🤗 Welcome to the Embedded System to Anomaly Detection on GitHUB!

Some infomation about us 😎
🙋‍♀️HUST Industry-Academia Collaboration Project betwwen Signal Processing and Radio Communication Laboratory, HUST and SI Synergy Technology Company Ltd, Japan
🧙 Instructor: Ph.D Huy-Dung Han
👩‍💻 Research Orientation:

• Amomaly Detection in Sound for Industrial, Smart Traffic and etc.
• Develop the Embedded Linux Platform to make comprehensive IoT system includes: Clouds Architecture, Embedded Linux OS by Yocto Project, Anomaly Detection Algorithm, etc.
• The target machines: Jetson Nano, Jetson Xavier, Jetson Orin Nano, Rasberry Pi 4, ...

The documentation ESCA provides on github includes:

• ESCA Information
  • Environmental sound collection and analysis embedded application
• ESCA Version: 2024

To run project:

Jetson Orin Nano Optimization

The inference module now supports TensorRT. Enable this by setting "USE_TENSORRT": true in python_ai/config/default.json and specify the TRT_MODEL_PATH where the optimized model will be stored. On the first run the model will be converted using TF‑TRT; subsequent runs load the optimized version directly. If conversion is not possible the code automatically falls back to TensorFlow execution.

Performance Evaluation

When running python_ai/inference/inference.py, the processing time of each cycle is written to REALTIME.LOG_PATH/processing_time.log (as configured in python_ai/config/default.json). Compare these timings before and after enabling TensorRT on Jetson to quantify the speed‑up for your report.

Data Storage

Set the environment variable ESCA_DATA_PATH to control where files such as config.json, health.log and processing_time.log are stored. If not set, all logs and configuration will be kept under ~/data_storage.

GPU Gammatone Pre-processing

Set "USE_TF_GTGRAM": true in the same configuration to offload the gammatone spectrogram calculation to TensorFlow running on GPU. This replaces the slower Python implementation, further improving throughput on Jetson.

Architecture Overview

The project follows a Layered + Hexagonal architecture:

1. UI Layer – implemented in QML files under ESCA/qml, strictly communicates with the application core via exposed properties and signals.
2. Core Orchestrator – C++ controllers located in ESCA/src/modules coordinate application logic and launch Python processes. They interact with the domain through interfaces such as InferenceEngine.
3. Domain Modules – reusable processing logic (e.g. training and inference) implemented mostly in python_ai and a few C++ classes.
4. External Adapters – infrastructure code like TensorFlow, file system access or device drivers.

To keep the layers independent, runtime paths are now resolved relative to the executable rather than hard coded. The configuration loader in python_ai/config/config_manager.py also accepts an ESCA_CONFIG_PATH environment variable for custom setups.