State-of-the-art artificial intelligence approaches for anomaly detection and remaining useful life prediction: a review

RQ1 What are the range of studies associated with predicting the RUL of equipment?

Table 8 measures studies interest distributed based on stage in life span where the bathtub curve model was used to identify the range of study because it defines components failure rate over time (Titu-Marius, 2021). In Fig. 4, The bathtub curve typically comprises of three stages: an initial diminishing failure rate (infant mortality), a consistent failure rate (useful life), and a rising failure rate (wear-out) (Ikonen, Corona & Harjunkoski, 2023). Numerous models and methodologies have been devised to precisely elucidate and conform to this curve to empirical data. For example, the adjusted exponential-Weibull (MEW) model, which merges the exponential and Weibull distributions, presents a versatile approach to accommodating failure time data with a bathtub-shaped hazard rate, delivering superior outcomes in comparison to other models (Al-Essa et al., 2023).

The early stage is supported by one study Zheng, Liao & Zhu (2023), indicating relatively less research focus on the high initial failure rates caused by manufacturing defects or early use issues. The absence of operational data is identified as a significant challenge, as RUL predictions heavily depend on this data, which is not available during the design and production stage since the equipment has not yet been put into use. In contrast, The operation stage has the most extensive research, with 19 studies. This extensive coverage indicates a well-established understanding of the low and normal constant failure rates during this phase, supporting effective maintenance strategies and reliability planning based on a broad consensus regarding random failures often due to external factors. The wear-out stage, which has been covered in eight studies, is mentioned divided into main two subphases: early wear out and late wear out stage. The transition phase marks the onset of increasing Major failure rates, accumulation of fatigue and outdated components. The late wear-out stage is defined by a sharp increase in failure frequency, signifies the approaching End of Life. Maintenance costs escalate, and decisions must be made regarding asset replacement, decommissioning, or life extension investments. This stage is particularly relevant for economic analysis comparing continued operation vs. replacement costs. Overall, the distribution of studies highlights well-researched middle and late stages, while the early stage presents an opportunity for further investigation to enhance early failure mitigation.

RQ2 How can hybrid models combining physical and data-driven methods improve RUL prediction for complex systems?

Accurately predicting the RUL is crucial for ensuring equipment achieves its expected life span. Hybrid models, which integrate two or more methods, offer a promising approach to enhance prediction accuracy by combining both physical principles with data-centric techniques particularly in a complex system. A common hybrid approaches, integrates physical-based with data-driven model. Physical-based models rely on mathematical representations of failure mechanisms, providing a fundamental understanding of system behavior (Wang & Zhao, 2023). Combining support vector regression (SVR) and random forest regression (RFR) alongside an exponential weighted moving average (EWMA) control chart for anomaly detection (Arunthavanathan et al., 2023; Wang & Mamo, 2019).

Meanwhile data-driven models, utilized historical data to identify patterns and predict future states, offering flexibility and adaptability to varying conditions (Wang & Zhao, 2023). In time frequency data transformations, Short-time Fourier Transform (STFT) and Wigner Ville Distribution (WVD), used to analyze non-stationary signals, which are common in complex systems (Rosero, Silva & Ribeiro, 2022). While in DL, CNN and RNN utilize in extracting multilevel features from data. PdM is a common example application of data-driven models. The integration with sensor and AI data processing technique enhanced the prediction and allow for faster decision making (Sayyad et al., 2021). Plus, a combine method will increase computational requirement and the accuracy of data-driven models hinges on the quality of their input data. Noisy data, often stemming from environmental influences, can significantly impact prediction accuracy (Biondi, Sand & Harjunkoski, 2017; Fan, Nowaczyk & Rögnvaldsson, 2020; Ma, Xu & Yang, 2023). Effective data pre-processing techniques are crucial to mitigate these issues and enhance model reliability. Furthermore, data-driven models must exhibit adaptability to varying operating conditions prevalent in industrial settings. Robust feature extraction methods and domain adaptation techniques are crucial to maintain prediction accuracy across diverse scenarios (Wang & Zhao, 2023).

RQ3 What are the comparative accuracies between Artificial Intelligence approaches in RUL prediction?

Among the most effective machine learning methods for predicting equipment’s RUL, deep learning techniques such as LSTM networks have demonstrated remarkable results (Table 9). LSTM combine with Autoencoders (LSTM-AE) produce high accuracy, provides monotonicity (0.38), trend ability (0.95), and prognosability (0.94). This highlights its ability to provide consistent health indicators outperforming other autoencoders such as gated recurrent unit autoencoder (GRU-AE) and bidirectional long short-term memory autoencoder (BiLSTM-AE) (de Pater & Mitici, 2023; Wu et al., 2022). On fault detection, the use of Robust-ResNet combined with LSTM and CNN architectures, achieves up to 99.53% accuracy (Zheng, Liao & Zhu, 2023). LSTM with COSMO features reduce mean absolute percentage error (MAPE) between 13–15%, compared to 25% with traditional methods (Duan et al., 2023). This approach capable in managing large amounts of temporal data and detecting subtle changes indicative of degradation or failure (de Pater & Mitici, 2023; Fan, Nowaczyk & Rögnvaldsson, 2020; Wang & Zhao, 2023; Zheng, Liao & Zhu, 2023).

In a situation of understanding data patterns in sequence, recursive neural network (RNN) is used. These in combination with the extended Kalman filter (EKF) reduces mean absolute error (MAE) by 15–25% and MAPE by 10–20% compared to standalone RNNs or traditional models, improving prediction accuracy for nonlinear, noisy datasets. While differentiating abnormality, OC-SVM used to distinguish abnormalities by separating them from normal operating conditions. The fault margin was dynamically adjusted from data patterns by considering the highest anomaly count in a window and incorporating a noise margin for accurate anomaly detection (Arunthavanathan et al., 2023).

CNN architecture was developed that incorporates domain adaptation techniques to minimize distribution discrepancies across different failure modes. Notable studies, such as those by Cheng et al. (2021), Rosero, Silva & Ribeiro (2022), Solís-Martín, Galán-Páez & Borrego-Díaz (2023) and Ture et al. (2024), have highlighted the utility of CNNs in this domain. In TCNN, metrics assess higher accuracy and robustness RUL of bearings under multiple failure behaviors with a low MAE of 0.10, root mean square error (RMSE) of 0.12, and a high R-squared (R²) of 0.82 (Cheng et al., 2021). CNN has achieved an accuracy of 93.93% performed well among DL models but slightly lower than stacking based ensemble learning (95.72%) (Ture et al., 2024).

Dynamic Bayesian networks (DBN), and Wiener process models offer robust frameworks for modeling temporal and degradation processes in corrective maintenance. DBN incorporates time-dependent variables and their probabilistic dependencies, enabling effective representation of dynamic systems and their evolution over time. However, Wiener process models characterize degradation trajectories by accounting for both gradual wear and negative jumps caused by imperfect maintenance. These approaches as discussed in studies Wang et al. (2018), Liu et al. (2020) and Zhang et al. (2020) provide robust tools for understanding and predicting equipment performance under varying operational conditions. Through a comparison with experimental data, DBN shows an error margin lower than 8.5% in the first 4 years, lower than 20.4% between 5–10 years and lower than 11.3% after 10 years (Liu et al., 2020).

Moreover, the use of Kalman filtering not only made an average improvement 18% in prediction results, but also reduce 10–15% compared to static methods like maximum likelihood estimation (MLE) (Pei et al., 2019). The least squares support vector machine (LS-SVM) algorithm with the artificial fish swarm—artificial bee colony algorithm (AFS-ABC) increase prediction in potential failures and assess equipment health. This model has accomplished lower error forecast compared to other methods (Ma, Xu & Yang, 2023). Bayesian methods, which involve developing exponential degradation models and updating parameters using real-time condition monitoring data, offer another effective approach (Liao & Tian, 2012).

Bar chart review methods in grouped categories based on functions (Fig. 5). For time-series analysis, methods like LSTM and RNN help track changes over time (four methods). Feature extraction, like CNN and ResNet, pulls useful patterns from data (two methods). Statistical modeling, such as Kalman filtering and Wiener process, uses mathematical approaches to predict outcomes (two methods). Anomaly detection, like OC-SVM and LS-SVM, identifies unusual patterns (two methods). Ensemble methods, like stacking, combine multiple models for better accuracy (one method) and general models, such as MLP and ResNet, are flexible tools for general predictions (two methods). These visualizations offer a well-rounded understanding of how different machine learning techniques, with components ranging from 1 to 4, utilize varying levels of complexity in their approaches to RUL prediction.

The high dependency on AI algorithms raises the risk of algorithmic biases, impact the fairness and reliability of these systems. Despite their prevalence, none of this research mentioned on the influence of these biases or proposed mitigation strategies. Algorithmic models may also reinforce historical inequalities by assuming that past trends predict future outcomes, as seen in recommender systems that limit diversity by continuously suggesting similar products.

RQ4 How effective are FDD approaches in detecting early indicators of equipment failure?

Figure 6 shows an overall schematic flow of FDD integration to anomaly detection. The integration of advanced technique such as FDD has significantly enhanced the ability to detect early indicators of equipment failure (Rosero, Silva & Ribeiro, 2022). Over the years, the technique has evolved and categorize into mathematical, analytical, data-driven, statistical, computational, and hybrid approaches (Zhao et al., 2019). To establish into one of these methods, variables associated with the fault need to be identified in advance. Abnormal values are detected by manually setting a threshold for these variables. These thresholds vary across different system components and must be adjusted for each specific application (Arunthavanathan et al., 2023; Nelson & Culp, 2023).

Table 10 encapsulates classification for fault detection studies in FDD strategy. In data-driven methods Carroll et al. (2019) and Wang & Mamo (2019) demonstrates the real-time sensor to reflect the dynamic behavior of systems efficiency, combine with artificial neural networks (ANNs) in predicting gearbox failures using SCADA and vibration data to achieved higher accuracy particularly when using high-frequency vibration data (Carroll et al., 2019). de Pater & Mitici (2023) focus on LSTM autoencoders, showing that the reconstruction error, which increases with system degradation, can effectively identify early faults. Statistical methods are highlighted by Aydemir & Acar (2020) discuss the use of the CUSUM control chart for anomaly detection, which helps identify the degradation onset point and subsequently improves the performance of RUL estimators by focusing on active degradation periods. Wang & Mamo (2019) employ the EWMA control chart for anomaly detection, which is crucial in their hybrid approach, where detected anomalies trigger RUL prediction models, thereby enhancing early fault detection. Model-based methods such as those described by Ma, Xu & Yang (2023) utilize deterministic reasoning and fuzzy inference for precise and uncertain data respectively, achieving effective fault detection. Hybrid methods, including the approaches by Arunthavanathan et al. (2023), Lee, Kim & Lee (2023), combine techniques like OC-SVM for fault detection with neural network permutation algorithms for classification, and use alarm rules alongside contextual diagnosis to enhance fault detection accuracy.

System performance monitoring further model known faults (as supervised learning), while unsupervised learning detects to unknown anomalies or unusual patterns in the data. This real-time monitoring ensures that even early signs of failure are noticed. Finally, the models are tested during validation to ensure they work accurately, and the insights are stored in a knowledge base for future use. Once validated, the system is deployed in the implementation stage, where it operates as part of real-time monitoring to predict and prevent equipment failure effectively. Unlike existing reviews, this assesses the FDD efficiency, systematically evaluates their performance across various contexts, accentuating pragmatic applications. Whereas the most of the reviews concentrate on singular methodologies or particular domains, highlighting the integration of hybrid and data-driven techniques, which are often overlooked in similar studies, the review underscores the amalgamation of data-centric and hybrid strategies, offering an all-encompassing view of their real-world ramifications.

RQ5 What challenges are associated with the implementation of sensor technology for fault detection?

One of the methods in early detection equipment malfunction is implementing sensors as a primary data in FDD. This proactive approach minimizes downtime, reduces costs, and improves safety across various applications. Table 11 summarizes the integration of multiple sensors for continuous monitoring.

RQ6 What are the most effective methods use for monitoring condition in the predictive maintenance and how can root cause analysis be effectively determined in fault tolerance systems?

Table 12 highlight method currently practice in monitoring asset and system. One of the most common effective methods is predictive maintenance (PdM), which utilizes sensor data and advanced analytics to predict equipment failures before they occur. Wang et al. (2018) emphasized the importance of predictive maintenance of degrading systems, thereby improving overall reliability and maintenance scheduling. Aydemir & Acar (2020) demonstrated that anomaly monitoring significantly improves RUL predictions, ensuring timely and effective maintenance interventions. Condition-based monitoring (CBM) is another effective method that involves continuous or periodic monitoring of equipment condition using sensors to detect deviations from normal operation. Zheng, Liao & Zhu (2023) developed a fault detection model for internal gear pumps, which enhances the effectiveness of CBM by accurately detecting faults and predicting RUL. These advanced monitoring techniques, combined with the use of machine learning and AI, such as the work by Cheng et al. (2021) using transferable convolutional neural networks, provide robust solutions for fault detection and RUL predictions. Root cause analysis (RCA) is a systematic method that involves collecting data, analyzing failure modes, and identifying the underlying reasons for faults. This approach ensures that the real cause of the problem is addressed rather than just treating the symptoms. Arunthavanathan et al. (2023) highlighted the significance of RCA in estimating RUL and transforming fault-to-failure processes in process systems. Fault tree analysis (FTA) and failure mode and effects analysis (FMEA) are additional methods that support RCA by providing structured frameworks for identifying and prioritizing potential causes of failures.

RQ7 How do life cycle analysis and survival analysis frameworks influence the selection of maintenance strategies and health indices in reliability driven maintenance?

Life cycle analysis (LCA) and survival analysis frameworks significantly influence the selection in reliability-driven maintenance by providing a structured proactive and reactive approach to evaluate and optimize maintenance decisions. Preventive maintenance (PM) is characterized by its proactive approach, involving regular inspections, servicing, and timely interventions to prevent equipment failures. While corrective maintenance (CM) is reactive approach, initiated only after equipment has failed. This approach solves the issues in the short term but leads to higher overall costs and reduced equipment lifespan. According to Wang et al. (2018), continuous PM can substantially extend the lifespan of equipment, prolong RUL and predict failures before it occurs (Pei et al., 2019). Effective maintenance scheduling and consequently, a longer equipment lifespan (Aydemir & Acar, 2020). CM is reactive, initiated only after equipment has failed. While this strategy might seem cost-effective in the short term due to lower initial maintenance expenditures, leads to higher overall costs and reduced equipment lifespan. de Pater & Mitici (2023) assert that PM’s focus on timely and planned interventions not only improves reliability but also optimizes maintenance resources. Cheng et al. (2021) illustrate how integrating AI with PM protocols enhances RUL predictions and fault detection accuracy.

CM results in extended downtimes and higher repair costs because failures are addressed only after they have occurred, often leading to significant damage (Sayyad et al., 2021). Carroll et al. (2019) mentioned the unpredictable nature of failures under CM necessitates expensive emergency repairs and replacements, further diminishing the equipment’s operational life. Ture et al. (2024), demonstrated that the implementation of predictive maintenance algorithms within PM frameworks significantly reduces unexpected failures and maintenance costs, thereby extending the operational life of assets.

In the context of equipment reliability, life cycle analysis helps identify potential failure points and maintenance needs at different stages of the equipment’s life. de Pater & Mitici (2023) demonstrated by understanding the degradation effect of wear and tear item, the implementation of predictive maintenance strategies will help preventing unexpected failures and enhancing overall reliability. Integrating LCA with cost analysis allows for the identification of the most cost-effective maintenance interventions. Aydemir & Acar (2020) use CUSUM in anomaly detection techniques, an integration LCA with cost analysis to enhances the accuracy of RUL estimation.

In survival analysis, reliability measured in statistical approach, focusing on predicting the time until a system fails based on its current condition and operational history (Arunthavanathan et al., 2023). The analytical approach places emphasis on survival function analysis, time to failure analysis and hazard function analysis.

Predicting RUL multiple asset under different stage of life span

Predicting the life span in the early stages of a component’s life is crucial for newly develop product, this to preventing unexpected failures in the early design stages (Haobin, Zhang & Sinha, 2024). Table 13 display predictable test used in different stages of the equipment’s life (Qian, Yan & Gao, 2017; Kim et al., 2004). Accelerated life testing (ALT) is a critical methodology used in manufacturing with its primary goal to accelerate the aging process of components, thereby obtaining significant life span data in a shorter period (Qiu & Li, 2024). In the middle stage, multiple factors involve in life prediction from utilization of usage, corrective maintenance, replacement parts and upgrading is a multifaceted challenge that requires integrating various advanced methodologies (Noot, Martin & Birmele, 2025). A multi-stage maintenance-impact degradation model based on the Wiener process can account for dynamic maintenance and failure thresholds, thereby improving the precision of RUL predictions (Li et al., 2024). Addressing uncertainties in RUL estimation is essential for improving the reliability and accuracy of predictive methods. These uncertainties fall into two main categories: epistemic and aleatory. Epistemic uncertainty comes from limited knowledge or incomplete information about the system, while aleatory uncertainty is due to the inherent randomness and variability in the system’s behavior. Various strategies have been proposed to tackle these uncertainties (Cao & Peng, 2023). Model-based strategies frequently encounter difficulties with complicated connections and uncertainties, whereas data-driven approaches sometimes neglect previous knowledge and struggle with restricted data (Liang, Liu & Xiao, 2024).

Integration of FDD and RUL estimation for the early detection of system faults and the prediction of the system’s future operational life, facilitating timely maintenance actions and reducing unexpected downtimes. A novel tree network framework can address fault classification and RUL prediction in parallel, improving model selection accuracy and prediction efficiency (Chai et al., 2024). Similarly, a joint learning model that simultaneously performs failure mode recognition and RUL prediction by leveraging multiple sensor signals has shown promising results (Wang, Xian & Song, 2023). Combining system modeling methods with regression-based approaches and genetic programming algorithms to predict fault occurrences and estimate RUL, even in the presence of measurement noise (Bahareh & Jørn, 2023). A range of studies have explored the use of probabilistic, highlight the importance of considering uncertainties in input variables. Zamzam et al. (2021) established a ranking assessment, prioritizing and predictive systems both for medical equipment maintenance, using machine learning algorithms for medical equipment. These studies collectively underscore the potential of probabilistic models in improving the accuracy and effectiveness of life cycle cost analysis in the medical equipment domain.

Uncertainty in equipment deterioration

There is often an incomplete information about all the factors influencing equipment deterioration. Models used to predict deterioration often rely on assumptions and historical data, which may not account for all possible future scenarios, leading to uncertainty in predictions (Zhu et al., 2024). State of degradation on equipment may consist of multiple types, including natural degradation, static degradation, and force-induced degradation, each contributing differently to the equipment’s end-of-life (Fig. 7). Natural degradation refers to the gradual wear and tear that occurs over time due to regular usage and environmental factors (Li et al., 2023).

Performance degradation assessment (PDA) methods, which utilize statistical and intrinsic energy features, are crucial for evaluating such static degradation states by constructing high-dimensional feature sets and reducing them to sensitive health indices (Lv, Hu & Wang, 2023). Zhang et al. (2020) emphasize the role of health indicators in maintenance strategies by enabling more effective scheduling of maintenance activities. Zheng, Liao & Zhu (2023) develops health indicators based on the fault types classification and RUL stages, thus allowing for timely intervention. Further models and methods enhancement have been developed to enhance the predictions to support maintenance planning. For example, Pei et al. (2019) and Duan et al. (2023) highlight the importance of accurate RUL estimation for maintenance management, remarking that reliable predictions enable better scheduling of maintenance activities and improve the overall reliability of systems. The integration of ML and Internet of Things (IoT) has transformed maintenance management systems into a comprehensive solution, allowing for real-time data processing and more sophisticated analysis of health indices (Cheng et al., 2021; Lee, Kim & Lee, 2023).

Digital twin and virtual assistance

The integration of digital twin (DT) and virtual assistance (VA) technologies ML is revolutionizing fault identification, monitoring prediction in asset management. This is crucial in assisting decision making related to RUL identification. DT technology involves creating a digital replica of physical assets, capturing real-time data to mirror their operational behavior (Abdullahi, Longo & Samie, 2024). This process includes data acquisition from sensors and IoT devices, data integration to form a cohesive dataset, model development using simulation tools and ML algorithms, continuous monitoring, and fault identification and prediction through ML analysis (Alam & El Saddik, 2017; Solari, Lysova & Montanari, 2023).

While, virtual assistants provide interactive support and decision-making capabilities, aiding maintenance teams in managing RUL and detecting anomalies. VA technology uses AI to process data, interact with users through conversational interfaces and implement automated actions based on predictive analytics thus to predict RUL estimation. VA tools are categorized into chatbots, virtual advisors, and autonomous agents, each providing varying degrees of interaction and decision-making capabilities. Research indicates significant benefits of integrating DT and VA with ML for predictive maintenance. Lu & Li (2023) show that for rolling element bearings, a hybrid DT and LSTM model significantly improved RUL prediction accuracy by integrating simulation data with experimental data, achieving over 97.5% accuracy.