Renewable energy sources such as wind and solar power are inherently intermittent and cannot provide stable continuous power output. Energy storage technology has therefore become a key enabler for the large-scale integration of renewable generation. Among various storage options, the energy storage battery based on lithium-ion chemistry offers high round-trip efficiency, long cycle life, and flexible deployment, making it the fastest-growing technology for stationary energy storage. Within the lithium-ion family, lithium iron phosphate (LFP) batteries are particularly attractive because of their excellent thermal stability, long calendar life, and relatively low cost. In the context of utility-scale energy storage power stations, LFP batteries have become the dominant choice for the energy storage battery market.
However, the safety of LFP-based energy storage battery systems remains a major technical challenge. The long-term operation of these batteries, combined with unavoidable inconsistency among cells, degradation of performance, and possible failure of the battery management system, can lead to abusive conditions such as overcharge, over-discharge, and internal short circuit. Overcharge is one of the most common triggers of thermal runaway in LFP energy storage battery modules. During overcharge, the negative electrode may deposit lithium dendrites, the positive electrode structure may be damaged, and the electrolyte may decompose, generating flammable gases. When the internal pressure exceeds the design limit, the safety vent opens, and the battery may eventually undergo thermal runaway, accompanied by fire or explosion. Therefore, early detection and warning of overcharge-induced thermal runaway is of utmost importance for the safe operation of large-scale energy storage battery stations.
Conventional monitoring techniques in battery energy storage plants rely on module-level voltage, current, and temperature measurements, together with gas or smoke detectors at the container level. These methods are useful, but they usually detect anomalies after the battery has already vented or after significant external changes have occurred. The response of external parameters to internal changes is often delayed, which limits the effectiveness of early warning. In recent years, ultrasonic testing has emerged as a promising non-destructive technique for the in-situ diagnosis of lithium-ion batteries. Because ultrasound waves propagate through the internal structure of the battery, they are sensitive to changes in material properties, state of charge, gas generation, and mechanical deformation. Most prior ultrasonic studies have focused on soft-pouch or small-format cells, whose thin casing and layered aluminum-plastic film allow relatively easy ultrasonic penetration. In contrast, large-format prismatic aluminum-shell LFP cells, which are widely used in grid-scale energy storage battery systems, have a thickness greater than 20 mm and a rigid aluminum housing that strongly reflects ultrasound waves. This makes ultrasonic monitoring considerably more challenging, but also more valuable if it can be achieved.
In this work, we propose an in-situ ultrasonic transmission method for monitoring the internal state of a prismatic aluminum-shell LFP energy storage battery during thermal runaway. We construct a comprehensive test system that combines ultrasonic transmission, temperature, voltage, current, and surface strain measurements. The system allows us to simultaneously record the evolution of internal and external parameters under static, cycling, and overcharge conditions. We analyze the response mechanism of ultrasonic signals to internal structural changes and demonstrate that the attenuation of the ultrasonic signal can serve as an early indicator of overcharge-induced thermal runaway. Based on these insights, we develop an enhanced Mahalanobis-Taguchi system (IMTS) for predicting the opening of the safety vent. The proposed algorithm provides a reliable early warning approximately 30 minutes before the vent opens, offering a substantial improvement in safety for energy storage battery applications.
Experimental Setup
To represent the type of energy storage battery used in actual utility-scale systems, we selected a commercial 20 Ah prismatic aluminum-shell LFP cell as the research object. Before any test, each battery was preconditioned with three complete charge-discharge cycles at 0.5 C using the constant-current constant-voltage charging and constant-current discharging protocol. The battery was then rested for more than 12 hours to ensure initial consistency. The key specifications of the test cell are listed in Table 1.
| Parameter | Value |
|---|---|
| Rated capacity | 20 Ah |
| Nominal voltage | 3.2 V |
| Charge cut-off voltage | 3.65 V |
| Cell format | Prismatic |
| Case material | Aluminum |
| Thickness | ~27 mm |
| Anode material | Graphite |
| Cathode material | LiFePO4 |
The complete thermal runaway testing platform consisted of several integrated subsystems. An explosion-proof temperature chamber provided a controlled ambient environment and isolated the test cell for safety. A battery charge-discharge tester was used to apply the required current profiles and to record voltage and current data automatically. The temperature on the battery surface was measured with multiple K-type thermocouples attached at the center, at the positive and negative terminals, and at the two bottom corners. Surface strain was measured with high-temperature resistance strain gauges placed near the center and the upper corner of the battery. A ducted fan extracted gas from above the safety vent into a gas collection box containing electrochemical sensors for CO, CO2, and H2. The ultrasonic system consisted of a pair of ultrasonic transducers, a signal generator, an amplifier, and a digital oscilloscope. Two ultrasonic transducers were placed on the two largest opposite faces of the prismatic cell, so that the ultrasonic wave propagated through the entire thickness of the cell in a transmission configuration. A five-cycle Gaussian modulated sine pulse generated by the signal generator and amplified by the amplifier was used to drive the transmitting transducer. The receiving transducer detected the wave transmitted through the cell, and the waveform was recorded by the oscilloscope.
The sensor layout on the battery surface was carefully arranged. The ultrasonic transducer, thermocouple, and strain gauge were distributed without interfering with each other, and the positions are schematically indicated in the figure above. The voltage and current were recorded by the charge-discharge tester at a sampling rate of 10 Hz. Temperature, gas concentration, and strain signals were logged at 1 Hz. The ultrasonic waveform was captured at 0.2 Hz during normal cycling and at 1 Hz during overcharge tests to capture fast dynamic changes.
Ultrasonic Signal Processing
Two characteristic parameters were extracted from the received ultrasonic waveform: the root-mean-square voltage, denoted as Urms, and the time of flight (TOF), denoted as TOF. The Urms value represents the overall intensity of the transmitted ultrasonic signal and is calculated as
$$U_{\text{rms}} = \sqrt{\frac{1}{N}\sum_{i=1}^{N} U_i^2},$$
where Ui is the voltage of the i-th sampled data point and N is the total number of data points in the time window. The TOF is defined as the time delay between the transmitted pulse and the first arriving received pulse. In practice, we determined TOF by finding the time at which the received signal envelope first crosses a threshold that is proportional to the maximum amplitude. This method is robust against noise and variations in signal amplitude.
Because different battery cells may have slightly different internal structures and acoustic properties, we normalized the ultrasonic features to their initial values at time t = 0, yielding relative parameters:
$$R_U = \frac{U_{\text{rms}}(t)}{U_{\text{rms}}(t=0)}, \qquad R_T = \frac{T_{\text{OF}}(t)}{T_{\text{OF}}(t=0)}.$$
This normalization allowed us to compare the evolution of ultrasonic signals across different cells and different test conditions. All results presented below use these dimensionless relative values unless otherwise stated.
Ultrasonic Response under Static Conditions
Effect of Battery Temperature
To understand how temperature affects the ultrasonic propagation inside a prismatic aluminum-shell energy storage battery, we first performed controlled temperature experiments. The battery at a fixed state of charge (SOC) was placed inside the temperature chamber, and the chamber temperature was increased from 20 °C to 45 °C in steps. At each temperature setpoint, the ultrasonic waveform was recorded after thermal equilibrium was reached. The normalized ultrasonic parameters as functions of surface temperature are shown in Figure 4 of the original paper (not reproduced here). In our experiments, we observed that RU continuously decreased with increasing temperature. This can be explained by the reduction in the viscosity and density of the electrolyte as the temperature rises, which lowers the acoustic impedance of the electrolyte and increases the impedance mismatch between the electrolyte and the electrodes. The increased impedance mismatch causes more reflection of the ultrasound wave at the interfaces, thus reducing the transmitted signal intensity.
In contrast, RT increased with temperature. Three factors contribute to this behavior. First, the electrodes expand upon heating, increasing the total path length through the cell. Second, the elastic modulus and density of the electrode materials change with temperature, reducing the effective ultrasonic wave velocity. Third, the battery temperature rise also alters the compressibility of the electrolyte, further reducing the wave speed. The combined effect is an increase in TOF. Notably, the curves of both RU and RT showed a distinct inflection point near 35 °C. Above this temperature, the attenuation of the signal increased sharply and the TOF growth became nonlinear. This inflection is attributed to the phase transition of ethylene carbonate, one of the main components of the electrolyte, from the solid to the liquid state. The phase transition significantly changes the elastic properties of the electrolyte and thus the acoustic impedance matching at the interfaces.
Comparing the magnitudes of the changes, we found that at 45 °C the TOF increased to about 110% of its value at 20 °C, whereas the Urms decreased to about 20% of its initial value. Therefore, the ultrasonic intensity is far more sensitive to temperature changes than the TOF. This suggests that RU is the preferred indicator when temperature variations are significant, such as during the early stage of thermal runaway.
Effect of State of Charge
We also investigated the influence of SOC on the ultrasonic signals under quasi-static conditions. The battery was charged at a low rate to specific SOC levels, and the ultrasonic signals were recorded after the battery had rested to allow the concentration gradients to relax. The results are summarized in Table 2. The relative ultrasonic intensity RU increased by up to 11.56% as the SOC was raised from 0% to 100%, while the relative TOF RT decreased by up to 0.67% over the same range. Both parameters changed monotonically in most SOC intervals, but the sensitivity of RU to SOC was significantly higher than that of RT.
| Parameter | Change from 0% to 100% SOC |
|---|---|
| RU | +11.56% |
| RT | −0.67% |
The observed SOC dependence arises from the complex interplay of lithium concentration, electrode stiffness, and interfacial impedance. When lithium ions are extracted from the positive electrode and inserted into the negative electrode, the lattice parameters and elastic properties of both electrodes change. The positive electrode (LiFePO4) becomes stiffer as lithium is removed, while the negative graphite electrode becomes softer and expands upon lithiation. The density and elastic modulus variations alter the acoustic impedance of the electrode materials and thereby affect the transmission coefficient at each interface. In addition, the state of lithiation changes the damping characteristics of the electrode, which influences the attenuation coefficient. The slight decrease in RT with increasing SOC indicates that the effective wave velocity increases, which can be attributed to the dominant effect of the graphite negative electrode, whose elastic modulus increases with lithium content, leading to higher wave speed despite the increased electrode thickness. Overall, the ultrasonic intensity is a reliable indicator of SOC for the energy storage battery under static or quasi-static conditions.
Ultrasonic Behavior during Normal Charge-Discharge Cycling
To evaluate the applicability of ultrasonic monitoring in realistic operating scenarios, we performed continuous charge-discharge cycling at different C-rates (0.25 C, 0.5 C, and 0.75 C). Figure 6 of the original paper (not reproduced here) showed the simultaneous evolution of RU, RT, and battery surface temperature as a function of time during a full cycle. The values of RU during the charge and discharge phases followed the same general trend as that observed in the static SOC test. However, at the very beginning of charge or discharge, a sudden jump or drop in RU occurred, which was more pronounced at higher C-rates. This transient effect is caused by concentration polarization. When a current is applied, lithium ions are rapidly consumed or produced at the electrode surfaces, creating a concentration gradient because the diffusion rate inside the electrodes is finite. The establishment of the concentration gradient changes the mechanical and acoustic properties of the electrode-electrolyte interface almost instantaneously. When the current is interrupted, the concentration gradient gradually relaxes, and RU returns to its quasi-equilibrium value. The strong dependence on C-rate clearly demonstrates that the initial transient is governed by mass transport kinetics.
The TOF response during cycling showed a different behavior. The relative TOF RT increased during both charging and discharging, and the increase was faster at higher currents. This pattern closely followed the surface temperature curve. Since ohmic and polarization heat generate heat during current flow, the battery temperature rises during charge and discharge, and then decreases during rests. The correlation between RT and temperature suggests that the TOF is predominantly controlled by temperature in dynamic operation, which alters both the electrode thickness and the wave velocity. Although lithium concentration also affects the TOF, its influence is secondary compared to the thermal effect during cycling. These findings indicate that for the energy storage battery under normal operating conditions, RU can provide reliable information on the internal state, whereas RT mainly reflects thermal variations.
Overcharge Thermal Runaway Characteristics
Overcharge experiments were conducted on fully charged (100% SOC) LFP energy storage battery cells at constant current rates of 0.25 C, 0.5 C, and 0.75 C. The cells were charged continuously until thermal runaway occurred or until the safety vent opened. The experimental platform safely contained the venting gases and recorded all parameters at high sampling rates. The behavior was consistent across all tests, with three distinct stages identified based on the voltage profile and the ultrasonic signal loss.
Voltage and Current Response
Figure 7(a) of the original paper (not reproduced here) shows a typical voltage and current trace during 0.5 C overcharge. In the first two stages, the voltage increased monotonically. This is because the negative electrode of the LFP cell has an excess capacity designed for safety. During overcharge, the lithium ions extracted from the positive electrode are partially intercalated into the negative electrode, while the remainder forms lithium dendrites on the negative electrode surface. The negative electrode potential eventually reaches a plateau, and the cell voltage increases rapidly. In the third stage, the voltage started to decrease. This decrease results from severe lithium depletion at the positive electrode, structural distortion of the cathode, and the reaction of lithium dendrites with the electrolyte, which dissolves lithium and increases the internal resistance in a manner that temporarily lowers the terminal voltage. It is noteworthy that at the moment of vent opening, the cell voltage under all three C-rates was remarkably similar: 4.99 V, 4.98 V, and 5.01 V for 0.25 C, 0.5 C, and 0.75 C, respectively. This consistent value provides a useful reference for voltage-based diagnostics.
Gas Generation and Surface Strain
The battery surface strain began to increase slowly during the first stage of overcharge, mainly because of the rise in temperature and the consequent contraction of the insulation blue film. In the second and third stages, gas generation inside the airtight cell caused the internal pressure to increase, leading to visible expansion of the aluminum case. The strain gauges recorded a gradually increasing surface strain until the safety vent opened. At the moment of vent opening, the gas was rapidly released, and the surface strain abruptly dropped. The gas sensors detected a sharp increase in H2, CO2, and CO concentrations simultaneously with the vent opening. H2 showed the most dramatic change, increasing immediately to the sensor saturation limit of 1000 ppm. CO2 increased from a background level of about 420 ppm to 573 ppm after the vent opened and then gradually reached 720 ppm. CO increased to 300 ppm and then slowly decayed. These gas species are formed through multiple mechanisms: lithium dendrites react with water and binder materials to produce H2; the decomposition of the SEI layer and reaction of the electrolyte with the negative electrode generate CO2; and further reduction of CO2 by lithium produces CO. Organic gases may also be generated if the local temperature is high enough to decompose the electrolyte. However, these gases are only detectable after the vent opens, so they cannot provide an early warning before venting.
Ultrasonic Signal Evolution
Figure 7(d) of the original paper (not reproduced here) illustrates the change of RU and RT during a 0.5 C overcharge test. As overcharging proceeded, RU continuously decreased, while RT continuously increased until the ultrasonic signal became so weak that the TOF could not be reliably extracted. The rate of decrease in RU accelerated substantially in the second and third stages. The physical reasons for the ultrasonic attenuation are manifold. In the first stage, the positive electrode is delithiated beyond its stable composition, causing structural distortion and increased absorption of ultrasound. Lithium dendrites form on the negative electrode, which lowers the effective elastic modulus and increases the impedance mismatch. In the second and third stages, gas bubbles are generated inside the cell, and these gas bubbles cause strong scattering of the ultrasonic waves. As the gas volume fraction increases, the ultrasonic wave is completely attenuated. Similarly, the increase in TOF during overcharge is associated with the softening of the electrodes due to lithium plating, the expansion of the battery caused by gas generation and temperature rise, and the reduction of the wave velocity in the gas-filled porous structure.
We performed overcharge tests at three different C-rates to compare the time sequences. Table 3 summarizes the key time points for each test. The ultrasonic signal completely disappeared at 1150 s, 586 s, and 257 s for 0.25 C, 0.5 C, and 0.75 C, respectively. In contrast, the safety vent opened at 5154 s, 2418 s, and 1537 s for the same C-rates. Thus, the ultrasonic signal was lost far earlier than the vent opening, by a margin of approximately 4000 s, 1830 s, and 1280 s for the three rates. This dramatic gap demonstrates that the ultrasonic attenuation is an extremely early indicator of an abnormal internal condition inside the energy storage battery.
| C-rate | Ultrasonic signal loss time (s) | Vent opening time (s) | Early warning margin (s) |
|---|---|---|---|
| 0.25 C | 1150 | 5154 | 4004 |
| 0.5 C | 586 | 2418 | 1832 |
| 0.75 C | 257 | 1537 | 1280 |
It is worth noting that the ultrasonic signal loss occurs while the external indicators such as temperature, voltage, and gas concentration still show little change. The battery surface temperature at the time of ultrasonic signal loss was still relatively low, and no gas had been released. This confirms that ultrasonic monitoring probes the internal state of the cell, providing information that cannot be obtained by conventional external sensors. Therefore, the ultrasonic signal is an excellent candidate for early warning of overcharge-induced thermal runaway in prismatic aluminum-shell LFP energy storage battery cells.
Early Warning System Based on Improved Mahalanobis-Taguchi System
Although the ultrasonic signal itself is a powerful indicator, we can further improve the reliability and robustness of the warning system by combining it with other measurable parameters. We developed an early warning algorithm based on the improved Mahalanobis-Taguchi system (IMTS). The Mahalanobis-Taguchi system is a pattern recognition technique designed for handling imbalanced data. It uses a set of normal samples to establish a reference space, and then evaluates the deviation of new samples from this reference space using the Mahalanobis distance (MD). Compared with the Euclidean distance, the MD accounts for the correlation between variables and is scale-invariant, making it well suited for battery condition monitoring.
However, the traditional Mahalanobis-Taguchi system has several limitations. First, the reference space may not be robust when the sample size is small. Second, the MD cannot be computed if the covariance matrix is singular due to multicollinearity among the features. Third, all feature variables are treated equally, which is not always appropriate because some parameters are more sensitive to overcharge than others. To overcome these problems, we employed the improved Mahalanobis-Taguchi system described below.
Feature Selection and Preprocessing
The first step was to choose the feature variables that best reflect the battery overcharge state. Based on the experiments, we selected six features: cell voltage, center temperature, ultrasonic Urms, ultrasonic TOF, H2 concentration, and center strain. These features are listed in Table 4. The selected features cover both external and internal quantities and provide complementary information about the thermal, mechanical, and acoustic state of the energy storage battery.
| No. | Feature variable | Symbol |
|---|---|---|
| 1 | Voltage | V |
| 2 | Center temperature | T |
| 3 | Ultrasonic intensity | Urms |
| 4 | Ultrasonic time of flight | TOF |
| 5 | H2 concentration | cH2 |
| 6 | Center strain | ε |
Some features increase during overcharge (e.g., temperature, TOF, H2 concentration, strain) and others initially increase then decrease (e.g., voltage), while ultrasonic intensity decreases. To make all features vary in a consistent direction, we applied a sign transformation. For features that should decrease when the battery becomes abnormal, we took the negative of the normalized value. After this preprocessing, all features were standardized using the Z-score method, i.e.,
$$z_{ij} = \frac{x_{ij} – \mu_j}{\sigma_j},$$
where xij is the i-th sample of the j-th feature, and μj and σj are the mean and standard deviation of feature j computed from the normal reference samples.
Construction of the Reference Space
The normal reference samples were collected during one complete charge-discharge cycle of a healthy LFP energy storage battery at 0.5 C, including the constant-current constant-voltage charge, the constant-current discharge, and the rest periods. The sampling frequency was 0.2 Hz, giving a total of 4965 normal samples. The abnormal samples were obtained from a separate overcharge test at 0.5 C, with a sampling rate of 1 Hz, totaling 3644 samples. These two sets formed the training dataset.
To avoid the singularity problem of the covariance matrix, we used the Gram-Schmidt orthogonalization process (GSP) to compute the Mahalanobis distance. The GSP method transforms the standardized variables into an orthogonal set, allowing the MD to be calculated as the sum of the squared orthogonal components. The Mahalanobis distance for a sample vector z is given by
$$MD = \sqrt{\sum_{k=1}^{p} \frac{y_k^2}{s_k^2}},$$
where yk is the k-th orthogonal component, sk2 is the variance of the k-th orthogonal component in the reference space, and p is the number of features. In this way, the MD value is a positive scalar that quantifies the deviation of the sample from the normal condition. We computed the MD for the normal samples, obtaining a mean reference distance MDnormal of approximately 1. For the abnormal samples, the MD was on the order of 104, which demonstrates that the reference space clearly separates normal and overcharged conditions.
Feature Weighting Using Fisher Criterion
Because the six features do not contribute equally to the fault detection, we assigned a weight to each feature based on the Fisher criterion. The weight wj for feature j was calculated as
$$w_j = \frac{(\bar{z}_{j,\text{abnormal}} – \bar{z}_{j,\text{normal}})^2}{\sigma_{j,\text{normal}}^2 + \sigma_{j,\text{abnormal}}^2},$$
where z̅j,abnormal and z̅j,normal are the means of the standardized feature j in the abnormal and normal sets, respectively, and σj,normal2 and σj,abnormal2 are the corresponding variances. A larger weight indicates that the feature distinguishes better between normal and abnormal states. We normalized the weights so that they sum to one. The weighted Mahalanobis distance (WMD) was then defined as
$$WMD = \sqrt{\sum_{k=1}^{p} w_k \frac{y_k^2}{s_k^2}}.$$
Using the training data, we computed the WMD for both normal and abnormal samples. The mean WMD for normal samples was close to 1, while for abnormal samples it was significantly larger, confirming that the weighted reference space is effective.
Health Index and Threshold
To provide an intuitive single scalar for battery health, we converted the WMD into a health index (HI) scaled between 0 and 1. The definition is
$$HI = \frac{WMD_{\max} – WMD}{WMD_{\max} – WMD_{\min}},$$
where WMDmax and WMDmin are the maximum and minimum WMD values observed in the normal training set. During healthy operation, HI is close to 1. As the battery degrades or experiences overcharge, WMD increases and HI decreases. A low HI indicates an abnormal condition.
We employed the 3σ criterion to determine the alarm threshold. The mean μHI and standard deviation σHI of the HI values for the normal samples were calculated. The lower threshold was set as
$$HI_{\text{th}} = \mu_{HI} – 3 \sigma_{HI}.$$
However, because the distribution of HI is not perfectly Gaussian, we slightly adjusted the threshold using a confidence level derived from the actual normal data. Specifically, we found that 99.15% of the normal samples had HI values above the preliminary threshold. We therefore defined the final threshold as the 0.15th percentile of the normal HI distribution, which was 0.9222. If three consecutive samples have HI values below this threshold, the algorithm declares an overcharge condition. Requiring three consecutive samples reduces the probability of false alarms due to transient noise.
Warning System Implementation and Validation
The complete workflow of the proposed warning system is shown in Figure 9 of the original paper (not reproduced here). The algorithm can be summarized in the following steps:
| Step | Action |
|---|---|
| 1 | Select six features and preprocess them (sign handling and Z-score standardization). |
| 2 | Construct the reference space from normal samples using GSP. |
| 3 | Verify that the MD of abnormal samples is much larger than normal MD. |
| 4 | Compute Fisher weights and define WMD. |
| 5 | Compute WMD for normal samples and derive HI. |
| 6 | Set alarm threshold based on 3σ rule and adjust based on confidence level. |
| 7 | Evaluate new samples; if three consecutive HI values are below threshold, raise overcharge alarm. |
To validate the algorithm, we performed an independent overcharge experiment on a fresh LFP energy storage battery. The cell was charged from 0% SOC at a constant current of 0.5 C up to 110% SOC. The six feature variables were measured continuously at a sampling rate of 1 Hz. Figure 11 in the original paper (not reproduced here) shows the HI index as a function of charging time. During the initial stage of charging, the HI value showed a transient increase due to the onset of polarization. Afterwards, the HI remained relatively stable and well above the threshold. At around the latter stage of charge, the HI began to decline rapidly and eventually dropped below the alarm threshold. The algorithm detected the overcharge at t = 7262 s, when three consecutive HI values were below HIth. According to Ah counting, the SOC at this moment was 104.2%.
At the warning time, the cell voltage was 4.46 V, the center temperature was 21.91 °C, the ultrasonic Urms had decreased to 0.677 V from an initial value close to 1 V, the TOF was 20.95 μs, the H2 concentration was still zero, and the center strain was about 53×10-6. In comparison, the vent opening in this type of cell under the same overcharge condition occurred approximately 30 minutes later. This demonstrates that the ultrasonic-based combined warning system can provide an early alarm with a 30-minute lead time, which is sufficient for operators to take preventive measures such as interrupting the charging current or disconnecting the faulty module.
Discussion
The results of this study clearly show that ultrasonic testing can reveal the internal state of a thick, prismatic, aluminum-shell LFP energy storage battery with high sensitivity. The key advantage of ultrasonic monitoring is that it is non-destructive, in-situ, and can be performed continuously in real time. The ultrasonic intensity is particularly sensitive to early gas generation, which occurs well before the safety vent opens. In all overcharge tests, the ultrasonic signal was completely lost thousands of seconds before the vent opened, making it one of the earliest observable indicators.
It is important to note that while the ultrasonic signal provides excellent early warning, it does not uniquely identify the type of fault. For example, a mechanical deformation or a local thermal event could also cause a reduction in Urms. Therefore, we combined ultrasonic parameters with voltage, temperature, gas concentration, and strain measurements in the Mahalanobis-Taguchi framework. This multi-parameter fusion improves robustness and reduces the probability of false alarms caused by a single parameter fluctuation. The improved MTS algorithm gives a single health index that captures the overall deviation of the energy storage battery from its healthy baseline. The weighting scheme based on the Fisher criterion automatically emphasizes the features that are most discriminative for overcharge detection, such as the ultrasonic intensity and the voltage deviation.
One practical consideration for field deployment is the coupling between the ultrasonic transducer and the battery surface. In our laboratory experiments, we used a coupling medium to ensure good acoustic contact between the transducer and the aluminum case. For real-world applications, it is desirable to design a fixture that can maintain consistent coupling over the lifetime of the battery, perhaps using dry coupling or a spring-loaded transducer. Another issue is the temperature dependence of the ultrasonic transducers themselves. In the range of interest (20–45 °C), the effect was small, but for higher ambient temperatures, compensation may be required.
Nevertheless, the current work demonstrates that the combination of ultrasonic transmission sensing and advanced pattern recognition can significantly enhance the safety of LFP-based energy storage battery systems. The warning time of 30 minutes provides a valuable window for emergency response, far exceeding the capabilities of conventional gas sensors or voltage-based methods. We believe this technique can be integrated into battery management systems for grid-scale energy storage plants, improving the overall safety and reliability of the energy storage battery infrastructure.
Conclusion
In this paper, we presented a comprehensive experimental and algorithmic study on the use of ultrasonic detection for early warning of overcharge thermal runaway in a 20 Ah prismatic aluminum-shell lithium iron phosphate energy storage battery. The main conclusions are as follows:
(1) The ultrasonic signal intensity Urms is more sensitive to temperature and SOC than the ultrasonic time of flight. The normalized relative parameters RU and RT can effectively characterize the internal state changes of the battery. During normal charge-discharge cycling, RU follows the SOC trend, while RT is mainly governed by temperature. This means that a single ultrasonic measurement can provide two complementary indicators.
(2) During overcharge, the ultrasonic signal is progressively attenuated due to lithium dendrite growth, structural distortion of electrodes, and gas generation. The complete loss of the ultrasonic signal occurs significantly earlier than the activation of the safety vent under all tested C-rates (0.25 C, 0.5 C, and 0.75 C). Therefore, ultrasonic attenuation is a valid early indicator of internal anomalous states in the energy storage battery.
(3) By combining the ultrasonic features with voltage, temperature, H2 concentration, and surface strain in an improved Mahalanobis-Taguchi system, we developed a health index that can reliably discriminate between normal operation and overcharge. The algorithm successfully issued an overcharge warning at a SOC of 104.2%, approximately 30 minutes before the safety vent opened. This provides a sufficiently long lead time for practical interventions.
The findings of this work advance the understanding of the spatiotemporal evolution of thermal runaway in large-format energy storage battery cells and provide a technological foundation for integrating ultrasound-based sensors into future battery management systems. Further research will focus on improving the coupling robustness, exploring the influence of battery aging, and validating the algorithm on larger battery modules and stacks.