Integrated Intelligence for Lithium-Ion Battery Management and Spatial Fault Detection

Lithium-ion (Li-ion) batteries underpin much of modern portable and stationary energy storage, including smartphones, medical devices, renewable-energy buffering, and electric vehicles (EVs). Their commercial success reflects a favourable combination of energy density (150–300 Wh/kg), cycle life, low self-discharge, and manufacturing cost, driven by continued progress in electrode chemistry and cell design [1, 2]. Global Li-ion production capacity exceeded several hundred gigawatt-hours annually by 2024 and continues to grow with EV demand.

Safe, long-lived operation of Li-ion systems depends on two battery management system (BMS) capabilities: (i) accurate estimation of State of Charge (SoC), the fraction of remaining usable capacity, since SoC cannot be measured directly and must be inferred from terminal voltage, current, and temperature; and (ii) timely detection and spatial localization of internal faults—short circuits, dendrite growth, and thermal-runaway precursors—before they escalate into hazardous events.

SoC estimation is difficult because the voltage-SoC relationship is nonlinear, hysteretic, temperature-dependent, and drifts with aging [3, 4]. Coulomb counting accumulates sensor and self-discharge error over time [5]; open-circuit-voltage (OCV) methods require prolonged rest that is impractical in service [6]; and equivalent circuit model (ECM) approaches combined with observers such as the Extended Kalman Filter (EKF) are accurate but require careful parameter identification and re-tuning as the cell ages [7, 8].

Fault detection is difficult for a different reason: faults such as lithium plating, separator damage, or localized overheating are inherently spatial, but conventional voltage- and impedance-based monitoring collapses the whole cell to a handful of terminal measurements and therefore cannot resolve where a fault is developing until it has already propagated [9, 10].

This paper reports two neural-network-based modules addressing these two problems on a shared experimental cell population: a feed-forward ANN for SoC estimation, and a physics-regularized graph attention network (GAT) for spatial fault classification and localization on a 72-node material-aware battery graph. The two modules are evaluated independently; we do not claim a single jointly-trained model, and we make explicit where the evidence for each claim comes from. The term "integrated intelligence" in the title refers to this shared sensing pipeline and the fact that both modules target the same BMS use case, not to a system that has been demonstrated running as a unified, embedded controller in a vehicle or grid installation; no such deployment or integration test is reported in this paper.

The specific contributions of this work are:

(1) A feedforward ANN SoC estimator trained and evaluated across six temperatures, three C-rates, and three aging states on a cell-wise held-out split of 3.7 million samples, achieving 1.43% Mean Absolute Error (MAE) with sub-5 ms embedded inference.

(2) A 72-node, material-aware battery graph representation (anode, cathode, separator, current-collector regions) with physics-derived edge weights (ionic, electronic, thermal transport coefficients) for spatially-resolved fault representation.

(3) A GAT trained with a composite loss combining cross-entropy classification, a spatial-consistency term, and a thermodynamic-consistency regularization term, achieving 94.3% classification accuracy and ±2.1 mm mean Euclidean localization error, cross-validated against post-mortem computed tomography (CT), scanning electron microscopy (SEM), and four-point-probe measurements on held-out cells (Section 3).

(4) An early-detection result showing the GAT flags all three fault types 15–20 cycles before a conventional ΔV > 50 mV voltage-threshold baseline (Section 4.4).

(5) An honest accounting of what is and is not demonstrated: SoC-SoH (state of health) co-estimation, remaining useful life (RUL) prediction, few-shot/meta-learning, and federated learning are discussed only as future work (Section 4) and are not claimed as results of this study.

The remainder of the paper is organized as follows. Section 2 reviews related work in three thematic areas and states the specific gaps this work targets. Section 3 describes the methodology, including the experimental cells, data acquisition, and hardware, the problem formulation, the battery graph model, the ANN and GAT architectures, term definitions, pseudocode, and the integrated system architecture. Section 4 reports and discusses the results, including a qualitative comparative analysis against prior work, limitations, and future work. Section 5 concludes the paper.

3.1 Data and experimental setup

3.1.1 Experimental cells and data acquisition

All experiments used commercial cylindrical 18,650-format NMC811/Graphite cells (nominal capacity 3.0 Ah, nominal voltage 3.6 V) sourced from a single production batch to minimise inter-cell variability. Data were assembled from three complementary sources: (i) controlled laboratory cycling, (ii) drive-cycle emulation, and (iii) accelerated aging and fault induction.

The ANN dataset (approximately 3.7 million time-step samples) was partitioned cell-wise—not sample-wise—into 70% train/15% validation/15% held-out test, so that the model is evaluated on cells never seen during training. Each split proportionally represents all six temperatures and three C-rates. Input features [V(t), I(t), T(t), Q_cum(t), τ(t)] were standardised to zero mean/unit variance using training-split statistics only; Q_cum was reset to zero at the start of each charge or discharge half-cycle. The ANN dataset draws on 60 physical cells spanning the full six-temperature × three C-rate × three-aging-state design (a mean of slightly over one cell per condition, with additional replicate cells at the temperature/C-rate extremes), split cell-wise into 42 train/9 validation/9 test (70/15/15). The three aging states are defined by capacity retention at the start of each cycling block, measured by a full-rate reference discharge: fresh (≥97% of rated capacity), mid-life (90–97%), and end-of-life (80–90%), consistent with the states. Reference SoC(t) was obtained by Coulomb counting anchored at each aging checkpoint by a low-rate (C/20) OCV rest, not by an independent ground-truth sensor; this reference is therefore subject to the same integration drift as any Coulomb counting method, bounded between successive OCV recalibrations. Sensor noise was not filtered or denoised before use: raw voltage, current, and temperature readings (accuracy specifications) were fed to the ANN as recorded, so the reported MAE/Root Mean Square Error (RMSE) reflects performance under native instrument noise rather than a smoothed signal. In the drive-cycle emulation segments, a half-cycle boundary for the Q_cum reset is defined by the emulation protocol's segment markers rather than by a current-sign reversal; the model's behaviour when the initial SoC at such a reset is itself uncertain—a genuine field cold start with no prior calibration—has not been tested in this study and is stated as a limitation in Section 4.4. Unlike the GNN dataset below, the ANN inputs do not include electrochemical impedance spectroscopy (EIS)-derived features, so the two modules carry different sensing requirements, a distinction discussed further in Section 4.4.

The GNN dataset comprised 48 fault-labelled cells, partitioned 70/15/15 cell-wise across the four fault classes. Each cell contributed 72 graph snapshots (one per EIS measurement cycle), yielding 3,456 labelled graphs. Node features were normalised per channel using training-set statistics, and class imbalance was addressed with weighted random sampling (weights inversely proportional to class frequency). Table 3 discusses the hardware details used in this paper. Ground-truth fault location was established by a single trained operator per cell, who registered the CT/SEM/four-point-probe coordinate frame to the graph's axial-radial coordinate frame using physical fiducial marks on the can and tab as common reference points. The spatial error reported as ±2.1 mm is the mean Euclidean distance between predicted and ground-truth locations in this unrolled axial-radial coordinate space, not a purely radial or purely axial distance. We selected the per-class attention thresholds in Section 4.1.2 on the validation split by sweeping the threshold and choosing the value that maximised node-level F1 for that class, rather than fixing them arbitrarily. With 48 fault-labelled cells and 3,456 labelled graphs, each cell contributes many correlated snapshots; per-cell performance variability, and whether a subset of cells drives the reported accuracy, have not been separately reported and are stated as a limitation in Section 4.4, along with a recommendation for future leave-one-cell-out evaluation.

Table 3. Hardware and equipment

Note: Electrochemical impedance spectroscopy (EIS), artificial neural network (ANN), graph attention network (GAT).

3.2 Problem formulation and methods

3.2.1 State of Charge estimation problem

SoC at time t is the ratio of remaining charge capacity Q(t) to nominal capacity Q_nom (Eq. (1)).

$\operatorname{SoC}(t)=Q(t) / Q_{\text {nom }} \times 100 \%$   (1)

where, Q(t) is obtained by integrating the net current flow. For the ANN target, the reference SoC was generated by Coulomb counting and periodically anchored at the aging checkpoints using a low-rate OCV. The reference SoC was not obtained from an independent SoC sensor. It remains subject to Coulomb-counting integration drift between successive OCV recalibrations. The resulting reference SoC was used as the target for ANN training and evaluation. SoC is bounded in [0%, 100%] by definition, though the practical operating range is constrained to [10%, 95%] to preserve cycle life. The estimation problem seeks the mapping given in Eq. (2).

$f:\left[V(t), I(t), T(t), Q_{\text {cum }}(t), \tau(t)\right] \rightarrow \operatorname{SoC}(t)$   (2)

This mapping is difficult to learn because of (i) strong nonlinearity and hysteresis of V(t) with respect to SoC(t); (ii) temperature dependence of the OCV curve and internal impedance; (iii) drift with aging and cycling history; and (iv) sensor noise. Training minimizes the loss in Eq. (3) across all operating conditions in the training split.

$L_{\mathrm{MAE}}=(1 / N) \Sigma_i\left|\operatorname{SoC}_{i, \text {pred}}-\operatorname{SoC}_{i, \text {ref}}\right|$   (3)

3.2.2 Equivalent circuit model—Background only

To motivate the SoC estimation problem and to clarify what is and is not integrated into the ANN, Figure 1 presents the ECM adopted in this study. The terminal voltage predicted by this model is given by Eq. (4).

$V_{\mathrm{T}}(t)=\operatorname{OCV}(\operatorname{SoC}, T)-I(t) \cdot R_0-V_1(t)-V_2(t)$   (4)

$\begin{aligned} & \mathrm{d} V_1 / \mathrm{d} t=-V_1 /\left(R_1 C_1\right)+I(t) / C_1, \\ & \mathrm{~d} V_2 / \mathrm{d} t=-V_2 /\left(R_2 C_2\right)+I(t) / C_2\end{aligned}$   (5)

$\operatorname{SoC}(t)=\operatorname{SoC}\left(t_0\right)-\left(\eta / Q_{\text {nom}}\right) \int I(t) \mathrm{d} t$   (6)

image010.png

Figure 1. Two resistor–capacitor (RC) branch Thevenin equivalent circuit model (ECM) for lithium-ion (Li-ion) battery State of Charge (SoC) estimation

3.2.3 Spatial fault detection problem

Fault detection is formulated as spatially resolved node classification on a battery graph G = (V, E), where each node v_i represents a spatial segment of the cell's internal structure and each edge (v_i, v_j) represents a physical interaction (ionic, electronic, or thermal) between adjacent segments. Each node carries a 12-dimensional feature vector x_i ∈ ℝ¹² encoding: normalized terminal-voltage contribution, normalized current density, local surface temperature, EIS-derived charge-transfer resistance, R_ct, solid electrolyte interphase (SEI) layer resistance R_SEI, constant phase element (CPE) exponent, material porosity, particle-size distribution, Brunauer-Emmett-Teller (BET) surface area, state of lithiation, cycle count, and capacity retention. The task assigns each node a label y_i ∈ {0: Healthy, 1: Internal Short Circuit, 2: Dendrite Formation, 3: Thermal Runaway Precursor}. The composite training loss is given by Eq. (7).

$L=L_{\mathrm{CE}}+\lambda_1 \cdot L_{\text {spatial}}+\lambda_2 \cdot L_{\text {physics}}$   (7)

where, L_CE is the node-level cross-entropy classification loss, L_spatial is the spatial-consistency term, L_physics is the thermodynamic-consistency regularization term (Table 4), and λ₁ and λ₂ are non-negative weighting coefficients that set the relative contributions of L_spatial and L_physics to the total loss, respectively.

The node features comprise both dynamic sensing variables and cell/material descriptors. Voltage contribution, current density, surface temperature, EIS-derived R_ct, R_SEI, and CPE exponent represent measurement-dependent features, whereas porosity, particle-size distribution, and BET surface area are material-characterization descriptors obtained during offline cell/electrode characterization. These offline descriptors are treated as cell-specific metadata and remain fixed for a given cell during GAT inference; they are not assumed to be measured continuously by the embedded BMS. State of lithiation, cycle count, and capacity retention are updated according to the operating history and characterization protocol. Thus, the term "material-aware" refers to the incorporation of both dynamic electrochemical measurements and pre-characterized material descriptors into the graph representation.

3.2.4 Battery graph model

The cell is represented as a heterogeneous graph with 72 nodes: 24 anode nodes, 24 cathode nodes, 12 separator nodes, and 12 current-collector nodes, each carrying the 12-dimensional feature vector described above. The mapping from the physical 18,650 jelly-roll to this node set is obtained by conceptually unwinding the jelly-roll into a flat anode/separator/cathode/current-collector strip along its winding length (≈250 mm) and cell height (≈65 mm). This unrolled strip is then divided into 6 axial bands along the height and 4 winding-depth bands along the length for the anode and cathode layers (6 × 4 = 24 nodes each, with each node spanning roughly 11 mm axially by 62 mm along the winding direction). The thinner separator and current-collector layers are divided into 6 axial × 2 winding-depth bands (12 nodes each, with each node spanning roughly 11 mm axially by 125 mm along the winding direction). This approach represents a simplified, uniform voxelization rather than an exact anatomical reconstruction of the jelly-roll geometry, and the approximation this introduces is noted as a limitation in Section 4.4.

image012.png

Figure 2. Spatially structured battery graph G = (V, E)

Figure 2 discusses the spatially structured battery, where Node types are: Anode (blue), Separator (green), Cathode (red), Current Collector (amber); edges encode ionic (dashed), electronic (solid), and thermal (dotted) transport pathways.

Edges carry physics-informed weights derived from the relevant transport coefficient between adjacent regions:

$w_{i j}=\sigma_{i j} / \max _k\left(\sigma_k\right)$   (8)

where, σ_ij ∈ {σ_ionic = 0.6 S/m, σ_elect, λ_therm = 1.0 W/(m·K)}.

The 72-node representation is therefore a coarse computational discretization of the unrolled cell rather than a one-to-one representation of millimeter-scale physical voxels. The GAT predicts the fault-associated node using class-specific attention weights, while the localization error is calculated in the registered unrolled axial-radial coordinate system against the independently identified post-mortem fault location. The reported localization error should therefore be interpreted as a coordinate-space localization error associated with the predicted fault region, rather than as the physical size or resolution of an individual graph node.

3.2.5 Artificial neural network and graph attention network architectures

The ANN consisted of 5 input neurons, 2 hidden layers with 20 and 10 rectified linear unit (ReLU) neurons, respectively, and one linear output neuron. It was trained to minimize Eq. (3) using the Adam optimizer (α = 0.001, β₁ = 0.9, β₂ = 0.999), with dropout (rate = 0.2) and early stopping (patience = 10 epochs). The ANN architecture is shown in Figure 3.

Weighted message passing is computed as shown in Eq. (9):

$h_i^{(n+1)}=\sigma\left(\Sigma_{j \in N(i)} \alpha_{i j} W h_j^{(n)}\right)$   (9)

$\alpha_{i j}=\frac{\exp \left(\operatorname{LeakyReLU}\left(a^{\top}\left[W h_i \| W h_j\right]\right)\right)}{\Sigma_{k \in N(i)} \exp \left(\operatorname{LeakyReLU}\left(a^{\top}\left[W h_i \| W h_k\right]\right)\right)}$   (10)

where, α_ij is the learned attention coefficient computed in Eq. (10), h_i⁽ⁿ⁾ the embedding of node i at layer n, W the trainable weight matrix, a the trainable attention vector, ‖ the concatenation operator, ⊤ the transpose, N(i) the neighborhood of node i, and σ the LeakyReLU activation. The GAT is trained on the composite loss of Eq. (7) using Adam (α = 0.005, weight decay 10⁻⁴), batch size 32 graphs, up to 100 epochs with early stopping (patience 15 epochs).

The ANN converged at epoch 87 (best checkpoint at epoch 77), requiring approximately 4.2 hours on the hardware described in Section 3.1.1 (Table 3), with a fixed random seed of 42. For the GAT, the graphs contained a mean of 198 ± 12 edges per graph, and early stopping was triggered at epoch 72, with final training and validation losses of 0.092 and 0.118, respectively. The definitions of the key terms are given in Table 4.

image017.png

Table 4. Definitions of key terms

3.2.6 Pseudocode: Preprocessing, training, and inference

The following pseudocode summarizes the ANN and GAT pipelines end-to-end.

3.3 Integrated system architecture

Figure 4 shows the shared preprocessing pipeline: raw voltage, current, temperature, and EIS measurements are collected from battery sensors and passed through preprocessing to extract features, which feed in parallel into the ANN (SoC) and GAT (fault detection) modules. The spatial fault map optionally feeds into post-mortem physical analysis via SEM/CT characterisation.

image021.png

Figure 4. Shared sensing and preprocessing architecture

This flowchart outlines a battery monitoring system architecture. Sensor data (voltage, current, temperature, EIS) undergoes preprocessing, then splits into two parallel branches: SoC estimation via MLP and fault detection via GAT. Both feed into a Decision Fusion & Alert Layer, producing three outputs: a spatial fault map, SoC estimate, and protective actions. The fault map optionally triggers offline Post-Mortem Analysis (SEM/XRD) for deeper material-level investigation, shown as a dashed, secondary pathway. This layered pipeline combines real-time ML inference with physics-based diagnostics, enabling both immediate operational responses (like protective actions) and deeper offline investigation, bridging predictive monitoring with root-cause failure analysis for battery health management systems.

The two modules share only the upstream sensing/preprocessing stage; they are trained and evaluated independently on their respective datasets (Section  4), and this paper does not present a jointly optimised or jointly validated combined model.

This paper reported two independently evaluated modules for Li-ion battery management: a feedforward ANN for SoC estimation and a physics-regularized GAT for spatial fault classification and localization. On held-out, cell-wise test data, the ANN achieved 1.43% MAE with sub-5 ms embedded inference, and the GAT achieved 94.3% fault-classification accuracy with ±2.1 mm mean Euclidean localization error, detecting faults 15–20 cycles earlier than a voltage-threshold baseline; localization was cross-validated against post-mortem CT, SEM, and four-point-probe measurements. These results are specific to the NMC811/Graphite 18,650 cell population, temperature range, and fault-induction protocol described in Section IV, and the two modules were trained and validated independently rather than as a single jointly optimised system. Claims regarding SoC-SoH co-estimation, RUL prediction, federated learning, and the isolated contribution of the physics-informed loss term are stated as future work rather than as results of this study.