An Autoencoder-Based Deep Learning Approach for Load Identification in Structural Dynamics

Abstract

In civil engineering, different machine learning algorithms have been adopted to process the huge amount of data continuously acquired through sensor networks and solve inverse problems. Challenging issues linked to structural health monitoring or load identification are currently related to big data, consisting of structural vibration recordings shaped as a multivariate time series. Any algorithm should therefore allow an effective dimensionality reduction, retaining the informative content of data and inferring correlations within and across the time series. Within this framework, we propose a time series AutoEncoder (AE) employing inception modules and residual learning for the encoding and the decoding parts, and an extremely reduced latent representation specifically tailored to tackle load identification tasks. We discuss the choice of the dimensionality of this latent representation, considering the sources of variability in the recordings and the inverse-forward nature of the AE. To help setting the aforementioned dimensionality, the false nearest neighbor heuristics is also exploited. The reported numerical results, related to shear buildings excited by dynamic loadings, highlight the signal reconstruction capacity of the proposed AE, and the capability to accomplish the load identification task.

A Machine Learning Enabled Hall-Effect IoT-System for Monitoring Building Vibrations

“…For instance, in [8], the authors proposed a data-driven SHM method based on convolutional neural networks and fast Fourier transform to identify structural damage conditions from vibration data. An autoencoder architecture targeting nonlinear dimensionality reduction of input vibration signals has been recently introduced for the task of load identification in [9].…”

“…Our contribution with respect to [8], [9] is represented by the implementation of two deep learning models on the embedded platform chosen as reference, hence showing the possibility of executing non-trivial machine learning tasks directly on the hardware platform in charge of collecting vibration data; given the integration with fully-fledged IoT protocols, the system can be reconfigured on the fly to also support cloud-enabled machine learning. We also benchmark the performances of the two implemented neural networks in a hypothetical smart building application (namely counting people's steps during walking).…”

Parametric study on the accuracy of full-field reconstruction from sparse measurements using autoencoders

“…An AE network was also applied to study impulsive components in vibration signals for feature extraction and damage classification. For these applications, AE showed accuracy comparable to deep neural networks 10 and superior performance than traditional data fusion approaches 11 . The combination of multiple AEs, stacked either in parallel or series, showed improved performance compared to linear techniques to denoise accelerometer data 12 and classify the conditions of roller bearing 13 , especially when the AEs are pre-trained using data augmented with generative adversarial networks 14 .…”

“…In these contexts, AEs have been utilized to extract features from sensor data and identify damage or structural changes 7 . AE-based methods have shown promising results in applications such as motor bearings' fault detection via spectrogram generation 8 and loading condition prediction using a false nearest neighbor heuristic 9 . An AE network was also applied to study impulsive components in vibration signals for feature extraction and damage classification.…”

Combined Use of Cointegration Analysis and Robust Outlier Statistics to Improve Damage Detection in Real-World Structures

“…To solve the problem of inclusive outliers in the training data, robust statistical methods are considered in a wide range of researches, including SHM [ 13 , 14 , 15 ], to provide unbiased estimates of mean and covariance parameters computed from a smaller subset of data whose behaviour is assumed to be close to the true population values. Alternatively, data normalization techniques such as principal component analysis [ 9 ], autoencoders [ 16 ] or cointegration [ 17 , 18 , 19 , 20 , 21 ], are adopted to project the data into a different space to remove or at least reduce the effect of environmental and operational changes. Although previous methods proved their effectiveness in creating a normal condition training set clear from the influence of external factors, they are unable to distinguish which of the outliers indicated in the data are “benign” and which are “malign”.…”