Research Team

Project Leader

Photo of Bouzrara Kais

Bouzrara Kais

Position: Professor

Involved Faculty Members

Mbarek Abdelkader

Position: Associate Professor

Krifa Abdelkader

Position: Associate Professor

Maraoui Saber

Position: Assistant Professor

Maraoui Safa

Position: Assistant Professor

Ben Mabrouk Walid

Position: Assistant Professor

PhD Students to be mobilized within the project

Kerkeni Rochdi

Ben Gharat Ibtissem

Oussaifi Mejdi

El Wavi Sidi Abdallah

Summary and Objectives

Summary

The studied theme mainly concerns the use of artificial intelligence methods for two main research axes:

  • Modeling and identification of complex systems.
  • Detection and diagnosis of faults in industrial systems.

Artificial intelligence (AI) has become an essential field in the development of innovative solutions for various scientific and technical disciplines. In the context of automatic control and systems control, modeling and diagnosis of dynamic systems are crucial steps to understand and control complex processes. Traditionally, these processes have been approached using classical mathematical models such as linear and non-linear models, which have shown their limitations when they must describe increasingly complex and unpredictable systems.

Regarding the modeling of complex systems, the use of artificial intelligence algorithms opens new perspectives by combining the expressiveness of NARX-Laguerre type models or, more generally, models exploiting orthogonal bases, with the learning capacity of AI algorithms (such as neural networks, optimization by evolutionary algorithms, or reinforcement learning). Such algorithms could allow a better understanding and modeling of complex systems, where classical approaches often fail. It should be noted that models based on orthogonal bases present interesting potential to represent a wide class of dynamic systems, whether linear or non-linear. These orthogonal bases allow describing complex behaviors, while retaining interesting algebraic properties for system parameter identification.

Fault detection and diagnosis play an essential role in maintaining product quality and safety, as well as improving the overall efficiency of industrial processes. These techniques allow quickly identifying anomalies or deviations from normal operating conditions, thus enabling rapid corrective actions to prevent or mitigate potential faults. A fault is any unauthorized deviation of at least one parameter or process variable in the facility. Thus, it is necessary to use process monitoring methods to detect faults and maintain processes in a safe and reliable state. Fault detection, diagnosis, and prognosis are the main phases of process monitoring. The detection step aims to determine the existence of a fault in the system, the diagnosis step aims to determine which fault occurred, and fault prognosis aims to predict the future behavior of the process anomaly. Fault detection and diagnosis are two important research areas that have attracted great attention in academic and industrial fields. By proactively planning maintenance activities based on fault detection, costly downtime and unplanned stoppages can be minimized. Fault detection and diagnosis can be classified into model-based, knowledge-based, and data-based approaches. Model-based approaches rely on comparing the observed process behavior, as detected by sensors, with the expected behavior predicted by a mathematical process model. This model generally stems from a comprehensive understanding of the process under normal and fault-free conditions, often based on physical and chemical principles. The comparison between actual process measurements and model predictions produces a residual that indicates whether the process is operating normally or encountering faults. However, the effectiveness of model-based methods depends on the accuracy of the mathematical model, which can be difficult to develop due to the complexity of industrial processes and the many parameters involved. Knowledge-based approaches refer to systems or methods that rely on human expertise, domain knowledge, or established principles to make decisions or solve problems. Knowledge-based systems often use rules, heuristics, or knowledge bases to guide their reasoning and decision-making processes. On the other hand, data-based techniques use historical data collected during fault-free processing operations. These methods exploit training data to build an empirical model, which is then used to detect faults by analyzing future measurement data.

Among the applications studied in this project, it is important to mention photovoltaic systems, wind systems, and electrical networks. The latter were built to supply electricity from centralized production to fixed customers and expected loads. With the integration of distributed generation sources from microgrid and energy storage technologies, distribution systems become more decentralized and bilateral in nature to achieve system reconfiguration and self-healing capability of the smart grid. Many disputes accompany this new structure and involve regulating bidirectional energy flow, voltage, and oscillation damping in a balanced manner within the network. In these situations, many types of faults can occur in the network, such as source and load side faults, converter faults (inverter and converter), cable faults, data communication faults, cybersecurity, failure of the Internet of Things (IoT) protocol, data leakage, insufficient data, smart meter failures, etc., which are difficult to manage, detect, and control. The increase in faults in various elements of smart grids is a major problem that includes elements of the power production and distribution framework. Moreover, faults can manifest through different failure modes in the same element.

Project Objectives

The objectives of this project are multiple:

Development of AI algorithms for modeling

Develop artificial intelligence algorithms for modeling and identification of linear and non-linear systems, using orthogonal bases, while ensuring accuracy, robustness, and generalization of the identified models.

Advanced fault detection and diagnosis methodologies

Develop advanced data-based fault detection and diagnosis (FDD) methodologies in industrial systems to improve their operational performance. To do this, different solutions can be studied:

  • First, improved fault detection using reduced kernel principal component analysis (KPCA) approaches can be applied to detect faults in industrial applications. Reduced kernel PCA-based models only consider specific features that show stronger correlation while preserving the main statistical characteristics of the original dataset.
  • Second, develop improved techniques based on fault diagnosis to distinguish operating modes in certain and uncertain systems. Methodologies can improve diagnostic performance at three levels: data preprocessing, feature extraction and selection, and decision making.
  • Third, validate the developed methodologies using simulated datasets.

Application on recognized complex systems

Apply the developed techniques for modeling and diagnosis on recognized complex systems such as:

  • The Tennessee Eastman Process (TEP)
  • Autonomous two-degree-of-freedom flying vehicles
  • The IEEE 9-bus electrical network

Research Program and Methodology

Methodological Approach

The team working on this theme of modeling and diagnosis of complex systems using intelligence algorithms first develops optimization and regression algorithms to model complex systems. In a second step, these algorithms will be used for fault detection. The last step is dedicated to developing classification algorithms to identify the type of fault.

Development of optimization algorithms

Develop algorithms for optimizing the parameters of orthogonal bases using artificial intelligence tools. The optimization of these parameters plays a key role in reducing model complexity.

Application of AI algorithms

Application of artificial intelligence algorithms, particularly machine learning and deep learning, for regression of load flow results on the IEEE 9-bus network. Among these methods, we mention MLP, RBFN, SVM, and CNN.

Enhanced FDD techniques

For the fault detection and diagnosis (FDD) phase, enhanced data-driven machine learning techniques are developed. These FDD techniques consist of two main steps: feature extraction and selection (FES) and fault classification (FC). In the FES step, the goal is to extract the most relevant and effective features from the data. Different strategies based on feature extraction using RKPCA methods are developed. Once the features are extracted, only the most relevant features must be selected. In this regard, multivariate statistical methods and statistical measures are used. During the FC phase, the selected features are passed to a Random Forest (RF) classifier.

AI-based FDD approaches

FDD approaches using artificial intelligence methods, including machine learning and deep learning, are developed to manage electrical quantities in power systems. These techniques rely on the analysis of load flow data, enabling more robust anomaly detection and more accurate fault diagnosis. The feasibility and effectiveness of the proposed AI models are evaluated under both normal conditions and in the presence of failures. Through simulations and experimental validations, AI-based approaches demonstrate optimized performance, both in terms of computational speed and accuracy of diagnostic metrics, thus offering adaptive solutions for power grid supervision.

Project Implementation Timeline

First Year

Literature Review

Literature review on:

  • Orthogonal basis modeling of complex systems
  • Development of control laws for complex systems
  • Artificial intelligence methods dedicated to regression and classification
  • Fault detection techniques based on data-driven methods
  • Electrical networks
Second Year

Algorithm Development

  • Proposals for robust predictive control algorithms for complex systems
  • Mathematical formulation of optimization problems
  • Proposal of regression methods for modeling complex systems
  • Proposal of new fault detection methods based on data-driven techniques
Third Year

Optimization and Validation

  • Proposal of optimization algorithms for predictive control of large and complex systems
  • Analysis of system stability and performance
  • Proposal of new data-driven process modeling methods
  • Proposal of data-driven diagnostic methods using artificial intelligence tools
Fourth Year

Validation and Finalization

  • Validation of the proposed methods on real processes and benchmarks to prove the robustness of these methods against real data
  • This can be carried out in collaboration with Tunisian industry or even international industry

Cooperation and Partnership

Partners

  • Sultan Qaboos University, Muscat, Oman

Expected Results (publications, patents, theses, habilitations, ...)

Scientific production

Expected results include:

  • Scientific publications in international journals
  • Communications in international conferences
  • Doctoral theses
  • Possibility of filing patents for the developed innovations

Socio-economic Benefits of the Project

Benefits for industry

The socio-economic benefits of the project include:

  • Providing the concerned industry with expertise in modeling industrial systems
  • Providing the concerned industry with expertise in monitoring large and complex systems
  • Providing the concerned industry with expertise in implementing artificial intelligence monitoring algorithms on embedded systems