Transmission Line State Recognition Method Based on Dual-Branch Convolution Neural Network Structure and Multi-Attention Mechanism
Objective Real-time monitoring and evaluation of transmission lines are essential to ensure the safety and stability of the power grid.Phase-sensitive optical time domain reflectometry(Φ-OTDR)offers advantages such as long detection range,strong resistance to electromagnetic interference,and low cost.It can directly utilize redundant fibers in optical fiber composite overhead ground wire(OPGW)and can effectively determine the operational status of transmission lines.This technology has broad application prospects in the field of transmission line status monitoring.Currently,the commonly used signal recognition methods in optical fiber vibration sensing are primarily based on machine learning and deep learning.In traditional machine learning,signal features are often extracted based on human experience before being input into a classifier,meaning that recognition performance is heavily dependent on the selection and quantity of these features.Deep learning,however,can automatically extract and select useful features from raw data,reducing the influence of manual intervention on recognition accuracy.Nonetheless,most existing deep learning methods primarily analyze a single mode of the vibration signal,limiting the model's ability to fully extract effective feature information.To more comprehensively describe signal characteristics and improve model recognition performance,we propose a transmission line signal recognition method that integrates one-dimensional time-domain signal analysis with two-dimensional image signal analysis,using a dual-branch convolution neural network(CNN)structure and a multi-attention mechanism.Six types of on-site data from transmission lines under different operating environments are collected using the Φ-OTDR system to construct a dataset for algorithm verification.The proposed method's recognition rate is compared with five other deep learning models,and the recognition performance of machine learning and deep learning methods on large-scale datasets is also analyzed.Methods In this model,the time-sequence branch uses a one-dimensional convolutional neural network(1DCNN)to directly extract temporal features from the raw vibration signal,while the image branch uses continuous wavelet transform(CWT)to convert the vibration signal into a two-dimensional time-frequency image,which is then processed by a two-dimensional convolutional neural network(2DCNN)to learn the time-frequency features of the image.To enhance the focus on critical information,channel attention mechanisms are incorporated into both branches.The proposed branch attention mechanism effectively addresses the issue of insufficient recognition accuracy caused by differences in modal information during feature fusion by assigning weights to branch features.Parameter comparison experiments are conducted to select the appropriate wavelet function for the dataset,comparing the effects of different wavelet functions on final recognition accuracy(Fig.8).In addition,ablation experiments are designed to verify the effectiveness of the attention mechanism included in the algorithm(Fig.9).The proposed method is also compared with five deep learning models:single branch 1DCNN,single branch 2DCNN,1DCNN-LSTM,1DCNN BiLSTM,and ATCN-SA BiLSTM.Furthermore,it is compared with three common machine learning methods:SVM,KNN,and decision tree.Before being input into the machine learning classifier,the six types of vibration signals are decomposed using variational mode decomposition with seven layers,and features are extracted for each decomposed modal component,resulting in four datasets with different features(Table 7).Results and Discussions In the parameter comparison experiment,the model achieves the highest recognition accuracy for the time-frequency image data set using the Morlet wavelet function(Fig.8,Table 4).In the ablation experiments,the model without the added attention mechanism has the lowest recognition accuracy,while our method,which includes channel and branch attention mechanisms,achieves the highest recognition accuracy by enabling the model to focus on key information(Fig.9,Table 5).The recognition performance is measured by comparing the accuracy,precision,recall,and F1 scores of different deep learning models on the test set(Fig.10,Fig.11,Table 6).The results demonstrate that the proposed method,by integrating two different types of modal information,compensates for the limitations of one-dimensional vibration signal and two-dimensional image data,providing a more comprehensive description of data characteristics.The classification accuracy on the test set reaches 94.92%,and the overall evaluation indices are optimal.In addition,a comparison between 1DCNN and 2DCNN shows that 2DCNN offers a modest improvement in recognition accuracy and a slight advantage in convergence speed,suggesting that converting one-dimensional signals into two-dimensional images can improve recognition performance to some extent.In the comparison experiment with machine learning methods,the machine learning methods are trained and tested using the extracted four feature datasets(Fig.12).The results show that the recognition performance of the machine learning method is significantly affected by the type and number of extracted features,and their recognition accuracy is far lower than that of our method,highlighting the insufficient generalization performance and classification efficacy of machine learning algorithms when processing large-scale datasets.Conclusions To address the challenge of real-time monitoring of transmission line operational status,we propose a deep learning model based on a dual-branch CNN structure and multi-attention mechanism.The model's time-sequence branch extracts time-domain features from vibration signals,while the image branch extracts time-frequency features from two-dimensional images.Channel and branch attention mechanisms are incorporated to enhance the model's focus on key feature information.On a real transmission line vibration signal dataset,the proposed method achieves a recognition accuracy of 94.92%,outperforming single branch networks,1DCNN-LSTM,1DCNN-BiLSTM,and ATCN-SA-BiLSTM models.When handling large-scale datasets,the proposed method demonstrates superior generalization and recognition performance compared to machine learning methods.The method provides a valuable reference for monitoring the operational conditions of transmission lines.
NETL
NSTL
万方数据
