Strojno učenje za zaznavanje anomalij in napovedno vzdrževanje robotov

Kerpan, Maruša (author), Mihelj, Matjaž (mentor), Šlajpah, Sebastjan (comentor)

Napovedno vzdrževanje predstavlja pomembno področje v industriji 4.0, saj omogoča pravočasno odkrivanje potencialnih okvar in zmanjšuje stroške vzdrževanja. Namen magistrske naloge je raziskati uporabo metod strojnega učenja za zaznavanje anomalij in napovedovanje preostale življenjske dobe robotskih sistemov. V ta namen smo razvili dinamični model robota, ki omogoča simulacijo gibanja in ustvarjanje sintetičnih podatkov s spreminjanjem parametrov, kot je trenje. Ti podatki so bili uporabljeni za testiranje in primerjavo različnih algoritmov strojnega učenja za zaznavanje odstopanj od normalnega delovanja. V raziskavi smo preizkusili tri različne metode strojnega učenja: izolacijski gozd, samodejni kodirnik in enorazredno metodo podpornih vektorjev. Modeli so bili ovrednoteni glede na njihovo sposobnost zaznavanja anomalij in razlikovanja med normalnimi in nenormalnimi podatki. Na podlagi teh primerjav smo ugotovili, da je metoda enorazrednih podpornih vektorjev dosegla najboljše rezultate, saj je zagotavljala visoko stopnjo natančnosti in zanesljivosti pri zaznavanju odstopanj. Poleg tega smo rezultate te metode uporabili za izdelavo linearnega regresijskega modela, ki napoveduje preostalo življenjsko dobo robotskih sistemov. S pomočjo kazalnika stanja, ki upošteva povprečno razliko med normalnimi in odstopajočimi podatki ter delež vrednosti, ki presegajo določen prag, smo lahko napovedali, kdaj bo robot dosegel kritično stanje, kar omogoča pravočasno intervencijo. V simulacijskem okolju smo pokazali, da lahko napovedno vzdrževanje, podprto z metodami strojnega učenja, prispeva k zanesljivosti in podaljšanju življenjske dobe robotskih sistemov. Metode, ki temeljijo na simulacijskih podatkih, so prilagodljive za uporabo v realnih okoljih brez potrebe po obsežnih zgodovinskih podatkih. Na realnem modelu bi se trening začel na podatkih novega robota, vključili pa bi lahko dodatne parametre, kot so vibracije ali temperaturne spremembe. Naš pristop omogoča prilagodljivost, zmanjšuje tveganje nenadnih okvar in optimizira vzdrževalne procese, kar je ključno za industrijo 4.0.

• Language: Slovenian
• Keywords: napovedno vzdrževanje, strojno učenje, zaznavanje anomalij, robot, simulacija, dinamični model, preostala življenjska doba
• Work type: Master's thesis/paper
• Organization: FE - Faculty of Electrical Engineering
• Year: 2024
• PID: 20.500.12556/RUL-164501
• Publication date in RUL: 28.10.2024

Secondary language

• Language: English
• Title: Machine learning for anomaly detection and predictive maintenance of robots

Abstract

Predictive maintenance is a key area in Industry 4.0, enabling timely detection of potential failures and reducing maintenance costs. The purpose of this master's thesis is to explore the use of machine learning methods for anomaly detection and prediction of the remaining useful life of robotic systems. To this end, we developed a dynamic model of a robot, which allows for the simulation of movements and the generation of synthetic data by varying parameters such as friction. These data were used to test and compare different machine learning algorithms for detecting deviations from normal operation. In the research, we tested three different machine learning methods: Isolation Forest, Autoencoder, and One-Class Support Vector Machine. The models were evaluated based on their ability to detect anomalies and distinguish between normal and abnormal data. Based on these comparisons, we found that the One-Class Support Vector Machine method produced the best results, providing a high level of accuracy and reliability in detecting deviations. Additionally, we used the results of this method to develop a linear regression model that predicts the remaining useful life of robotic systems. Using a health indicator that considers the average difference between normal and deviating data, as well as the proportion of values exceeding a certain threshold, we were able to predict when the robot would reach a critical state, allowing for timely intervention. In the simulation environment, we demonstrated that predictive maintenance supported by machine learning methods contributes to the reliability and extended lifespan of robotic systems. Methods based on simulated data are flexible for use in real-world environments without the need for extensive historical data. In real-world applications, the training would begin with data from a new robot, and additional parameters such as vibrations or temperature changes could be included. Our approach offers adaptability, reduces the risk of sudden failures, and optimizes maintenance processes, which is crucial for Industry 4.0.

• Keywords: predictive maintenance, machine learning, anomaly detection, robot, simulation, dynamic model, remaining useful life