acceptodds
Under review as a conference paper at ICLR 2027

RELATE: Health-Relative Evidence for Structured Industrial Diagnosis with LLMs

Abstract

Industrial diagnosis spans signals with different sampling rates, fault vocabularies, and diagnostic attributes. We introduce RELATE, a framework that pairs source-rate query features with healthy training references for condition classification and structured record prediction. RELATE-MLP classifies the complete query–reference representation; RELATE-Gen augments the same representation with temporal features and maps the fused evidence to 4 continuous tokens for candidate-conditioned Qwen generation. We evaluate both routes on MOSAIC (Multi-source Operational Signals with Annotated Industrial Conditions), which covers 12 acoustic, vibration, current, and voltage settings under fixed group-level splits. RELATE-MLP reaches 80.60% setting-macro accuracy and 78.36% setting-macro F1 on 3,598 frozen-test groups. RELATE-Gen reaches 78.65/77.01% accuracy/F1 and 68.86% setting-macro Joint EM on records assembled by per-field voting across views. Across three fixed validation runs under the selected autoregressive objective, health-relative evidence improves accuracy by 5.07 points on average; the complementary temporal evidence and learned fusion add 4.56 points.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Industrial fault diagnosis, Healthy-reference learning, Multimodal industrial signals, Large language models, Structured prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_T9xvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper is about industrial fault diagnosis when the sources are heterogeneous: acoustic, vibration, current and voltage records with different sampling rates, different fault vocabularies, and different sets of annotated fields. Each query comes with machine context and a candidate list, and the system has to pick a condition and fill whatever record fields the source supports. The proposed setup takes a frozen FISHER embedding of the query, q, retrieves a healthy prototype r from a train-only bank by cosine similarity (up to 16 groups, one view per group), and feeds [q, r, q-r, |q-r|] to one of two heads. RELATE-MLP is a 6.30M-parameter MLP over the full 24,576-d representation with setting-local candidate masks. RELATE-Gen adds a MiniROCKET branch, fuses the two through a gated residual block, squeezes the result into 4 continuous tokens, and uses them to condition a frozen 4-bit Qwen2.5-7B-Instruct with LoRA that generates the record as JSON.

The benchmark, MOSAIC, is 11 RMIS settings plus one MIMII-DG setting with frozen group-level splits, 15,477/1,218/3,598 groups. On test, RELATE-MLP gets 80.60/78.36 setting-macro accuracy/F1 and RELATE-Gen 78.65/77.01, with 68.86 Joint EM after per-field voting and 79.95 Field EM. A readout on the same frozen tokens gets 69.07 Joint EM at 45.95 views/s against 0.20 for generation. Three-run validation controls give +5.07 for the healthy reference and +4.56 for the temporal branch with fusion; single-run controls cover token count, input blocks and prefix size. There is also a CWRU adaptation compared against a locally adapted BearLLM.

Critical strengths
  1. The classifier result holds up under both aggregations the paper reports, and the harness behind it is unusually careful. Every method, including the three proprietary zero-shot APIs, is scored by the same group-level scorer with the same denominators and metric weights. The voting rules are written down rather than left implicit: missing values enter the per-field vote and only win if they beat every non-missing value, ties among non-missing values break by canonical JSON lexical order, a field is emitted only when its winning value is non-missing, and paths outside a group's supported set are dropped from the record metric but counted in a separate audit, which reports that "none of the 11,171 emitted group-level field paths fell outside their corresponding setting-level schemas". Gold is joined after generation, and the prompt carries no dataset IDs, signal UIDs, group IDs or field mask. I recomputed the tables instead of trusting them: the twelve per-setting rows of Table 1 reproduce the setting-macro row to within 0.005 in every column, both panels of Table 11 and all four columns of Table 10 reproduce theirs, the twelve rows of Table 8 sum exactly to 15,477/33,476, 1,218/2,079 and 3,598/4,430, and the two within-condition fields of Table 23 average to exactly the Severity EM column of Table 2. What is solid is that RELATE-MLP's lead does not depend on the weighting questions I raise below: it is +5.36 on setting-macro and +2.00 on group-micro, and it is ahead on 10 of 12 settings. A 6.30M-parameter model whose whole route bundle trains in about ten minutes on one consumer GPU and beats every baseline on both aggregations is the strongest thing in the paper. To be clear about where the value is: most of the machinery is borrowed, including the query-reference-residual construction that Section 2 attributes to BearLLM, so the contribution worth having is the protocol, the harness, and the finding that a frozen representation plus a small MLP is enough to lead the table. That is a real result, but it is not a new architecture.

  2. The paper is candid about the results that go against it, and the controls are matched properly. FISHER 1-NN is reported as beating the generator on group-micro (89.83 vs 88.85), the classifier-minus-generator interval is reported as spanning zero, going from 4 to 8 tokens is reported as hurting, the expanded prefix is reported as worse than the compact one, the auxiliary-loss control is reported below the autoregressive objective, and the input control is reported as favouring [q, r] over the full four blocks. The controls themselves are better matched than I am used to seeing here: the three input arms share one saved initialisation and differ only in masking, the AR and Direct interfaces share features, sampler, schedule and endpoint, and Qwen and its readout receive literally identical token vectors exported from the same checkpoint. I raise this as credit, but it also shapes how I read the paper. The contradictions I complain about below are between the abstract and your own tables, not between the paper and reality, which is why my concerns are about framing and coverage rather than about the work itself.

Critical weaknesses
  1. The arithmetic behind the headline runs one way for the classifier and the other way for the generator, and the abstract reports only the setting-macro side of both. Setting-macro gives each setting one twelfth, but the settings are not remotely the same size: Table 8 has 16 test groups for umged_cur, 28 for umged_vol, 36 each for pu_cur and pu_vib, and 1,000 for mafaulda_vib, so the biggest is 62.5x the smallest. Since every per-setting accuracy is a count over n groups, every margin in Table 1 is an exact multiple of 100/n, so the margins can be converted back into groups. I did that. RELATE-MLP comes out +72 groups against FISHER 1-NN, which is exactly what its group-micro row implies (+2.00% of 3,598 = 72). RELATE-Gen comes out -35 groups (-0.98% of 3,598 = 35), so the same Table 1 that gives it +3.41 points of setting-macro gives it 35 fewer groups correct, and it is ahead on only 5 of 12 settings. Four small settings carry 83% of the classifier's macro lead and 118% of the generator's; for the classifier the group-level count agrees with the macro, so the concentration does not change its sign, and for the generator it does. Your own intervals say the same thing. RELATE-Gen minus 1-NN is "+3.41 [0.02,6.94]", whose lower end is two hundredths of a point, and the two settings carrying that margin have intervals spanning more than 28 points each (PU current +30.56 [13.89,47.22], UMGED voltage +21.43 [7.14,35.80]). The consequence is narrow but it matters for contribution (3): the claim that the classifier leads the generator rests on "[-0.56,4.60]", which crosses zero. On the record side, Field EM is the mean of seven paths and one of them is handed to the model: Appendix G says "Machine type is provided in the prompt", and on the Table 24 validation rows machine type scores 99.94 and 100.00 against a seven-path mean of 78.17, so dropping that one path would leave 74.54, with three of the remaining paths described by the authors themselves as "largely determined by the primary taxonomy". Severity EM (52.48 against 51.78) has no interval row in Table 13 at all, and Table 23 has the frozen-token readout beating the generator on eccentricity, 32.89 against 27.63. That last comparison is like-for-like, since the taxonomy mapping only covers non-severity fields.

None of this is a numerical error, and I want to say that plainly because my recommendation leans on the tables being trustworthy. What I would ask for here is a re-scoring. One comparison in Table 2 is not like-for-like, and there the asymmetry runs against the generator rather than in its favour: the caption says both readouts use the same train-only taxonomy rules, and those rules fill in the non-severity fields that the primary condition determines uniquely, whereas the generator predicts the complete record itself. Its 68.86 Joint EM is therefore not directly comparable to the readout's 69.07, and until the generator is scored under the same mapping the paper cannot say what the generative route is worth. That is my first question below.

  1. Each component claim rests on a single three-run comparison, and both numbers move when the recipe moves. Section 5 says "Validation selects RELATE configurations". The frozen-test tables contain no ablation rows at all: no query-only, no reference-free, no other objective. Every component comparison sits on the same 1,218 validation groups that chose the components, in Tables 3, 4, 19, 20, 21 and 22, and every one of them except the two headline ablations is single-seed. The headline ablations themselves are three-run, but the auxiliary-loss versions of the same two are single-seed, and both come out far lower: the reference at +0.59 against +5.07 (71.20 to 71.79 in Table 22), and the temporal branch with fusion at +1.37 against +4.56 (70.42 to 71.79 for FISHER-small in Table 21). With three runs on one side and one on the other I cannot tell whether that gap belongs to the objective or to the seed, and the abstract carries neither the numbers nor the caveat. There is a second issue in how the reference ablation was built. The control "keeps the temporal branch and input geometry, replacing the FISHER blocks with [q, q,0,0]", which zeroes the difference blocks as well, so the +5.07 credited to the reference also contains whatever the difference blocks contribute; Table 20 measures those separately at +2.43. The same tension runs through the method. Eq. (1) and Figure 1(b) define the input as [q, r, q-r, |q-r|], but in the matched control [q, r] gets 78.38/76.79 against 77.31/75.08 for the full four blocks, so the blocks the framework is named after cost 1.07 and 1.71 points, and the main model keeps all four anyway while reporting 78.38, the same number the two-block arm reaches. Some choices are also being made inside your own error bars: 4 tokens is selected over 8 (75.75 against 75.36, a gap of 0.39) against standard deviations of 1.20 to 1.90 that the paper reports across its three-run configurations, on a sweep the paper calls "single-seed". None of this overturns the classifier result, which is a frozen-representation result and does not depend on the autoregressive objective at all. It does mean that contribution (1), the claim that the reference and the fusion are what makes the method work, is supported only on the split that selected the configuration, and by one comparison per component.

That is why I would reject the paper as written rather than ask for a prose revision. What the paper claims is that the health-relative components are what makes the method work, and that the generative route holds its own against the non-language alternatives. On the evidence submitted neither is established: the component gains come from the split that selected the configuration, the auxiliary-loss versions of the same two ablations come out several times lower (single-seed, Tables 21 and 22), the difference blocks that Eq. (1) is built around cost accuracy in the paper's own matched control, the classifier-minus-generator interval crosses zero, and the generator's macro lead rests on 3.2% of the groups. What does hold is the protocol, the harness, and the fact that a frozen representation plus a small MLP leads every aggregate metric. The protocol is claimed as contribution (2) but not delivered, while the harness and the classifier result are delivered but never claimed. So the three things I would need are a claim and a delivery that match the evidence: re-frame contributions (1) and (3) around the classifier result and the harness, scoring the generator under the same taxonomy rules as the readouts so that anything still claimed for it can be checked like for like; release the split manifests and the evaluation code, without which contribution (2) cannot be reproduced; and report per-setting n with per-setting intervals beside a group-weighted column, carrying the auxiliary-loss numbers next to the autoregressive ones wherever the abstract's two gains appear. I would reconsider on those three, and none of them needs a new model: two are writing and release work, and the third is a re-score and arithmetic on data you already have.

Questions for authors

Q1. Table 2's caption says both readouts use the same train-only taxonomy rules, but the generator predicts the complete record itself, so its 68.86 Joint EM is computed under a harder rule than the readout's 69.07. Can you re-score the generator under the same taxonomy mapping so that the three rows are comparable? And can you break its remaining errors into malformed JSON, out-of-candidate primary labels, and well-formed records with wrong field values? If format dominates, grammar-constrained decoding would remove most of the gap and the 4-token interface would have to be argued on other grounds.

Q2. Can you report the healthy-reference ablation and the [q, r] versus [q, r, q-r, |q-r|] comparison on the frozen test split, rather than only on the split that selected the configuration? If that is not available, what should a reader conclude about contribution (1) from validation-only evidence?

Q3. Table 22 has the auxiliary-loss version of the reference ablation (+0.59) and Table 21 the auxiliary-loss version of the fusion ablation (+1.37), against +5.07 and +4.56. The autoregressive numbers are three-run and the auxiliary ones single-seed, so I cannot separate the objective from the seed. Can you run the auxiliary variant with three runs, or state that both gains are conditional on the objective? Related: since the no-reference control zeroes the difference blocks as well as swapping the FISHER blocks for [q, q,0,0], could you add a control that removes the reference while keeping the four-block geometry?

Q4. FISHER 1-NN beats the generator on group-micro, and the generator's macro lead sits in four settings holding 116 groups. Could you give per-setting test n and per-setting intervals, plus setting-macro under group weighting? Which of the two aggregate conclusions should a reader carry to an unseen setting, and does the protocol in contribution (2) commit to one of them?

Q5. Will the MOSAIC split manifests, benchmark-group identifiers and evaluation code be released? Section 3 points to Appendix A for "the split construction and source identifiers", but Appendix A defines the source-family keys rather than listing identifiers, and Appendix A.1 describes a class-balanced subset drawn by quota from that partition rather than the full native test set. For a benchmark contribution I would need the manifests themselves.

Q6. Why does the main RELATE-MLP keep the four input blocks when the matched control gives them -1.07/-1.71, and is its 78.38 distinguishable from the 78.38 of the [q, r] arm, given that the two use different recipes (group-level 80 epochs against view-level 100,424 steps)?

Q7. RELATE-CWRU trains with an auxiliary candidate-masked cross-entropy head over the ten classes, removed at inference, and the description of the locally adapted BearLLM baseline does not mention a corresponding supervised head, on 222 test segments with different base-model sizes. Can you either align the two recipes or restrict the conclusion to what the aligned part supports?

4 comments
AuthorsOpenReview ↗

Response to T9xv (1/4): Q1 — comparable record scoring

Thank you for your careful review and constructive feedback.

Substantive revisions in the manuscript are marked in blue.

Q1 Same-taxonomy scoring and remaining errors

Yes. We applied the same training-only field-filling rule to all three predictors. For the originally reported seed-111 Gen model, Joint EM remains 68.86% under this common rule. Gen and the AR readout use the same signal tokens learned through RELATE's structured-record generation training. Direct learns its signal tokens through direct condition and field supervision. All three cover 4,430 views, 3,598 benchmark groups and 12 settings in each training seed. The three-seed results are as follows. Here ± denotes sample standard deviation across seeds 111/222/333.

MethodJoint EM mean±SDField EM mean±SD
Gen67.43±1.36%78.90±0.90%
AR readout68.20±0.75%79.70±0.50%
Direct readout66.73±0.79%79.62±0.72%

For the original seed-111 comparison, raw/mapped Joint EM is 68.86/68.86% for Gen, 68.47/69.07% for AR and 65.08/66.70% for Direct. Gen raw/mapped Field EM is 79.9503/79.9368%.

A paired comparison of all 3,598 groups finds 29 more fully correct Gen records than AR records at seed 111, 13 more at seed 222 and one fewer at seed 333. These are equally weighted group counts. The table's Joint EM gives each setting equal weight. The three UMGED settings contribute −1.25 points to the three-seed mean Gen-minus-AR Joint EM difference, while the other nine settings together contribute +0.48 points. Both summaries use the same common-mapped predictions.

The seed-111 error decomposition contains zero malformed JSON records and zero out-of-candidate primary labels among all 4,430 generated records. After aggregation, 3,110 groups are fully correct. Among the other 488 groups, 401 have a wrong primary condition, including two that also lack a severity field. The other 87 have the correct condition and a wrong severity value. These are group counts, whereas Joint EM averages the twelve setting-level proportions equally. Condition and severity estimation account for these errors even when the output is syntactically valid, so grammar constraints alone would leave their main cause in place.

With the same readout architecture, the AR-trained representation gives higher test Joint EM than Direct in all three seeds, with paired gains of +2.37, +1.84 and +0.20 points. These results support using RELATE's autoregressive training scheme to learn signal tokens for record prediction. Gen and the AR readout then compare two ways to decode that learned representation, allowing deployment to use either Qwen generation or a lightweight field predictor.

Before scoring, Gen, the AR readout and the Direct readout use the same rule to fill non-severity fields whose values are uniquely determined by a setting and condition in the training data. We construct this rule from training annotations alone. For example, if every training record for outer-race fault in a setting has health state anomalous and location outer race, a prediction of outer-race fault sets these two fields to those values. Non-severity fields without a unique training value retain the model's prediction. The rule leaves the predicted condition, damage size and eccentricity unchanged.

When a benchmark group contains multiple views, we first combine their predictions into one group-level prediction. Gen selects the condition receiving the most votes. The two readouts average the candidate logits across views and select the condition with the highest mean logit. We then apply the field-filling rule using that group-level predicted condition.

The readouts restore numeric fields to the data types declared during training before scoring.

Joint EM requires the condition and every supported field to match. Field EM scores each field over its supporting groups and then averages the seven field scores equally. This puts the three methods on the same taxonomy and evaluation basis.

Revised Table 2 and Appendix G, Tables 34–36, report the common-rule record comparisons.

AuthorsOpenReview ↗

Response to T9xv (2/4): Q2–Q3 — three-seed controls

Continuing our response to Reviewer T9xv.

Q2 Input controls on the frozen test split

We evaluated the MLP and Gen three-arm input comparisons on the frozen test split across three training seeds. The results use setting-macro accuracy on the same 3,598 groups and 12 settings.

MLP inputTest accuracy mean±SD
Query only [q,0,0,0]79.15±2.08%
Query/reference [q,r,0,0]78.89±1.74%
Full [q,r,q-r,abs(q-r)]79.45±0.90%

Within a seed, these MLP arms share the saved initialization, architecture, 80-epoch recipe, sampler and 24,576-dimensional input width. Full minus query-only averages +0.30±2.36 points, and full minus pair averages +0.56±1.28. The explicit differences have a small mean effect for this classifier, with variation across seeds.

Gen inputTest accuracy mean±SD
Self-reference [q,q,0,0]74.96±1.03%
Query/reference [q,r,0,0]75.92±0.82%
Full [q,r,q-r,abs(q-r)]76.76±1.72%

Full input improves accuracy in each seed, with a paired gain of +1.80±1.47 points over self-reference. Retaining the healthy reference and masking only the differences gives 75.92±0.82%, versus 76.76±1.72% for full input. The paired full-minus-pair accuracy difference is +0.85±2.28 points. The individual differences are as follows.

Training seedFull-input test accuracyPair-only [q,r,0,0] test accuracyFull minus pair-only, pp
11178.65%75.17%+3.48
22275.28%75.78%−0.50
33376.35%76.80%−0.45

Pair-only is 0.70 points below self-reference at seed 111. Across three seeds, full and pair-only F1 is 75.07±1.75% and 74.96±0.88%, with a paired difference of +0.11±2.44 points.

Gen uses the fixed 100,424-microstep endpoint, greedy JSON generation and the same group scorer. These frozen-test comparisons supplement the validation evidence and distinguish the complete reference construction from the difference blocks with the reference retained.

The complete frozen-test controls are in revised Appendix F, Tables 28–29.

Q3 Objective dependence and the four-block control

The reference and fusion effects depend on the training objective. Both auxiliary comparisons now have three seeds at the same fixed 100,424-microstep endpoint. All entries below are validation setting-macro accuracy differences on 1,218 groups and 12 settings.

ContrastAutoregressive objective, three seedsAuxiliary objective, three seeds
Full versus self-reference+5.07 pp+0.95±1.32 pp
Temporal branch with learned fusion versus FISHER-only+4.56 pp−0.58±2.10 pp

For the reference comparison, AR full/self-reference reaches 74.73±1.20% / 69.66±1.79%, with a positive gain in each seed. Auxiliary full/self-reference reaches 71.39±0.43% / 70.44±1.66%. Its paired differences are +0.5903, +2.4048 and −0.1548 points, averaging +0.95±1.32.

For auxiliary-objective fusion at the same 100,424-microstep endpoint, full and FISHER-only accuracy is 71.39±0.43% and 71.97±2.01%, respectively. Their paired differences are +1.37, −0.30 and −2.80 points at seeds 111/222/333, averaging −0.58±2.10 points. These ± values are sample standard deviations across training seeds. The fusion comparison therefore gives different results under the two training objectives.

The self-reference control retains the four-block formula by setting r*=q. Then [q,r*,q−r*,abs(q−r*)] is [q,q,0,0]. Input width, temporal branch and projection are held fixed. The resulting gain measures the whole healthy-reference construction, including the deviations derived from the reference. To distinguish the explicit differences, [q,r,0,0] keeps the healthy prototype and masks only those two blocks. Full minus this pair-only arm measures the difference blocks with the reference retained. We report these contrasts separately because they answer different questions about the construction.

The three-seed auxiliary-objective results are in revised Appendix F, Tables 30–31.

AuthorsOpenReview ↗

Response to T9xv (3/4): Q4–Q5 — weighting and reproducibility

Continuing our response to Reviewer T9xv.

Q4 Weighting, sample sizes and uncertainty

The protocol uses group-level setting-macro accuracy as the primary metric, giving each diagnostic setting equal weight. Group-micro accuracy gives each test group equal weight. The fixed submitted models give the following results on the same 3,598 groups.

Fixed modelSetting-macro accuracyGroup-micro accuracy
FISHER 1-NN75.24%89.83%
Main RELATE-MLP80.60%91.83%
Seed-111 RELATE-Gen78.65%88.85%

The MLP gains 72 correct groups over 1-NN and leads on 10 of 12 settings. Its paired gains are +5.36 macro points [2.51, 8.43] and +2.00 micro points [1.11, 2.86]. The MLP-minus-Gen macro difference is +1.94 [−0.56, 4.60]. These results support the full-evidence classification route under both benchmark weightings.

The fixed Gen model leads 1-NN on 5 of 12 settings and has 35 fewer correct groups overall. Its macro difference is +3.41 [0.02, 6.94], while its micro difference is −0.98 points. The macro gain is concentrated in small settings. Both weightings are needed to interpret that result. Per-setting group counts and intervals are given here.

SettingnMLP accuracy [95% CI]Gen accuracy [95% CI]1-NN accuracy [95% CI]
idmt_air40089.50 [86.50, 92.25]85.75 [82.25, 89.00]89.50 [86.50, 92.26]
mafaulda_sound29085.52 [81.38, 89.32]82.07 [77.59, 86.21]80.00 [75.17, 84.14]
mafaulda_vib100097.60 [96.60, 98.40]93.20 [91.60, 94.70]97.50 [96.50, 98.40]
pu_cur3661.11 [44.44, 77.78]75.00 [61.11, 88.89]44.44 [27.78, 61.11]
pu_vib3683.33 [69.44, 94.44]83.33 [69.44, 94.44]80.56 [66.67, 91.67]
sdust_bearing40097.50 [96.00, 98.75]97.00 [95.25, 98.50]97.25 [95.75, 98.75]
sdust_gear700100.00 [100.00, 100.00]99.29 [98.71, 99.86]99.43 [98.86, 99.86]
umged_cur1662.50 [37.50, 81.25]43.75 [18.75, 68.75]50.00 [25.00, 75.00]
umged_vib3259.38 [43.67, 75.00]68.75 [53.12, 84.38]68.75 [53.12, 84.38]
umged_vol2860.71 [42.86, 78.57]60.71 [42.86, 78.57]39.29 [21.43, 57.14]
wt_plane_gearbox6090.00 [81.67, 96.67]80.00 [71.63, 90.00]81.67 [70.00, 90.04]
mimii_dg60080.00 [76.67, 83.17]75.00 [71.50, 78.50]74.50 [70.83, 78.00]

The per-setting intervals use the existing 1,000-replicate percentile bootstrap, with benchmark groups sampled within settings and the seed-111 models fixed. The paired aggregate intervals use 10,000 shared within-setting group resamples. Training-seed variation is reported separately as SD. A perfect observed result, such as the MLP on SDUST gear, produces a degenerate empirical-bootstrap interval.

For the matched three-seed MLP controls, query-only, pair-only and full input have group-micro accuracy of 91.15±0.56%, 90.73±0.26% and 91.74±0.16%. Full adds 44, 32 and 33 correct groups over pair-only, averaging +1.01±0.19 micro points. Their setting-macro difference remains +0.56±1.28. Setting-macro is the protocol's primary summary for equal weighting of the twelve known settings, while group-micro describes their observed test mixture. Evaluating transfer to an unseen setting requires a held-out-setting experiment.

Revised Appendix B, Table 11, lists the per-setting sample counts and intervals.

Q5 Release of the actual benchmark materials

Yes. We will release the exact train/validation/test group and view manifests, benchmark-group and view identifiers, fixed candidate order, label/schema definitions and executable scoring code in anonymous supplementary materials. The manifests specify 15,477/1,218/3,598 groups and 33,476/2,079/4,430 views. They identify the actual class-balanced quota subset used here, including its relationship to the upstream RMIS partition.

The materials also contain Gen predictions for all 4,430 seed-111 test views, with gold annotations stored separately. The README provides commands for primary scoring, raw structured scoring and the common train-only taxonomy rule. Those commands reproduce Joint EM of 68.8642% and raw/mapped Field EM of 79.9503/79.9368%. Source attribution, upstream conditions and a dated environment snapshot accompany the scorers.

AuthorsOpenReview ↗

Response to T9xv (4/4): Q6–Q7, fields and token count

Continuing our response to Reviewer T9xv.

Q6 Why expose the differences, and which MLP recipe is used?

We retain the four blocks to give both predictors the query state q, healthy-reference state r, signed deviation q−r and deviation magnitude abs(q−r) as explicit inputs. The MLP uses the complete representation, while Gen compresses it into four tokens after temporal fusion. The difference blocks are deterministic functions of the pair, and their measured contribution is evaluated separately for each predictor below.

The matched three-seed MLP validation accuracy is 76.08±1.89% for full input and 76.16±1.97% for pair-only. Its test full-minus-pair difference is +0.56±1.28 points. Pair-only is competitive for this classifier. The Gen results in Q2 distinguish the complete reference construction from the explicit difference blocks with the reference retained.

The two 78.38% validation values are from separate training runs under the same nominal group-level, 80-epoch primary-MLP recipe. The matched input arms share a saved initialization and sampler. Input effects are estimated within those arms. The 100,424 view-microstep recipe in the neighboring appendix paragraph belongs to the separate structured-MLP experiment.

ExperimentTraining recipeValidation accuracyTest accuracy
Submitted main full MLPGroup-level, 80 epochs78.38%80.60%
Matched query/reference controlSame saved initialization and 80-epoch control recipe78.38%77.64%
Matched full-input controlSame saved initialization and 80-epoch control recipe77.31%79.57%

The main classifier's 80.60% test result is its fixed-run benchmark result. The matched full-input result of 79.57% is used with the other matched arms to estimate input effects.

Revised Appendix F, Tables 27–28, report the fixed-seed and three-seed input controls separately.

Q7 Scope of the CWRU comparison

We restrict the conclusion to the two complete local adaptations evaluated on the same 222 one-second CWRU test segments and ten-class candidate labels. RELATE-CWRU uses a frozen FISHER-relative signal branch, Qwen2.5-7B and a training-only auxiliary candidate-masked primary-condition head. Adapted BearLLM uses its released signal components and Qwen2.5-1.5B. The auxiliary head is removed for RELATE-CWRU inference.

Adapted systemSegment accuracySegment macro F1
RELATE-CWRU83.33%80.95%
BearLLM43.24%31.49%

The observed advantage applies to these systems under their respective adaptation recipes. Encoder, model size and supervision are part of that comparison.

The two adaptations are documented in revised Appendix C, Table 15.

Supplied fields, severity uncertainty and token count

Machine type is given in the prompt. For seed-111 Gen it is correct on all 600 supporting test groups. Removing that field from the raw seven-field average changes Field EM from 79.95% to 76.61% over the other six fields. Health state, location and mechanism largely express the condition taxonomy. We therefore report those record fields separately from damage size and eccentricity, which can vary within a condition, and use the common taxonomy rule in Q1 for the three methods.

Across three seeds, Gen-minus-AR severity results use all 300 damage-size groups and 76 eccentricity groups in each seed. Severity EM weights the two fields equally.

FieldGen mean±SDAR readout mean±SDPaired Gen−AR difference, pp
Damage size74.44±2.83%72.11±1.50%+2.33±3.79
Eccentricity24.12±4.02%32.02±2.74%−7.89±5.74
Equal-field Severity EM49.28±2.85%52.06±1.75%−2.78±3.84

The damage-size differences are +6.67, +0.67 and −0.33 points across seeds. Eccentricity favors the readout in all three. For the fixed seed-111 models, Gen obtains 77.33% damage-size and 27.63% eccentricity EM, versus 56.00% and 23.68% for the learned MLP severity heads. The paired 95% intervals below use 10,000 within-setting benchmark-group resamples and identical draws for the compared models.

Fixed seed-111 contrastDamage-size EM difference, pp [95% CI]Eccentricity EM difference, pp [95% CI]Severity EM difference, pp [95% CI]
Gen minus AR readout+6.67 [2.31, 10.97]−5.26 [−13.16, 2.63]+0.70 [−3.95, 5.26]
Gen minus MLP severity heads+21.33 [15.36, 27.33]+3.95 [−7.89, 15.79]+12.64 [5.85, 19.47]

In the matched single-seed token sweep, four tokens reach 75.75% validation accuracy and eight reach 75.36%, a 0.39-point difference. We choose four for half the signal-prefix length with slightly higher observed validation accuracy in that sweep. Separately, the final four-token configuration has a three-seed validation accuracy SD of 1.20 points. That SD describes repeated training at four tokens.

Reviewer_YFefOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RELATE for structured industrial diagnosis using health-relative evidence from frozen FISHER representations and retrieved healthy training references. RELATE-MLP directly classifies this evidence, while RELATE-Gen fuses it with MiniROCKET features into four continuous tokens for structured record generation. The paper also introduces MOSAIC, covering twelve known-setting diagnostic settings. On 3,598 test groups, RELATE-MLP achieves 80.60% setting-macro accuracy, and RELATE-Gen achieves 68.86% setting-macro Joint EM.

Critical strengths
  1. RELATE compares query features with retrieved healthy references. This comparison improves generator validation accuracy by 5.07 percentage points on average across three runs.
  2. RELATE-Gen compresses fused FISHER and MiniROCKET features into four continuous tokens, allowing one decoder to generate diagnostic records across twelve settings using their respective candidate labels.
Critical weaknesses

1.Appendix A reports that 2,395 of 2,998 identifiable test groups share acquisition families with training. Please distinguish shared operating configurations from shared recordings, clarify whether the latter satisfy RMIS split rules, and report performance separately for overlapping and non-overlapping families where feasible.

2.The benefit of LLM generation over the token readout is unclear. Their Joint EM scores are similar (68.86% versus 69.07%, Table 2), but generation has substantially higher latency and memory use (Table 5). Please explain what additional diagnostic capability justifies this cost.

Questions for authors

For the acquisition-family overlap reported in Appendix A, which cases involve independent recordings under the same operating conditions, and which involve data from the same recording? How do these cases relate to RMIS split rules, and does performance differ between overlapping and non-overlapping families? Given the similar Joint EM of generation and the AR-token readout (Table 2), what practical diagnostic benefit does generation offer that warrants its higher latency and memory use (Table 5)?

2 comments
AuthorsOpenReview ↗

Response to YFef (1/2): Q1 — split units and source sharing

Thank you for your careful review and constructive feedback.

Substantive revisions in the manuscript are marked in blue.

Q1 What is shared across splits, and how do the two subsets perform?

Thank you for pointing out the need to clarify the data split. RMIS splits data by the same sensor channel of the same recording. We checked the manifests actually used: segments from the same recording and channel do not cross the training/test boundary, while different channels from the same original multichannel file can appear on opposite sides.

To show how the methods perform under this sharing relationship, we report results separately for the 2,998 RMIS test groups: 2,395 share original files with training through different channels, and the remaining 603 contain 1,019 input views. Keeping the models and predictions fixed, we compute accuracy with equal weight for each benchmark group:

Method2,395 groups sharing original files with trainingRemaining 603 groups
RELATE-MLP95.62%88.56%
FISHER 1-NN94.28%87.40%
RELATE-Gen92.99%86.24%

RELATE-MLP exceeds 1-NN by 1.34 and 1.16 percentage points in the two subsets, respectively. The subsets differ in their settings and class composition, so we compare methods within each subset. We will provide the exact training/validation/test manifests, source-identifier definitions and scoring code in the anonymous supplementary materials.

AuthorsOpenReview ↗

Response to YFef (2/2): Q2 — generation, readout and cost

Continuing our response to Reviewer YFef.

Q2 What does generation contribute relative to its cost?

For the fixed-schema task evaluated here, we recommend the AR readout when deployment cost matters. It uses the same four signal tokens learned through structured-record generation as the Qwen decoder.

Record accuracy and decoding cost. Gen and the AR readout have three-seed common-taxonomy Joint EM of 67.43±1.36% and 68.20±0.75%, respectively. Field EM is 78.90±0.90% and 79.70±0.50%. Each model and seed covers the same 4,430 views, 3,598 groups and 12 settings. Joint EM averages the twelve setting scores equally, while Field EM averages the seven supported fields. All ± values are sample SDs across training seeds 111/222/333.

DecoderMedian latency per viewPeak allocated GPU memory
Gen4,789.82 ms, approximately 4.79 s5.428 GiB
AR readout23.61 ms0.078 GiB

These measurements use one NVIDIA A10, 92 validation views, batch size one and cached evidence. They include the adapter and decoder, with feature extraction and reference retrieval outside the timing scope. The readout makes the generation-trained representation available for low-cost field prediction.

A paired comparison of the complete records finds 29 more fully correct Gen records than AR records at seed 111, 13 more at seed 222 and one fewer at seed 333. These counts give every group equal weight. Under setting-macro weighting, the three UMGED settings contribute −1.25 points to the three-seed mean Gen-minus-AR Joint EM difference, while the other nine settings together contribute +0.48 points, summing to −0.77 points over all twelve settings. Both summaries describe the same predictions.

The two severity fields show where the output routes differ. Every seed includes all 300 damage-size groups and 76 eccentricity groups, with equal-field Severity EM weighting the two fields equally.

FieldGen mean±SDAR readout mean±SDPaired Gen−AR difference, pp
Damage size74.44±2.83%72.11±1.50%+2.33±3.79
Eccentricity24.12±4.02%32.02±2.74%−7.89±5.74
Equal-field Severity EM49.28±2.85%52.06±1.75%−2.78±3.84

Gen's damage-size differences are +6.67, +0.67 and −0.33 points across seeds. Eccentricity favors the readout in all three. The fixed seed-111 damage-size difference has a paired group-bootstrap 95% interval of [2.31, 10.97]. Its eccentricity and equal-field severity differences are −5.26 [−13.16, 2.63] and +0.70 [−3.95, 5.26]. Those intervals condition on the trained models, separately from the seed variation in the table.

Learning the representation. With the same readout architecture and train-only taxonomy rule, three-seed test Joint EM is 68.20±0.75% for AR-trained tokens and 66.73±0.79% for directly supervised tokens. The paired differences are +2.37, +1.84 and +0.20 points, averaging +1.47±1.13. The test gains in all three seeds support using RELATE's autoregressive training scheme to learn signal tokens for record prediction. AR trains the signal adapter and Qwen LoRA through record generation, while Direct trains its adapter through condition and field supervision. The common readout evaluates the frozen representations learned by these complete training schemes.

On validation, the corresponding means are 65.65% and 65.52%, a +0.13-point difference with a paired 95% interval of [−2.08, 2.35]. The interval uses 10,000 shared within-setting group-bootstrap resamples with six fixed checkpoints, three per model, and describes validation-group uncertainty separately from training-seed SD.

Qwen receives the four signal tokens together with machine context, natural-language instructions and candidate conditions. This puts signal evidence directly into the language model's input. Future work can extend this route to follow-up questions about a signal, combining signal evidence with maintenance-manual passages, or selecting tools and work-order actions. Both output record types can also be passed to downstream language systems.

The record and field comparisons are in revised Table 2 and Appendix G.

Reviewer_iFwmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This is a time-series representation learning problem, primarily focused on the industrial domain. The paper discusses a pipeline to learn representations of healthy and anomalous samples.

Critical strengths

The paper addresses an interesting problem, traditionally known as unsupervised anomaly detection. In general, the paper is in the right direction; however, as highlighted in the weak sections, the paper needs significant improvement.

Critical weaknesses
  • I started reading the paper by looking at the references and working backward. I appreciated the author's care in ensuring all references are well documented. I loved it and suggest a few more improvements to make it better: a) I found that the field is evolving faster, and some cited papers are a bit outdated. I suggest searching the internet for PHM, a community in the industrial domain focused on asset health, and you will find many papers like PDMBench, PHMBench, ITFormer, PHM-Health, etc. I also noted that the author's primary area is unsupervised representation, so I will base my evaluation on that.

  • The conclusion talks about the Qwen model, which makes me wonder why. The talk is about the relation of the paper’s title to the word model? This is my internal thinking and does not affect my evaluation. The conclusion could be improved; does it really align with the paper title?

  • Section 6.6 is only four lines, and the author did not make the effort to ensure the reader can understand several terms like BearLLM, or what is in Appendix C. It is unusual for a paper to cite a reference in the last paragraph of the experimental work.

  • Section 6.5 - In general, please don't divert us to the appendix; briefly tell us what it covers so we're motivated. Seems a lot of observation is written by an LLM and not in full sentences. The same applies to lines 445-449.

  • Table 4: please highlight the highest value (bold it)

  • Section 6.4 did a good job of conducting an ablation study. It gives insight into why certain things are happening. The conclusion is missing, as is an explanation of why certain things are happening; any retrospective analysis in Section 6.4 will help. A similar style is noticed in Section 6.3. We get the outcomes of the experiments but not the reasoning. The same applies to Section 6.2. Table 2 also lacks bold numbers, as in Tables 1 and 3.

  • Section 6.1 has a lot of redirection. The results need a clearer presentation style. Following this convention, Section 5 is good but needs to be linked to the corresponding section in Section 6. Only one table, Table 1, is not very well explained: the Setting (especially current, bearing, vibration, etc. I looked at Figure 1; there is some confusion in the representation: multiple (b) and (c), one inside (a) and another outside (a).

  • I suggest that a relevant paper (ITFormer) also fine-tuned and/or refined the representation for anomaly detection in the given multivariate time series data. Section 4 also has a paragraph that could define the training pipelines, etc., and a table that captures the parameters, which could help link to Section 5 for an ablation study. This part is missing. Also, while describing the methodology, what aspect is innovative? The intuition is missing. A scientific paper needs to give us a clear picture of what it does and why. This is about technical innovation.

  • In Section 3, define the problem formulation and describe the MOSAIC dataset. Much of the text is confusing. Adding a few MoSAIC examples could make the paper much clearer, such as Slide, Valve, and Gearbox, along with their health states and fault mechanisms. This dataset description provides a solid foundation for understanding the topic. Ideally, a table with a few examples would improve the paper. The first paragraph is unclear and poorly written. It might be easier to explain this using MOSAIC as an example. Given that this paper is based on MOSAIC, I did not see any state-of-the-art accuracy reported for the same dataset and setting.

  • The Introduction section should include an example input query (see ITFormer), as well as a few examples of predictions on the test dataset, explaining why they were tough earlier and how the proposed method solved them. Typically, representation papers showcase what the embedding layer tends to perform well. One common issue in the introduction is not citing a concept the first time it's mentioned; for example, the introduction doesn't cite MiniROCKET.

  • Figure 1, Structured record, should be explained.

  • In the introduction, I would have expected you to include recent papers focused on Industrial monitoring (to give a status update on what is going on, especially on agentic approaches, etc.). Please use the term " Industrial Diagnostic with LLM and you will get plenty of recent work in the AI community. As well, I feel general-purpose Anomaly Detection is also missing (Fault Diagnostic) as a baseline like DAEMON,

  • Overall, the paper did a good job; however, I think it needs peer review by a group of experts, as LLM-written experimental work needs a deeper look to be more compelling.

  • No LLM was used to read the paper, write the review, or upload it anywhere. We printed the paper, read it in print, and drafted the review.

Note on post-rebuttal revision: I helped the author improve the paper's clarity and readability to an ICLR-level technical standard. The revised manuscript has a clear problem formulation and flow.

Questions for authors

None. My assessment is final.

5 comments
AuthorsOpenReview ↗

Response to iFwm (1/3): task, contribution and examples

Thank you for your careful review and constructive feedback.

Substantive revisions in the manuscript are marked in blue.

Task, supervision and the meaning of a diagnostic record

The task is supervised condition classification and prediction of source-supported diagnostic attributes. Training supplies condition and record targets, and normal labels define the healthy-reference bank. FISHER stays frozen while the classifier or the signal adapter and Qwen LoRA learn from these targets.

For a bearing query, the model receives a vibration signal, the machine context and candidates such as healthy, inner-race fault, outer-race fault and rolling-element fault. It predicts a condition and supported attributes, for example outer-race fault with 0.6 mm damage. The condition and damage size are targets, while machine type is given context. Gold diagnoses and source identifiers remain outside the input.

The MOSAIC test set contains 4,430 views from 3,598 benchmark groups across twelve settings. Views of a group are first combined into one prediction. Primary accuracy averages the twelve setting accuracies equally. Joint EM requires a group's condition and all supported fields to match, then averages the twelve setting scores equally. A record with the right condition but the wrong damage size receives credit for condition accuracy and fails Joint EM.

Contribution beyond healthy-reference learning

BearLLM uses a fault-free reference to condition an LLM for bearing diagnosis. RELATE constructs and tests health-relative evidence across acoustic, vibration, current and voltage settings, combining a full-evidence classifier, a four-token record model and a shared benchmark protocol.

The frozen FISHER frontend preserves ordered frequency-band positions at the evaluated source sampling rates, so query and reference features are compared at aligned coordinates. A prototype averages up to sixteen distinct training-normal groups. The inputs retain q and r and expose q−r and abs(q−r), giving the learned projection the original states, signed changes and their magnitudes directly. MiniROCKET adds temporal-pattern statistics to the time-averaged FISHER features. Gated fusion combines the two sources and produces four continuous tokens of width 3,584 for Qwen.

On the common benchmark, the fixed MLP reaches 80.60% setting-macro and 91.83% group-micro accuracy, against 75.24% and 89.83% for FISHER 1-NN. In the generator input control, the full reference construction exceeds self-reference by 1.80±1.47 test accuracy points across three seeds. Here ± denotes the sample standard deviation across the three training seeds. With a common readout, AR-trained tokens exceed directly supervised tokens by 1.47±1.13 test Joint-EM points. The benchmark supplies the same labels, groups and scoring rules for comparing these models.

Concrete queries and test predictions

For a bearing query, the model receives vibration, bearing context and the four candidate conditions listed above. It predicts the condition and damage size from the signal evidence. The six saved single-view test predictions are also in Appendix H, Table 40.

Signal/contextAnnotated diagnosisActual prediction
Air-leak sound, four candidates from normal operation to low-pressure vent leakLow-pressure vent leakFISHER 1-NN predicts normal operation, main MLP predicts low-pressure vent leak, and Gen predicts vent leak
Bearing vibrationOuter-race fault, 0.6 mm damageGen predicts outer-race fault, anomalous health, outer-race location and defect mechanism, and 0.6 mm damage
Bearing vibrationInner-race fault, 0.4 mm damageGen predicts the correct condition, health, location and defect mechanism, but 0.6 mm damage. FISHER 1-NN gets the condition correct, while the main MLP predicts healthy
Slider sound, normal/anomalous slider candidatesAnomalous slider operationGen predicts anomalous slider operation, anomalous health and slider machine type
Valve sound, normal/anomalous valve candidatesAnomalous valve operationGen predicts normal valve operation, normal health and valve machine type
Gearbox sound, normal/anomalous gearbox candidatesNormal gearbox operationGen predicts anomalous gearbox operation, anomalous health and gearbox machine type

The air-leak example is a correct MLP prediction where 1-NN predicts normal and Gen confuses the two leak conditions. The two bearing examples show a fully correct damage-size record and a correct condition with an incorrect damage size. These outcomes explain why we report Joint EM alongside condition accuracy.

In the slider, valve and gearbox examples, machine type is already supplied. Their source annotations support health state and machine type, while a specific failure mechanism is unavailable. The slider prediction is correct. The valve and gearbox predictions make opposite health-state errors.

AuthorsOpenReview ↗

Response to iFwm (2/3): training and experimental evidence

Continuing our response to Reviewer iFwm.

What the training and ablations establish

The classifier learns condition labels from the full FISHER representation of the query, reference and their differences. The generator trains the signal adapter and LoRA using structured records while the 4-bit Qwen base stays frozen. The readout learns record fields from the frozen four-token representation.

RouteInput and trained parametersTraining
Main RELATE-MLP24,576 coordinates, one 256-unit hidden layer, 6.30M parameters80 group-level epochs, AdamW, batch 256, learning rate 10^-3
RELATE-GenFour 3,584-wide tokens, trainable signal adapter and LoRA, frozen 4-bit Qwen2.5-7B base100,424 view microsteps, batch 1, accumulation 8, LoRA rank 8 and alpha 16
Frozen-token readoutThe same four tokens, 7.74M parameters100,424 view microsteps with the matched sampler

The input construction has different measured effects for the two predictors. With initialization and recipe matched within each MLP seed, full minus pair-only test accuracy is +0.56±1.28 points, while validation accuracy is close at 76.08% / 76.16%. Gen gains 1.80±1.47 points over self-reference across three test seeds, supporting use of the complete healthy-reference construction for that predictor.

Adding MiniROCKET together with learned fusion improves AR-objective validation accuracy by 4.56 points over FISHER-only evidence across three runs. This measures the benefit of the combined temporal branch and fusion change. MiniROCKET alone reaches 35.67% validation accuracy. In the single-seed validation token sweep, four tokens reach 75.75% and eight reach 75.36%, a 0.39-point difference. We use four for half the prefix length with slightly higher observed validation accuracy. Separately, the final four-token configuration has a three-run validation accuracy SD of 1.20 points.

To test attribute prediction beyond a condition lookup, the validation analysis uses 88 groups from nine setting/condition/field combinations with multiple training values, across SDUST bearing and the three UMGED settings. Each baseline looks up the training-mode attribute value for its model's predicted condition, and evaluation averages groups within a combination, combinations within each field, then the two fields equally. Across three fixed checkpoints per model, AR scores 49.43% and Direct 48.96%, versus 18.75% and 19.31% for their respective lookup baselines.

For record prediction, the readout on RELATE's generation-trained tokens reaches 68.20±0.75% test Joint EM, compared with 66.73±0.79% for directly supervised tokens under the same readout architecture. The paired test gains are positive in all three seeds and average +1.47±1.13 points. Direct learns its adapter through condition and field supervision. Once learned, the same AR-trained tokens support Qwen generation and lightweight field prediction, giving two deployment options for the signal representation.

On one NVIDIA A10 with 92 cached validation views and batch size one, generation takes a median 4,789.82 ms with 5.428 GiB peak allocated memory, versus 23.61 ms and 0.078 GiB for the readout. The timing includes the adapter and decoder and excludes feature extraction and reference retrieval. The framework therefore offers a low-cost readout option for fixed-schema deployment using the same learned signal interface. Four sensor tokens also enter Qwen with machine context, instructions and candidates. Interactive diagnosis, maintenance-manual retrieval, assistants and tool or work-order workflows are potential extensions, and both record types can serve downstream language systems.

AuthorsOpenReview ↗

Response to iFwm (3/3): related work and presentation

Continuing our response to Reviewer iFwm.

Recent PHM work and comparable baselines

  • ITFormer connects temporal encoders to a frozen LLM for time-series question answering on EngineMT-QA. It informs the design of a learned signal/text interface. A matched comparison would require adapting its question-answer targets to MOSAIC’s condition/attribute labels under the same partitions. ITFormer, ICML 2025
  • PDMBench provides predictive-maintenance datasets, models and evaluation infrastructure for classification and remaining-life prediction. It is relevant to reproducible PHM evaluation and classification baselines. PDMBench
  • PHMbench and PHM-Vibench organize PHM workflows, including vibration diagnosis, remaining-life prediction and anomaly detection. PHM-Bench separately evaluates PHM-oriented large-model capabilities. PHMbench PHM-Vibench PHM-Bench
  • DAEMON learns normal multivariate patterns with an adversarial autoencoder and detects anomalies from reconstruction evidence. We have added it to the related-work discussion to distinguish normal-only anomaly detection from MOSAIC's supervised condition-classification task. Table 1 compares condition-classification baselines under the same supplied labels and fixed splits. DAEMON
  • PHMForge evaluates LLM agents executing PHM tools and coordinating prognostic workflows, providing context for the potential tool-using applications of a signal-conditioned language model. PHMForge

Table 1 compares the implemented signal classifiers, pretrained representations and adapted LLM methods on the same MOSAIC labels, frozen partitions and 3,598 benchmark groups. The main MLP exceeds FISHER 1-NN by 5.36 setting-macro points with a paired fixed-model 95% interval of [2.51, 8.43], and by 2.00 group-micro points.

BearLLM and CWRU. BearLLM is a reference-conditioned bearing-diagnosis system. Our local comparison uses the same ten-class candidates and 222 non-overlapping, one-second, 12-kHz CWRU test segments for both adaptations, with 1,554 training and 444 validation segments. RELATE-CWRU obtains 83.33% accuracy / 80.95% macro F1, compared with 43.24% / 31.49% for adapted BearLLM. RELATE-CWRU uses a frozen FISHER-relative branch with a training-normal reference, Qwen2.5-7B and a training-only auxiliary primary-condition head removed at inference. BearLLM uses its released signal adapters and DCN with Qwen2.5-1.5B. This conclusion concerns the two complete adaptations with their stated encoders, model sizes and training recipes.

Presentation and manuscript changes

The revised Figure 1 uses unique panel labels and explains how the healthy-reference and fusion panels relate to the overview. Its structured-record card contains selected fields from an actual generated validation prediction. The caption identifies its predicted condition, health state and mechanism, distinguishes them from schematic feature symbols, and defines the frozen/trainable markers.

The revised text explains Table 1's source/modality names locally, including PU bearing current/vibration and the UMGED current/vibration/voltage settings. Tables 2 and 4 show the best value for each metric in bold. The parameter table separates the 80-epoch primary classifier from the 100,424-microstep models. Section 5 links each research question to its results sections. The results discussions identify the tested contrasts and explain what their values support.

MiniROCKET is cited at first mention. The CWRU paragraph defines BearLLM and identifies Appendix C as the CWRU Bearing Comparison before the cross-reference. The introduction describes a candidate-conditioned diagnostic input and its output fields. The conclusion uses RELATE's model names and the supervised known-setting scope, with Qwen identified as the decoder implementation.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.