PulseGuard: 1D-CNN Denoising Autoencoder for Edge Devices
An edge model that cleans noisy sensor signals using a 1D-CNN denoising autoencoder trained on aggressive synthetic noise profiles.
+18 dB
Signal-to-Noise Improvement
< 2 MB
Model Size
< 20 ms
Inference Latency
Wearable and remote health devices collect noisy biosignals. PulseGuard was developed to clean those signals at the edge before they are used for downstream analysis, reducing cloud round-trips and improving real-time responsiveness.
Why was a 1D-CNN denoising autoencoder chosen?
A 1D convolutional autoencoder learns to reconstruct clean sensor windows from corrupted inputs. The encoder compresses the signal into a latent representation, and the decoder reconstructs a denoised version. This design is compact, fast, and easier to quantize for edge hardware than larger transformer-based models.
How was the model trained on synthetic noise?
We generated aggressive synthetic noise profiles that mimic real-world artifacts: motion blur, baseline wander, power-line interference, and sensor dropout. By training on these controlled distortions, the model generalizes to unseen real-world noise without needing large labeled clinical datasets.
What are the edge deployment results?
The final model is under two megabytes and runs in less than twenty milliseconds on edge hardware. Signal quality improved by approximately eighteen decibels, making downstream anomaly detection and diagnosis more reliable.
How does PulseGuard protect privacy?
Because denoising runs locally, raw sensor data never needs to leave the device. Only cleaned, summarized features are transmitted when required, which reduces bandwidth and lowers exposure to data breaches.