Inertial sensor based human behavior recognition in modal testing using machine learning approach

Abstract

Adaptive phase control impact device (APCID) was developed for performing in-service modal analysis using impact synchronous modal analysis. However, this device is large and heavy, making it unsuitable for real world applications. This automated impact device can be replaced with human hand but the randomness in human behavior can reduce the accuracy of APCID control scheme. To replace APCID with a smart semi-automated device while still using APCID control scheme, machine learning models are presented in this paper to recognize human behavior by classifying 13 different impact types and predicting impact time using the impact classification. The impact classification model gave classification accuracy of over 96% with 130 real time impacts. With successful classification of different impact types, randomness in human behavior can be reduced by two to three times by associating a range of impact time with each impact type. However, the impact time ranges may differ person to person. To address this issue and to further reduce variations in impact time, a time prediction machine learning model was developed to make compensations in the control scheme of APCID by predicting impact time. The model gave reasonable accuracy with mean prediction errors of 5.2% in real time testing compared to measured time for 100 impacts.

Behavior Recognition of College Students Based on Improved Deep Learning Algorithm

“…used deep learning techniques to learn the sample data collected from in-vehicle sensors in their study, proposed a joint data enhancement scheme, designed a new multiview convolutional neural network (CNN) model to construct a sample dataset that better matches the complex real driving environment, and developed a new multiview convolutional neural network model for training, learning, and recognition of driving behavior. Zahid et al (2021) proposed a machine learning model to recognize human behavior by classifying 13 different crash types and using crash classification to predict crash times and developed a time prediction machine learning model to compensate the control scheme of APCID by predicting the collision time. Byeon et al (2021) proposed a four-stream integrated CNN based on ROI, where the data consisted mainly of images and skeletons, by converting 3D skeleton sequences into pose evolution images, inputting RGB videos into 3D-CNN to extract temporal and spatial features, limiting the body ROI of RGB video into 3D-CNN and RGB video limited to ROI of hand-object interaction into 3D-CNN.…”

“…Image classification is to classify images according to the semantic information contained in them. The steps of image classification are generally as follows: First, the features of the image are extracted, then the target model is obtained by training, and finally the extracted feature map is input into the target model for classification (Zahid et al, 2021). With the development of deep learning, more and more people gradually have studied the application of deep learning related technologies to image classification (Zheng et al, 2020).…”

Implementation of a Secure Storage Using Blockchain for PCA-FRF Sensor Data of Plate-Like Structures

“…For that reason, ISMA is able to measure the FRF data of in-service structures because of denoising properties. Several developments and improvements on the ISMA technique have been made over the decade [79]- [82], with the recent development method includes the scope of inertial sensor based human behavior recognition using machine learning approach, which enable semi-automated features for ISMA modal testing [83]. Generally, the phase difference information between acceleration response and cyclic load component suppresses the non-synchronous components, enabling ISMA to be conducted on a structure that is exposed to noise or is in-operation without compromising the quality of the modal parameters extracted [84], [85].…”

Behavior Recognition and Exhibition Tour Scene Classification Using In-Depth Learning

“…Mathai and other experts use a dual-resolution convolutional neural network to input and process video images into two sets of highly independent data streams, namely, the original resolution and the low resolution, which are alternately composed of regular, convolution, and pooling layers in feature extraction, and fuse the data on the full connection layer, which makes it easier to understand and identify the subsequent signs based on the results obtained [7]. Zahid et al scholars used human behavior recognition technology to train national team divers and established models to track and identify human behavior [8].…”