From Failure Prediction to Component Selection in Semantic Segmentation

Minhui Kang; Yiyao Zhang; Peizhe Long; Jiafeng Qi; Hongyu Jin; Jiayun Zheng

cs.CVOct 12, 2026

From Failure Prediction to Component Selection in Semantic Segmentation

Authors: Minhui Kang, Yiyao Zhang, Peizhe Long, Jiafeng Qi, Hongyu Jin, Jiayun Zheng

Organizations: School of Power Engineering, Jiangxi Electric Vocational and Technical College, Nanchang, Jiangxi, China · College of Information Science and Engineering, Shandong Agricultural University, Tai’an, Shandong, China · College of Aviation Manufacturing Industry, Nanchang Hangkong University, Jingdezhen, Jiangxi, China

Abstract

Reliable scene understanding requires deciding which predictions to trust, retain, or inspect. In semantic segmentation, component failure predictors support these decisions, but their predictive performance does not fully describe the value of the resulting selection. We develop a decision-centered evaluation framework that connects three questions: what an available score predicts, what can be learned from its existing fields, and what its ranking delivers. Across three segmentation systems on Cityscapes and ACDC at 90% retention, adapted MetaSeg reduces retained failure relative to forecast entropy in all twelve architecture-population comparisons, but does not establish a failure advantage over direct probability ranking. Image-nested source experiments reveal a different opportunity: reusing the full field set improves predictive and retained-loss performance under both learners in all three systems, while the raw OCRNet probe improves prediction with unresolved selective gains. Mapping a fixed prediction pair to conditional log-loss risk produces mixed selection effects, which we examine through the components exchanged by the two rankings. These findings distinguish the performance of a fitted score from the value obtainable from its inputs and the consequences of its use. The framework provides a practical basis for deciding whether to adopt an additional score, relearn existing fields, or revise a selection rule, and for explaining the gains and failures of component-level reliability methods.

Figures & tables

Figure 1: Three evidence questions for component reliability. Predict evaluates forecasts; Realize tests added field information; Select compares retained sets under the same budget and loss. Arrows connect the shared component record. The Select example compares native $m$ with mapped $r_{\mathrm{LL}}(m,q)$ , evaluating fixed $p$ . Gray/orange/blue masks denote shared/native-only/mapped-only members; these unordered subsets use independent display scales.

Figure 2: Frozen DeepLab components and failure forecasts in ten Cityscapes/ACDC scenes. Rows show RGB, predicted classes, base failure probabilities $p$ and adapted MetaSeg probabilities $m$ . Crops align within columns; probability maps share the $[0,1]$ scale. Gray marks excluded components under complete-image eligibility; yellow outlines mark predicted components.

Panel of Figure 3 — Figure 3: Probe construction and matched source controls. (a) Matched learners produce diagnostic forecasts; CE evaluates prediction and mapped-risk ranking evaluates selection using fixed $p$ . (b) Training calibration losses supply the ridge target; raw $b$ is a source field. (c) Image-nested inputs exclude $V$ labels from fitting. Thumbnails are archived P0 examples; other objects are schematic.

Figure 4: MetaSeg gains against entropy (blue) and direct probability (peach) at 90% all-valid retention. Rows denote architectures; columns show failure and forecast-loss gains. Positive favors MetaSeg; bars show pointwise 95% paired-image intervals. Exact values: Supplement S4, Table S8.

SegFormer
Population	Images	$N$	AUROC $p\uparrow$	AUROC $m\uparrow$	CE $p\downarrow$	CE $m\downarrow$
P0	300	18,728	0.8105	0.7157	0.4993	0.6076
P1	200	12,209	0.8115	0.7361	0.5121	0.5927
P2	406	32,429	0.7562	0.7090	0.5834	0.6825
P3	400	29,614	0.7403	0.6973	0.5953	0.6880
DeepLab

Table 1: Absolute levels on all outcome-valid components. (a) AUROC pools components; CE is image-macro in nats;

N

counts eligible components. (b) Retained failure (%) and forecast LL (nats) at 90%; LL evaluates fixed

p

. P0: source out-of-fold; P1/P2/P3: source-development/adverse-development/final populations.

Figure 5: Predictive and selective gains from G4 (all fields) and G2 (raw probe). Peach/blue denote learned

p

-only (G0) and dimension-matched

p

-derived controls (PC); circles/squares denote D1-R/D2-R. Positive favors added fields; bars show pointwise 95% paired-image intervals. CE evaluates diagnostic forecasts; retained LL evaluates fixed

p

. Full values: Supplement S3, Tables S4–S5.

Figure 6: Native-minus-mapped gains at 90% retention; positive favors mapping. Filled blue circles denote all-valid components; open peach circles denote nonempty interiors, each with its own eligible set and matched budget. Bars show pointwise 95% paired-image intervals. Forecast LL evaluates fixed

p

Panel of Figure 7 — Figure 7: Exchanged supports in six fixed DeepLab scenes at 90% all-valid retention. Orange: native-only; blue: mapped-only. Crosses mark centroids. Paired labels locate mask insets, magnified independently. Selection uses the full image. Descriptive outcomes: Table 2 ; scene and inset rules: Supplement S5.

Scene	$n$	$k$	$x$	Native	Mapped	Native	Mapped
				LL [nats]		Failures/ $k$
Source 1	118	107	10	0.381016	0.397064	19/107	23/107
Source 2	91	82	9	0.345995	0.325607	12/82	16/82
Fog	116	105	9	0.543883	0.575054	63/105	61/105
Night	241	217	22	0.541156	0.533147	164/217	169/217
Rain	178	161	17	0.535679	0.501192	83/161	89/161

Table 2: Descriptive outcomes for the six scenes in Figure 7 at 90% all-valid retention.

n

: eligible count;

k

: retained count;

x

: removed and added count. LL is mean retained loss of fixed

p

(nats); failures are counts over

k

Figure 8: Exclusive-component losses in Figure 7 a: ten members per group, with failures filled. Boxes show medians and interquartile ranges; whiskers span minimum–maximum. Groups are unpaired and describe one image. Component identities: Supplement S5.

Figure 9: Native-minus-mapped gains across all-valid retention budgets; positive favors mapping. Columns show failure gain (percentage points) and LL gain (

10^{-3}

nats). Filled markers denote 90%; lines connect archived estimates. Colors denote P0–P3. Primary intervals: Figure 6 .

Appendix figures & tables25 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 12 — Figure S1: Component identities and probe output for main Figure 2a, with a wider crop and the same complete-image eligibility. (a) Repeating categorical colors encode identities, not classes or confidence; yellow marks component 121. (b) Source-standardized probe $b$ has a signed scale, not probability semantics. Gray marks excluded components.

1	Input the original five image folds and fixed upstream/downstream learner definitions.
2	For each outer validation fold $V$ , set $T$ to the other four folds.
3	For each image fold $U\subset T$ , fit upstream F2/F3 on $T\setminus U$ , including preprocessing, and predict $U$ .
4	Join the inner-held-out rows to form downstream training features on $T$ . All fits exclude labels from $V$ .
5	Fit upstream F2/F3 on all $T$ and generate outer-validation features on $V$ .
6	Fit each diagnostic’s transformations and learner on the cross-fitted rows in $T$ , using its specified image weights.
7	Apply the frozen transformations and predictors to $V$ .

Appendix

Table 13

Index	Field	Formula	Unit
1	area	size(K)	pixels
2	boundary_size	size(B)	pixels
3	interior_size	size(I)	pixels
4	mean_max_probability	mean_K u	probability
5	mean_entropy	mean_K h	nats
6	mean_probability_margin	mean_K g	probability difference

Appendix

Table S2: Ordered numerical fields of the base failure predictor. The nineteen class indicators follow these eighteen fields.

Object	Features	Target	Learner
Base forecast $p$ (F2)	18 numeric + 19 indicators	Y	fixed L1 liblinear
Frozen score F3	11 descriptors	global-minus-image affine opportunity	robust-scaled ridge
Descriptors (G3)	p + raw 11 descriptors	Y	D1-R / D2-R
All fields (G4)	p + raw 37 fields	Y	D1-R / D2-R
MetaSeg	15 E/D/S + 19 class means	Y	nested-C L1 saga
Fixed-pair mapping	available frozen (p,m)	no fit	deterministic mapping

Appendix

Table S3: Model and ranking specifications. Each row in (b) corresponds to the same object in (a). Analysis timing and evidence status are given in S8.

Fields	Learner	CE vs G0	CE vs PC	LL vs G0	LL vs Entropy
SegFormer (A1)
G1	Spline	+0.004 [-0.386, +0.371]	+0.126 [-0.246, +0.487]	+0.086 [-0.087, +0.272]	-0.028 [-0.260, +0.202]
G1	Boosting	-1.454 [-2.627, -0.163]	-1.454 [-2.627, -0.163]	-0.888 [-1.718, -0.119]	-1.274 [-2.191, -0.411]
G2	Spline	-0.010 [-0.222, +0.203]	+0.112 [-0.150, +0.381]	+0.060 [-0.099, +0.217]	-0.055 [-0.267, +0.164]
G2	Boosting	-0.708 [-1.942, +0.558]	-0.708 [-1.942, +0.558]	-0.493 [-1.240, +0.251]	-0.878 [-1.578, -0.168]
G3	Spline	-0.043 [-0.893, +0.838]	+0.174 [-0.671, +1.105]	+0.013 [-0.218, +0.262]	-0.101 [-0.363, +0.171]

Appendix

Table S4: G1–G4 source gains with pointwise 95% paired-image intervals, in

10^{-3}

nats. CE compares against learned

p

-only (G0) and dimension-matched

p

-derived controls (PC); LL compares against G0 and entropy. Positive favors added fields. Table S5 completes the retained-LL controls.

Learner / fields	PC columns	control minus Gz LL [95% CI]
SegFormer (A1)
Spline / G1	1	+0.095 [-0.087, +0.289]
Spline / G2	1	+0.068 [-0.100, +0.232]
Spline / G3	11	+0.028 [-0.221, +0.278]
Spline / G4	37	+2.238 [+0.418, +4.260]
Boosting / G1	1	-0.888 [-1.718, -0.119]

Appendix

Table S5: Retained-LL gains against dimension-matched

p

-derived controls (PC), completing Table S4. All 24 field–learner–architecture comparisons are shown, in

10^{-3}

nats with pointwise 95% intervals.

Figure S2: Full-field (G4) gains against learned

p

-only,

p

-derived controls and entropy references, shown separately for each learner. Positive favors G4; bars show pointwise 95% paired-image intervals.

Population / eligibility	N components	MetaSeg AUROC	$p$ AUROC	MetaSeg CE	$p$ CE
SegFormer (A1)
Source / All	18728	0.715702	0.810513	0.607565	0.499287
Source / Interior	15869	0.798600	0.803125	0.485110	0.483277
Src. dev. / All	12209	0.736090	0.811533	0.592712	0.512086
Src. dev. / Interior	10277	0.805113	0.803678	0.503272	0.501892
Adv. dev. / All	32429	0.708956	0.756237	0.682546	0.583443

Appendix

Table S6: Native predictive performance under all-valid and nonempty-interior eligibility. CE is in nats.

Population / eligibility	MetaSeg minus $p$ : AUROC $\times 1000$ [95% CI]
SegFormer (A1)
Source / All	-94.811 [-103.044, -86.755]
Source / Interior	-4.525 [-7.242, -1.720]
Src. dev. / All	-75.443 [-85.645, -65.133]
Src. dev. / Interior	+1.435 [-2.501, +5.529]
Adv. dev. / All	-47.281 [-54.037, -40.452]

Appendix

Table S7: Paired MetaSeg-minus-

p

AUROC differences, with 95% image-cluster bootstrap intervals.

Population	LL vs entropy	LL vs $p$	Failure vs entropy	Failure vs $p$
SegFormer (A1)
Source	-8.404 [-10.368, -6.455]	+10.100 [+7.773, +12.596]	+1.885 [+1.511, +2.252]	-1.147 [-1.414, -0.866]
Src. dev.	-8.010 [-10.215, -5.774]	+11.683 [+8.312, +15.035]	+1.896 [+1.426, +2.340]	-1.018 [-1.353, -0.679]
Adv. dev.	-18.003 [-19.594, -16.344]	+6.758 [+4.350, +9.058]	+1.771 [+1.501, +2.032]	-0.254 [-0.478, -0.031]
Final	-16.282 [-17.984, -14.515]	+4.607 [+1.878, +7.553]	+1.617 [+1.360, +1.888]	-0.106 [-0.345, +0.137]
DeepLab (A2)

Appendix

Table S8: MetaSeg gains from the original matched-analysis stream (main Figure 4). LL:

10^{-3}

nats; failure: percentage points. Positive favors MetaSeg; brackets show pointwise 95% intervals.

Population	Reference	Failure gain [95% CI]	LL gain [95% CI]
SegFormer (A1)
Source	Entropy	1.885 [1.530, 2.251]	-8.404 [-10.356, -6.360]
	Probability	-1.147 [-1.428, -0.873]	10.100 [7.635, 12.470]
Src. dev.	Entropy	1.896 [1.436, 2.331]	-8.010 [-10.287, -5.819]
	Probability	-1.018 [-1.342, -0.657]	11.683 [8.309, 14.798]
Adv. dev.	Entropy	1.771 [1.504, 2.044]	-18.003 [-19.515, -16.347]

Appendix

Table S9: All-valid components at 90% retention. Gains are reference minus MetaSeg: failure in percentage points; LL in

10^{-3}

nats. Brackets show pointwise 95% paired-image intervals. Comparators share eligible IDs and the recomputed budget within each row.

Population	Reference	Failure gain [95% CI]	LL gain [95% CI]
SegFormer (A1)
Source	Entropy	2.075 [1.675, 2.473]	-9.217 [-11.313, -7.129]
	Probability	-0.289 [-0.506, -0.081]	1.776 [0.667, 2.988]
Src. dev.	Entropy	1.721 [1.256, 2.162]	-8.177 [-10.676, -5.758]
	Probability	-0.141 [-0.400, 0.122]	1.336 [-0.306, 2.989]
Adv. dev.	Entropy	1.850 [1.557, 2.141]	-19.219 [-20.954, -17.398]

Appendix

Table S10: Nonempty-interior components at 90% retention. Gains are reference minus MetaSeg: failure in percentage points; LL in

10^{-3}

nats. Brackets show pointwise 95% paired-image intervals. Comparators share eligible IDs and the recomputed budget within each row.

Population	All: native-map LL	All: Entropy-map LL	Interior: native-map LL	All: native-map fail
SegFormer (A1)
Source	-5.547 [-9.010, -2.389]	-13.952 [-16.651, -11.387]	+6.420 [+4.218, +8.601]	+0.643 [+0.277, +0.997]
Src. dev.	-8.863 [-12.422, -5.104]	-16.873 [-20.011, -13.581]	+8.757 [+6.083, +11.598]	+0.712 [+0.326, +1.094]
Adv. dev.	-0.813 [-3.323, +1.855]	-18.816 [-21.148, -16.560]	+24.582 [+22.424, +26.871]	-0.197 [-0.453, +0.055]
Final	-0.926 [-3.876, +2.083]	-17.209 [-20.057, -14.374]	+23.236 [+20.954, +25.544]	-0.126 [-0.386, +0.125]
DeepLab (A2)

Appendix

Table S11: Mapping at 90% retention: LL gains in

10^{-3}

nats and failure gains in percentage points, with pointwise 95% image intervals. Positive favors the mapping.

Population	All-valid: $p$ minus mapped	Interior native minus map
SegFormer (A1)
Source	-0.504 [-0.716, -0.301]	-1.596 [-2.006, -1.178]
Src. dev.	-0.306 [-0.498, -0.107]	-1.409 [-1.856, -0.973]
Adv. dev.	-0.450 [-0.598, -0.297]	-1.566 [-1.849, -1.278]
Final	-0.232 [-0.382, -0.089]	-1.377 [-1.671, -1.082]
DeepLab (A2)

Appendix

Table S12: Failure comparisons completing Table S11. Entries are percentage points with pointwise 95% intervals. The interior column uses its recomputed restricted budget.

Figure S3: Native-minus-mapped LL gains across coverage, in

10^{-3}

nats: three architectures, four populations and two eligibility sets. All 168 archived estimates are shown; lines connect estimates. The dashed line marks primary 90% coverage.

Population / eligibility	50%	60%	70%	80%	90%	95%	100%
SegFormer (A1)
Source / All	-5.962	-8.611	-9.861	-7.419	-5.547	-2.931	+0.000
Source / Interior	+6.837	+9.285	+10.218	+9.392	+6.420	+4.299	+0.000
Src. dev. / All	-9.458	-11.981	-10.937	-9.857	-8.863	-3.773	+0.000
Src. dev. / Interior	+9.006	+12.366	+14.174	+12.687	+8.757	+5.282	+0.000
Adv. dev. / All	+6.653	+6.239	+4.849	+0.521	-0.813	-0.801	+0.000

Appendix

Table S13: Native-minus-mapped LL gains across coverage, in

10^{-3}

nats. Entries are archived point estimates.

Population / eligibility	native-map LL	Entropy-map LL	native-map failure
SegFormer (A1)
Adv. dev. / All	-0.813 [-3.759, +1.597]	-18.816 [-21.481, -16.916]	-0.197 [-0.421, +0.076]
Adv. dev. / Interior	+24.582 [+21.474, +27.545]	+5.363 [+3.087, +7.252]	-1.566 [-1.749, -1.314]
Final / All	-0.926 [-3.916, +2.537]	-17.209 [-19.616, -14.402]	-0.126 [-0.412, +0.135]
Final / Interior	+23.236 [+21.019, +25.543]	+5.288 [+3.925, +6.676]	-1.377 [-1.682, -1.087]
DeepLab (A2)

Appendix

Table S14: Target mapping comparisons with whole-sequence 95% intervals: 21 groups in P2 and 33 in P3. LL is in

10^{-3}

nats; failure is in percentage points.

Figure S4: Retained membership for the DeepLab source-development image in main Figure 7a. Each selector retains 107 of 118 components: 97 shared (gray), ten native-only (orange), ten mapped-only (blue); component 53 is rejected by both (open). Integer labels are original full-image identities.

Primary all-valid quantity	Minimum	Maximum
Changed-image fraction (%)	+98.500	+100.000
Mean exchanged retained fraction (%)	+9.356	+10.348

Appendix

Table S15: Primary exchange ranges across the twelve architecture–population cells. These ranges summarize the cells and are not confidence intervals.

Figure S5: Exchange descriptors for twelve primary comparisons at 90% retention. P0/P1/P2/P3: source/source-development/adverse-development/final. Axes show mean exchanged fraction

v_{j}

(% of retained components) and mean exchanged LL difference

d_{j}

(nats). Overall gain averages within-image products

v_{j}d_{j}

, not the product of these plotted means.

Component ID	$Y$	$\ell_{\rm LL}(Y,p)$ [nats]
Native-only
38	0	0.414215
43	0	0.282678
44	0	0.435519
45	0	0.628635
79	0	0.399822

Appendix

Table S16: Twenty exclusive components from main Figure 8 (DeepLab/P1, Source 1). The native-only and mapped-only groups are unpaired and ordered by component ID. LL evaluates fixed

p

(nats), rounded to six decimals; the accompanying numeric record preserves original precision. Values describe this single image.

Population	F1	F2=p	F3
SegFormer (A1)
Src. dev.	-10.929 [-13.887, -7.947]	-19.693 [-23.135, -16.199]	-0.167 [-0.912, +0.568]
Adv. dev.	-19.496 [-21.347, -17.682]	-24.761 [-27.233, -22.264]	-0.168 [-0.648, +0.315]
Final	-18.026 [-20.176, -15.839]	-20.890 [-23.930, -17.662]	-0.911 [-1.369, -0.416]
DeepLab (A2)
Src. dev.	-6.826 [-8.958, -4.561]	-7.703 [-9.963, -5.242]	+0.194 [-0.341, +0.756]

Appendix

Table S17: Complete original frozen-score LL@90 gains, R0 minus candidate, in

10^{-3}

nats. Brackets are the original pointwise 95% paired-image intervals. Source OOF is not pooled with deployment evaluation.

Cell	Gain x1000 nats [90% CI]	Positive single-fold operators
Seg_dev	-1.340 [-2.706, +0.061]	0/5
Deep_dev	+1.467 [+0.313, +2.692]	5/5
Seg_pros	-0.213 [-0.597, +0.196]	1/5
Deep_pros	-0.245 [-0.715, +0.234]	1/5

Appendix

Table S18: Additional development-analysis gains with primary 90% intervals. Favorable fold counts describe the fitted fold operators, not independent replications.

Contrast	Point [interval]
Image-weighted / T_Seg	1.126 [-0.446, 2.898]
Image-weighted / T_Deep	-1.712 [-3.248, -0.202]
Image-weighted / Cdev	2.807 [0.616, 5.123]
Image-weighted / Clater	-0.031 [-0.771, 0.738]
Image-weighted / D	-2.838 [-5.269, -0.500]
Equal-city / Seg_dev	-1.340 [-2.996, 0.137]

Appendix

Table S19: Additional direct-change and training-target controls, in

10^{-3}

nats.

T

is later minus development gain;

C

is DeepLab minus SegFormer gain;

D=C_{\rm later}-C_{\rm dev}

. Direct-change and equal-city intervals are 95%; direct-LL controls retain their original 90% intervals. Equal-city weighting is a sensitivity estimand.

Research line and unit	Established evaluation target	Connection to this study
Predictive uncertainty; samples or pixels [ 1 , 2 , 3 , 4 , 5 ]	Model and data uncertainty; calibration, error detection and distribution shift.	Frozen segmentation outputs; component losses and per-image retention budgets.
Failure prediction; pixels, components or images [ 6 , 7 , 8 , 9 , 10 ]	Correctness or quality forecasts; confidence aggregation to the decision unit.	Component MetaSeg; entropy and probability references; prediction and selection assessed separately.
Selective prediction; samples with rejection [ 11 , 12 , 13 , 14 , 15 , 16 , 17 ]	Risk–coverage, conditional-risk selection, joint predictor/rejector learning and shift.	A fixed predictor and matched retained counts within each image.
Probability calibration; class forecasts [ 18 , 19 , 20 , 21 , 22 ]	Class-probability calibration, spatial scaling and decision-specific calibration.	Component-failure forecasts; diagnostic CE; retained LL of unchanged base $p$ .
Evaluation; accepted sets and thresholds [ 23 , 24 , 25 , 26 , 27 ]	Calibration, discrimination and selection measures; rejection; multi-threshold alignment.	Absolute retained levels, paired gains, coverage and eligibility sensitivity.
Loss alignment; forecasts and image-level segmentation [ 28 , 15 , 29 ]	Proper probability scores, conditional-risk ordering and Dice-aligned confidence.	Component failure versus forecast loss; native versus LL-risk ranking on fixed $(p,m)$ .

Appendix

Table S20: Literature map by evaluation unit and target. The final column identifies the corresponding comparison in this study.

Selective Prediction

Selective prediction lets models answer only when sufficiently reliable, trading coverage for lower error through abstention, referral, or extra evidence gathering.

See all 287 papers

Semantic Feature Segmentation for Interpretable Predictive Maintenance in Complex Systems
May 14, 2026Emilio Mastriani, Alessandro Costa, Federico Incardona +2Interpretable ML Predictive Maintenance
INAF, Osservatorio Astrofisico di Catania, Via S. Sofia 78, Catania, 95123, Italy
LOCUS: A Distribution-Free Loss-Quantile Score for Risk-Aware Predictions
Mar 2, 2026Matheus Barreto, Mário de Castro, Thiago R. Ramos +2Uncertainty Calibration
Department of Statistics, Federal University of S˜ao Carlos, S˜ao Carlos, S˜ao Paulo, Brazil · Institute of Mathematics and Computer Science, University of S˜ao Paulo, S˜ao Carlos, S˜ao Paulo, Brazil · School of Forest, Fisheries, and Geomatics Sciences, University of Florida, Gainesville, Florida, United States of America
Rethinking Uncertainty in Segmentation: From Estimation to Decision
Apr 14, 2026Saket MagantiImage Segmentation Segmentation Uncertainty Estimation

May 14, 2026J/K move · Enter open · S save

Semantic Feature Segmentation for Interpretable Predictive Maintenance in Complex Systems

Emilio Mastriani, Alessandro Costa, Federico Incardona, Kevin Munari+1

INAF, Osservatorio Astrofisico di Catania, Via S. Sofia 78, Catania, 95123, Italy

Interpretable ML Predictive Maintenance Semantic Segmentation

Open paper AskPDF arXiv

Predictive maintenance in complex systems is often complicated by the heterogeneity and redundancy of monitored variables,which can obscure fault-relevant information and reduce model interpretability. This work proposes a semantic feature segmentation framework that decomposes the monitored feature space into a canonical component,expected to retain the dominant predictive information, and a residual component containing structurally peripheral signals. The segmentation is defined through domain informed criteria and sets up monitoring variables into functional groups reflecting operational mechanisms such as throughput,latency,pressure,network activity,and structural state. To evaluate the effectiveness of this decomposition, we adopt a predictive perspective in which expected predictive risk is used as an operational proxy for task-relevant information. Experimental results obtained through time-aware cross-validation show that the canonical space consistently achieves lower predictive risk than the residual space across multiple temporal configurations, indicating that the semantic segmentation concentrates the most relevant information for fault anticipation. In addition, the canonical segments exhibit significantly stronger intra-segment coherence than inter-segment dependence, and this structural organization remains stable after redundancy reduction. When compared with the full feature space and with a Principal Component Analysis (PCA) representation, the canonical space carries out comparable predictive performance and furthermore preserves the semantic meaning of the original variables. These findings suggest that semantic feature segmentation provides an interpretable and information-preserving decomposition of monitoring signals, enabling competitive predictive performance without sacrificing the operational interpretability required in predictive maintenance applications.

Abstract

Selective Prediction

Explore similar work

Semantic Feature Segmentation for Interpretable Predictive Maintenance in Complex Systems

LOCUS: A Distribution-Free Loss-Quantile Score for Risk-Aware Predictions

Rethinking Uncertainty in Segmentation: From Estimation to Decision