Different Facets of Artificial Intelligence-Based Predictive Maintenance for Electric Powertrains

The multifaceted nature of predictive maintenance for electric powertrains entails comprehensive approaches to diagnose the internal physical phenomena of various components within electric vehicle (EV) powertrains. By integrating advanced sensing technologies, data analytics, and machine learning algorithms, predictive maintenance enables real-time monitoring and analysis of key performance indicators, facilitating early detection of faults or abnormalities. This proactive approach not only optimizes the performance, longevity, safety, and reliability of electric powertrains but also mitigates unplanned downtime, lowers maintenance costs, and enhances overall operational efficiency. Through the utilization of multidimensional predictive maintenance strategies, stakeholders in the electric vehicle industry can achieve significant advancements in powertrain maintenance and contribute to the development of sustainable transportation solutions.

This study aims to investigate the facets and benefits of implementing predictive maintenance strategies, specifically focusing on artificial intelligence-based approaches, in the context of electric powertrains. The European Union-funded ESCALATE and OPEVA approach aims to assure the electric vehicle user reaches its current destination to boost customer satisfaction, as it offers not only a safe journey but also promises no more breakdowns before the destination is reached. By examining the multifaceted aspects of predictive maintenance, including advanced sensing technologies, data analytics, and machine learning algorithms, this research aims to explore how these building blocks are brought together to enhance the performance, longevity, safety, and reliability of electric powertrains. Furthermore, the study seeks to evaluate the impact of advanced predictive maintenance on reducing unplanned downtime, optimizing maintenance costs, and improving overall operational efficiency. Through data-driven insights, this study aims to provide valuable insights and recommendations for stakeholders in the electric vehicle industry, contributing to the development of sustainable transportation solutions.

Concurrently, extensive research is underway to enhance BMS performance, particularly in terms of predicting battery ageing and ensuring safety. These techniques also aim to predict battery lifespan and detect faults. To address these challenges, James C. Chen et al. [1] propose empirical mode decomposition (EMD), grey relational analysis (GRA), and deep recurrent neural networks (RNN), whilst Y. Che et al. [2] propose optimized health indicators and online model correction with transfer learning for the RUL prediction of lithium-ion batteries.

The practical integration of AI into power electronic applications encompasses deterministic and stochastic environments, data collection and analytics methods, forecasting techniques, and cost-effective algorithms and hardware implementations [3]. In recent years, physics-informed PdM methods aligned seamlessly with the strategy of making AI a frontrunner in e-mobility applications. Among these, Li et al. [4] propose a model to calculate the cumulative fatigue damage of IGBT modules in EVs to evaluate the reliability. Rao et al. [5] propose a novel approach to PdM of power electronics by integrating optical and quantum-enhanced AI techniques.

The electric motor, a pivotal component within the electrical powertrain of electric vehicles powered by batteries, operates under demanding conditions characterized by significant temperature fluctuations, pronounced vibrations and voltage stress stemming from inverter power supply [6]. Wotawa et al. [7] employ model-based, simulation-based and machine learning-based real-time PdM approach for e-motors. Rjabtšikov et al. [8] and Dettinger et.al [9] use a Digital Twin-based approach when dealing with the predictive maintenance of e-motors.

Recent developments [10] employ AIoT-based preventive diagnostic by enabling recommendation and notification prediction for Run-to-Failure maintenance, planned preventive maintenance, and predictive maintenance for EV thermal management system. Yi et al. [11] utilize Long Short-Term Memory (LSTM) based Digital Twin to achieve real-time temperature forecasting and analyze degradation models for lithium-ion batteries. Dettinger et al. [9] introduce a machine learning-based model for fault detection in EV powertrains, employing a Digital Twin. In parallel, Bhatti et al. [12] advocate the application of Digital Twin technology in EVs through their systematic review.

Integrating technologies like IoFog, Arrowhead, and Mimosa into predictive maintenance (PdM) for Electric Vehicle (EV) powertrains can significantly enhance real-time data processing, communication, and overall management of maintenance activities by leveraging the capabilities of edge computing, interoperable automation, and standards for enterprise integration respectively. The PdM services are designed to be used as online services, in the form of PdM-as-a-Service, within a digital twin framework as depicted in Fig. 1. The framework has a data acquisition and management module which is built on a secure multi-agent system. This module is implemented as a heterogeneous data flow backbone based on an Arrowhead-ioFog-Mimosa architecture (See Fig. 2). This module is strengthened with cyber security solutions enabling very fast agent data encryption and person/node authentication. The multi-agent system enables effective data acquisition from EVs, grid infrastructure (i.e., charging infrastructure) and environmental data.

The acquired data can be classified as JSON files and stored in a cloud database, such as MariaDB, as proposed in the digital twin framework. There exist auxiliary backend services such as data and query management services, cloud services management, security and privacy management and Identity Management System (IDMS).

A two-way Kafka-based messaging system is designed to feed the Smart Services Modules which are utilized as Docker containers at the righthand side. The Docker containers receive the classified data in JSON format or other messages or requests from the secure multi-agent system and backend services. When the Digital Twin services including AI-powered PdM solutions produce an output, these outputs are also delivered to the visualization and user interfaces layer through a results gateway, or as feedback to the Secure Multi-Agent System (for instance, in case of an anomaly).

The framework is realized through a layered architecture (Fig. 2) that utilizes the combination of Arrowhead [13], Eclipse ioFog [14] and MIMOSA [15] and is composed of i) Edge layer (ioFog), ii) Communication & Automation Layer (Arrowhead), iii) Enterprise integration layer (MIMOSA), and iv) Execution layer.

The Communication & Automation Layer provides a framework for ensuring interoperable communication between various systems involved in a PdM solution space, such as between EVs, charging stations, and maintenance facilities. This layer is implemented by deploying an Arrowhead framework which is an open-source technology that establishes a standardized and secure interoperability platform for seamless communication and interaction between diverse services and devices. The Arrowhead-based solution facilitates automated workflows that enhance PdM, like scheduling maintenance activities, based on predictive insights while considering logistic and operational constraints.

Bild vergrößern

The four-layer architecture enables an effective data and action flow between layers. Data flow starts with raw data acquired from edge nodes like edge devices, gateways, smart grids, and/or onboard systems. Processed data and initial insights are transferred between the edge and communication layers. The standardized and structured data are propagated to the enterprise layer along with global insights and alerts. On the other hand, action flows can be instructions for actual maintenance activities that are streamed from the enterprise layer to the execution layer. Feedback, status updates and any corrective actions are exchanged between the execution and enterprise layer.

Bild vergrößern

In both edge and communication & automation layers, unsupervised learning (UL) algorithms can be deployed for reactive control and data preprocessing as well as maintenance scheduling and energy management. The proposed architecture at the edge layer implements UL algorithms to develop local control policies that react to real-time data, adjusting operating parameters to enhance performance or mitigate anomalies. The edge layer also enables optimal energy management and creating policies that consider both localized powertrain health and broader energy demands, like grid stability or energy prices. Physics-Informed Machine Learning (PIML) for more effective PdM can also be tacked both at the edge and global level. ioFog can be used to implement PIML for real-time analysis, using physics-based principles to enhance anomaly detection in powertrain components, ensuring that the alerts are not just data-driven but also physically plausible. Here, safety protocols leverage PIML to establish safe operating boundaries for the powertrain, using physics principles to prevent settings or actions that might risk component health or safety. At the automation layer, PIML models can be used for analysing long-term physical degradation patterns in powertrain components, ensuring that PdM strategies are aligned with anticipated wear and tear. A unified model can be developed to integrate physics-based principles with data-driven insights, presenting comprehensive models that ensure predictive insights and prescribed actions that are scientifically substantiated.

This paper presents a digital twin framework based on a 4-layered architecture that integrates Edge (ioFog), Communication & Automation (Arrowhead), Enterprise integration (MIMOSA), and Execution layers. IoFog, Arrowhead, and Mimosa are utilized for orchestrating physics-informed PdM and UL for EV powertrains so that organizations can create a robust, efficient, and scalable system. This integration not only enhances real-time data processing and communication but also ensures that the maintenance activities are streamlined, standardized, and optimized for the best possible performance and longevity of EV powertrains.