Imputation of Missing Values Affecting the Software Performance of Component-based Robots

Highlights

• Different regression techniques together with neuronal models have been applied to impute missing values in two data sets corresponding to anomalies detected in the operation of a robot.

• Best regression results are obtained in most of the cases analysed by the Radial Basis Function, with similar results to those offered by non-linear regression techniques and the Multilayer Perceptron.

• It is also interesting to highlight and analyze how not always the same technique obtains the best results, and it is necessary to apply more than one technique and compare them to reach the most optimal solution.

Abstract

Intelligent robots are foreseen as a technology that would be soon present in most public and private environments. In order to increase the trust of humans, robotic systems must be reliable while both response and down times are minimized. In keeping with this idea, present paper proposes the application of machine learning (regression models more precisely) to preprocess data in order to improve the detection of failures. Such failures deeply affect the performance of the software components embedded in human-interacting robots. To address one of the most common problems of real-life datasets (missing values), some traditional (such as linear regression) as well as innovative (decision tree and neural network) models are applied. The aim is to impute missing values with minimum error in order to improve the quality of data and consequently maximize the failure-detection rate. Experiments are run on a public and up-to-date dataset and the obtained results support the viability of the proposed models.

Introduction

Wide attention has been devoted to the development of intelligent robots in recent years. Although significant contributions have been done, it still is a challenging field where further progress is required to satisfy present and future demands. One of such demands is the fluent interaction with non-expert humans, that is required for robotic systems to be widely integrated in a variety of homes and workplaces [1]. In order to get such fluency, performance of both hardware and software is a keystone. However, the ever-increasing complexity of robots leads to a parallel increase in chances of experiencing a failure. Accurate and prompt detection of such events is required in order to improve performance and hence fluency. Full attention has been payed to advance in many subfields of the robotics arena but according to some authors [2], further effort must be devoted to anomaly detection in such systems. It is even more challenging when failures happen in a real-world context where complex phenomenon may interfere.

Accordingly, present paper focuses on the preprocessing of robot-performance data, whose importance is widely acknowledged. More precisely, the aim is the successful imputation of Missing Values (MV) in order to get as much data as possible for subsequent anomaly/failure detection. Thus, a wide variety of Artificial Intelligence (AI) models are applied in order to predict the MV of all the dataset components.

The successful detection of anomalies/faults is a challenging task that does not only apply to robots [3], [4], [5]. From a business perspective [6], AI in general, and Machine Learning (ML) in particular, can greatly contribute to anomaly detection and some other interesting tasks, maximizing companies benefits.

Since pioneer works [7] in the application of ML to robotics, unsupervised [8], supervised, and reinforcement [9] learning models have been previously applied. A variety of problems have been addressed so far such as control [10], [11] and communications [12] among others. In the case of anomaly detection, most ML previous work has been focused on the detection of hardware anomalies [13], while software anomalies have been scarcely investigated. Software failures often occur in robotic systems and their automatic detection requires training data. The problem comes from the difficulty of obtaining the data either because of the lack of execution traces or because the existing registers do not refer to the exact moment in which anomalies are produced. That is why it is difficult to find a dataset generated in a controlled environment where all the information is available. Furthermore, when data are gathered in a real-life environment, quite likely there will be some or many MV, that can not be processed by ML models.

One of the few works on the detection of software anomalies within the framework of component-based robots is [14]. In that paper, authors proposed the only publicly-available dataset (further details in Section 3) that gathers data from different performance indicators of a robot. The dataset [15] has been used in present paper as a benchmark dataset due to its interest and novelty. In this dataset, there are many MV associated to different data so a robust strategy must be followed in order to deal with them as most ML models can not process such data. One of the obvious preprocessing alternatives in order to solve such problem in the data is removing MV, either by deleting data instances or by deleting attributes. However, a more advanced proposal is to impute such values, keeping some information that could be useful for the subsequent anomaly detection. This approach is adopted in the present paper.

AI methods have been previously applied for imputation of MV [16]. However, scant attention has been devoted to the application of ML methods in order to solve such problem in robot datasets. One of the very few previous proposals is [17], where a probabilistic approach for classification using incomplete data was applied. The author performed a classification (for failure detection) of data samples by calculating the a priori probability of MV, determined from the data samples that are not missing. However, the author proposal was only applied to outdated (1999) datasets containing hardware anomalies. Differentiating from previous work, the present paper is the first approach to impute MV in a dataset containing information about the performance of the software components of a robot. A comprehensive benchmark comprising a wide variety of methods has been performed and some of the methods are applied to this problem for the first time.

The methods applied for imputation of MV are introduced in Section 2, while the analysed case study is described in Section 3. The performed experiments, together with their associated results are compiled in Section 4. Finally, the main conclusions and some proposals for further work are presented in Section 5.