DNN-based Digital Twin Framework of a DC-DC Buck Converter using Spider Monkey Optimization Algorithm

Component ageing is a critical concern in power electronic converter systems (PECSs). It directly impacts the reliability, performance, and operational lifespan of converters used across diverse applications, including electric vehicles (EVs), renewable energy systems (RESs) and industrial automation. Therefore, understanding and monitoring component ageing is crucial for developing robust converters and achieving long-term system reliability. This paper proposes a data-driven digital twin (DT) framework for DC-DC buck converters, integrating deep neural network (DNN) with the spider monkey optimization (SMO) algorithm to monitor and predict component degradation. Utilizing a low-power prototype testbed along with empirical and synthetic datasets, the SMO+DNN approach achieves the global optimum in 95% of trials, requires 33% fewer iterations, and results in 80% fewer parameter constraint violations compared to traditional methods. The DNN model achieves $R^2$ scores above 0.998 for all key degradation parameters and accurately forecasts time to failure ($t_{failure}$). In addition, SMO-tuned degradation profile improves the converter’s performance by reducing voltage ripple by 20-25% and inductor current ripple by 15-20%.

💡 Research Summary

The paper presents a data‑driven digital twin (DT) framework for a DC‑DC buck converter that couples a deep neural network (DNN) with the Spider Monkey Optimization (SMO) algorithm to monitor and predict component ageing. A low‑power prototype serves as the physical “multiphysics mechanism model” (MMM); its switching behavior, voltage, current, and temperature signals are captured with a high‑speed DAQ system. From these measurements a synthetic dataset of 10 000 samples is generated, covering realistic operating ranges and incorporating Gaussian sensor noise. Each sample consists of input variables (Vin, Iin, V_D, I_D, V_L, I_L, V_C, I_C, V_o) and output degradation parameters (L, C, r_L, r_C, r_ds‑ON, t_failure) defined by physics‑informed linear decay/growth equations.

SMO, a swarm‑intelligence meta‑heuristic, is employed offline to calibrate the Simulink‑based digital model (DM). The algorithm organizes candidate solutions into groups with local and global leaders, updating positions via a weighted combination of attraction to the local leader, random exploration, and a perturbation factor. The cost function is an integral‑squared‑time‑weighted‑squared‑error (ISTSE) that penalizes long‑term discrepancies between simulated and measured output voltage and inductor current. SMO converges in fewer iterations than conventional particle swarm optimization, achieving the global optimum in 95 % of trials, reducing iteration count by 33 %, and cutting parameter‑constraint violations by 80 %.

The optimized parameters (L*, C*, r*_L, r*_C, r*_ds‑ON) become the feature set for a five‑layer fully‑connected DNN regression model built with TensorFlow/Keras. Trained on a 70 %/15 %/15 % split (training/validation/testing) using the Adam optimizer and mean‑squared‑error loss, the DNN attains R² scores above 0.998 for all degradation variables and accurately forecasts the remaining useful life (t_failure).

Experimental validation on the physical buck converter under an accelerated thermal cycling regime (≈19 h equivalent ageing) confirms the DT’s predictive capability. The SMO‑tuned degradation profile reduces voltage ripple by 20‑25 % and inductor‑current ripple by 15‑20 % compared with the unoptimized baseline.

Overall, the study demonstrates that integrating SMO for robust parameter identification with a high‑capacity DNN for nonlinear regression yields a digital twin capable of near‑real‑time health monitoring and performance optimization of power electronic converters. Limitations include the lack of detailed hyper‑parameter tuning procedures for SMO and DNN, and the focus on a low‑power prototype rather than high‑voltage/high‑current industrial converters. Future work is suggested to extend the framework to larger power stages, incorporate online learning for adaptive DT updates, and explore multi‑objective optimization that balances efficiency, thermal stress, and electromagnetic interference.