In an article published in Sensors, researchers proposed a novel endpoint detection algorithm and a feature fusion technique based on a multi-scale feature fusion two-dimensional convolutional neural network (2DCNN) to identify fiber vibration signals.
Study: Optical Fiber Vibration Signal Recognition Based on the Fusion of Multi–Scale Features. Image Credit: AerialVision_it/Shutterstock.com
The Sagnac type distributed optical fiber sensing system suffered from the drawback of low precision in the intrusion vibration recognition events. The method demonstrated in the current paper combined the conventional optical fiber vibration signal recognition with the features of automatic feature fusion by a 2DCNN to achieve higher accuracy.
A new endpoint detection algorithm was employed to detect the original signal's vibration component to increase the effectiveness of endpoint detection. This algorithm combined the energy spectral entropy and the spectral centroid product to achieve the desired detection effect.
The signal's multi-scale and multi-level features were extracted using a 2DCNN of various scales. A novel technique for integrating differential pooling features was employed to address information loss during the pooling process.
The extracted features were recognized using a multi-layer perceptron (MLP). Experiments demonstrated that the efficiency of distributed optical fiber sensing systems was higher when compared to conventional variational mode decomposition (VMD) and empirical mode decomposition (EMD) pattern recognition and one-dimensional CNN (1D-CNN) approaches.
The Growing Need for Reliable Perimeter Security Systems
Demands for secure and dependable perimeter security systems are becoming increasingly urgent due to the rapid technological and scientific advancements and the rise in living standards. Thus, optical fiber perimeter security has drawn significant interest from researchers worldwide, a technology with significant growth potential.
Optical fiber perimeter security systems have been extensively employed in tunnel detection, border monitoring, offshore oil exploration, seismic monitoring, marine monitoring, perimeter security, and oil pipeline monitoring.
The Sagnac interferometer-based distributed optical fiber sensing system is small in size with high sensitivity and easy installation.
A distributed optical fiber sensing system’s basic construction and symmetrical design make fiber vibration signal detection appropriate for distributed deployment without requiring a high degree of light source coherence. There is no requirement for a reference fiber. Therefore, the recognition of the Sagnac type distributed optical fiber sensing sensor was performed in the current research.
The three basic processes in identifying optical fiber vibration signals are feature fusion extraction, classification, and preprocessing of optical fiber vibration signals. Denoising the vibration signal and endpoint detection are the primary goals of preprocessing.
Conventional optical fiber vibration signal recognition is primarily constrained to the inherent method of extracting features and then identifying and classifying, which frustrates the relationship between the two and results in the loss of certain signal information. The single-scale CNN can only extract a portion of the signal's information and is unable to extract data from various signal scales.
The researchers demonstrated the distributed optical fiber sensing technique for identifying optical fiber vibration signals in the present paper. The endpoint detection effect was effectively improved by a new endpoint detection algorithm that combined energy spectral entropy and spectral centroid products. This algorithm also extracted features at various scales and levels of the 2DCNN. It applied a novel technique for differential pooling features and feature fusion to further increase accuracy in distributed optical fiber sensing.
Constructing the Enhanced Vibration Signal Recognition Mechanism
Signal preprocessing, multi-scale feature fusion extraction, differential pooling features, and classification recognition made up the fundamental framework of the optical fiber vibration signal detection based on the feature fusion of multi-scale features.
Numerous ineffective silent signals were found in the unrefined data of the optical fiber vibration signal. Therefore, the endpoint detection of the original optical fiber vibration signals was conducted before creating the experimental data set. The vibration section of the signal was then captured to decrease the analysis of silent signals and enhance the operating rate. Signal endpoint detection was primarily utilized to ascertain the beginning and ending points of the vibration portion of the signal.
A 2DCNN classified and recognized the preprocessed optical fiber vibration signal to more comprehensively represent the exhaustive information of the original signal, enhance the recognition rate, and minimize the error rate.
A 2DCNN required less consideration of parameters than other shallow or deep neural networks, and the hidden layers of a 2DCNN processed input data in the filtered form. The 2DCNN network could intelligently extract pertinent features from the convolutional and differential pooling features layers and automatically reduce noise in the signal.
The present work proposed a differential pooling features structure that integrated average pooling and maximum pooling to make the recovered vibration signal features more thorough and discriminative, enhancing the precision of optical fiber vibration signal recognition. The differential pooling features minimized the impact of background noise on the signal while collecting the specific information of the fiber vibration signal, combining the benefits of maximum pooling and average pooling.
Comparative experiments were conducted using the gathered data sets to confirm the efficiency of the distributed optical fiber sensing method. In the experiments of the current work, the multi-scale feature fusion technique (Fusion 2DCNN) was compared to the first channel CNN (First 2DCNN), the dual-channel CNN (FS 2DCNN), and the second channel CNN (Second 2DCNN).
The indicators of the multi-scale feature fusion (Fusion 2DCNN) approach were better than those of the FS 2DCNN method, demonstrating that the differential pooling features approach combined the benefits of average pooling and maximum pooling. Thus, it was observed that the differential pooling features approach was critical in enhancing the performance of distributed optical fiber sensing.
The Novel Vibration Signal Recognition and a Safer Tomorrow
An innovative method for identifying fiber vibration signals was put forth in the present paper based on a new endpoint detection technique and the multiple scale feature fusion approach. The endpoint detection algorithm was based on combining the energy spectral entropy and spectral centroid product and merging their respective advantages. Therefore, the endpoint detection method effectively improved the detection precision of optical fiber vibration signals and had an improved detection impact on low-frequency signals.
The running and walking signals, which were difficult to distinguish due to the brief vibration time, could be effectively differentiated by the signal intercepted for 1 s. The method of multi-scale feature fusion was used to more effectively extract the multi-level and multi-scale feature information of optical fiber vibration signals, and MLP was utilized for classification and recognition.
The comparison of VMD and EMD conventional pattern recognition techniques, 1D-CNN, and multi-scale feature fusion methods demonstrated that the average precision of the method proposed in the present work was higher than the other traditional methods. Therefore, the distributed optical fiber sensing system can effectively reduce the false alarms of optical fiber vibration signal recognition and accurately identify the four kinds of optical fiber vibration signals for knocking, running, walking, and flapping.
Reference
X. Ma, J. Mo, J. Zhang, J. Huang, Optical Fiber Vibration Signal Recognition Based on the Fusion of Multi–Scale Features. 2022. Sensors.http://www.mdpi.com/1424-8220/22/16/6012/htm
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.