Automated peak picking using region‐based convolutional neural network for operational modal analysis

Abstract

Peak picking is used to determine locations of salient peaks in a graphical representation of a physical quantity. It is often used to extract possible natural frequencies from the frequency domain representation of structural responses.One of the challenges in peak picking is to establish a method to automatically distinguish peaks from data containing noise peaks. As selecting peaks intrinsically depends on human perception, algorithms for automated peak picking have exhibited only partial success to date. This paper presents an automated peak picking method that utilizes a region-based convolutional neural network with possible object locations to accurately identify the peaks. A tailored deep learning architecture is proposed, which is subsequently trained using only the peaks numerically generated by computer software. Laboratory and field tests are conducted using acceleration responses obtained from three test structures of beam, truss, and stay cable to verify the identification performance of the proposed peak detector. The proposed peak detector is shown to successfully identify most of the salient peaks with high accuracy.parameters of the physical modes are consistently found regardless of the provided system orders. 15 The modal parameters calculated from SSI or ERA with different system orders are grouped based on their similarity to each other, from which the frequently obtained are selected as the physical modes. [16][17][18][19] Neu et al. 20 and Cabboi et al. 21 further improved the approach in terms of automation by minimizing user-defined parameters. In addition, Afshar and Khodaygan 22 presented an enhanced stabilization chart for automated modal analysis. However, the identification performance is still highly dependent on the maximum system order, of which an optimal value should be provided as a priori information.Peak picking associated with frequency domain approaches has also been used to carry out automated modal analysis. [23][24][25] Peak picking on the frequency domain representation is considered as an intuitive way for operational modal analysis, as natural modes generally have salient peaks. A general approach for automated peak picking in the literature is to select maximum values of frequency domain representations of the structural responses (e.g., Fourier transform and power spectrum) from each of the user-defined subfrequency ranges, which are believed to include peaks. 26,27 This method is valid only when the subfrequency ranges are known; changes in peak frequencies due to structural damage or environmental effects can lead to the selection of undesired peaks. A user-defined threshold value was additionally introduced to filter out less important peaks in each subfrequency range. [28][29][30][31] This approach ignores peaks that are lower than the user-defined threshold and thus reduces the number of false positives; however, the imposition of thresholds for the peak amplitude and subfrequency range critically limits the practical applications of the approach to different ...

Automated wireless monitoring system for cable tension forces using deep learning

“…However, the user-defined threshold is dependent on the type of civil structure and may fail to remove undesirable peaks due to noise. [18][19][20] Peak-picking that identifies local maxima near predefined natural frequencies may identify undesirable peaks when the modal properties of the cable are altered due to structural damage resulting from environmental effects. 18 Currently, deep learning has shown considerable potential for automated peak-picking based on frequency domain representations of structural responses, making it possible to extract natural frequencies automatically.…”

“…[18][19][20] Peak-picking that identifies local maxima near predefined natural frequencies may identify undesirable peaks when the modal properties of the cable are altered due to structural damage resulting from environmental effects. 18 Currently, deep learning has shown considerable potential for automated peak-picking based on frequency domain representations of structural responses, making it possible to extract natural frequencies automatically. 18 However, the proposed peak-picking method may detect undesirable peaks when applied to the cable response, because this method is tailored to deal with responses from bridge components, such as beams, trusses, and cables.…”

“…18 Currently, deep learning has shown considerable potential for automated peak-picking based on frequency domain representations of structural responses, making it possible to extract natural frequencies automatically. 18 However, the proposed peak-picking method may detect undesirable peaks when applied to the cable response, because this method is tailored to deal with responses from bridge components, such as beams, trusses, and cables. In addition, the proposed technique requires high computation power to perform peakpicking, which can burden the embedded process of a system.…”

Identification of Grout Sleeve Joint Defect in Prefabricated Structures Using Deep Learning

“…A spatial timefrequency data set was established using multi-dimensional vibration signals, and different railway events were classified by monitoring data that contained environmental noise. Kim and Sim (2019) proposed a framework composed of a fast R-CNN and a region-suggestion network based on deep learning that can automatically extract peaks in frequency-domain pattern recognition. Tang et al (2020) presented an overview of recognition and localization methods for vision-based fruitpicking robots.…”

Vibration-Based Monitoring for Transverse Cooperative Working Performance of Assembled Concrete Multi-Girder Bridge: System Design, Implementation and Preliminary Application

“…have a similar form as matrix singular value decomposition (SVD)[40]. Therefore it's possible to pick peaks from the measured vibration signals' singular value spectrum using manual and automated algorithms[41] to identify modal frequency while the corresponding singular vectors is taken as mode shape. Generally, the first singular walue is considered in PSD matrix(dB) 1st singular value 2nd singular value 3rd singular value (b) Auto-power spectrum of strain Singular value spectrum of monitoring data: black points are peaks picked manualy this process.…”