Power transmission and distribution line passage anomaly detection method, system, device and computer program product

Abstract

The application discloses a power transmission and distribution line channel anomaly detection method, system, device and computer program product, and belongs to the technical field of anomaly detection. The method comprises the following steps: generating a multi-element structure field prior graph according to the structural characteristics of the power transmission and distribution line channel; constructing a target detection network to obtain a multi-scale feature set; performing spatial registration on the multi-element structure field prior graph and each scale feature to generate a prior gating graph, and then performing gating modulation to obtain enhanced features; performing boundary box regression based on the enhanced features to output candidate boundary boxes and regression uncertainty; performing scale self-adaptive supervision construction and regular constraint on the candidate boundary boxes based on the multi-element structure field prior graph to obtain spatial consistency scores and calibration confidence; and performing deterministic filtering according to the spatial consistency scores, the regression uncertainty and the calibration confidence, and performing suppression processing to obtain a final detection result. The method can significantly reduce false positives outside the corridor and improve the stability and reliability of long-range small target detection and alarm.

Description

Technical Field

[0001] This invention relates to the field of anomaly detection technology, and in particular to a method, system, device, and computer program product for anomaly detection of power transmission and distribution line channels. Background Technology

[0002] Transmission and distribution line corridors are crucial areas for ensuring the safe operation of power systems. To promptly detect risks such as intrusion from internal and external forces, illegal construction, fire hazards, and foreign object attachments, power operation and maintenance units commonly employ drone inspections, fixed monitoring, and mobile terminal image or video capture methods for corridor patrols. Due to the large volume of inspection data and high timeliness requirements, relying solely on manual frame-by-frame inspection is inefficient and susceptible to subjective factors. Existing technologies for identifying abnormal targets in corridors (such as construction vehicles, cranes, tower cranes, smoke / burning, and bird nests) mainly include traditional image processing methods and deep learning-based target detection methods. Traditional methods typically rely on manual rules such as color, edge, texture, morphology, or background modeling. They lack robustness in complex conditions such as low light, fog, backlight, compressed noise, and changing viewing angles, and are poorly adaptable to changes in target scale and morphological diversity. With the development of deep learning and convolutional neural network training, the one-stage / two-stage object detection framework based on the network has been widely used in power line inspection scenarios due to its advantages such as end-to-end training and a balance between accuracy and speed. The detection performance has been improved through data augmentation, feature pyramids, multi-scale training, attention mechanisms, and loss function improvements.

[0003] However, the inspection scenario of power transmission and distribution line corridors has distinct industry characteristics, resulting in the following shortcomings of general detection models in practical applications: First, the surrounding environment of power transmission and distribution line corridors is often accompanied by complex backgrounds such as roads, buildings, vegetation, water bodies, shadows, and reflections, as well as interference objects that resemble the appearance of abnormal targets. General detection models are prone to generating false alarms outside the corridor or in non-interested areas. Furthermore, engineering alarms emphasize "accurate alarms," ​​and false alarms will lead to repeated verification and ineffective handling, reducing system availability. Second, drone inspections are characterized by long-distance shooting and large changes in the top-down perspective. Targets in the images are often small or even extremely small in scale, making it easy to miss vehicle targets. Smoke targets have blurred boundaries and irregular shapes, making it difficult to guarantee positioning accuracy and stability. Third, imaging conditions such as low light, fog, and backlighting significantly change contrast and texture distribution, resulting in large performance fluctuations of the model when deployed across time, weather, and equipment.

[0004] Furthermore, to improve positioning accuracy, some detection frameworks employ distributed bounding box regression (representing position regression as discrete distribution prediction) to enhance performance in high Intersection over Union (IoU) intervals. However, in cases of small targets, annotation errors, or uncertain boundaries, the regression distribution is prone to overfitting to noise, and the regression-related loss may rebound on the validation set in the later stages of training, thus limiting the improvement of high IoU metrics. Meanwhile, current inference stages generally employ fixed thresholds and general suppression strategies to screen candidate boxes, lacking a mechanism to spatially constrain and calibrate the confidence level of detection results using the structural feature of the "effective area of ​​the power transmission and distribution corridor." This leads to false alarm control relying more on empirical parameter tuning, making it difficult to form stable and reproducible alarm strategies. Therefore, how to effectively suppress interference, improve the detection capability of distant small targets, and enhance positioning accuracy and stability in cases of small targets and blurred boundaries remains a key technical challenge that urgently needs to be addressed in the field of power transmission and distribution line corridor anomaly detection. Summary of the Invention

[0005] The purpose of this application is to overcome the shortcomings of the prior art and provide a method, system, device and computer program product for detecting abnormalities in power transmission and distribution line channels, so as to achieve high accuracy and low false alarm detection and alarm of abnormal targets.

[0006] In a first aspect, this application provides a method for detecting anomalies in power transmission and distribution line channels, comprising the following steps:

[0007] The image to be analyzed is preprocessed to obtain input data;

[0008] Based on the structural characteristics of the power transmission and distribution line corridor, structural information of the effective area of ​​the corridor is extracted from the input data to generate a multi-element structural field prior map.

[0009] Construct an object detection network, input the input data into the object detection network, and obtain a multi-scale feature set;

[0010] The prior map of the multi-structure field is spatially registered with features at each scale to generate a prior gating map. Based on the prior gating map, gating modulation is performed to obtain enhanced features.

[0011] Based on the enhanced features, a forward operation for bounding box regression is performed, and candidate bounding boxes and their corresponding regression uncertainties are output.

[0012] Based on the prior graph of the multivariate structure field, the candidate bounding box is constructed and regularized by scale-adaptive distribution regression under supervision to obtain spatial consistency score and calibration confidence.

[0013] Deterministic filtering is performed based on the spatial consistency score, regression uncertainty, and calibration confidence, and suppression processing is applied to the filtered candidate bounding boxes to obtain the final detection result.

[0014] Optionally, based on the structural characteristics of the transmission and distribution line corridor, structural information of the effective area of ​​the corridor is extracted from the input data to generate a multi-element structural field prior map, including:

[0015] Edge extraction and line segment detection are performed on the input data to obtain a set of line segments. The direction angles of the line segments are statistically analyzed or clustered to determine the main direction of the corridor. The set of line segments is fitted based on the main direction of the corridor to obtain the center line of the corridor.

[0016] Determine the corridor half-width function based on perspective effect;

[0017] The probability field of the corridor region is determined based on the distance of each pixel from the center line of the corridor.

[0018] The corridor region is determined based on the probability field of the corridor region, and the distance from each pixel to the corridor region is calculated to obtain the distance field;

[0019] The gradient direction is calculated on the distance field, and a unit vector pointing towards the direction of decreasing distance is taken to obtain the direction field.

[0020] Optionally, the multi-structured prior map includes one or more of the following: the corridor region probability field, the distance field to the corridor centerline or boundary, and the direction field of the conductor or the main direction of the corridor.

[0021] Optionally, the prior map of the multi-dimensional structure field is spatially registered with features at each scale to generate a prior gating map. Based on the prior gating map, gating modulation is performed to obtain enhanced features, including:

[0022] The scale-aligned prior is obtained by aligning the multi-structure field prior map to the spatial resolution of each scale feature in the multi-scale feature set.

[0023] The scale-aligned prior is mapped to a prior gating graph;

[0024] Based on the prior gating map, affine gating modulation is performed on each scale feature to obtain enhanced features.

[0025] Optionally, a forward operation for bounding box regression is performed based on the enhanced features to output candidate bounding boxes and their corresponding regression uncertainties, including:

[0026] The enhanced features are input into the detection head, and the output is the class probability and the discrete distribution of the four sides of the bounding box;

[0027] The candidate bounding box is obtained by performing expectation decoding on the discrete distribution of the four sides of the bounding box;

[0028] Calculate the regression uncertainty of the candidate bounding box.

[0029] Optionally, based on the prior graph of the multivariate structure field, supervised construction and regularization constraints are applied to the candidate bounding boxes using scale-adaptive distribution regression to obtain spatial consistency scores and calibration confidence scores, including:

[0030] Determine the target scale based on the real frame;

[0031] Establish a mapping function from the target scale to the temperature parameter, and construct a soft label distribution based on the mapping function from the target scale to the temperature parameter;

[0032] The distribution regression loss is determined based on the soft-label distribution;

[0033] Set a distribution smoothing term, establish a mapping function from the target scale to the regularization weights, determine the regularization loss based on the mapping function from the target scale to the regularization weights, and obtain the total loss;

[0034] The corridor probability consistency, distance consistency, and orientation consistency of pixels inside the candidate bounding box are calculated respectively, and the spatial consistency score is obtained by weighted fusion.

[0035] The original category confidence score is calibrated using the spatial consistency score and regression uncertainty to obtain the calibrated confidence score.

[0036] Optionally, it also includes calculating a risk score based on calibration confidence, spatial consistency score, and the degree of intrusion of the candidate bounding box relative to the corridor, determining an alarm level based on the risk score, and outputting the final result.

[0037] Secondly, this application also provides a transmission and distribution line channel anomaly detection device for performing the transmission and distribution line channel anomaly detection method as described in any one of the first aspects, comprising: a preprocessing module, a channel structure space prior construction module, a prior multi-scale alignment and gating modulation module, a detection head module, a scale adaptive distribution regression training module, an inference calibration and post-processing module, and an alarm evaluation and output module. The preprocessing module preprocesses the image to be analyzed to obtain input data; the channel structure space prior construction module generates a multi-dimensional structure field prior map based on the input data output by the preprocessing module; the prior multi-scale alignment and gating modulation module extracts and fuses features from the input data to obtain a multi-scale feature set, and applies it to the multi-dimensional structure field prior map. The system performs prior alignment and gated modulation to output enhanced features. The detection head module takes these enhanced features as input, calculates the class probability and the discrete distribution of the four sides of the bounding box, and performs expected decoding to output candidate bounding boxes and regression uncertainty. The scale-adaptive distribution regression training module performs supervised construction and regularization constraints on the candidate bounding boxes to obtain spatial consistency scores, calibration confidence, and calculates the total loss. The inference calibration and post-processing module calculates consistency indices, confidence calibration, deterministic filtering, and coordinate remapping based on the candidate bounding boxes and regression uncertainty, and outputs the final output bounding boxes. The alarm evaluation and output module calculates risk scores and classifies alarm levels based on calibration confidence and corridor half-width, and outputs graded alarm information and the final result.

[0038] Thirdly, this application also provides a transmission and distribution line channel anomaly detection system, comprising: one or more processors; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the transmission and distribution line channel anomaly detection method as described in any one aspect.

[0039] Fourthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the power transmission and distribution line channel anomaly detection method as described in any one of the first aspects.

[0040] This application provides a method, system, device, and computer program product for detecting anomalies in power transmission and distribution line corridors. By constructing a multi-dimensional structured field prior map and implementing gating modulation on multi-scale feature flows, the target detection network becomes more sensitive to the effective area of ​​the corridor and more suppresses interference outside the corridor, thereby reducing false alarms caused by complex backgrounds. Through scale-adaptive distribution regression, differentiated distribution supervision / constraints are applied to targets of different scales, making the regression of small targets more stable and noise-resistant, and the localization of large targets more precise, which can alleviate the rebound of regression loss at the validation end caused by overfitting of the regression distribution in the later stage of training. By combining prior gating and prior calibration, the detection stability under conditions such as low light, haze, and backlight is improved, and the missed detection of small targets in the distance is also improved. Through deterministic filtering and suppression processing using multi-dimensional indicators, false detection results can be effectively eliminated, significantly improving the accuracy, reliability, and robustness of anomaly detection in power transmission and distribution line corridors, providing strong protection for the safe and stable operation of power transmission and distribution lines. At the same time, the solution process is clear and the modules are clearly divided, which can be integrated with existing target detection frameworks, making it easy to implement in engineering and facilitating subsequent expansion in terms of alarm strategies, model lightweighting, and deployment optimization.

[0041] To make the above-mentioned features and advantages of the invention more apparent and understandable, specific embodiments are described below, and detailed descriptions are provided in conjunction with the accompanying drawings. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart of a method for detecting abnormalities in power transmission and distribution line channels provided in one embodiment of this application.

[0044] Figure 2 This is a flowchart of step S2 in the power transmission and distribution line channel anomaly detection method provided in one embodiment of this application.

[0045] Figure 3 This is a flowchart of step S4 in the power transmission and distribution line channel anomaly detection method provided in one embodiment of this application.

[0046] Figure 4 This is a flowchart of step S5 in the power transmission and distribution line channel anomaly detection method provided in one embodiment of this application.

[0047] Figure 5 This is a flowchart of step S6 in the power transmission and distribution line channel anomaly detection method provided in one embodiment of this application.

[0048] Figure 6 This is a flowchart of step S8 in the power transmission and distribution line channel anomaly detection method provided in one embodiment of this application.

[0049] Figure 7 This is a schematic diagram of the structure of a power transmission and distribution line channel anomaly detection device provided in another embodiment of this application.

[0050] Figure 8 This is a schematic diagram of the prior construction module of the channel structure space in the power transmission and distribution line channel anomaly detection device provided in another embodiment of this application.

[0051] Figure 9 This is a schematic diagram of the prior multi-scale alignment and gating modulation module in a power transmission and distribution line channel anomaly detection device provided in another embodiment of this application.

[0052] Figure 10 This is a schematic diagram of the structure of the detection head module and the scale adaptive distribution regression training module in the power transmission and distribution line channel anomaly detection device provided in another embodiment of this application.

[0053] Figure 11 This is a schematic diagram of the inference calibration and post-processing module in the power transmission and distribution line channel anomaly detection device provided in another embodiment of this application.

[0054] Figure 12 This is a schematic diagram of the alarm evaluation and output module in the power transmission and distribution line channel anomaly detection device provided in another embodiment of this application. Detailed Implementation

[0055] To make the objectives and technical solutions of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.

[0056] In one embodiment, see Figure 1 This application provides a method for detecting anomalies in power transmission and distribution line channels, which may include the following steps: steps S1 to S7.

[0057] Step S1: Preprocess the image to be analyzed to obtain input data.

[0058] Step S2: Based on the structural characteristics of the power transmission and distribution line corridor, extract the structural information of the effective area of ​​the corridor from the input data and generate a multi-element structural field prior map.

[0059] Step S3: Construct an object detection network by inputting the input data into the object detection network to obtain a multi-scale feature set.

[0060] Step S4: Spatial registration is performed between the multi-dimensional structure field prior map and the features at each scale to generate a prior gating map. Gating modulation is then performed based on the prior gating map to obtain enhanced features.

[0061] Step S5: Perform forward bounding box regression based on enhanced features, and output candidate bounding boxes and their corresponding regression uncertainties.

[0062] Step S6: Based on the prior graph of the multivariate structure field, supervise the construction and regularization of the candidate bounding box by scale-adaptive distribution regression to obtain the spatial consistency score and calibration confidence.

[0063] Step S7: Perform deterministic filtering based on the spatial consistency score, regression uncertainty, and calibration confidence, and perform suppression processing on the filtered candidate bounding boxes to obtain the final detection result.

[0064] In the power transmission and distribution line channel anomaly detection method of this application, a multi-element structure field prior map is constructed, and prior gating modulation is performed in the multi-scale feature flow of the target detection network. This makes the target detection network more focused on the effective area of ​​the corridor during training and inference, thereby reducing false alarms caused by complex backgrounds and reducing interference from invalid areas. Through a scale-adaptive distributed bounding box regression optimization mechanism, the noise resistance and stability of small target regression are improved and the high IoU positioning accuracy is enhanced, achieving high-precision detection and alarm of abnormal targets in the channel. While ensuring alarm accuracy, it also takes into account missed detection control and adaptability to complex operating conditions. By using gating modulation to enhance or retain the feature response of the effective area of ​​the corridor and suppress the feature response of the invalid area outside the corridor, false alarms caused by background interference are reduced and the detection capability of abnormal targets in the channel is improved. Through multi-dimensional index deterministic filtering and suppression processing, false detection results can be effectively eliminated, significantly improving the accuracy, reliability and robustness of power transmission and distribution line channel anomaly detection, and providing strong protection for the safe and stable operation of power transmission and distribution lines.

[0065] In step S1, please refer to Figure 1 In step S1, the image to be analyzed is preprocessed to obtain input data.

[0066] Specifically, the image to be analyzed is acquired, and the image to be analyzed is preprocessed to obtain input data. The preprocessing may include normalization and data augmentation.

[0067] As an example, the images to be analyzed include, but are not limited to, images from drone inspections, images from fixed surveillance cameras, and images extracted from video streams.

[0068] As an example, the drone inspection images can be acquired by camera equipment mounted on the drone; the fixed monitoring images can be acquired by fixed monitoring equipment; and the video stream frame-sampling images can be obtained by extracting frames from the original video stream acquired by the mobile inspection terminal or video acquisition equipment at a preset frame rate or time interval.

[0069] As an example, the camera device could be a visible light camera.

[0070] As an example, the input data can be an RGB color image, input in RGB three-channel format.

[0071] For example, the input size can be set to... To analyze the image Perform proportional scaling to normalize the input size. Proportional scaling factor. for:

[0072]

[0073] in, These are the height and width of the image to be analyzed, respectively. These are the input height and width, respectively.

[0074] Furthermore, the scaled image is padded to obtain the initial input image. It also records the scaling factor and fill offset. The fill offset can be set as follows:

[0075]

[0076]

[0077] in, These are the fill offsets in the horizontal and vertical directions, respectively, in pixels; This is the scaling factor; These are the height and width of the image to be analyzed, respectively. These are the input height and width, respectively.

[0078] Furthermore, for the initial input image Linear normalization is performed to obtain the normalized input data, with a pixel range of [missing information]. .

[0079] Furthermore, the normalized input data can be randomly transformed (e.g., rotated, flipped, or color-dithered) to obtain the input data. This is to increase the diversity of training samples, improve the generalization ability of the model, prevent overfitting, and achieve data augmentation.

[0080] In step S2, please refer to Figure 1 In step S2, based on the structural characteristics of the power transmission and distribution line channel, the structural information of the effective area of ​​the corridor is extracted from the input data to generate a multi-element structural field prior map.

[0081] For example, please refer to Figure 2 Step S2 may include the following steps: Step S21 to Step S25.

[0082] Step S21: Perform edge extraction and line segment detection on the input data to obtain a set of line segments. Statistically analyze or cluster the direction angles of the line segments to determine the main direction of the corridor. Fit the set of line segments based on the main direction of the corridor to obtain the center line of the corridor.

[0083] Step S22: Determine the corridor half-width function based on perspective effect.

[0084] Step S23: Determine the probability field of the corridor area based on the distance from the pixel to the center line of the corridor.

[0085] Step S24: Determine the corridor region based on the corridor region probability field, calculate the distance from each pixel to the corridor region, and obtain the distance field.

[0086] Step S25: Calculate the gradient direction on the distance field and take the unit vector pointing to the direction of decreasing distance to obtain the direction field.

[0087] As an example, the structural features of the power transmission and distribution line corridor may include linear features of the corridor centerline, strip-shaped distribution features of the corridor area, width gradient features caused by perspective effect, and spatial continuity features.

[0088] As an example, the multivariate structure field prior map may include one or more of the following: a corridor region probability field, a distance field map to the corridor centerline or boundary, and a direction field map of the traverse or the main direction of the corridor. The multivariate structure field prior map is not a binary region of interest (ROI), but a continuous field that expresses both "whether it is in the corridor" and "how far it is from the corridor and in which direction it returns to the corridor," for subsequent gating and calibration.

[0089] As an example, the probability field of the corridor area, the distance field map to the center line / boundary of the corridor, and the direction field map of the main direction of the guide / corridor are spatially aligned with the input data in the pixel coordinate system.

[0090] The following section uses a ternary structured field prior diagram, which includes the probability field of the corridor region, the distance field diagram to the center line of the corridor, and the direction field diagram of the main direction of the traverse / corridor, as an example to specifically introduce steps S2 to S7.

[0091] Specifically, in step S21, the input data can be processed. Grayscale conversion, smoothing, and edge extraction are performed to obtain the edge image.

[0092] As an example, the image undergoes grayscale conversion using one or more combinations of component weighting, median filtering, or extreme value filtering based on the RGB color space to transform the color image into a single-channel grayscale image. Next, the grayscale image is smoothed and denoised using linear or nonlinear filtering algorithms such as mean filtering, median filtering, Gaussian filtering, or bilateral filtering, or a multi-scale filtering fusion approach, to suppress image noise while preserving image edges and details as much as possible. Finally, edge extraction is performed on the smoothed grayscale image, employing one or more fusion detection methods from gradient operator classes (Sobel, Prewitt, Roberts), Laplacian operator classes, and Canny edge detection operators. By calculating the gradient of changes in image pixel grayscale values, determining gradient extrema, and setting edge response thresholds, accurate extraction of target edges in the image is achieved. Furthermore, the algorithm parameters for each step can be adaptively adjusted and optimized according to the noise characteristics and edge detail requirements of the actual image.

[0093] Furthermore, line segment detection is performed on the edge map to obtain a set of line segments, and the direction angles of the line segments are statistically analyzed or clustered to determine the main direction of the corridor.

[0094] As an example, the line segment detection can be probabilistic Hough line segment detection, and the statistics can be main peak statistics.

[0095] Furthermore, the set of line segments is fitted based on the main direction of the corridor, assuming the line segments The direction angle is The main direction of the corridor is Select the one that satisfies The line segments are used as the set of supporting line segments, where, A preset angle threshold is used to constrain the consistency between the support line segment and the main direction of the corridor.

[0096] As an example, to adapt to different viewing angles and noise conditions, the preset angle threshold... It can be adaptively determined based on the robust dispersion of the orientation angle deviation, and the expression is:

[0097]

[0098] in, .

[0099] As an example, a preset angle threshold It can be set to Preferred .

[0100] Furthermore, robust straight-line fitting is performed on the set of support segments to obtain the corridor centerline, the expression of which is:

[0101]

[0102]

[0103] in, For image coordinates, These are the fitting parameters.

[0104] As an example, the centerline of the corridor can also be normalized and parameterized to simplify calculations and facilitate distance measurement.

[0105] Furthermore, in step S22, considering the perspective effect, the corridor is wider at the bottom of the image and narrower at the top. The corridor width varies with the vertical position of the image. The half-width of the corridor is inversely proportional to the normalized ordinate and is limited to a preset minimum half-width. With preset maximum half width Between, i.e., truncating the boundary, such that the half-width of the corridor in the foreground region of the image is greater than the half-width of the corridor in the background region of the image, then the corridor half-width function... for:

[0106]

[0107] in, Input data height; This is the scaling factor; It is a very small constant; These are the preset minimum half widths. With preset maximum half width .

[0108] As an example, the corridor half-width function makes the corridor width vary with the longitudinal direction, which conforms to the imaging patterns of drones or surveillance.

[0109] Further, in step S23, the distance from each pixel to the center line of the corridor is calculated. The expression is:

[0110]

[0111] in, For image coordinates, These are the fitting parameters.

[0112] Furthermore, based on the normalized distance from each pixel to the center line of the corridor, a probability field for the corridor region is generated according to a continuous function that monotonically decreases with distance. This ensures that the corridor probability of pixels near the corridor centerline is greater than that of pixels outside the corridor, and that the corridor probability changes continuously in space. The corridor region probability field... The expression is:

[0113]

[0114] in, The attenuation coefficient is... ; This is a function for the half-width of the corridor. for Distance to the center line of the corridor; It is a very small constant; .

[0115] Furthermore, in step S24, the probability field of the corridor area is set to be greater than or equal to a preset probability threshold. The pixels in the input data represent the corridor region. The distance from any pixel in the input data to the corridor region is calculated to obtain the distance field. .

[0116] As an example, the preset probability threshold The threshold used to control the spatial extent of the corridor area can be adaptively determined based on the attenuation coefficient of the corridor area probability field and the corridor half-width function in step S23, so that the threshold corresponds to "the distance from the pixel to the corridor centerline is equal to the corridor half-width". The probability value at the given location, i.e., taking This allows the corridor bandwidth and half-width to be maintained under different viewing angles and width variations. Consistency is key; avoid over-segmentation or under-segmentation.

[0117] For example, when the probability field adopts an exponential decay form, a preset probability threshold can be used. ,in This is the attenuation coefficient in step S23.

[0118] As an example, the preset probability threshold It can be set to In this specific embodiment, it is preferred that... .

[0119] As an example, the distance from any pixel in the input data to the corridor region can be an Euclidean distance.

[0120] As an example, step S24 may also include truncating and normalizing the range field to obtain a normalized range field. This is to eliminate extreme value interference in areas far from the corridor, so that it has a uniform dimension and numerical range.

[0121] Further, in step S25, the gradient direction is calculated on the distance field and a unit vector pointing towards the direction of decreasing distance is taken to obtain the direction field. Finally, the prior diagram of the ternary structure field was obtained. The ternary structure field prior map subsequently participates in both network internal control and inference output calibration, forming a closed loop. This is the key difference between this technical solution and solutions that only use binary Region of Interest (ROI) clipping or only use post-processing sieving.

[0122] As an example, the orientation field The expression is:

[0123]

[0124] in, For gradient operators, It is a 2-norm. for The distance field of a pixel. It is a very small constant.

[0125] In step S3, please refer to Figure 1 In step S3, a target detection network is constructed by inputting the input data into the target detection network to obtain a multi-scale feature set.

[0126] Specifically, a target detection network is constructed to adapt to both small distant targets and large near targets.

[0127] As an example, the target detection network may include a backbone network and a neck network.

[0128] Furthermore, the input data is fed into the target detection network to obtain a multi-scale feature set.

[0129] As an example, the step size of the feature layers at different scales in the multi-scale feature set can be set to... The corresponding feature map is used for subsequent detection head output of categories and regression results.

[0130] It should be noted that this embodiment does not limit the specific implementation form of the backbone network and neck network, and can adopt the backbone and feature pyramid structure of the existing one-stage detection framework.

[0131] In step S4, please refer to Figure 1 In step S4, the prior map of the multi-dimensional structure field is spatially registered with the features at each scale to generate a prior gating map. Based on the prior gating map, gating modulation is performed to obtain enhanced features.

[0132] For example, please refer to Figure 3Step S4 may include the following steps: Step S41 to Step S43.

[0133] Step S41: Align the multi-scale structure field prior map to the spatial resolution of each scale feature in the multi-scale feature set to obtain the scale-aligned prior.

[0134] Step S42: Map the scale-aligned prior to a prior gating graph.

[0135] Step S43: Perform affine gating modulation on each scale feature based on the prior gating map to obtain enhanced features.

[0136] Specifically, in step S41, in order to ensure that the channel structure prior can exert a consistent constraint and enhancement effect on feature layers of different scales, this embodiment uses the multi-element structure field prior map. Spatial registration with feature layers at various scales is performed to obtain the prior map of the multi-element structure field. According to size Perform scaling and alignment to obtain the scale alignment prior. The expression is:

[0137]

[0138] in, Indicates the scaling method. These represent the height and width of the input data, respectively.

[0139] As an example, the scaling method can employ bilinear interpolation scaling to ensure geometric alignment.

[0140] Furthermore, in step S42, the scale is aligned with the prior. Mapped to a priori gating graph This allows for gated modulation of multi-scale feature flows, thereby enabling explicit guidance of effective corridor regions and suppression of interference from ineffective regions outside the corridors within the network. The prior gating graph... The expression is:

[0141]

[0142] in, Representing scale-aligned priors respectively The three components, For the Sigmoid function, These represent the learnable scalar parameters, Let the main direction of the corridor be the unit vector. The vector dot product is used to measure directional consistency.

[0143] As an example, the prior gating graph probability field of corridor region (i.e., whether the area is within the corridor), normalized distance field (i.e., how far is the area from the corridor), directional field The determination of whether the area points to the corridor (i.e., whether the area points to the corridor) together can avoid false suppression or false suppression caused by relying on a single mask.

[0144] As an example, the prior gating graph The pixel values ​​are continuous values ​​between 0 and 1, and the prior gating map The following conditions must be met: the higher the corridor probability, the larger the gate value; the larger the normalized distance, the smaller the gate value; and the higher the consistency between the direction field and the main direction of the corridor, the larger the gate value.

[0145] As an example, the prior gating diagram is a single-channel door air conditioner.

[0146] Furthermore, in step S43, based on the prior gating graph For each scale feature Affine gating modulation is performed to obtain enhanced features. The expression is:

[0147]

[0148] in, For element-wise multiplication (broadcast to the channel dimension); It is a learnable scalar.

[0149] As an example, the affine modulation includes: gating a priori gating graphs. As spatial weights for corresponding scale features Multiplicative enhancement and gating-related additive bias are applied to enhance or preserve the feature responses within the effective area of ​​the corridor and suppress the feature responses outside the effective area of ​​the corridor, thereby reducing false alarms in complex backgrounds and improving the detection rate of abnormal targets in the corridor.

[0150] In step S5, please refer to Figure 1 In step S5, a forward operation for bounding box regression is performed based on the enhanced features, outputting candidate bounding boxes and their corresponding regression uncertainties.

[0151] As an example, the bounding box regression can be a distributed regression.

[0152] For example, please refer to Figure 4 Step S5 may include the following steps: Step S51 to Step S53.

[0153] Step S51: Input the enhanced features into the detection head, and output the class probability and the discrete distribution of the four sides of the bounding box.

[0154] Step S52: Decode the discrete distribution of the four sides of the bounding box to obtain the candidate bounding box.

[0155] Step S53: Calculate the regression uncertainty of the candidate bounding box.

[0156] Specifically, in step S51, the enhanced features are... Input the detector head, output the class probability. .

[0157] As an example, if the category set (Including background), the output matching score is Then the category Category probability for: .

[0158] Furthermore, perform [further steps] on each of the four sides. Discretely divide the bounding box into several intervals (bins) to obtain the discrete distribution of the four sides. That is, predicting the distribution. The expression is:

[0159]

[0160]

[0161] in, For the largest discrete interval index, Indicates the first Each interval These represent the left margin, top margin, right margin, and bottom margin of the bounding box, respectively. Indicates the first The edge at the 1st Predicted probabilities over discrete intervals.

[0162] Furthermore, in step S52, the discrete distribution of the four sides of the bounding box is... Perform expectation decoding to obtain continuous margins. For the continuous edge distance By step size Converted to pixel scale Candidate bounding box B is determined based on the current center point, where, , respectively representing left, top, right, and bottom.

[0163] As an example, the continuous margin The expression is:

[0164]

[0165] in, For the largest discrete interval index, For the first The edge at the 1st Predicted probabilities over discrete intervals .

[0166] As an example, the pixel scale The expression is:

[0167]

[0168] in, For continuous margins, Step size, .

[0169] For example, if the center point of the current location is The bounding box is then:

[0170]

[0171] in, In pixels .

[0172] Furthermore, in step S53, for continuous edge spacing... Calculate variance Normalization yields regression uncertainty. ,in, It is used for inference calibration and alarm stability.

[0173] As an example, variance The expression is:

[0174]

[0175] in, Indicates the first The edge at the 1st Predicted probabilities over discrete intervals; This indicates the continuous margin of the edge.

[0176] As an example, regression uncertainty The expression is:

[0177]

[0178] in, Represents variance. For the largest discrete interval index, .

[0179] As an example, regression uncertainty The larger the value, the more dispersed the distribution and the less stable the location. Regression uncertainty. It will directly participate in confidence calibration and candidate box filtering during the inference stage, demonstrating the value of scale-adaptive distribution regression for noise reduction of small targets from the output mechanism level.

[0180] In step S6, please refer to Figure 1 In step S6, the candidate bounding box is constructed and regularized based on the prior graph of the multivariate structure field by supervising the scale-adaptive distribution regression, and the spatial consistency score and calibration confidence are obtained.

[0181] For example, please refer to Figure 5 Step S6 may include the following steps: Step S61 to Step S66.

[0182] Step S61: Determine the target scale based on the real bounding box.

[0183] Step S62: Establish a mapping function from the target scale to the temperature parameter, and construct a soft label distribution based on the mapping function from the target scale to the temperature parameter.

[0184] Step S63: Determine the distributional regression loss based on the soft-label distribution.

[0185] Step S64: Set a distribution smoothing term, establish a mapping function from the target scale to the regularization weights, determine the regularization loss based on the mapping function from the target scale to the regularization weights, and obtain the total loss.

[0186] Step S65: Calculate the corridor probability consistency, distance consistency, and orientation consistency of pixels inside the candidate bounding box, and obtain the spatial consistency score through weighted fusion.

[0187] Step S66: Use the spatial consistency score and the regression uncertainty to calibrate the original category confidence score to obtain the calibrated confidence score.

[0188] Specifically, in step S61, based on the real frame Determine the target scale .

[0189] As an example, the target scale It can be scaled as the square root of the area, with units in pixels, and the expression is:

[0190]

[0191] in, , These represent the coordinates of the actual bounding box.

[0192] Furthermore, in step S62, the target scale is established. to temperature parameters mapping function The expression is:

[0193]

[0194] in, These represent the minimum and maximum values ​​of the temperature parameter, respectively. ; It is a scale decay constant.

[0195] As an example, the mapping function from the target scale to the temperature parameter It can determine the width of the soft label and the target size of small targets. Small, then Larger means smoother supervision; larger target scale Large, then Smaller means more acute supervision.

[0196] Furthermore, based on the mapping function from the target scale to the temperature parameter... Constructing soft label distribution .

[0197] For example, the bin size of the true left margin can be denoted as... Then the left-side scale adaptive soft label distribution The expression is:

[0198]

[0199] in, For the largest discrete interval index, Indicates the first Each interval Indicates the summation index.

[0200] As an example, the specific method for constructing the right, top, and bottom scale-adaptive soft label distribution can refer to the specific method for constructing the left scale-adaptive soft label distribution in step S62, which will not be repeated here.

[0201] Further, in step S63, based on the soft label distribution Deterministic regression loss The expression is:

[0202]

[0203] in, For the first The edge at the 1st Predicted probabilities over discrete intervals For soft label distribution, For the largest discrete interval index, Indicates the first Each interval .

[0204] Further, in step S64, a distribution smoothing term is set. This is to suppress overfitting of small targets in the later stages. The distribution smoothing term... The expression is:

[0205]

[0206] in, For the first The edge at the 1st Predicted probabilities over discrete intervals For the largest discrete interval index, Indicates the first Each interval .

[0207] Furthermore, establish target scales. To regularization weights mapping function According to the mapping function from the target scale to the regularized weights Determine the regular loss The expression is:

[0208]

[0209] in, This is the distribution smoothing term.

[0210] As an example, the mapping function from the target scale to the regularized weights The expression is:

[0211]

[0212] in, These represent the minimum and maximum values ​​of the regularization weight, respectively. For the target scale, Let be the scale constant. The mapping function from the target scale to the regularization weights for small targets. Larger targets exhibit stronger ability to smoothly suppress noise collapse; the mapping function from target scale to regularized weights for larger targets. Smaller sizes allow for more precise positioning by reducing constraints.

[0213] As an example, step S64 may also include classification loss. and positioning loss The classification loss can be cross-entropy with Focal modulation, and the localization loss can be IoU class loss (such as CIoU), denoted as... .

[0214] As an example, the classification loss The expression is:

[0215]

[0216] in, For one-hot tags; For Focal parameters, Indicate category The class probability, It is a set of categories.

[0217] Furthermore, the total loss can be obtained. for:

[0218]

[0219] in, For classifying losses, To pinpoint the loss, For the loss of cloth return, For normal loss, It is a fixed-weight constant.

[0220] As an example, steps S62 to S64 explicitly define the specific implementation method of scale adaptation: supervising the mapping function from the target scale to the temperature parameter for the width of the distribution. The determination of the smoothing regularization strength is achieved by the mapping function from the target scale to the regularization weights. The decision can reduce regression overfitting caused by small target annotation errors and unclear boundaries, and improve the localization quality in high IoU intervals.

[0221] Further, in step S65, the corridor probability consistency of pixels inside the candidate bounding box is calculated respectively. Distance consistency Consistency of direction Spatial consistency scores are obtained through weighted fusion. The corridor probability consistency Candidate bounding boxes Internal pixel probability field in corridor region Average value; distance consistency Candidate bounding boxes Internal pixels in normalized distance field Average value; directional consistency Candidate bounding boxes The direction field at the center point is consistent with the direction pointing towards the center line of the corridor.

[0222] As an example, the corridor probability consistency The expression is:

[0223]

[0224] in, Candidate bounding boxes The set of pixels, that is, all pixels within the candidate bounding box; For area; probability field for corridor region .

[0225] As an example, the distance consistency The expression is:

[0226]

[0227] in, For the normalized distance field, Candidate bounding boxes The set of pixels, For area.

[0228] As an example, setting candidate bounding boxes The center point is The centerline normal is The unit vector pointing to the corridor is Then the direction consistency The expression is:

[0229]

[0230] in, Indicates the center point directional field, Indicates the center point Distance to the center line of the corridor.

[0231] As an example, the spatial consistency score The expression is:

[0232]

[0233] in, , The fusion coefficient is... + .

[0234] Further, in step S66, the spatial consistency score is utilized. With the aforementioned regression uncertainty Confidence of the original category Perform calibration to obtain calibration confidence level. The expression is:

[0235]

[0236] in, The consistency enhancement coefficient, ; The uncertainty penalty coefficient, .

[0237] As an example, the original category confidence score The expression is:

[0238]

[0239] in, Indicates the candidate bounding box category The class probability.

[0240] As an example, the calibration confidence is obtained by jointly calibrating the original class confidence using the monotonic gain of the spatial consistency score and the monotonic penalty of the regression uncertainty. A higher spatial consistency score results in a higher calibration confidence, while higher regression uncertainty results in a lower calibration confidence. Even if high original class confidence occurs in areas outside the corridor... False positives can also affect spatial consistency scores. Low visibility and suppression; candidate bounding boxes with high uncertainty due to extremely small distant targets or unclear boundaries will be suppressed. Suppressed output makes the output more stable and the alarms more reliable.

[0241] As an example, the inference phase combines the original classification confidence of the candidate boxes with the consistency of the channel structure and the regression uncertainty to perform deterministic calibration, screening and suppression. This forms a closed loop of gating and calibration with the prior calibration, which can reduce false alarms outside the corridor and improve alarm reliability.

[0242] As an example, the "width" and "smoothing" intensity of the supervision distribution are adaptively adjusted for different target scales, thereby achieving more stable and noise-resistant regression for small targets and more detailed localization for large targets.

[0243] As an example, scale-adaptive distribution regression must satisfy the following: the smoothing width of the supervised distribution is determined based on the true target scale, and the smaller the true target scale, the smoother the supervised distribution; and the distribution smoothing regularity is determined based on the true target scale, and the distribution smoothing regularity is stronger when the true target scale is smaller and weaker when the true target scale is larger.

[0244] In step S7, please refer to Figure 1 In step S7, deterministic filtering is performed based on the spatial consistency score, regression uncertainty, and calibration confidence, and suppression processing is performed on the filtered candidate bounding boxes to obtain the final detection result.

[0245] Specifically, based on the spatial consistency score Regression uncertainty Calibration confidence Candidate bounding boxes must satisfy the first condition. Deterministic filtering is performed based on the first condition to obtain the filtered candidate bounding boxes.

[0246] As an example, the first condition is:

[0247]

[0248]

[0249]

[0250] in, This is the system threshold constant. The spatial consistency score. To return to uncertainty, To calibrate the confidence level.

[0251] Furthermore, suppression processing is performed on similar candidate bounding boxes within the filtered candidate bounding boxes to obtain suppressed candidate bounding boxes.

[0252] As an example, the continuous processing can employ Soft-NMS suppression, expressed as:

[0253]

[0254] in, Preserve the current bounding box; The bounding box to be suppressed; To suppress the scale constant, To calibrate the confidence level.

[0255] Furthermore, based on the scaling factor and padding offset recorded in step S1, the coordinates of the suppressed candidate bounding boxes are... The image coordinates are then remapped back to the original image coordinate system to obtain the final output bounding box. The remapping process is as follows:

[0256]

[0257]

[0258]

[0259]

[0260] in, These are the fill offsets in the horizontal and vertical directions, respectively, in pixels; This is the scaling factor; This is the final output bounding box.

[0261] In one specific embodiment, the power transmission and distribution line channel anomaly detection method of this application may further include step S8: calculating a risk score based on the calibration confidence, spatial consistency score and the degree of intrusion of the candidate bounding box relative to the corridor, determining the alarm level based on the risk score, and outputting the final result.

[0262] For example, please refer to Figure 6 Step S8 may include the following steps: Step S81 to Step S82.

[0263] Step S81: Determine the degree of intrusion based on the distance of each pixel to the center line of the corridor and the corridor half-width function.

[0264] Step S82: Determine the risk score based on the degree of intrusion, determine the alarm level based on the risk score, and output the final result.

[0265] Specifically, in step S81, the degree of intrusion can be calculated by strongly binding it to the corridor geometry. This is based on the distance of each pixel from the center line of the corridor. With corridor half-width function Determine the extent of the intrusion , The expression is:

[0266]

[0267] in, for The distance from the point to the center line of the corridor for The corridor half-width function at the location, It is a very small constant. Intrusion level The larger the value, the closer to the center of the corridor and the higher the risk; the degree of intrusion outside the corridor. .

[0268] As an example, for the set of exception categories A fixed weight table with preset values ​​can be used. Then, during output, the weights of the abnormal categories can be obtained by looking up a table by category. .

[0269] As an example, anomalous targets may include at least one of the following: engineering vehicles, cranes, tower cranes, smoke / burning, and bird nests.

[0270] Furthermore, in step S82, based on the degree of intrusion... Calibration confidence Spatial consistency score Determine the risk score The expression is:

[0271]

[0272] in, For fixed constants, ; For abnormal categories The weight.

[0273] Furthermore, risk score thresholds can be set as follows: Then it can be based on the risk score Determine the alarm level and output the final result.

[0274] As an example, the risk score threshold can be set. .

[0275] As an example, the alarm levels may include Level I Emergency Alarm, Level II High-Risk Alarm, Level III General Alarm, and Level IV Warning Alarm. Specifically, when the risk score... When a Level I emergency alarm is triggered; When a Level II high-risk alarm is triggered; When a Level III general alarm is triggered; A Level IV alert is triggered at this time; if If so, only logging will be performed without triggering an alarm.

[0276] In one example, for a video stream frame-by-frame image, a trajectory is formed by cross-frame correlation of similar bounding boxes using IoU, and an exponential moving average is applied to the risk score. The expression is:

[0277]

[0278] in, For a fixed smoothing coefficient, ; Let t be the risk score at time t. An alarm must be triggered if the condition is continuous. frame ,in, It is a fixed constant.

[0279] In one example, the final output can be a set of fields, which may include, but are not limited to, data identification information, detection result information, risk results, and visual evidence.

[0280] As an example, the data identification information may include, but is not limited to, camera identifiers. Image number or frame number and timestamp .

[0281] As an example, the detection result information may include, but is not limited to, categories. The final output is the bounding box coordinates. Original category confidence With calibration confidence Corridor probability consistency Distance consistency Consistency of direction Spatial consistency score With regression uncertainty .

[0282] As an example, the risk outcome may include, but is not limited to, the degree of intrusion. Risk score And the corresponding alarm level.

[0283] As an example, the visual evidence may include, but is not limited to, the original image with boxed annotations and the overlay of corridor probability heatmaps, for alarm interpretation and verification.

[0284] As an example, the method of this application has a clear process and well-defined modules, which can be integrated with existing target detection frameworks, making it easy to implement in engineering and facilitating future expansion in terms of alarm strategies, model lightweighting, and deployment optimization.

[0285] In the transmission and distribution line channel anomaly detection method of this application, by simultaneously applying the channel structure prior in the form of a continuous structure field to network feature learning and inference calibration, and introducing scale-adaptive distribution regression with uncertainty constraints, false alarms outside the corridor and the stability and reliability of distant small target detection and alarm can be significantly reduced. By enhancing or retaining the feature response of the effective area of ​​the corridor and suppressing the feature response of the ineffective area outside the corridor through gating modulation, false alarms caused by background interference are reduced and the detection capability of abnormal targets within the channel is improved. By applying differentiated distribution supervision / constraints to targets of different scales through scale-adaptive distribution regression, the regression of small targets is more stable and noise-resistant, and the localization of large targets is more precise. This can alleviate the rebound of regression loss at the validation end caused by overfitting of the regression distribution in the later stages of training, and improve high I-values ​​such as mAP50-95. The oU interval performance can explicitly utilize channel structure characteristics to suppress interference from invalid regions, improve the detection capability of distant small targets, enhance detection stability under low light, haze, and backlight conditions, and improve the detection of missed distant small targets, thereby achieving high-precision and low-false-alarm anomaly detection and alarm. By outputting candidate bounding boxes based on enhanced features and regression uncertainty, combined with supervised construction and regularization constraints of scale-adaptive distribution regression using multivariate structure field prior maps, the bounding box regression error can be reduced, and the spatial consistency and confidence calibration accuracy of candidate boxes can be improved. Through deterministic filtering and suppression processing using multi-dimensional indicators, false detection results can be effectively eliminated, significantly improving the accuracy, reliability, and robustness of anomaly detection in power transmission and distribution line channels, providing strong protection for the safe and stable operation of power transmission and distribution lines.

[0286] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the accompanying drawings may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0287] In another embodiment, please refer to Figure 7This application also provides a transmission and distribution line channel anomaly detection device, which may include: a preprocessing module 1, a channel structure space prior construction module 2, a prior multi-scale alignment and gating modulation module 3, a detection head module 4, a scale adaptive distribution regression training module 5, an inference calibration and post-processing module 6, and an alarm evaluation and output module 7. The preprocessing module 1 preprocesses the image to be analyzed to obtain input data; the channel structure space prior construction module 2 generates a multi-dimensional structure field prior map based on the input data output by the preprocessing module 1; the prior multi-scale alignment and gating modulation module 3 extracts and fuses features from the input data to obtain a multi-scale feature set, performs prior alignment and gating modulation on the multi-dimensional structure field prior map, and outputs an enhanced feature map. The enhanced features are input into the detection head module 4, which calculates the category probability and the discrete distribution of the four sides of the bounding box, and performs expected decoding to output candidate bounding boxes and regression uncertainty. The scale-adaptive distribution regression training module 5 performs supervised construction and regularization constraints on the candidate bounding boxes for scale-adaptive distribution regression, obtains spatial consistency score and calibration confidence, and calculates the total loss. The inference calibration and post-processing module 6 calculates consistency index, confidence calibration, deterministic filtering, and coordinate back mapping based on the candidate bounding boxes and regression uncertainty, and outputs the final output bounding box. The alarm assessment and output module 7 calculates risk score and classifies alarm levels based on calibration confidence and corridor half-width, and outputs graded alarm information and final results.

[0288] As an example, preprocessing module 1 performs preprocessing operations such as proportional scaling, boundary padding, and normalization on the image to be analyzed, resulting in the network input image. The proportional scaling factor is calculated and stored during the preprocessing process. With fill offset This is used for subsequent output coordinate remapping.

[0289] As an example, the multi-element structure field prior map may include one or more of the following: a corridor region probability field, a distance field to the corridor centerline or boundary, and a direction field of the conductor or the main direction of the corridor. The multi-element structure field prior map is aligned with the input data in pixel coordinates.

[0290] For example, please refer to Figure 8The channel structure space prior construction module 2 includes a centerline extraction unit 21, a corridor half-width modeling unit 22, a corridor region probability field generation unit 23, a distance field generation unit 24, and a direction field generation unit 25. The centerline extraction unit 21 performs edge detection and line segment extraction on the input data to obtain a set of line segments, performs statistical analysis or clustering on the direction angles of the line segments to determine the main direction of the corridor, and fits the set of line segments to obtain the centerline of the corridor. The corridor half-width modeling unit 22 constructs a corridor half-width function based on the set of line segments output by the centerline extraction unit 21, the main direction of the corridor, and the corridor centerline, and outputs the corridor half-width corresponding to each longitudinal position. The corridor region probability field generation unit 23 calculates the distance from each pixel to the corridor centerline and outputs the corridor region probability field. The distance field generation unit 24 determines the corridor region based on the corridor region probability field, calculates the distance from any pixel in the input data to the corridor region to obtain the distance field, and truncates and normalizes the distance field to obtain a normalized distance field. The direction field generation unit 25 determines the direction field based on the gradient of the normalized distance field.

[0291] For example, please refer to Figure 9 The prior multi-scale alignment and gating modulation module 3 may include: a multi-scale feature set construction unit 31, a prior alignment unit 32, and a gating modulation unit 33. The multi-scale feature set construction unit 31; the prior alignment unit 32 aligns the multi-element structure field prior map to the spatial resolution of each scale feature in the multi-scale feature set to obtain the scale-aligned prior; the gating modulation unit 33 maps the scale-aligned prior to a prior gating map, and performs affine gating modulation on each scale feature based on the prior gating map to obtain the corresponding enhanced feature.

[0292] As an example, the prior multi-scale alignment and gating modulation module 3 can enhance the feature response of the effective area of ​​the corridor and suppress the feature response of the area outside the corridor.

[0293] For example, please refer to Figure 10 The detection head module 4 may include a detection head unit 41, an expectation decoding unit 42, and a regression uncertainty calculation unit 43. The detection head unit 41 receives enhanced features and outputs the class probability and the discrete distribution of the four sides of the bounding box. The expectation decoding unit 42 performs expectation decoding on the discrete distribution of the four sides of the bounding box and outputs candidate bounding boxes. The regression uncertainty calculation unit 43 calculates the regression uncertainty of the candidate bounding boxes.

[0294] As an example, please continue reading Figure 10The scale-adaptive distribution regression training module 5 may include a soft-label distribution construction unit 51, a distribution regression loss unit 52, a regularization loss unit 53, and a total loss unit 54. The soft-label distribution construction unit 51 establishes a mapping function from the target scale to the temperature parameter based on the target scale and outputs the soft-label distribution. The distribution regression loss unit 52 calculates the distribution regression loss based on the soft-label distribution. The regularization loss unit 53 determines the regularization loss based on the mapping function from the distribution smoothing term to the target scale to the regularization weight. The total loss unit 54 performs weighted fusion of the distribution regression loss, regularization loss, classification loss, and localization loss and outputs the total loss.

[0295] For example, please refer to Figure 11 The inference calibration and post-processing module 6 may include a consistency index calculation unit 61, a confidence calibration unit 62, a deterministic filtering unit 63, a suppression processing unit 64, and a coordinate back-mapping unit 65. The consistency index calculation unit 61 calculates a consistency index for candidate bounding boxes, which may include corridor probability consistency, distance consistency, and orientation consistency. The confidence calibration unit 62 uses the consistency index and regression uncertainty to calibrate the original category confidence and outputs the calibration confidence. The deterministic filtering unit 63 performs deterministic filtering based on the consistency index, regression uncertainty, and calibration confidence and outputs the filtered candidate bounding boxes. The suppression processing unit 64 performs suppression processing on the same type of candidate bounding boxes in the filtered candidate bounding boxes and outputs the suppressed candidate bounding boxes. The coordinate back-mapping unit 65 back-maps the coordinates of the suppressed candidate bounding boxes back to the original image coordinate system based on the recorded proportional scaling factor and fill offset to obtain the final output bounding boxes.

[0296] As an example, please continue reading Figure 12 The alarm assessment and output module 7 may include an intrusion degree calculation unit 71, a risk score calculation unit 72, and an alarm assessment unit 73. The intrusion degree calculation unit 71 determines the intrusion degree based on the distance from each pixel to the center line of the corridor and the corridor half-width function; the risk score calculation unit 72 determines the risk score based on the intrusion degree; and the alarm assessment unit 73 determines the alarm level based on the risk score.

[0297] In another specific embodiment, the power transmission and distribution line channel anomaly detection device of this application may further include a result encapsulation and publishing module, which generates a structured result record for each final output result and publishes it to an alarm platform or storage system, and generates an evidence image corresponding one-to-one with the structured result record. The fields of the structured result record are fixed.

[0298] As an example, the final output result can be a set of fields, which may include, but are not limited to, data identification information, detection result information, risk results, and visual evidence.

[0299] As an example, the evidence image includes at least the original image with superimposed frame and the corridor probability heat map.

[0300] In yet another specific embodiment, the power transmission and distribution line channel anomaly detection device of this application may further include an image acquisition module for acquiring the image to be analyzed and outputting the original input data and the identification information corresponding to the input data.

[0301] As an example, the identification information includes at least the image number or frame number and the timestamp.

[0302] In the aforementioned power transmission and distribution line corridor anomaly detection device, the preprocessing module and the corridor structure space prior construction module work together to extract structural spatial information such as towers and conductor corridors from the original image, generating a multi-dimensional structural field prior map with clear physical meaning, providing robust geometric constraints for subsequent detection. The prior multi-scale alignment and gating modulation module achieves feature extraction fusion and prior alignment modulation, enhancing the key feature representation capability and improving feature quality. The discrete distribution regression mechanism of the detection head module quantifies the regression uncertainty of the four sides of the bounding box, and combined with the supervised constraints applied by the scale adaptive distribution regression training module, it can obtain spatial information from the original image. The system calibrates detection results using two dimensions: consistency and confidence, addressing the issues of inaccurate positioning and high false alarm rates in complex environments. Supervised construction and regularization constraints through a scale-adaptive distribution regression training module improve the consistency of candidate bounding boxes and the accuracy of confidence calibration. Multi-stage optimization via inference calibration and post-processing modules eliminates false detection results, ensuring the accuracy of the final bounding box. Alarm assessment and output modules classify alarm levels and output tiered information, enabling precise assessment and efficient feedback of abnormal risks. This comprehensively enhances the accuracy, reliability, and practicality of the device's detection capabilities, providing strong support for the safe operation and maintenance of power transmission and distribution lines.

[0303] In another embodiment, this application also provides a power transmission and distribution line channel anomaly detection system, including: one or more processors; a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the power transmission and distribution line channel anomaly detection method described in the above embodiments.

[0304] As an example, the power transmission and distribution line channel anomaly detection system pre-sets and solidifies the constant parameters in the above embodiments in the storage device; when the processor runs, it reads the constant parameters in the above embodiments and strictly performs the formula calculation and judgment according to the above embodiments, and the output result is uniquely determined by the constant parameters in the above embodiments.

[0305] In another embodiment, this application also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, it implements the various steps of the power transmission and distribution line channel anomaly detection method provided in the above embodiments.

[0306] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0307] Although this application has been disclosed above with reference to embodiments, it is not intended to limit this application. Anyone skilled in the art can make some modifications and refinements without departing from the spirit and scope of this application.

Claims

1. A method for detecting anomalies in power transmission and distribution line corridors, characterized in that, Includes the following steps: The image to be analyzed is preprocessed to obtain input data; Based on the structural characteristics of the power transmission and distribution line corridor, structural information of the effective area of ​​the corridor is extracted from the input data to generate a multivariate structural field prior map. This includes: edge extraction and line segment detection of the input data to obtain a set of line segments; statistical analysis or clustering of the line segment direction angles to determine the main direction of the corridor; fitting the set of line segments based on the main direction of the corridor to obtain the corridor centerline; determining the corridor half-width function based on perspective effect; determining the corridor region probability field based on the distance from each pixel to the corridor centerline; determining the corridor region based on the corridor region probability field; calculating the distance from each pixel to the corridor region to obtain a distance field; and calculating the gradient direction on the distance field and taking a unit vector pointing towards the direction of decreasing distance to obtain a direction field. Construct an object detection network, input the input data into the object detection network, and obtain a multi-scale feature set; The process involves spatially registering the multi-dimensional structure field prior map with features at each scale to generate a prior gating map, and then performing gating modulation based on the prior gating map to obtain enhanced features. This includes: aligning the multi-dimensional structure field prior map to the spatial resolution of each scale feature in the multi-scale feature set to obtain a scale-aligned prior; mapping the scale-aligned prior to a prior gating map; and performing affine gating modulation on each scale feature based on the prior gating map to obtain enhanced features. Based on the enhanced features, a forward bounding box regression operation is performed to output candidate bounding boxes and their corresponding regression uncertainties; including: inputting the enhanced features into the detection head, outputting the class probability and the discrete distribution of the four sides of the bounding box; performing expectation decoding on the discrete distribution of the four sides of the bounding box to obtain candidate bounding boxes; and calculating the regression uncertainty of the candidate bounding boxes. Based on the prior map of the multivariate structure field, supervised construction and regularization constraints are applied to the scale-adaptive distribution regression of the candidate bounding boxes to obtain spatial consistency scores and calibration confidence scores. This includes: determining the target scale based on the ground truth bounding boxes; establishing a mapping function from the target scale to temperature parameters, and constructing a soft label distribution based on the mapping function; determining the distribution regression loss based on the soft label distribution; setting a distribution smoothing term, establishing a mapping function from the target scale to regularization weights, and determining the regularization loss based on the mapping function to obtain the total loss; calculating the corridor probability consistency, distance consistency, and orientation consistency of pixels within the candidate bounding boxes, and obtaining the spatial consistency score through weighted fusion; and calibrating the original category confidence score using the spatial consistency score and regression uncertainty to obtain the calibration confidence score. Deterministic filtering is performed based on the spatial consistency score, regression uncertainty, and calibration confidence, and suppression processing is applied to the filtered candidate bounding boxes to obtain the final detection result.

2. The method for detecting anomalies in power transmission and distribution line channels according to claim 1, characterized in that, The multivariate structured prior map includes one or more of the following: the probability field of the corridor region, the distance field to the center line or boundary of the corridor, and the direction field of the conductor or the main direction of the corridor.

3. The method for detecting anomalies in power transmission and distribution line corridors according to any one of claims 1 to 2, characterized in that, It also includes calculating a risk score based on calibration confidence, spatial consistency score, and the degree of intrusion of the candidate bounding box relative to the corridor, determining the alarm level based on the risk score, and outputting the final result.

4. A device for detecting anomalies in power transmission and distribution line channels, characterized in that, A method for detecting anomalies in transmission and distribution line channels as described in any one of claims 1 to 3, comprising: a preprocessing module, a channel structure space prior construction module, a prior multi-scale alignment and gating modulation module, a detection head module, a scale-adaptive distribution regression training module, an inference calibration and post-processing module, and an alarm evaluation and output module. The preprocessing module preprocesses the image to be analyzed to obtain input data; the channel structure space prior construction module generates a multi-dimensional structure field prior map based on the input data output by the preprocessing module; the prior multi-scale alignment and gating modulation module extracts and fuses features from the input data to obtain a multi-scale feature set, performs prior alignment and gating modulation on the multi-dimensional structure field prior map, and outputs... The enhancement features are used to calculate the category probability and the discrete distribution of the four sides of the bounding box, and to perform expected decoding to output candidate bounding boxes and regression uncertainty. The scale-adaptive distribution regression training module performs supervised construction and regularization constraints on the candidate bounding boxes to obtain spatial consistency scores, calibration confidence, and calculate the total loss. The inference calibration and post-processing module calculates consistency indices, confidence calibration, deterministic filtering, and coordinate back mapping based on the candidate bounding boxes and regression uncertainty, and outputs the final output bounding boxes. The alarm evaluation and output module calculates risk scores and classifies alarm levels based on calibration confidence and corridor half-width, and outputs graded alarm information and the final result.

5. A power transmission and distribution line corridor anomaly detection system, characterized in that, include: One or more processors; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the transmission and distribution line channel anomaly detection method as described in any one of claims 1 to 3.

6. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the power transmission and distribution line channel anomaly detection method according to any one of claims 1 to 3.

Images