Accurately predicting the State of Health (SOH) of lithium-ion power batteries is a cornerstone of modern battery management systems (BMS). It is essential for ensuring the reliability and safety of everything from electric vehicles to grid-scale energy storage. While data-driven SOH prediction methods have gained popularity due to their flexibility, they typically rely on complete, high-quality performance data. In real-world scenarios, however, sensor malfunctions, communication drops, and inconsistent sampling often result in incomplete or noisy datasets.
This paper introduces a novel framework designed to achieve reliable SOH prediction even when faced with significant data gaps. By leveraging incomplete tensor representation learning and Transformer-based architectures, the researchers provide a robust solution to the "messy data" problem in battery health monitoring.
The Challenge of Incomplete Data
Traditional data-driven models, such as standard deep learning architectures, often assume a perfect input stream of voltage, current, and temperature readings. When data is missing, common fixes include simple imputation (filling gaps with averages) or deleting incomplete records. These methods are problematic because they often destroy the complex, multi-dimensional relationships inherent in battery degradation. A voltage drop at a specific second in one cycle is structurally related to similar drops in subsequent cycles; simple imputation fails to capture these cross-dimensional correlations.
A Three-Stage Prediction Framework
To address these challenges, the proposed method utilizes a three-stage pipeline:
1. Incomplete Tensor Representation Learning
Rather than treating battery data as a flat list or a simple time-series, the researchers model it as a fourth-order incomplete tensor. The dimensions (modes) are defined as: Battery Cell × Sampling Time × Monitoring Indicator × Cycle Index.
By using a Tucker2 decomposition-based approach, the model learns a "complete" feature representation directly from the incomplete raw data. This allows the system to capture structural correlations across all dimensions simultaneously, effectively "understanding" the underlying patterns of degradation despite the missing values.
2. Feature Extraction
The high-dimensional tensor is then compressed into a lower-dimensional "effective feature" space using factor matrices and mode-wise tensor-matrix multiplication. This step is crucial for reducing computational noise and focusing the model on the most significant signals of battery aging, filtering out the "distractions" caused by sensor inconsistencies.
3. Performance Prediction via Transformers
These refined features are fed into a Transformer-based model. Transformers are uniquely suited for this task because of their self-attention mechanisms, which excel at capturing long-term dependencies in time-series data. The model looks across historical cycles to predict performance over the next $|K|$ cycles, enabling a highly accurate SOH forecast.
Experimental Validation and Robustness
The framework was evaluated using two prominent benchmarks:
- Toyota Battery Dataset: 124 cells with 7 monitoring indicators over 100 cycles.
- HUST Dataset: 77 LFP/graphite cells tracked over 1,500 cycles.
To simulate real-world failures, the researchers randomly masked (removed) 10%, 20%, 30%, and 40% of the data. Even with 40% of the data missing, the proposed model maintained high precision. On the Toyota dataset, it reduced the Mean Squared Error (MSE) by 11% and the Mean Absolute Error (MAE) by 36% compared to the best-performing baseline models, such as FilterNet and CycleNet.
Key Insights
- Representation over Imputation: The study demonstrates that learning the representation of data in its incomplete state is superior to trying to fix the data before learning. This prevents the introduction of systematic biases or "fake" features.
- Structural Awareness: Modeling battery data as a multi-dimensional tensor allows the model to preserve the relationship between different monitoring indicators (like voltage and temperature) and how they evolve over hundreds of cycles.
- Superiority of Transformers: Compared to traditional Recurrent Neural Networks (RNNs) or LSTMs, the Transformer architecture proved more capable of handling the complex, non-linear curves typical of lithium-ion battery degradation.
Future Directions
While the framework is highly effective at handling missing data, future work will focus on addressing outliers—data points that are present but incorrect due to sensor noise. Additionally, the researchers aim to optimize the computational efficiency of handling 4th-order tensors to make the model more suitable for deployment on edge devices, such as on-board vehicle BMS, where processing power is limited.
Summary updated