National Engineering School of Sfax

The Kebili oasis, located in southern Tunisia, is currently suffering from a serious salinisation problem affecting wells that draw from the Complex Terminal shallow groundwater aquifer. For this reason, an investigation study was carried out, including interviews with 68 well owners, as well as an analysis of historical drilling and logging data, and a campaign of 51 water samples was conducted for physicochemical and major ions analyses. The aim of this study was twofold: firstly, to characterise the extent and evolution of groundwater salinisation in the Kebili region for CT and CI wells; and secondly, to identify the causes. The results show that integrity problems in deep Continental Intercalaire (CI) wells are responsible for widespread Complex Terminal (CT) groundwater salinisation. Indeed, six deep wells (DW-1, 16, 35, 42, 43, 58) have shown an integrity issue responsible for a sudden and significant increase in the total dissolved salts, reaching a maximum concentration of 169.7 g/l. Four deep wells have already affected nearby shallow wells. This has led to increases in salinity and sometimes temperature, including 36 shallow water wells. In these wells, salinity has locally reached 25.1 g/l. Workovers carried out on some deep wells have resulted in significant improvements in water quality, but have not restored the original quality. In conclusion, an ongoing ecological disaster has been identified in the Kebili region, which is threatening the Complex Terminal groundwater resource on a regional scale. In order to regain control of this issue, urgent inspections and proactive measures should be implemented in all deep wells to prevent the serious consequences of integrity loss. Furthermore, it is recommended that a well integrity investigation be conducted. This will highlight the weaknesses in the current well design and improve it in the future.

With unperceived faults often leading to costly failures, safety hazards, and operational downtime, anomaly recognition is very important to the current engineering systems. Conventional rule-based and supervised learning approaches suffer from problems such as lack of labeled fault data and the inability to generalize to new anomaly patterns. This paper provides a systematic and reproducible comparative analysis of four state-of-the-art unsupervised machine learning methods for anomaly detection in industrial engineering systems: Isolation Forest (IF), One-Class Support Vector Machine (OC-SVM), Autoencoder neural networks (AE), and DBSCAN clustering. We test these techniques in a standard experimental environment and cross-validate them on multivariate time-series data in the publicly available SKAB industrial sensor benchmark. In addition to regular classification metrics, we study distributions of anomaly scores, Precision-Recall trade-offs, and computational efficiency. For the practicalities, experimental results indicate that Autoencoder delivers higher detection performance in terms of $\mathbf{F}$-score $\mathbf{= 0. 9 4}$ and $\mathbf{A U C}-\mathbf{R O C}=\mathbf{0. 9 7}$, and Isolation Forest gives the best accuracy/cost trade-off for real-time deployment. These results offer empirical guidelines for algorithm choice in resource-limited engineering environments on water-circulation industrial systems, and set a reproducible baseline for the field of hybrid and physics-informed anomaly detection. Generalization to other industrial domains remains a direction for future work.

Integrating Renewable Energy Sources (RESs) into the electrical grid is a pertinent and essential energy strategy in the context of the energy transition. Weather fluctuations have a significant impact on the solar system's and wind turbine's maximum output power. The Artificial Neural Network (ANN), will be used to determine this maximum output power in real time despite climate fluctuations. Using the Particle Swarm Optimization (PSO) technique, the current work suggests a way to maximize the strategic integration of Distributed Generators (DGs) in Low-Voltage Distribution Grids (LV-DGs). The research mixes solar and wind sources in a 70 / 30 ratio and focuses on a 33-bus radial distribution system in the Sfax area of Tunisia. Finding the best DG sites to enhance voltage profiles and reduce energy losses is the methodology's main goal. The method for load flow analysis is based on a backward/forward sweep. To show the remarkable effectiveness of the PSO algorithm in maximizing the integration of distributed generators into Low-Voltage Distribution Grids (LV-DGs), a number of MATLAB scenarios are provided and thoroughly studied.

Modern digital twins are emerging as new paradigms for operational monitoring, simulation, and predictive maintenance in industrial systems. However, constructing dynamic and realistic digital twins for complex engineering assets remains difficult due to system nonlinearities, imperfect physical models, and high-dimensional sensor data. This paper introduces a data-driven digital twin framework utilizing Long Short-Term Memory (LSTM) networks for temporal state prediction, Autoencoder-based modules for anomaly detection, and Physics-Informed Neural Networks (PINNs) for hybrid physics-data modeling. The framework is validated on the UCI Hydraulic Systems Condition Monitoring dataset. Results averaged over five independent runs show a Root Mean Squared Error (RMSE) of 0.032 ± 0.002, an F1-score of $0.93 \pm 0.01$ for fault detection, and a 40% reduction in false alarm rate compared to purely data-driven baselines. The paper also discusses class-imbalance handling, sensitivity to physical model simplifications, computational cost, and deployment feasibility, providing an interpretable foundation for autonomous predictive monitoring at industrial scale.

This study presents a regularized phase-field model for simulating quasi-static brittle fracture in 2D polycrystalline materials, grounded in Griffith’s theory and the Ambrosio–Tortorelli variational approximation. The quadratic degradation formulation (AT2) is employed to represent diffuse cracks through a phase-field variable d, ensuring computational linearity and intrinsic compliance with damage bounds. A quadratic degradation function with residual stiffness is incorporated to capture stiffness loss, while an additive decomposition of the elastic energy prevents non-physical crack propagation under compression. For polycrystalline materials exhibiting weak elastic anisotropy, a dedicated formulation is developed, integrating a spectral decomposition of the strain tensor and a modified effective driving force to describe mixed-mode (I–II) fracture under plane stress conditions. This approach, implemented in the finite element software Abaqus via a user-defined subroutine, enables the simulation of complex crack propagation patterns by accounting for differences in the critical energy release rates of modes I and II. The results highlight the influence of the degradation function on model accuracy, providing a foundation for advanced numerical simulations in computational material mechanics.