Physics-Constrained Autoencoder for Interpretable Health Representation Learning in Turbofan Engines
Authors: Leonardo Baldo, Jokin Alcibar Ibarzabal, Daniel Cisneros Acevedo, Kai Goebel, Manuel Arias Chao
Organizations: Institute for Data Science, Zurich University of Applied Science, Winterthur, 8401, Switzerland · Computer Science Department, Electronics, Mondragon University, Loramendi, 4, Mondragón, 20500, Spain · Faculty of Aerospace Engineering, Delft University of Technology, HS 2926 Delft, The Netherlands · Fragum Global, Mountain View, CA 94040, USA · Luleå University of Technology, Luleå, 971 87, Sweden
Health Indices (HIs) summarize degradation from operational data and support diagnostic and prognostic tasks in Prognostics and Health Management (PHM). Their usefulness depends on capturing degradation and interpreting system condition. Autoencoders learn such representations from unlabelled monitoring data; however, when training is driven only by reconstruction, the physical meaning of latent variables remains undefined, limiting their interpretation and diagnostic value. This work proposes a physics-constrained autoencoder that embeds system knowledge as an inductive bias in the decoder architecture and as a learning bias in the loss function. The decoder contains differentiable physical equations that receive latent variables and operating conditions as inputs, structurally guiding each latent variable to represent a predefined physical quantity and enabling component-level analysis before aggregation into a system-level HI.The methodology is benchmarked on the N-CMAPSS dataset against alternative representations. Reconstruction reaches an R2 of 0.9993 and an MAE of 0.0036, comparable to fully data-driven models. After operating-condition correction, the inferred physical quantities track degradation, and their aggregation produces an HI with a Spearman correlation of 0.9791 with the ground truth, an RMSE of 0.0286, and an MAE of 0.0233. The methodology is further evaluated using a single model trained across four datasets representing distinct degradation scenarios. This model preserves component-level signatures that distinguish compression-side from expansion-side degradation without diagnostic labels.Finally, the proposed framework enables degradation trends to be visualized on component maps, providing a health representation in a form familiar to engine operators and domain experts.
Figures & tables
Method family
HI construction and knowledge integration
Interpretability and scope
Representative refs.
Expert-designed HIs
Statistical or signal-processing quantities selected from known degradation signatures.
Explicit meaning, but dependent on operating conditions, failure mode, and feature selection.
Lei et al. (2018) ; Zhou et al. (2022)
Model- and residual-based HIs
Residuals or estimated parameters based in regression based models. Knowledge enters through the model structure and reference behaviour.
Interpretation follows from the model and residual definition; performance depends on reference quality and model fidelity.
Jardine et al. (2006) ; Coble and Hines (2009)
Autoencoder residual HIs
Input or latent reconstruction errors or sensor reconstruction losses, possibly conditioned.
Quantify departure from learned behaviour without identifying the physical source of the change.
González-Muñiz et al. (2022) ; de Pater and Mitici (2023)
Knowledge-constrained learned HIs
Latent representations or HI outputs constrained by causal structure, monotonicity, temporal ordering, or boundary conditions.
Degradation-consistent trajectories; meaning depends on the imposed constraints.
Bajarunas et al. (2023) ; Bajarunas et al. (2024) ; Tefera et al. (2025)
Hybrid physics-decoder HIs
A scalar latent space with predefined meaning enters physics-derived decoder terms, possibly combined by later trainable layers.
Provides a system-level physical interpretation, but estimates a single HI and requires a system-specific model.
Huang et al. (2024) ; Plonczak et al. (2026)
Proposed component-resolved physics-decoder AE
A latent vector modifies specific quantities within coupled equations that directly reconstruct multiple measurements.
Coordinates are structurally associated with component and performance degradation effects; interpretability depends on model fidelity and latent identifiability/observability.
This work
Table 1: Representative HI construction strategies and positioning of the proposed component-resolved physics-decoder autoencoder.
Figure 1: Generic architecture of the proposed physics-constrained autoencoder. The encoder maps the normalized measurements xt into a structured latent representation zt . Each latent variable is associated with a predefined physical quantity and is propagated, together with the operating conditions wt , through the differentiable physical decoder to reconstruct the measured variables x^t .
Figure 2: Implementation of the physics-constrained autoencoder for the N-CMAPSS turbofan application. The 15 -dimensional latent space represents physically assigned engine parameters, which are combined with the operating conditions in the physical decoder. Model training minimizes the reconstruction error together with the residuals associated with the turbine shaft-power balances.
Figure 3: Schematic of the physics-based turbofan decoder and its correspondence with the N-CMAPSS variables and latent health modifiers. The engine stations and component decomposition are shown together with the measured N-CMAPSS variables used as model inputs and reconstruction targets, highlighted in red. Component colors identify the fan, LPC, HPC, compressor, HPT, and LPT. The latent variables z parameterize component-level health modifiers associated with efficiencies, pressure ratios, flow capacity, and selected thermodynamic quantities, while the operating conditions enter through the free-stream variables.
Category
Symbol
Meaning
Symbol
Meaning
Symbol
Meaning
Operating conditions
H
Altitude
M
Flight Mach number
T2
Fan-inlet total temperature
TRA
Throttle-resolver angle
TRA
Normalised throttle setting
Atmospheric quantities
Tamb
Ambient static temperature
Pamb
Ambient static pressure
PSL
Sea-level pressure
P0
Freestream total pressure
kH
ISA altitude coefficient
R
Specific gas constant
Gas properties
γ
Specific-heat ratio, air
γt
Specific-heat ratio, turbine gas
cp
Specific heat, air
cp,t
Specific heat, turbine gas
cp,mix
Mixed-stream specific heat
MFP
Mass-flow parameter
Table 2: Symbols and physical quantities used in the turbofan decoder.
Figure 4: Inference and health-assessment workflow of the proposed framework. The inferred physical latent variables are used to investigate the latent space and construct the system-level HI after removal of the operating-condition contribution. In parallel, the inferred turbine quantities are used to evaluate degradation-induced changes in the component operating maps.
Figure 5: Measured and reconstructed trajectories for the 12 sensor variables obtained with the proposed physics-constrained autoencoder. The reported R2 values are computed over all test samples.
Variable
R2
MAE
Variable
R2
MAE
T24
0.99920±0.00018
0.00310±0.00044
P21
0.99842±0.00008
0.00750±0.00015
T30
0.99897±0.00039
0.00371±0.00069
P24
0.99976±0.00006
0.00208±0.00026
T48
0.99884±0.00017
0.00306±0.00022
Ps30
0.99938±0.00022
0.00311±0.00065
T50
0.99785±0.00095
0.00323±0.00071
P40
0.99961±0.00012
0.00241±0.00031
P15
0.99900±0.00005
0.00598±0.00015
P50
0.99987±0.00007
0.00189±0.00047
P2
0.99986±0.00000
0.00277±0.00000
Wf
0.99838±0.00037
0.00395±0.00065
Table 3: Per-variable reconstruction performance of the proposed physics-constrained autoencoder on the test set. Results are reported as mean ± half-range.
Figure 6: Latent trajectories before and after operating-condition removal for Unit 14 . Grey lines indicate the original latent variables, blue lines their corrected residuals, and the green dashed line the engine ground-truth HI. The panel titles report the correlation before and after the correction.
Figure 7: Cross-engine evolution of the inferred physical modifiers against normalised life fraction, over the six DS03 test engines of five independently seeded trainings (30 trajectories). Grey curves are the individual traces, the pale band their total range, the darker band the range across the five per-seed means, and the heavy curve the pooled mean. Shading marks the two coordinates degraded in DS03.
Method
Extrapolation
W -fit R2
Spearman
HI RMSE
Composite
Fit time
No removal
–
–
0.6684±0.1345
0.1833±0.0845
0.5601±0.1220
–
Linear ridge
Yes
0.9551±0.0335
0.9409±0.0219
0.0310±0.0054
0.8001±0.0141
0.8s
Degree- 2 polynomial ridge
Yes
0.9980±0.0009
0.9775±0.0029
0.0265±0.0036
0.8436±0.0106
0.9s
Degree- 3 polynomial ridge
Yes
0.9991±0.0003
0.9791±0.0059
0.0286±0.0033
0.8555±0.0133
1.1±0.1s
MLP regressor
Yes
0.9983±0.0006
0.9609±0.0035
0.0271±0.0041
0.8358±0.0191
85.2±9.7s
Table 4: Comparison of the methods used to remove the operating-condition contribution from the latent representation. All operating-condition models are fitted using healthy samples only. For each method, four coordinates are selected and weighted using development trajectories, and the resulting HI is evaluated on the test engines. Fit times are retained from the original measurements of the unchanged fitting procedure. Each method was evaluated over five independent runs using different random seeds; results are reported as mean ± half-range over five seeds. The best and second-best results in each column are shown in bold and underlined, respectively.
Quantity
Value
Quantity
Value
Monotonicity
0.6272±0.0503
Trendability
0.9879±0.0062
Prognosability
0.9998±0.0005
Mutual information
0.8072±0.0134
Composite score
0.8555±0.0133
Spearman with GT
0.9791±0.0059
Pearson with GT
0.9958±0.0016
MAE
0.0233±0.0028
RMSE
0.0286±0.0033
Table 5: Health-index construction and test performance for the proposed physics-constrained representation, with coordinate orientation, selection, and fusion weights determined on development trajectories. Results are reported as mean ± half-range over five seeds.
Figure 8: Overview of the benchmarking framework used to compare the proposed physics-constrained representation with the considered data-driven and reduced-order baselines.
Figure 9: Channel-wise reconstruction performance obtained by the considered representations over all test samples, reported in terms of (a) coefficient of determination R2 and (b) mean absolute error (MAE).
Representation
k
Recon. R2
Recon. MAE
Spearman
HI RMSE
Composite
PCA, 1 component
1
0.83108
0.05527
0.9643
0.0286
0.8266
Data-driven AE, 1 -D
1
0.99928±0.00000
0.00208±0.00001
0.7545±0.3664
0.1300±0.1691
0.6758±0.2291
Data-driven AE, 15 -D
4
0.99928±0.00000
0.00208±0.00001
0.9525±0.0389
0.0305±0.0172
0.8392±0.0599
Physics AE, bottleneck 1
1
0.94811±0.00074
0.02745±0.00028
0.9428±0.0170
0.0398±0.0077
0.8148±0.0147
Physics AE, 15 -D (proposed)
4
0.99927±0.00006
0.00357±0.00015
0.9791±0.0059
0.0286±0.0033
0.8555±0.0133
Supervised baseline †
–
–
–
0.8530
0.0820
0.6114
Table 6: Comparison of the considered representations on the N-CMAPSS DS03 test set. Reconstruction metrics are computed over all test samples, while HI metrics are computed over the 438 flight cycles. The column k indicates the number of latent coordinates retained for HI construction. For one-dimensional representations, k=1 by construction; k is not applicable to the supervised baseline. For all unsupervised representations, coordinate orientation and, where applicable, coordinate selection and fusion weights are determined on development trajectories and applied unchanged to the test engines. Results with spread are reported as mean ± half-range over five seeds. The best and second-best results in each column are shown in bold and underlined, respectively.
Figure 10: Estimated HIs and ground-truth HI for the six test engines. The grey regions identify the healthy reference intervals, while the dotted lines indicate the degradation-onset cycles. Almost all methodologies show a similar trend, except for the supervised model in yellow and the data-driven autoencoder with 1 dimension in the latent space. This justifies the goodness of the physics decoder while providing an explicitly physics-structured latent representation.
Figure 11: Relative changes in the five inferred efficiency-related coordinates for DS03. Thin blue curves show individual test engines, and the thick blue curve shows their mean. Dashed red curves show the imposed HPT- and LPT-efficiency degradation. No degradation is imposed on the fan, LPC, or HPC.
Training set
Test R2
Test MAE
Jacobian rank
HP shaft residual
DS03 (5 seeds)
0.99927±0.00006
0.00357±0.00015
9
0.00111±0.00028
DS04
0.99930
0.00339
9
0.00116
DS05
0.99932
0.00374
9
0.00080
DS06
0.99907
0.00383
9
0.00094
DS03–DS06 pooled (5 seeds)
0.99912±0.00009
0.00383±0.00024
9
0.00093±0.00023
Table 7: Reconstruction on the test engines for single-scenario models and the pooled model. The DS03 and pooled entries are reported as mean ± half-range over five seeds; the remaining entries are single runs.
Figure 12: Degradation signatures of the 18 test engines from the pooled DS03–DS06 model. (a) Normalized changes in the five efficiency-related coordinates, with individual engine profiles and regional means with min–max bands. (b) Score difference Scomp−Sturb , correctly assigning all engines to the compression or expansion side. (c) PCA projection of the same signatures, with labels added after projection. Colour denotes the degradation region and marker shape denotes the source dataset.
Figure 13: Degradation-induced evolution of the inferred HPT and LPT operating lines in turbine-map coordinates for three representative units with a single seed. Grey points show the complete set of inferred operating points, while coloured lines represent the corrected-flow–expansion-ratio relation at increasing life fractions. The progressive upward displacement indicates an increase in inferred corrected flow at a given expansion ratio with degradation. The percentages reported in each panel quantify the corresponding increase in corrected flow. These representations describe the evolution of operating lines in component-map coordinates and do not constitute complete turbine characteristic maps.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Variable
Our model
GasTurb
TurboGrad
Variable
Our model
GasTurb
TurboGrad
ηfan
→
↓
↓
ΠHPC
↓
↑
↑
ηLPC
↓
↓
↓
ΠHPT
↓
↓
↓
ηHPC
↑
↑
→
ΠLPT
↓
↑
↓
ηHPT
↓
↓
↓
Wmod
↑
↑
↑
ηLPT
↓
↓
↓
BPR
→
→
↑
Πfan
↑
↑
↑
T41
↑
↑
↑
Appendix
Table 8: Qualitative comparison of the degradation-induced evolution of the physical quantities inferred by the proposed model and obtained with GasTurb and TurboGrad. ↑ , ↓ , and → indicate an increase, decrease, and no clear systematic variation, respectively.
Figure 14: Sensitivity of the latent coordinates and effect of retaining the discarded combustor cooling fraction. (a) Norm of each coordinate’s column of the decoder Jacobian, in normalized sensor units per unit of latent change, on a logarithmic axis. The retained coordinates are shown in blue and βcomb in red. (b) Reconstruction R2 and HI composite score for the adopted 15-dimensional representation and the 16-dimensional representation retaining βcomb . Bars show means over five independently seeded training runs, error bars show the spread across seeds, and dots show individual runs. Reintroducing the coordinate provides no improvement in either metric.
Coordinate
∥∂x^/∂zk∥
∥∂L/∂zk∥
uk
Status
βcomb
3.77×10−5
7.86×10−18
–
removed
βLPT
1.80×10−3
1.63×10−10
0.998
kept
βHPT
9.52×10−3
1.04×10−9
0.993
kept
ηLPC
1.49×10−2
7.23×10−10
0.646
kept
ηfan
2.03×10−2
9.73×10−10
0.354
kept
BPR
2.13×10−2
1.81×10−9
0.669
kept
Appendix
Table 9: Sensitivity and local identifiability of the latent coordinates. The sensitivity and loss-gradient norms refer to the initial 16-dimensional representation. The null-space share uk refers to the adopted 15-dimensional representation and is therefore not defined for the removed coordinate. It quantifies the squared projection of each coordinate onto the null space of the evaluated mean Jacobian.
Variable
Initial
Calibrated
Variable
Initial
Calibrated
T24
−160.8
0.312
P24
−121.4
0.775
T30
−45.8
−0.218
Ps30
−32.9
0.324
T48
−0.661
0.051
P40
−33.2
0.331
T50
−78.5
−1.007
P50
−11.8
0.465
P15
0.486
0.919
Wf
0.167
0.175
P2
0.9998
0.9998
P21
0.36945
0.91920
Appendix
Table 10: Reconstruction R2 of the nominal physical model before and after calibration on healthy development samples.
Parameter
Initial
Calibrated
Bounds
At bound
Πfan0
1.500
1.306
[1.10,2.20]
ΠLPC0
3.500
1.222
[1.05,5.00]
ΠHPC0
11.429
15.594
[5.00,25.00]
BPR0
9.000
10.235
[3.00,15.00]
ηfan0
0.880
0.918
[0.60,0.95]
ηLPC0
0.880
0.863
[0.60,0.95]
Appendix
Table 11: Initial and calibrated nominal-engine parameters. Parameters marked as bounded lie at, or within numerical tolerance of, one of the prescribed optimization bounds.
N-CMAPSS variable
Decoder quantity
Physical interpretation
T24
T27
LPC outlet total temperature
T30
T3
HPC outlet total temperature
T48
T44
HPT outlet total temperature after HPT cooling remix
T50
T5
LPT outlet temperature after downstream flow remix
P15
P17
Bypass-duct total pressure
P2
P2
Fan-inlet total pressure
Appendix
Table 12: Correspondence between the N-CMAPSS measurement labels and the internal quantities used by the physical decoder.
Autoencoders learn compressed representations by encoding data into latent codes and decoding them for reconstruction, generation, denoising, or anomaly detection.