Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DeepFilter: A Transformer-style Framework for Accurate and Efficient Process Monitoring

Hao Wang, Zhichao Chen, Licheng Pan, Xiaoyu Jiang, Yichen Song, Qunshan He, Xinggao Liu

TL;DR

DeepFilter replaces self-attention in Transformer-based process monitoring with a real-FFT–based filtering block to capture long-term discriminative patterns efficiently. The method achieves higher accuracy and lower computational cost than a range of baselines, including Transformer variants, on nuclear monitoring datasets, while maintaining real-time applicability. Theoretical analysis links the filtering operation to a circular convolution with a history-length kernel and demonstrates decorrelation in the frequency domain, supporting improved representations. Empirical results and ablations confirm the necessity of the spectral filtering components and demonstrate favorable scalability and deployment viability in industrial settings.

Abstract

The process monitoring task is characterized by stringent demands for accuracy and efficiency. Current transformer-based methods, characterized by self-attention for temporal fusion, exhibit limitations in accurately understanding the semantic context and efficiently processing monitoring logs, rendering them inadequate for process monitoring. To address these limitations, we introduce DeepFilter, which revises the self-attention mechanism to improve both accuracy and efficiency. As a straightforward yet versatile approach, DeepFilter provides an instrumental baseline for practitioners in process monitoring, whether initiating new projects or enhancing existing capabilities.

DeepFilter: A Transformer-style Framework for Accurate and Efficient Process Monitoring

TL;DR

DeepFilter replaces self-attention in Transformer-based process monitoring with a real-FFT–based filtering block to capture long-term discriminative patterns efficiently. The method achieves higher accuracy and lower computational cost than a range of baselines, including Transformer variants, on nuclear monitoring datasets, while maintaining real-time applicability. Theoretical analysis links the filtering operation to a circular convolution with a history-length kernel and demonstrates decorrelation in the frequency domain, supporting improved representations. Empirical results and ablations confirm the necessity of the spectral filtering components and demonstrate favorable scalability and deployment viability in industrial settings.

Abstract

The process monitoring task is characterized by stringent demands for accuracy and efficiency. Current transformer-based methods, characterized by self-attention for temporal fusion, exhibit limitations in accurately understanding the semantic context and efficiently processing monitoring logs, rendering them inadequate for process monitoring. To address these limitations, we introduce DeepFilter, which revises the self-attention mechanism to improve both accuracy and efficiency. As a straightforward yet versatile approach, DeepFilter provides an instrumental baseline for practitioners in process monitoring, whether initiating new projects or enhancing existing capabilities.
Paper Structure (22 sections, 2 theorems, 17 equations, 8 figures, 4 tables)

This paper contains 22 sections, 2 theorems, 17 equations, 8 figures, 4 tables.

Table of Contents

  1. Introduction
  2. Related Works
  3. Preliminaries
  4. Problem Definition
  5. Fourier Transform
  6. Methodology
  7. Motivation
  8. Global Filtering Block
  9. Theoretical Analysis
  10. DeepFilter Architecture and Learning Objective
  11. Experiments and Results
  12. Dataset Collection and Background
  13. Experimental Setup
  14. Datasets
  15. Baselines
  16. ...and 7 more sections

Key Result

Lemma 4.1

Suppose $\mathbf{W}=\mathcal{F}^{-1}(\mathbf{W}^\mathrm{(F)})$, "$*$" is the circular convolution operator, the filtered sequence in eq:barzk can be acquired via circular convolution below oppenheim1999discrete

Figures (8)

  • Figure 1: Overview of the core components in DeepFilter.
  • Figure 2: Overview of the nuclear monitoring stations in our project.
  • Figure 3: Dynamics of the quality variable (a) and some important process variables (b-d) in the collected dataset from Jinan station. We also provide the amplitude characteristics of their FFT in (b-d).
  • Figure 4: In-depth comparison on the monitoring performance of DeepFilter and Transformer.
  • Figure 5: Efficiency analysis results given varying settings, with solid lines for mean values of 10 trials and shaded areas for 90% confidence intervals. The default values of T, D and batch size are 16, 32, and 64, respectively.
  • ...and 3 more figures

Theorems & Definitions (3)

  • Lemma 4.1
  • proof
  • Lemma 4.2: Modified from wang2025iclrfredf