STEP: Learning STructured Embeddings for Progressive Time Series

arXiv:2605.31061 · cs.LG, cs.AI · Submitted 2026-05-29 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "STEP: Learning STructured Embeddings for Progressive Time Series".

Jane: Here's the gist of the paper. We present STEP (STructured Embeddings for Progressive Time Series) a novel method for learning interpretable representations of progressive time series, that is, data capturing irreversible state transitions such as degradation or task completion Our approach uses a self-supervised contrastive objective to learn a low-dimensional latent space whose geometry is itself the interpretation From this structure we read a latent compass, the polar coordinates (θ, r) of the latent vector, in which θ tracks the progression of the underlying state (e.g., from healthy to failed) and r identifies the active mode (e.g., the operating condition), without any proxy labels We evaluate the approach against the state of the art on diverse domains, including industrial degradation, robotic tasks, and neural activity, validating three key capabilities: (1) end-state prediction, (2) multi-step forecasting, and (3) interpretable phase separation Core Methodology The central idea behind our approach is to make the latent space itself the interpretation Each observation xt is encoded into a point zt on a learned manifold (latent space), and thus an entire trajectory becomes a path across that manifold, and two polar coordinates of zt, namely the angle θt and the radius rt, together form a latent compass We learn this by combining three forces in a composite objective The full objective is Ltotal = λ1 Lrecon + λ2 Ltrip + λ3 Lproto, (1) Learning the Geometry We use a triplet loss with cosine similarity to organize z-space by state similarity Geometrically cos(za, zp) = cos(θap), so Eq.

Tom: (2) softly encourages θap < θan, ordering embeddings angularly along a manifold that follows the state sequence s1,..., sT The sampling strategy constructs triplets using temporal proximity within a trajectory as a sampling proxy for state similarity The encoder is per-observation by design Establishing the Reference Frame To fix this frame we set the start and end anchors to two standard basis vectors of R d, zinit = e1, zend = e2, (4) which are orthogonal by construction This instantiates the standard partially-observable picture of a progressive process: the latent space contains an initial-state region (around e1) and an absorbing terminal region (around e2), with trajectories navigating between them The radial direction is also left free so that distinct modes spread around the anchors rather than onto them Performance and Interpretability Our method matches or improves over black-box counterparts on all of these while providing transparency about the underlying mechanisms A simple linear regressor on top of the latent compass coordinates is competitive with deep architectures, direct quantitative evidence that the underlying state is encoded in a geometrically accessible form The use of θ and r as diagnostics for AE and SoftCLT (which were not designed for circular geometry) is motivated in Appendix C.1 Key Findings on Decoupling The angular coordinate θ does not collapse to t/T (FD002) Adding t/T to θ does not improve, and slightly hurts, downstream RMSE/R2, evidence that θ already encodes the state The marginals of θ and t/T are nearly disjoint near the trajectory extremes (KS=1.00 early, 0.98 late) Robustness and Scaling The non-regression assumption concerns the latent health state, not noisy sensors The instability previously observed at dim(z)=4 is resolved via β-scheduling, and results are stable across dim(z) ∈ 2, 4, 8, 16, 32 with no collapse Even with a single zend, unsupervised K-means (k=2) on terminal embeddings recovers distinct failure modes on FD003 (1 OC, 2 FMs) and FD004 (6 OCs, 2 FMs) with high silhouette scores We have presented STEP, a method for learning interpretable representations from progressive time series by framing state evolution as geometric paths in a contrastively learned latent space anchored to fixed orthogonal prototypes Our approach yields a latent compass, the polar coordinates (θ, r), that tracks progression without proxy supervision, exposes underlying system mechanisms, and disentangles progression from active mode Such information is captured in a relatively low-dimensional bottleneck, as little as 2 dimensions, and remains stable up to dim(z)=32 via βscheduling We performed experiments on progressive time series from industrial degradation, robotic task completion, and neural activity In all cases interpretability does not come at the cost of performance: simple linear models on the latent compass coordinates match or exceed complex black-box approaches, providing direct quantitative evidence that the underlying state is encoded in a geometrically accessible form Limitations and Future Work Our method is designed for progressive (non-regressive) sequences and does not natively handle systems that revert to earlier concept states (e.g., after maintenance actions) Extensions to (i) reversible state transitions with intervention modeling, (ii) learning intermediate prototypes from the discovered latent structure as a refinement stage, and (iii) integrating latent SDEs to reduce forecasting drift on longer horizons, are promising directions Notation Reference Table 2: Notation and terminology used throughout the paper Symbol Description xt Observation at time t st Unobserved latent (e.g.

Jane: health) state at time t zt = fenc(xt) Learned latent: z˜ Mean-wise scaled latent set z ¯ Trajectory-wise normalized latent z init, cend Fixed latent prototype vectors Xinit, Xend Initial/terminal observation sets Lrecon Reconstruction loss (MSE) Ltriplet Triplet loss with cosine similarity Lproto Prototype alignment loss Ltotal Combined objective Model Definition and Theoretical Framework B.1 RUL problem formulation We have a dataset of observation-trajectories (x1,..., xT), each sampled independently (and possibly of varying length T) B.2 Formulation as a Progressive Latent Markov Process The underlying mechanism behind a degradation/progression process can be expressed via a nonregressive Markov chain (st)Tu t=1 Model Architecture Overview Table 3: Model architecture components and specifications Component Description Specification Input Projection Linear projection to transformer dimension R input dim → R dmodel Transformer Encoder Stack of transformer layers Layers / Heads / FF= 2×d, GELU Global Pooling Sequence aggregation Adaptive average pooling Latent Projection MLP to latent space R dmodel →R latent dim Decoder MLP reconstruction R latent dim→R input dim Parameter configuration.

Tom: Default values: latent dim = 2 (or 4–32), dmodel = 32, 8 attention heads, 4 transformer layers, dropout 0.1, GELU activation Theoretical Basis for Polar Coordinate Diagnostics Our use of θ and r as the latent compass coordinates is grounded in contrastive representation learning Wang and Isola

2020: identify two key properties: alignment (positive pairs should be close) and uniformity (features should be roughly uniform on the hypersphere) Implementation Details D.4 Ablation: Prototype Anchoring (k = 0) Removing prototype anchoring degrades performance primarily on heterogeneous subsets (FD004: 12.10→19.01 RMSE, +57%) Experimental Framework Our goal is to investigate interpretability gains without losing performance We test our approach on three tasks: semantic interpretation, RUL estimation, and forecasting We compare against two baselines with the same backbone architecture: AE (trained with Lrecon only) and SoftCLT (trained with the SoftCLT loss

Lee et al., 2024: ), which is state of the art on general time-series datasets Results and Discussion Table 1 shows our method achieves SOTA on FD002 (12.34 RMSE) and FD004 (12.10 RMSE) and competitive performance on the other subsets Among the compared methods, only I-GLIDE also offers HIs for downstream prognostics, but it is trained only on data deemed “healthy” through novelty detection and thus fails to capture the full degradation progression Notably, a simple linear regressor trained on the latent compass coordinates (L.Reg.) achieves competitive performance, sometimes surpassing specialized deep learning models (e.g., RVE, VAE+RNN) Geometric Interpretation of Latent Spaces The latent manifolds learned by each method reveal a clear progression in structure and interpretability (Fig.

Paper summary: Tom: So we're looking at this paper called "STEP: Learning STructured Embeddings for Progressive Time Series," and what it claims is that they've created a new way to learn representations for time series data that capture things like degradation or task completion.

Jane: That sounds pretty big, Tom. What exactly is the main idea behind this approach? What problem are they trying to solve with these new representations?

Tom: Well, the core idea is that they use a self-supervised contrastive objective to learn a low-dimensional latent space where the geometry itself gives us the interpretation. They're not just learning embeddings; they're learning how things progress.

Lu: It’s about making the latent space itself an interpretation, which means we can read some kind of compass from it. Each observation becomes a point on a manifold anchored between two fixed orthogonal prototype vectors, and a whole trajectory becomes a path across that manifold <ref:2605.31061#pg1>.

Meng: A compass? So they're trying to find some way to map time into directions like "healthy to failed," which is something traditional methods struggle with when things get messy in the real world.

Lalam: They claim that from this structure, you can read polar coordinates, called theta and r. Theta tracks the progression of the underlying state, like moving from healthy to failed, and r identifies what mode you're currently in <ref:2605.31061#pg3>.

Jane: So if I'm driving or cooking, what does that mean for me? What does this latent compass actually tell me about my situation?

Tom: It means you get these polar coordinates, theta and r. Theta tells you how the state is progressing—maybe from healthy to failed. And r identifies the active mode, like whether you're in a certain operating condition right now <ref:2605.31061#pg3>.

Lu: The geometry they learn is structured by using a triplet loss with cosine similarity to organize the latent space by state similarity, where the cosine of the angle between two points tells you about their state difference <ref:2605.31061#pg3>.

Meng: That sounds like they're forcing the model to group things that are similar in terms of their progression together, which is smart if we want to find distinct failure modes without any pre-labeled data.

Lalam: They set up a reference frame by anchoring the start and end points of this space to two standard basis vectors, z init and z end <ref:2605.31061#pg4>, which are orthogonal by construction. This sets up an initial state region around one anchor and an absorbing terminal region around the other <ref:2605.31061#pg4>.

Paper summary: Jane: So, if we think of it geometrically, we have a starting point and a stopping point that the data has to travel between as it degrades or completes a task?

Tom: Exactly. The trajectory becomes this path across the manifold connecting those two fixed points <ref:2605.31061#pg1>. And they use this structure to validate three key capabilities: end-state prediction, multi-step forecasting, and interpretable phase separation <ref:2605.31061#pg4>.

Meng: That’s the practical part. So they're not just making a pretty visualization; they're proving that this structure helps us predict what happens next or tell us what mode we're in, without needing those messy proxy labels that usually get noisy data.

Lu: They test this against several domains, like industrial degradation and robotic tasks, to show it works across different types of progression <ref:2605.31061#pg1>. The goal is to show that this geometric structure is robust enough for real-world applications.

Lalam: And they found that the angular coordinate theta doesn't just collapse to time divided by total time, t over T <ref:2605.31061#pg5>, and adding t over T actually hurts the performance a little, which suggests that theta is already encoding the state information <ref:2605.31061#pg5>.

Jane: That's interesting because it means we don't need to manually feed in time as a separate input if this structure is working well on its own. But what about the numbers, Tom? How does it compare to what’s already out there?

Tom: They match or improve over black-box counterparts on all those tasks they tested <ref:2605.31061#pg8>. And here's a concrete piece of evidence: a simple linear regressor built on top of these latent compass coordinates is competitive with deep architectures <ref:2605.31061#pg8>.

Meng: A linear model beating a deep architecture sounds promising for deployment because it’s way easier to certify and debug at scale. But what about the instability when they test it on different dimensions of the latent space?

Lu: They addressed that instability by using beta scheduling, which keeps things stable across latent space dimensions from two up to thirty-two, showing no collapse even with just a single end point vector <ref:2605.31061#pg8>.

Jane: So it seems like the geometry itself is the key ingredient here, and it’s not overly sensitive to how much detail we pack into that space. But what are they actually suggesting this means for us in terms of understanding systems?

Paper summary: Tom: It means we get transparency about the underlying mechanisms. You see exactly how the system moves through its states, which is a huge deal when you're dealing with something safety-critical <ref:2605.31061#pg2>.

Meng: I wonder about running this at scale. If we have massive streams of data, how does the cost of training this complex contrastive objective compare to just using a standard autoencoder?

Lu: The architecture itself is designed to be relatively efficient, with a linear projection into the latent space and then an MLP for decoding, keeping the parameter count manageable depending on your choice of latent dimension Table three.

Lalam: And they noted that even if you only have one end point vector, unsupervised clustering using k equals two on the terminal embeddings can successfully recover distinct failure modes in datasets like FD003 and FD004 with high silhouette scores <ref:2605.31061#pg8>.

Jane: So, to wrap up what we’ve heard about "STEP: Learning STructured Embeddings for Progressive Time Series," it seems this method gives us a way to visualize the progression of time series data not just as numbers, but as a geometric path with clear directional information.

Tom: Right. It uses that latent compass—theta for state and r for mode—to give you direct diagnostics without needing any proxy labels to tell you what's going on <ref:2605.31061#pg1>.

Meng: I think the real implication is that we can start using these learned structures to certify models because we can see the progression path, not just a final prediction number.

Lu: They point out a limitation, though, and it’s important: this method is designed for progressive sequences only; it doesn't natively handle systems that might revert back to an earlier concept state after some intervention Limitations.

Jane: That makes sense. If you have a system that gets fixed and then starts degrading again later, this specific structure might not capture the full picture right away. So where does this leave us in terms of future research?

Tom: The authors are looking at extending it to reversible state transitions, modeling interventions like maintenance actions or learning intermediate prototypes from the discovered latent structure Limitations.

Meng: And integrating latent stochastic differential equations to help reduce prediction drift on longer time horizons, because they noted performance can degrade when forecasting over very long periods Results and Discussion.

Lalam: So this paper shows a path toward making complex time series representations interpretable by turning the data evolution into a geometric journey with clear angular and radial components.

Conclusion: Tom: So we've looked at this paper called STEP, which stands for Structured Embeddings for Progressive Time Series.

Jane: Right, so what's the main idea here? What are they actually trying to do with all this time series data?

Tom: Essentially, they’re learning how things change progressively—like equipment getting worn out or a robotic task finishing—and they want to make sure that the way the AI represents those changes has a clear structure.

Jane: So instead of just throwing all the data into a big black box, they’re trying to find some hidden geometric path that shows us exactly how the system is evolving.

Tom: Exactly. They use this contrastive objective to build a low-dimensional space where the very shape of that space tells you something about the state progression.

Jane: And they call this structure a latent compass, with an angle theta and a radius r, which gives them clues about what’s happening without needing any labels.

Tom: That’s the core takeaway for me. They're using geometry to extract information that usually requires messy proxy labels.

Jane: So what does this mean for us? What are the real-world implications of having a compass instead of just a final prediction number?

Tom: It means we get transparency into the underlying mechanisms, so we can actually see *why* the model thinks something is failing, not just *that* it is failing.

Jane: And they show that this works across different areas like industrial equipment and even neural activity. That’s pretty broad.

Tom: Yeah, and they found that a simple linear model built on these compass coordinates can match or even beat deep learning models on the actual tasks they tested.

Jane: That’s huge for deployment because it means we can get high performance without needing those massive deep networks running all the time.

Tom: But there are still caveats, like this method only works for things that degrade in one direction and doesn't handle systems that might actually revert to an earlier state after some maintenance.

Jane: Right, so while the geometry is powerful for tracking a single progression, we still have to be careful about real-world systems that aren't strictly progressive.

Tom: That’s where the next part of this paper gets really interesting—the authors are already looking at how to handle those reversible transitions.

Jane: So they aren't just stopping at this structure; they’re already thinking about fixing the limits so it can handle more complex, real-world scenarios.

Lucas Thil, Jesse Read, Rim Kaddah, Guillaume Doquet

École Polytechnique LIX · IRT SystemX · Safran Tech

cs.LG, cs.AI

Submitted: 2026-05-29

Updated: 2026-11-06

Comments: NeurIPS 2026, 10 pages

Journal ref: Advances in Neural Information Processing Systems Main Conference (NeurIPS 2026)

Code: https://github.com/LucasStill/STEP

License: http://creativecommons.org/licenses/by-nc-sa/4.0/

Importance score: 91/100

The gist: We present STEP (STructured Embeddings for Progressive Time Series) a novel method for learning interpretable representations of progressive time series, that is…

Key concepts

Latent Compass
This is the core interpretation mechanism where each data point's latent vector has two polar coordinates: theta and r. Theta represents the progression of the underlying state, like moving from healthy to failed. R represents the active operating mode, such as a specific condition or failure type, allowing researchers to diagnose system behavior geometrically.
Triplet Loss with Cosine Similarity
This loss function organizes the latent space by enforcing state similarity. It ensures that embeddings of similar states are close together in the latent space by minimizing the angular difference (cosine similarity) between positive pairs, effectively structuring the manifold according to the temporal sequence of degradation.
Polar Coordinates (θ, r)
These coordinates provide a diagnostic tool for system health. Theta tracks how far along a progression path an observation is, while r identifies which specific operating condition or mode the system is currently in. This allows for reading the system's state without relying on external labels.
Progressive Time Series
This refers to data that captures irreversible state transitions, such as equipment degradation or task completion. The method is designed specifically to model these non-regressive sequences where the system moves through distinct, ordered stages over time.

Terminology

Summary

We present STEP (STructured Embeddings for Progressive Time Series) a novel method for learning interpretable representations of progressive time series, that is, data capturing irreversible state transitions such as degradation or task completion Our approach uses a self-supervised contrastive objective to learn a low-dimensional latent space whose geometry is itself the interpretation From this structure we read a latent compass, the polar coordinates (θ, r) of the latent vector, in which θ tracks the progression of the underlying state (e. p. 0

Where the paper says it

fr Jesse Read LIX, École Polytechnique Palaiseau, France Rim Kaddah IRT SystemX Palaiseau, France Guillaume Doquet Safran Tech Chateaufort, France Abstract We present a novel method for learning interpretable representations of progressive time series, that is, data capturing irreversible state transitions such as degradation or task completion. Our approach uses a self-supervised contrastive objective to learn a low-dimensional latent space whose geometry is itself the interpretation: each observation becomes a point on a manifold anchored between two fixed orthogonal prototype vectors, and a trajectory becomes a path across that manifold.

g., from healthy to failed) and r identifies the active mode (e. p. 0
Where the paper says it

From this structure we read a latent compass, the polar coordinates (θ, r) of the latent vector, in which θ tracks the progression of the underlying state (e. , from healthy to failed) and r identifies the active mode (e.

g., the operating condition), without any proxy labels We evaluate the approach against the state of the art on diverse domains, including industrial degradation, robotic tasks, and neural activity, validating three key capabilities: (1) end-state prediction, (2) multi-step forecasting, and (3) interpretable phase separation

Core Methodology

The central idea behind our approach is to make the latent space itself the interpretation Each observation xt is encoded into a point zt on a learned manifold (latent space), and thus an entire trajectory becomes a path across that manifold, and two polar coordinates of zt, namely the angle θt and the radius rt, together form a latent compass We learn this by combining three forces in a composite objective The full objective is Ltotal = λ1 Lrecon + λ2 Ltrip + λ3 Lproto, (1)

Learning the Geometry

We use a triplet loss with cosine similarity to organize z-space by state similarity Geometrically cos(za, zp) = cos(θap), so Eq. (2) softly encourages θap < θan, ordering embeddings angularly along a manifold that follows the state sequence s1,. p. 3

Where the paper says it

Geometrically cos(za, zp) = cos(θap), so Eq. (2) softly encourages θap < θan, ordering embeddings angularly along a manifold that follows the state sequence s1, .

.., sT The sampling strategy constructs triplets using temporal proximity within a trajectory as a sampling proxy for state similarity The encoder is per-observation by design

Establishing the Reference Frame

To fix this frame we set the start and end anchors to two standard basis vectors of R d, zinit = e1, zend = e2, (4) which are orthogonal by construction This instantiates the standard partially-observable picture of a progressive process: the latent space contains an initial-state region (around e1) and an absorbing terminal region (around e2), with trajectories navigating between them The radial direction is also left free so that distinct modes spread around the anchors rather than onto them

Performance and Interpretability

Our method matches or improves over black-box counterparts on all of these while providing transparency about the underlying mechanisms A simple linear regressor on top of the latent compass coordinates is competitive with deep architectures, direct quantitative evidence that the underlying state is encoded in a geometrically accessible form The use of θ and r as diagnostics for AE and SoftCLT (which were not designed for circular geometry) is motivated in Appendix C.1

Key Findings on Decoupling

The angular coordinate θ does not collapse to t/T (FD002) Adding t/T to θ does not improve, and slightly hurts, downstream RMSE/R2, evidence that θ already encodes the state The marginals of θ and t/T are nearly disjoint near the trajectory extremes (KS=1.00 early, 0. p. 5

Where the paper says it

We analyze the views z˜ and z¯ to interpret trajectory progression across all datasets, asking whether the learned representations disentangle OCs/FMs (C-MAPSS, Appendix Table 4), task-solving strategies (robotics), and intruder-driven behavioral phases (mouse). t-SNE of latent z (Visual purpose only) Angular Progression AngleRadius Space (polar coordinates) BaseAE Latent Dimension 2 (scaled) Unit Trajectory Start (early) * End (late) ^ 2 150฿ Polar Angle 0 Latent Dimension 1 (scaled) | | | h l Radius r from origin (0,0) Latent Dimension 2 (scaled) Polar Angle 0 Radius r from origin (0,0) SoftCLT Latent Dimension 1 (scaled) Latent Dimension 2 (scaled) Polar Angle O Radius r from origin (0,0) STEP w=1 2 Latent Dimension 1 (scaled) STEP w=10, dim(z)=2 STEP w=10, dim(z)=4 2 1 2 Latent Dimension 2 (scaled) Latent Dimension 2 (scaled) Polar Angle 6 Polar Angle 6 Radius r from origin (0,0) Radius r from origin (0,0) Figure 2: Latent compass comparison on C-MAPSS FD002 (6 operating conditions).

98 late)

Robustness and Scaling

The non-regression assumption concerns the latent health state, not noisy sensors The instability previously observed at dim(z)=4 is resolved via β-scheduling, and results are stable across dim(z) ∈ 2, 4, 8, 16, 32 with no collapse Even with a single zend, unsupervised K-means (k=2) on terminal embeddings recovers distinct failure modes on FD003 (1 OC, 2 FMs) and FD004 (6 OCs, 2 FMs) with high silhouette scores

Conclusion

We have presented STEP, a method for learning interpretable representations from progressive time series by framing state evolution as geometric paths in a contrastively learned latent space anchored to fixed orthogonal prototypes Our approach yields a latent compass, the polar coordinates (θ, r), that tracks progression without proxy supervision, exposes underlying system mechanisms, and disentangles progression from active mode Such information is captured in a relatively low-dimensional bottleneck, as little as 2 dimensions, and remains stable up to dim(z)=32 via βscheduling We performed experiments on progressive time series from industrial degradation, robotic task completion, and neural activity In all cases interpretability does not come at the cost of performance: simple linear models on the latent compass coordinates match or exceed complex black-box approaches, providing direct quantitative evidence that the underlying state is encoded in a geometrically accessible form

Limitations and Future Work

Our method is designed for progressive (non-regressive) sequences and does not natively handle systems that revert to earlier concept states (e.g., after maintenance actions) Extensions to (i) reversible state transitions with intervention modeling, (ii) learning intermediate prototypes from the discovered latent structure as a refinement stage, and (iii) integrating latent SDEs to reduce forecasting drift on longer horizons, are promising directions

Notation Reference

Table 2: Notation and terminology used throughout the paper Symbol Description xt Observation at time t st Unobserved latent (e.g. health) state at time t zt = fenc(xt) Learned latent: z˜ Mean-wise scaled latent set z ¯ Trajectory-wise normalized latent z init, cend Fixed latent prototype vectors Xinit, Xend Initial/terminal observation sets Lrecon Reconstruction loss (MSE) Ltriplet Triplet loss with cosine similarity Lproto Prototype alignment loss Ltotal Combined objective

Model Definition and Theoretical Framework

B.1 RUL problem formulation We have a dataset of observation-trajectories (x1,. p. 4

Where the paper says it

Training trajectories contain complete run-to-failure sequences; test sequences are right-censored at arbitrary points with known RUL ground truth. We also use NVIDIA GROOT’s Can2Drawer dataset [NVIDIA et al.

.., xT), each sampled independently (and possibly of varying length T) B. p. 3
Where the paper says it

The encoder is per-observation by design. The encoder fenc processes each observation xt independently; temporality enters only through the sampling, not the architecture.

2 Formulation as a Progressive Latent Markov Process The underlying mechanism behind a degradation/progression process can be expressed via a nonregressive Markov chain (st)Tu t=1

Model Architecture Overview

Table 3: Model architecture components and specifications Component Description Specification Input Projection Linear projection to transformer dimension R input dim → R dmodel Transformer Encoder Stack of transformer layers Layers / Heads / FF= 2×d, GELU Global Pooling Sequence aggregation Adaptive average pooling Latent Projection MLP to latent space R dmodel →R latent dim Decoder MLP reconstruction R latent dim→R input dim Parameter configuration. Default values: latent dim = 2 (or 4–32), dmodel = 32, 8 attention heads, 4 transformer layers, dropout 0. p. 5

Where the paper says it

We train three variants of the backbone autoencoder (STEP, SoftCLT, AE), each using the same architecture but different training objectives. For each variant we extract the latent compass coordinates and train a transformer-based RUL estimator on these features, as well as a simple linear regressor head (L.

1, GELU activation

Theoretical Basis for Polar Coordinate Diagnostics

Our use of θ and r as the latent compass coordinates is grounded in contrastive representation learning Wang and Isola [2020] identify two key properties: alignment (positive pairs should be close) and uniformity (features should be roughly uniform on the hypersphere)

Implementation Details

D.4 Ablation: Prototype Anchoring (k = 0) Removing prototype anchoring degrades performance primarily on heterogeneous subsets (FD004: 12. p. 4

Where the paper says it

The radial direction is also left free so that distinct modes spread around the anchors rather than onto them: in our FD002 experiments, r recovers all six operating conditions as concentric rings without any supervision (Sec. Removing the prototype loss (λ3=0) degrades performance primarily on heterogeneous subsets (FD004: 12.

10→19.01 RMSE, +57%)

Experimental Framework

Our goal is to investigate interpretability gains without losing performance We test our approach on three tasks: semantic interpretation, RUL estimation, and forecasting We compare against two baselines with the same backbone architecture: AE (trained with Lrecon only) and SoftCLT (trained with the SoftCLT loss [Lee et al., 2024]), which is state of the art on general time-series datasets

Results and Discussion

Table 1 shows our method achieves SOTA on FD002 (12.34 RMSE) and FD004 (12. p. 7

Where the paper says it

3 RUL Estimation Performance Table 1 shows our method achieves SOTA on FD002 (12. 34 RMSE) and FD004 (12.

p. 7
Where the paper says it

3 RUL Estimation Performance Table 1 shows our method achieves SOTA on FD002 (12. 34 RMSE) and FD004 (12.

p. 7
Where the paper says it

3 RUL Estimation Performance Table 1 shows our method achieves SOTA on FD002 (12. 34 RMSE) and FD004 (12.

10 RMSE) and competitive performance on the other subsets Among the compared methods, only I-GLIDE also offers HIs for downstream prognostics, but it is trained only on data deemed “healthy” through novelty detection and thus fails to capture the full degradation progression Notably, a simple linear regressor trained on the latent compass coordinates (L. p. 7
Where the paper says it

Unlike most prior works, our approach provides interpretable HIs without compromising predictive accuracy. Among the compared methods, only I-GLIDE also offers HIs for downstream prognostics, but it is trained only on data deemed “healthy” through novelty detection and thus fails to capture the full degradation progression.

Reg.) achieves competitive performance, sometimes surpassing specialized deep learning models (e. p. 7
Where the paper says it

Notably, a simple linear regressor trained on the latent compass coordinates (L. ) achieves competitive performance, sometimes surpassing specialized deep learning models (e.

g., RVE, VAE+RNN)

Geometric Interpretation of Latent Spaces

The latent manifolds learned by each method reveal a clear progression in structure and interpretability (Fig. 2) STEP, in contrast, produces a structured manifold: with window size w=1, six crisp rings emerge (one per OC); θ advances monotonically with deterioration, forming a continuous SOH proxy; r indicates the active OC

θ encodes state, not elapsed time

Predictive gain: a downstream transformer trained on t/T alone collapses to RMSE 58.1 (R2=−1.75) on FD002, while training on θ alone achieves RMSE 14. p. 6

Where the paper says it

1 (R2=−1. 75) on FD002, while training on θ alone achieves RMSE 14.

1 (R2=0.84); concatenating t/T to θ does not help (Appendix Fig. p. 6
Where the paper says it

75) on FD002, while training on θ alone achieves RMSE 14. 84); concatenating t/T to θ does not help (Appendix Fig.

5), evidence that θ already encodes the state

RUL Estimation Performance

Table 9 shows our method achieves SOTA on FD002 (12.34 RMSE) and FD004 (12.10 RMSE) and competitive performance on the other subsets Among the compared methods, only I-GLIDE also offers HIs for downstream prognostics, but it is trained only on data deemed “healthy” through novelty detection and thus fails to capture the full degradation progression

Forecasting Performance

Fig. 4 shows that STEP outperforms vanilla representations (AE) and the general-time-series SOTA (SoftCLT) on forecasting of high-dimensional, noisy mouse brain data Performance degrades on longer horizons compared to Latent SDE models (SING-GP, VDP-GP), warranting future work to reduce prediction drift

The non-regression assumption concerns the latent health state, not noisy sensors The instability previously observed at dim(z)=4 is resolved via β-scheduling, and results are stable across dim(z) ∈ 2, 4, 8, 16, 32 with no collapse Even with a single zend prototype, unsupervised K-means (k=2) on terminal embeddings recovers distinct failure modes on FD003 (1 OC, 2 FMs) and FD004 (6 OCs, 2 FMs) with high silhouette scores

Table 9 shows our method achieves SOTA on FD002 (12.34 RMSE) and FD004 (12.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Geometric Interpretation of Latent Space: The system can read a latent compass, which is described as the polar coordinates (theta, r) of the latent vector, in which theta tracks the progression of the underlying state (e.g., from healthy to failed) and r identifies the active mode. This allows for reading a latent compass that tracks progression while providing grounded coordinates for interpretability.

  2. State-Specific Health Indicator Discovery: The system can discover underlying mechanisms by deriving formula-based features, specifically the angular position theta, which serves as a smooth, full-trajectory proxy for SOH. This feature allows for the discovery of distinct failure modes and is validated by showing that a simple linear regressor on top of the latent compass coordinates is competitive with deep architectures.

  3. Robustness to Noise and Scale: The system exhibits graceful degradation when sensor noise is scaled, as seen in Table 7, confirming that the cosine triplet loss operates on distributional proximity and that STEP degrades gracefully under amplified noise. Furthermore, interpretability scales with dimension, as interpretability scales with latent dimension (full diagnostics in Appendix G).

  4. Phase Localization and Mode Separation: The system can recover interpretable task phases from continuous data, such as locating or grabbing in robotics, where the unit-normalized view z bar projects trajectories onto the unit circle, exposing the angular progression theta t and the residual radius r as polar coordinates. This enables action-phase localization and soft phase labeling without manual annotation.

  5. Decoupling of State from Time: The system can distinguish between state progression and elapsed time because theta already encodes the state rather than time, as evidenced by a downstream transformer trained on theta alone achieving R2=0.84 on FD002, whereas training on t/T alone yields a significantly lower score.

Abstract

We present a novel method for learning interpretable representations of progressive time series, that is, data capturing irreversible state transitions such as degradation or task completion. Our approach uses a self-supervised contrastive objective to learn a low-dimensional latent space whose geometry is itself the interpretation: each observation becomes a point on a manifold anchored between two fixed orthogonal prototype vectors, and a trajectory becomes a path across that manifold. From this structure we read a latent compass, the polar coordinates (θ, r) of the latent vector, in which θ tracks the progression of the underlying state (e.g., from healthy to failed) and r identifies the active mode (e.g., the operating condition), without any proxy labels. We evaluate the approach against the state of the art on diverse domains, including industrial degradation, robotic tasks, and neural activity, validating three key capabilities: (1) end-state prediction, (2) multi-step forecasting, and (3) interpretable phase separation. Our method matches or improves over black-box counterparts on all of these while providing transparency about the underlying mechanisms. A simple linear regressor on top of the latent compass coordinates is competitive with deep architectures, direct quantitative evidence that the underlying state is encoded in a geometrically accessible form.

Sources

Related papers