Mohamed Abbes

The rapid adoption of photovoltaic (PV) technology has made it a keycomponent of the renewable energy transition. However, PV panels are af-fected by surface and structural anomalies, including soiling, snow cover,cracks, and electrical faults, which reduce performance and accelerate degra-dation. Since manual inspection is costly, labor-intensive, and intermittent,automated visual inspection has become an important research direction.Although convolutional neural networks (CNNs) and transfer learning haveimproved PV defect classification, their practical use remains limited by thelack of model interpretability. This paper proposes an explainable deep learning framework for multi-class PV surfacebcondition recognition that balances predictive performanceand interpretability. Three transfer learning architectures, namely VGG16,MobileNetV2, and ResNet50, are evaluated within a unified framework. Inaddition, Gradient-weighted Class Activation Mapping (Grad-CAM), Shap-ley Additive Explanations (SHAP), and Local Interpretable Model-AgnosticExplanations (LIME) are used to provide local and global explanations. Ex-planation stability and faithfulness are further quantified using cosine similarity and deletion/insertion area under the curve (AUC), respectively. Exper-imental results show that transfer learning effectively distinguishes six PVsurface classes, while the proposed explainable artificial intelligence (XAI)framework provides stable and faithful explanations that support human ver-ification in automated PV inspection systems.

This paper presents the design procedure of a high-performance sensorless control strategy for the widely used brushless DC (BLDC) motors. Generally, conventional sensorless techniques are based on detecting the zero-crossing instants (ZCP) of the back electromotives forces (back-EMFs) of the three phases. These methods, although widely adopted and marketed on an industrial level, involve many limitations such as filtering delays, difficulty to operate at low speeds and immunity against Electromagnetic Interferences (EMI). In this paper, the main objective is to develop a sensorless control technique integrally independent from the zero-crossing points of the back-EMFs. In the proposed method, a zero-delay adaptive filter was used to extract the fundamental and the quadrature components of the line-to-line voltage of the motor. Then, the Synchronous Reference Frame Phase Locked Loop (SRF-PLL) is used to estimate the electrical angle of phase-to-phase back-EMF along with the rotor speed. The conventional SRF-PLL was implemented using a second-order loop filter (type-3 PLL) in order to improve synchronization performances and for better rejection of voltage spikes induced in motor phases during commutations. The benefits of the control technique are brought to light through simulation results. An experimental prototype was designed to confirm the validity of the proposed method.

In drive systems based on Brushless DC motors (BLDC), sensorless detection of rotor position is preferred to achieve variable speed operation. In the literature, the most popular sensorless strategies are based on the zero-crossing detection of the back electromotive forces or back-EMFs Therefore, this work proposes a new sensorless control technique which ensures maximum torque generation and continuous synchronization with the rotor. In this technique, the rotor angular position is estimated based on a three-phase PLL. Compared with conventional techniques, the proposed method offers an increased immunity against Electromagnetic Interferences (EMI) which may cause erroneous switching sequences and unwanted oscillations of the electromagnetic torque.