Ultra-high voltage transmission line defect detection enhancement method based on one-way lstm online learning

Abstract

The application discloses an enhanced method for defect detection of an ultrahigh-voltage transmission line based on one-way LSTM online learning. The method comprises collecting an ultrahigh-voltage transmission line image and pre-processing the image; using an image segmentation algorithm to segment and extract a core area of the transmission line from the pre-processed ultrahigh-voltage transmission line image; using a feature extraction model to extract multi-scale features in the segmented image and fuse the features to obtain an initial defect map; inputting the initial defect map into a defect recognition model for defect detection to obtain a final defect map and performing threshold segmentation to obtain a defect area position, and combining a transmission line defect type feature to obtain a defect type and a confidence level. The application is aimed at the repetitive and linearly extended defects of the ultrahigh-voltage transmission line, uses time sequence dependence and an attention mechanism to improve precision and robustness, and through online self-adaption, light weight and explainable design, realizes dynamic optimization, low-power deployment and traceability, and balances performance, flexibility and practicability.

Description

Technical Field

[0001] This invention belongs to the field of ultra-high voltage transmission line defect detection, specifically involving an enhanced method for ultra-high voltage transmission line defect detection based on unidirectional LSTM online learning. Background Technology

[0002] Ultra-high voltage (UHV) transmission lines are the core transmission carriers of power systems. Defects such as broken conductor strands, damaged insulators, deformed hardware, and foreign object entanglement directly threaten the safety of power transmission. Currently, UHV transmission line inspections mainly rely on drones and inspection robots to collect images, and then use intelligent algorithms to identify defects. However, the inspection scenario faces two major technical challenges: First, transmission lines exhibit significant repetitive defect characteristics. Standardized components such as conductors, insulators, and hardware of the same specification tend to have repeated defects at fixed locations due to consistent manufacturing processes and stress characteristics. Furthermore, fixed line sections are continuously affected by geographical and climatic environments, leading to repeated occurrences of the same type of defect. Second, the inspection environment is complex and variable. Factors such as lighting intensity, shooting angle tilt, and background obstruction by trees and buildings can cause significant differences in the characteristics of the collected images, placing high demands on the environmental adaptability of the detection algorithms.

[0003] Existing defect detection technologies for ultra-high voltage transmission lines are mainly divided into traditional machine vision methods (such as template matching and edge detection) and deep learning methods (such as YOLO and PatchCore). Traditional methods rely on manually designed features, which have poor adaptability to complex environments and subtle defects. Although mainstream deep learning methods have improved detection accuracy through network structure improvements and have partially achieved lightweight design, they have not fully explored the inherent patterns of repetitive defects in similar components and fixed sections of transmission lines, and also lack dynamic online optimization capabilities, making it difficult to balance detection accuracy, environmental adaptability, and deployment flexibility.

[0004] For example, the existing technology CN119515821A lacks the ability to mine temporal information and repetitive defect patterns, resulting in a high rate of missed detections for minor defects and foreign object entanglement in similar components and fixed sections. At the same time, the offline static model has weak generalization ability and lacks dynamic optimization capability after deployment, making it unable to adapt to complex environmental changes and new defects, and the detection effect is difficult to continuously improve. Summary of the Invention

[0005] To address the problems existing in the background technology, this invention provides an enhanced method for detecting defects in ultra-high voltage transmission lines based on unidirectional LSTM online learning. This method solves the technical problems of existing ultra-high voltage transmission line defect detection technologies, which generally fail to explore the temporal patterns of repetitive defects and lack dynamic online optimization capabilities, resulting in high false negative rates, weak environmental generalization, and difficulty in balancing detection accuracy and deployment flexibility in complex dynamic scenarios.

[0006] The technical solution adopted in this invention is:

[0007] I. An Enhanced Method for Defect Detection of Ultra-High Voltage Transmission Lines Based on One-Way LSTM Online Learning:

[0008] S1. Acquire images of ultra-high voltage transmission lines and preprocess the images.

[0009] S2. A pre-trained image segmentation algorithm is used to segment and extract the core region of the transmission line from the pre-processed ultra-high voltage transmission line image to obtain the segmented transmission line image.

[0010] S3. A pre-trained feature extraction model is used to extract multi-scale features from the power transmission line segmentation image. The initial defect map is obtained after fusing the multi-scale features. .

[0011] S4. Initial defect map The data is input into a defect recognition model for defect detection, resulting in the final defect map. .

[0012] S5. Final Defect Diagram Threshold segmentation is performed to obtain the location of the defect area, and the defect type and confidence level are obtained by combining the characteristics of the transmission line defect type.

[0013] The pre-trained image segmentation algorithm uses the Mask R-CNN algorithm; the core region of the transmission line includes the area where the conductors, insulators, and fittings are located; the pre-trained feature extraction model uses MobileViT-S, ViT-L / 14@336px, EfficientNet-B0, or YOLOv8n backbone network.

[0014] The defect identification model employs an anomaly repetition learner model, and the specific processing steps of the anomaly repetition learner model are as follows:

[0015] S41. Initial defect map After one convolutional layer, the context-enhanced defect map is obtained. .

[0016] S42. Initial defect diagram Flattened into a one-dimensional defect time series .

[0017] S43, Transform the one-dimensional defect time sequence Each element The input is fed into two cascaded unidirectional LSTM networks for processing.

[0018] S44. The hidden state sequence in the second unidirectional LSTM network... Converted into attention weights Then, the context-enhanced defect map Weighted modulation is performed to obtain the attention-enhanced transmission line defect map. .

[0019] S45. Initial defect diagram With enhanced attention, the defect diagram of the power transmission line was obtained. Residual fusion is performed to obtain the final defect map. .

[0020] Two cascaded unidirectional LSTM networks process a one-dimensional defect time series according to the following formula. The Middle element Processing:

[0021]

[0022]

[0023]

[0024]

[0025] in, For indexing; A one-dimensional defect time sequence The Middle One element; and These represent the results obtained by applying activation functions to the outputs of the second and first unidirectional LSTM networks, respectively. and These represent the second and first Sigmoid activation functions, respectively. and These represent the second unidirectional LSTM. The hidden state and cell state corresponding to each element; and These represent the second unidirectional LSTM. The hidden state and cell state corresponding to each element; and These represent the first unidirectional LSTM. The hidden state and cell state corresponding to each element; and These represent the first unidirectional LSTM. The hidden state and cell state corresponding to each element; and These are the weights in the second and first unidirectional LSTMs, respectively. and These are the biases in the second and first unidirectional LSTMs, respectively.

[0026] Final Defect Diagram The result is obtained by processing it using the following formula:

[0027] ; ;

[0028]

[0029] in, This is the final defect diagram; A diagram of power transmission line defects after attention enhancement; This is an adaptive scaling factor; This is the initial defect diagram; Attention weights; This is a context-enhanced defect graph; Represents the Hadamard product; Use the Sigmoid activation function; It is a multilayer perceptron; This is the hidden state sequence in the second unidirectional LSTM network; , and Representing the first element and the second element in the second unidirectional LSTM network, respectively. The element and the first The hidden state corresponding to each element.

[0030] A hybrid self-supervised loss function is used when training the defect identification model. The hybrid self-supervised loss function Set it according to the following formula:

[0031]

[0032]

[0033]

[0034]

[0035] in, Represents the hybrid self-supervised loss function; This results in a loss of time-series consistency. This is for orientation alignment loss; To enhance the contrast and increase the loss; , and All are weighted parameters; and These represent the height and width of the initial defect map, respectively; and All are indexes; For the final defect diagram medium to high Width Element; This is the previous initial defect map input into the defect identification model; For the previous initial defect diagram medium to high Width Element; It is an L2 norm; It is a constant to prevent the denominator from being 0; and This represents the previous final defect map output by the defect identification model. Defective and normal regions were obtained separately after threshold segmentation; This indicates the previous final defect diagram. The defective region of the element Seeking expectations; The final defect diagram shown above Normal region for elements Seeking expectations.

[0036] The characteristics of the transmission line defect types include linear characteristics, irregular block characteristics, edge gap characteristics, and local fracture characteristics; if the defect area has linear characteristics, the defect type is broken strand; if the defect area has irregular block characteristics, the defect type is foreign object attachment; if the defect area has edge gap characteristics, the defect type is damage; if the defect area has local fracture characteristics, the defect type is cracking.

[0037] II. A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0038] 3. A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above method.

[0039] The beneficial effects of this invention are:

[0040] 1. Highly targeted defect detection: Specifically adapted to the "repetitive and linear extension" characteristics of defects in ultra-high voltage transmission lines (such as repeated foreign object entanglement and long-distance conductor strand breakage in the same line segment), it captures temporal dependencies through unidirectional LSTM to enhance the detection response of similar defects. Compared with traditional single-frame detection, it has higher detection accuracy for repetitive defects. At the same time, the attention weighting mechanism can accurately focus on the extension direction of linear defects, avoiding defects such as strand breakage and long cracks from being misjudged as noise.

[0041] 2. Outstanding robustness in complex inspection environments: Addressing issues such as variable lighting (sunny / cloudy / nighttime), tilted shooting angles, and background interference (trees / buildings obstructing the view) in power transmission line inspections, a hybrid self-supervised online optimization mechanism is used. Without manual annotation, parameters can be updated in real time using only continuous frame time sequence information, dynamically adapting to environmental changes and solving the "model drift" problem of offline models in complex scenarios.

[0042] 3. Online adaptive capability to meet the needs of power transmission line operation and maintenance: The types of defects in power transmission lines will increase with changes in the environment (such as seasonal foreign objects, new types of cable damage). The ARL model of this invention can learn new recurring abnormal patterns in real time during the detection process, without the need to re-label data and retrain offline, which greatly reduces the operation and maintenance cost, and is especially suitable for long-term inspection scenarios of power transmission lines in remote areas.

[0043] 4. Lightweight design meets the essential needs of drone deployment: The ARL model has only 1.2k parameters at the very least. With lightweight backbones such as MobileViT-S and YOLOv8n, the total number of parameters is ≤30M and the single-frame inference time is ≤10ms. It is fully compatible with the computing power constraints of drone embedded chips and has low inference power consumption, which does not affect the drone's battery life.

[0044] 5. Interpretability Adapts to Power Industry Audit Requirements: Attention weight vectors can clearly mark the confidence level of defect areas, and the residual fusion mechanism only amplifies highly repeatable abnormal signals, making it easy for maintenance personnel to trace the detection basis, which meets the power industry's management requirements of "verifiable and traceable detection results". Attached Figure Description

[0045] Figure 1 This is a flowchart of the method of the present invention.

[0046] Figure 2 This is a flowchart of the abnormal repetition learner model of the present invention.

[0047] Figure 3 This diagram illustrates the hybrid self-supervised loss function structure of the defect identification model of this invention. Detailed Implementation

[0048] The present invention will now be described in more detail with reference to the accompanying drawings and embodiments. However, the present invention is not limited thereto. For those skilled in the art, several improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention. Contents not described in detail in this specification are prior art known to those skilled in the art.

[0049] like Figure 1 As shown, the enhanced defect detection method for ultra-high voltage transmission lines in this embodiment is implemented according to the following steps:

[0050] S1. Use drones / inspection robots to collect images of ultra-high voltage transmission lines and preprocess the images; preprocessing involves removing invalid images that are blurry, overexposed, or severely occluded.

[0051] In the subsequent training process of the defect identification model (ARL model), a large number of ultra-high voltage transmission line images will be collected first to construct an ultra-high voltage transmission line image dataset.

[0052] Specifically, the method of the present invention is a zero-shot detection method. The model does not need to be trained by labeling images of ultra-high voltage transmission lines, and the defect type is only used for the custom rules of defect classification.

[0053] S2. A pre-trained image segmentation algorithm is used to segment and extract the core region of the transmission line from the pre-processed ultra-high voltage transmission line image to obtain the segmented transmission line image.

[0054] The core area of ​​a power transmission line includes the area where the conductors, insulators, and fittings are located.

[0055] In this embodiment, the pre-trained image segmentation algorithm uses the Mask R-CNN algorithm.

[0056] S3. A pre-trained feature extraction model is used to extract multi-scale features from the power transmission line segmentation image. After fusing the multi-scale features, a pixel-level initial defect map is obtained. .

[0057] The pre-trained feature extraction model uses MobileViT-S, ViT-L / 14@336px, EfficientNet-B0, or YOLOv8n backbone networks. In this embodiment, the pre-trained feature extraction model uses the ViT-L / 14@336px model.

[0058] S4. Initial defect map The data is input into a defect recognition model for defect detection, resulting in the final defect map. .

[0059] The defect identification model employs an Anomaly Repeat Learner (ARL) model. The ARL model is the core module of this invention, responsible for capturing the spatiotemporal dependencies of defects in ultra-high voltage transmission lines and enhancing detection performance. Applying the ARL model to ultra-high voltage transmission line detection is not simply about performing temporal modeling on inspection frames, but rather about customizing spatiotemporal feature capture and enhancement based on the repetitive defect patterns of "similar components and fixed sections" in transmission lines. This matches the characteristics of transmission line operation and maintenance inspection, specifically consisting of two core adaptation layers:

[0060] Layer 1: Adapting to the "Inherent Repetitive Defect Patterns of Similar Components in Transmission Lines". Ultra-high voltage transmission lines are composed of standardized, similar components such as conductors, insulators, fittings, and grading rings. These components, with the same specifications and structure, exhibit recurring defects at fixed locations due to consistent manufacturing processes and stress characteristics. The ARL model addresses this pattern by initializing an independent ARL model for each type of component through category-specific instantiation. This allows the model to specifically learn the "high-probability defect area" features of a particular component type—for example, the ARL model for insulators focuses only on repetitive defect features at "skirt edges and bonding points," while the ARL model for conductors only captures strand breakage / wear patterns at "joints and mid-span sections." This avoids interference from cross-component features, achieving targeted enhancement for repetitive defects at fixed locations of similar components. This makes the model's detection sensitivity for these high-probability defect areas higher than traditional single-frame detection methods.

[0061] The second layer: Adapting to the "environmentally induced repetitive defect patterns in fixed sections of transmission lines". Due to the fixed geographical environment and climate conditions, fixed sections of ultra-high voltage transmission lines will have a continuous and fixed external effect on the transmission lines they pass through, causing various components in the section to repeatedly exhibit the same type of defect (such as foreign object entanglement in mountainous windy sections, insulator corrosion in coastal salt spray sections, and conductor icing in river crossing sections). This is a unique regional repetitive defect pattern in ultra-high voltage transmission line inspection, and it is also the core optimization direction of the ARL model.

[0062] One-way LSTM network modeling: Accumulate the historical features of defects in a fixed section through the hidden state of LSTM---for example, in a windy section of a mountain, foreign objects such as plastic film and tree branches are repeatedly entangled. The one-way LSTM network will continuously accumulate the foreign object defect features of this section. When inspecting this section, the model will strengthen the defect detection response in advance. Even if the foreign object defect is minor, it can be accurately identified and avoid missed detection.

[0063] Attention weighting: The features of high-probability defect areas in fixed segments / components learned by the LSTM network are transformed into attention weights, which amplify the features of these areas in a targeted manner, while suppressing noise in normal areas without defect history, so that the detection results are more focused on "defect-prone parts / segments".

[0064] like Figure 2 As shown, based on the aforementioned repetitive defect patterns, the overall implementation logic of the ARL model in ultra-high voltage transmission line inspection can be summarized as follows: First, define the inspection unit according to "component category + line section" → then capture the repetitive defect patterns within the inspection unit → finally perform targeted enhancement on high-probability defect areas. This is specifically achieved step-by-step through the following 5 components:

[0065] 1. Spatial context extraction: extracting the initial defect map After a 3×3 convolutional layer, the context-enhanced defect map is obtained. A 3×3 convolutional layer can extract the initial defect map. Local spatial structure information is used to smooth noise while preserving defect contours: .

[0066] 2. Time sequence conversion: Convert the initial defect map... Flattened into a one-dimensional defect time series .

[0067] Specifically, ,in , and These represent the initial defect diagrams. The first, second and third The value corresponding to each element.

[0068] 3. Unidirectional LSTM network modeling: transforming one-dimensional defect time series... Each element The input is fed into two cascaded unidirectional LSTM networks for processing. The two cascaded unidirectional LSTM networks process the one-dimensional defect time series according to the following formula. The Middle element Processing:

[0069]

[0070]

[0071]

[0072]

[0073] in, For indexing; A one-dimensional defect time sequence The Middle One element; and These represent the results obtained by applying activation functions to the outputs of the second and first unidirectional LSTM networks, respectively. and These represent the second and first Sigmoid activation functions, respectively. and These represent the first and second unidirectional LSTM networks, respectively. and These represent the second unidirectional LSTM. The hidden state and cell state corresponding to each element; and These represent the second unidirectional LSTM. The hidden state and cell state corresponding to each element; and These represent the first unidirectional LSTM. The hidden state and cell state corresponding to each element; and These represent the first unidirectional LSTM. The hidden state and cell state corresponding to each element; and These are the weights in the second and first unidirectional LSTMs, respectively. and These are the biases in the second and first unidirectional LSTMs, respectively.

[0074] 4. Attention Weighting: Weighting the hidden state sequence in the second unidirectional LSTM network. Converted into attention weights Then, the context-enhanced defect map Weighted modulation is performed to obtain the attention-enhanced transmission line defect map. .

[0075] Enhanced attention to power line defect diagram The result is obtained by processing it using the following formula:

[0076] ;

[0077] in, A diagram of power transmission line defects after attention enhancement; Attention weights; Use the Sigmoid activation function; It is a multilayer perceptron; This is the hidden state sequence in the second unidirectional LSTM network; This is a context-enhanced defect graph; This represents the Hadamard product.

[0078] 5. Residual Fusion: This involves combining the initial defect map... With enhanced attention, the defect diagram of the power transmission line was obtained. Perform residual fusion to obtain the current initial defect map. Corresponding final defect diagram ;

[0079] Final Defect Diagram The result is obtained by processing it using the following formula:

[0080]

[0081] in, This is the final defect diagram; A diagram of power transmission line defects after attention enhancement; For adaptive scaling factor, The default value is 0.5; This is the initial defect diagram.

[0082] S5. Final Defect Diagram Threshold segmentation is performed to obtain the location of the defect area (the remaining locations are normal areas), and the defect type and confidence level are obtained by combining the characteristics of the power transmission line defect type.

[0083] The characteristics of transmission line defects include linear features, irregular block features, edge gap features, and local fracture features. If the defect area has linear features, the defect type is broken strand; if the defect area has irregular block features, the defect type is foreign object attachment; if the defect area has edge gap features, the defect type is damage; if the defect area has local fracture features, the defect type is cracking.

[0084] Specifically, the image segmentation algorithm and feature extraction model need to be trained separately to obtain the corresponding pre-trained models. Then, steps S1-S5 are repeated to train the defect recognition model (ARL model) to obtain the trained ARL model.

[0085] like Figure 3 As shown, a hybrid self-supervised loss function is further employed when training the defect identification model (ARL model). Using the current final defect diagram of ultra-high voltage transmission lines Compared with the previous initial defect diagram Unsupervised optimization, which requires no manually labeled data, includes three loss terms:

[0086] 1. Temporal consistency loss ( To ensure the stability of continuous frame defect detection results for ultra-high voltage transmission lines and reduce fluctuations caused by changes in the environment where the transmission lines are located, the formula is as follows:

[0087]

[0088] 2. Orientation alignment loss ( To enhance the directional consistency of defect characteristics in ultra-high voltage transmission lines (such as the direction of linear strand breakage), the formula is as follows:

[0089]

[0090] This loss term is specifically designed to enhance the directional consistency of linear defects in transmission lines (such as broken strands and long-distance scratches), ensuring that the model can accurately identify the extension trajectory of defects even when the shooting angle is tilted, thus avoiding positioning deviations caused by angle changes.

[0091] 3. Contrast enhancement loss ( To increase the response difference between defective and normal areas of ultra-high voltage transmission lines and improve defect discrimination, the formula is as follows:

[0092]

[0093] Total loss (hybrid self-supervised loss function) The sum of the three losses is expressed as follows:

[0094]

[0095] in, Represents the hybrid self-supervised loss function; This results in a loss of time-series consistency. This is for orientation alignment loss; To enhance the contrast and increase the loss; , and All are weighted parameters; and These represent the height and width of the initial defect map, respectively; and All are indexes; For the final defect diagram medium to high Width Element; This is the previous initial defect map input into the defect identification model; For the previous initial defect diagram medium to high Width Element; It is an L2 norm; It is a very small constant to prevent the denominator from being 0; and This represents the previous final defect map output by the defect identification model. Defective and normal regions were obtained separately after threshold segmentation; This indicates the previous final defect diagram. The defective region of the element Seeking expectations; The final defect diagram shown above Normal region for elements Find the expected value. Specifically, The smaller the value, the higher the response of the defective area, and the more obvious the defect.

[0096] in The model parameters are updated in real time using the AdamW optimizer (initial learning rate 1×10−4).

[0097] This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method. This embodiment further provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described method.

[0098] To further demonstrate the beneficial effects of the method of the present invention, comparative experiments were conducted on the MVTec AD (15 classes) public dataset and the VisA (12 classes) public dataset, respectively. The experimental results are shown in Table 1 and Table 2, respectively:

[0099] Table 1. Comparison results on MVTec AD (15 categories)

[0100]

[0101] Table 2. Comparison results on VisA (12 categories)

[0102]

[0103] The results in Tables 1 and 2 clearly demonstrate that the method of the present invention has a significant advantage over the prior art.

[0104] In summary, the method of this invention significantly improves the detection accuracy of recurring defects (such as foreign object entanglement and broken strands) and long-distance linear defects in ultra-high voltage transmission lines by using unidirectional LSTM to capture temporal dependencies and combining them with an attention weighting mechanism, thus overcoming the missed detection problem of traditional single-frame detection. Through hybrid self-supervised online optimization, parameters can be updated in real time without manual annotation, dynamically adapting to complex environmental changes such as lighting, angle, and occlusion, completely solving the "model drift" problem of offline models. Simultaneously, the model can learn new recurring anomaly patterns online during the detection process, eliminating the need for retraining and significantly reducing long-term operation and maintenance costs in remote areas. The lightweight design is fully compatible with embedded chips in drones, with low power consumption and no impact on battery life. Furthermore, the attention weighting and residual fusion mechanism provides interpretable defect confidence, facilitating review and traceability by maintenance personnel, and meeting the management requirements of the power industry. In conclusion, this invention achieves comprehensive breakthroughs in detection accuracy, environmental robustness, online adaptability, lightweight deployment, and interpretability.

[0105] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. An enhanced method for defect detection in ultra-high voltage transmission lines based on unidirectional LSTM online learning, characterized in that, Including the following methods: S1. Acquire images of ultra-high voltage transmission lines and preprocess the images; S2. A pre-trained image segmentation algorithm is used to segment and extract the core region of the transmission line from the pre-processed ultra-high voltage transmission line image to obtain the transmission line segmentation image. S3. A pre-trained feature extraction model is used to extract multi-scale features from the power transmission line segmentation image. The initial defect map is obtained after fusing the multi-scale features. ; S4. Initial defect map The data is input into a defect recognition model for defect detection, resulting in the final defect map. ; S5. Final Defect Diagram Threshold segmentation is performed to obtain the location of the defect area, and the defect type and confidence level are obtained by combining the characteristics of the transmission line defect type.

2. The enhanced method for detecting defects in ultra-high voltage transmission lines based on unidirectional LSTM online learning according to claim 1, characterized in that: The pre-trained image segmentation algorithm uses the Mask R-CNN algorithm; the core region of the transmission line includes the area where the conductors, insulators, and fittings are located; the pre-trained feature extraction model uses MobileViT-S, ViT-L / 14@336px, EfficientNet-B0, or YOLOv8n backbone network.

3. The enhanced method for ultra-high voltage transmission line defect detection based on unidirectional LSTM online learning according to claim 1, characterized in that, The defect identification model employs an anomaly repetition learner model, and the specific processing steps of the anomaly repetition learner model are as follows: S41. Initial defect map After one convolutional layer, the context-enhanced defect map is obtained. ; S42. Initial defect map Flattened into a one-dimensional defect time series ; S43, Transform the one-dimensional defect time sequence Each element The input is fed into two cascaded unidirectional LSTM networks for processing; S44. The hidden state sequence in the second unidirectional LSTM network... Converted into attention weights Then, the context-enhanced defect map Weighted modulation is performed to obtain the attention-enhanced transmission line defect map. ; S45. Initial defect diagram With enhanced attention, the defect diagram of the power transmission line was obtained. Residual fusion is performed to obtain the final defect map. .

4. The enhanced method for detecting defects in ultra-high voltage transmission lines based on unidirectional LSTM online learning according to claim 3, characterized in that: Two cascaded unidirectional LSTM networks process a one-dimensional defect time series according to the following formula. The Middle element Processing: in, For indexing; A one-dimensional defect time sequence The Middle One element; and These represent the results obtained by applying activation functions to the outputs of the second and first unidirectional LSTM networks, respectively. and These represent the second and first Sigmoid activation functions, respectively. and These represent the second unidirectional LSTM. The hidden state and cell state corresponding to each element; and These represent the second unidirectional LSTM. The hidden state and cell state corresponding to each element; and These represent the first unidirectional LSTM. The hidden state and cell state corresponding to each element; and These represent the first unidirectional LSTM. The hidden state and cell state corresponding to each element; and These are the weights in the second and first unidirectional LSTMs, respectively. and These are the biases in the second and first unidirectional LSTMs, respectively.

5. The enhanced method for ultra-high voltage transmission line defect detection based on unidirectional LSTM online learning according to claim 3, characterized in that: Final Defect Diagram The result is obtained by processing it using the following formula: ； ； in, This is the final defect diagram; A diagram of power transmission line defects after attention enhancement; This is an adaptive scaling factor; This is the initial defect diagram; Attention weights; This is a context-enhanced defect graph; Represents the Hadamard product; Use the Sigmoid activation function; It is a multilayer perceptron; This is the hidden state sequence in the second unidirectional LSTM network; , and Representing the first element and the second element in the second unidirectional LSTM network, respectively. The element and the first The hidden state corresponding to each element.

6. The enhanced method for detecting defects in ultra-high voltage transmission lines based on unidirectional LSTM online learning according to claim 1, characterized in that: A hybrid self-supervised loss function is used when training the defect identification model. The hybrid self-supervised loss function Set it according to the following formula: in, Represents the hybrid self-supervised loss function; This results in a loss of time-series consistency. This is for orientation alignment loss; To enhance the contrast and increase the loss; , and All are weighted parameters; and These represent the height and width of the initial defect map, respectively; and All are indexes; For the final defect diagram medium to high Width Element; This is the previous initial defect map input into the defect identification model; For the previous initial defect diagram medium to high Width Element; It is an L2 norm; It is a constant to prevent the denominator from being 0; and This represents the previous final defect map output by the defect identification model. Defective and normal regions were obtained separately after threshold segmentation; This indicates the previous final defect diagram. The defective region of the element Seeking expectations; The final defect diagram shown above Normal region for elements Seeking expectations.

7. The enhanced method for detecting defects in ultra-high voltage transmission lines based on unidirectional LSTM online learning according to claim 1, characterized in that: The characteristics of the transmission line defect types include linear characteristics, irregular block characteristics, edge gap characteristics, and local fracture characteristics; if the defect area has linear characteristics, the defect type is broken strand; if the defect area has irregular block characteristics, the defect type is foreign object attachment; if the defect area has edge gap characteristics, the defect type is damage; if the defect area has local fracture characteristics, the defect type is cracking.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images