Multimodal Data-Driven Abnormal Condition Detection for Energy Storage Batteries Using EAA-Informer

Jun Xie, Zezhong Sun, Yan Li, Erik Cambria, Xianxun ZhuResearch Square 2025 · DOI

Abstract

To enhance the operational safety of energy storage batteries under complex conditions, this study proposes a multimodal data-driven abnormal condition identification method based on Enhanced Attention Allocation (EAA) Informer. First, an analysis of thermal experimental data under various abnormal conditions reveals that temperature changes during the early stages are often subtle. As a result, relying on a single temperature parameter is insufficient to accurately reflect the battery state. This requires the integration of multiple parameters into a joint modeling framework to improve identification accuracy. Subsequently, a multimodal temporal feature extraction network based on LSTM-MFM is constructed. This network enables deep modeling of multi-source data—such as voltage, current, temperature, and state of charge—and facilitates cross-modal correlation learning. To further enhance the model's ability to capture both critical time steps and long-range temporal dependencies, an EAA mechanism is introduced for feature optimization. This is followed by an improved Informer network to perform temporal modeling and abnormal condition identification. Experimental results show that the proposed method accurately identifies abnormalities with low false rates, highlighting its potential for energy storage system safety monitoring.

Introduction

As energy storage systems (ESS) scale globally to support the green energy transition, battery safety—specifically the prevention of thermal runaway—has become a paramount concern. Traditional Battery Management Systems (BMS) often rely on single-parameter monitoring, typically waiting for a significant temperature spike to trigger an alarm.

However, by the time a temperature rise is significant enough to be detected by standard thresholds, the battery may already be in an irreversible state of failure. Early-stage anomalies like slight overcharging or high-rate discharge produce subtle variations that are often lost in sensor noise. To capture these "weak but significant" indicators, a method is needed to model the complex, non-linear relationships between multiple physical parameters simultaneously over long durations.

The EAA-Informer Framework

The researchers propose the EAA-Informer, a framework that moves beyond simple thresholds by fusing multimodal data (voltage, current, temperature, and SOC). The system follows a three-step pipeline designed for high-sensitivity anomaly detection.

Workflow of the EAA-Informer framework, illustrating the three-step process from data acquisition to condition recognition.

1. Multimodal Feature Extraction (LSTM-MFM)

Instead of concatenating all sensor data immediately, the system uses independent LSTM networks for each modality. A Multimodal Factorization Mechanism (MFM) then decomposes these into "modality-specific" factors (unique to that sensor) and "shared discriminative" factors (cross-sensor correlations).

2. Feature Optimization (EAA)

The framework introduces Enhanced Additive Attention (EAA), which acts as a filter to accentuate critical time steps while suppressing redundant data. This ensures that the feature vectors are information-dense before entering the main temporal model.

3. Temporal Modeling (Improved Informer)

The core is an upgraded Informer network. While standard Transformers struggle with the computational cost of long sequences, the Informer uses ProbSparse Attention to maintain efficiency. Crucially, the authors added Modality-Aware Encoding to ensure the model retains the context of which data came from which sensor during fusion.

Experimental Setup and Datasets

The study utilized real-world BMS monitoring data from 280Ah Lithium Iron Phosphate ($LiFePO_4$) batteries. The dataset covers seven distinct operating states: normal charging/discharging, overcharging, overdischarging, high-rate charging/discharging, and high-temperature environments. Data was sampled at a frequency of 1Hz to capture the long-term temporal dependencies required for safety analysis.

Key Insights and Results

The EAA-Informer demonstrated a clear performance lead over industry-standard baselines:

  • Superior Accuracy: The model achieved 96.5% accuracy, significantly outperforming standard LSTM (81.0%) and basic Transformer models (83.0%).
  • Stability and Reliability: The "metric fluctuation range" (the gap between accuracy, precision, and recall) was only 0.3%, indicating a highly balanced model with low false-alarm rates.
  • Early Detection of Subtle Cues: In conditions like overcharging, the initial temperature rise is minimal. By analyzing the coupling between voltage and temperature simultaneously, the model catches anomalies much earlier than thermal-only sensors.
  • Edge-Ready Efficiency: By using ProbSparse attention, the model reduces computational complexity, making it a viable candidate for deployment within actual BMS hardware rather than relying on cloud servers.

Conclusion and Future Work

The EAA-Informer framework provides a robust solution for proactive battery safety monitoring. While highly effective for $LiFePO_4$ batteries, future research will focus on adapting the framework for other chemistries (like NCM) and exploring how these early detections can trigger active mitigation strategies, such as dynamic current limiting, to extend battery life and prevent degradation.

Summary updated

Abstract

<title>Abstract</title> To enhance the operational safety of energy storage batteries under complex conditions, this study proposes a multimodal data-driven abnormal condition identification method based on Enhanced Attention Allocation (EAA) Informer.First, an analysis of thermal experimental data under various abnormal conditions reveals that temperature changes during the early stages are often subtle. As a result, relying on a single temperature parameter is insufficient to accurately reflect the battery state. This requires the integration of multiple parameters into a joint modeling framework to improve identification accuracy. Subsequently, a multimodal temporal feature extraction network based on LSTM-MFM is constructed. This network enables deep modeling of multi-source data—such as voltage, current, temperature, and state of charge—and facilitates cross-modal correlation learning.To further enhance the model's ability to capture both critical time steps and long-range temporal dependencies, an EAA mechanism is introduced for feature optimization. This is followed by an improved Informer network to perform temporal modeling and abnormal condition identification. Experimental results show that the proposed method accurately identifies abnormalities with low false rates, highlighting its potential for energy storage system safety monitoring.
BibTeX
@inproceedings{xie2025multimodal,
  author = {Jun Xie and Zezhong Sun and Yan Li and Erik Cambria and Xianxun Zhu},
  title = {{Multimodal Data-Driven Abnormal Condition Detection for Energy Storage Batteries Using EAA-Informer}},
  booktitle = {Research Square},
  year = {2025},
  doi = {10.21203/rs.3.rs-6502419/v1},
}
Figure