💻 computer science

LIBAD: A Multimodal Anomaly Detection Benchmark for Li-Ion Battery Electrode Manufacturing

This paper introduces LIBAD, the first multimodal anomaly detection benchmark for Li-ion battery electrode manufacturing that addresses challenges in continuous process manufacturing and cross-modal inconsistency, alongside the proposed DA-Core method which significantly improves detection accuracy and efficiency by optimizing normal feature coreset selection.

Original authors: Wenbo Sui, Daniel Lichau, Harold Phelippeau, Zhao Liu

Published 2026-08-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Wenbo Sui, Daniel Lichau, Harold Phelippeau, Zhao Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to catch a thief in a massive, high-tech factory. Usually, when we think of factory inspections, we picture a robot arm checking a single, distinct object—like a car engine or a smartphone—looking for scratches or dents. This is called "discrete" manufacturing. But what if the factory doesn't make individual items? What if it's a giant, continuous conveyor belt spitting out miles of material, like a never-ending roll of wallpaper? This is "process manufacturing," and it's much harder to police. In this world, the "thief" is a tiny defect hidden inside the material, and the "detective" is a computer vision system. The challenge is that these systems often rely on two eyes: one that sees the surface (like a regular camera) and one that sees inside (like an X-ray). In many factories, these two eyes agree perfectly. But in the world of lithium-ion batteries, they often disagree. One eye might see a flaw clearly, while the other sees nothing at all. This paper dives into that confusing, messy reality to see if we can teach computers to be better detectives when their eyes don't agree.

The researchers behind this study, Wenbo Sui and their team, realized that while we have great tools for checking individual objects, we are terrible at checking continuous rolls of battery material. To fix this, they built something new called LIBAD. Think of LIBAD as a giant, specialized training school for AI detectives. It's the first-ever benchmark (a standardized test) specifically designed for spotting defects in the electrodes of lithium-ion batteries as they are being made on a real, rolling production line.

The "school" uses 744 real-life samples of battery electrode patches. These aren't fake computer-generated images; they are physical pieces cut from actual factory lines. The students (the AI algorithms) are trained using three different "senses":

VIS (Visible Light): A high-resolution grayscale camera that sees the surface texture, like a human looking at a piece of paper.
X-rayH (High-Resolution X-ray): A super-detailed internal scan, like a medical CT scan, which shows the inside structure but takes a long time.
X-rayL (Low-Resolution X-ray): A faster, slightly blurrier internal scan that can be used while the belt is moving (inline), making it practical for real factories.

The big twist in this story is that these "eyes" often see different things. In a normal factory, a scratch might be visible to both the camera and the X-ray. But in battery manufacturing, a defect might be a tiny bubble deep inside the material. The X-ray sees it clearly, but the surface camera sees nothing. Conversely, a surface scratch might be obvious to the camera but invisible to the X-ray. This creates a "cross-modal inconsistency," where the evidence is strong in one sense but weak or missing in another. The authors found that existing AI methods, which were trained to expect both eyes to agree, got confused and started screaming "THIEF!" at perfectly good battery patches. They produced too many false alarms.

To solve this, the team proposed a new method called DA-Core. Imagine you are trying to memorize what a "perfect" battery patch looks like. A standard method (called Farthest Point Sampling) tries to pick examples that are as different from each other as possible to cover all the bases. However, the authors realized this was like a detective who only memorizes the rare, weird-looking normal patches and forgets the common, everyday ones. Because the "normal" battery patches are so similar to each other (homogeneous), the rare ones are easy to spot, but the common ones are where the real trouble lies. If the AI doesn't have enough examples of the common "normal" look, it gets scared and thinks a normal patch is a defect.

DA-Core changes the strategy. Instead of just looking for the most different examples, it looks for the densest clusters of normal examples. It's like a detective who realizes, "Most normal patches look exactly like this group of 100," and makes sure to memorize that group thoroughly. By paying attention to where the normal data is crowded together, DA-Core builds a better memory bank.

The results of their experiments were promising but cautious. When they tested their new method against the old ones using the practical setup of Visible Light plus the fast Low-Resolution X-ray, the old methods had a false-positive rate (saying a good battery is bad) of about 60.4%. This means if you checked 100 good batteries, the old AI would wrongly reject about 60 of them, which is a disaster for a factory. DA-Core managed to lower this number to 54.3%. While that might not sound like a huge jump, in the world of industrial inspection, reducing false alarms is a massive deal. It suggests that by understanding how the data is clustered, we can make the AI less jumpy.

The paper also explicitly ruled out some popular ideas. They found that methods designed to "reconstruct" images (where the AI tries to redraw the picture from memory and sees what's missing) failed miserably here. Because the surface and the inside look so different, the AI couldn't figure out what a "normal" combined image should look like, leading to confusion. Similarly, methods that rely on the relationship between 3D shapes and surface colors (common in car or phone inspection) didn't work because battery electrodes are flat sheets, not 3D objects with complex geometry.

In the end, the authors suggest that LIBAD is still a very hard challenge. Even with their new method, the false-positive rate is still too high for a factory to run without human help. They conclude that we can't just copy-paste solutions from other industries; we need to design systems that understand the specific, messy way that battery defects hide and reveal themselves across different types of sensors. It's a step forward, suggesting that if we want safer electric vehicles and energy storage, we need to teach our AI detectives to listen to the quiet clues, not just the loud ones.

Technical Summary: LIBAD: A Multimodal Anomaly Detection Benchmark for Li-Ion Battery Electrode Manufacturing

Problem Statement

Industrial anomaly detection (IAD) has traditionally focused on discrete, object-centric products using strongly correlated RGB and 3D modalities. However, continuous process manufacturing, such as roll-to-roll production, remains underexplored. Furthermore, existing multimodal benchmarks often assume that anomalies produce consistent evidence across all sensing modalities. This assumption fails in scenarios where modalities measure distinct physical properties (e.g., surface appearance vs. bulk density), leading to "cross-modal anomaly inconsistency" where a defect is visible in one modality but weak or absent in another.

The paper addresses these gaps by introducing LIBAD, a benchmark for Li-ion battery electrode manufacturing. The dataset highlights two critical challenges:

Process Manufacturing Context: Unlike discrete objects, electrode patches exhibit highly homogeneous material appearances, making subtle intra-class variations difficult to distinguish from defects.
Modality Inconsistency: Defects may be prominent in visible-light (VIS) imaging but invisible in X-ray, or vice versa. Existing methods relying on cross-modal reconstruction or RGB-3D priors (appearance-geometry consistency) struggle in this setting, often misinterpreting normal modality differences as anomalies, resulting in high false-positive rates (FPR).

Methodology

The authors propose DA-Core, a memory-based anomaly detection method designed to handle the specific distribution of normal features in LIBAD.

1. Dataset Construction (LIBAD):

Data Source: 744 electrode patches collected from real roll-to-roll production lines.
Modalities: Three spatially aligned modalities:
- VIS: Double-sided visible-light grayscale imaging (surface appearance).
- X-rayH: High-resolution X-ray radiography (used offline for labeling and analysis).
- X-rayL: Low-resolution inline X-ray radiography (density-sensitive, used for the benchmark).
Characteristics: 11 defect categories. The dataset emphasizes that normal patches are visually homogeneous, and defect evidence is often modality-selective.

2. DA-Core Architecture:
DA-Core improves upon standard memory-based approaches (like PatchCore) by modifying the coreset selection process.

Feature Extraction: Uses frozen DINOv3 ViT-S/16 encoders to extract patch-level features from VIS and X-rayL inputs.
Density-Aware Coreset Selection (DA FPS):
- Standard Farthest Point Sampling (FPS) prioritizes global coverage but ignores local feature density, often under-representing dense clusters of normal variations.
- DA-Core introduces a density score ( $\rho_i$ ) for each feature based on its $k$ -nearest neighbors.
- The selection metric combines the distance to the current memory bank ( $r_i$ ) and the local density ( $\rho_i$ ): $q_i = \tilde{r}_i / (1 + \lambda \tilde{\rho}_i)$ .
- This strategy ensures the memory bank retains representatives from dense normal regions, reducing nearest-neighbor distances for common normal variations and suppressing false positives.
Scoring & Fusion:
- Anomaly scores are computed via nearest-neighbor matching for each modality.
- A One-Class Support Vector Machine (OCSVM) performs late fusion of VIS and X-rayL scores to generate the final image-level anomaly score.

Key Contributions

LIBAD Benchmark: The first multimodal IAD benchmark for Li-ion battery electrode manufacturing, featuring 744 real-world samples, three aligned modalities, and 11 defect categories. It explicitly addresses the challenges of continuous process manufacturing and modality-dependent defect visibility.
Characterization of Inconsistency: The paper identifies and analyzes "cross-modal anomaly inconsistency," demonstrating that defect evidence is often modality-selective rather than shared, challenging the assumptions of existing RGB-3D methods.
DA-Core Method: A density-aware coreset selection strategy that constructs a more representative memory bank from non-uniform normal feature distributions. It explicitly balances global coverage with local density to reduce false positives.
Comprehensive Benchmarking: Evaluation of representative multimodal IAD methods (including M3DM, CFM, G2SF, CFR) under a unified protocol, revealing their limited transferability to the LIBAD setting.

Results

Baseline Performance: Existing multimodal methods (adapted for VIS+X-rayL) show limited transferability. For instance, G2SF achieved an FPR95 of 59.5%, and CFR reached 68.1%, barely improving over unimodal PatchCore (60.4% FPR95). Reconstruction-based methods suffered from large residuals even on normal samples due to weak modality correlations.
DA-Core Performance:
- FPR95: DA-Core achieved 54.3%, a 6.1 percentage point reduction compared to standard multimodal PatchCore (60.4%).
- Efficiency: At a coreset ratio of 0.05, DA-Core outperformed the best standard coreset result (obtained at a 0.20 ratio) while reducing inference time by 43.9% (13.12 ms vs. 23.39 ms).
- Ablation: The density weight ( $\lambda$ ) was found to be critical; $\lambda=0.7$ yielded the best trade-off between coverage and density representation.
Unimodal vs. Multimodal: While combining VIS and X-rayL improved performance over single modalities, the primary gains came from the improved memory construction (DA-Core) rather than the fusion architecture itself.

Significance and Claims

The paper claims that LIBAD represents a significant step forward by shifting the focus from discrete object inspection to continuous process manufacturing, where data distributions and modality relationships differ fundamentally.

The authors assert that LIBAD is still an open and challenging benchmark, noting that even the best method (DA-Core) yields an absolute FPR95 (54.3%) that remains too high for direct industrial deployment without further refinement. The significance of the work lies in:

Highlighting that data distribution (specifically the density of normal features) and modality relationships (inconsistency) must be explicitly considered when designing anomaly detection methods for process manufacturing.
Providing a realistic testbed that exposes the limitations of current RGB-3D centric methods when applied to weakly correlated sensing modalities like VIS and X-ray.
Proposing a practical, efficient solution (DA-Core) that reduces false alarms, a critical requirement for inline inspection systems, thereby contributing to safer battery manufacturing for electric vehicles and energy storage systems.