💻 computer science

Algebraic Prognostics via fr-Codes: Tracking Truncated Factor Dimensions in Turbofan Degradation

This paper proposes a training-free, algebraic prognostic framework using fr-codes and finitely presented groups to track turbofan engine degradation by quantizing sensor data into discrete states and monitoring drops in the dimension of a truncated factor algebra as a deterministic indicator for remaining useful life prediction.

Original authors: MAKSIM KHOTINSKIY

Published 2026-09-04

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: MAKSIM KHOTINSKIY

Original paper licensed under CC BY 4.0 (https://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

Keeping a jet engine healthy is a high-stakes challenge for anyone who relies on air travel. Engineers constantly monitor these massive machines, listening to the hum of twenty-one different sensors that track temperature, pressure, and rotation speed. The goal is to predict when an engine will fail so it can be fixed before it breaks, a practice known as prognostics. For years, the standard approach has been to feed vast amounts of historical data into complex computer programs called deep learning models. These programs learn to recognize patterns that precede a breakdown, but they require massive amounts of labeled examples, need to be retrained whenever conditions change, and often act as "black boxes" where the internal logic remains a mystery. A new approach, however, seeks to solve this problem without learning from examples at all. Instead of training a computer to recognize failure, this method treats the engine's data as a set of rules that build a mathematical structure, watching how that structure changes as the machine wears down.

The researchers behind this study, Maksim Khotinsky, focused on a specific dataset of simulated jet engines that run until they fail. Rather than feeding the raw sensor numbers into a neural network, they translated the engine's behavior into a simple language of states. Imagine the engine's condition at any given moment as a snapshot of all twenty-one sensors. If a sensor reading is steady, it gets a zero; if it rises, a one; if it falls, a two. This turns a complex, continuous stream of data into a sequence of simple, discrete steps. As the engine runs, it moves from one state to another. The researchers treated these movements as a story being told. Every time the engine moved from one specific state to another, they recorded it. If a particular move happened only once, it might have been a fluke or a glitch. But if the engine repeated that same move, the researchers treated it as a rule, a permanent law of that engine's current behavior.

By collecting these repeated moves, the team built a growing mathematical object, a kind of group, that represented the engine's history. At the start of the engine's life, there were very few rules because the machine was exploring many new paths. As the engine aged and began to degrade, its movements became more restricted and repetitive. New rules were added to the mathematical group, but these new rules often conflicted with or depended on the old ones. The researchers then applied a specific algebraic test to this group, a method borrowed from a branch of mathematics that studies how shapes and structures relate to one another. They measured the "dimension" of the structure, which essentially counts how many independent directions or possibilities the engine still had left.

The key discovery was that as the engine got closer to failure, this dimension began to drop. It was not a smooth, steady decline, but a series of sudden steps down. Each time a new rule was added that made the engine's behavior more constrained, the dimension fell. The researchers counted these drops. They found a clear pattern: the more times the dimension dropped, the less life the engine had left. Across hundreds of simulated engines, they observed a consistent relationship where a higher number of drops signaled a shorter remaining lifespan. This allowed them to create a simple health indicator that required no training data, no labels for what kind of failure was happening, and no complex feature engineering. They could place an engine into one of five zones, ranging from "Normal" to "Emergency," simply by counting how many times this mathematical dimension had shrunk.

The results were promising but not perfect. The method successfully identified that engines with more drops were closer to failure, with a correlation that was statistically significant. However, the researchers were careful to note that this approach did not outperform the simplest possible baseline, which was just counting how many unique states the engine had visited. A simple count of unique states actually predicted the remaining life slightly better than the complex algebraic drops. The value of this new method, the author suggests, lies in its ability to see the structure of the engine's behavior rather than just the volume of its activity. It offers a way to understand how the engine is failing, not just when. While the current accuracy is not yet good enough to replace the high-performance deep learning models used in industry, the method provides a transparent, training-free tool that works without needing to know the specific type of failure in advance. It is a proof of concept that the language of algebra can describe the physical decay of a machine, offering a new lens through which to view the slow, inevitable wear of a jet engine.

Technical Summary: Algebraic Prognostics via fr-Codes

Problem Statement
Predicting the Remaining Useful Life (RUL) of turbofan engines is a critical challenge in prognostics and health management. While state-of-the-art approaches utilizing deep learning (LSTMs, transformers, CNNs) achieve mean absolute errors (MAE) of 15–25 cycles on the NASA CMAPSS dataset, they rely heavily on large volumes of labeled failure data, require retraining for new failure modes, and lack structural interpretability. This paper proposes a fundamentally different, training-free approach that avoids learning a direct mapping from sensor readings to RUL. Instead, it constructs an algebraic object—a finitely presented group—from sensor data and tracks the evolution of a structural statistic derived from this object as the engine degrades.

Methodology
The proposed method, termed "Algebraic Prognostics," relies on the theory of fr-codes developed by Ivanov, Mikhailov, and Pavutnitskiy. The core workflow involves three stages:

Data Quantization and State Construction:
Multivariate sensor data (21 sensors in the CMAPSS FD002 subset) are converted into a discrete sequence of state vectors. At each cycle $t$ , the change in each sensor reading is quantized into a ternary alphabet $\{0, 1, 2\}$ representing steady, increasing, or decreasing trends. The global state $\sigma_t$ is the Cartesian product of these ternary values across all sensors.
Construction of an Expanding Filtration:
A sequence of group presentations $G_t = \langle V_t | R_t \rangle$ is built as the engine operates.
- Generators ( $V_t$ ): Each unique state vector observed up to cycle $t$ becomes a generator.
- Relations ( $R_t$ ): Transitions between consecutive states ( $\sigma_{t-1} \to \sigma_t$ ) are recorded. A transition is promoted to a defining relation only after it has been observed at least twice, filtering out noise.
- Filtration: As cycles accumulate, the set of relations $R_t$ grows strictly monotonically, creating a filtration $G_1 \twoheadrightarrow G_2 \twoheadrightarrow \dots \twoheadrightarrow G_T$ .
Algebraic Computation (Truncated Factor Dimensions):
The method computes a structural statistic $S_c(G_t)$ based on the fr-code dictionary.
- Magnus Expansion: Relations are mapped into the truncated free algebra via the Magnus expansion (truncated at degree $d=3$ ).
- Boundary Matrices: For specific fr-codes (combinations of the augmentation ideal $f$ and relation ideal $r$ ), boundary matrices $M_c$ are constructed. The paper focuses on the code $c = ffr + rff$.
- Factor Dimension: The statistic is defined as $S_c(G_t) = \dim(f) - \text{rank}(M_c)$ .
- Degradation Indicator: As new relations are added, the rank of the matrix may increase, causing $S_c(G_t)$ to decrease. The cumulative count of these decreases ("factor dimension drops") serves as the prognostic indicator.

Key Contributions

Training-Free Structural Indicator: The paper introduces a method that requires no supervised training, no failure mode labels, and no feature engineering beyond basic quantization.
Algebraic Filtration from Time Series: It establishes a procedure for converting multivariate time series into a filtered sequence of group presentations, bridging discrete sensor dynamics with algebraic topology.
Pragmatic Approximation of Derived Limits: While the full theory of fr-codes involves computationally intractable derived limits over a category of presentations, this work proposes a pragmatic approximation: evaluating the functor on a single, data-induced presentation. The author explicitly distinguishes this statistic ( $S_c$ ) from the exact homological invariant ( $\lim^1$ ).
Five-Zone Health Indicator: Based on the cumulative drop count, the author defines a five-zone operational status (Normal, Warning, Critical, Severe, Emergency) to guide maintenance actions.

Results
The method was evaluated on the NASA CMAPSS FD002 dataset (260 training engines, 259 test engines; 244 test engines yielded sufficient relations for analysis).

Correlation: The cumulative number of factor dimension drops for the code $ffr + rff$ correlates with the true RUL with a Pearson coefficient of $r = -0.53$ ( $p < 0.001$ ). The relationship is monotonic: more drops indicate less remaining life.
Predictive Accuracy: A linear predictor based on drop counts achieved a Mean Absolute Error (MAE) of 38 cycles on the test set. This is lower than the 15–25 cycles typical of deep learning methods but superior to a baseline (predicting the mean RUL, MAE 45).
Comparison with Baselines: The raw count of distinct states ( $K$ ) alone showed a stronger correlation with RUL ( $r = -0.68$ ) than the algebraic drop count. However, within sub-cohorts of engines with identical state counts, the factor dimension provided additional structural information, correlating with RUL at $r = 0.86$ for a specific cohort ( $K=38$ ).
Zonal Separation: The method successfully separated engines into distinct health zones, with a mean RUL difference of 87 cycles between the "Normal" and "Severe" zones.

Significance and Claims
The paper positions its contribution as an initial exploration of a theoretical framework rather than a replacement for high-accuracy deep learning models.

Interpretability: The primary value lies in the method's structural interpretability. The "factor dimension drops" are rigorously defined algebraic events representing the introduction of linear dependencies in the truncated Magnus expansion, corresponding to physical constraints on the engine's state space as it degrades.
Robustness: The approach is robust to operating condition offsets because the quantization operates on first differences, neutralizing baseline shifts.
Modest Claims: The author explicitly acknowledges that the current predictive accuracy (MAE 38) does not match state-of-the-art supervised methods. They note that the correlation is moderate ( $r^2 \approx 0.28$ ) and that the exact timing of the final drop is highly variable.
Future Directions: The paper identifies the computation of the full derived limit via cosimplicial resolution as the necessary next step to validate whether the single-presentation approximation converges to the true homological invariants. It also suggests future work on higher Magnus degrees and hybrid frameworks with continuous Topological Data Analysis (TDA).

In conclusion, the paper demonstrates that algebraic invariants derived from fr-codes can serve as a viable, training-free structural health indicator for turbofan engines, capturing degradation trends through the evolution of group presentation relations, even if the raw predictive power is currently modest compared to deep learning.

Technical Summary: Algebraic Prognostics via fr-Codes

More like this