In the context of the carbon-neutrality goal, electrochemical energy storage systems centered around lithium-ion batteries are being deployed at large scale in electric power systems, electric vehicles, and other fields. However, during manufacturing, transportation, and long-term cycling, energy storage cells are prone to various defects—such as delamination, gas bubbles, cracks, and lithium plating—which are key triggers for thermal runaway, capacity fade, and safety hazards. Reliable state detection is the core support for safe and efficient large-scale energy storage applications.
Traditional battery condition monitoring methods are limited to identifying macroscopic faults such as thermal runaway and lack effective detection capabilities for subtle features of hidden faults. Among nondestructive testing (NDT) techniques, ultrasonic scanning stands out for its high penetration, high resolution, rapid, and low-cost advantages. By analyzing the changes in acoustic characteristics of ultrasonic waves propagating inside an energy storage cell, we can image internal structures and defects effectively. However, conventional ultrasonic inspection relies heavily on manual interpretation of images, which is subjective and inefficient.
Recent advances in deep learning and machine learning have shown great promise in battery image recognition. For example, convolutional neural networks and support vector machines (SVM) have been applied to detect defects and estimate battery state of health. Nevertheless, a closed-loop chain from NDT to defect diagnosis to state evaluation is still lacking.
In this work, we propose a method that integrates ultrasonic longitudinal wave reflection–transmission fusion NDT with artificial intelligence algorithms for defect diagnosis and state evaluation of energy storage cells. The key contributions are:
- Construction of an ultrasonic image database covering the entire lifecycle of energy storage cells, including various defect types and aging conditions.
- Introduction of the YOLOv12 algorithm for intelligent identification and classification of internal and external defects, optimized for multi-scale detection.
- Development of a multi-dimensional evaluation index system incorporating infiltration, gas accumulation, surface quality, internal structure, and lithium plating regions. Combined with TOPSIS-based weighting and an SVM model, the system achieves quantitative state classification into “normal”, “mildly damaged”, and “severely damaged”.
This forms a closed-loop technical framework of “ultrasonic NDT → intelligent defect diagnosis → comprehensive state evaluation”, providing an integrated solution for health management and intelligent maintenance of energy storage systems.
Ultrasonic Image Database Construction
Ultrasonic Testing Principle and Platform
Ultrasonic NDT works by emitting high-frequency ultrasonic pulses and analyzing reflected or transmitted signals to identify structural conditions inside materials. For energy storage cells, longitudinal wave reflection signals are used to scan thickness-direction defects (e.g., delamination, dents, cracks), while longitudinal wave transmission signals detect overall consistency and internal structure anomalies (e.g., electrolyte wetting state, lithium plating, gas generation). The fusion of reflection and transmission modes enables comprehensive NDT of a cell in a short time.
The experimental platform is the Feli-300 ultrasonic inspection system. Its detection range is 300 mm × 250 mm, maximum cell thickness ≤ 30 mm, imaging resolution 0.1–1 mm, and detection resolution 0.5 mm. It can achieve high-precision imaging of both pouch cells and prismatic cells.
Model Construction and Data Collection
To ensure representativeness and diversity of the ultrasonic images, we selected both pouch cells and prismatic cells as samples. We designed a series of accelerated aging and defect simulation experiments to cover the entire lifecycle from incoming inspection to operation, maintenance, and retirement. The experiments included:
- Normal charging/discharging at 0.5C in silicone oil at room temperature.
- High-temperature cycling aging at 50°C and 1C.
- Low-temperature cycling aging at –5°C and 1C.
- Micro-short circuit introduced by opening the valve and adding iron powder, then cycling at 25°C and 1C.
- Lithium plating induced by opening the valve, adding iron powder, and cycling at –2.5°C.
- Mechanical dents from drop, impact, and squeeze.
- Scratches from friction and abrasion.
All cells were scanned using the Feli-300 with standardized procedures, with parameters (frequency, driving voltage, scan speed, step size) adjusted dynamically according to cell geometry. The resulting ultrasonic images were augmented by flipping, rotation, and noise injection to build a structured database of 2400 images.
The following table summarizes the typical defect/state types, their causes, and ultrasonic manifestations.
| Defect/State Type | Cause/Manifestation |
|---|---|
| Gas bubbles | Internal side reactions or overcharging cause shell expansion or local bulging. |
| Dents, scratches | Physical damage during manufacturing, transportation, or installation. |
| Lithium plating | Lithium dendrites form on electrode surfaces or separators, often with swelling or local hardening. |
| Charge/discharge abnormality | Burn marks or discoloration from local overheating or electrochemical imbalance. |
| Micro-short circuit | Positive and negative electrode contact causing short, leading to heat, swelling, and drastic performance degradation. |
| High-temperature cycling damage | Electrolyte decomposition reduces wettability; gas products form bubbles. |
| Low-temperature cycling damage | Increased electrolyte viscosity reduces wettability and charge/discharge efficiency. |
Defect Recognition and Classification Using YOLOv12
To achieve intelligent, high-precision identification of defects in ultrasonic images of energy storage cells, we adopt the latest YOLOv12 algorithm. Compared with earlier YOLO versions, YOLOv12 introduces innovations in network structure, feature fusion, and loss function, significantly improving detection accuracy and speed for multi-scale and complex-shaped defects.
Image Preprocessing and Annotation
All ultrasonic images are first converted to the YOLO format with unified resolution. To preserve aspect ratio, we apply gray padding. Gaussian filtering is used to suppress random noise. Then, using the Labelme tool, we annotate each defect with a bounding box and a category label (e.g., “bubble”, “scratch”, “dent”, “lithium plating”). The annotated dataset is organized into a standard structure.
YOLOv12 Architecture
The YOLOv12 architecture consists of three parts: Backbone, Neck, and Head. It utilizes region attention mechanisms, residual efficient layer aggregation networks (R-ELAN), and Flash Attention to reduce computational cost while maintaining a large receptive field. This enables efficient and robust detection of multi-scale defects in energy storage cells.
Defect Diagnosis Model
We build the defect diagnosis model using YOLOv12-Nano, which has only 2.56 M parameters and 6.3 GFLOPs, with an inference latency of 1.64 ms on an NVIDIA T4 GPU (FP16). The model structure includes 32 convolutional layers, 16 attention layers, 5 concatenation layers, and 3 detection heads. The dataset is split into 70% training, 20% validation, and 10% testing.
Training is performed on an NVIDIA GeForce RTX 4090 GPU. We load pre-trained weights and fine-tune with transfer learning. Training hyperparameters: 200 epochs, input size 640×640, batch size 4. A two-stage augmentation strategy is used: first, Albumentations for geometric, color, and blur transformations; second, Pillow for pixel-level perturbations (brightness, noise injection). The best weights are saved as best_model.pt.
A Tkinter-based desktop application allows users to load the trained model and perform inference on new images. The interface shows detection results with bounding boxes and confidence scores. Users can also re-run detection or save results, and even add new annotated samples to the training set for online learning.
The performance is measured by mean average precision (mAP). The formula is:
$$ \text{mAP} = \frac{1}{M} \sum_{i=1}^{M} \text{AP}_i $$
where AP is the area under the precision-recall curve for class i, and M is the total number of defect classes.
State Evaluation Using SVM
Feature Extraction from Ultrasonic Images
We extract key acoustic features from the ultrasonic signals that reflect the internal state of the energy storage cell. The features and their extraction methods are summarized in the table below.
| Core Feature | Definition | Extraction Method and Quantification |
|---|---|---|
| Wetting degree I | Degree of electrolyte wetting of electrode plates; dry areas contain air. | Let v₀ be the sound velocity in fully wetted region, vᵢ in the region of interest. Then I = vᵢ / v₀ (0 < I < 1). v₀ is calibrated from a standard fully wetted cell. |
| Gas accumulation area | Gas has extremely low acoustic impedance, causing near-total reflection or severe attenuation. | Set adaptive threshold T = 0.3 × S_normal, where S_normal is the mean signal intensity of normal cells. Signal intensity S < T indicates gas accumulation. |
| Interface and material continuity | Any defect in surface or internal structure produces extra echo signals. | Let N₀ be the number of main interface echoes in a defect-free cell, Nt the total echoes in the test cell. Then defect echoes Nd = Nt – N₀. Larger Nd means more severe defects. |
| Lithium metal deposition area L | Uneven deposition of lithium on the negative electrode, typical of fast charging and low-temperature charging. | L = 1 – vᵢ / v₀, where v₀ is the sound velocity in fully wetted region and vᵢ in the region of interest. |
Evaluation Index Quantification
We design a multi-index evaluation system covering wetting, gas accumulation, surface quality, internal structure, and lithium plating region. The weights of these indices are determined by the Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS) combined with the entropy weight method. The procedure is:
- Positive transformation: Convert low-optimal and neutral indicators into high-optimal ones using:
- Normalization: Construct normalized matrix Z.
- Determine ideal solutions: Positive ideal solution Z⁺ from the maximum per column, negative ideal solution Z⁻ from the minimum per column.
- Calculate distances:
- Calculate closeness:
- Indicator weight fusion via entropy method:
$$ x_{ij}^{\prime} = \begin{cases} x_{ij}, & \text{if high-optimal} \\ 1/x_{ij}, & \text{if low-optimal} \\ \frac{M}{M + |x_{ij} – M|}, & \text{if neutral} \end{cases} $$
where M is the expected value for the neutral indicator.
$$ D_i^{+} = \sqrt{\sum_{j=1}^{n} (z_{ij} – z_j^+)^2}, \quad D_i^{-} = \sqrt{\sum_{j=1}^{n} (z_{ij} – z_j^-)^2} $$
$$ C_i = \frac{D_i^{-}}{D_i^{+} + D_i^{-}} $$
For the closeness matrix of n objects × 5 indicators, compute:
$$ P_{ij} = \frac{C_{ij}}{\sum_{i=1}^{n} C_{ij}} $$
$$ E_j = -\frac{1}{\ln n} \sum_{i=1}^{n} P_{ij} \ln P_{ij} $$
$$ G_j = 1 – E_j $$
$$ w_j = \frac{G_j}{\sum_{j=1}^{5} G_j} $$
The final weight vector is W = (w₁, w₂, w₃, w₄, w₅) satisfying ∑w_j = 1.
SVM-Based Evaluation Model
We build an SVM classifier with a radial basis function (RBF) kernel. The input features are the extracted four core features (wetting degree I, gas area ratio, defect echo count, lithium plating degree L) from each ultrasonic image. In addition, we incorporate HOG descriptors automatically optimized over a parameter search space (gradient orientations {6,9,12,15}, cell size {(4,4),(6,6),(8,8),(10,10)}, block size {(2,2),(3,3),(4,4)}). The best HOG parameters are orientations=6, cell size=(10,10), block size=(4,4). For the SVM, we perform grid search with 5-fold cross-validation, obtaining optimal parameters C=10 and gamma=0.01.
The ground truth labels for training are “normal”, “mildly damaged”, and “severely damaged”. These labels are determined by the threshold criteria in the table below, which were established based on statistical analysis of experimental data.
| Health State | Wetting Degree I | Gas Area (S<t) | Defect Echoes Nd | Lithium Plating L | |
|---|---|---|---|---|---|
| Normal | I ≥ 0.9 | < 1% | ≤ 2 | ≤ 0.03 | |
| Mildly Damaged | 0.75 ≤ I < 0.9 | 1% to < 5% | 2 – 5 | 0.03 – 0.08 | |
| Severely Damaged | I < 0.75 | ≥ 5% | > 5 | > 0.08 |
The SVM model classifies each energy storage cell into one of these three states. The accuracy of the model is evaluated using:
$$ \text{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN} $$
where TP, TN, FP, FN are true positive, true negative, false positive, false negative, respectively.
Experimental Validation and Results
The overall process of the proposed method is illustrated in the workflow diagram (not shown here due to image reference). We performed ultrasonic scanning on pouch cells and prismatic cells (Figure 5 in the original work, but we do not include the figure number) to build the multi-stage battery state models and the image database. Then, YOLOv12 and SVM were applied sequentially.
Defect Diagnosis Results
Using the mAP@0.5 metric, the YOLOv12 model achieved an average precision above 90% across all defect categories. For clearly visible defects (bubbles, dents, scratches), the confidence scores were above 85%; for subtle defects (small gas pockets, minor lithium plating), the confidence was generally above 70%. The visual interface (analogous to Figure 7) shows real-time detection results.
State Evaluation Results
The SVM evaluation model achieved an accuracy exceeding 95% on the test set. The confusion matrix (not shown) indicated very few misclassifications between “normal” and “mildly damaged”, and zero misclassifications between “normal” and “severely damaged”. The visual interface (analogous to Figure 8) displays the detected health state for each energy storage cell, along with the feature values. For example, a cell with I=0.85, gas area=2%, Nd=3, L=0.05 is correctly classified as “mildly damaged”.
These results demonstrate that the proposed method can reliably diagnose defects and evaluate the state of energy storage cells using ultrasonic NDT combined with AI.
Conclusion and Outlook
We have presented a method that integrates ultrasonic longitudinal wave reflection–transmission NDT with artificial intelligence algorithms for defect diagnosis and state evaluation of energy storage cells. By constructing a comprehensive ultrasonic image database, applying the YOLOv12 algorithm for defect identification, and developing an SVM-based multi-index evaluation model, we achieved:
- Accurate detection and classification of bubbles, dents, scratches, lithium plating, and other defects with mAP > 90%.
- Reliable state evaluation into “normal”, “mildly damaged”, and “severely damaged” with accuracy > 95%.
This closed-loop framework provides a solid foundation for intelligent health management of energy storage systems, enabling early warning of potential failures and optimized maintenance strategies. Future work will focus on:
- Extending the database with more cell chemistries and form factors.
- Deploying the model on edge devices for real-time in-field inspection.
- Integrating transfer learning for few-shot defect recognition.
- Exploring multi-modal fusion (ultrasonic + X-ray + thermography) for even more comprehensive evaluation.
The methodology demonstrated here has the potential to significantly enhance the safety, reliability, and longevity of energy storage systems, ultimately supporting the large-scale adoption of battery energy storage in the power grid and electric vehicle sectors.