Normal-Data-Only Fault Detection and Faulty-Cell Localization in Heterogeneous Electric-Vehicle Battery Packs Using a Hotspot-Suppressed and Alarm-Consistent Dual-Branch Attention Transformer

Abstract

This paper presents HAC-DBAT, a normal-data-only dual-branch Transformer framework for record-level fault recognition and faulty-cell localization in heterogeneous electric-vehicle battery packs. The dataset comprises 88 vehicle records, including 34 normal training records and a held-out test partition of 50 records containing seven labeled faults and 13 labeled faulty cells. On the test partition, the record-level F1-score is 0.9333, the conditional faulty-cell F1-score is 0.8889, and the end-to-end faulty-cell F1-score is 0.8276. The framework integrates validity-masked pack alignment, residual-conditioned dual-branch reconstruction, normal-reference thresholding and hotspot-aware cell ranking. Its evaluation concerns fault-record identification and faulty-cell localization, rather than verified thermal-runaway lead time.

1. Introduction

Faults in electric-vehicle (EV) battery systems pose operational and safety risks [1,2,3]. The data provider labeled the affected records as thermal-runaway-related faults in ternary lithium-ion batteries; these labels do not independently confirm that thermal runaway occurred. Identifying an abnormal operating record also does not establish advanced prediction of thermal runaway. A verified warning lead time requires event-onset and corresponding alarm timestamps, which are not established here. Accordingly, this study evaluates record-level fault detection and faulty-cell localization, rather than the pre-event warning capability investigated in related work [4,5].

Battery fault-diagnosis methods include engineering thresholds, physics-informed models and data-driven approaches [6,7,8]. Studies of thermal runaway also highlight the complexity of battery failure mechanisms [9,10]. Fixed thresholds can be sensitive to operating conditions, detailed mechanism-based models can be difficult to calibrate across vehicle fleets, and supervised classifiers require representative labeled fault examples. Normal-data-only anomaly reconstruction instead learns regular behavior and identifies deviations during inference.

Recurrent sequence models and dropout-based uncertainty estimation provide foundations for multivariate time-series analysis [11,12,13]. Attention and reconstruction methods offer further foundations for anomaly detection [14,15,16]. Stochastic recurrent models and Transformer anomaly detectors, including Anomaly Transformer and TranAD, provide related methods [17,18,19]. USAD and DCdetector illustrate additional unsupervised and attention-based approaches [20,21]. One-class and local-outlier methods offer complementary strategies [22,23]. Nonparametric dynamic thresholding informs alarm calibration [24]. Other attention-based temporal modeling methods also provide relevant context [25,26,27]. HAC-DBAT does not claim novelty in attention itself; it combines validity-masked battery layouts, residual-conditioned reconstruction, normal-calibrated record decisions and hotspot-aware cell ranking. Recent studies of elevator, aircraft fuel-system and railway-actuator diagnosis offer cross-domain methodological context [28,29,30]. Work on nuclear-plant sensor reconstruction, recurrence-based fault diagnosis and marine-engine thermal-fault diagnosis provides additional context [31,32,33]. These studies are not presented as directly comparable EV battery benchmarks.

The dataset includes heterogeneous U95_T34 and U96_T32 pack structures. Naive padding of missing channels can create artificial reconstruction discrepancies; isolated voltage spikes and recurrent high-response channels can also confound localization. HAC-DBAT addresses these concerns by masking structurally absent channels, aggregating window-level evidence, using uncertainty and alarm consistency in record scoring, and suppressing normal-state hotspots during the ranking of eligible voltage-cell channels.

The main contributions are: (i) validity-masked mapping and normal-reference layout compensation for the two observed voltage/temperature configurations; (ii) residual-conditioned dual-branch reconstruction with masked losses, alternating parameter updates and count-independent record decisions; and (iii) record-level fault detection and conditional/end-to-end faulty-cell localization with baseline, ablation and sensitivity comparisons. Both favorable and unfavorable changes are included in the analysis. Figure 1 summarizes these paths: the normal-only reference supports layout alignment, model training, record scoring and cell localization. Voltage windows are reconstructed by the dual-branch Transformer, while the temperature descriptor contributes at record level; stable-alarm windows feed the record decision and, for records classified as faulty, inform faulty-cell localization.

Figure 1. Overall workflow of the normal-data-only HAC-DBAT framework for record-level fault detection and faulty-cell localization. Solid arrows indicate data and alarm-window pathways; dashed arrows indicate normal-reference-dependent operations.

2. Materials and Methods

2.1. Dataset, Labels, and Train–Test Independence

The dataset contains 88 de-identified operating records from 88 distinct vehicles and battery packs. The authors confirmed from acquisition-batch records that training and test partitions contain no shared vehicle, battery-pack or acquisition-batch identifiers. The original training-side partition comprised 34 normal and four fault-associated records; the latter were excluded from model fitting, parameter selection and normal-reference calibration. Testing used 50 held-out records (43 normal; seven provider-annotated fault records containing 13 annotated faulty cells) (Table 1).

Table 1. Dataset composition and pack-layout distribution.

Data Split	Records	Fault Records	Normal Records	U95_T34	U96_T32	Purpose
Original training set	38	4	34	25	13	Dataset construction
Unsupervised training set	34	0	34	22	12	Normal-only training
Test set	50	7	43	30	20	Independent evaluation
All data	88	11	77	55	33	Total dataset

The data provider labeled the affected records as thermal-runaway-related faults in ternary lithium-ion batteries and supplied the associated faulty-cell indices. These annotations do not independently confirm that thermal runaway occurred. The provider-supplied record labels and cell list form the reference annotations for evaluation; faulty-cell labels are not derived from model predictions. The source records contain acquisition timestamps, and the observed duration of each record is defined by its last timestamp minus its first timestamp; durations are thus record-specific rather than assumed constant. The nominal sampling interval is 10 s. Although fault-category and faulty-cell labels were provided, independently verifiable thermal-runaway onset timestamps and detailed failure-mechanism labels are not available for this analysis. No alarm-to-thermal-runaway lead time or mechanism-specific classification accuracy is claimed.

2.2. Alignment, Temperature Roles and Dual-Branch Architecture

For U95_T34 and U96_T32 layouts, 95/96 voltage channels are mapped to 96 slots with validity mask m^v; 32/34 temperature channels are separately mapped to 34 slots with mask m^t. The Transformer receives voltage windows only; valid temperatures contribute to a record-level concordance descriptor, not cell-level temperature mapping. Layout offsets are estimated from observed channels of the 34 normal-reference records, grouped by structure, in Equations (1) and (2). Figure 2 illustrates the channel mapping, validity masks and layout compensation for the two pack structures.

$${\mu }_{g,j}={\mathrm{med}}_{{\mathcal{N}}_g}\left(X_j\right), {\mu }_{\ast ,j}={\mathrm{med}}_{\mathcal{N}}\left(X_j\right)$$

(1)

$${\Delta }_{g,j}={\mu }_{g,j}-{\mu }_{\ast ,j}, {\tilde{X}}_{i,t,j}=m_{i,j}^v\left[X_{i,t,j}-{\eta }_{\mathrm{DA}}{\Delta }_{g,j}\right], {\eta }_{\mathrm{DA}}=0.60$$

(2)

Figure 2. Unified channel mapping, validity-mask construction, and layout compensation for heterogeneous EV battery packs: (a) original U95_T34 and U96_T32 pack structures; (b) channel mapping, padding and layout correction; and (c) unified masked representation. Gray-filled slots indicate valid channels (mask = 1), whereas white slots indicate masked or structurally absent channels (mask = 0).

The layout compensation in Equation (2), with η_DA = 0.60, masks absent slots after the normal-reference offset correction. Padded channels are excluded from both reconstruction loss and cell ranking.

For each voltage window X ∈ ℝ^{(L × 96)}, the primary branch reconstructs O₁ and forms a masked residual E₁ = m^v⊙(O₁ − X). The auxiliary branch receives the projected concatenation [X, E₁] ∈ ℝ^{(L × 192)}. Its lower-triangular attention mask prevents direct future-position access within that layer; because E₁ comes from a full-window primary branch, this does not establish battery-variable causation or a globally causal network (Figure 3). The symbol ⊙ denotes element-wise (Hadamard) multiplication.

Figure 3. Residual-conditioned dual-branch voltage-reconstruction architecture with a temporally masked auxiliary attention operation. Panel (a) shows the primary reconstruction branch, and panel (b) shows the residual-conditioned auxiliary branch. Pale-blue boxes indicate voltage inputs and reconstruction outputs; light-gray boxes indicate masked loss operations.

For minibatch size B and valid-slot count M_i = Σ_j m^v_i,j, Equations (3) and (4) define the two masked reconstruction losses and their weighted objective.

$${\mathcal{L}}_b={\displaystyle \frac{1}{B}}{\displaystyle \sum _{i=1}^B}{\displaystyle \frac{{\displaystyle {\sum }_{t=1}^L}{\displaystyle {\sum }_{j=1}^{96}}{m}_{i,j}^v{\left(O_{b,i,t,j}-X_{i,t,j}\right)}^2}{L{\displaystyle {\sum }_{j=1}^{96}}{m}_{i,j}^v}}, b=1,2$$

(3)

$ℒ$_rec = λ$ℒ$₁ + (1 − λ)$ℒ$₂,  λ = 0.60

(4)

The objective uses λ = 0.60 (1 − λ = 0.40). Alternating updates train the primary branch with the auxiliary branch frozen, then train the auxiliary branch using detached E₁ with the primary branch frozen. The hidden size is 64 with four heads and one encoder layer.

2.3. Window Alarms and Count-Independent Record Decision

Sampling occurs nominally every 10 s; a 10-sample window spans 90 s between its first and last observations. Training and inference strides are eight (80 s) and four (40 s) samples. Three consecutive window alarms are required; in the uninterrupted case the third window ends 170 s after the first window starts. These intervals describe detection timing, not lead time before thermal runaway.

For record i and window h, a_i,h and u_i,h are the mean and standard deviation of the masked absolute voltage-reconstruction score over four MC Dropout passes. A provisional alarm uses the normal-window q99 threshold. Table 2 defines six record-level features f₁–f₆, including voltage–temperature concordance at record level. Each feature is standardized using normal-reference median/MAD statistics (Equation (5)), not cell-level [0, 1] scaling.

Table 2. Operational definitions of the six record-level fusion features.

Feature	Computed Quantity	Meaning
f1	95th percentile of a_i,h across all inference windows	validity-masked voltage residual
f2	longest run of provisional window alarms/max(1, number of windows)	temporal persistence
f3	mean_h[u_i,h/(a_i,h + ε)]	MC-pass normalized uncertainty
f4	max_j V_i,j/(sum_{j:valid} V_i,j + ε), V_i,j = mean_h absolute cell residual	voltage evidence concentration
f5	max(0, Pearson corr_h(a_i,h, mean valid |ΔT_i,h,c|)); zero if variance is zero	voltage/pack-temperature temporal concordance
f6	fraction of windows for which both branch residual scores exceed their respective normal q99 limits	cross-branch alarm agreement

With w = (0.40, 0.18, 0.12, 0.15, 0.08, 0.07), the fused record score is S_i = Σ_k w_k z_i,k. Normal-reference q99 calibrates the record threshold τ_record; Equation (6) jointly requires threshold exceedance and a three-consecutive-window gate, independently for each record. Standardized features are clipped to [−8, 8], so τ_record can exceed one. Neither the number of faulty test records nor their labels enter calibration.

$$z_{i,k}=\mathrm{clip}\left({\displaystyle \frac{f_{i,k}-{\mathrm{med}}_N\left(f_k\right)}{1.4826{\mathrm{MAD}}_N\left(f_k\right)+{10}^{-6}}},-8,8\right)$$

(5)

$$A_{i,h}=\mathbb{1}[a_{i,h}>{\tau }_{\mathit{window}}],{\tau }_{\mathit{window}}=Q_{0.99}\left(a_{\mathit{normal}}\right)$$

$$G_i=\mathbb{1}[\exists h : A_{i,h}=A_{i,h+1}=A_{i,h+2}=1]$$

$${\hat{y}}_i=\mathbb{1}[S_i>{\tau }_{\mathit{record}}]·G_i$$

(6)

Equation (6) defines the final count-independent record decision. As shown in Figure 4, normal-only reference statistics calibrate the robust feature normalization and decision threshold. The fused record severity must exceed that threshold, while a separate stable-three-window alarm gate must also hold. These are parallel requirements of the joint record decision, not sequential re-scoring stages. Only fault-positive records proceed to faulty-cell localization. No test-set ranking, fixed fault count, or top-seven selection is used.

Figure 4. Multi-source anomaly fusion and count-independent record-level fault decision. Fused-severity thresholding and the stable-alarm gate are parallel conditions of the joint decision; fault-positive records are forwarded to faulty-cell localization. Solid arrows indicate evidence and decision pathways; dashed arrows indicate normal-reference calibration; light-gray shading highlights normal-reference and normalization operations.

2.4. Cell Fusion, Settings and Evaluation

For eligible voltage cells, Equation (7) combines five support terms g₁–g₅ defined in Table 3, with weights α = (0.42, 0.21, 0.14, 0.10, 0.13) and normal-state hotspot penalty β = 0.28. Only valid cells exceeding τ_cell = 0.30 are ranked; at most K_cell = 5 are reported. Voltage-index neighbors do not necessarily represent physical cell adjacency.

Table 3. Operational definitions of five cell-level evidence groups and hotspot penalty.

Signal	Computable Operation	Interpretation
g1	u1 = mean_h r_i,h,j across alarm-associated windows; normalize by layout-specific normal q05–q95	reconstruction support
g2	g2 = fraction of alarm windows where r_i,h,j exceeds its normal q99	multiscale/multiwindow persistence
g3	g3 = mean-normalized g1 of valid index neighbors j − 1 and j + 1	index-neighborhood consistency; not physical adjacency
g4	u4 = mean_h v_i,h,j/normal valid-slot MAD (+ε); normalize by layout-specific normal q05–q95	raw-voltage departure
g5	g5 = fraction of stable-alarm windows in which j is in the preliminary top five	alarm-window cell consistency
P_hot	normal-only top-five frequency of voltage slot j, over all normal-reference windows	recurring benign hotspot penalty

The record-level temperature descriptor is excluded from cell ranking. Continuous cell features g₁ and g₄ use normal-reference 5th/95th percentile scaling within each layout; g₂ and g₅ are window proportions and g₃ is the valid-neighbor mean of normalized g₁. Hotspot frequency is the normal-reference proportion of windows where a voltage index enters the preliminary top-five set.

For the unscaled continuous evidence u_r(i,j), r ∈ {1, 4}, the normal-only layout-specific bounds L_(r,g) and U_(r,g) are the 5th and 95th percentiles. The clipped [0, 1] mapping is given below.

$$g_r\left(i,j\right)={\mathrm{clip}}_{\left[0,1\right]}\left({\displaystyle \frac{u_r\left(i,j\right)-L_{r,g}}{U_{r,g}-L_{r,g}+{10}^{-6}}}\right), r\in \{1,4\}.$$

For g₂ and g₅ an empty applicable-window set produces zero; g₃ averages existing valid adjacent-index evidence, or zero when neither neighbor exists. Invalid slots are excluded throughout. All percentile bounds and hotspot frequencies come only from the normal reference.

$$C_{i,j}={\displaystyle {\sum }_{r=1}^5}{\alpha }_r{g}_{i,j,r}-\beta P_{i,j}^{\mathrm{hot}}, {\widehat{\mathcal{C}}}_i={\mathrm{TopK}}_5\{j:m_{i,j}^v=1,\hspace{0.25em}{C}_{i,j}\ge {\tau }_{\mathrm{cell}}\}$$

(7)

The cell threshold is applied before descending ranking and the K_cell = 5 inspection-list cap; fewer than five cells may be output. Sensitivity to β, K_cell and q is reported in Section 3.3. Table 4 summarizes the evaluation parameters and their interpretation in actual sampling time.

Table 4. Evaluation parameters and actual-time interpretation.

Parameter	Summary	Interpretation
Source sampling period	10 s	Nominal sample interval
window_len/train/infer stride	10/8/4	90 s span; 80 s/40 s start offsets
stable_k/MC Dropout	3/4	3 adjacent alarms; 4 stochastic passes
Architecture	D = 96; 64 hidden; 4 heads	One-layer specification
Fusion/threshold	6 w values; q = 0.99	Normal-reference calibration, no fault count
Cell fusion	5 α values; β = 0.28; K = 5	Hotspot penalty and report cap
Layout compensation	η_DA = 0.60	Normal-only group vs pooled median correction, Equations (1) and (2)
Masked losses/optimization	λ = 0.60; alternating updates	Masked L1/L2, frozen opposing branch, residual detached, Equations (3) and (4)
Training schedule	16 epochs; batch 256; lr 0.001	Training hyperparameters reported in the original manuscript
Dropout	0.1	Network regularization reported in the original manuscript

Training used 16 epochs, a batch size of 256, a learning rate of 0.001 and a dropout of 0.1. Conditional localization evaluates each true fault record independently of the record decision; end-to-end localization emits candidates only for records flagged faulty and counts cells on false-positive records as false positives. Both definitions are applied consistently to compared methods. The computational environment used Python 3.10.19, PyTorch 2.9.1+cu128 (CUDA 12.8), NumPy 2.1.2, pandas 2.3.3, scikit-learn 1.7.1, SciPy 1.15.3, Matplotlib 3.10.6 and tqdm 4.67.3 on an NVIDIA GeForce RTX 5060 GPU (NVIDIA Corporation, Santa Clara, CA, USA).

Parameter Selection and Pre-Test Configuration Control

The authors verified from retained code configurations and run records that w, α, λ, β = 0.28, τ_cell = 0.30, K_cell = 5 and q = 0.99 were fixed before held-out test evaluation. Each model obtained its own window and record thresholds from normal-only reference scores. K_cell = 5 limits inspection workload rather than presuming a true fault count. Later β/K/q sweeps are descriptive sensitivity checks, not retrospective selection evidence; detailed control procedures are in Supplementary Tables S1 and S2.

3. Results

3.1. Record-Level and Cell-Level Evaluation Results

Table 5 and Table 6 report the main counts and record-wise predictions under the count-independent revised evaluation.

Table 5. Corrected fixed-partition record-level and cell-level outcomes.

Endpoint	TP	FP	FN	TN	Precision	Recall	F1
Record	7	1	0	42	0.8750	1.0000	0.9333
Conditional cell	12	2	1	—	0.8571	0.9231	0.8889
End-to-end cell	12	4	1	—	0.7500	0.9231	0.8276

Note: “—” indicates that a true-negative count is not applicable to cell-level localization.

Table 6. Record-level and cell-level outcomes for the revised evaluation.

Record	GT Flag	Pred.	GT Cells	Cond. Predicted	E2E Predicted
JBGS25000002622	0	1	—	—	U39, U60
JBGS25000001450	1	1	U21	U21, U71	U21, U71
JBGS25000005848	1	1	U63	U63, U85	U63, U85
JBGS25000005921	1	1	U92, U68	U92, U68	U92, U68
JBGS25000007855	1	1	U31	U31	U31
JBGS25000007863	1	1	U6	U6	U6
JBGS25000008429	1	1	U5, U74, U92, U48, U44	U5, U74, U92, U48, U44	U5, U74, U92, U48, U44
JBGS25000009894	1	1	U21, U34	U21	U21

Note: “—” indicates that no applicable annotated or predicted faulty-cell entry is available for that record and column.

On the 50-record test partition, the model detected all seven fault records with one false alarm (TP/FP/FN/TN = 7/1/0/42). Conditional cell localization yielded 12/2/1 TP/FP/FN and end-to-end localization 12/4/1; the two extra end-to-end false positives arise from the falsely flagged normal record.

3.2. Baselines and One-Factor Ablations

Table 7 compares models under the revised protocol. HAC-DBAT gives record F1 = 0.9333 and conditional cell F1 = 0.8889, against 0.8750/0.8462 for DBAT and 0.8750/0.8571 for Anomaly Transformer. Historical GRU-AE results are not included as revised-protocol runs.

Table 7. Comparison of record-level detection and faulty-cell localization across methods.

Method	Pred. +	Record F1	Conditional Cell F1	E2E Cell F1
sAE	4	0.3636	0.5714	0.4800
LSTM-AE	6	0.9231	0.7407	0.7200
Transformer-AE	9	0.8750	0.8276	0.7500
DBAT	9	0.8750	0.8462	0.7586
Anomaly Transformer	9	0.8750	0.8571	0.7742
HAC-DBAT	8	0.9333	0.8889	0.8276

For fairness, all revised baselines use the same normal training/test records, validity masking, window/stride settings, model-specific normal q99 thresholds, three-alarm gate and cell-selection rule. Each architecture retains its native anomaly score; unavailable fusion features are omitted and the remaining weights renormalized. Protocol controls and the separately retained historical GRU-AE result are documented in Supplementary Table S3.

Table 8 summarizes six single-factor ablations. In particular, removing hotspot suppression (β = 0) reduces conditional cell F1 from 0.8889 to 0.7273; removing cell-level alarm consistency lowers it to 0.8571 without changing record F1. The cell-level consistency feature is distinct from the record-level three-window decision gate.

Table 8. Single-factor ablation results for record-level detection and faulty-cell localization.

Variant	Record F1	Cell F1	E2E F1	Cell TP/FP/FN
Full	0.9333	0.8889	0.8276	12/2/1
w/o DA	0.8235	0.8000	0.7273	12/5/1
w/o UQ	0.8750	0.8462	0.7333	11/2/2
w/o EW	0.8750	0.8276	0.7273	12/4/1
w/o HS	0.9333	0.7273	0.6667	12/8/1
w/o AC	0.9333	0.8571	0.8000	12/3/1
w/o CF	0.9333	0.4889	0.4583	11/21/2

All ablations keep the same cohort, metrics and validity policy. Disabled fusion terms are omitted with remaining weights renormalized; affected record thresholds are recalibrated from normal records at q = 0.99.

The six changes are: w/o DA, η_DA = 0 while retaining masks; w/o UQ, deterministic scoring without f₃; w/o EW, removal of f₂ and g₂ while retaining the record alarm gate; w/o HS, β = 0; w/o AC, removal of g₅ while retaining the record gate and f₆; and w/o CF, cell support g₁−βP_hot instead of five-term fusion. All use the same τ_cell and K_cell.

For additional count checks, DBAT and Anomaly Transformer each yield record TP/FP/FN/TN = 7/2/0/41; their conditional cell counts are 11/2/2 and 12/3/1, respectively. Cell-level end-to-end metrics additionally penalize candidates from false-positive records.

3.3. Sensitivity and Statistical Uncertainty

Table 9 shows that the conditional cell F1 changes from 0.7273 to 0.8889 to 0.8800 for β = 0, 0.28 and 0.35, respectively. Varying K_cell and the normal-reference q quantile also changes performance; these retrospective sensitivity checks cannot establish when parameters were chosen.

Table 9. Parameter sensitivity and statistical uncertainty summaries.

Analysis	Summary	Interpretation
β: 0 → 0.28 → 0.35	Cell F1: 0.7273 → 0.8889 → 0.8800	F1 response as hotspot penalty increases
K_cell: 3 → 5 → 7	Cell F1: 0.8000 → 0.8889 → 0.8276	Sensitivity of the ranked-cell report limit
q: 0.90 → 0.99	Record F1: 0.8235 → 0.9333	Sensitivity of normal-reference threshold quantile
Eight stratified partitions (n = 8)	Record F1 = 0.9249 ± 0.0753 Conditional cell F1 = 0.8881 ± 0.0533 End-to-end cell F1 = 0.8441 ± 0.0917	Performance variation across eight stratified partitions, with wider record-level uncertainty due to seven positive records.
95% exact sensitivity CI	7/7: [0.590, 1.000]	Small positive sample count

Across eight within-cohort normal-record re-partitionings, record, conditional cell and end-to-end F1 were 0.9249 ± 0.0753, 0.8881 ± 0.0533 and 0.8441 ± 0.0917 (mean ± sample SD). The seven fault test records remained fixed; these are not eight independent fault-cohort validations. Record sensitivity 7/7 has an exact 95% binomial interval of approximately [0.590, 1.000]. The individual seeds, confusion counts and q99 thresholds are provided in Supplementary Tables S4 and S5.

3.4. Training-Data-Size Sensitivity

Training-size sensitivity used N = 9, 17, 26 and 34 normal records, with eight runs per size and the same held-out 50-record test cohort. Smaller training sets varied vehicle membership; full-size runs varied initialization and window sampling. Other reported training settings were held constant.

Table 10 summarizes mean ± sample SD for record, conditional cell and end-to-end F1 across the four training sizes. Performance generally improved with more normal records; all 32 run-level confusion counts and F1 values are retained in Supplementary Table S6. Supplementary Information contains the unchanged detailed tables relocated from the expanded revision.

Table 10. Training-data-size sensitivity (eight runs per size; mean ± sample SD).

Normal Training Records	Repeated Runs	Record F1 Mean ± SD	Conditional Cell F1 Mean ± SD	End-to-End Cell F1 Mean ± SD
9	8	0.6708 ± 0.0553	0.7177 ± 0.0293	0.5008 ± 0.0577
17	8	0.7675 ± 0.0498	0.7962 ± 0.0279	0.6272 ± 0.0536
26	8	0.8738 ± 0.0464	0.8388 ± 0.0262	0.7330 ± 0.0529
34	8	0.9354 ± 0.0473	0.8924 ± 0.0299	0.8335 ± 0.0649

4. Discussion

The revised protocol produces record F1 = 0.9333 and conditional cell F1 = 0.8889; the baseline and ablation comparisons suggest a contribution from layout compensation, residual reconstruction and hotspot-aware cell fusion in this cohort. Results from related sensor-diagnosis fields [29,31,33] are contextual rather than EV-pack benchmarks.

Unlike the earlier evaluation, which contained an expected_fault_files = 7 count-prior setting, the revised procedure makes independent normal-threshold and stable-alarm decisions for every record. The old count-prior setting was removed before the revised evaluation; the two procedures are therefore not presented as a matched comparison.

Limitations include only seven annotated fault records and reuse of these fault records across the eight normal-data re-partitionings. Although the provider labeled the records as thermal-runaway-related faults in ternary lithium-ion batteries and supplied faulty-cell labels, the available evidence does not independently establish whether thermal runaway occurred, its onset time, or the detailed electrochemical failure mechanism; hence, a pre-event warning lead time cannot be evaluated. Individual record durations are defined by source-file timestamp spans but are not numerically tabulated here. Pre-test parameter settings were checked by the authors against retained run artifacts; independent reproduction would require access to these artifacts. Larger independent fault cohorts and external-source validation are still needed.

5. Conclusions

HAC-DBAT provides a validity-masked, normal-data-only reconstruction and cell-ranking framework for heterogeneous EV battery-pack monitoring. On the held-out partition, record F1 = 0.9333, conditional cell F1 = 0.8889 and end-to-end cell F1 = 0.8276. The baseline comparisons, one-factor ablations and sensitivity analyses characterize performance at the record and faulty-cell levels. Because the test set includes only seven positive records, broader generalization and verified lead time to thermal-runaway onset remain outside the conclusions of this study.