A survey on AI Approaches for Internet of Things Devices Failure Prediction

. The use of Internet of Things (IoT) devices has experienced a substantial surge in various sectors, including manufacturing, healthcare, agriculture, and transportation. Nonetheless, the susceptibility of these devices to failures has emerged as a significant concern, contributing to costly periods of inactivity and diminished productivity. Consequently, the development of sophisticated and precise techniques for forecasting device failures in advance has become imperative. This research paper thoroughly investigates and analyses the most recent advancements and scholarly inquiries pertaining to the implementation of artificial intelligence methodologies, notably machine learning and deep learning, in the realm of predicting and averting IoT device failures. These AI-based approaches can be trained on extensive historical datasets, enabling the detection of distinctive patterns and anomalies that serve as potential precursors to device malfunctions. By incorporating these innovative failure prediction techniques into their operations, organizations can actively identify and address potential issues, thereby minimizing the adverse repercussions of device failures on their overall performance and functionality.


Introduction
Over the last few years, the industrial sector has played a significant role in driving economic growth and innovation in Europe, with 75% of all exports and 80% of all innovations coming from this sector [1].In 2011, the concept of "Industry 4.0" was introduced at the Hanover Trade Fair as part of Germany's high-tech strategy.The four core components of Industry 4.0 are cyber-physical systems (CPS), the Internet of Things (IoT), the Internet of Services (IoS), and smart factories [1].The advent of Industry 4.0 marks the fourth industrial revolution, which has progressed quickly due to advances in technology, specifically the Internet of Things.Over recent decades and after the appearance of the IoT concept in 1999 by Kevin Ashton, IoT has lately become the main topic of research in the wider research community these years.It aims to make physical devices "Smart" by connecting them to the internet to enable remote control and monitoring [2].Three key features that identify a device as "Smart" are autonomy, context-awareness, and connectivity [3].However, no IoT device is perfect, it can be affected by various internal and external factors.Furthermore, Maintenance nowadays was founded to meet the needs of manufacturers' information about when services repair, or replace unhealthy equipment components, and autonomous services prepared the right conditions for efficient production and improved equipment availability.Proactive maintenance refers to a maintenance strategy in which equipment and systems are regularly inspected and maintained before they fail.This approach aims to prevent equipment breakdowns and prolong the Remaining Useful Life (RUL) of the equipment.Examples of proactive maintenance include regular equipment inspections, scheduled maintenance, and the implementation of predictive maintenance (PdM) techniques such as vibration analysis or oil analysis.Mathematical and statistical modeling based on Artificial Intelligence (AI), Machine Learning (ML), and Deep Learning (DL), can be used to enhance a proactive maintenance approach by providing more accurate and timely predictions of equipment failures.By using data from sensors and other sources, AI models can analyze equipment breakdowns and reduce downtime (see Fig. 1).This paper is organized as follows.Section 1 of the article is the introductory phase which sheds light on the transformative impact of industry 4.0 and IoT on the industrial sector, focusing on proactive maintenance empowered by AI to prevent equipment failures and optimize production.Section 2 discusses the details of the crucial concept of failure prediction in industry 4.0.Section 3 focuses on studying similar state-of-art work.Section 4 highlights the methodology adopted to investigate failure prediction and proactive maintenance strategies in the context of Industry 4.0.The results are discussed in Section 5. Finally, the paper's conclusion.

Internet of Things and Failure Prediction in Proactive Maintenance
The "Internet of Things" is integral to Industry 4.0, connecting physical objects with electronic devices, sensors, and connectivity for data exchange across industries.It enhances manufacturing by real-time monitoring and predictive maintenance, thereby improving efficiency, reducing costs, and increasing safety in various industries [2,4].IoT's application spans sectors such as smart farming [5,6], industry [3], smart healthcare [7], transportation [8], energy, and utilities [9].[10], and Predictive Maintenance collaborate toward this goal.It enhances safety, efficiency, reliability, data-driven decision-making, and reduces downtime and costs [11].RUL prediction estimates time until failure, and PdM schedules based on this, promoting better equipment performance [12,13].

AI-Powered Failure Prediction in Industry 4.0
In industry 4.0, AI is pivotal for modernizing production systems.Integrating of IoT, big data, and AI enables real-time monitoring and data-based decisions.Machine learning algorithms are commonly employed for failure prediction, using historical data to identify component health trends.AI models forecast system behavior, chosen based on factors like data quality, computational resources, and memory.AI models encompass techniques such as artificial neural networks (ANNs), neuro-fuzzy systems, Bayesian approaches with hidden Markov models, and logistic regression algorithms [9][14].
The efficacy of ML-based failure prediction models has been demonstrated for various systems such as Smart grids [15], industrial machines [16], agriculture [17], and Photovoltaic parks [18].These models are divided into four subprocesses, which will be discussed in detail in the methodology review section.

Related work
The evaluation of significant literature contributions in this section highlights various approaches to predictive maintenance and anomaly detection in the context of Industry 4.0.
J. Dalzochio et al. [14] offer a comprehensive overview of machine learning and reasoning techniques in predictive maintenance within Industry 4.0.Their focus is on both the current state of the field and the challenges it faces.The paper emphasizes architectures and frameworks proposed to tackle predictive maintenance using methods like neural networks, machine learning, deep learning, and ontologies.The article acknowledges the need to integrate these techniques with existing systems and to develop more robust and accurate models.
C. Ferreiraa et al. [10] delve into the prediction of RUL using machine learning methods.They highlight challenges like data volume, equipment complexity, and prediction accuracy.Various machine learning techniques, such as supervised, unsupervised, and DL, are reviewed for RUL prediction.
A Systematic Literature Review by M. Fahim [19] systematic literature review underscores the dominance of AI, particularly machine learning and deep learning, in anomaly detection and prediction within IoT environments.The review identifies gaps in research, including handling high-dimensional data, real-time anomaly detection, scalability, dealing with missing data, combining techniques, and applying them to specific domains.
The article [20] focuses on challenges in establishing comprehensive data-driven systems for PdM.It addresses issues like noisy data, model generalizability, and data collection in industrial settings.The review examines anomaly detection, prognostics, and architectural perspectives, advocating for improved anomaly detection, holistic predictive models, and the utilization of both cloud and edge resources to meet real-time requirements in Industry 4.0.The literature review showcases multiple studies that examine the use of machine learning and reasoning methods for PdM in Industry 4.0, as well as detecting and predicting anomalies.

Methodology review
This section explores the outcome of the search process, the selection process of Predictive Failure and the qualitative examination of the chosen paper.

IoT and AI interest exploration in Scopus Database
Our search and retrieval of around 10,040 document results when we used the criteria (ALL (internet AND of AND things) AND ALL (failure AND prediction)), which demonstrates the high level of interest among researchers in the topic of IoT.Afterwards, we focused our investigation using the criterion (TITLE-ABS-KEY (internet AND of AND things) AND TITLE-ABS-KEY (failure AND prediction) AND TITLE-ABS-KEY (machine AND learning) OR TITLE-ABS-KEY (artificial AND intelligence)) and got around 231 documents results.228 documents were published between 2016 and 2023.Our focus in the summary of reviewed literature is presented in Table 1 (see Fig. 2).First, the prediction problem is defined as determining when a system or component will fail.Next, data collection involves acquiring a large set of sensor data that represents both healthy and faulty operation, taking into account factors such as time period, attributes, sampling rate, and records of failures.Data preprocessing involves transforming and selecting features from the dataset using statistical techniques like correlation and Principal Component Analysis (PCA).The next step, model training, involves searching for an algorithm such as Support Vector Machines (SVM), Random Forest (RF), and optimizing the algorithm's hyperparameters using techniques like grid search, random search, gradient-based optimization, and Bayesian optimization.Finally, the last strategic step is to evaluate the model's performance by using factors such as cross-validation, validation set, a reporting form, and performance metrics such as precision and recall.By following these steps, a prediction model that utilizes ML technology for failure prediction, anomaly detection, and Remaining Useful Life in industrial maintenance can be achieved [10].

Summary of the reviewed literature
The literature review presents a comprehensive analysis of numerous studies that investigate the application of machine learning and reasoning techniques in the context of Predictive Maintenance within the Industry 4.0 framework.These studies also focus on the detection and prediction of anomalies.To provide a succinct overview of the reviewed literature, the Table 1 offers a summary of the key findings:

Results
As previously stated, Failure Prediction in the Internet of Things field can be accomplished through a combination of sensor data, machine learning, and artificial intelligence algorithms.Table 1 presents a comparison study using the following criteria: • Models: Machine learning models used in mentioned paper enhance the prediction of IoT device failures.
• Application: The main topic of each section study in the field of Failure Prediction.
• Context/ dataset: The dataset or of the dataset used to train, test, and evaluate the proposed model.
• Features: Various capabilities and functions that the device or system is able to perform in the selected articles.
• Evaluation methods/ statistical measure: Multiple methods and measure are used to assess and optimize the performance of the model, compare different models, and prevent overfitting.
• Best performance results: the best accuracy for the proposed model.
It is noted that majority of models focus on classification and regression.The literature review reveals that many studies opt for using multiple algorithms instead of one, with the aim of enhancing the robustness, generalizability, and accuracy of the results.Furthermore, a significant number of papers employ "hybrid" or "ensemble" learning methods (see Fig. 4).These methods can be broadly classified into two categories: • Homogenous ensemble methods: which use the same type of base models such as LSTM-RNN [23].
In addition, The hybrid method in AI proves to be more accurate than other propose models [8,23,27,29].This approach leverages the strengths of multiple algorithms, resulting in better performance, scalability and accuracy.
Regarding the use of these studies, the key findings pertain to failure prediction, prediction of RUL, predictive maintenance, diagnostic, and monitoring with a common goal: predicting failures before they occur.
Nevertheless, the context and dataset are unique, presenting a major challenge for these studies as each article employs its own system-specific dataset, such as an autoclave sterilizer [28], an MEP component [26], light bulbs [23], Smart Grid [15] .The dataset comprises samples of varying sizes collected at different time intervals.
Features are essential in AI model training and prediction process.The right choice and representation of these features has significant impact on the model's performance, accuracy, and interpretability.As shown in Evaluation is an essential aspect of the modeling process, it is performed to assess the model's effectiveness, identify areas for improvement, and determine its suitability to the problem at hand.There are several evaluation metrics used in machine learning, including Confusion Matrix [15,18], Precision and recall, F1-score, ROC Curve and AUC [24], and Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) [9,26].These metrics offer a numerical assessment of the model's effectiveness and facilitate comparison with other models, facilitating the selection of the optimal model for the specific problem.As result, a combination of data and artificial intelligence can be used to predict IoT device failure.Many studies opt for using multiple algorithms in a hybrid approach, which results in better performance and accuracy compared to using a single algorithm.

Conclusion
The use of AI approaches for failure prediction in IoT devices has been the subject of significant research in recent years.This article has highlighted the key findings and contributions of the most relevant studies in the field, including the different AI approaches that have been used, such as machine learning and deep learning, neural networks, as well as specific techniques and algorithms that have been employed.However, it's important to note that challenges still exist, such as a lack of publicly available data sets for IoT devices, and a need for more robust evaluation metrics.Furthermore, the ethical implications of using AIbased failure prediction in IoT devices must be considered, as it could have an impact on users' privacy and security.
As future work, we plan to diagnose failures and factors that affect IoT devices and proposing a ML model for IoT device failure.

Fig. 2 .
Fig. 2. Growth of published and Scopus indexed IoT papers per year

Table 1 .
Summary of reviewed literature.