cs.CVOct 25, 2026

Real-Time Conveyor Belt Defect Detection via Edge TPU-Accelerated YOLOv11 on Raspberry Pi 5

Authors: Muhammad Nur Bijak Bestari, Mulki Indana Zulfa

Organizations: Department of Electrical Engineering, Faculty of Engineering, Universitas Jenderal Soedirman, Purwokerto 53123, Central Java, Indonesia

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Conveyor belts are vital arteries in continuous bulk material handling, particularly in heavy process industries such as cement manufacturing. Undetected structural anomalies (longitudinal tears, holes, degraded patch repairs, and impact punctures) cause catastrophic belt snapping, costly downtime, and severe safety hazards. While deep learning offers high accuracy, existing models rely on power-hungry workstation GPUs, hindering practical deployment at remote industrial edge points. In this paper, we propose, implement, and validate an end-to-end, low-power edge AI vision system for real-time conveyor belt defect detection deployed on a Raspberry Pi 5 accelerated by a Google Coral Edge TPU coprocessor. The system leverages an optimized YOLOv11 nano (YOLOv11n) architecture trained on a multi-source industrial corpus exceeding 3,000 augmented samples from an operational cement plant in West Java, open-source datasets, and automated captures. The model undergoes post-training full integer quantization (INT8) compiled for systolic tensor execution. Experimental results demonstrate robust detection capability, achieving precision of 0.742, recall of 0.730, and [email protected] of 0.764 across four defect classes. Deployed on-site in an IP-rated enclosure interfaced with an RTSP PoE industrial camera, the system achieves stable throughput of 30-40 FPS at 320x320 resolution with an end-to-end latency of 70-90 ms. Operational profiling under harsh clinker dust confirms thermal equilibrium at 55 +/- 5 deg C on the host CPU and ~65 deg C on the Edge TPU with zero frame loss. This demonstrates that edge AI bridges the gap between high-accuracy computer vision and cost-effective predictive maintenance in continuous manufacturing.

Figures & tables

Figure 1: End-to-end research methodology flowchart, delineating data acquisition, annotation conversion, YOLOv11 model training, INT8 Edge TPU compilation, and embedded industrial RTSP deployment.
Figure 2: Three-tier edge computing system architecture, depicting the RTSP IP camera feed, PoE injector, Raspberry Pi 5 host, Google Coral Edge TPU coprocessor, and real-time monitoring display.
Figure 3: Industrial hardware deployment: (a) on-site IP-rated protective metal enclosure containing the Raspberry Pi 5, Edge TPU, and PoE supply at the cement facility; (b) laboratory hardware testing bench.
Figure 4: Representative raw conveyor belt samples: (a) direct in-situ field inspection captures at the cement plant; (b) public industrial dataset samples; (c) automated frames from the stream sampler.
Figure 5: Corresponding annotated training samples showing ground-truth bounding box delineations across classes: (a) field inspection defects; (b) transformed public annotations; (c) sampled conveyor damage.
Figure 6: Training and validation performance progression of the YOLOv11 model across 100 epochs: (a) Precision converging to 0.742; (b) Recall converging to 0.730; (c) [email protected] reaching 0.764; (d) [email protected]:0.95 reaching 0.553.
Figure 7: Normalized confusion matrix of the trained YOLOv11 model across the four defect categories and background, demonstrating robust true-positive localization with minimal inter-class confusion.
Model FormatDevice280 px320 px640 pxOperational Behavior & Observations
PyTorch ( .pt )RPi 5 CPU∼\sim 20 FPS∼\sim 10 FPS3–7 FPSHigh CPU load; at 640 px bounding boxes appear with noticeable visual lag.
ONNX (CPU)RPi 5 CPU——∼\sim 20 FPSEvaluated on CPU runtime; steady frame rate but elevated core temperatures.
TFLite + Edge TPUCoral USB—30–40 FPS∼\sim 10 FPSHighest throughput achieved at 320 px; stable real-time Patch Work and Tear detection.
Table 1: Comparative Inference Throughput (Frames Per Second) on Raspberry Pi 5 Across Runtime Formats and Spatial Resolutions.
Figure 8: Field deployment results: (a) system graphical user interface displaying real-time video feed and defect count; (b) positive real-time detection of a Patch Work anomaly on the running industrial conveyor.
ParameterMeasured ValueOperational Context & Stability Status
Maximum Edge TPU FPS30–40 FPSMeasured using RTSP sub-stream at 320×320320\times 320 px; smooth defect tracking.
Minimum PyTorch FPS3–7 FPSHost CPU at 640×640640\times 640 px; inadequate for fast conveyor speeds.
ONNX CPU Frame Rate∼\sim 20 FPSGeneral benchmark baseline on host CPU.
End-to-End Latency70–90 ms\mathbf{70\text{--}90\text{ ms}}Cumulative latency from camera sensor exposure, network transport, preprocessing, TPU inference, to display render.
Raspberry Pi 5 Core Temp.55±5∘C55\pm 5^{\circ}\text{C}Thermally stable within enclosed metallic panel without active refrigeration.
Google Coral USB Temp.∼65∘C\sim 65^{\circ}\text{C}Consistent thermal equilibrium under continuous 100% TPU utilization.
Table 2: System Operational and Environmental Parameters Recorded During Sustained Industrial Field Testing.
More in this topic

Industrial Anomaly Detection

Industrial anomaly detection finds defective products and abnormal equipment behavior from visual, sensor, audio, and other production data.

See all 260 papers

Explore similar work

CardsList
Jun 3, 2026

Real-Time Industrial Defect Detection on Edge Hardware Using Fine-Tuned YOLOv8: A Systematic Benchmark on the NEU Surface Defect Database and MVTec AD with Automotive & Battery Manufacturing Extensions

Automated surface defect detection is critical for ensuring rigorous quality control in high-speed manufacturing environments. While deep learning models offer remarkable accuracy, deploying them on resource-constrained edge hardware without introducing significant latency remains a persistent challenge. This paper presents Industrial-YOLO, an edge-optimized framework built upon a fine-tuned YOLOv8 architecture specifically engineered for real-time industrial defect detection. We conduct a systematic benchmark utilizing the NEU surface defect database for steel sheets and the MVTec AD dataset, supplemented with custom automotive manufacturing extensions representing real-world structural anomalies (scratches, pits, and inclusions). To bridge the gap between algorithmic complexity and edge hardware constraints, target-specific optimizations are introduced via TensorRT and OpenVINO acceleration engines. Experimental results demonstrate that Industrial-YOLO achieves a high-velocity inference speed exceeding 120 FPS on the NVIDIA Jetson Orin platform while maintaining an exceptional mean Average Precision (mAP) of 98.5%. The proposed framework showcases highly robust, zero-latency performance when deployed directly onto an active automotive assembly line, offering a scalable blueprint for next-generation automated optical inspection (AOI) systems.

Emmanuel Ezeji Somtochukwu, Nitesh Rijal

Zema AI Labs, Abuja, Nigeria

Edge InferenceIndustrial InspectionObject DetectionReal-Time Object Detection
AskPDF
Mar 17, 2026

TinyGLASS: Real-Time Self-Supervised In-Sensor Anomaly Detection

Anomaly detection plays a key role in industrial quality control, where defects must be identified despite the scarcity of labeled faulty samples. Recent self-supervised approaches, such as GLASS, learn normal visual patterns using only defect-free data and have shown strong performance on industrial benchmarks. However, their computational requirements limit their deployment on resource-constrained edge platforms, especially within in-sensor processing architectures. This work introduces TinyGLASS, a lightweight adaptation of the GLASS framework designed for real-time edge and in-sensor anomaly detection. The proposed architecture replaces the original WideResNet-50 backbone with a compact ResNet-18 and introduces deployment-based modifications that enable static graph tracing and INT8 quantization. We evaluate the proposed approach on the Sony IMX500 intelligent vision sensor, exploiting the in-sensor processor using the Sony Model Compression Toolkit. In addition to evaluating performance on the MVTec-AD benchmark, we investigate robustness to contaminated training data and introduce a custom industrial dataset, named MMS Dataset, for cross-device evaluation. Experimental results show that TinyGLASS achieves 8.6x parameter compression while maintaining competitive detection performance, reaching 94.2% image-level AUROC on MVTec-AD and operating at 20 FPS within the 8 MB memory constraints of the IMX500 platform. System profiling showcases low power consumption (4.0 mJ per inference), real-time end-to-end throughput (20 FPS), and high energy efficiency (470 GMAC/J). Furthermore, the model demonstrates stable performance under moderate levels of training data contamination.

Pietro Bonazzi, Rafael Sutter, Luigi Capogrosso +2

ETH Zurich, Zurich, Switzerland · Interdisciplinary Transformation University of Austria, Linz, Austria · Swiss Engineering Partners AG, Zurich, Switzerland

Anomaly DetectionIndustrial Anomaly DetectionUnsupervised Anomaly DetectionOnline Anomaly DetectionEdge Inference
AskPDF
Sep 25, 2024

A Comprehensive Evaluation of Deep Learning Object Detection Models on Heterogeneous Edge Devices

Modern applications such as autonomous vehicles, intelligent surveillance, and smart city systems increasingly require object detection on resource-constrained edge devices. Yet, there is still limited understanding of how different object detection models behave across heterogeneous edge devices and under varying scene complexity. In this paper, we benchmark YOLOv8 (Nano, Small, Medium), EfficientDet Lite (Lite0, Lite1, Lite2), and SSD (SSD MobileNet V1, SSDLite MobileDet) on Raspberry Pi 3, 4, 5 with/without Coral TPU accelerators, Raspberry Pi 5 with AI HAT+, Jetson Nano, and Jetson Orin Nano. We evaluate energy consumption, inference time, and accuracy, and further examine how accuracy changes with the number of objects in the input image. The results reveal clear trade-offs among accuracy, latency, and energy efficiency across model-device combinations. SSD MobileNet V1 achieves the lowest latency and energy consumption but the lowest accuracy, whereas YOLOv8 Medium achieves the highest accuracy at higher computational cost. TPU-based Raspberry Pi devices improve the efficiency of SSD and EfficientDet Lite while reducing YOLOv8 accuracy. Orin Nano offers the most favorable overall balance across most model families. The object-count-based analysis further shows that models achieve more similar accuracy on simpler images, while the accuracy gap widens as scene complexity increases.

Daghash K. Alqahtani, Muhammad Aamir Cheema, Maria A. Rodriguez +1

DisNet Lab, School of Computing and Information Systems, The University of Melbourne, Melbourne, VIC, Australia · Monash University, Melbourne, VIC, Australia

Energy-Efficient MLEdge InferenceEfficient Object DetectionObject Detection
AskPDF