External defect detection technology and application status of transmission and transformation equipment based on video images

Power transmission and distribution equipment plays a critical role in supplying electricity to consumers. However, these assets are susceptible to external defects, such as corrosion, mechanical damage, and wear, which can lead to failures and disruptions in the electrical grid. Traditional inspection methods for detecting these defects often rely on manual inspections, which are time-consuming, costly, and subjective. To overcome these limitations, this paper explores the current state of video image-based external defect detection techniques for power transmission and distribution equipment. This makes up for the deficiencies of conventional approaches to inspecting and maintaining power transmission and transformation equipment by decreasing the waste of human resources and increasing the frequency and efficiency of intelligent operation and maintenance of power systems. This work investigates a completely convolutional block detection-based defect identification method to address the issue of defect recognition. The fully convolutional neural network is enhanced with the concept of block detection thanks to this approach. The local discrimination mechanism may be realized, and the drawbacks of the conventional block detection receptive field are avoided in the process. This approach offers improved generalization and fault identification over the original ResNet image classification system.


Introduction
Power transmission and distribution equipment serve as the backbone of electrical grids, ensuring the reliable supply of electricity to consumers [1].However, these critical assets are prone to external defects, including corrosion, mechanical damage, and wear, which can compromise their performance and lead to failures or disruptions in the grid [2].Traditionally, the detection of external defects in power transmission and distribution equipment relied on manual inspections carried out by skilled technicians.However, this approach is often time-consuming, costly, and subjective [3].Moreover, manual inspections may not be able to identify hidden or hard-to-reach defects, which can go undetected until they escalate into major issues.
To overcome these limitations, there is a growing interest in exploring automated defect detection techniques that leverage computer vision and image analysis [4].These techniques utilize video data collected from cameras installed on or around the equipment of power transmission and distribution systems [5].By analyzing these videos, it becomes possible to detect and classify external defects in real-time or during post-processing.The use of video-based defect detection techniques offers numerous advantages.Firstly, it can significantly reduce inspection time and costs by eliminating the need for manual inspections.Secondly, video analysis can provide more objective and consistent defect identification, minimizing the potential for human error or bias.Thirdly, automated defect detection can enable early detection and diagnosis of defects, allowing for timely preventive maintenance and repairs [6].
In recent years, there has been a growing body of research on video-based defect detection techniques for power transmission and distribution equipment [7].These techniques typically involve the application of computer vision algorithms, such as object detection, image segmentation, and pattern recognition, to identify and characterize external defects accurately [8].This paper aims to provide an overview of the current state of video-based defect detection techniques for power transmission and distribution equipment.It will explore the challenges faced by traditional inspection methods, investigate the potential benefits of video analysis in defect detection, and discuss various methodologies and algorithms employed in this field [9].
The findings of this study are expected to contribute to the development and implementation of efficient and reliable defect detection systems for power transmission and distribution equipment.By leveraging the advancements in computer vision and image analysis, these systems have the potential to revolutionize the inspection and maintenance practices in the power industry, enhancing the reliability of electrical grids and ensuring the continuous supply of electricity to consumers.

Related work
In the field of defect detection, several studies have explored different approaches to improve the accuracy and robustness of detection models.One of the challenges encountered in detecting defects in power line insulators based on aerial image backgrounds is the low detection accuracy and recall rate when using only convolutional neural networks (CNN).To address this issue, a literature [10] proposed a cascade of two Faster R-CNN networks.The cascade architecture consists of two stages.In the first stage, the Faster R-CNN network is used to locate the insulators in the aerial images.This stage is crucial as it accurately localizes the regions of interest, which are the power line insulators.The second stage focuses on detecting defects on the located insulators.By breaking down the detection process into two stages, the model achieves a higher level of accuracy and recall rate compared to using a single CNN.The proposed method in this literature has been evaluated on a complex background insulator dataset and demonstrated impressive results.It achieved a detection accuracy of 91% and a recall rate of 96%, meeting the requirements for robust and accurate defect detection in power line insulators.The cascade architecture proves to be effective in addressing the limitations of using only CNNs for this specific task.In literature [11], researchers took a different approach by embedding Faster R-CNN into a classification CNN.The outer classification network is responsible for classifying the defect samples, while the inner network locates the defects.This combination allows the model to leverage the advantages of both classification and detection networks.The proposed method achieved a remarkable detection accuracy of 99.8% on the DAGM2007E dataset [12], highlighting its effectiveness in defect detection.
Another study, literature [13], utilized k-means clustering to design reasonable Anchors and introduced a feature pyramid network [14] into the Faster R-CNN architecture.This integration enhanced the connection between different-level features and benefited from shallow information, making it more suitable for detecting small defects.The evaluation on a PCB defect dataset showed an average detection accuracy of 98.90%, indicating better generalization capabilities.Improvements were also made to the YOLOv3 model in literature [15].The researchers incorporated an Anchor clustering algorithm and a dual-density convolutional structure to increase the model's prediction scale.This modification resulted in a 26.1% higher detection accuracy compared to the previous version when applied to X-ray images.The enhanced YOLOv3 model demonstrated its effectiveness in the field of defect detection.In literature [16], a three-stage cascade detection network was constructed with decreasing scales to detect fastener defects on arm nodes.The first stage utilized an SSD network to locate the arm nodes and fasteners.Then, the located regions were cropped and fed into YOLO for fastener defect detection.Finally, a deep convolutional neural network (DCNN) was employed to classify the target regions.This multi-stage approach proved to be successful in achieving accurate detection and classification of fastener defects.Overall, these studies showcase various approaches to enhance defect detection models.By combining different networks, incorporating clustering algorithms, and introducing specialized architectures, researchers have been able to achieve higher detection accuracy, recall rates, and better generalization capabilities in different domains and datasets.

Methods
In contrast to the classification task, the defect recognition task presents a little fault detection challenge.Very modest changes in area ratio provide no discernible difference between small and regular area pixels.All-image evaluations are the norm for the current crop of convolutional neural networks used for image categorization.When this kind of comprehensive inspection is used directly to defect identification activities, it typically results in an extremely low detection rate for even the smallest of problems.Block detection is a widespread and efficient approach for locating local, tiny faults.Blocking a picture allows for a larger proportion of the examined area to be devoted to problems, making even minute flaws easier to see.However, the conventional block detection method ignores all of the data collectively, making it vulnerable to errors in judgment.This paper incorporates the concept of block detection into a complete convolution network and presents a defect recognition method using full convolution blocks to address the limitations of the standard block detection approach.

Advantages and bottlenecks of traditional methods
Defect detection algorithms frequently employ the technique of block detection.The three benefits listed below all apply to the conventional block detection method.The detection accuracy may be enhanced by increasing the fraction of the fault area using block detection.Two) Large amounts of training data may be quickly obtained by block division.Data is the fuel for deep learning models, which necessitates an enormous amount of data for training.Thirdly, blocking can protect the classifier-learning process from unintentional causes.In order to protect the learning process from being disrupted by random events, the block detection technique may clearly indicate to the classifier which regions are faulty and which are normal.
However, conventional approaches to block detection suffer from inadequate receptive fields and so have their limits.Therefore, this research enhances the standard block identification method by combining its features with those of a fully convolutional neural network.

Block detection with FCN
An essential notion in CNN is the receptive field, which stands for the dynamic range of the initial picture of neurons.The formula for determining the receptive field by recursion is: By incorporating fully convolutional neural networks, the limitations of traditional block identification can be overcome.With a one-to-one mapping between the source image and the segmented area in the final feature map, the network can capture both local and global information effectively.In traditional block detection methods, the individual receptive fields of neurons can only cover limited regions within an image.This means that global context and information beyond the examined portion of the image may be overlooked.However, with the use of fully convolutional neural networks, each pixel in the final feature map is connected to a larger receptive field that extends beyond the boundaries of the tested picture block.This expanded receptive field allows the network to incorporate contextual and global data, enabling a more comprehensive understanding of the image.The multi-layer convolutional neural network extracts features from the input image, gradually capturing hierarchical representations of the image.Each pixel in the final feature map contains a feature vector that represents the relevant area in the original image.This means that instead of solely relying on localized information, the network can leverage the collected features to make predictions based on both local details and global context.By utilizing both local and global information, the fully convolutional neural network can improve the accuracy and robustness of block detection.It becomes more effective in identifying flaws or abnormalities in image blocks, as it can consider the holistic structure of the image rather than relying solely on local characteristics.This enhanced capability empowers applications in various domains, such as image recognition, object detection, and quality control in manufacturing processes.
In figure 1, we see the network architecture used by the defect identification technique in this article.To create a completely convolutional neural network, the FCN was swapped out for a convolutional layer.Bilinear interpolation is used to scale the input picture down to 224×224 pixels, and a fully convolutional neural network is used to generate the prediction image.The prediction picture's receptive field encompasses the whole input image, and each pixel represents a 32×32 image block in the source image.The method anticipates (0,1) for the output value of the damaged picture block and (1,0) for the normal image block.This research uses the ResNet-50 network, which has very tiny parameters, as its backbone network because of the limited data characteristics of the defect identification job.

Optimization and training
In this research, we optimize a network with Logistic cross entropy loss, L2 regularization, and the Adam optimization method: The loss of an N-block picture batch trained using stochastic gradient descent is: During the training phase, the approach used in this study deviates from the conventional method of calculating the loss using all the picture blocks.Instead, a limited number of positive and inferior image blocks are randomly selected.This novel approach aims to enhance the unpredictability of the network, thereby mitigating the issue of overfitting.By adopting this strategy, the training samples can be effectively expanded based on the specific requirements.Through this approach, the variety of samples encountered during each iteration of the network model is significantly enhanced.This improvement is due to the fact that the samples seen in each iteration are nearly independent of each other.Consequently, this method provides a greater diversity of samples for the network to learn from.
The method proposed in this study for selecting picture blocks is as follows.Firstly, for each flawed sample, a healthy replacement is randomly chosen.This ensures that the network encounters both flawed and normal samples during training, promoting a balanced learning process.Then, the defective and normal samples are utilized independently to select the defective and normal blocks, respectively.This random selection process enables the network to expose itself to a diverse range of picture blocks, allowing it to learn and generalize better.By employing this innovative approach to sample selection, the study aims to address the limitations of using all picture blocks for calculating loss in the traditional training phase.The proposed method not only expands the training samples as needed but also improves the variety and independence of the samples seen by the network model.
These enhancements contribute to reducing the over-fitting problem and enhancing the overall performance of the network during the training phase.Following are the steps used throughout the fault recognition algorithm's training process: First, use a random seed when setting up the network.Second, pick n valid samples and n invalid samples, and then pick a mark for each pair of samples from the mark set.Third, feed the chosen samples into the network to obtain predictive predictions from the network.Fourth, generate the loss function by randomly picking faulty and healthy blocks.Fifth, adjust the network's settings such that the loss is as small as possible by applying the Adam optimization method.Lastly, keep going until the loss stops increasing.

Data enhancement
In the context of deep learning, the robustness and effectiveness of a model hinge greatly on the diversity and richness of the training samples utilized during its construction.Particularly in scenarios involving tasks such as defect detection, where the available sample size is limited (typically ranging in the tens or hundreds), the application of data augmentation becomes essential.It is aimed at maximizing sample diversity while simultaneously minimizing the risk of overfitting.In the present work, several data enhancement strategies have been adopted to achieve these objectives.One of the fundamental techniques employed for data augmentation and increasing sample variety is the utilization of rotation transformations.By introducing rotation-invariance into the model, it becomes adept at detecting flaws in objects such as buttons and other similarly structured items.The application of angular transformations involves the rotation of the input picture around its center, thereby expanding the diversity of perspectives from which the model learns to recognize defects.This not only aids in mitigating the impact of variations in the orientation of objects in the input data but also facilitates the creation of a more generalized and robust defect detection model.Furthermore, the integration of rotation transformations as part of the data augmentation process allows the model to be trained on a wider range of object orientations, thus enhancing its ability to detect defects irrespective of their orientation within the input data.This approach contributes significantly to the overall effectiveness and versatility of the defect detection model, particularly when the training sample size is constrained.By leveraging rotation transformations as a data augmentation strategy, the model becomes more resilient to variations in object orientation and is better equipped to generalize its defect detection capabilities across diverse real-world scenarios.
′ = / (6) Chroma, saturation, and luminosity are all adjusted as part of the color space transition.In this work, we employ both brightness and saturation transformations.

Result
This paper evaluates the efficacy of defect identification algorithms by their TPR (true positive recognition), TNR (true negative recognition), and ACC (accuracy).The indicator is derived in the following manner: In this part, we evaluate the method's detection performance in comparison to ResNet-50, an image classification system.Since there were so few samples in each experiment, there was a lot of room for variation in the results.The results of an experiment are considered reliable only after they have been subjected to cross-validation.The study has five independent cross-validation cohorts.The outcomes are depicted in figure 2, figure 3, and figure 4.   .ACC result.Overall, using ResNet-50 as a foundation for product fault identification yields considerable improvements.Using experimental data, we can deduce that the data enhancement technique can successfully boost the defect identification algorithm's accuracy, and that the ResNet-50 picture classification method benefits even more from the strategy's implementation.The algorithm's accuracy has dropped on the dataset regardless of whether or not the data improvement approach is implemented.The technique uses a full-convolutional network block approach to compensate for imperfect data, allowing it to keep its superior detection effect even in the face of limited variety in the training set, as compared to the ResNet-50 image classification method.
Small faults are notoriously challenging for defect identification algorithms.This is due to two primary factors.To begin, little faults have less of an overall influence on the input image and a lower percentage of the image's area as compared to larger defects.Second, a minor defect might be hard to spot because pattern can be similar to noise pattern of real region.One crucial metric for judging a defect recognition algorithm is how well it identifies tiny flaws.Defect samples are considered tiny if their combined defect area is less than 100 pixels in the studies shown here.Similar to the last part, this one makes use of a cross-validation test.Comparing with ResNet-50, the results of the experiment demonstrate a significant enhancement in the detection performance of small defects using this method.

Conclusion
In conclusion, video image-based external defect detection techniques have emerged as a promising solution to address the limitations of traditional inspection methods for power transmission and distribution equipment.This paper has reviewed and analyzed the current state of video image-based defect detection, considering the challenges faced, methodologies employed, and their applications in the power industry.To solve the problem of flaw detection, this study looks at a defect identification strategy based only on convolutional block detection.This method adds the idea of block detection to the ResNet fully convolutional neural network.It is possible to implement the local discrimination mechanism, while avoiding the problems associated with the more common block detection receptive field.This method outperforms the original ResNet image classification system in terms of generalization and error detection.The findings of this study demonstrate the potential of utilizing video image analysis for accurate and efficient identification of external defects.By leveraging computer vision and image processing technologies, these techniques offer several advantages over manual inspections, including reduced inspection time, increased objectivity, and improved accuracy.However, there are still several challenges that need to be addressed.For instance, variations in lighting conditions, camera angles, and equipment configurations can affect the performance of videobased defect detection systems.Additionally, the detection of subtle or hidden defects remains a challenging task that requires further research and development.

Figure
Figure TNR result.

Figure 4
Figure 4. ACC result.Overall, using ResNet-50 as a foundation for product fault identification yields considerable improvements.Using experimental data, we can deduce that the data enhancement technique can successfully boost the defect identification algorithm's accuracy, and that the ResNet-50 picture classification method benefits even more from the strategy's implementation.The algorithm's accuracy has dropped on the dataset regardless of whether or not the data improvement approach is implemented.The technique uses a full-convolutional network block approach to compensate for imperfect data, allowing it to keep its superior detection effect even in the face of limited variety in the training set, as compared to the ResNet-50 image classification method.Small faults are notoriously challenging for defect identification algorithms.This is due to two primary factors.To begin, little faults have less of an overall influence on the input image and a lower percentage of the image's area as compared to larger defects.Second, a minor defect might be hard to spot because pattern can be similar to noise pattern of real region.One crucial metric for judging a defect recognition algorithm is how well it identifies tiny flaws.Defect samples are considered tiny if their combined defect area is less than 100 pixels in the studies shown here.Similar to the last part, this one makes use of a cross-validation test.Table1displays the aggregate results from the five crossvalidation test sets.Table1.Small defect recognition experiment.
Changing one's viewpoint is another typical technique for improving data quality.The basic structure of the original image will not shift even if the viewpoint is slightly altered.The formula for a viewpoint transformation is:

Table 1
displays the aggregate results from the five crossvalidation test sets.

Table 1 .
Small defect recognition experiment.