between this feature vector and all templates in the hazard feature template library. The weighted Euclidean distance between the feature vector to be identified and each template feature vector is calculated. The weighting coefficient is preset according to the different features and the distinguishability of various hazards. After calculating a set of distance values ​​with all templates, the system selects the hazard type corresponding to the template with the smallest distance as the preliminary identification result. At the same time, it checks whether the minimum distance is less than the preset matching success threshold. If it is less than the threshold, the preliminary identification result is confirmed and the corresponding classification label is generated. If it is greater than the threshold, the identification result is marked as "unknown type" and needs to be submitted for manual review. The classification label is finally written into the structured hazard report and the hazard location description field to form a complete hazard record.

[0039] See Figure 4 This is a bar chart evaluating the correlation strength of grid cells in power transmission line hazard detection. It quantifies the matching degree between the anomaly feature map and the environmental baseline map through three dimensions: cosine similarity, normalized Euclidean distance, and correlation strength index. The correlation strength index is a comprehensive reflection of cosine similarity and normalized Euclidean distance; a higher value indicates a greater likelihood of a hazard in the grid cell. Grid cells with high correlation strength (such as (14,46)) can be directly located as priority areas for hazard investigation, improving inspection efficiency. By analyzing the distribution of correlation strength, the hazard judgment threshold can be dynamically adjusted to reduce missed and false detections. It provides a quantitative evaluation basis for the output results of the lightweight detection network and verifies the detection accuracy of the model on different grid cells. Long-term monitoring of the changing trend of correlation strength can identify the evolution pattern of hazards and provide early warning of potential risks.

[0040] In one embodiment of the present invention, in a specific implementation, the data acquisition agent services deployed at each monitoring point along the transmission line are polled. In an example scenario, multiple monitoring points are deployed along the transmission line, and each monitoring point is equipped with a data acquisition agent service. The polling process sends status query instructions to each data acquisition agent service sequentially according to a predetermined order list. Data comparison shows that under normal network conditions, the average response time of the data acquisition agent service to the status query instructions is 200 milliseconds. When the data acquisition agent service returns to the ready state, a data retrieval instruction with a timestamp is sent to it. The timestamp is used to mark the precise time of the data request to ensure subsequent data synchronization. The data retrieval instruction specifies the type and time range of the required data. The packaged data block returned by the data acquisition agent service according to the data retrieval instruction is received. The packaged data block is an encapsulated binary data packet containing image data files, time-series sensor data files, and metadata for verification.

[0041] The packaged data blocks are unpacked and verified to separate independent image data files and time-series sensor data files. The unpacking process parses the structure of binary data packets according to a predefined communication protocol. The verification process compares the calculated cyclic redundancy check (CRC) code of the packaged data blocks with the check code recorded in the metadata. The data comparison shows that the calculated CRC code of the packaged data blocks that are successfully verified is completely consistent with the recorded value, while the packaged data blocks that fail to verify are inconsistent.

[0042] After unpacking and verifying the packaged data block, if verification fails, the device identifier and failure time of the data acquisition agent service corresponding to the packaged data block are recorded. The device identifier is a unique code of the data acquisition agent service, and the failure time is the system time when the verification operation occurred. The recorded information is written to the local log file. A data retransmission request is sent to the data acquisition agent service, and a retransmission timer is started. The data retransmission request includes the timestamp information of the previously failed data retrieval command. The timeout duration of the retransmission timer is dynamically set according to network conditions; in the example scenario, the initial timeout duration is 5 seconds. If a correct data packet is received before the retransmission timer expires, the processing continues. The correct data packet refers to the packaged data block that has passed verification. If a correct data packet is not received after the retransmission timer expires, the device identifier is marked as abnormal and temporarily removed from the current polling list. At the same time, a backup data acquisition agent service in the same area is activated to supplement data collection. The backup data acquisition agent service is physically close to the main service and can collect monitoring data from the same transmission line section. In some embodiments, the integrity verification formula used in the verification process is: in: This represents the final check value. This represents the original content of the packaged data block, excluding the validation fields. This represents a 32-bit cyclic redundancy check function. This refers to the SHA-256 hash function. This represents a bitwise XOR operation. It can be understood that Cyclic Redundancy Check (CRC) codes are used to detect random errors during transmission. Optionally, the timeout duration of the retransmission timer can be dynamically adjusted based on historical network latency.

[0043] In practical implementation, if the verification is successful, the image data files and time-series sensor data files are synchronized and stitched together according to a unified time base to form the original monitoring data stream of the transmission line. The unified time base adopts Coordinated Universal Time (UTC). The synchronization process pairs time-aligned data frames by comparing the timestamps in the image file metadata with the timestamp sequences in the sensor data files. The stitching process combines the paired image frames and sensor data frames into a continuous data stream object in chronological order. Data comparison shows that the alignment error between the synchronized image data and vibration sensor data on the time axis is less than 10 milliseconds. In some embodiments, time base synchronization is calibrated using a network time protocol. It can be understood that a unified time base is the foundation for ensuring accurate correlation of multi-source data. Optionally, the data stream object is stored in memory using a circular buffer structure.

[0044] See Figure 5 This is a performance evaluation chart of the lightweight model for detecting hidden dangers in power transmission lines. It quantifies the resource consumption and detection effect of different levels of the model from three dimensions: model size, inference time, and detection accuracy. Model size and inference time are positively correlated; the later the level, the smaller the model and the shorter the inference time, reflecting the effect of lightweight pruning. After the accuracy of the feature extraction layer 1 drops due to information compression, the accuracy of subsequent layers gradually recovers and approaches the theoretical peak through feature purification and fusion. The decision layer and output layer achieve the optimal balance in terms of model size, inference time, and accuracy, and are the core layers for deploying the lightweight model. This provides a basis for the allocation of computing power and memory for edge devices, such as reserving more resources for the input layer and feature extraction layer to ensure detection accuracy.

[0045] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A model-based lightweight method for detecting hidden dangers in transmission lines, characterized in that, include: Acquire raw monitoring data streams of power transmission lines that include image data and time-series sensor data; Based on the predefined hidden danger characterization rules, the original monitoring data stream of the transmission line is deconstructed to separate the abnormal mode data stream and the background mode data stream. The abnormal mode data stream and the background mode data stream constitute a heterogeneous data pair. The heterogeneous data pairs are input into a preloaded lightweight detection network, which includes parallel processing branches and fusion decision branches. In the parallel processing branch, a dynamic convolution kernel is used to extract features from the abnormal pattern data stream, while a static convolution kernel is used to model the environment of the background pattern data stream, generating an abnormal feature map and an environmental baseline map respectively. The abnormal feature map and the environmental baseline map are input into the fusion decision branch to perform cross-map correlation calculation; Based on the correlation calculation results, a structured hazard report containing the location and type of the hazard is generated.

2. The model-based lightweight transmission line hazard detection method according to claim 1, characterized in that, The step of performing pattern deconstruction on the original monitoring data stream of the transmission line based on predefined hidden danger characterization rules specifically includes: For the image data in the original monitoring data stream of the transmission line, the edge saliency detection operator is invoked to extract high gradient change regions and low texture smooth regions in the image; For the time-series sensor data in the original monitoring data stream of the transmission line, a sliding window differential algorithm is applied to identify abrupt change points and periodic fluctuation segments in the data sequence. The high gradient change region and the mutation point are spatiotemporally aligned and logically bound to form an initial candidate set for the abnormal mode data stream; The low-texture smooth region and the periodic fluctuation segment are fused to form a stable reference set for the background pattern data stream; A confidence level filter is applied to the initial candidate set of the abnormal pattern data stream to remove pseudo-abnormal data with a confidence level lower than a preset threshold, thus forming the final abnormal pattern data stream.

3. The model-based lightweight transmission line hazard detection method according to claim 2, characterized in that, Before the step of applying confidence filtering to the initial candidate set of the anomalous pattern data stream, the method further includes: A multi-level confidence model is constructed to evaluate the pseudo-anomaly data. The multi-level confidence model integrates historical false alarm statistics and real-time environmental interference factors. From the initial candidate set of the abnormal mode data stream, the spatiotemporal attributes and physical quantity attributes are extracted item by item; The spatiotemporal attributes and physical quantity attributes of each item are input into the multi-level confidence model to calculate the individual confidence score; Based on all individual confidence scores, generate a confidence distribution histogram of the initial candidate set of the abnormal pattern data stream; Based on the confidence distribution histogram, the value of the preset threshold is dynamically adjusted.

4. The model-based lightweight transmission line hazard detection method according to claim 1, characterized in that, In the step of inputting the heterogeneous data pair into the preloaded lightweight detection network, the loading process of the lightweight detection network includes: Download the basic network topology that matches the current transmission line tower model from the cloud model repository; Read the hardware performance configuration file deployed on the local edge device, which contains memory capacity and processor computing power information; Based on the hardware performance configuration file, an online pruning operation is performed on the basic network topology to remove redundant connection layers and activation layers; Using the network topology generated after pruning, the corresponding model weight parameter files are incrementally pulled from the cloud model repository; The model weight parameter file is loaded into memory to complete the instantiation and warm-up of the lightweight detection network.

5. The model-based lightweight transmission line hazard detection method according to claim 4, characterized in that, The step of performing online pruning on the basic network topology specifically includes: Analyze the contribution weight of each layer of the basic network topology to the historical hidden danger detection task; Based on the contribution weight, the network layers are sorted, and the set of network layers whose contribution is lower than the pruning threshold is marked. Check the data dependencies between the network layers in the set of network layers to be removed and their adjacent network layers; For network layers with strong data dependencies, their structure is retained but their weights are reset to zero; for network layers without strong data dependencies, they are directly removed from the basic network topology. Traverse and update the data flow paths of the entire network to generate a simplified network topology description file.

6. The model-based lightweight transmission line hazard detection method according to claim 1, characterized in that, The step of using dynamic convolutional kernels to extract features from the abnormal pattern data stream specifically includes: Based on the statistical characteristics of data segments in the abnormal mode data stream, the size parameters and stride parameters of the convolution kernel are calculated in real time. Using the calculated size and stride parameters, convolutional computation units are dynamically assembled in memory; The abnormal mode data stream is divided into multiple data blocks, which are then sequentially input into the dynamically assembled convolutional computation unit for convolutional operations. Collect the intermediate feature maps output after each data block undergoes convolution operations; All intermediate feature maps are subjected to multi-scale pooling and normalization processing, and then integrated to form the abnormal feature map.

7. The model-based lightweight transmission line hazard detection method according to claim 1, characterized in that, The steps for performing cross-graph correlation calculations specifically include: The anomaly feature map and the environmental baseline map are aligned by gridding to ensure that the anomaly feature map and the environmental baseline map have a one-to-one correspondence of grid cells in spatial coordinates; For each aligned grid cell, feature vectors are extracted from the anomaly feature map and the environmental baseline map, respectively. Calculate the cosine similarity and Euclidean distance between the feature vectors from the anomaly feature map and the feature vectors from the environmental baseline map; The cosine similarity and the Euclidean distance are combined into a correlation strength index; Traverse all grid cells to generate a correlation intensity distribution map with the same spatial resolution as the original map.

8. The model-based lightweight transmission line hazard detection method according to claim 7, characterized in that, After generating a correlation strength distribution map with the same spatial resolution as the original map, the process further includes: The correlation strength distribution map is segmented by region growing to identify connected regions whose correlation strength exceeds the judgment threshold; Calculate the geometric center coordinates and the bounding rectangle of each connected region; Based on the geometric center coordinates, backtrack to the corresponding position in the original anomaly feature map and extract the original fragment of multimodal sensing data corresponding to the position; By combining the circumscribed rectangle boundary with the extracted original fragments of multimodal sensor data, a descriptive field for the location of the hazard and a classification label for the hazard type are generated in the structured hazard report.

9. The model-based lightweight transmission line hazard detection method according to claim 1, characterized in that, The step of acquiring the raw monitoring data stream of the transmission line, which includes image data and time-series sensor data, specifically includes: Poll the data acquisition proxy service deployed at each monitoring point on the transmission line; When the data acquisition agent service returns to the ready state, a data retrieval command with a timestamp is sent to it; Receive the packaged data block returned by the data acquisition agent service according to the data retrieval instruction; The packaged data blocks are unpacked and verified to separate independent image data files and time-series sensor data files; The image data file and the time-series sensor data file are synchronized and spliced ​​according to a unified time reference to form the original monitoring data stream of the power transmission line.

10. The model-based lightweight transmission line hazard detection method according to claim 9, characterized in that, After the step of unpacking and verifying the packaged data block, the method further includes: If the verification fails, the device identifier and failure time of the data acquisition agent service corresponding to the packaged data block are recorded. Send a data retransmission request to the data acquisition agent service and start the retransmission timer; If the correct data packet is received before the retransmission timer expires, the processing continues; If the retransmission timer times out and no correct data packet is received, the device identifier is marked as abnormal and temporarily removed from the current polling list. At the same time, the backup data acquisition agent service in the same area is activated to collect supplementary data.

Images