Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Causal-HM: Restoring Physical Generative Logic in Multimodal Anomaly Detection via Hierarchical Modulation

Xiao Liu, Junchen Jin, Yanjie Zhao, Zhixuan Xing

TL;DR

The paper addresses causal blindness in multimodal industrial anomaly detection by enforcing a unidirectional Process to Result ($P \rightarrow R$) causal prior. It introduces Sensor-Guided CHM modulation and a Causal-Hierarchical encoding pipeline with a Mamba state-space encoder and anti-generalization decoder to bridge modality heterogeneity while preserving physical generative logic. The method achieves state-of-the-art I-AUROC of $90.7\%$ on the Weld-4M benchmark and demonstrates strong performance on process-hidden defects, along with improved inference efficiency over memory-bank baselines. This work offers a practical framework for robust, real-time industrial QA by leveraging physical causality to detect deep-seated process anomalies.

Abstract

Multimodal Unsupervised Anomaly Detection (UAD) is critical for quality assurance in smart manufacturing, particularly in complex processes like robotic welding. However, existing methods often suffer from causal blindness, treating process modalities (e.g., real-time video, audio, and sensors) and result modalities (e.g., post-weld images) as equal feature sources, thereby ignoring the inherent physical generative logic. Furthermore, the heterogeneity gap between high-dimensional visual data and low-dimensional sensor signals frequently leads to critical process context being drowned out. In this paper, we propose Causal-HM, a unified multimodal UAD framework that explicitly models the physical Process to Result dependency. Specifically, our framework incorporates two key innovations: a Sensor-Guided CHM Modulation mechanism that utilizes low-dimensional sensor signals as context to guide high-dimensional audio-visual feature extraction , and a Causal-Hierarchical Architecture that enforces a unidirectional generative mapping to identify anomalies that violate physical consistency. Extensive experiments on our newly constructed Weld-4M benchmark across four modalities demonstrate that Causal-HM achieves a state-of-the-art (SOTA) I-AUROC of 90.7%. Code will be released after the paper is accepted.

Causal-HM: Restoring Physical Generative Logic in Multimodal Anomaly Detection via Hierarchical Modulation

TL;DR

The paper addresses causal blindness in multimodal industrial anomaly detection by enforcing a unidirectional Process to Result () causal prior. It introduces Sensor-Guided CHM modulation and a Causal-Hierarchical encoding pipeline with a Mamba state-space encoder and anti-generalization decoder to bridge modality heterogeneity while preserving physical generative logic. The method achieves state-of-the-art I-AUROC of on the Weld-4M benchmark and demonstrates strong performance on process-hidden defects, along with improved inference efficiency over memory-bank baselines. This work offers a practical framework for robust, real-time industrial QA by leveraging physical causality to detect deep-seated process anomalies.

Abstract

Multimodal Unsupervised Anomaly Detection (UAD) is critical for quality assurance in smart manufacturing, particularly in complex processes like robotic welding. However, existing methods often suffer from causal blindness, treating process modalities (e.g., real-time video, audio, and sensors) and result modalities (e.g., post-weld images) as equal feature sources, thereby ignoring the inherent physical generative logic. Furthermore, the heterogeneity gap between high-dimensional visual data and low-dimensional sensor signals frequently leads to critical process context being drowned out. In this paper, we propose Causal-HM, a unified multimodal UAD framework that explicitly models the physical Process to Result dependency. Specifically, our framework incorporates two key innovations: a Sensor-Guided CHM Modulation mechanism that utilizes low-dimensional sensor signals as context to guide high-dimensional audio-visual feature extraction , and a Causal-Hierarchical Architecture that enforces a unidirectional generative mapping to identify anomalies that violate physical consistency. Extensive experiments on our newly constructed Weld-4M benchmark across four modalities demonstrate that Causal-HM achieves a state-of-the-art (SOTA) I-AUROC of 90.7%. Code will be released after the paper is accepted.
Paper Structure (20 sections, 1 equation, 4 figures, 3 tables, 1 algorithm)

This paper contains 20 sections, 1 equation, 4 figures, 3 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Reconstruction-based and Multimodal Anomaly Detection
  4. Heterogeneous Signal Modulation and State Space Models
  5. Method
  6. Preliminary
  7. Overall Framework
  8. Sensor-Guided Causal Heterogeneous Modulation
  9. Causal-Hierarchical Encoding and Multi-Angle Aggregation
  10. Anti-Generalization Causal Decoder
  11. Objective Functions and Anomaly Scoring
  12. Experiments
  13. Experimental Setup and Dataset
  14. Compared Baselines
  15. Implementation and Hyperparameters
  16. ...and 5 more sections

Figures (4)

  • Figure 1: Comparison between (left) Traditional Flat Fusion and (right) our Causal-Hierarchical Fusion.
  • Figure 2: The overall architecture of Causal-HM. The framework operates on fixed representations extracted by Frozen Backbones (L0, in gray) and is organized into a causal hierarchy. (a) Functional Pathways: Distinct colors denote trainable components (Green: Causal Encoders, Gold: Sensor-Guided CHM Modulation, Blue: Anti-Generalization Decoder). (b) Mechanism: The model explicitly learns the unidirectional mapping $P \to R$.
  • Figure 3: Robustness analysis under environmental noise. This figure illustrates the comparative I-AUROC degradation of Causal-HM and the M3DM baseline as the sensor noise level increases.
  • Figure 4: Qualitative results, where sampled image and our predicted anomaly map are shown for each class in Weld-4M8.