Time-Warping: A Time Series Data Augmentation of IMU Data for Construction Equipment Activity Identification

Abstract

Automated, real-time, and reliable equipment activity identification on construction sites can help to minimize idle times, improve operational efficiencies, and reduce emissions. Many previous efforts in activity identification have explored different machine learning algorithms that use time-series sensor data collected from inertial measurement units mounted on the equipment. However, machine learning algorithms requires large volume of training data collection from the field, as inadequate and smaller amounts of data results in model overfitting. This study proposes an automatic and real-time activity recognition framework by using data from multiple IMUs attached to equipment's moving and articulated parts. In doing so, first a time-series data augmentation technique called window-warping (WW) is introduced to generate synthetic training data from a smaller volume of field-collected data. Two supervised machine learning algorithms, artificial neural network (ANN), and K-nearest neighbour (KNN) were trained and evaluated using the augmented training data to identify equipment activity. The developed data augmentation methodology is validated using a case study of an earthmoving excavator. The results show the potential for using time-series data augmentation in training machine learning algorithms for construction equipment activity recognition using minimal data collected from the field.

An empirical survey of data augmentation for time series classification with neural networks

“…There are also hybrid methods that use multiple domains. It should be noted that multiple transformation techniques can be used to augment the data set in serial [30] and in parallel [41,42]. In the following subsections, we will detail each of these domains and the random transformation-based data augmentation methods associated with them.…”

“…where R is an element-wise random rotation matrix for angle y � N ð0; s 2 Þ for multivariate time series [30] and flipping for univariate time series [42]. While rotation data augmentation can create plausible patterns for image recognition, it might not be suitable for time series since rotating a time series can change the class associated with the original sample [47].…”

“…The scaling parameter α can be determined by a Gaussian distribution a � N ð1; s 2 Þ with σ as a hyperparameter [30], or it can be from a random value from a pre-defined set [42]. It should be noted that "scaling" in terms of time series is different than in the image domain.…”

“…For time series, it refers to just increasing the magnitude of the elements and not enlarging the time series. Some examples of using scaling as data augmentation include classification of sensor data [30,42]. Escano et al [49] and Tran and Choi [50] used a combination of scaling with jittering and element-wise interpolation for Gait recognition.…”

Deep neural networks for vehicle driving characterization by means of smartphone sensors

“…Imagen obtained from (Cui, Chen, & Chen, 2016). (Rashid & Louis, 2019) also use WW in time series. In the following image (Figure 5.35), an example speeding up and slowing down the activity X is illustrated.…”