Wahyu Sapto Aji creator author. text my 2025 2025 monographic eng xv, 155 pages : illustrations ; 30 cm. + 1 CD-ROM. Unstripped Sterilised Bunches (USB) represent a critical yet persistently underestimated source of oil loss within palm oil milling operations. Industry standards, notably those established by the Malaysian Palm Oil Board (MPOB), traditionally estimate USBrelated oil losses at approximately 0.05% per tonne of Fresh Fruit Bunch (FFB) processed. However, recent empirical audits conducted across multiple Malaysian mills between 2020 and 2023 reveal a more concerning reality: actual losses consistently range between 0.05% and 0.1% per tonne FFB. For a medium-sized mill operating at 60 tonnes FFB per hour, this translates to an annual economic loss of 3,000 to 6,000 tonnes of recoverable palm oil. This significant discrepancy underscores USB not merely as a minor inefficiency, but as a substantial source of economic wastage demanding precise monitoring and intervention. Current industry reliance on manual visual inspection of sterilised bunch conveyor belts for USB detection is fraught with limitations that impede accurate loss accounting. This method is inherently subjective, heavily dependent on the individual operator's expertise and vigilance. Crucially, manual inspections struggle with low resolution, frequently missing smaller USBs or those partially obscured by EFB (Empty Sterilised Fruit Bunch). The impracticality of continuous human monitoring on high-speed conveyors (typically moving at 2–3 m/s) inherently limits scalability. Conventional USB detection through manual visual inspections of conveyor belts suffers from subjectivity, fatigue-induced errors, and poor scalability, creating an urgent need for automated solutions. The drive towards Industry 4.0 adoption within the palm oil sector underscores the urgent need for data-driven solutions to mitigate losses like those caused by USBs. A reliable automated USB detection system represents a crucial technological advancement. It would deliver realtime, objective, and auditable data on oil loss, enabling immediate process adjustments to optimize thresher performance and sterilisation efficiency. This study also fill the gap that previous studies have focused more on the detection of FFB, not USB after sterilisation and threshing processes This study addresses three critical research gaps: First, while deep learning shows promise for industrial object detection, existing architectures lack optimization for USB recognition. To resolve this limitation, the research develops an enhanced VGG16-based Faster R-CNN framework through Objective 1, integrating concatenate layers for multi-scale feature fusion and a Gaussian High-Pass Filter (GHPF) salient attention mechanism within the RoI head to prioritize USB texture patterns. Second, the scarcity of diverse training data from operational mills impedes model generalization, which Objective 2 addresses by employing Progressive Growing GANs (PGGAN) to synthesize USB images. Third, prior research neglects validation of USB detectors in dynamic production settings, prompting Objective 3 to quantitatively evaluate the framework’s efficacy using conveyor system footage. Validation was performed by comparing the proposed method for counting USB in video, compared to the manual method and with a USB counter based on the comparison method. The proposed method achieved 94.3% detection accuracy(surpassing benchmarks set by Hourglass, ResNet50, EfficientNet-SSD, and baseline VGG16 models)while video-based USB counting tests on USB conveyor footages attained 88.67% mean accuracy. These results demonstrate the first viable automated USB detection system, offering mills a scalable solution to reduce oil losses through synthetic data augmentation and attention-driven feature extraction. specialized Wahyu Sapto Aji Faculty of Electrical & Electronics Engineering Technology Thesis (Doctor of Philosophy) -- Universiti Malaysia Pahang - 2025 Include bibliographical reference Faculty of Electrical & Electronics Engineering Technology Dissertations Universities and colleges Dissertations Theses Dissertations THE0010444 (Local) UMPSA 260407 20260407144846.0 eng