Prognostics and health management (PHM) increasingly play a constructive role throughout the entire lifetime of industrial equipment, significantly benefiting from extensive research in physical modelling and machine learning techniques. This has led to the development of hybrid approaches that seamlessly integrate both domains through physics-informed machine learning (PIML). PIML ensures the generation of cohesive solutions encompassing various aspects of physics knowledge across different stages of the machine-learning pipeline, substantially contributing to detection, diagnostics, and prognostics. However, PIML’s design relies heavily on expert experience and demands rigorous interdisciplinary expertise, requiring a profound understanding of machine learning and physical principles. Inadequate design of PIML often leads to suboptimal outcomes, where the combined effect is less than the sum of its parts. Currently, PIML lacks a scalable and engineered application architecture to effectively utilise its embedded results. To address this challenge, our paper introduces a novel parallel architectural approach that employs pre-training and fine-tuning strategies for optimising the different model parts. Its data-driven branch is first trained with zero output in the PI branch, then fine-tuning the physics-informed branch. It takes the frozen data-driven model as a fixed feature extractor to get physics-consistency prediction. This approach proposes a generic solution for embedding physics knowledge into ML that guarantees performance improvement. The effectiveness of our approach is validated in the context of Remaining Useful Life (RUL) prediction using MIT-Stanford battery data.

In the prognostics and health management (PHM) of industrial systems, prediction of remaining useful life (RUL) is a crucial task. RUL prediction is based on data collected from the industrial system, and involves learning underlying health indicator trends. As industrial systems are complex and can be monitored by different sensors, time alignment of multiple temporal data streams and extraction of their underlying characteristics are essential to perform an accurate prognostics. Hence, this paper aims to develop an efficient method to address the above issue. The proposed method is based on the attention and convolution mechanisms of deep neural networks. Its performance is highlighted when compared to other state of the art models such as RNN and LSTM using the C-MAPSS datasets. Numerous experiments demonstrate that our model provides better results in some situations, as well as an ability to capture both local short term contexts and long term associations.

Fault detection and diagnostics are important steps in the predictive maintenance of industrial systems, especially faults in the mechanical parts most susceptible to fail, such as bearings and gears in rotating machines. These two components represent more than 50% of causes of the operational downtime. Therefore, the detection of their appearance allows anticipating the total failure of the machine and schedule in advance maintenance actions. However, in the presence of a combined gear and bearing faults, it is difficult to isolate their states. To remedy this situation, this paper proposes a data processing methodology that exploits the three-phase current signals of the rotating machine and build a health indicator (HI) from each current phase that reveals the different health states. This indicator is constructed by extracting features from the collected raw data in frequency and time domains, and then they properly combined with a physical significance. After that, all health indicators (HIs) of the three phase current data are fed to a machine learning model for an online pattern recognition of the bearing and gear states, including the combined faults. The proposed approach is demonstrated through a test bench that studies bearing and gear defects of a gearbox under different operating conditions.

During the two last decades, the failure prognostics has been given an increasing importance by the researchers and industrials thanks to the benefits its results can induce in terms of availability, security, and maintenance costs. However, a detailed study of the literature in the domain shows that most of the published works deal with prognostics at component level, where the failure of the system is linked to the failure of its critical components. Even if this assumption can be tolerated in some cases, it cannot be generalized. This is because, in practice, the components are interconnected and, consequently, a degradation of one component can affect the others leading to a modification in the system's health state and then in its remaining useful life (RUL). To take into account this reality, a new model for prognostics is proposed in this article, namely inoperability input-output model (IIM). This model allows representing multicomponent systems while considering the interactions between their components. More than that, the IIM enables to integrate the mission profile effects on the system degradations to make its RUL predictions more accurate. To illustrate the effectiveness of the proposed modeling approach, a mechatronic system is considered as a case study.

This paper addresses the problem of uncertainty quantification in system-level prognostics. To this purpose, a three-step methodology, based on the inoperability input-output model, is presented. The first step concerns the estimation of the system inoperability, using a new adapted particle filtering method, while considering the interactions between its components. The second step focuses on the long-term prediction of the system inoperability in order to determine its evolution. Finally, in the third step, a method for calculating the remaining useful life of the system, based on its configuration, is formulated. The proposed methodology is applied on data obtained from the Tennessee Eastman Process simulations to predict the shutdown due to violation of process constraints.

Health monitoring data are vital for failure prognostic and maintenance planning. The continuous monitoring data or frequent inspections can provide a large amount of information on degradation evolution and therefore ensure the quality of deterioration modeling and the lifetime prognostic accuracy. However, they are usually very costly and sometimes impracticable in real engineering applications. Therefore, it is essential to address the issue of the appropriate amount of monitoring data. This paper proposes a new methodology to help the companies improving their actual inspection/monitoring policy to reduce the operation and maintenance costs but also ensure the information quality. We investigate different types of inspection policies including the periodic or nonperiodic ones by considering multiples functions of the system degradation state that are linear, concave, or convex. The best policies are chosen based on a multiobjective optimization problem dealing with the inspection cost and the information level. The advantages and disadvantages of the proposed methodology are discussed through numerous numerical examples for different types of degradation process, particularly the Wiener and Gamma processes that have been largely addressed in the framework of degradation modeling.

Wireless communication links tend to be employed more and more in safety-critical railway applications. Their safe use in an advanced train control system (TCS) is an issue that is addressed in this paper by characterizing the TCS service interruption due to communication errors. More precisely, occurrence probabilities of single errors are first discussed. Then, we obtain probabilistic analytical expressions of several temporal conditions that lead to a TCS service interruption, here a train emergency braking (the critical event). The accuracy of this analytical approach is proved when the results are compared with those given by a simulation approach with a Petri net model. Additionally, as the use case related to the “trains' separation” is considered in this paper, an analytical evaluation process is proposed to discuss the tolerated time margins that can be fixed to limit the critical-event occurrence probability due to the wireless communication errors.