Abstract
Thermal runaway in lithium iron phosphate (LFP) energy storage batteries poses a critical safety challenge for large-scale electrochemical energy storage systems. This study proposes an in-situ ultrasonic transmission method for monitoring the internal state evolution of prismatic aluminum-cased LFP batteries during thermal runaway. By integrating ultrasonic signal intensity (Urms) and time of flight (TOF) with multi-parameter measurements including temperature, voltage, current, and surface deformation, we constructed a comprehensive internal-external parameter testing system for energy storage batteries. Our experimental results demonstrate that Urms is highly sensitive to temperature and state-of-charge (SOC) variations, while TOF effectively captures structural changes. Overcharging significantly enhances ultrasonic attenuation, with the signal being submerged by noise substantially earlier than the venting valve activation. This finding establishes ultrasonic signals as a reliable early warning indicator for overcharge thermal runaway in energy storage batteries. Building on this insight, we developed an early warning algorithm based on an improved Mahalanobis-Taguchi system (MTS) that predicts venting valve opening up to 30 minutes in advance. This work advances the understanding of spatiotemporal evolution during thermal runaway and provides a technological foundation for enhancing safety in energy storage battery systems.
1 Introduction
The rapid integration of renewable energy sources such as wind and solar power into the electrical grid has intensified the demand for efficient and reliable energy storage solutions. Among various storage technologies, lithium-ion batteries—particularly lithium iron phosphate (LFP) chemistry—have become dominant due to their high conversion efficiency, long cycle life, and cost advantages. LFP energy storage batteries are now widely deployed in grid-scale applications, offering improved thermal stability compared to other lithium-ion chemistries. However, issues such as cell inconsistency, performance degradation over prolonged use, and potential protection system failures can lead to overcharging and over-discharging, ultimately triggering thermal runaway. The safety of energy storage batteries has therefore become a pressing concern for both researchers and industry practitioners.
Current monitoring approaches for energy storage battery systems primarily rely on external parameters such as module-level voltage, current, and temperature, as well as gas and smoke detection at the cluster or prefabricated cabin level. While these methods have achieved some success in early warning, the inherent delay between internal anomalies and external manifestations limits their effectiveness. For instance, recent studies have shown that large-capacity LFP energy storage batteries exhibit a stable time interval between venting valve activation and full thermal runaway. Early warning schemes based on voltage and temperature can provide 660 seconds of advance notice, and gas sensors can detect venting events 940 seconds before short-circuit conditions. Nevertheless, direct internal monitoring would significantly improve detection timeliness.
Ultrasonic testing (UT) is a well-established non-destructive technique offering high sensitivity, rapid response, and operational flexibility. In recent years, UT has been increasingly applied to lithium-ion battery diagnostics. Previous work on pouch cells has demonstrated that ultrasonic amplitude correlates positively with SOC and that overcharge-induced gas generation causes dramatic signal attenuation. However, most studies have focused on thin pouch cells (under 10 mm thickness) where ultrasound penetration is easier. For large-format prismatic aluminum-cased LFP energy storage batteries commonly used in engineering applications (thickness >20 mm), the aluminum casing strongly reflects ultrasonic waves, making internal state monitoring challenging. In this work, we carefully select ultrasonic frequencies that balance attenuation and penetration to ensure sufficient depth coverage while maintaining adequate resolution. We establish the relationship between ultrasonic transmission signal variations and the spatiotemporal evolution of battery thermal runaway under different operating conditions.
We develop a comprehensive in-situ testing platform that integrates ultrasonic transmission, temperature, voltage, current, and surface deformation measurements for large-capacity prismatic aluminum-cased LFP energy storage batteries. Experimental studies are conducted under static, normal charge-discharge cycling, and overcharge conditions. We find that ultrasonic signal attenuation serves as an effective early warning indicator for overcharge thermal runaway. Based on the improved Mahalanobis-Taguchi system (MTS), we propose an early warning algorithm that reliably estimates the battery’s health state and predicts venting events 30 minutes in advance. This paper is organized as follows: Section 2 describes the experimental setup and methodology. Section 3 presents results and discussion, including static behavior, cycling characteristics, and overcharge dynamics. Section 4 introduces the early warning system based on the improved MTS. Section 5 concludes the work.
2 Experimental Setup and Methodology
2.1 Battery Samples
Commercial 20 Ah prismatic aluminum-cased LFP energy storage batteries are used as test objects. To ensure initial consistency, each battery undergoes three preconditioning cycles of 0.5 C constant current constant voltage (CCCV) charging followed by constant current (CC) discharging, with a rest period of at least 12 hours before any formal experiment. The nominal voltage is 3.2 V, and the operating voltage range is 2.5–3.65 V under normal conditions.
2.2 Experimental Platform
We construct an internal-external parameter testing system for LFP energy storage batteries, as illustrated conceptually in the experimental setup. The core component is an ultrasonic transmission detection system consisting of two ultrasonic transducers (transmitter and receiver) positioned on opposite sides of the battery, a signal generator generating a five-cycle Gaussian-modulated sinusoidal voltage pulse, a signal amplifier, and an oscilloscope for data acquisition. The ultrasonic waves propagate through the battery thickness, and the received signals are recorded for analysis. Simultaneously, we integrate the multi-parameter platform previously developed in our group, which includes an explosion-proof temperature chamber, a charge-discharge tester, thermocouples, a gas analysis system (monitoring CO, CO2, and H2), and high-temperature strain gauges for surface deformation measurement. This platform enables synchronized recording of voltage, current, temperature (at multiple locations including near the ultrasonic probe, under the terminals, and at corners), gas concentrations, and surface strain during overcharge thermal runaway experiments.
2.3 Ultrasonic Signal Normalization
The received ultrasonic signal is characterized by two primary parameters: the root mean square voltage (Urms) representing signal intensity, and the time of flight (TOF) representing the propagation time through the battery. The Urms is computed as follows:
$$U_{\text{rms}} = \sqrt{\frac{1}{N}\sum_{i=1}^{N} U_i^2}$$
where \(U_i\) is the voltage value of the \(i\)-th sampling point, and \(N\) is the total number of points in the received signal. The TOF is defined as the time from the start of the transmitted pulse to the first zero crossing or peak of the received signal.
To enable comparison across different batteries and experiments, we normalize these parameters relative to their initial values at time \(t=0\) (the beginning of each test). The relative Urms (RU) and relative TOF (RT) are defined as:
$$R_U = \frac{U_{\text{rms}}(t)}{U_{\text{rms}}(0)}, \quad R_T = \frac{T_{\text{OF}}(t)}{T_{\text{OF}}(0)}$$
2.4 Test Conditions
Three categories of experiments are conducted: (1) Static tests: the battery is placed in a temperature chamber and its surface temperature is varied from 20 °C to 45 °C at different fixed SOC levels (25%, 50%, 75%, 100%). (2) Normal cycling tests: batteries are cycled at 0.25 C, 0.5 C, and 0.75 C rates through a full charge-discharge cycle (CCCV charge to 3.65 V, CC discharge to 2.5 V) inside the chamber maintained at 25 °C. (3) Overcharge tests: starting from 100% SOC, batteries are continuously charged at 0.25 C, 0.5 C, and 0.75 C until thermal runaway occurs. All overcharge tests are conducted inside the explosion-proof chamber with the gas and strain monitoring systems active.
3 Results and Discussion
3.1 Static State: Temperature and SOC Effects
3.1.1 Influence of Temperature
Figure in the experimental section shows the variation of RU and RT with battery surface temperature at different SOC levels. As temperature increases, RU decreases monotonically. For example, at 45 °C, RU drops to approximately 20% of its value at 20 °C. This is attributed to the decrease in electrolyte viscosity and density with rising temperature, which reduces the acoustic impedance of the electrolyte. The impedance mismatch between the electrolyte and electrode materials increases, lowering the transmission coefficient and thus Urms. Conversely, RT increases with temperature, reaching about 110% at 45 °C relative to 20 °C. Two factors contribute: thermal expansion of electrodes increases the propagation path length, and temperature-induced changes in material properties (density and elastic modulus) reduce the effective sound speed, both increasing TOF.
A notable inflection point is observed around 35 °C for both RU and RT. Below 35 °C, the variations are nearly linear; above 35 °C, RU decreases more rapidly and RT increases nonlinearly. This behavior can be explained by the phase transition of ethylene carbonate (a major component of the electrolyte) from solid to liquid at around 35 °C, which significantly reduces the elastic modulus of the electrolyte, thereby altering ultrasonic transmission characteristics. The sensitivity of RU to temperature is substantially higher than that of RT.
3.1.2 Influence of SOC
The relationship between ultrasonic signals and SOC under static conditions is shown in Figure 5. As SOC increases from 0% to 100%, RU first increases by up to 11.56%, then slightly decreases. RT decreases approximately linearly, with a maximum reduction of 0.67%. The change in RU is more pronounced than that in RT, indicating that Urms is a more sensitive indicator of SOC than TOF.
The mechanism involves lithium concentration changes in the electrodes. During charging, lithium ions deintercalate from the cathode and intercalate into the anode. This alters the density and elastic modulus of both electrodes. At low SOC, the cathode (typically LiFePO4) has its highest elastic modulus while the anode (graphite) has its lowest, leading to a maximum impedance mismatch and minimum ultrasonic transmission. As SOC increases, the cathode modulus decreases, the anode modulus increases, and both densities decrease, reducing impedance mismatch. Additionally, the ultrasonic attenuation coefficient of the battery materials decreases with SOC due to lattice relaxation effects. These factors cause RU to initially increase. After a certain SOC, further changes in the anode thickness and material properties may reverse the trend. The RT decrease indicates that the effective sound speed increases (dominated by the anode’s modulus increase) despite the slight swelling of the anode thickness.
3.2 Normal Charge-Discharge Cycling
During normal cycling at different rates (0.25 C, 0.5 C, 0.75 C), the RU and RT profiles are consistent with the static behavior but with additional transient effects due to current-induced polarization. Figure 6 presents characteristic parameter changes over a full cycle at the three rates.
3.2.1 RU during Cycling
The RU variation during charging and discharging follows the same trend as the static SOC dependence. However, at the onset and termination of current, sharp jumps in RU are observed. These jumps are attributed to concentration polarization: when current is applied, lithium ion diffusion cannot keep pace, causing local accumulation at the electrode surface and depletion in the interior. This non-uniform lithium distribution changes the local mechanical properties, leading to abrupt changes in ultrasonic transmission. The magnitude of the jumps increases with C-rate. After the current stops (during rest), the polarization relaxes and RU returns to its equilibrium value.
3.2.2 RT during Cycling
During both charging and discharging, RT increases, and the increase is more pronounced at higher C-rates. The RT variation shows a strong correlation with the battery surface temperature change (Figure 6c). In fact, the RT curve nearly overlaps the temperature curve in shape. This indicates that the RT change during cycling is dominated by temperature-induced expansion and sound speed reduction, rather than by lithium concentration effects. The internal heating from resistive losses causes electrode swelling and reduces wave speed, increasing TOF. The contribution of lithium migration to elasticity change is secondary under dynamic conditions.
3.3 Overcharge Thermal Runaway
3.3.1 Voltage, Gas, and Deformation
Overcharge experiments are conducted at 0.25 C, 0.5 C, and 0.75 C from 100% SOC. The voltage evolution (Figure 7a) shows three distinct phases: Phase I – voltage increases steadily as lithium plating on the anode increases internal resistance; Phase II – voltage continues increasing but at a slower rate as the cathode becomes lithium-deficient and structure distortion begins; Phase III – voltage decreases due to severe lithium consumption by side reactions and SEI breakdown. The venting valve opens at the end of Phase III, with voltage around 4.99–5.01 V across different rates.
Gas monitoring (Figure 7b) reveals that H2, CO2, and CO are not detected before venting because the aluminum casing confines the gases. Immediately after venting, gas concentrations surge: H2 exceeds the sensor range (1000 ppm), CO2 rises from ambient 420 ppm to 720 ppm, and CO peaks at 300 ppm. These gases originate from lithium plating reactions with electrolyte moisture, binder decomposition, and electrolyte reduction.
Surface strain (Figure 7c) increases gradually during Phase II and III as internal gas pressure builds. The strain at the center of the battery increases earlier than at the corner, indicating that gas accumulation initiates near the core. Upon venting, strain drops abruptly as pressure is released.
3.3.2 Ultrasonic Signal Evolution
The most striking observation is the ultrasonic signal behavior during overcharge (Figure 7d). RU decreases continuously, and RT increases continuously, until at a certain point the signal is completely attenuated (noise floor). This occurs significantly earlier than venting valve activation. For 0.25 C, the signal disappears at 1150 s into overcharge, compared to venting at 5154 s. For 0.5 C, the times are 586 s vs. 2418 s. For 0.75 C, 257 s vs. 1537 s. The higher the C-rate, the faster the ultrasonic signal is lost.
The mechanism involves several overlapping effects: (1) Cathode over-lithiation (actually de-lithiation in LFP) causes structural distortion increasing ultrasonic absorption; (2) Lithium plating on the anode reduces anode elastic modulus, increasing impedance mismatch; (3) Plating and subsequent gas generation (from SEI breakdown and electrolyte decomposition) create gas bubbles that strongly reflect and scatter ultrasound. As gas accumulates, the ultrasonic path is blocked, leading to complete signal attenuation. The sudden drop of RU to near zero is a clear, early indicator of internal gas evolution, well before the external manifestation of venting.
Table 1 summarizes the key timing differences observed:
| Overcharge Rate | Time to Ultrasonic Signal Loss (s) | Time to Venting Valve Opening (s) | Advance Warning Time (min) |
|---|---|---|---|
| 0.25 C | 1150 | 5154 | 66.7 |
| 0.5 C | 586 | 2418 | 30.5 |
| 0.75 C | 257 | 1537 | 21.3 |
3.4 Mechanism Summary
The ultrasonic response to overcharge in LFP energy storage batteries is a combined result of electrode property changes, gas generation, and temperature rise. Among all monitored parameters (voltage, temperature, gas concentration, surface strain), the ultrasonic RU and RT changes are the earliest and most dramatic. Therefore, ultrasonic attenuation can serve as an ideal early warning indicator for thermal runaway in energy storage batteries.
4 Early Warning System Based on Improved Mahalanobis-Taguchi System
4.1 Methodology
To develop a robust early warning algorithm, we adopt an improved Mahalanobis-Taguchi system (IMTS). The standard MTS constructs a reference space (Mahalanobis space) from normal data and measures the Mahalanobis distance (MD) of unknown samples. An MD exceeding a threshold indicates an anomaly. The Mahalanobis distance is computed using the covariance matrix to account for correlations among features. However, the standard MTS suffers from instability and inability to handle multicollinearity. We apply the improved version using Gram-Schmidt orthogonalization (GSP) to compute MDs without matrix inversion, and we assign feature weights via Fisher’s criterion to emphasize the most discriminative parameters.
4.2 Feature Selection and Preprocessing
Based on our experimental findings, we select six features that best characterize the onset of overcharge: battery voltage (V), mid-surface temperature (T), ultrasonic Urms, ultrasonic TOF, H2 concentration, and mid-surface strain. Some features increase during overcharge (e.g., temperature, strain) while others decrease (Urms). To ensure monotonicity, we forward-transform features that decrease (e.g., Urms becomes 1/Urms or complement). Then all features are normalized using Z-score standardization based on the mean and standard deviation of the normal training set. The Mahalanobis distance for sample vector x is computed as:
$$MD(\mathbf{x}) = (\mathbf{x} – \boldsymbol{\mu})^T \mathbf{C}^{-1} (\mathbf{x} – \boldsymbol{\mu})$$
In our GSP approach, we avoid direct inversion by orthogonalizing the normalized data.
4.3 Training and Validation
Normal training data consists of 4965 samples collected during one complete 0.5 C cycle (including charge, discharge, and rest) from a healthy LFP energy storage battery. Anomalous data consists of 3644 samples from a 0.5 C overcharge experiment to thermal runaway. After preprocessing, we first compute the standard MD for normal samples (MD1) and abnormal samples (MD2). The results show:
$$\text{MD1} = 0.9998, \quad \text{MD2} = 7.0497 \times 10^4$$
Since MD2 >> MD1, the reference space is effective in distinguishing normal from abnormal states.
Next, we compute feature weights using Fisher’s criterion. The weight for each feature i is given by:
$$w_i = \frac{(\bar{y}_{i,\text{normal}} – \bar{y}_{i,\text{abnormal}})^2}{s_{i,\text{normal}}^2 + s_{i,\text{abnormal}}^2}$$
where \(\bar{y}\) is the mean and \(s^2\) is the variance of the normalized feature. The weighted Mahalanobis distance (WMD) is then:
$$WMD = \sqrt{\sum_{i=1}^{p} w_i \cdot (z_i)^2}$$
After weighting, the WMD for normal samples (WMD1) remains 0.9998, while for abnormal samples (WMD2) increases to 7.2525 × 104, confirming improved separability.
4.4 Health Index and Threshold Determination
We define a health index (HI) that normalizes the WMD to the range [0, 1], with 1 indicating perfectly healthy. The HI is computed as:
$$HI = \exp\left(-\frac{WMD – \mu_n}{\sigma_n}\right)$$
Using the 3σ rule (99.73% confidence), we calculate a threshold HIt such that only 0.27% of normal samples fall below it. In practice, taking a confidence level of 99.15% (based on the distribution of normal sample WMDs), we determine the threshold to be HIt = 0.9222. An overcharge alarm is triggered when three consecutive samples have HI < HIt.
4.5 Algorithm Validation
We validate the algorithm using a separate test: an initialized LFP energy storage battery is charged at 0.5 C from 0% SOC to 110% SOC (overcharge). The HI computed from the test data is plotted against charging time. Initially, HI rises slightly due to normal SOC changes, then remains stable above HIt. Near the end of charging, HI drops sharply, crossing below the threshold at 7262 s into charge. At this moment, the battery SOC is estimated at 104.2% (by coulomb counting). This is during Phase I of overcharge, well before venting (which occurs at 2418 s from 100% SOC, i.e., about 7270 s total charge time). The early warning provides an advance notice of approximately 30 minutes compared to venting activation.
Table 2 lists the values of each feature at the warning moment:
| Feature | Value at Warning |
|---|---|
| Voltage (V) | 4.46 |
| Mid-temperature (°C) | 21.91 |
| Urms (V) | 0.677 |
| TOF (μs) | 20.95 |
| H2 concentration (ppm) | 0 |
| Mid-strain (×10-6) | 53 |
4.6 Discussion of the Early Warning System
The IMTS-based algorithm successfully leverages the high sensitivity of ultrasonic signals to internal degradation. Compared to conventional methods that rely on voltage or temperature thresholds, our system provides a significantly earlier alarm, which is critical for taking corrective actions (e.g., disconnecting the battery) to prevent catastrophic thermal runaway. The use of weighted Mahalanobis distance enhances robustness by emphasizing the most informative features (primarily Urms and TOF) and suppressing noise from less sensitive parameters. The algorithm is computationally efficient and can be deployed in battery management systems (BMS) for real-time monitoring of energy storage batteries.
5 Conclusion
This research demonstrates that ultrasonic transmission monitoring is a powerful tool for early detection of thermal runaway in large-format prismatic aluminum-cased LFP energy storage batteries. Key findings are:
- Relative ultrasonic intensity (RU) is more sensitive to temperature and SOC changes than time-of-flight (RT). During normal cycling, RT is dominated by temperature effects, while RU effectively reflects internal state.
- During overcharge, ultrasonic signals are lost significantly earlier than venting valve activation, with the time advantage ranging from 21 to 67 minutes depending on C-rate. This makes ultrasonic attenuation a reliable early indicator for overcharge thermal runaway in energy storage batteries.
- A health index based on an improved Mahalanobis-Taguchi system, incorporating six features (voltage, temperature, Urms, TOF, H2 concentration, strain), can predict venting events up to 30 minutes in advance with high reliability.
The proposed methodology provides a robust foundation for integrating ultrasonic sensors into battery management systems, thereby enhancing the safety and operational reliability of energy storage battery installations. Future work will focus on long-term cycling effects, multi-cell pack integration, and field validation in actual energy storage plants.