DEVELOPMENT OF SENSOR NODE, MOBILE ROBOT AND UNMANNED AERIAL VEHICLE BASED VOLCANO ERALY DETECTION SYSTEM FOR TYPE-A VOLCANO IN JAVA ISLAND INDONESIA

• Main Author: Evita, Maria
• Format: Dissertations
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/78467
• Institution: Institut Teknologi Bandung

A combination of sensor node, mobile robot and UAV for volcano monitoring has been developed, hence the eruption could be early detected to minimize the after-effects. The sensor node consists of TGS-2602 (SO2), MG-811 (CO2), DHT-11 (temperature), ADXL-345 (seismicity) and MPU6050 (landslides) sensors; solar panel controller (power supply), WiFi ESP-8266 and LoRa-02 (data communication) and Raspberry Pi for data and volcano warning status processing using fuzzy logic. The solar cell could generate 18.81 Volt voltage and 0.23 Ampere current in laboratory testing. The temperature sensor has been calibrated for until 51oC, while the gas sensor could properly detect the gas until 8 ppm. After FFT analysis the seismic sensor detected almost zero frequency with 0.168 m/s2 , 0.0168 m/s2 and 1.125 m/s2 X, Y and Z axis zero offsets respectively. The landslide sensor had a maximum of about 4o land slope error. Moreover, only a 4% (maximum) PER is shown by the data communication. Furthermore, a simulation to decide the volcano status for the volcano in fuzzy control where SO2 4 ppm, CO2 500 ppm, temperature 36oC, seismicity 2 Hz and landslides 0o were the inputs for the controller resulted in Normal to Waspada (Alert) warning status. The system was then tested in Kelud, Tangkuban Parahu and Guntur Volcanoes (type-A stratovolcanoes). Sulfur dioxide was detected under 1 ppm, while CO2 was 0,306 – 0, 3 ppm (only measured in Guntur Volcano). Moreover, only a 1 Hz (0,32 gRMS) seismic wave was detected in Kelud, whereas nothing in Tangkuban Parahu and Guntur. Meanwhile, the temperature was in the range of 23 – 35 oC. These sensor data generated a Normal warning status for all the volcanoes. Hereinafter, the robot with the same monitoring sensor, driven by 24-volt DC motors (PG 28) of skid-steering mechanism and PID controller in STM-32 Nucleo F446RE microcontroller connected to Raspberry Pi 3 for velocity and ROS-RTOS operating systems. PID Controller has been successfully applied with average deviations of 2.5% (left motor), and 2.75% (right motor). The robot could reach the waypoints generated by the highlevel planner and a real trajectory was depicted followed by the low-level navigation layer while avoiding the obstacles in MATLAB performing near-optimal navigation algorithm. This algorithm then was tested on a flat surface of the laboratory floor and a grassy surface of a field. The robot successfully reached the destination while avoiding 1 and 2 obstacles automatically. The same result also was shown in field testing at Tangkuban Parahu on a rocky and sloppy uneven terrain. The obstacles had been detected beforehand by performing the YOLOv5s algorithm for objects (trees, persons, rocks and stairs) in the early step for four models. the best model (batch = 16, epochs = 500) resulted in mAP_0.5 = 25.7% and mAP_0.5:0.95 = 12.3%. A further training then was applied for trees and rocks to be implemented in Tangkuban Parahu volcano. The dataset was trained in 3 epochs (100, 300 and 500) and 16 batches of YOLOv5s. The last variant yielded the best result of 63.4% mAP_0.5 and 40.4% mAP_0.5:0.95 with almost zero loss. This model then was implemented in the Raspberry Pi3 to detect trees and rocks captured by the camera in Tangkuban Parahu Volcano. Most trees and rocks were successfully detected shown by 90.9% recall, 79.9% precision and 91.5% accuracy. Moreover, data from monitoring sensor in mobile robot decided Normal status of Tangkuban Parahu, Guntur, Merapi and Galunggung volcanoes in the fuzzy controller where SO2 (0 – 0,639 ppm), CO2 (0 – 0,219 ppm), seismicity frequencies (0 - 2 Hz at 0,05 gRMS), temperatures (18,65 – 31,25 oC) and landslides (0,010 – 0,085 o /s) in normal ranges. The maps generated from UAV (DJI Phantom 4 Pro) images processed by 3D survey, Pix4D and Agisoft softwares resulted in a few centimeters of resolutions, and total errors between 0 – 8 meters for 20,000 – 545,000 m2 monitored areas including laboratory and field areas where the vegetative and non-vegetative areas could be distinguished. Hereinafter, the RMS errors implied a few centimeters of horizontal accuracies in 95% of the confidence level