Industrial automation is evolving rapidly, and increasingly adopted by manufacturers Industry 4.0 Practices that become smarter and more efficient. At the same time, advances in hardware have made computing power more accessible and compact.
I developed a using both trends Real-time Volt Detection and Counting System Using Nvidia Jetson Nano. The solution enables industrial processes to automate conversions not only for developing robust machine learning models, but also for optimization and deployment on edge devices like Jetson Nano.
ProcedureDataset Collection: The dataset was collected using mobile and USB cameras and captured images of industrial bolts located in different locations. Bolts have been placed Different angles And below Multiple lighting Conditions that simulate actual industrial environments. this Diversity It ensures that the model can be properly generalized and accurately predicted, even when external conditions vary, such as low light or unusual viewing angles.
Data Annotation: I used it because the purpose is to detect and count industrial bolts Edge Impulse Application Creates annotations that mark each volt in the image. Annotations were made manually for higher accuracy using a rectangular box around each bolt. Proper labeling is important for training object detection models that can identify and distinguish multiple bolts in a single frame.
Feature Extraction: Function extraction was performed to convert raw image data into meaningful functions. This step includes normalization, resizing and transformations RGB Channel Depending on the model pipeline you choose. These features are important inputs for training object detection models.
ML Model: I used it to train the model MobileNETV2 SSD FPN-LITE 320×320a highly efficient object detection architecture tailored to edge devices. This model architecture is ideal for detecting small objects like industrial bolts in complex scenes.
- mobileNetv2 It acts as a feature extractor and provides a lightweight, fast backbone.
- SSD (Single-Shot Multi-Box Detector) Enables real-time object detection by predicting bounding boxes and class scores on a single forward pass.
- fpn-lite (Functional Pyramid Network Light) Improves the discovery of small objects by aggregating features from multiple levels of the network.
- Input image resolution of 320×320 Balance detection accuracy and edge expansion calculation efficiency.
Training structure:
- Number of training cycles: twenty five
- Learning rate: 0.15
- Batch size: 32
- Training Processor: CPU
- Verification set size: 20% of the total dataset
- How to split data: Randomized splitting
- Quantization: INT8
After training, this model was profiled to run efficiently on edge hardware using INT8 quantization, further reducing memory footprint and inference latency. This configuration allowed us to create compact yet accurate models that could be run in real time on resource-constrained devices like the Nvidia Jetson Nano.
ML Model Conversion: After successful training, the model was converted to a format compatible with edge devices. The trained ML model is converted to Tensorflow Lite and custom C++ library formats, depending on the deployment goal. In this case, the model was converted to a Tensortort compatible format for optimized performance with Jetson nano.
ML Model Deployment on Nvidia Jetson Nano: The deployment environment was prepared by flashing Ubuntu 18.04 Using Nvidia Jetson Nano Nvidia SDK Manager. This ensures a stable and compatible base system for inference of AI models.
After OS setup, Python 3.6 Installed and configured to support runtime requirements for the model. Transformed models have been integrated with custom C and Python Scripts To enable real-time inference directly on the device.
The model was optimized using TensortNvidia's high-performance deep learning inference library. This allowed the Jetson Nano to utilize onboard GPUs for acceleration processing. A lightweight inference application has been created to handle camera input, image preprocessing, model execution, and output visualization all in real time.
This deployment setup ensures that the system runs autonomously at the edge, providing low latency detection and volt counting without the need for an internet connection.
ML Model Inference: Once deployed, the model performs real-time inference on video frames. Detects and counts bolts present in the field of view. The output includes the volt count and its location. This can be further used for quality control, automated inspection, or operational analysis in industrial environments.
