Implementasi Two-Stage Detection YOLOv11 pada Sistem Robot Pembersih Panel Surya Otonom Berbasis Klasifikasi Tingkat Kekotoran
Muhammad Ridho Assidiqi, Dr. Ir. Atikah Surriani, S.T., M.Eng., IPM.
2026 | Tugas Akhir | D4 Teknologi Rekayasa Instrumentasi dan Kontrol
Akumulasi debu pada panel surya
menurunkan efisiensi konversi energi, sementara metode konvensional berbasis
sensor Light Dependent Resistor (LDR) atau sensor arus tidak dapat
membedakan kotoran dari gangguan cuaca, dan sistem berbasis cloud rentan gagal
koneksi di lokasi terpencil. Penelitian ini merancang robot pembersih panel
surya otonom yang memadukan deteksi visual two-stage YOLOv11, kendali Proportional-Integral-Derivative
(PID) closed-loop, dan komunikasi serial lokal agar mandiri pada
Pembangkit Listrik Tenaga Surya (PLTS) off-grid tanpa internet.
Sistem menggunakan arsitektur two-tier:
Raspberry Pi 5 menjalankan pipeline two-stage detection (deteksi panel
YOLOv11n, klasifikasi tingkat kekotoran YOLOv11n-cls ke empat kelas: bersih,
ringan, sedang, berat) yang dikonversi menjadi weighted score untuk
strategi pembersihan adaptif, sedangkan ESP32 mengeksekusi siklus pembersihan
melalui finite state machine berkendali PID dengan umpan balik rotary
encoder dan kompensasi feedforward. Kedua unit berkomunikasi via
UART kabel langsung, dan hasil pembersihan diverifikasi otomatis melalui
deteksi ulang.
Model deteksi panel mencapai
mAP@0.5 99,43% (precision 97,58%, recall 99,46%) dan model
klasifikasi mencapai akurasi 98,00% (450 sampel validasi), sementara validasi
lapangan (60 sampel) mencapai 96,67%. Inferensi two-stage 494,77 ms/frame
(2,02 FPS); kendali PID dengan feedforward menjaga overshoot di
bawah 10?n galat tunak dalam deadband 0,5 RPM meski beban gesek
bervariasi; komunikasi UART mencapai success rate 100?ngan latensi
rata-rata 52,49 ms. Sistem berhasil menjalankan siklus pembersihan adaptif tiga
level dengan verifikasi otomatis, layak diterapkan pada PLTS off-grid.
Dust accumulation on solar
panel surfaces reduces energy conversion efficiency, while conventional methods
based on LDR or current sensors cannot visually distinguish soiling from
weather disturbances, and cloud-based systems are prone to connection failures
in remote locations. This research designs an autonomous solar panel cleaning
robot integrating YOLOv11 two-stage visual detection, closed-loop PID control,
and local serial communication for independent operation on off-grid solar
power plants (PLTS) without relying on the internet.
The system uses a two-tier
architecture: a Raspberry Pi 5 runs the two-stage detection pipeline (panel
detection with YOLOv11n, dirt-level classification with YOLOv11n-cls into four
classes: clean, light, moderate, heavy) converted into a weighted score for
adaptive cleaning strategy, while an ESP32 executes the cleaning cycle via a
finite state machine with PID control using rotary encoder feedback and
feedforward compensation. The units communicate via direct wired UART, with cleaning
outcomes automatically verified through re-detection.
The panel detection model
achieves mAP@0.5 of 99.43% (precision 97.58%, recall 99.46%) and the
classification model reaches 98.00?curacy (450 validation samples), while
field validation (60 samples) obtained 96.67?curacy. Two-stage inference
takes 494.77 ms per frame (2.02 FPS); PID control with feedforward keeps
overshoot below 10% and steady-state error within a 0.5 RPM deadband despite
varying friction loads; UART communication achieves a 100% success rate with
52.49 ms average latency. The system successfully performs adaptive three-level
cleaning with automatic verification, suitable for off-grid solar power plant
installations.
Kata Kunci : panel surya, YOLOv11, two-stage detection, kendali PID, robot otonom, edge computing