NOCE-Net: Representation Learning for Battery Operational Context via Nested Sequence Modelling

summary

Video file (mp4)

The gist

The gist: NOCE-Net is a deep nested sequence model that learns control-enabling representations directly from raw cyclic operational or telemetric data to predict arbitrarily long voltage responses,

In short

NOCE-Net is a deep nested sequence model that learns useful representations directly from raw battery operational data to predict long voltage responses. It standardizes diverse battery datasets into Full Equivalent Cycles (FECs) and uses a combination of transformers and GRUs to capture both short-term cycle dynamics and long-term degradation trends. This method achieves strong zero-shot performance, showing learned states encode transferable degradation physics.

Key concepts

re-BatteryData
This data pipeline standardizes different battery aging datasets into a uniform format called Full Equivalent Cycles (FECs). FEC is defined by the total charge throughput over twice the nominal capacity. This standardization allows the model to learn consistently across various battery types and usage patterns, enabling end-to-end learning.
NOCE-Net Architecture
This framework uses a nested structure: multi-patch transformer encoders handle short intra-cycle dynamics (what happens within one cycle), while a GRU backbone models the long inter-cycle state evolution (how the battery degrades over many cycles). This combination allows it to capture both fine details and overall trajectory changes.
Full Equivalent Cycle (FEC)
An FEC is a standardized measure of battery aging based on absolute charge throughput, specifically when the battery has accumulated twice its nominal capacity. By defining sequences in terms of FECs, the model can learn dynamics that are independent of how many cycles an actual battery has undergone.
PID-R Features
These features are a compact summary derived from the frozen hidden states of NOCE-Net, combining Proportional, Derivative, Integral, and Relative components. These features provide a fixed-dimensional representation of the battery's degradation trajectory that is useful for estimating State of Health (SOH) without needing further training.

Terminology used across episodes

This episode discusses

The paper

NOCE-Net: Representation Learning for Battery Operational Context via Nested Sequence Modelling · Read on arXiv

Arnab Bhattacharjee, *, Wayes Tushar, *, Tapan K. Saha

Capalo AI · The University of Queensland

Learning effective representations from battery operational history is essential for predictive health management and control-oriented applications. Existing approaches rely on response-dependent engineered features or short early-life windows, limiting generalisability and precluding use in prospective settings where future responses are unavailable. We propose NOCE-Net, a deep nested sequence model that learns control-enabling representations directly from raw cyclic operational or telemetric data. Our data pipeline, re-BatteryData, standardises heterogeneous aging datasets into semantically equivalent Full Equivalent Cycles (FECs) defined by absolute charge throughput, enabling end-to-end learning across diverse operational regimes. NOCE-Net combines multi-patch transformer encoders for intra-cycle dynamics with a GRU backbone for inter-cycle state evolution, operating on approximately 700 batteries across ten public datasets spanning over 300 operational profiles. In zero-shot evaluation on unseen battery chemistries and usage patterns, NOCE-Net achieves 13.7% MAPE in voltage response prediction, outperforming strong long-sequence baselines by over 43% while using 40 times fewer parameters. Shape and component diagnostics of the learned hidden states reveal that the GRU's principal direction of evolution achieves Spearman ρ at least 0.95 with the capacity fade trajectory on average across all datasets with monotone degradation profiles, including zero-shot test batteries, providing evidence that the representations encode degradation dynamics transferable to downstream tasks beyond battery response modelling. Code is available at https://github.com/BeArnab96/NOCE-Net.

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "NOCE-Net: Representation Learning for Battery Operational Context via Nested Sequence Modelling".

Dev: The gist: NOCE-Net is a deep nested sequence model that learns control-enabling representations directly from raw cyclic operational or telemetric data to predict arbitrarily long voltage responses,

Rosa: First, who's behind it and why it matters.

Paper summary: Rosa: So we've covered how NOCE-Net learns control enabling representations by standardizing data into Full Equivalent Cycles and using a nested transformer and GRU architecture to capture both intra cycle dynamics and inter cycle state evolution re-BatteryData. We saw it performs well in zero shot evaluations with a thirteen to fourteen percent MAPE re-BatteryData <ref:2610.21391#pg2>.

Dev: And we also looked at how the learned hidden states encode degradation structure without capacity labels, showing that the principal component of the hidden state trajectory achieves near perfect rank correlation with capacity fade across all splits including zero shot batteries re-BatteryData. This means we can infer degradation shape even without knowing the actual capacity numbers.

Taro: The paper also flagged that while it captures direction, it doesn't capture absolute scale, which means they need per dataset calibration to get a full SOH estimate, and there are some cases like evTOL style partial discharge profiles where capacity trajectories aren't monotone re-BatteryData.

Rosa: So the overall picture from NOCE-Net is that it provides a non invasive framework for long horizon voltage forecasting by coupling operational inputs with latent dynamics to capture degradation driven changes in battery behavior without relying on RPTs re-BatteryData.

Dev: The authors highlight that the recurrent backbone and task and FEC conditioning are both critical for learning the long term dynamics governing voltage response over battery life re-BatteryData. They also point out that they compare it against models like TimerXL, AutoTimes, MAMBA, and TinyTimeMixers re-BatteryData.

Taro: It really shows the value of preserving that nested structure when you're dealing with variable length inputs, as the Transformer based models like AutoTimes and TimerXL don't generalize well to irregular sequences re-BatteryData.

Rosa: So what this paper means for us in the long run is that we have a new way to learn representations directly from raw operational data for battery health management and control applications re-BatteryData. It’s about building systems that can handle future predictions even when the world misbehaves.

Conclusion: Rosa: So we're wrapping up on NOCE-Net, that paper titled "NOCE-Net: Representation Learning for Battery Operational Context via Nested Sequence Modelling."

Dev: Yeah, it's about how they learned these control enabling representations directly from raw operational or telemetric data.

Rosa: The core idea is using this deep nested sequence model to predict long voltage responses without needing those response-dependent engineered features.

Dev: They standardized all the different aging datasets into something called Full Equivalent Cycles, which makes the learning work across totally different operational regimes.

Rosa: And what's really interesting is that these learned hidden states reveal degradation dynamics—the way things are fading—even without using any capacity labels for training.

Dev: That’s a big deal because it means we can infer the shape of the degradation trajectory, not just the direction it's going.

Rosa: But there are some caveats, like they still need some per-dataset calibration to nail down the absolute scale of that fading.

Dev: And they mentioned that for certain battery types, like those with partial discharge profiles, you can get non-monotone capacity trajectories, which means scalar SOH estimation gets tricky.

Rosa: So in simple terms, NOCE-Net gives us a robust way to predict long voltage responses using raw data and learns how the battery degrades on its own.

Dev: It moves us away from relying solely on pre-defined features and towards learning representations that are truly transferable across different battery chemistries.

Rosa: This kind of representation learning could really change how we build predictive health management systems for everything from electric vehicles to grid storage.

More episodes

← Home