Paper deep dive
Candidate-Fate Accounting for Transparent Sensor Diagnostic Pipeline Search
Haotao Xie, Yutian Chen, Yangqi Liu, Xiaoyu Jiang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 8/20/2026, 4:52:46 AM
Summary
The paper introduces 'candidate-fate accounting,' a framework for auditing automated machine learning (AutoML) search processes in industrial sensor diagnostics. It addresses the gap where existing AutoML reports omit non-fitted candidates (invalid, pruned, skipped, cached), limiting transparency. The framework uses typed legality checks, stable hashing, and a closed fate ledger to assign terminal fates to all observed candidates, ensuring complete and auditable trace records. Experiments on bearing-diagnostic datasets demonstrate its ability to detect invalid candidates and account for omitted trials while maintaining diagnostic performance.
Entities (8)
Relation Signals (7)
Candidate-Fate Accounting → evaluatedon → Ottawa Dataset
confidence 95% · Experiments on three bearing-diagnostic datasets... Ottawa
Candidate-Fate Accounting → evaluatedon → CWRU Dataset
confidence 95% · Experiments on three bearing-diagnostic datasets show that the framework detects invalid candidates... CWRU
Candidate-Fate Accounting → evaluatedon → SEU Dataset
confidence 95% · Experiments on three bearing-diagnostic datasets... SEU
Candidate-Fate Accounting → uses → Hash-based Ledger
confidence 95% · a hash-based ledger consolidates repeats and assigns one terminal fate—including non-fitted outcomes—to each observed candidate.
Candidate-Fate Accounting → uses → Typed Legality Check
confidence 95% · Typed legality checks classify generated candidates; a replaceable allocator prioritizes eligible ones
Candidate-Fate Accounting → addressesproblemof → AutoML
confidence 90% · existing automated machine/deep learning (AutoML/AutoDL) reports typically retain only fitted trials... omitting generated candidates
Guided Ledger Search (GLS) → isallocatorfor → Candidate-Fate Accounting
confidence 90% · guided ledger search is one allocator designed in our experiments
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Industrial sensor diagnostics relies on preprocessing, representation, and classification pipelines, making automated pipeline search useful for reducing manual design cost. However, existing automated machine/deep learning (AutoML/AutoDL) reports typically retain only fitted trials, scores, and winners, omitting generated candidates that are invalid, pruned, skipped, cached, or unfitted. This omission limits reviewers' ability to check signal constraints, budget use, and unevaluated legal alternatives. To address this, we propose candidate-fate accounting, a candidate-level audit framework for diagnostic search traces. It records each observed candidate as auditable evidence: hashes merge repeated observations, legality checks flag invalid candidates, allocation rationales explain budget decisions, and a closed fate ledger assigns one terminal fate to each candidate. Experiments on three bearing-diagnostic datasets show that the framework detects invalid candidates and identifies 30--41 candidates omitted by fitted-trial-only reports, with closed fate records verifying complete candidate accounting while maintaining competitive diagnostic performance. The code is available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2608.18665v1
- Canonical: https://arxiv.org/abs/2608.18665v1
Trouble viewing inline? Open PDF directly →
Full Text
35,277 characters extracted from source content.
Expand or collapse full text
Candidate-Fate Accounting for Transparent Sensor Diagnostic Pipeline SearchDOI: 10.1145/3799682.3839938Conference: Proceedings of the 35th ACM International Conference on Information and Knowledge Management; November 7–11, 2026; Rome, ItalyProceedings of the 35th ACM International Conference on Information and Knowledge Management (CIKM ’26), November 7–11, 2026, Rome, ItalyISBN: 979-8-4007-2539-5/2026/11CCS: Computing methodologies Machine learningCCS: Information systems Data miningCCS: Applied computing Industry and manufacturing Haotao Xie OrcID: 0009-0002-3061-8090 Affiliation: Hangzhou International Innovation Institute, Beihang University , Hangzhou , China email: haotaoxie@buaa.edu.cn , Yutian Chen OrcID: 0009-0006-6114-8756 Affiliation: Hangzhou International Innovation Institute, Beihang University , Hangzhou , China email: 25601153@buaa.edu.cn , Yangqi Liu OrcID: 0009-0003-7688-1190 Affiliation: College of Cyber Security, Jinan University , Guangzhou , China email: yangqilau@163.com and Xiaoyu Jiang OrcID: 0000-0003-4170-5579 Affiliation: Hangzhou International Innovation Institute, Beihang University , Hangzhou , China email: jiangxiaoyu@buaa.edu.cn 2026© , 2026; Abstract. Industrial sensor diagnostics relies on preprocessing, representation, and classification pipelines, making automated pipeline search useful for reducing manual design cost. However, existing automated machine/deep learning (AutoML/AutoDL) reports typically retain only fitted trials, scores, and winners, omitting generated candidates that are invalid, pruned, skipped, cached, or unfitted. This omission limits reviewers’ ability to check signal constraints, budget use, and unevaluated legal alternatives. To address this, we propose candidate-fate accounting, a candidate-level audit framework for diagnostic search traces. It records each observed candidate as auditable evidence: hashes merge repeated observations, legality checks flag invalid candidates, allocation rationales explain budget decisions, and a closed fate ledger assigns one terminal fate to each candidate. Experiments on three bearing-diagnostic datasets show that the framework detects invalid candidates and identifies 30–41 candidates omitted by fitted-trial-only reports, with closed fate records verifying complete candidate accounting while maintaining competitive diagnostic performance. The code is available at https://github.com/XXIE999/candidate-fate-accounting. Keywords: industrial sensor diagnostics, search transparency, candidate-fate accounting †c-license: by Figure 1. Motivating comparison among expert-designed workflows, AutoML/AutoDL search, and candidate-fate accounting.A three-part diagram contrasts expert modeling, AutoML/AutoDL search, and candidate-fate accounting. Expert modeling is shown as interpretable and checkable but manually costly; AutoML/AutoDL is shown as automated but limited in search-process visibility; candidate-fate accounting is shown as tracking generated candidates through legality checks, allocation rationales, and final fates. 1. Introduction Industrial sensor diagnostics depends on pipelines that convert machine signals into reliable fault decisions (31). In deployed maintenance settings, these pipelines must be both accurate and inspectable: engineers need to know which signal transforms are admissible, which classifiers can consume each representation, and why a selected model is plausible (14; 25). Expert-built pipelines provide this reviewability, but the same reasoning is costly to repeat across machines, channels, and operating regimes (32). Automated pipeline search (20; 5; 24) reduces this manual burden by generating and evaluating candidate preprocessing, representation, and classification pipelines. However, automation should not remove the evidence needed to review the search process. A useful diagnostic search report should show not only which pipeline won, but also what happened to generated candidates that were never fitted. Figure 1 summarizes this contrast. Existing work (13; 33; 19) addresses parts of this problem but does not close the generated-candidate audit gap. AutoML and AutoDL methods report fitted configurations; hyperparameter-optimization frameworks such as Optuna record scheduled-trial states such as completed, failed, or pruned trials (1; 12; 26; 7; 16); grammar and validity methods reject illegal programs; and provenance systems track executed artifacts (22; 9; 10; 29; 8; 18; 21; 2; 27). These tools are useful, but their reporting units are usually fitted configurations, scheduled trials, rejected programs, or executed artifacts rather than canonical generated candidates. As a result, repeated proposals, type-invalid candidates that never become trials, and legal candidates skipped by budget or cache decisions may lack one terminal explanation, leaving search validity and budget use difficult to audit at the candidate level. Candidate-fate accounting complements rather than replaces these systems: it adds a canonical candidate-level partition that also covers invalid and non-executed alternatives. This common reporting unit supports reproducible, optimizer-independent comparison of legality and budget traces across AutoML search policies. A simple diagnostic trace illustrates the gap. Suppose a search generates a short-time Fourier transform (STFT) map with logistic regression, a z-score plus log-mel support vector machine (SVM) pipeline, and a highpass raw-vector XGBoost pipeline, but fits only the third candidate. A fitted-trial report records only the evaluated pipeline and score, hiding that the first is type-invalid and the second is legal but cost-blocked. These hidden outcomes matter under small budgets because they expose legality checks, budget allocation, cache reuse, and untested legal alternatives. To address this gap, we design candidate-fate accounting, a candidate-level audit framework for emitted diagnostic search traces. The framework closes the record over observed canonical candidates rather than enumerating candidates that never appear. It uses stable hashes to merge repeated observations, typed legality checks to expose type and semantic failures before fitting, allocation rationales to explain budget decisions for legal candidates, and a closed fate ledger with Δclose _close to verify that each observed candidate receives exactly one terminal fate. Thus, non-fitted candidates become reportable evidence rather than optimizer bookkeeping. Allocation policies remain replaceable: guided ledger search is one allocator designed in our experiments, while candidate-fate accounting is the paper’s main contribution. Figure 2 summarizes the workflow. We evaluate candidate-fate accounting on three bearing-diagnostic datasets: the Case Western Reserve University (CWRU) bearing dataset, the University of Ottawa bearing dataset (Ottawa), and the Southeast University (SEU) bearing dataset. The experiments test whether the framework detects invalid candidates before fitting, accounts for candidates omitted by fitted-trial-only reports with complete fate accounting, and preserves useful diagnostic performance under a controlled protocol. The main contributions are: (1) We define the generated-candidate audit gap in automated industrial diagnostic search, showing why non-fitted candidates should be treated as evidence about validity, budget use, and untested legal alternatives. (2) We design a typed candidate-fate accounting framework that maps observed canonical candidates to auditable legality, allocation rationale, terminal fate, and closure evidence. (3) We evaluate the framework on CWRU, Ottawa, and SEU through invalid-candidate probes, closed fate ledgers, allocation checks, and controlled-protocol diagnostic performance. Figure 2. Overview of candidate-fate accounting. Typed legality checks classify generated candidates; a replaceable allocator prioritizes eligible ones; and a hash-based ledger consolidates repeats and assigns one terminal fate—including non-fitted outcomes—to each observed candidate. The closure condition Δclose=0 _close=0 verifies complete, non-overlapping accounting.A three-stage candidate-fate accounting workflow. First, a generated diagnostic program is decomposed into preprocessing, signal-representation, and estimator components and checked for input/output compatibility and semantic validity. Second, a budget allocator uses replaceable search strategies to prioritize generated, valid, uncached candidates that have not received a terminal fate. Third, normalized candidate hashes merge repeated observations into a ledger that assigns one of six fates: invalid, duplicate or cache, pruned, skipped, full-evaluated, or residual legal-unfitted. A zero closure gap yields auditable search evidence. 2. Method We formalize candidate-fate accounting as a reporting contract over emitted diagnostic search traces. As shown in Figure 2, the framework records typed legality evidence, allocation rationale, and terminal fates in closed fate ledger. These records make signal constraints, budget use, and unevaluated legal alternatives inspectable while keeping the audit layer optimizer-independent. Section 2.1 defines legal diagnostic candidates, Section 2.2 defines budget allocation, and Section 2.3 assigns final fates. 2.1. Typed Legality Evidence A generated candidate is a diagnostic program c=(p,r,e)c=(p,r,e), where p is a preprocessing chain, r is a signal representation, and e is an estimator. The objects exchanged by these components have declared types, such as waveform, vector, time–frequency map, and patch sequence. Each primitive u declares an input type τin(u) _in(u), an output type τout(u) _out(u), and lightweight metadata for semantic checks. The first legality condition is type compatibility: type=c=(p,r,e):τout(p)=τin(r),τout(r)=τin(e).C_type=\c=(p,r,e): _out(p)= _in(r),\, _out(r)= _in(e)\. The set typeC_type contains candidates whose adjacent signatures agree. It captures syntactic feasibility before any search policy is considered. This rule turns signal-type constraints into pre-fit evidence by rejecting errors such as feeding a map representation to a vector-only classifier. A deterministic semantic map ρ:type→0,1ρ:C_type→\0,1\ resolves conflicts by conservative rejection: a missing-value primitive is invalid when dataset metadata reports zero missingness, while a second z-score normalization or filtering step is invalid as redundant. Fixed rule precedence records one reason when checks overlap. The second step distinguishes intrinsic legality from budget admissibility. Given budget B and its cost guard CostOK(c,B)CostOK(c;B), we define Legal(c) (c) ≡c∈type∧ρ(c)=1, ≡ c _type ρ(c)=1, Admissible(c,B) (c;B) ≡Legal(c)∧CostOK(c,B). (c) (c;B). Thus leg=c:Legal(c)C_leg=\c:Legal(c)\ is the legal candidate space independent of budget. Type/semantic-incompatible candidates are invalid before fitting, while legal candidates blocked by cost, cache, or termination remain legal and later receive a non-evaluated fate. 2.2. Allocation Rationale Allocation rationale makes budget use reviewable without changing candidate legality. At search step t, fitting budget (17) may be spent only on candidates that are generated, legal, not cached, and not already assigned a terminal fate. Let tG_t be generated candidates, ℋt−1H_t-1 be cached canonical hashes, h(c)h(c) be the stable hash of candidate c, and ℓ(c) (c) be its current ledger state. The allocatable set is t=c∈t∩leg:h(c)∉ℋt−1,ℓ(c)unassigned.A_t=\c _t _leg:h(c) _t-1,\ (c)\ unassigned\. Search policies operate only on tA_t: RLS (4) samples uniformly, the TPE-style sampler (3) adapts primitive distributions from observed validation macro-F1, and GLS (15) uses MCTS-style selection over grammar-approved macro actions. They change only legal-candidate priority, not legality, budget eligibility, cache handling, or terminal fate. Optional training-split, label-free descriptors such as impulsiveness, spectral entropy, support, and missingness define rguider_guide by adjusting primitive-family priority, but cannot bypass legality or ledger rules. Table 1. Closed B=50B=50 GLS ledger counts under typed legal search, averaged over three seeds. Dataset Unique Generated Invalid Pruned Skipped Duplicate/ Cache Full Evaluated Residual Legal- Unfitted Closure Gap CWRU 49.7±0.549.7±0.5 0±00±0 6.7±0.96.7±0.9 23.0±2.823.0±2.8 0±00±0 20.0±2.920.0±2.9 0±00±0 0 Ottawa 49.7±0.549.7±0.5 0±00±0 20.3±2.920.3±2.9 18.7±4.718.7±4.7 2.3±0.92.3±0.9 8.3±2.18.3±2.1 0±00±0 0 SEU 49.7±0.549.7±0.5 0±00±0 6.3±2.16.3±2.1 23.7±3.123.7±3.1 4.3±2.14.3±2.1 15.3±2.515.3±2.5 0±00±0 0 2.3. Closed Fate Ledger The closed ledger reports the fate of every observed canonical candidate, including candidates that were generated but never fitted. The reporting unit is a canonical hash because one diagnostic program may appear through proposal, cache, probe, or evaluation records. In our implementation, h(c)h(c) is a SHA-256-derived identifier over a deterministic sorted-key serialization of the ordered preprocessing steps, representation, estimator, and their parameters. Each hash bucket stores the serialized candidate, legality result, allocation reason, and observed event types. Repeated proposal, cache, probe, or evaluation records update the existing bucket rather than increasing U. Because component order and parameters are included, the same configuration maps to one identity independently of its event path, while structurally different pipelines remain distinct. Ledger construction therefore separates identity resolution from fate assignment: records are first consolidated by h(c)h(c), precedence is then applied per bucket, and closure is computed over U unique entries. This prevents repeated observations from inflating generated-candidate counts while preserving the event evidence needed to justify the final state. Let ¯T G_T be the observed canonical generated set at the end of a run. The fate function ℓ:¯T→ : G_T maps each candidate to one state in =sinv,sdup,sprn,sskp,seval,sresS=\s_inv,s_dup,s_prn,s_skp,s_eval,s_res\: invalid, duplicate/cache, pruned, skipped, full-evaluated, or residual legal-unfitted. Fates are assigned by a fixed precedence order so that each observed canonical candidate contributes to one table cell. The order makes late evidence decisive when a candidate is first proposed cheaply and later fitted. Candidates failing type or semantic checks become invalid. Any candidate that consumes fitting budget becomes full-evaluated, even if earlier records only proposed or probed it. Legal candidates observed only through cache reuse become duplicate/cache. Legal non-duplicates removed by cost, complexity, or multi-fidelity guards become pruned. Legal candidates selected for execution but blocked before fitting become skipped. Remaining legal non-duplicates that are observed but never fitted become residual legal-unfitted. The closure check tests whether terminal fates form a mutually exclusive and exhaustive partition of the observed set: ¯T=⋃˙s∈¯Ts,Δclose=|¯T|−∑s∈|¯Ts|=0. G_T= _s G_T^s, _close=| G_T|- _s | G_T^s|=0. Here ⋃˙ denotes a disjoint union, ¯Ts G_T^s is the subset assigned fate s, and Δclose _close is zero only when no observed canonical candidate is missing or double-counted. For each observed canonical candidate, the evidence tuple is ℰ(c)=(rtype(c),rguide(c),ℓ(c)),E(c)=(r_type(c),r_guide(c), (c)), where rtyper_type is the legality rationale, rguider_guide is optional allocation metadata or rationale, and ℓ(c) (c) is the terminal fate. The run-level report is ℛT=(c⋆,ℰ(c):c∈¯T,BT,Δclose).R_T=(c ,\E(c):c∈ G_T\,B_T, _close). Here c⋆c is the selected pipeline and BT=|¯Tseval|≤B_T=| G^s_eval_T|≤ B. This report states which pipeline won, why candidates were ruled out, where budget was spent, and which legal alternatives remained unseen by fitting. For N trace records, U unique candidates, and serialized candidate size L, ledger construction costs O(NL)O(NL) hashing plus expected O(N)O(N) hash-table updates; the in-memory unique-candidate ledger needs O(U)O(U) storage, and closure is O(U)O(U), with no additional model fitting. Each incoming record requires one hash and an expected O(1)O(1) lookup/update. In our B=50B=50 GLS runs, the main JSON ledger occupies 32.8–37.7 KB per run; exact disk use depends on schema and serialization. At the accounting layer, a new domain supplies a canonical serializer, type/semantic rules, and a mapping from search events to the six terminal fates; hash consolidation, precedence, and closure remain unchanged. 3. Experiments We evaluate four research questions (RQs) aligned with the audit framework: whether typed legality exposes invalid candidates before fitting (RQ1), whether the candidate-fate ledger closes over observed generated candidates (RQ2), whether allocation behavior is comparable under the same legal space (RQ3), and whether audited search still returns useful diagnostic pipelines under a controlled protocol (RQ4). 3.1. Experimental Setup Protocol. We use a controlled small-budget protocol to compare audit counts and diagnostic utility across search policies. For each dataset and seed, all main search conditions share the same typed legal space, primitive cost model, split, and macro-F1 metric, with at most B=50B=50 full model-fitting attempts. This cap applies to fitted candidates, not generated candidates: generated candidates may instead be skipped, pruned, cached, or recorded as residual legal-unfitted. Results are averaged over seeds 42, 43, and 44. Datasets. We evaluate three bearing-diagnostic datasets with dataset-appropriate window-level splits. CWRU (11; 23; 28) and Ottawa use approximately 60/20/20 train/validation/test splits. SEU uses a cross-condition split 30_2→20_030\_2→ 20\_0, with the source condition for training and the target condition split equally for validation and testing. 3.2. Typed Legality Probe RQ1 tests whether typed legality exposes invalid candidates before fitting. We generate 100 weakly constrained skeletons over the shared primitive inventory and label each skeleton with type and semantic checks. The probe yields 48 type-invalid, 20 semantic-invalid, and 32 legal skeletons. Type failures capture object mismatches such as time–frequency maps followed by vector-only classifiers; semantic failures capture conservative metadata conflicts such as duplicate normalization or repeated filtering. The 68/100 invalid count is a legality-layer stress test, not an estimate of typed main-run invalidity: because the main searches operate within the typed legal space, zero invalid candidates there is expected. Rather than weakening the audit claim, this zero-invalid main-run outcome makes the invariant checkable: the ledger confirms that invalid candidates do not consume fitting budget, and the weak probe identifies the failures that less constrained generation would need to catch and explain. 3.3. Closed Ledger Accounting RQ2 tests whether candidate-fate accounting accounts for emitted candidates dropped by fitted-trial reporting and verifies complete accounting over observed canonical candidates. Each record stores legality evidence, allocation rationale when available, a terminal fate, and a reason. Table 1 gives a B=50B=50 GLS accounting example aggregated over three seeds. The main finding is that fitted-trial-only reporting omits substantial trace evidence: the ledger records about 30, 41, and 34 non-full-evaluated canonical candidates on CWRU, Ottawa, and SEU, respectively. The closure check verifies that these fate counts form a complete and non-overlapping partition, with all GLS ledger rows in Table 1 satisfying Δclose=0 _close=0. For example, fitted-trial-only reporting on Ottawa exposes only 8.3±2.18.3±2.1 full-evaluated candidates, whereas the ledger attributes 41.341.3 others on average to pruning, skipping, or duplicate/cache. The latter comprise 20.320.3 pruned, 18.718.7 skipped, and 2.32.3 duplicate/cache candidates, so the shortfall from generated candidates to fitted trials becomes attributable rather than unexplained. The interpretability is process-level: legality decisions, candidate fate, and budget consequence, not post-hoc explanations of a fitted model. These fates make the search reviewable but do not rank optimizers. 3.4. Early Allocation RQ3 uses GLS as an allocator case study and compares allocation behavior with the legal space and budget fixed. All rows share typed legality, cache rules, semantic guards, and fitting budget. Best F1 ≤30≤ 30 is the best validation macro-F1 within 30 full evaluations; Target Success counts seeds that reach the final RLS validation score, used as a dataset-specific reference target; Evaluations to Target measures the first full-evaluation index at which a run reaches the final RLS validation score, computed only over successful seeds. Low Target Success therefore indicates unstable early allocation. Table 2. Early allocation under a shared typed legal space. Bold marks the best F1 and, among 3/3-success rows, the fewest Evaluations to Target. Dataset Method Best F1≤ 30 ↑ Target Success Evaluations to Target ↓ CWRU RLS 0.9942± 0.0009 3/3 26.00± 7.87 CWRU TPE 0.9982± 0.0009 3/3 19.00± 4.32 CWRU GLS 0.9982± 0.0002 3/3 7.00± 4.55 Ottawa RLS 0.9673± 0.0063 3/3 9.00± 5.89 Ottawa TPE 0.9639± 0.0038 1/3 3.00± 0.00 Ottawa GLS 0.9856± 0.0026 3/3 3.67± 1.70 SEU RLS 0.6751± 0.0458 3/3 28.00± 3.56 SEU TPE 0.7090± 0.0167 1/3 27.00± 0.00 SEU GLS 0.7646± 0.0185 3/3 7.33± 6.85 Three panels for CWRU, Ottawa, and SEU plot the best validation macro-F1 within 30 full evaluations against Evaluations to Target. Gray circles denote random legal search, orange squares denote Tree-structured Parzen Estimator-style sampling, and blue triangles denote guided ledger search. Guided ledger search lies in the high-F1, low-evaluation region in all three panels. Figure 3. Best validation macro-F1 within 30 full evaluations versus Evaluations to Target; upper-left is better.Three panels for CWRU, Ottawa, and SEU plot the best validation macro-F1 within 30 full evaluations against Evaluations to Target. Gray circles denote random legal search, orange squares denote Tree-structured Parzen Estimator-style sampling, and blue triangles denote guided ledger search. Guided ledger search lies in the high-F1, low-evaluation region in all three panels. Table 2 and Figure 3 summarize early-allocation speed and budget-30 utility. GLS gives the most consistent early-allocation profile: it reaches the random-search target in 3/3 seeds on all datasets, requires the fewest Evaluations to Target among 3/3-success rows, attains the best budget-30 F1 on Ottawa and SEU, and ties TPE to four decimals on CWRU. This should be read as an auditability check, not as a broad optimizer ranking: allocation choices can be reported alongside fitted-trial outcomes under the same legal space. 3.5. Protocol Utility RQ4 asks whether the audit framework retains controlled-protocol diagnostic utility. Table 3 reports controlled-protocol final-test macro-F1 for selected same-space search pipelines, fixed diagnostic recipes, a one-dimensional convolutional neural network (1D-CNN), and AutoML references. Random forest (RF) denotes the estimator used in two fixed recipes. Search rows are matched by split, channel, windowing, label map, metric, and data limit; neural and AutoML rows use different input representations and serve only as scale references. Table 3. Protocol-scoped final-test macro-F1; search rows show three-seed mean (standard deviation), with the best matched search policy per dataset in bold. Method CWRU F1 ↑ Ottawa F1 ↑ SEU F1 ↑ RLS 0.9951(0.0004) 0.9851(0.0046) 0.7130(0.0514) TPE 0.9978(0.0028) 0.9691(0.0119) 0.7128(0.0242) GLS 0.9968(0.0000) 0.9862(0.0032) 0.7341(0.0134) Statistical + RF 0.8200 0.6216 0.0748 Spectral classical 0.7068 0.6181 0.0593 Envelope classical 0.5364 0.5926 0.5899 Wavelet + RF 0.8324 0.7490 0.0901 1D-CNN 0.7414 0.0667 0.2380 AutoGluon (6) 0.9172 0.8776 0.2647 FLAML (30) 0.9022 0.8821 0.1371 Table 3 shows that accountable search retains controlled-protocol diagnostic utility. Same-space search rows achieve high macro-F1 on CWRU and Ottawa; among matched search policies, TPE is highest on CWRU while GLS is highest on Ottawa and SEU. Because selected pipelines vary across settings, the report records the selected family, transform, estimator, and seed. These results support the scoped claim that candidate-fate accounting can expose terminal fates and allocation outcomes while still selecting plausible final models. 4. Conclusion Candidate-fate accounting addresses the generated-candidate audit gap in automated diagnostic search. By merging repeated trace records, recording legality and allocation rationales, and checking ledger closure with Δclose _close, it turns rejected, skipped, cached, and unfitted candidates into reviewable evidence. Experiments on CWRU, Ottawa, and SEU show that the framework exposes invalid skeletons, attributes non-full-evaluated candidates to explicit fates, and retains useful final-test performance under a controlled protocol. The accounting contract can be adapted through domain-specific candidate schemas and semantic rules, but empirical cross-domain validation remains future work. Acknowledgements. This work was supported in part by the National Natural Science Foundation of China under Grant 62403425, in part by the Zhejiang Provincial Natural Science Foundation of China under Grant LMS26F030019, in part by the Hangzhou Natural Science Foundation under Grant 2025SZRJJ2330, in part by the Youth Talent Support Project of the Zhejiang Provincial Association for Science and Technology, and in part by the Jiangsu Provincial Scientific Research Center of Applied Mathematics under Grant BK20233002. Generative AI (GenAI) Usage Disclosure The authors used generative AI tools in a limited assistive capacity for manuscript language polishing and for code debugging/checking. These tools were not used to generate experimental data, alter results, or make scientific decisions; all code, results, technical claims, and final manuscript text were reviewed and verified by the authors. References Akiba et al. (2019) T. Akiba, S. Sano, T. Yanase, T. Ohta, and M. Koyama Optuna: a next-generation hyperparameter optimization framework. Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining. External Links: Link Cited by: §1. Belhajjame et al. (2013) K. Belhajjame, R. B’Far, J. Cheney, S. Coppens, S. Cresswell, Y. Gil, P. Groth, G. Klyne, T. Lebo, J. McCusker, S. Miles, J. D. Myers, S. S. Sahoo, and C. Tilmes PROV-dm: the prov data model. External Links: Link Cited by: §1. Bergstra et al. (2011) J. Bergstra, R. Bardenet, Y. Bengio, and B. Kégl Algorithms for hyper-parameter optimization. In Neural Information Processing Systems, External Links: Link Cited by: §2.2. Bergstra and Bengio (2012) J. Bergstra and Y. Bengio Random search for hyper-parameter optimization.. Journal of machine learning research 13 (2). Cited by: §2.2. Drori et al. (2021) I. Drori, Y. Krishnamurthy, R. Rampin, R. Lourenço, J. P. Ono, K. Cho, C. T. Silva, and J. Freire AlphaD3M: machine learning pipeline synthesis. ArXiv abs/2111.02508. External Links: Link Cited by: §1. Erickson et al. (2020) N. Erickson, J. W. Mueller, A. Shirkov, H. Zhang, P. Larroy, M. Li, and A. Smola AutoGluon-tabular: robust and accurate automl for structured data. ArXiv abs/2003.06505. External Links: Link Cited by: Table 3. Falkner et al. (2018) S. Falkner, A. Klein, and F. Hutter BOHB: robust and efficient hyperparameter optimization at scale. ArXiv abs/1807.01774. External Links: Link Cited by: §1. Ferreira et al. (2022) L. Ferreira, A. Pilastri, F. Romano, and P. Cortez Using supervised and one-class automated machine learning for predictive maintenance. Applied Soft Computing 131, p. 109820. External Links: ISSN 1568-4946, Document, Link Cited by: §1. Feurer et al. (2015) M. Feurer, A. Klein, K. Eggensperger, J. Springenberg, M. Blum, and F. Hutter Efficient and robust automated machine learning. In Advances in Neural Information Processing Systems, C. Cortes, N. Lawrence, D. Lee, M. Sugiyama, and R. Garnett (Eds.), Vol. 28, p. . External Links: Link Cited by: §1. Hadi et al. (2023) R. H. Hadi, H. N. Hady, A. M. Hasan, A. A. L. Al-Jodah, and A. J. Humaidi Improved fault classification for predictive maintenance in industrial iot based on automl: a case study of ball-bearing faults. Processes. External Links: Link Cited by: §1. Hendriks et al. (2022) J. Hendriks, P. Dumond, and D. Knox Towards better benchmarking using the cwru bearing fault dataset. Mechanical Systems and Signal Processing. External Links: Link Cited by: §3.1. Hutter et al. (2011) F. Hutter, H. H. Hoos, and K. Leyton-Brown Sequential model-based optimization for general algorithm configuration. In Learning and Intelligent Optimization, External Links: Link Cited by: §1. Hutter et al. (2019) F. Hutter, L. Kotthoff, and J. Vanschoren Automated machine learning - methods, systems, challenges. External Links: ISBN 978-3-030-05317-8, Document Cited by: §1. Jiang et al. (2026) X. Jiang, H. Xie, J. Wang, Z. Yang, Y. Zhou, L. Yao, and Z. Zhu Agentic ai for safety-aware process monitoring and fault diagnosis: a review. Processes 14 (13). External Links: Link, ISSN 2227-9717, Document Cited by: §1. Kocsis and Szepesvari (2006) L. Kocsis and C. Szepesvari Bandit based monte-carlo planning. In European Conference on Machine Learning, External Links: Link Cited by: §2.2. Li et al. (2018) L. Li, K. G. Jamieson, A. Rostamizadeh, E. Gonina, J. Ben-tzur, M. Hardt, B. Recht, and A. Talwalkar A system for massively parallel hyperparameter tuning. arXiv: Learning. External Links: Link Cited by: §1. Li et al. (2016) L. Li, K. G. Jamieson, G. DeSalvo, A. Rostamizadeh, and A. Talwalkar Hyperband: a novel bandit-based approach to hyperparameter optimization. J. Mach. Learn. Res. 18, p. 185:1–185:52. External Links: Link Cited by: §2.2. Marinescu et al. (2021) R. Marinescu, A. Kishimoto, P. Ram, A. Rawat, M. Wistuba, P. Palmes, and A. Botea Searching for machine learning pipelines using a context-free grammar. In AAAI Conference on Artificial Intelligence, External Links: Link Cited by: §1. Mitchell et al. (2018) M. Mitchell, S. Wu, A. Zaldivar, P. Barnes, L. Vasserman, B. Hutchinson, E. Spitzer, I. D. Raji, and T. Gebru Model cards for model reporting. Proceedings of the Conference on Fairness, Accountability, and Transparency. External Links: Link Cited by: §1. Mohr et al. (2018) F. Mohr, M. Wever, and E. Hüllermeier ML-plan: automated machine learning via hierarchical planning. Machine Learning 107, p. 1495–1515. External Links: Link Cited by: §1. Nguyen et al. (2020) T. Nguyen, B. Gabrys, and K. Musial AutoWeka4MCPS-avatar: accelerating automated machine learning pipeline composition and optimisation. Expert Syst. Appl. 185, p. 115643. External Links: Link Cited by: §1. Olson and Moore (2019) R. S. Olson and J. H. Moore TPOT: a tree-based pipeline optimization tool for automating machine learning. In Automated Machine Learning: Methods, Systems, Challenges, F. Hutter, L. Kotthoff, and J. Vanschoren (Eds.), p. 151–160. External Links: ISBN 978-3-030-05318-5, Document, Link Cited by: §1. Rosa et al. (2024) R. K. Rosa, D. Braga, and D. Silva Benchmarking deep learning models for bearing fault diagnosis using the cwru dataset: a multi-label approach. External Links: Link Cited by: §3.1. Salvador et al. (2016) M. M. Salvador, M. Budka, and B. Gabrys Automatic composition and optimization of multicomponent predictive systems with an extended auto-weka. IEEE Transactions on Automation Science and Engineering 16, p. 946–959. External Links: Link Cited by: §1. Shi et al. (2024) L. Shi, J. Guo, and H. Wang A precision machining equipment fault diagnosis based on CWT and improved ResNeXt. Instrumentation 11 (2), p. 36–43. External Links: Document Cited by: §1. Snoek et al. (2012) J. Snoek, H. Larochelle, and R. P. Adams Practical bayesian optimization of machine learning algorithms. In Neural Information Processing Systems, External Links: Link Cited by: §1. Thornton et al. (2012) C. J. Thornton, F. Hutter, H. H. Hoos, and K. Leyton-Brown Auto-weka: combined selection and hyperparameter optimization of classification algorithms. Proceedings of the 19th ACM SIGKDD international conference on Knowledge discovery and data mining. External Links: Link Cited by: §1. Vieira et al. (2025) J. P. Vieira, V. A. Bauler, R. K. Rosa, and D. Silva Towards a more realistic evaluation of machine learning models for bearing fault diagnosis. ArXiv abs/2509.22267. External Links: Link Cited by: §3.1. Wagner et al. (2023) T. Wagner, A. Gepperth, and E. Engels A framework for the automated parameterization of a sensorless bearing fault detection pipeline. Journal of Applied Research on Industrial Engineering 10 (4). External Links: Link, Document Cited by: §1. Wang et al. (2019) C. Wang, Q. Wu, M. Weimer, and E. Zhu FLAML: a fast and lightweight automl library. In Conference on Machine Learning and Systems, External Links: Link Cited by: Table 3. Wang and Liu (2024) H. Wang and X. Liu Research on rotating machinery fault diagnosis based on improved multi-target domain adversarial network. Instrumentation 11 (1), p. 38–50. External Links: Document Cited by: §1. Yang et al. (2026) C. Yang, J. Yan, Y. Feng, L. Li, and J. Tan Hybrid deep learning for hydraulic cylinder fault diagnosis under complex conditions via multi-source signal fusion. Instrumentation 13 (1), p. 40–56. External Links: Document Cited by: §1. Zöller and Huber (2019) M. Zöller and M. F. Huber Benchmark and survey of automated machine learning frameworks. J. Artif. Intell. Res. 70, p. 409–472. External Links: Link Cited by: §1.