Overview & Objective
The objective of this project was to establish a highly reliable, low-latency machine learning pipeline capable of operating entirely on edge hardware. By bypassing cloud processing, the system provides real-time alerts for impending mechanical degradation, solving critical bandwidth and security constraints in industrial environments.
Technical Implementation
Feature Engineering Pipeline
Built preprocessing and feature extraction workflows in Python, computing rolling-window statistics, fast Fourier transforms (FFT) for vibration frequency analysis, and temperature gradient shifts to capture impending hardware degradation.
Model Exploration & Benchmarking
Evaluated several predictive architectures—including deep Autoencoders, Multi-Layer Perceptrons (MLPs), and regularized Logistic Regression baselines—measuring inference latency, memory footprint, and reconstruction error metrics (MSE).
Hyperparameter Optimization with Optuna
Automated tuning routines using Optuna to discover optimal latent dimensionalities, activation functions, learning rates, and pruning strategies, balancing detection precision against the compute constraints of edge hardware.
Edge Deployment Readiness
Structured lightweight, low-footprint inference modules designed for minimal latency execution on constrained single-board and IoT gateway microprocessors.
Challenges & Problem Solving
Edge Compute Constraints
Early iterations of the Autoencoder were too large to run efficiently on edge hardware, leading to high inference latency. I used Optuna to prune the network, aggressively reducing the latent space dimensionality and hidden layer count until the model maintained precision without exceeding the memory footprint.
Sensor Noise
Raw vibration data contained high-frequency mechanical noise that triggered false positive anomalies. I implemented a digital low-pass filter and utilized Fast Fourier Transforms (FFT) in the preprocessing pipeline to isolate the dominant frequencies before feeding the data into the model.
Results & Future Improvements
Results
Successfully deployed a lightweight anomaly detection model that identifies thermodynamic and vibration anomalies with high precision, executing inferences in under 50ms on constrained hardware.
Future Improvements
I want to export the PyTorch models to ONNX or TensorRT format to leverage hardware-accelerated inference in C++, further reducing latency and power consumption.