Causal Convolutional Neural Networks as Finite Impulse Response Filters

Kiran Bacsa, Wei Liu, Xudong Jian, Huangbin Liang, Eleni N. ChatziarXiv 2025 · DOI · arXiv

Abstract

Operational Modal Analysis (OMA) is essential for assessing the structural integrity of systems like bridges and buildings. While deep learning models, particularly Convolutional Neural Networks (CNNs), have shown promise in time-series analysis, their "black box" nature often limits their utility in engineering contexts where physical interpretability is paramount. This paper presents a novel framework that treats Causal CNNs (CCNNs) with quasi-linear activations as equivalent to classical Finite Impulse Response (FIR) filters. By leveraging the associative property of convolution, the authors demonstrate that these networks can be collapsed into a single, interpretable filter that effectively identifies natural frequencies and vibration patterns in spectrally sparse data.

Introduction

In structural health monitoring, identifying "modal properties"—natural frequencies and vibration patterns—is critical for safety and maintenance. Standard CNN architectures are typically optimized for image processing where frequency content is broad. In contrast, structural vibrations are "spectrally sparse," meaning the critical information is concentrated in narrow frequency bands. This research addresses the gap between deep learning and classical signal processing by "transparentizing" neural networks, allowing them to be used reliably for system identification.

Methodology

The proposed architecture utilizes Causal CNNs (CCNNs) characterized by three key features:

  1. Left-padding: This ensures temporal causality, meaning the network only uses past data to predict future states.
  2. Extended Kernels: Unlike the small kernels typical in computer vision, this approach uses kernels longer than 35 taps to capture narrow-band spectral peaks.
  3. Quasi-linear Activations: By using hyperbolic tangent (tanh) functions, the network operates linearly near zero, maintaining the interpretability of a linear system while retaining the learning capacity of a neural network.

The core innovation lies in the associative property of convolution. Since the layers are quasi-linear, a multi-layer network can be mathematically collapsed into a single equivalent FIR filter post-training. This allows for direct comparison with optimal filters designed via classical least-squares criteria.

Overview of the proposed scheme: training a CCNN on sparse data and collapsing it into an equivalent FIR filter.

Experimental Validation

The framework was validated across three distinct sources:

  • Simulated Band-pass Tasks: Synthetic data used to verify if CCNNs could learn specific spectral peaks.
  • Multi-Degree-of-Freedom (MDOF) Simulations: Numerical models of 2-DOF and 9-DOF vibrating systems to test regression and unsupervised learning capabilities.
  • The Z24 Bridge Benchmark: A real-world dataset from a monitored bridge in Switzerland, serving as the gold standard for structural health monitoring.

Results and Discussion

The study revealed that CNNs possess an implicit spectral bias, naturally acting as "matched filters" that align with dominant energy frequencies in training data. Interestingly, deeper networks with multiple layers (which can be collapsed) were found to be more effective at learning these features than single-layer networks with many parameters.

Quantitatively, the trained CCNNs achieved a linearity score ($R^2$) of 0.975, justifying their treatment as linear FIR filters. As shown in the comparison below, the CCNN-derived filters closely match the weights and frequency responses of ideal Least-Squares FIR filters.

Comparison of learned CCNN weights and frequency response against an ideal LS-FIR filter at 10 Hz.

On the Z24 bridge data, the CCNN successfully identified natural frequencies with higher data efficiency than classical methods. While traditional Frequency Domain Decomposition (FDD) might require 28 minutes of data, the CCNN approach achieved stable results with only 50-second segments. This suggests that CCNNs can visualize exactly what frequencies the model is "listening" to, turning a deep learning model into a tool that acts like a classical engineering filter.

Conclusion

By collapsing deep learning models into equivalent FIR filters, this research provides a bridge between advanced machine learning and classical engineering principles. While the current model faces challenges in isolating individual modes in complex systems, it offers a significant step toward interpretable, data-efficient structural health monitoring. Future work will focus on refining the architecture to isolate individual structural modes and integrating these models into real-time monitoring systems.

Summary updated

BibTeX
@misc{bacsa2025causal,
  author = {Kiran Bacsa and Wei Liu and Xudong Jian and Huangbin Liang and Eleni N. Chatzi},
  title = {{Causal Convolutional Neural Networks as Finite Impulse Response Filters}},
  year = {2025},
  eprint = {2510.24125},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.24125},
}
Figure