A Deep Learning Approach for Autonomous Compression Damage Identification in Fiber-Reinforced Concrete Using Piezoelectric Lead Zirconate Titanate Transducers

Effective damage identification is paramount to evaluating safety conditions and preventing catastrophic failures of concrete structures. Although various methods have been introduced in the literature, developing robust and reliable structural health monitoring (SHM) procedures remains an open research challenge. This study proposes a new approach utilizing a 1-D convolution neural network to identify the formation of cracks from the raw electromechanical impedance (EMI) signature of externally bonded piezoelectric lead zirconate titanate (PZT) transducers. Externally bonded PZT transducers were used to determine the EMI signature of fiber-reinforced concrete specimens subjected to monotonous and repeatable compression loading. A leave-one-specimen-out cross-validation scenario was adopted for the proposed SHM approach for a stricter and more realistic validation procedure. The experimental study and the obtained results clearly demonstrate the capacity of the introduced approach to provide autonomous and reliable damage identification in a PZT-enabled SHM system, with a mean accuracy of 95.24% and a standard deviation of 5.64%.

Do Neural Network Weights Account for Classes Centers?

The exploitation of deep neural networks (DNNs) as descriptors in feature learning challenges enjoys apparent popularity over the past few years. The above tendency focuses on the development of effective loss functions that ensure both high feature discrimination among different classes, as well as low geodesic distance between the feature vectors of a given class. The vast majority of the contemporary works rely their formulation on an empirical assumption about the feature space of a network's last hidden layer, claiming that the weight vector of a class accounts for its geometrical center in the studied space. This article at hand follows a theoretical approach and indicates that the aforementioned hypothesis is not exclusively met. This fact raises stability issues regarding the training procedure of a DNN, as shown in our experimental study. Consequently, a specific symmetry is proposed and studied both analytically and empirically that satisfies the above assumption, addressing the established convergence issues. More specifically, the aforementioned symmetry suggests that all weight vectors are unit, coplanar, and their vector summation equals zero. Such a layout is proven to ensure a more stable learning curve compared against the corresponding ones succeeded by popular models in the field of feature learning.

Gait analysis comparison between manual marking, 2D pose estimation algorithms, and 3D marker-based system

Introduction

Recent advances in Artificial Intelligence (AI) and Computer Vision (CV) have led to automated pose estimation algorithms using simple 2D videos. This has created the potential to perform kinematic measurements without the need for specialized, and often expensive, equipment. Even though there's a growing body of literature on the development and validation of such algorithms for practical use, they haven't been adopted by health professionals. As a result, manual video annotation tools remain pretty common. Part of the reason is that the pose estimation modules can be erratic, producing errors that are difficult to rectify. Because of that, health professionals prefer the use of tried and true methods despite the time and cost savings pose estimation can offer.

Methods

In this work, the gait cycle of a sample of the elderly population on a split-belt treadmill is examined. The Openpose (OP) and Mediapipe (MP) AI pose estimation algorithms are compared to joint kinematics from a marker-based 3D motion capture system (Vicon), as well as from a video annotation tool designed for biomechanics (Kinovea). Bland-Altman (B-A) graphs and Statistical Parametric Mapping (SPM) are used to identify regions of statistically significant difference.

Results

Results showed that pose estimation can achieve motion tracking comparable to marker-based systems but struggle to identify joints that exhibit small, but crucial motion.

Discussion

Joints such as the ankle, can suffer from misidentification of their anatomical landmarks. Manual tools don't have that problem, but the user will introduce a static offset across the measurements. It is proposed that an AI-powered video annotation tool that allows the user to correct errors would bring the benefits of pose estimation to professionals at a low cost.

Deep Feature Space: A Geometrical Perspective

One of the most prominent attributes of Neural Networks (NNs) constitutes their capability of learning to extract robust and descriptive features from high dimensional data, like images. Hence, such an ability renders their exploitation as feature extractors particularly frequent in an abundance of modern reasoning systems. Their application scope mainly includes complex cascade tasks, like multi-modal recognition and deep Reinforcement Learning (RL). However, NNs induce implicit biases that are difficult to avoid or to deal with and are not met in traditional image descriptors. Moreover, the lack of knowledge for describing the intra-layer properties -and thus their general behavior- restricts the further applicability of the extracted features. With the paper at hand, a novel way of visualizing and understanding the vector space before the NNs' output layer is presented, aiming to enlighten the deep feature vectors' properties under classification tasks. Main attention is paid to the nature of overfitting in the feature space and its adverse effect on further exploitation. We present the findings that can be derived from our model's formulation and we evaluate them on realistic recognition scenarios, proving its prominence by improving the obtained results.