A Bayesian based Intelligent Troubleshooting System

Available online at www.sciencedirect.com[pic 1][pic 2][pic 3]

ScienceDirect

Procedia Computer Science 200 (2022) 602–610

3rd International Conference on Industry 4.0 and Smart Manufacturing

A Bayesian based Intelligent Troubleshooting System

I Yung∗, Federico D’ambrosio, Alissa Zaccaria, Fabio Floreani

Consorzio Intellimech, Via Stezzano 87, Bergamo 24126, Italy

[pic 4]

Abstract

Troubleshooting systems can bring different benefits in assets management, particularly for service operations, facilitating the di- agnosis of problems and faulty components identification. However, these systems are commonly based on rigid computation logic unable to handle uncertainties. In this work, a knowledge-based system exploiting the Bayesian theorem was developed and applied in a troubleshooting tool that relies on human-machine interaction. The required knowledge and the algorithm were analyzed and tested to ensure robustness and self-learning capabilities. Subsequently, the system was implemented in an industrial environment, specifically from a crane manufacturing company. The algorithm is robust to errors and provides the possibility of not answering some questions. However, the system performance is highly dependent on the questions, both in terms of quantity (adequate num- ber compared to possible failures) and quality (effective to discriminate among failures). Indeed, this work shows how the system knowledge enhancement by introducing additional questions can significantly improve the troubleshooting performance. Future developments may involve user-friendliness enhancement and self-learning implementation to add and update questions over time.

© 2022 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)

Peer-review under responsibility of the scientific committee of the 3rd International Conference on Industry 4.0 and Smart Manufacturing

Keywords: Knowledge based system; Bayesian theorem; troubleshooting system; human support; self-learning

[pic 5]

1. Introduction

The potentialities of expert systems in assisting the human workforce have been investigated over the years in different fields. The first computer-based consultation system, Mycin, was proposed in the medical field. Mycin aimed to support physicians in bacterial infections diagnosis, allowing the selection of the most suitable therapy [1]. In the manufacturing field, computer-aided troubleshooting systems have been deemed suitable to overcome the scarcity of experienced troubleshooters [2].

Highly skilled troubleshooters are commonly trained on the job over a long period. Indeed, it is very challeng- ing for companies to replace such a skilled workforce. Moreover, the evolution of machines complexity introduces additional challenges to the troubleshooting process. In this regard, several computer-based troubleshooting systems

[pic 6]

∗ Corresponding author. Tel.: +39-035-069-0366.

E-mail address: i.yung@intellimech.it

1877-0509 © 2022 The Authors. Published by Elsevier B.V.

This is an open access article under the CC BY-NC-ND license (https://creativecommons.org/licenses/by-nc-nd/4.0)

Peer-review under responsibility of the scientific committee of the 3rd International Conference on Industry 4.0 and Smart Manufacturing 10.1016/j.procs.2022.01.258

were developed based on the knowledge collected by experts specialized in the identification of machine failures (e.g., [3]). However, these systems are commonly based on rigid computation logic and unable to handle the uncertainties (different from what human is typically capable of).

Machine learning algorithms, which represent the basis of modern approaches to artificial intelligence, can be exploited to solve this problem. Indeed, these algorithms, based on statistical rules, can compensate for typical uncer- tainties and human errors. The objective of the present work is to investigate and exploit the possible algorithms behind a well-known artificial intelligence game, Akinator, for a robust troubleshooting system implementation, tolerant to mistakes and uncertain answers.

The game Akinator is able to guess a character, either a real or an imaginary person, after asking the player several closed-ended questions even in the presence of some wrong answers. Besides its robustness, other significant features of Akinator include the following:

• Available answers are limited: yes, no, don’t know, probably, probably not. The last three possible answers might allow the system to handle uncertainties;
• The questions order is not randomly selected. The game mostly starts with general questions and ends with specific ones;
• There are several repetitive questions. Namely, two different questions obtaining the same information can exist (e.g., a question asking the character’s gender is female and another question asking if the character’s gender is male).

If the system is not able to identify the character, a list of possible characters is presented, allowing the user to select the one they are referring to. If the character is not present in the list, the user is asked to introduce a new one with a brief description. The character in Akinator refers to a person which the system should identify among all the possible candidates. The candidates can basically be any form of entity that the system is expected to identify. that can also be in the form of objects, numbers or failures in the case of troubleshooting.

Akinator presents advantageous features that could be exploited for troubleshooting purposes, such as simple an- swers, question order, learning capability and robustness to wrong answers. In this paper, the characteristics of several similar systems are first illustrated (section 2). Then, the most promising algorithm is selected and applied to a test dataset (section 3) before its validation on an industrial dataset (section 4). Finally, the simulation results are presented (section 5), followed by the conclusions and future developments (section 6).

1. Applicable expert systems

The troubleshooting systems under consideration involve two subsystems. In particular, one subsystem is aimed at updating the probability of the characters after each answer, while the second one manages the selection of the order of the questions. In the following, several algorithms which accomplish similar functionalities are presented.

The expert system in the medical field, Mycin, was developed to assist physicians in the diagnosis of bacterial infections and, hence, in the identification of the most suitable therapy for the patients. This system asks the doctor a long series of closed-ended and simple textual questions, and, based on the answers, it ranks the possible diseases from the most to least probable ones. From the user perspective, the functionality is close to the expected system. However, decision rules are the backbone of this system. As such, it cannot handle either wrong answers or automatic learning capability.

Renyi-Ulam is a mathematical game where a player tries to guess an unnamed object or a number thought by the other player by asking yes/no questions and considering that one or more of the given answers may be wrong. A rigorous analysis of the strategies proposed in the literature to find the correct answer in the most efficient way possible, meaning with the least number of questions, was outlined in [4]. Possible questions are straightforward for specific environments, such as guessing an integer, in which the selection of the questions is predictable. In several other cases, the questions are commonly related to the characteristics of objects, such as people’s attributes in Akinator and symptoms of faults in troubleshooting.

A similar game called 20Q is a long-time case study in the information technology community [5]. One proposal to solve this game was based on a simple neural network [6]. More specifically, the system was based on simple

operations, namely, additions and subtractions. The algorithm identifies the character or object that has the highest value over a course of answered questions. Note that the selection of each question is similar to a binary search method, meaning that the next question is chosen to approximately subdivide the group of characters/objects with the highest value into two subgroups. The algorithm is robust and features an ongoing self-learning mechanism based on the result of the character/object identification.

...