Samuela Rokocakau

Mr. Samuela Rokocakau

orcid

Research Focus & Metrics

Academic Career Length

9Years (2018-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
4 June 2024 in IEEE Transactions on Industry Applications
Verified
Journal Article
Closed access
Samuela RokocakauSamuela Rokocakau
Samuela Rokocakau
JRJacopo Riccio
Jacopo Riccio
GTGiulia Tresca
Giulia Tresca
Rahul Ranjeev KumarRahul Ranjeev Kumar
Rahul Ranjeev Kumar
GCGiansalvo Cirrincione
Giansalvo Cirrincione
Maurizio CirrincioneMaurizio Cirrincione
Maurizio Cirrincione
This paper introduces a novel fault diagnosis approach for two-level Voltage Source Inverters in Motor Drives, using Shallow Neural Networks. A key feature of this study is the implementation of three different Classifiers: two are designed to both detect and diagnose faults, while the third is specialized in fault detection and localization. The strategic adoption of Shallow Neural Networks plays a pivotal role in substantially reducing computational demands while enabling the direct utilization of raw, normalized data obtained from both simulation and experimental activities. The research includes a comparative analysis of three distinct shallow neural network architectures against other machine learning methods to underscore their effectiveness. Specifically, the use of shallow Long Short Term Memories networks is a pioneering effort in the field of power converter fault detection, especially for their reduced memory requirement (one hidden layer with 64 units). The final results show that the average time per observation for the localization classifier(7 Class-Classifier) across the entire test set is equal to 39 $\mu$ s. This efficiency is coupled with promising metrics, including a precision, recall and F1-score of 94.50%, 94.24%, and 94.35%, respectively.

20 October 2024
Verified
Conference Paper
Closed access
Samuela RokocakauSamuela Rokocakau
Samuela Rokocakau
GTGiulia Tresca
Giulia Tresca
BMBehrouz Mohammadzadeh
Behrouz Mohammadzadeh
PZPericle Zanchetta
Pericle Zanchetta
GCGiansalvo Cirrincione
Giansalvo Cirrincione
MCMaurizio Cirrincione
Maurizio Cirrincione
In this work, a novel Hybrid 1-D Convolutional Transformer Neural Network is utilized for diagnosing faults in a Cascaded H-Bridge PMSM drive system. This innovative method exhibits high precision in open-circuit switch fault detection and localization for multi-class classification. The neural network model is used to assess the overall health status, locate faulty modules, faulty phase leg localization and identify individual switch faults in two different Cascaded H-Bridge topologies: one H-Bridge per phase (Topology I) and three H-Bridges per phase (Topology II) PMSM drive systems. The study showcases the prediction accuracy, reliability, and robustness of the neural network models for both topologies. In Topology I, the model achieved accuracy rates above 99% in its fault detection and localization tasks, with an average prediction time of around 1ms per sample for all considered diagnostic scenarios. Despite Topology II being more complex. the model had prediction accuracy rates exceeding 98% and an average prediction time between 1.5ms and 9ms per sample for all its considered diagnostic scenarios. Both drive system topologies resulted in minimal memory usage of approximately 2MB and 2.1MB, respectively.

20 October 2024
Verified
Conference Paper
Closed access
Samuela RokocakauSamuela Rokocakau
Samuela Rokocakau
GTGiulia Tresca
Giulia Tresca
BMBehrouz Mohammadzadeh
Behrouz Mohammadzadeh
PZPericle Zanchetta
Pericle Zanchetta
GCGiansalvo Cirrincione
Giansalvo Cirrincione
MCMaurizio Cirrincione
Maurizio Cirrincione
In this work, the 1-D Convolutional Transformer Hybrid Neural Network is compared to other deep learning and non-neural techniques for Open Switch Fault Detection and Localization in a Cascaded H-Bridge Converter. The novel approach outperforms others in multiclass classification switch fault detection and localization. Two H-Bridge topologies are analyzed, with the hybrid neural network achieving high prediction accuracy in both. Key metrics like prediction accuracy, speed and memory size highlight the architecture’s advantages. The models detect overall system health, locate faulty phase modules and identify single switch faults, reaching accuracy rates of up to 99% and 98% for Topologies I and II, respectively. The model size is approximately 2MB for both topologies, with average prediction times of 1ms for Topology I. Topology II’s 2-class scenario for overall health status has an average prediction time of 9ms, while other scenarios take 1.5ms. The advancement of AI-based technologies can help improve the reliability and robustness of inverter systems by understanding and implementing strategies discussed in this paper.

1 September 2023
Verified
Conference Paper
Closed access
Samuela RokocakauSamuela Rokocakau
Samuela Rokocakau
GTGiulia Tresca
Giulia Tresca
GCGiansalvo Cirrincione
Giansalvo Cirrincione
PZPericle Zanchetta
Pericle Zanchetta
RKRahul Kumar
Rahul Kumar
LFLucia Frosini
Lucia Frosini
Ensuring reliability and safety in multilevel power converter topologies requires effective fault diagnostics, especially considering the greater number of components involved compared to traditional converters. Recent research has recognized the potential of Artificial Neural Networks (ANN) in enhancing diagnostic strategies. In this context, this study proposes an innovative approach for detecting devices faults in multilevel converters using Convolutional Neural Networks (CNN) through 2D image-based classification. The simulation results demonstrate the efficacy of this method, enabling more dependable and efficient multilevel power electronic systems.

27 July 2025 in Energies
Verified
Journal Article
Open access
PDF file
AVAndrea Volpini
Andrea Volpini
Samuela RokocakauSamuela Rokocakau
Samuela Rokocakau
GTGiulia Tresca
Giulia Tresca
FGFilippo Gemma
Filippo Gemma
PZPericle Zanchetta
Pericle Zanchetta
With the increasing integration of renewable energy sources (RESs) into power systems, batteries are playing a critical role in ensuring grid reliability and flexibility. Among them, vanadium redox flow batteries (VRFBs) have emerged as a promising solution for large-scale storage due to their long cycle life, scalability, and deep discharge capability. However, achieving optimal control and system-level integration of VRFBs requires accurate, real-time modeling and parameter estimation, challenging tasks given the multi-physics nature and time-varying dynamics of such systems. This paper presents a lightweight physics-informed neural network (PINN) framework tailored for VRFBs, which directly embeds the discrete-time state-space dynamics into the network architecture. The model simultaneously predicts terminal voltage and estimates five discrete-time physical parameters associated with RC dynamics and internal resistance, while avoiding hidden layers to enhance interpretability and computational efficiency. The resulting PINN model is integrated into a modulated model predictive control (MMPC) scheme for a dual-stage DC-AC converter interfacing the VRFB with low-voltage AC grids. Simulation and hardware-in-the-loop results demonstrate that adaptive tuning of the PINN-estimated parameters enables precise tracking of battery parameter variations, thereby improving the robustness and performance of the MMPC controller under varying operating conditions.

1 October 2018
Verified
Conference Paper
Closed access
HMH. K. Mudaliar
H. K. Mudaliar
DKD. M. Kumar
D. M. Kumar
Samuela RokocakauSamuela Rokocakau
Samuela Rokocakau
KRK. Ram
K. Ram
MAM. Assaf
M. Assaf
AMAli MOHAMMADI
Ali MOHAMMADI