Prescriptive Maintenance and Intelligent Operation

Lines of Research
Prescriptive Maintenance and Intelligent Operation

This research line aims to explore methodologies and methods of Artificial Intelligence focused on prescriptive decision-making, based on diagnostic data input and prediction of future states of systems with multiple components. It utilizes a multi-layered approach, distinguishing between the component and system level. In the medium term, the goal is to create a collaborative internet environment for prescriptive maintenance studies by companies and researchers.

Within this research line, the challenges of identifying and eliminating spurious measures with outliers due to potential measurement equipment issues are also included. In addition to integrating data from endogenous and exogenous data sources to the industrial production process, prescriptive maintenance of any physical system can be translated as the real-time automation of decision-making regarding the type and timing of its maintenance, as well as the optimized planning, scheduling, control, and operation of multiple units using any combination of optimization, heuristics, machine learning, and cyber-physical systems.

For prescriptive maintenance to have a more comprehensive application within the context of Industry 4.0, some barriers need to be overcome. At the most basic level, it involves the phenomenological understanding of each stage of an industrial production line and the implementation of appropriate sensors for real-time data acquisition of the process. It is important to note that there is no publicly available and scientifically validated compendium for all modern equipment in production lines. However, reliability analysis techniques such as Failure Mode and Symptoms Analysis (FMSA) can be used to define sensors to monitor specific variables associated with equipment degradation commonly used in production processes. Additionally, there is a challenge in implementing artificial intelligence algorithms for autonomous operational decision-making by these cyber-physical agents. A vital aspect is the design decision to allocate these AI algorithms to the most suitable processing stratum (fog, cloud, and edge computing), considering the required level of security for the industrial process it is embedded in. These AI agents need to act at different levels and negotiate an ideal outcome at the system level, which may include economic considerations where maintenance impact can be translated into costs.

Additionally, there are the following challenges:

• Online data visualization: involves the study and development of new ways of presenting data, promoting better interpretability and understanding of the displayed information flows.
• Development of machine learning algorithms for optimizing the operating points of equipment or production lines with the aim of optimizing the production system.
• Development of an Artificial Intelligence platform that provides an environment for data analysis with appropriate machine learning protocols, pre-customized for each type and model of equipment in their industrial park, using data provided by the IIoT platform. The objective is to create an environment that can be coupled with the IIoT platform, simplifying the user's task of conducting both exploratory studies for prescriptive maintenance and finding the optimal operating point of their production line. Additionally, when the study or improvement of the algorithm is mature enough for production, the user can update the cloud-based AI agent.
• Online learning: involves continuous and incremental updating of predictive and decision-making machine learning models to maintain accuracy over time, in contrast to changes in data patterns and industrial processes (including changes in regulations and standards and expert and managerial knowledge) that cause degradation of learned models, for example, phenomena that are not necessarily indicative of a predicted failure but rather a change in the machine's operating regime, lack of calibration, etc. (MITCHELL et al., 2018).
• Transfer learning: involves creating a predictive prognostic model developed for a specific industrial sector that can be partially reused as a starting point for predicting failures in another sector/company, even if there are differences in the context in which these sectors operate. Creating a portfolio of techniques can be an effective solution to the scarcity of labeled prognostic data in industrial facilities.

The short to medium-term goal of this research line is to identify, analyze, and select among the available IIoT platforms in the market the one that best fits the focus of this line of study, namely, Prescriptive Maintenance with the application of sophisticated artificial intelligence agents. Such a platform should allow the use of data from any equipment and manufacturer with commonly used communication protocols. In this platform, it should be possible to load data from company equipment, conduct studies based on pre-trained AI agents, and provide customizable parameters for various types of equipment.

In the short term, studies will be conducted based on the set of real-time measurement and monitoring systems, using the techniques and methodologies developed and explored in the research line of the same name. Prescriptive maintenance will employ an approach based on cyber-physical agents that will be developed according to the commonly used data analysis methodology, Cross Industry Standard Process for Data Mining (CRISP). This approach encompasses understanding the business and data to be analyzed, as well as the phases of data preparation and modeling, culminating in the deployment of the produced models. Once the models are deployed in an online monitoring platform, their output can be a recommendation, a warning, a critical alarm, or an optimized planning and scheduling of maintenance operations.

Machine learning algorithms and data fusion strategies are essential for recognizing patterns linked to the data itself or for detecting and inferring possible faults or abnormal conditions (ROUSSEAUX, 2017).

Classification aims to identify the most likely causes of confirmed abnormal behaviors and failure episodes to predict and prevent them in the future (BANKS et al., 2011). When labels are available, supervised learning and reinforcement learning can be applied to train a classifier and act on the monitored asset. In (LI et al., 2016) and (RUIZ et al., 2018), fault detection and classification in wind turbines are performed by combining acoustic and vibrational signals and texture features of the signals in the time domain, respectively. Similarly, Yin and Zhao (2016) adopt a Deep Learning approach for diagnosing faults in high-speed train equipment.

Research around data fusion and analysis for industrial prognostics is highly necessary. The use of data fusion techniques and machine learning algorithms to explore all available information allows for the learning of complex behaviors and prognostic models from historical data; terabytes of data can be analyzed in real-time, and production processes can be intelligently monitored online.

Among the expected impacts of this research line are:

• Creation of an asset performance management system with data collection and interpretation protocols focused on prescriptive maintenance, which will be used for decision recommendation. Part of this content will be public, another part will be exclusively accessible to participating companies in pre-competitive projects, and other parts will be premium, exclusively accessible to the company funding a competitive project. This information database will focus on the types and specifications of sensors required for effective data processing protocols to create the mentioned artificial intelligence agents.
• Development of a library of cyber-physical artificial intelligence agents for the industrial environment in the context of edge processing, with generic hardware and sufficient capability to execute AI algorithms.