2.1. Transmission Line Image Detection Model Based on YOLOv8-CBAM

Traditional image recognition mostly uses edge detection operators, such as Prewitt, Sobel, and Canny. These methods often rely heavily on the design of Softmax and are not universally applicable to solving computable problems ^[15]. YOLOv8 is a single-stage detecting algorithm using regression thinking, which can directly convolve the obtained image data and label the category of the image ^[16]. YOLOv8 not only has good recognition accuracy but also has high real-time performance, which can better adapt to image detection of transmission lines. Therefore, this study introduces YOLOv8 to detect foreign object images in transmission lines. Fig. 1 shows the structure of YOLOv8.

Fig. 1. Network structure of YOLOv8.

In Fig. 1, first, the input layer is responsible for inputting the foreign object image information of the transmission line to be detected into the network backbone module (CSPDarkNet-53). Then, the CSPDarkNet-53 module layer extracts features from the foreign object image information of the transmission line. Secondly, the feature enhancement module layer is responsible for pooling and fusing the input transmission line feature foreign object image information. Finally, the output layer outputs abnormal image information. The CSPDarkNet-53 module layer of YOLOv8 includes four modules: feature extraction layer (Focus), convolutional layer (Conv), Spatial Pyramid Pooling Fast (SPPF), and BottleneckCSP. Fig. 2 shows the CSPDarkNet-53 module layer of YOLOv8.

Fig. 2. CSPDarkNet-53 module.

In Fig. 2, the Focus module mainly consists of a 3×3 sized Conv, which is responsible for cutting the input transmission line foreign object image data into 4 parts and performing channel dimension concatenation and feature map convolution operations. The Focus module serves to reduce spatial resolution while simultaneously increasing the number of channels. This is achieved by reducing the costs associated with convolution computations and employing tensor reshaping operations. The Conv module is responsible for 2D convolution, Batch Normalization (BN), and parity functions. The BottleneckCSP module is responsible for extracting deep semantic information from transmission line images. SPPF is mainly composed of three 5×5 convolutional kernels, responsible for increasing receptive fields and separating important contextual features. BN calculation is represented by Eq. (1).

In Eq. (1), $\mu_{B}$ and $\sigma_{B}^{2}$ correspond to tensors to calculate the mean and variance, respectively. $x_{i}^{*}$ represents the normalized output value. $x_{i}$ represents the values of the elements in the input feature map. $\varepsilon$ is a value that prevents variance from being 0. However, the values after performing BN operations will be concentrated around 0. Therefore, to enhance the nonlinearity of the network, the BN calculation equation is scaled and translated. The BN calculation after scaling and translation is represented by Eq. (2).

In Eq. (2), $\gamma$ and $\beta$ represent the scaling ratio and offset, respectively. The maximum pooling operation is represented by Eq. (3).

In Eq. (3), $y_{m,n}^{d}$ refers to the maximum output transmission line characteristic map value. $R_{m,n}^{d}$ refers to the area of the transmission line characteristic map. The calculation of Conv is represented by Eq. (4).

In Eq. (4), $N$ represents the size of the output transmission line characteristic map. $M$, $F$, $P$, and $S$ correspond to the input image size, convolution kernel size, fill factor, and convolution kernel movement step in the transmission line, respectively. However, YOLOv8 still cannot meet the detection requirements for foreign object images on transmission lines. The attention mechanism can facilitate YOLOv8 to successfully extract feature image information from transmission line diagrams and suppress irrelevant image features, greatly improving image recognition efficiency ^{[17, 18]}. As a lightweight attention module, CBAM can effectively reduce the number of parameters and computational resources required while improving the overall performance of the model. This enables YOLOv8 to identify and locate targets with greater accuracy when processing images. Accordingly, the study introduces the CBAM into the YOLOv8 network algorithm, enhancing the CSPDarkNet-53 layer within the CBAM module and the C2f module within the Neck layer. Furthermore, it proposes a transmission line image detection model based on the improved YOLOv8 network algorithm. Fig. 3 shows the proposed YOLOv8-CBAM transmission line image detection model.

Fig. 3. The YOLOv8-CBAM transmission line image detection model.

In Fig. 3, the proposed YOLOv8-CBAM transmission line image detection model mainly consists of four parts: CSPDarkNet-53, SPPF, multi-scale fusion structure, and YOLO v8 detection head. Firstly, the model extracts transmission line image features from the CSPDarkNet-53 layer and fixes the image size. The improved C2f module extracts channel information and spatial information from the input transmission line feature map. This can further enhance the ability of multi-scale feature fusion, better capture feature expressions at different levels, and improve the accuracy and stability of object detection networks. Then, a fixed image size feature map is used to construct a feature pyramid and transmit semantic information forward. Finally, the deep features will be upsampled and fused with the features of the previous layer, resulting in higher resolution and richer feature representations. This improved process not only enhances the resolution of the features but also leads to a richer and more comprehensive feature representation, which provides more accurate target localization and classification information for the final detector head. The output feature value is represented by Eq. (5).

In Eq. (5), $w$ represents the convolutional kernel's weight value. $x$ and $R$ correspond to the characteristic map and sampling point interval positions of the transmission line to be tested, respectively. $p_{0}$ and $p_{n}$ correspond to output feature values at a certain sampling point, respectively. $R$ is a positional element. The overall implementation process of CBAM is represented by Eq. (6).

In Eq. (6), $M_{x}$ and $A_{x}$ represent maximum pooling and average pooling, respectively. $f$ represents the fully connected layer.

2.2. Construction of External Force Damage Prevention System for Transmission Lines

The transmission line image detection model based on YOLOv8-CBAM can recognize the damage caused by foreign objects in transmission line equipment through image recognition. However, it is not yet sufficient to address the prevention of transmission line network security issues. Therefore, a transmission line image detecting model using YOLOv8-CBAM is utilized as the basic framework. A new transmission line external force damage prevention system integrating YOLOv8 and IDA is constructed using IDA. Fig. 4 shows the basic process of IDA.

Fig. 4. Basic flow of intrusion detection algorithms.

In Fig. 4, IDA mainly includes the stages of information analysis, response decision-making, and information collection. Firstly, in the information collection stage, the network raw data information is collected and preprocessed according to the set rules, providing data level support for the subsequent information analysis. Secondly, in the information analysis stage, in-depth analysis is conducted on the data processed in the information collection stage to make accurate judgments on intrusion behavior. Finally, during the response decision stage, intrusion behavior is immediately prevented in accordance with IDA's pre-set methods, such as prohibiting IP access, forcing shutdown, etc. Fig. 5 shows the proposed transmission line external force damage prevention system.

Fig. 5. Structure of the transmission line external damage prevention system.

In Fig. 5, the proposed transmission line external force damage prevention system structure mainly consists of four parts, namely data preprocessing, mixed sampling, improved YOLOv8 transmission line image detection model, and classification results. Firstly, the transmission line foreign object image dataset is used as the test data source. The data preprocessing part is responsible for merging the input data information to ensure that the transmission line foreign object damage prevention system model can learn adequately. This includes cropping, scaling, and color correction of the image data to fit the model inputs. To better train the transmission line external damage prevention system model as well as not affecting the classification results, the study also uses the Maximum-Minimum (Max-Min) method for numerical and normalization operations to newly divide the data. This step is designed to eliminate the effect of different scales, thereby facilitating the comparison of data at a uniform scale. This, in turn, enhances the efficiency and effectiveness of model training. At the same time, the image data are enhanced by rotating, scaling, and shearing techniques to increase the diversity of the dataset and the generalization ability of the model. Due to the possible imbalance between the various types of samples in the transmission line foreign object image dataset, a mixed sampling method is used to balance the dataset to ensure that the model is not biased against a particular category. Moreover, the Principal Component Analysis (PCA) algorithm is used to downscale the data information to reduce the dimensionality of the data and retain the most important features. Subsequently, the processed data are fed into the YOLOv8 network for model training. During the training process, the model is evaluated using a validation set to monitor the performance of the model and prevent overfitting. Finally, the classification results are obtained through the transmission line image detection model. The calculation of Max-Min is represented by Eq. (7).

In Eq. (7), $x^{'}$ represents the result of normalizing the characteristics of the transmission line data. $M_{min}$ and $M_{max}$ correspond to the minimum and maximum values after feature processing, respectively. PCA information content is generally measured by removing mean, calculating contribution rate, mean deviation, and covariance matrix. The mean removal of data feature decentralization is represented by Eq. (8).

In Eq. (8), $\mu$ and $n$ represent mean removal and sample data quantity, respectively. The mean deviation is represented by Eq. (9).

In Eq. (9), $\phi_{i}$ represents the deviation from the mean $\mu$ for each sample $x_{i}$. The covariance matrix is represented by Eq. (10).

In Eq. (10), $S$ represents the dataset sample's covariance matrix. $T$ represents matrix. The calculation contribution rate $\eta$ is represented by Eq. (11).

In Eq. (11), $m$ and $k$ represent matrix positions. $\alpha_{i}$ represents an eigenvector. $n$ represents the number of samples. Fig. 6 shows the training of the proposed transmission line external force damage prevention system.

Fig. 6. Training process.

In Fig. 6, the proposed system mainly has 5 training processes. Firstly, the input data are processed using mixed sampling and PCA, with an 8*8 grayscale image as the input layer. Then, Conv is used to extract features from grayscale images. BN performs batch normalization on the extracted feature data to accelerate the convergence speed. Secondly, maximum pooling is used to pool the processed data. The fully connected layer is used to recombine the output features to reduce feature loss. Finally, the data information is output using Softmax.

3.1. Performance Testing of Transmission Line Image Detection Model

A suitable experimental environment was established to test the proposed transmission line external force damage prevention system. The Windows 10 operating system and Python language were used. TensorFlow was used to build a network model. The CPU was Intel Core i7, the GPU was NVIDIA GeForce, and the memory was 64GB. The iterations were 80, the learning rate was $10^{-3}$, the batch size was 6, and the IoU threshold was 0.3. The dataset of foreign object images on transmission lines was divided into training and testing sets in an 8:2 ratio. Table 1 shows the specific quantity of various foreign object images in the transmission line foreign object image dataset.

In Table 1, a batch of 6010 images were collected using image processing techniques as a dataset of foreign object images, including 8 types of foreign object images of transmission lines. The foreign object images in the dataset mainly include: plastic bags, ice and snow, bird nests, advertising ribbons, kites, balloons, leaves, and banners. In addition, the study also introduced commonly used image detection models and compared their accuracy with the proposed YOLOv8-CBAM image detection model, such as YOLOv8, YOLOv5, and YOLOv5-CBAM image detection models. Fig. 7 shows foreign object images' detecting accuracy on transmission lines using four different models.

Fig. 7. Model detection accuracy of kite and balloon foreign object images.

Fig. 7(a) shows the variation curve of foreign object image detecting accuracy in transmission lines for four models under the training set. Fig. 7(b) shows the variation curve of the accuracy for four models under the test set. As the sample size continued to increase, the detection accuracy of all four models showed a trend of slowly increasing first and then stabilizing. The detection accuracy of YOLOv8, YOLOv5, YOLOv5-CBAM and the proposed YOLOv8-CBAM on the training set was 88.15%, 82.31%, 91.24%, and 98.88%, respectively. The detection accuracy in the test set was 87.24%, 79.87%, 90.58%, and 97.64%, respectively. By comparison, the proposed YOLOv8-CBAM image detection model performed the best. Compared with the unimproved YOLOv8, the proposed model's detecting accuracy in two sets was improved by 10.73% and 10.40%, respectively. Therefore, the improvement of YOLOv8 had a positive effect on the model's overall performance, with significant effect. In addition, the study conducted classification tests on foreign object images of bird nests, balloons, kites, and leaves using classification accuracy as a test indicator to verify YOLOv8-CBAM's classification effect on foreign object images. Fig. 8 shows the test results.

Fig. 8. Model classification accuracy.

Figs. 8(a) and 8(b) show the YOLOv8 and YOLOv8-CBAM image detection model's classification accuracy for foreign objects images: bird nests, balloons, kites, and leaves. The highest classification accuracy of the YOLOv8 image detection model for images of bird nests, balloons, kites, and leaves was 86.79%, 87.12%, 79.58%, and 88.12%, respectively. The highest classification accuracy of the YOLOv8-CBAM-PCA image detection model for images of bird nests, balloons, kites, and leaves was 95.45%, 89.78%, 90.02%, and 96.03%, respectively. Overall, the proposed YOLOv8-CBAM image detection model had higher classification accuracy for images of bird nests, balloons, kites, and leaves compared to the YOLOv8 image detection model. The proposed YOLOv8-CBAM test curve was also closer to the ideal curve, which was more in line with the current detection of foreign object images in transmission lines, having certain feasibility and effectiveness.

3.2. Simulation Testing of External Force Damage Prevention System for Transmission Lines

The study used DD Cup 99 and NSL-KDD datasets as testing data sources. DD Cup 99 is a well-known dataset in the field of network security, which contains a large number of network connection records. Each record has a detailed feature description, including time and content-based features and labels indicating whether each connection belongs to an attack. NSL-KDD includes over 20,000 pieces of network attack data information, including DOS, U2R, R2L, etc. The study selected four types of network attacks as inputs: Normal, Denial of Service (DOS), illegal access from remote machines (R2L), and illegal access from local superuser privileges by ordinary users (U2R). The classification performance of network attack types for four models was tested in Fig. 9.

Fig. 9. Effectiveness of the model in categorizing types of network attacks.

Figs. 9(a) and 9(b) show the classification test results of four transmission line external force damage prevention systems against Normal, DOS, R2L, and U2R attacks in DD Cup 99 and NSL-KDD. After multiple tests, the proposed new YOLOv8-CBAM-PCA-intrusion transmission line external force damage prevention system had a higher accuracy in network attack classification. The highest classification accuracy of YOLOv8-CBAM-PCA-intrusion for four network attacks: Normal, DOS, R2L, and U2R was 97.36%, 95.23%, 96.31%, and 95.01%, respectively. Compared with the unimproved YOLOv8, the classification accuracy increased by 28.59%, 16.32%, 18.13%, and 8.85%, respectively. In addition, the study also conducted comparative experiments on the predicted and actual values' differences of these four network attacks mentioned above in Fig. 10.

Fig. 10. Discrepancy between the predicted and true values.

Figs. 10(a) and 10(b) show the difference between the predicted and true values of the four network attacks of YOLOv8 and YOLOv8-CBAM-PCA-intrusion. In Fig. 10, the difference between the predicted and actual values of the four types of network attacks on the YOLOv8-BAM-PCA intrusion transmission line external force damage protection system was lower than that of YOLOv8. The predicted and true values' difference of the Normal, DOS, R2L, and U2R network attack modes of YOLOv8-CBAM-PCA-intrusion was only 0.2%, 1.1%, 0.9%, and 0.1%, respectively. The above data indicate that the proposed model has good robustness and is more adaptable to changes in data. Finally, the study also conducted a multi-indicator test on the YOLOv8-CBAM-PCA-intrusion system using Precision (P), Recall (R), F1 value, and average detection time as indicators. The results are shown in Table 2.

In Table 2, among these four indicators detected, the performance of the YOLOv8 system without improvement was the worst, followed by YOLOv8-CBAM, YOLOv8-CBAM-PCA, and YOLOv8-CBAM-PCA-intrusion systems. The highest P of YOLOv8 was 90.08%, the highest R was 88.76%, the highest F1 value was 88.35%, and the average detection time was 5.53s. The proposed new transmission line external force damage prevention system had a maximum P of 98.15%, a maximum R of 98.73%, a maximum F1 value of 98.02%, and an average detection time of 2.53 s. In addition, it is also the research model that performs best when compared to the rest of the commonly used transmission line external damage prevention models. Therefore, the proposed YOLOv8-CBAM-PCA-intrusion had relatively good performance and was more suitable for the current prevention of external damage to transmission lines.