Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Improving Robustness of Spectrogram Classifiers with Neural Stochastic Differential Equations

Joel Brogan, Olivera Kotevska, Anibely Torres, Sumit Jha, Mark Adams

TL;DR

This work explores the idea of Neural Stochastic Differential Equations (NSDE's) to improve the robustness of models trained to classify time series data and the effect of NSDE's on the explainability of outputs and test the effectiveness of these approaches by applying them to a non-intrusive load monitoring dataset.

Abstract

Signal analysis and classification is fraught with high levels of noise and perturbation. Computer-vision-based deep learning models applied to spectrograms have proven useful in the field of signal classification and detection; however, these methods aren't designed to handle the low signal-to-noise ratios inherent within non-vision signal processing tasks. While they are powerful, they are currently not the method of choice in the inherently noisy and dynamic critical infrastructure domain, such as smart-grid sensing, anomaly detection, and non-intrusive load monitoring.

Improving Robustness of Spectrogram Classifiers with Neural Stochastic Differential Equations

TL;DR

This work explores the idea of Neural Stochastic Differential Equations (NSDE's) to improve the robustness of models trained to classify time series data and the effect of NSDE's on the explainability of outputs and test the effectiveness of these approaches by applying them to a non-intrusive load monitoring dataset.

Abstract

Signal analysis and classification is fraught with high levels of noise and perturbation. Computer-vision-based deep learning models applied to spectrograms have proven useful in the field of signal classification and detection; however, these methods aren't designed to handle the low signal-to-noise ratios inherent within non-vision signal processing tasks. While they are powerful, they are currently not the method of choice in the inherently noisy and dynamic critical infrastructure domain, such as smart-grid sensing, anomaly detection, and non-intrusive load monitoring.
Paper Structure (14 sections, 2 equations, 4 figures, 4 tables)

This paper contains 14 sections, 2 equations, 4 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related work
  3. Model Explainability
  4. SDEs and Neural SDEs
  5. Noise Shaping
  6. Methods
  7. Model Types
  8. Noise Shaping
  9. Explanation Generation
  10. Experiments
  11. Data Source
  12. Processing and Training
  13. Results
  14. Conclusion

Figures (4)

  • Figure 1: Surface representations of the 2D Brownian surface noise injected into our Neural SDE
  • Figure 2: A general overview of how shaped stochastic noise is utilized in our ConvNext architecture to produce more robust explanation attributions.
  • Figure 3: Data collection locations in power system for injection dataset adams2023harmonic
  • Figure 4: A comparison of Non-Stochastic (top) and Stochastic (bottom) variants of ConvNext spectrogram explanations. A spectorgram input (left) into the Stochastic Neural SDE variant provides more coherent frequency attribution bands for both IG (center) and NoiseTunnel (right) explanations