**1. Introduction**

Failure modes and e ffects analysis (FMEA) is a proactive analytical technique for identifying, tracking and mitigating product and process potential failures in a systematic way by determining its potential occurrence, root causes, consequences, and impact [1]. FMEA provides a quantitative score to evaluate failures where every failure is transformed into a numerical value that is called risk priority number (RPN). RPN is the result of multiplying three parameters namely severity, occurrence, and detectability. Severity is the risk or damage that may a ffect the machine, product, next operator or the end-user. On the other hand, occurrence is the likelihood of this failure that may occur again. Finally, detectability is the degree to which this failure could be detected [2–4]. Higher RPN value represents a higher priority of risk [5]. Appropriate corrective actions are usually determined based on RPN threshold value. If this threshold is reached, a risk mitigation procedure is applied accordingly [6]. Moreover, RPN value is used as a tool for optimal resource allocation by giving focus on risks that have the highest RPN or the most critical issues [3,7].

FMEA was firstly developed by NASA in 1963 to enhance the performance of the devices that are used in the aerospace industry [8]. Later, FMEA was adopted and promoted by Ford Motors in 1977 [3]. Currently, FMEA is being used in the automotive industry to ensure the quality and reliability of production systems [9]. Daimler Chrysler, Ford, and General Motors have developed an international standard called SAE J1739\_200006 as general guidance for implementing FMEA techniques to avoid

failures and enhance system reliability and safety [10]. FMEA documents are classified into two types namely design FMEA, and process FMEA [11]. Design FMEA is constructed during product design to define product weaknesses, critical components and their respective potential failure modes, root causes, and e ffects [1]. Meanwhile, process FMEA focuses on potential failures that may occur during the manufacturing process and incurred risks at each process step [3].

FMEA is a robust tool for quality improvement in both manufacturing and services industries. It can be used at the design stage of the product and during its implementation [9]. The aim of this is to avoid the end-user from experiencing unfavorable defects that may a ffect the reputation of the company negatively [3]. FMEA is also used as a process improvement technique to ensure consistency, reliability and avoid deviations. Moreover, it is also used to define and mitigate risks [12]. On the other hand, FMEA is used to improve maintenance managemen<sup>t</sup> by analyzing the maintenance requirements of the product and developing the maintenance plans that would be used to ensure that the system is doing what it is meant to do when it was created. Finally, FMEA is used to improve safety by conducting hazards analysis of components that have critical hazards on lives, property, or other losses that are identified and mitigated [7].

However, FMEA is criticized for many conceptual aspects. The most popular disadvantage of this method is the narrative and qualitative nature of its structure. For every product or process, FEMA documents are developed by engineers and experts using linguistic terms that are based on the personal evaluation. The RPN parameters' values are determined by engineers and experts which may include uncertainty and vagueness [12]. Moreover, the parameters that are used in FMEA are represented by (1–10) crisp scale which is an unreliable representation of real-application cases [5,13]. Additionally, Chang, et al. in [3] have criticized the RPN estimation by the inhomogeneous morphologic correlations between the three parameters. This criticism is based on the fact that each of these parameters is obtained and linearly multiplied by the other with an identical scale. This process is done despite the actual impact of every independent parameter and the di fferent qualitative interpretation of the scale. For example, high severity value should result an extremely high RPN value due to the critical hazard on the operator or the machine. In other words, once there is a risk on human, the other parameters shouldn't downgrade the overall value of RPN even if they are low.

Thus, in order to overcome this ambiguity, researchers proposed several approaches to improve the application of FMEA and the development of RPN. Several fuzzy techniques were examined to develop a new risk assessment approach to overcome the weaknesses of FMEA. Haktanır and Kahraman in [13] have summarized several fuzzy techniques and grey theory and proposed interval-valued neutrosophic (IVN) sets-based FMEA to eliminate the inaccuracy of human decisions and evaluations. Ayber and Erginel in [12] have proposed single-valued neutrosophic (SVN) Fuzzy FMEA as a new risk analysis tool to overcome the ambiguity of the linguistic terms. Al-Khafaji, et.al in [14] have proposed a fuzzy multicriteria decision-making model aligned with FMEA principles to obtain an efficient criterion for maintenance management. Liu et al. in [5] have used cloud model theory and hierarchical TOPSIS method to enhance FMEA e ffectiveness, overcoming bias probability of human judgment, and to facilitate the transformation of qualitative terms to quantitative values. Yang et al. in [2] have utilized a data mining-based method for isolating faults based on FMEA parameters in order to enhance predictive maintenance by using historical big-data to create data-driven models, by which future failure can be predicted e fficiently and accordingly avoid failures at a very critical operational item. Keskin and Özkan in [6] have applied a fuzzy adaptive resonance theory (ART) method for FMEA modeling in order to improve the classical methodology of calculating the RPN, which in total minimized cost and e fforts needed to respond to corrective actions alerts.

In the aforementioned research, the interpretation of FMEA documents was well addressed and resolved. However, the weakness of FMEA and RPN is not limited to the ambiguity of the FMEA textual description nor its quantitative representation, but it also extends to the importance of being proactive and responsive to failures. The flow of information once a failure is detected until the time it is ranked and resolved is important as well to guarantee minimum impact and limited implications. Another shortcoming of the conventional FMEA technique comes from the fact that its documents are prepared during the product or process design stages, which makes these documents obsolete after production starts ahead. Therefore, these documents need to be dynamically validated and updated on a continuous basis. Hence, utilizing new technologies is very vital to overcome these weaknesses and keep these documents updated and responsive [2].

In the era of Industry 4.0, connectivity o ffered instant communication and collaboration among the value chain. Artificial intelligence (AI), the internet of things (IoT), big-data, and cyber-physical systems (CPS) made a grea<sup>t</sup> leap in automation and optimization at all levels of manufacturing. Here, automation is not limited to machines and processes, but also to managemen<sup>t</sup> information systems such as enterprise resources planning (ERP), customer relationship managemen<sup>t</sup> (CRM) and quality managemen<sup>t</sup> systems (QMS) [15]. Additionally, the real-time flow of data among the value chain, which is instantly analyzed and transformed to user-friendly information, thanks here to the advanced supercomputing and analyzing power [16], resulted new paradigms of manufacturing systems which are being called nowadays by smart factory, smart machine, smart product and augmented operator [17]. These pillars changed the production systems from being reactive to be proactive and levered the human intervention from doing the work to supervise it while it is being done. Sensors, 3D cameras, radio frequency identifier (RFID), and Wi-Fi made monitoring processes more precise and accurate. Unseen defects or deviation of products or processes can be detected as soon as it is occurring. Defect elimination and processes re-adjustment are made autonomously at the micro and macro levels [15,18,19]. All these technologies, alongside the increased complexity of products and their manufacturing systems, generated a large volume of data, at a high velocity, veracity, and variety. The analysis of such big data requires advanced resources and techniques to classify data and detect patterns that cannot be detected using traditional analytical tools.

Automated machine learning (AutoML) are tools that automate the process of a machine learning workflow, o ffering the same capabilities of regular machine learning, without explicit knowledge of programming [20]. AutoML aims at reducing the human intervention in data preprocessing, feature selection and algorithm selection so as to make machine learning automated [21]. Google AutoML is a cloud machine learning platform that automates supervised machine learning in a very efficient way. It handles the tasks of data preprocessing, feature extraction, feature engineering, feature selection, algorithm selection, and hyperparameter optimization [22]. Google AutoML automatically develops models based on neural architecture search (NAS). It follows the try and error strategy by developing the model based on a random set of hyperparameters, then evaluates the performance of the model which is resulted by using this set of hyperparameters and finally concludes the most accurate model [22,23].

AutoML is increasingly used in scientific research areas. Faes et al. in [24] have evaluated the performance of AutoML hosted by google cloud platform against other machine learning methods and algorithms. It is claimed that AutoML has higher accuracy in medical image classification and can be used by people who are less experienced in coding and algorithms. Similarly, Hayashi et.al. in [25] utilized Google AutoML to identify pest aphid species and improving crop protection e ffectiveness. The authors concluded that such a tool provided an accuracy of 0.96 which allowed them to consider the AutoML as a useful and e ffective tool. Additionally, Li et al. in [26] have used AutoML to automate customer service activities by analyzing di fferent customers' information and respond to their inquiries based on historical frequent inquiries. According to the authors, the solution provided improved responsiveness and minimized the cost of customer service management. Moreover, Galitsky et.al in [27] proposed a novel approach to automate customer complaints processing and classification by training a machine learning algorithm for analyzing dialogues recorded between customers and company-agents.

Based on that, the aim of this paper is to examine a novel optimization approach applied to FMEA and RPN by classifying failures according to updated FMEA documents and generating the RPN automatically without human intervention. A successful FMEA is gained through optimized consistency, responsiveness, and accumulated experience. The suggested approach aims at solving the above-mentioned FMEA weaknesses by two steps: first, reviewing and re-evaluating a dataset containing reported failures manually by experts to ensure accuracy. Secondly, conducting supervised machine learning techniques on the updated data and develop machine learning models that can be deployed to evaluate and classify newly reported failures automatically with minimum processing time and enhanced consistency.
