Condition-based maintenance (CBM) has attracted significant attention in advanced manufacturing systems. In practice, degradation of production units is often influenced by their production rates. Unlike traditional CBM models that treat maintenance optimization as a stand-alone problem, we develop a joint optimization model of condition-based maintenance and production for multi-stage manufacturing systems. The resulting joint dynamic optimization problem is formulated as a Markov decision process (MDP) with the aim at maximizing the expected total production benefit. To address the curse of dimensionality and improve scalability in decision-making, a hierarchical multi-agent reinforcement learning framework is put forth to solve the MDP, leveraging a two-level architecture that decomposes the problem into manageable subtasks for maintenance and production. A distributed training strategy is introduced to independently train stage-level agents, significantly improving scalability. Experimental results from a semiconductor manufacturing facility demonstrate that the proposed joint dynamic optimization method significantly improves production benefit by maintaining a balance between production efficiency and system reliability.

Learning factories have been increasingly applied in engineering research and education. In this paper, we develop a small-scale, low-cost, fully open-sourced learning factory that leverages AI and digital twins to demonstrate smart and resilient manufacturing. The learning factory consists of an AGV, three robots, two conveyors and their digital twins. Three typical use cases are designed to demonstrate the capabilities of the learning factory, i.e., material sorting through computer vision-based robot control, coordinating AGV, robots and conveyors for assembly, and digital twin-driven predictive maintenance. The results show that the developed learning factory can adequately support the teaching and research activities on digital twins and AI on manufacturing. Its fully open-source architecture and relatively low deployment cost make it an ideal solution for universities and research institutions.

In this article, we develop a data-driven approach to predict the reliability of multicomponent repairable systems, considering component dependencies. We estimate component reliability functions from system-level time-to-failure data without prior knowledge of the system structure and use these estimates to generate training data for a deep long short-term memory network. This leads to system reliability prediction and addresses uncertainties through quantile regression. Validated through simulations of 500 systems and real-world data from GE HealthCare magnetic resonance imaging (MRI) machines, our model outperforms traditional methods (such as Cox model and random survival forest) in terms of accuracy, particularly for complex systems, by effectively learning from uncertainties.

International audience

International audience

5G networks are designed to create a new ecosystem for vertical industries such as health care, energy, and public transport. These novel applications, on the other hand, bring new challenges to network resilience. Among them, traffic variation is one of the most vital threats to the 5G network. With tens of thousands of devices connected to the network, network service resilience is threatened by the heavy traffic change induced by the end users or malicious attacks. While long timescale traffic variation can be easily predicted based on historical data, short timescale abnormal traffic is hard to forecast yet can significantly violate the service requirements. The impact of short timescale traffic variation can be mitigated by 5G management and control systems. However, the complexity and dynamics of the virtualized 5G system make it hard to estimate its resilience. This paper provides a 5G network model that captures the data traffic changes and network dynamic management mechanism. The model is able to evaluate the performance of different network services with different requirements under traffic variation events. We analyze the effectiveness of auto-scaling and compare different isolation strategies for traffic congestion. The simulation results on service resilience estimation can become strong supporting information for 5G network deployment and configuration.

Expert knowledge is an important information source for system reliability assessment, especially when historical data are limited. However, when elicited, expert knowledge is often imprecise with large uncertainty. Moreover, as experts usually own different expertise and knowledge, the elicited knowledge from different experts might be conflicting. In this article, a two-stage optimization model is put forth to fuse the imprecise and conflicting information from multiple sources to assess reliability of multistate systems. The degradation of the components in the multistate system is modeled via imprecise Markov models. Then, in the first-stage optimization, upper and lower bounds of the degradation model parameters are determined by minimizing the conflict between the prediction of the model and the multisource imprecise information elicited from experts. A particle swarm optimization algorithm is tailored to solve the computational problems brought by the presence of high-dimensional decision variables and resolve the optimization problem. The second-stage optimization is, then, conducted to identify the upper and lower bounds of the system reliability function given that the degradation model parameters are constrained in the bounds obtained from the first-stage optimization. A numerical example, along with a radar display and control system, is used to demonstrate the effectiveness and applicability of the proposed method.

A large variety of real engineering systems operate with multiple performance measures that are multistate in nature. These systems are usually modeled as multiperformance multistate systems (MPMSSs). However, existing MPMSS models fail to consider an important aspect, i.e., the performance conversion process. For example, in a combined heat and power (CHP) generating unit, apart from the output heat and electricity, decision-makers are also interested in the unit's capacity to convert gas into electricity and heat. The latter is related to the performance conversion process. This article proposes a framework for the reliability evaluation of performance conversion-based MPMSS. In the proposed MPMSS model, the couplings among different types of performances inside the components are quantified into the multistate performance conversion matrix. The performance conversion structure functions are proposed to derive system performance conversion capability based on the conversion capabilities of the components. Two reliability evaluation methods considering the steady-state performance conversion process and the continuous-time performance conversion process are proposed, respectively. Numerical examples are given to demonstrate the developed methods.

In engineering scenarios, expert judgments play an essential role in reliability assessment, especially for those systems with few historical data. To achieve a rational result, experts from different areas should be involved, and the uncertainties in their assessments should be properly addressed. Such information is often referred to as multisource imprecise information (MSII) and might contain high degree of conflicts, as different experts usually have different expertise and knowledge. Properly quantifying the conflicts among the MSII, then, becomes a critical issue, as the subsequent processing of MSII (e.g., combination and calibration), depends on the degree of conflict in the MSII. To this end, a new conflict measure is put forth based on the Dempster–Shafer theory (DST) to quantify and visualize the conflict in the MSII from a group of experts. In the first place, the MSII from each expert is used to construct the basic belief assignment (BBA) of the reliability estimates for the corresponding expert under the DST. A 2-D conflict measure, which combines the conflict factor and Jousselme distance in DST, is, then, proposed to measure the conflict between the experts’ BBAs. The conflict is quantified from two perspectives, viz., mutual conflict and total conflict. Finally, a Bhattacharyya distance-based method is developed to further quantify the informativeness of each expert's MSII to the system reliability estimate. A numerical example along with an engineering case is used to validate the effectiveness of the proposed approach.

In this article, we develop a mixture of Gaussians-evidential hidden Markov model (MoG-EHMM) to fuse expert knowledge and condition monitoring information for remaining useful life (RUL) prediction under the belief function theory framework. The evidential expectation-maximization algorithm is implemented in the offline phase to train the MoG-EHMM based on historical data. In the online phase, the trained model is used to recursively update the health state and reliability of a particular individual system. The predicted RUL is, then, represented in the form of its probability mass function. A numerical metric is defined based on the Bhattacharyya distance to measure the RUL prediction accuracy of the developed methods. We applied the developed methods on a simulation experiment and a real-world dataset from a bearing degradation test. The results demonstrate that despite imprecisions in expert knowledge, the performance of RUL prediction can be substantially improved by fusing expert knowledge with condition monitoring information.

A sequential Bayesian approach is presented for remaining useful life (RUL) prediction of dependent competing failure processes (DCFP). The DCFP considered comprises of soft failure processes due to degradation and hard failure processes due to random shocks, where dependency arises due to the abrupt changes to the degradation processes brought by the random shocks. In practice, random shock processes are often unobservable, which makes it difficult to accurately estimate the shock intensities and predict the RUL. In the proposed method, the problem is solved recursively in a two-stage framework: in the first stage, parameters related to the degradation processes are updated using particle filtering, based on the degradation data observed through condition monitoring; in the second stage, the intensities of the random shock processes are updated using the Metropolis-Hastings algorithm, considering the dependency between the degradation and shock processes, and the fact that no hard failure has occurred. The updated parameters are, then, used to predict the RUL of the system. Two numerical examples are considered for demonstration purposes and a real dataset from milling machines is used for application purposes. Results show that the proposed method can be used to accurately predict the RUL in DCFP conditions.

Traditional quantitative risk assessment methods (e.g., event tree analysis) are static in nature, i.e., the risk indexes are assessed before operation, which prevents capturing time-dependent variations as the components and systems operate, age, fail, are repaired and changed. To address this issue, we develop a dynamic risk assessment (DRA) method that allows online estimation of risk indexes using data collected during operation. Two types of data are considered: statistical failure data, which refer to the counts of accidents or near misses from similar systems and condition-monitoring data, which come from online monitoring the degradation of the target system of interest. For this, a hierarchical Bayesian model is developed to compute the reliability of the safety barriers and a Bayesian updating algorithm, which integrates particle filtering (PF) with Markov Chain Monte Carlo, is developed to update the reliability evaluations based on both the statistical and condition-monitoring data. The updated safety barriers reliabilities, are, then, used in an event tree (ET) for consequence analysis and the risk indexes are updated accordingly. A case study on a high-flow safety system is conducted to demonstrate the developed methods. A comparison to the DRA method which only uses statistical failure data shows that by introducing condition-monitoring data on the system degradation process, it is possible to capture the system-specific characteristics, and, therefore, provide a more complete and accurate description of the risk of the target system.