Zero velocity update is a common and efficient approach to bound the accumulated error growth for foot-mounted inertial navigation system. Thus a robust zero velocity detector (ZVD) for all kinds of locomotion is needed for high accuracy pedestrian navigation systems. In this paper, we investigate two machine learning-based ZVDs: Histogram-based Gradient Boosting (HGB) and Random Forest (RF), aiming at adapting to different motion types while reducing the computation costs compared to the deep learning-based detectors. A complete data pre-processing procedure, including a feature engineering study and data augmentation techniques, is proposed. A motion classifier based on HGB is used to distinguish “single support” and “double float” motions. This concept is different from the traditional locomotion classification (walking, running, stair climbing) since it merges similar motions into the same class. The proposed ZVDs are evaluated with inertial data collected by two subjects over a 1.8 km indoor/outdoor path with different motions and speeds. The results show that without huge training dataset, these two machine learning-based ZVDs achieve better performances (55 cm positioning accuracy) and lower computational costs than the two deep learning-based Long Short-Term Memory methods (1.21 m positioning accuracy).

https://hal.archives-ouvertes.fr/hal-02879906/document

Machine Learning-Based Zero-Velocity Detection for Inertial Pedestrian Navigation

Software defect prediction is an active research area in software engineering. Accurate prediction of software defects assists software engineers in guiding software quality assurance activities. In machine learning, ensemble learning has been proven to improve the prediction performance over individual machine learning models. Recently, many Tree-based ensembles have been proposed in the literature, and their prediction capabilities were not investigated in defect prediction. In this paper, we will empirically investigate the prediction performance of seven Tree-based ensembles in defect prediction. Two ensembles are classified as bagging ensembles: Random Forest and Extra Trees, while the other five ensembles are boosting ensembles: Ada boost, Gradient Boosting, Hist Gradient Boosting, XGBoost and CatBoost. The study utilized 11 publicly available MDP NASA software defect datasets. Empirical results indicate the superiority of Tree-based bagging ensembles: Random Forest and Extra Trees ensembles over other Tree-based boosting ensembles. However, none of the investigated Tree-based ensembles was significantly lower than individual decision trees in prediction performance. Finally, Adaboost ensemble was the worst performing ensemble among all Tree-based ensembles.

Software defect prediction using tree-based ensembles

(1) Background: Low back disorders are a leading cause of missed work and physical disability in manual material handling due to repetitive lumbar loading and overexertion. Ergonomic assessments are often performed to understand and mitigate the risk of musculoskeletal overexertion injuries. Wearable sensor solutions for monitoring low back loading have the potential to improve the quality, quantity, and efficiency of ergonomic assessments and to expand opportunities for the personalized, continuous monitoring of overexertion injury risk. However, existing wearable solutions using a single inertial measurement unit (IMU) are limited in how accurately they can estimate back loading when objects of varying mass are handled, and alternative solutions in the scientific literature require so many distributed sensors that they are impractical for widespread workplace implementation. We therefore explored new ways to accurately monitor low back loading using a small number of wearable sensors. (2) Methods: We synchronously collected data from laboratory instrumentation and wearable sensors to analyze 10 individuals each performing about 400 different material handling tasks. We explored dozens of candidate solutions that used IMUs on various body locations and/or pressure insoles. (3) Results: We found that the two key sensors for accurately monitoring low back loading are a trunk IMU and pressure insoles. Using signals from these two sensors together with a Gradient Boosted Decision Tree algorithm has the potential to provide a practical (relatively few sensors), accurate (up to r2 = 0.89), and automated way (using wearables) to monitor time series lumbar moments across a broad range of material handling tasks. The trunk IMU could be replaced by thigh IMUs, or a pelvis IMU, without sacrificing much accuracy, but there was no practical substitute for the pressure insoles. The key to realizing accurate lumbar load estimates with this approach in the real world will be optimizing force estimates from pressure insoles. (4) Conclusions: Here, we present a promising wearable solution for the practical, automated, and accurate monitoring of low back loading during manual material handling.

https://www.mdpi.com/1424-8220/21/2/340/pdf

A Promising Wearable Solution for the Practical and Accurate Monitoring of Low Back Loading in Manual Material Handling.

Robot calibration of an industrial robot is of importance for those applications requiring high positioning accuracy. The use of numerical optimization techniques to identify the accurate kinematic parameters and tackle the residual errors for improving positioning accuracy is still challenging. The difficulties in using the conventional optimization techniques is associated with the expensive computation of the second derivative terms in the Hessian matrix and even more linked to the additional improvement of the residual errors induced by the configuration and payload effects for the practical applications. A hybrid calibration procedure, which is robust and efficient, is proposed to handle the said problems in this study. The first step is to utilize the Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm together with the quasi-Newton method to optimize the cost function with non-smooth nature established using massive measurement data. In the second step, two kinds of artificial neural network (ANN) algorithms are employed to further improve the nonlinear residual errors induced by payload and configuration effects. After the completion of the hybrid calibration, improvements of 88.1% and 80.1% have been attained, respectively, for the mean and maximum positioning errors of the robot end-effector, i.e., the mean/maximum positioning errors are reduced from 2.613mm/6.294mm to 0.310mm/1.255mm for 840 untrained measurement data. The experimental results also validated the effectiveness of the proposed method.

/pdf/positioning-accuracy-improvement-of-industrial-robots-23ls4qrjr0.pdf

Positioning Accuracy Improvement of Industrial Robots Considering Configuration and Payload Effects Via a Hybrid Calibration Approach

Software defect prediction refers to the automatic identification of defective parts of software through machine learning techniques. Ensemble learning has exhibited excellent prediction outcomes in comparison with individual classifiers. However, most of the previous work utilized ensemble models in the context of software defect prediction with the default hyperparameter values, which are considered suboptimal. In this paper, we investigate the applicability of a stacking ensemble built with fine-tuned tree-based ensembles for defect prediction. We used grid search to optimize the hyperparameters of seven tree-based ensembles: random forest, extra trees, AdaBoost, gradient boosting, histogram-based gradient boosting, XGBoost and CatBoost. Then, a stacking ensemble was built utilizing the fine-tuned tree-based ensembles. The ensembles were evaluated using 21 publicly available defect datasets. Empirical results showed large impacts of hyperparameter optimization on extra trees and random forest ensembles. Moreover, our results demonstrated the superiority of the stacking ensemble over all fine-tuned tree-based ensembles.

/pdf/software-defect-prediction-using-stacking-generalization-of-2f7fchvc.pdf

Software Defect Prediction Using Stacking Generalization of Optimized Tree-Based Ensembles

One of the most popular machine learning algorithms is gradient boosting over decision trees. This algorithm achieves high quality out of the box combined with comparably low training and inference time. However, modern machine learning applications require machine learning algorithms, that can achieve better quality in less inference time, which leads to an exploration of grading boosting algorithms over other forms of base learners. One of such advanced base learners is a piecewise linear tree, which has linear functions as predictions in leaves. This paper introduces an efficient histogram-based algorithm for building gradient boosting ensembles of such trees. The algorithm was compared with modern gradient boosting libraries on publicly available datasets and achieved better quality with a decrease in ensemble size and inference time. It was proven, that algorithm is independent of a linear transformation of individual features.

Histogram-Based Algorithm for Building Gradient Boosting Ensembles of Piecewise Linear Decision Trees

Gradient boosting ensembles of decision trees are a very popular machine learning model with several popular implementations. Some of those implementations utilize oblivious decision trees - decision trees of specific structures that allow for improved regularization and prediction speed. Over the last years, several research works have been published related to piecewise-linear decision trees and their utilization in gradient boosting ensembles. In this paper, oblivious piecewise-linear decision trees were introduced, and it was shown that it is possible to train such trees in gradient boosting ensembles efficiently. Furthermore, it was shown that such oblivious piecewise-linear decision trees have significantly faster prediction time than regular decision trees and piecewise-linear decision trees of similar depths. Ensembles of oblivious piecewise-linear decision trees achieve competitive quality on open datasets. 

Aleksei Guryanov

Papers

Histogram-Based Algorithm for Building Gradient Boosting Ensembles of Piecewise Linear Decision Trees

Oblivious Piecewise-Linear Decision Trees