Battery internal temperature monitoring method and system with multi-feature decoupling and deep learning

Abstract

The application is suitable for the technical field of batteries, and provides a battery internal temperature monitoring method and system based on multi-feature decoupling and deep learning. The method comprises the following steps: obtaining a multi-dimensional acoustic feature set of an ultrasonic detection battery, wherein the multi-dimensional acoustic feature set comprises a time of flight, a surface wave amplitude and a bottom wave phase shift; performing decoupling processing on the multi-dimensional acoustic feature set to eliminate the coupling interference of SOC on the acoustic feature, to obtain a decoupled acoustic feature; and inputting the decoupled acoustic feature into a pre-trained deep learning model to output the internal temperature of the battery. The application eliminates the coupling interference of non-temperature factors such as SOC on the original ultrasonic feature, and realizes high-precision estimation of the internal temperature.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of batteries, and particularly relates to a battery internal temperature monitoring method and system based on multi-feature decoupling and deep learning. BACKGROUND

[0002] For lithium-ion batteries as core energy storage devices in modern society, accurately obtaining the internal core temperature is the top priority to ensure efficient, long-lasting and safe operation. The internal temperature is not only the key basis for the battery management system (BMS) to implement effective thermal management and avoid performance degradation caused by overheating or overcooling, but also the core index for early warning of thermal runaway risk under internal short circuit and other misuse conditions. In dangerous working conditions, the abnormal temperature rise in the core area of the battery occurs much earlier than the surface temperature change, and timely capturing this signal is the key to preventing thermal runaway and avoiding combustion and explosion. Therefore, accurately obtaining and real-time tracking the internal temperature change of the battery is of great significance to developing an efficient thermal management strategy, identifying abnormal temperature rise in advance, and ensuring the safety and reliability of the battery throughout its life cycle.

[0003] Current methods for monitoring the internal temperature of lithium-ion batteries mainly include temperature sensors, battery thermal models, and electrochemical impedance spectroscopy (EIS). Among them, surface sensors are difficult to accurately reflect the core temperature due to measurement lag and location limitations; while implantable sensors such as optical fibers and thin films are more accurate, but their invasive implantation method damages the integrity of the battery, and the process compatibility and long-term durability are insufficient, resulting in poor universality, and most of them are still in the experimental stage. The battery thermal model method based on heat transfer mechanism highly depends on complex parameter calibration, and has poor robustness under varying working conditions. The electrochemical impedance spectroscopy method is limited in practicality due to the tedious test process and interference from noise and aging state.

[0004] Currently, the research on monitoring the internal temperature of ultrasonic batteries mainly focuses on the surface temperature, and the ultrasonic features used generally rely only on a single ultrasonic feature for temperature monitoring. There is also a core bottleneck problem that limits the application accuracy of ultrasonic in real dynamic working conditions: the strong nonlinear coupling interference of battery temperature and state of charge (SOC) on ultrasonic features. In actual work, the temperature and SOC of the battery are always in dynamic change, and the influence of the two on ultrasonic acoustic features is intertwined and highly coupled, making it difficult to accurately distinguish their respective contributions.

[0005] In summary, there is an urgent need for an economical, efficient and non-destructive monitoring technology to achieve accurate monitoring of the internal temperature. SUMMARY

[0006] The purpose of the embodiments of the present application is to provide a battery internal temperature monitoring method based on multi-feature decoupling and deep learning, aiming to solve the limitations of existing battery temperature monitoring.

[0007] The battery internal temperature monitoring method based on multi-feature decoupling and deep learning is implemented as follows:

[0008] A multi-dimensional acoustic feature set of the battery is acquired by ultrasonic detection, and the multi-dimensional acoustic feature set includes a time of flight, a surface wave amplitude, and a bottom wave phase shift;

[0009] The multi-dimensional acoustic feature set is subjected to decoupling processing to eliminate the coupling interference of the SOC on the acoustic feature, and a decoupled acoustic feature is obtained.

[0010] The decoupled acoustic feature is input into a pre-trained deep learning model, and a battery internal temperature is output.

[0011] Further, the decoupling processing includes:

[0012] Based on static calibration experimental data, an FNN decoupling model is constructed, and the multi-dimensional acoustic feature set is input into the FNN decoupling model.

[0013] An estimation reference of the time of flight is estimated, and a residual error between a measured value and the estimation reference is calculated as a new decoupled feature.

[0014] A SOC threshold is set, and when the SOC threshold is exceeded, a residual error of the bottom wave phase shift is calculated, and when the SOC threshold is not exceeded, the original signal is retained.

[0015] The untreated surface wave amplitude, the full-range decoupled time of flight residual error, and the conditionally decoupled bottom wave phase shift are combined to obtain the decoupled acoustic feature.

[0016] Further, the input of the FNN decoupling model includes the SOC, the battery surface temperature, the SOC change rate, and the surface temperature change rate.

[0017] Further, the deep learning model includes:

[0018] A CNN feature extraction module is configured to extract local transient features from an input sequence.

[0019] A BiLSTM sequence modeling module is configured to capture long-term evolution trends and hysteresis effects of battery thermal behavior from both forward and reverse time sequences.

[0020] An attention mechanism module is configured to weight and focus on an output sequence of the BiLSTM sequence modeling module to obtain a feature vector.

[0021] A fully connected layer is configured to integrate and nonlinearly map the feature vector to obtain the battery internal temperature.

[0022] Further, the input of the deep learning model further comprises a moving average of the decoupled ultrasonic feature, a first-order difference and a second-order difference of the decoupled ultrasonic feature, and a battery surface temperature.

[0023] Further, in the deep learning model training stage, a battery implanted with a thermocouple is used to collect the internal temperature true value, and the root mean square error and the mean absolute error are used as the indicators for model training and evaluation.

[0024] Another purpose of the embodiment of the present application is a battery internal temperature monitoring system based on multi-feature decoupling and deep learning, which comprises:

[0025] An acquisition module is configured to acquire a multi-dimensional acoustic feature set of a detected battery, wherein the multi-dimensional acoustic feature set comprises a time of flight, a surface wave amplitude, and a bottom wave phase shift;

[0026] An FNN decoupling model is configured to decouple the multi-dimensional acoustic feature set to eliminate the coupling interference of the SOC on the acoustic feature, and obtain a decoupled acoustic feature;

[0027] A deep learning model is configured to input the decoupled acoustic feature into a pre-trained deep learning model, and output a battery internal temperature.

[0028] Further, the decoupling process comprises:

[0029] Based on static calibration experimental data, an FNN decoupling model is constructed, and the multi-dimensional acoustic feature set is input into the FNN decoupling model;

[0030] An estimation benchmark for estimating the time of flight is calculated, and a residual error between the measured value and the estimation benchmark is calculated as a new decoupled feature;

[0031] An SOC threshold is set, and when the SOC exceeds the SOC threshold, the residual error of the bottom wave phase shift is calculated, and when the SOC does not exceed the SOC threshold, the original signal is retained;

[0032] The unprocessed surface wave amplitude, the full-range decoupled time of flight residual error, and the conditionally decoupled bottom wave phase shift are combined to obtain the decoupled acoustic feature.

[0033] Further, the deep learning model comprises:

[0034] A CNN feature extraction module is configured to extract local transient features from an input sequence;

[0035] A BiLSTM sequence modeling module is configured to capture long-term evolution trends and hysteresis effects of battery thermal behavior from both forward and reverse time sequence directions;

[0036] An attention mechanism module is configured to weight and focus on the output sequence of the BiLSTM sequence modeling module to obtain a feature vector.

[0037] a full connection layer, used for information integration and nonlinear mapping of the feature vector, to obtain the internal temperature of the battery.

[0038] The battery internal temperature monitoring method provided by the embodiment of the present application has the beneficial effects that the battery ultrasonic signal is acquired, a multi-dimensional acoustic feature set containing time of flight (TOF), surface wave amplitude (SWA) and bottom wave phase shift (BPS) is constructed to capture the global and local thermal physical state of the battery, the multi-dimensional acoustic feature set is decoupled, the nonlinear coupling interference of the state of charge (SOC) and environmental factors is stripped by using a partition and dynamic compensation strategy according to the sensitive difference of different acoustic features to the SOC, and pure temperature features are extracted, the decoupled features are input into a CNN-BiLSTM-Attention hybrid neural network model, local morphological features are extracted by a convolution unit, time sequence evolution rules are captured by a bidirectional long short-term memory network, and attention mechanism is used to focus on temperature mutation key frames, and finally the high-precision internal temperature of the battery is output. BRIEF DESCRIPTION OF DRAWINGS

[0039] Figure 1 The flowchart of the battery internal temperature monitoring method provided by the embodiment of the present application is provided.

[0040] Figure 2 The lithium manganate battery ultrasonic waveform schematic diagram provided by the embodiment of the present application is provided.

[0041] Figure 3 The ultrasonic frequency domain feature map provided by the embodiment of the present application is provided.

[0042] Figure 4 The ultrasonic time domain feature map provided by the embodiment of the present application is provided.

[0043] Figure 5 The schematic diagram of the correlation coefficient of each ultrasonic feature and temperature provided by the embodiment of the present application is provided.

[0044] Figure 6 The ultrasonic feature and internal temperature normalized comparison provided by the embodiment of the present application is provided.

[0045] Figure 7 The CNN-BiLSTM-Attention hybrid neural network model provided by the embodiment of the present application is provided.

[0046] Figure 8 The FNN decoupling model provided by the embodiment of the present application is provided.

[0047] Figure 9 The multi-working-condition estimation result provided by the embodiment of the present application is provided.

[0048] Figure 10 A comparison of undissociated and single-feature results provided in the embodiments of the present invention;

[0049] Figure 11 The internal temperature estimation results of the DST under simulated operating conditions provided in the embodiments of the present invention;

[0050] Figure 12 Comparison of ablation experiment results provided in the embodiments of the present invention; Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0052] In the first embodiment, such as Figure 1 As shown, a method for monitoring the internal temperature of a battery using multi-feature decoupling and deep learning is proposed. The method includes steps S1 to S3:

[0053] Step S1: Obtain the multidimensional acoustic feature set of the probe battery, which includes flight time, surface wave amplitude and bottom wave phase shift.

[0054] In this embodiment, it is necessary to first acquire the ultrasonic testing signal of the lithium-ion battery, and then extract a multi-dimensional acoustic feature set from the ultrasonic testing signal. Ultrasonic testing technology, as an advanced non-destructive evaluation method, is highly sensitive to changes in the physicochemical state of materials due to the propagation characteristics of sound waves in the medium, and can effectively monitor the state of lithium-ion batteries.

[0055] The tested battery is a lithium-ion battery, specifically a lithium manganese oxide (LMO) battery. Its internal structure mainly consists of functional layers: a positive electrode (lithium manganese oxide), a negative electrode (graphite), an electrolyte, and a separator. These layers are arranged sequentially using a precision stacking process to form a multilayer cell with a periodic arrangement. This structure not only affects the battery's electrochemical performance but also significantly influences the propagation characteristics of ultrasonic signals. The ultrasonic reflection waveform of this battery is shown below. Figure 2 As shown, using the traditional bottom echo detection method, a coupling agent is added to the middle area of ​​the bottom of the battery. Then, a finger is moved in this area, and the waveform in the figure is observed to change as the finger approaches and moves away, thus identifying the wave as a bottom echo. The time of flight (TOF) of the ultrasound is taken between the surface wave amplitude and the bottom wave amplitude.

[0056] The speed of sound in a medium is affected by temperature. In an ideal gas, the expression for the speed of sound is as follows:

[0057] ;

[0058] Where C is the speed of sound in a specific medium, K is the adiabatic index of the gas, R is the gas constant, M is the molecular weight of the gas, and T is the temperature. When the gas composition is determined, the corresponding gas constant Z is also determined.

[0059] The core theory behind ultrasonic testing of battery temperature lies in the "acoustoelastic effect," which states that the acoustic properties of a material (such as wave velocity and attenuation) change with its physical state (such as temperature and stress). For lithium-ion batteries, temperature changes alter the physical properties of their internal components, thus affecting the propagation of ultrasonic waves. The propagation speed *c* of sound waves within a material is the most critical parameter for temperature monitoring. When sound waves propagate through a material, their propagation speed depends on the material's bulk modulus (K), shear modulus (G), and density (ρ), as shown in the equation:

[0060] ;

[0061] Meanwhile, acoustic impedance is related to the density and Young's modulus (E) of the material through which the wave passes. When an acoustic signal reaches the interface between two materials with different acoustic impedances, part of the signal is reflected and part is transmitted.

[0062] ;

[0063] When the internal temperature of a battery changes, the density and impedance of its internal structures will be affected by the temperature change, which will cause the sound wave velocity to change, and the change in acoustic impedance will cause the sound wave signal to attenuate.

[0064] Therefore, changes in the internal temperature of a battery directly lead to alterations in the elastic modulus and density of its constituent materials (such as electrode active materials, separators, and electrolytes). These changes in physical properties, in turn, systematically affect the propagation characteristics of ultrasound, specifically manifesting as significant changes in key acoustic parameters such as propagation speed, acoustic impedance, and signal attenuation coefficient. Based on this principle, by externally and precisely measuring characteristic parameters such as the flight time (corresponding to the speed of sound) and signal attenuation of ultrasound waves within the battery, and establishing a quantitative calibration relationship between these parameters and temperature, non-destructive and high-precision monitoring of the internal temperature field of the battery can be achieved.

[0065] Therefore, the following explanation is needed regarding the relationship between the battery's internal temperature and three acoustic characteristics: Time of Flight (TOF), Surface Acoustic Wave (SWA), and Bottom Wave Phase Shift (BPS).

[0066] Bottom wave phase shift (BPS), as a phase-sensitive parameter, reflects subtle changes in the propagation speed of sound waves within battery materials. These changes are closely related to the temperature effect, material elastic modulus, and acoustic impedance during charge-discharge cycles. The instantaneous phase is extracted using Hilbert transform, and the phase difference is calculated. Then, the average value is applied to obtain the phase shift characteristics of a single ultrasonic A-wave. The instantaneous phase difference sequence is first obtained, and then its average value is taken as the phase shift characteristic. The specific formula is as follows:

[0067] ;

[0068] In the formula Let i be the instantaneous phase of the i-th A wave. Let i be the phase of the i-th wave A. The phase of the reference wave is given, and N is the number of sampling points in the bottom wave region. The final phase shift characteristic is given by t, where t is time. The unwrap() function is used to expand the phase angle in radians of a vector or matrix to eliminate phase jumps.

[0069] Phase shift features directly reflect local phase delay changes, are sensitive to temperature variations, and also exhibit good smoothness and noise resistance. Time-of-flight (TOF) reflects the overall propagation time of a wave within a battery, while phase shift captures local phase modulation information during propagation. Combining phase shift features with TOF features creates a complementary feature set. TOF provides a global physical quantity reference, while phase shift provides locally sensitive information, holding great potential for temperature monitoring.

[0070] Figure 3 and Figure 4 By extracting the ultrasonic time-domain and frequency-domain features, it can be seen that the battery experienced a process of temperature first decreasing and then increasing. These features show that the ultrasonic features are affected by temperature changes, and some feature change trends even have a high degree of fit with temperature changes.

[0071] In the battery temperature rise experiment, the battery underwent a heating and cooling process. Pearson correlation analysis was performed on 36 ultrasonic features extracted during this process to identify ultrasonic features highly correlated with internal temperature changes. This step helps eliminate features with little explanatory power for the target variable and extracts features suitable for estimating the internal temperature.

[0072] The results of Pearson correlation coefficient calculation are as follows: Figure 5As shown in the figure, (a) is the correlation coefficient graph of the top 10 ultrasonic features with the highest correlation coefficient with the internal temperature in the temperature change experiment without charge / discharge, and (b) is the result obtained in the charge / discharge experiment at room temperature. The results show that under both conditions with and without charge / discharge, the three ultrasonic features, SWA, TOF, and BPS, all have a strong correlation with the internal temperature of the battery above 0.85, showing sensitivity to temperature. Further normalization comparison verification shows that SWA and TOF are highly fitted to the trend of internal temperature change, while the trend of BPS is opposite.

[0073] During the battery heating process, such as Figure 6 As shown, the increase in temperature and the time-of-flight (TOF) exhibit a stable synchronous growth relationship, a phenomenon mainly influenced by multiple physical factors. First, with rising temperature, the electrode materials, separator, and electrolyte undergo thermal expansion, leading to a decrease in material density. This decrease in density slows down the propagation speed of ultrasound waves, thus increasing TOF. Second, the elastic modulus of the battery's internal materials adjusts with temperature changes, particularly the mechanical properties of the electrodes and separator, which may alter the propagation characteristics of ultrasound waves. Furthermore, the viscosity of the electrolyte typically decreases with increasing temperature. As the primary medium for ultrasound propagation within the battery, the decrease in electrolyte viscosity alters the propagation efficiency of ultrasound waves, further affecting TOF. Finally, the battery may expand during heating, especially the volume expansion of the electrodes, which increases the ultrasound propagation path, thereby prolonging TOF. These factors combined result in a synchronous increase in TOF with rising temperature, and a synchronous decrease during the cooling phase. SWA exhibits the same trend. As temperature increases, the mechanical damping and energy dissipation of the battery's internal materials decrease, reducing the attenuation of ultrasonic waves during propagation and thus enhancing the amplitude of surface waves. During cooling, material damping increases, leading to greater sound wave energy attenuation and a decrease in amplitude. In contrast, BPS decreases with increasing temperature. The physical mechanism is that rising temperature causes thermal expansion, decreased density, and changes in the local elastic modulus of the battery's internal materials, slowing ultrasonic wave propagation and increasing local phase delay. Extracting the instantaneous phase using Hilbert transform and calculating the mean phase difference reflects a decrease in the phase shift characteristic value. During cooling, material contraction and increased propagation speed reduce local phase delay, resulting in an increase in the phase shift characteristic value. The combination of these three factors forms a complementary set of physical characteristics, providing comprehensive and reliable support for internal temperature monitoring.

[0074] Therefore, step S1 mainly focuses on analyzing the three ultrasound features used:

[0075] 1. Surface wave amplitude (SWA) exhibits a strong, nearly linear positive correlation with temperature. Most importantly, the SWA response curves almost completely overlap across all tested SOC levels. This indicates that SWA is a quasi-ideal thermosensitive characteristic, with its signal response almost entirely dominated by temperature within the studied temperature range, demonstrating extremely high robustness to SOC.

[0076] 2. Ultrasonic time-of-flight (TOF) also shows a strong positive correlation with temperature. However, the TOF curves representing different states of charge (SOC) exhibit clear vertical stratification; that is, on the same isotherm, the lower the SOC, the higher the TOF value. This indicates that the TOF signal contains explicit SOC information, and if it is not corrected, directly using it for temperature estimation will introduce systematic bias.

[0077] 3. The dependence of BPS on SOC is the most profound and complex. Its response exhibits obvious interval dependence: in the low to medium SOC range (0-50%), the BPS curve is tightly clustered and insensitive to changes in SOC; however, when SOC enters the high range (75-100%), not only does the absolute value of BPS undergo a sharp jump from positive to negative, forming a "bifurcation" in the feature space, but its sensitivity to changes in SOC also increases significantly.

[0078] The above results explain the response ambiguity observed in dynamic operating conditions and provide a solid physical basis for establishing an effective decoupling mechanism. The strong dependence of BPS and TOF characteristics on SOC is the root cause of the "different characteristics at the same temperature" phenomenon in dynamic cycling.

[0079] Step S2: Decouple the multidimensional acoustic feature set to eliminate the coupling interference of SOC on the acoustic features and obtain the decoupled acoustic features.

[0080] In this embodiment, features such as TOF, SWA, and BPS change synchronously with the internal temperature, but the relationship between them is not a simple mapping. A key finding is that during the charging and discharging phases, even when the internal temperatures are similar, the corresponding combinations of ultrasonic feature values ​​are drastically different. This reveals a significant coupling effect of the battery's operating state on the ultrasonic signal. The root cause of this response ambiguity is that the ultrasonic signal under dynamic operating conditions is a coupled output of the battery's internal thermophysical state (characterized by temperature T) and electrochemical state (characterized by SOC). Throughout the cycle, SOC undergoes a complete 0-100%-0 cycle, and its influence on the acoustic properties of electrode materials, electrolytes, and other media is intertwined with the temperature effect. This means that any attempt to directly infer the temperature from the raw ultrasonic signal will be affected to some extent. Therefore, to construct an accurate and robust temperature estimation model, it is necessary to first decouple these two mutually coupled physical effects.

[0081] Step S3: Input the decoupled acoustic features into the pre-trained deep learning model and output the internal temperature of the battery.

[0082] To accurately estimate the internal temperature of lithium batteries, this study first utilizes a feedforward neural network (FNN) decoupling model to eliminate the coupling interference of non-temperature factors such as state of charge (SOC) on the original ultrasonic features, thereby extracting pure features more relevant to temperature changes. Subsequently, the decoupled features are input into a deep estimation model fusing a convolutional neural network (CNN), a bidirectional long short-term memory network (BiLSTM), and an attention mechanism, ultimately achieving high-precision estimation of the internal temperature. The estimation model establishment process is as follows: Figure 7 As shown.

[0083] The specific structure and data flow of the model are as follows:

[0084] 1. Input Layer: Receives a fixed-length time series window as input. Each time step of the sequence consists of multiple features, including: decoupled ultrasonic features, their moving average, first and second order differences, and battery surface temperature.

[0085] 2. CNN Feature Extraction Module: The input sequence first passes through a one-dimensional convolutional layer. The CNN acts like a sliding filter here, automatically extracting key local patterns and shape features from the temporal data.

[0086] 3. BiLSTM Sequence Modeling Module: The deep feature maps extracted by CNN are fed into the BiLSTM layer. BiLSTM processes the sequence in parallel, enabling it to capture both past (historical information) and future (trend information) context at any given time point.

[0087] 4. Attention Mechanism Module: This is the key component connecting the BiLSTM and the final output. After processing the sequence, the BiLSTM outputs the hidden states at each time step. The attention mechanism learns a weight distribution, assigning different importance scores to these hidden states. For the temperature monitoring task, this means the model can adaptively focus on those moments in the sequence that have the greatest impact on the current temperature (e.g., the instant when the current changes abruptly or the temperature begins to rise rapidly), and give them higher weights. By weighted summing of all hidden states, a highly condensed "context vector" containing key information is generated.

[0088] 5. Output module: The "context vector" generated by the attention mechanism is fed into a fully connected layer for information integration and nonlinear mapping, and finally outputs an accurate estimate of the internal temperature at the current moment through a regression layer.

[0089] This estimation model is a deep learning model that integrates CNN, BiLSTM, and an attention mechanism. First, the model uses a CNN to extract key local and transient patterns from the decoupled feature sequence of the input. Then, a BiLSTM network performs deep temporal modeling on these features; its bidirectional structure can comprehensively capture the long-term evolution trend and hysteresis effects of battery thermal behavior. Finally, the attention mechanism dynamically weights the output sequence of the BiLSTM, enabling the model to focus on the most influential time steps, thereby generating a highly condensed feature vector, which outputs a high-precision estimation result through a fully connected layer.

[0090] In this embodiment, the multidimensional acoustic feature set also includes time-domain and frequency-domain features. In the experiment, a second-order Butterworth bandpass filter with a lower frequency limit of 0.84 MHz and an upper frequency limit of 1.65 MHz was used to denoise the original data to eliminate environmental noise caused by the temperature chamber, testing equipment, etc., during the experiment. To fully extract the features of the acquired ultrasonic signal, Hilbert and Fourier transforms were performed on the signal in the bass oscillation region, thereby obtaining its multidimensional feature set in the time and frequency domains:

[0091] (1) Time-Domain Features (TD): including mean (TDM), standard deviation (TDSTD), root mean square (TDRMS²), absolute mean (TDMAV), skewness (TDS), kurtosis (TDK), variance (TDV), maximum value (TDMax), minimum value (TDMin), peak difference (TDPP), integral of the square of the signal (TDEI), signal weighted average rate (TDWC), ratio of standard deviation to absolute mean (TDSAVR), ratio of maximum value to standard deviation (TDMSR), ratio of maximum value to absolute mean (TDMAR), ratio of maximum value to root mean square (TDMRR), higher-order features of skewness (TDHSK), ratio of kurtosis to standard deviation (TDKSR), and ratio of higher-order moments to energy (TDHMER).

[0092] (2) Frequency-Domain Features (FD): including mean (FDM), variance (FDV), skewness (FDS), kurtosis (FDK), frequency mean (FDAF), frequency standard deviation (FDSD), root mean square frequency (FDRMS), root mean square frequency (FDSR), kurtosis factor (FDCF), ratio of frequency standard deviation to mean (FDFSDR), frequency skewness (FDFS), frequency kurtosis (FDFK), frequency uniformity (FDEQ), and frequency weighted average rate (FDFWM).

[0093] (3) Other characteristics: flight time, surface wave amplitude and bottom wave phase shift.

[0094] In this embodiment, the decoupling process in step S2 includes:

[0095] Based on static calibration experimental data, an FNN decoupling model is constructed, and the multidimensional acoustic feature set is input into the FNN decoupling model.

[0096] An estimation benchmark for the estimated flight time is used, and the residual between the measured value and the estimation benchmark is calculated as a new feature after decoupling.

[0097] Set a SOC threshold. If the SOC threshold is exceeded, calculate the residual of the bottom wave phase shift. If the SOC threshold is not exceeded, retain the original signal.

[0098] The unprocessed surface wave amplitude, the time-of-flight residual after full decoupling, and the bottom wave phase shift after conditional decoupling are combined to obtain the decoupled acoustic characteristics.

[0099] In this embodiment, to effectively address the coupling challenge between ultrasonic features and SOC in battery internal temperature estimation, this study proposes and constructs a feature-independent and conditional decoupling mechanism. A targeted, physics-driven signal processing strategy is employed. The FNN decoupling model is constructed as follows: Figure 8As shown, the decoupling model is constructed based on comprehensive static calibration experimental data, covering the entire SOC range from 0% to 100%. This embodiment trains two independent FNNs as decoupling sub-models for SOC-sensitive features—TOF and BPS. Unlike traditional methods, the key improvement of this model lies in the introduction of dynamic features. The input to the FNN sub-model is a four-dimensional vector: [SOC, T_surface, dSOC / dt, dTemp / dt], representing the state of charge, surface temperature, rate of change of SOC, and rate of change of surface temperature. This allows the decoupling model to better capture and compensate for signal changes during dynamic charging and discharging processes. When processing dynamic operating condition data, this mechanism implements decoupling in the following way: For the TOF feature, which exhibits systematic dependence across the entire SOC range, the baseline value is estimated using the decoupling model, and then the residual between the measured value and the estimated baseline is calculated as the new feature after decoupling. For the BPS feature, which is sensitive only in the high SOC range, a conditional decoupling strategy is implemented. The decoupled model is activated to calculate its residuals only when the real-time SOC exceeds a preset threshold (SOC>60%); below this threshold, the original signal is retained. Finally, the unprocessed SWA, the fully decoupled TOF residuals, and the conditionally decoupled BPS signal / residuals are combined into a physically purified, more information-pure feature vector. This vector provides high-quality input for the subsequent CNN-BiLSTM deep learning model, enabling it to focus more on learning the complex temporal relationship between internal temperature and core thermal dynamics.

[0100] It is important to emphasize that this embodiment proposes and constructs a feature-independent and conditional decoupling mechanism based on the non-uniform coupling characteristics of different ultrasonic features (SWA, TOF, BPS) to the state of charge (SOC). A targeted, physics-driven feature decoupling strategy is employed: for the thermally robust SWA feature, it is directly retained as a temperature reference; for the TOF feature, which is affected by SOC interference across the entire domain, background removal decoupling is implemented throughout the entire cycle; and for the BPS feature, which only undergoes nonlinear jumps in the high SOC range, a conditionally triggered decoupling logic is designed. Furthermore, the dynamic rate of change of SOC and temperature is introduced as a compensation factor. Through reference prediction and residual extraction, high-fidelity separation of temperature-sensitive components is achieved at the physical level. This scheme effectively solves the multi-physics coupling problem under dynamic operating conditions, significantly improving the robustness and accuracy of temperature monitoring.

[0101] Specifically, the uniqueness and significant effect of the physical decoupling described in this embodiment can be represented by the following three aspects:

[0102] 1. Compared to treating all ultrasound features as homogeneous inputs or simply performing basic normalization, this method reveals fundamental differences in the response mechanisms of different acoustic features (TOF, SWA, BPS) to SOC and temperature. In particular, it discovers that the dependence of BPS features on SOC exhibits significant "range characteristics" (sensitive only in the high SOC range), while SWA remains almost unaffected.

[0103] 2. Instead of denoising all features indiscriminately, a "divide and conquer" approach was designed. For Time-of-Flight (TOF), due to systematic bias, global decoupling was implemented. For BPS, conditional decoupling was implemented, triggering residual calculation only when SOC > 60%, preserving the original physical information in the low SOC range, and avoiding noise introduction caused by over-processing. For SWA, it was directly retained as a pure thermal reference.

[0104] 3. The decoupling model is not just a static mapping, but introduces dynamic features as input to the FNN, which solves the hysteresis or nonlinear error caused by transient changes during dynamic charging and discharging.

[0105] In this embodiment, the Bidirectional Long Short-Term Memory (BiLSTM) network in step S3 is specifically as follows. BiLSTM is a variant of LSTM that captures more comprehensive temporal dependencies by simultaneously processing forward and backward time-series data. BiLSTM consists of two independent LSTM layers, a forward layer and a backward layer, effectively utilizing past and future information to understand the data at the current moment. This structure is particularly important for tasks requiring complete contextual information, such as natural language processing and time-series data estimation and prediction. In BiLSTM, the forward LSTM processes data sequentially from front to back, and its hidden state computation and updates follow the same gating mechanism as traditional LSTM. The backward LSTM layer processes data in reverse temporal order from back to front, and its gating mechanism and state updates are similar to the forward layer, but its input is the reverse order of the time series.

[0106] The final output of BiLSTM is obtained by concatenating the hidden states of the forward and backward layers at each time step. In this way, the model can simultaneously utilize information from time t-1 and earlier (from the forward layer) and information from time t+1 and later (from the backward layer), thus capturing richer contextual information. This bidirectional structure significantly enhances the model's ability to handle complex time-series data while balancing model complexity and computational cost; for example, in ultrasound data analysis, it can better capture long-term trend changes in the data.

[0107] BiLSTM can capture bidirectional dependencies in sequences by combining the outputs of two LSTM networks, forward and backward. Therefore, it not only inherits the advantages of LSTM in processing time-series data but also possesses stronger robustness and generalization ability. Compared to traditional unidirectional LSTM, the bidirectional structure of BiLSTM effectively improves data utilization and overcomes the limitations of traditional LSTM in processing time-series data. Therefore, in the field of time-series data processing, BiLSTM is generally more efficient and intelligent. For temperature monitoring, BiLSTM, by simultaneously utilizing historical and future time information, can effectively capture the nonlinear evolution trend of battery temperature under multiple stages and operating conditions, and is particularly suitable for identifying dynamic features at critical moments such as temperature abrupt changes and stage transitions. Compared to traditional LSTM, BiLSTM performs better in modeling transition points, handling delayed heating processes, and short-term fluctuations. It has excellent multi-dimensional feature processing capabilities and can collaboratively model variables such as TOF, SOC, ΔSOC, and operating stages to uncover their complex coupling relationships with temperature changes.

[0108] In the second embodiment, the model of the first embodiment is validated.

[0109] The model was validated under six operating conditions (Cycle 1-Cycle 6) under which the battery normally operates. These conditions include constant temperature (25℃ and 35℃) 0.5C, room temperature (fluctuating between 30-35℃) 0.5C, constant temperature (25℃ and 35℃) 1C, and room temperature (fluctuating between 30-35℃) 1C, representing normal temperature, high temperature, and uncontrolled temperature environments, respectively. Furthermore, exploratory testing under low-temperature conditions revealed that implanting a thermocouple caused severe attenuation of the ultrasonic signal, making effective acquisition impossible. Although the battery's charging and discharging functions were unaffected, low-temperature conditions were not included in this study due to the inability to obtain effective ultrasonic data for temperature correlation analysis. All results presented and discussed are based on temperature ranges where high-quality ultrasonic signals could be obtained. The model estimation results are as follows: Figure 9 As shown.

[0110] The training results are primarily evaluated using the following two metrics: Root Mean Square Error (RMSE) and Mean Absolute Error (MAE), as shown in the following formula:

[0111] ;

[0112] ;

[0113] In the formula For the i-th estimate, Let be the i-th true value, and n and N be the sample sizes.

[0114] likeFigure 9 As shown, the model in this embodiment exhibits stable and excellent accuracy under six test conditions: RMSEs of 0.220℃, 0.277℃, 0.359℃, 0.237℃, 0.249℃, and 0.275℃ under the conditions of 25℃ 0.5C, room temperature 0.5C, room temperature 1C, 25℃ 1C, 35℃ 1C, and 0.5C, respectively; and MAEs of 0.176℃, 0.222℃, 0.308℃, 0.190℃, 0.200℃, and 0.223℃, respectively. The average RMSE for all conditions is 0.270℃, and the average MAE is 0.220℃. Larger deviations occur at a few locations, indicating that the method possesses high accuracy and strong robustness under different magnification and temperature conditions. A significant feature is that the model's estimation curve effectively filters out high-frequency noise in the real signal collected by the temperature sensor, accurately capturing the core trend of temperature change, and the overall estimated trajectory highly matches the actual temperature change. In summary, the model maintains robust estimation performance under different ambient temperatures and current ratios, fully verifying the strong robustness and excellent generalization ability of the proposed method. The average RMSE results of the undecoupled model and the single ultrasonic TOF estimation are 0.371℃ and 0.309℃, respectively, and the average MAE results are 0.308℃ and 0.251℃, respectively. Comparing the model results of this embodiment with the two models mentioned above, the results are as follows... Figure 10 As shown, the model in this embodiment improved RMSE by 27.2% and 28.6%, respectively, and MAE by 12.6% and 12.4%, respectively. This demonstrates that decoupling and multiple ultrasonic features play a significant role in improving the accuracy of battery internal temperature monitoring.

[0115] In addition to the above operating conditions, actual driving simulation (DST) tests were also conducted. One cycle (70 DST cycles) was used for training, and the test set size was the same as the training set. The results show that the model can still effectively and accurately estimate the internal battery temperature under simulated driving conditions, with an RMSE of 0.184℃ and a MAE of 0.144℃. Figure 11 As shown.

[0116] To further evaluate the contributions of each key component in the battery internal temperature estimation model framework proposed in the first embodiment and to verify the advantages of the decoupled model and multiple ultrasonic features, ablation experiments were conducted to verify the role of each part of the model. The complete model containing all design modules was used as the baseline, and four variant models were constructed by systematically removing individual key components. All models were evaluated under four different test conditions (the horizontal axis in the figure represents Cycle 1 to Cycle 6, corresponding to different temperature and rate conditions) to examine their accuracy and robustness. The comparison results of the root mean square error (RMSE) and mean absolute error (MAE) of the models are presented in the figures. Figure 12 middle.

[0117] Experimental results clearly demonstrate that the proposed baseline model exhibits optimal performance across all test configurations, with mean root mean square error (RMSE) and mean absolute error (MAE) of 0.270℃ and 0.220℃, respectively, validating the superiority of the overall design. Specifically, the hybrid architecture demonstrates significant advantages compared to single-structure models. The baseline model's accuracy is far superior to the standalone "LSTM-Only" model (RMSE=0.378℃) and "CNN-Only" model (RMSE=0.376℃), confirming that combining the local feature extraction capability of CNN with the long-term dependency modeling capability of BiLSTM is key to improving internal temperature estimation performance. Furthermore, the necessity of each component is strongly demonstrated. When the CNN module is removed ("No-CNN"), the model performance deteriorates most severely, with the RMSE soaring to 0.393℃, highlighting the core role of CNN in effectively extracting key features. Similarly, removing the attention mechanism ("No-Attention") also significantly increases the model error to 0.307℃, indicating that this mechanism is crucial for focusing important information and improving estimation accuracy. In summary, the results of this ablation study collectively validate the rationality of the model design in this embodiment, and each component made an indispensable contribution to the final high-precision monitoring.

[0118] To further verify the performance advantages of the model proposed in this embodiment, this embodiment selects several classic machine learning and deep learning models for comparison, including Support Vector Regression (SVR), Multilayer Perceptron (MLP), Gated Recurrent Unit (GRU), and Random Forest (RF). All models were evaluated on the same dataset, and their average errors are shown in Table 1.

[0119] Table 1: Comparison of estimation results from different models

[0120]

[0121] The results show that the model in this embodiment exhibits an overwhelming performance advantage, with root mean square error (RMSE) and mean absolute error (MAE) of 0.270℃ and 0.220℃, respectively, both achieving the best performance among the proposed methods. This advantage is particularly significant compared to traditional machine learning methods: compared to the worst-performing SVR (RMSE=0.683℃), the error of this embodiment is reduced by 60.5%; even compared to the best-performing Random Forest model (RMSE=0.356℃), the error is further reduced by 24.2%. In comparison with advanced deep learning models such as GRU (RMSE=0.359℃), the carefully designed hybrid architecture of this embodiment—utilizing CNN to extract local key features, BiLSTM to capture long-term bidirectional dependencies, and an Attention mechanism to dynamically focus on important information—demonstrates its unique superiority. In summary, these comparative results strongly demonstrate that the model proposed in this embodiment, by organically integrating the advantages of each component, can more profoundly capture the complex dynamic laws of temperature changes inside the battery, thus standing out among different models and achieving the highest monitoring accuracy and robustness.

[0122] In the third embodiment, a battery internal temperature monitoring system based on multi-feature decoupling and deep learning is provided. The system includes:

[0123] The acquisition module is used to acquire a multidimensional acoustic feature set of the ultrasonic detection battery, the multidimensional acoustic feature set including time of flight, surface wave amplitude and bottom wave phase shift;

[0124] The FNN decoupling model is used to decouple the multidimensional acoustic feature set to eliminate the coupling interference of SOC on the acoustic features and obtain the decoupled acoustic features.

[0125] A deep learning model is used to input the decoupled acoustic features into a pre-trained deep learning model and output the internal temperature of the battery.

[0126] The deep learning model includes:

[0127] The CNN feature extraction module is used to extract local transient features from the input sequence;

[0128] The BiLSTM sequence modeling module is used to capture the long-term evolution trend and hysteresis effect of battery thermal behavior from both positive and negative time series directions.

[0129] The attention mechanism module performs weighted focusing on the output sequence of the BiLSTM sequence modeling module to obtain the feature vector;

[0130] A fully connected layer is used to integrate information and perform nonlinear mapping on the feature vectors to obtain the internal temperature of the battery.

[0131] The decoupling process includes:

[0132] Based on static calibration experimental data, an FNN decoupling model is constructed, and the multidimensional acoustic feature set is input into the FNN decoupling model.

[0133] An estimation benchmark for the estimated flight time is used, and the residual between the measured value and the estimation benchmark is calculated as a new feature after decoupling.

[0134] Set a SOC threshold. If the SOC threshold is exceeded, calculate the residual of the bottom wave phase shift. If the SOC threshold is not exceeded, retain the original signal.

[0135] The unprocessed surface wave amplitude, the time-of-flight residual after full decoupling, and the bottom wave phase shift after conditional decoupling are combined to obtain the decoupled acoustic characteristics.

[0136] This embodiment develops a novel ultrasound-based framework that integrates multi-dimensional acoustic feature extraction and interference decoupling strategies to achieve accurate temperature estimation. This embodiment constructs comprehensive acoustic features from ultrasonic echoes, including time-of-flight (TOF), surface wave amplitude (SWA), and bottom wave phase shift (BPS), and employs a feature-selective decoupling method to separate nonlinear interferences related to the state of charge (SOC). The refined features are then processed by a CNN-BiLSTM-Attention network to capture time dependencies and achieve accurate temperature estimation. Experimental validation demonstrates that the proposed method exhibits superior estimation accuracy and robustness at various ambient temperatures and magnifications (C-rates), achieving a mean root mean square error (RMSE) of 0.270°C and a mean absolute error (MAE) of 0.220°C. Compared to the undecoupled model and the single-feature (TOF-only) model, this performance reduces the RMSE by 27.2% and 12.6%, respectively, thus validating the necessity of employing decoupling strategies and multi-feature integration. Furthermore, the model exhibits strong robustness under highly dynamic DST (Dynamic Stress Test) conditions, achieving an extremely low RMSE of 0.184°C.

[0137] This framework provides a practical approach for non-destructive, real-time monitoring of battery internal temperatures. While the method has been validated on specific cell types, further research is needed under different chemical systems and states of health (SOH) to ensure broader applicability. Future work will focus on expanding dataset diversity, improving model efficiency for real-time integration into battery management systems (BMS), and combining data-driven decoupling methods with electrochemical-thermal physics models to enhance estimation accuracy and physical interpretability.

[0138] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0139] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0140] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A battery internal temperature monitoring method based on multi-feature decoupling and deep learning, characterized in that, The method includes: A multidimensional acoustic feature set of an ultrasonic detection battery is obtained, the multidimensional acoustic feature set including time of flight, surface wave amplitude and bottom wave phase shift; The multidimensional acoustic feature set is decoupled to eliminate the coupling interference of SOC on the acoustic features, and the decoupled acoustic features are obtained. The decoupled acoustic features are input into a pre-trained deep learning model, which outputs the internal temperature of the battery. The deep learning model includes: The CNN feature extraction module is used to extract local transient features from the input sequence. The BiLSTM sequence modeling module is used to capture the long-term evolution trend and hysteresis effect of battery thermal behavior from both positive and negative time series directions. The attention mechanism module performs weighted focusing on the output sequence of the BiLSTM sequence modeling module to obtain the feature vector; A fully connected layer is used to integrate information and perform nonlinear mapping on the feature vectors to obtain the internal temperature of the battery; The decoupling process includes: Based on static calibration experimental data, an FNN decoupling model is constructed, and the multidimensional acoustic feature set is input into the FNN decoupling model. An estimation benchmark for the estimated flight time is used, and the residual between the measured value and the estimation benchmark is calculated as a new feature after decoupling. Set a SOC threshold. If the SOC threshold is exceeded, calculate the residual of the bottom wave phase shift. If the SOC threshold is not exceeded, retain the original signal. The unprocessed surface wave amplitude, the time-of-flight residual after full decoupling, and the bottom wave phase shift after conditional decoupling are combined to obtain the decoupled acoustic characteristics.

2. The battery internal temperature monitoring method based on multi-feature decoupling and deep learning according to claim 1, characterized in that, The inputs to the FNN decoupling model include SOC, battery surface temperature, SOC change rate, and surface temperature change rate.

3. The battery internal temperature monitoring method based on multi-feature decoupling and deep learning according to claim 1, characterized in that, The inputs to the deep learning model also include the moving average of the decoupled ultrasonic features, the first-order and second-order differences of the decoupled ultrasonic features, and the battery surface temperature.

4. The battery internal temperature monitoring method based on multi-feature decoupling and deep learning according to claim 1, characterized in that, During the training phase of the deep learning model, the true internal temperature is collected using a battery with an implanted thermocouple, and the root mean square error and mean absolute error are used as indicators for model training and evaluation.

5. A battery internal temperature monitoring system based on multi-feature decoupling and deep learning, characterized in that, The system includes: The acquisition module is used to acquire a multi-dimensional acoustic feature set of the ultrasonic detection battery, the multi-dimensional acoustic feature set including time of flight, surface wave amplitude and bottom wave phase shift; The FNN decoupling model is used to decouple the multidimensional acoustic feature set to eliminate the coupling interference of SOC on the acoustic features and obtain the decoupled acoustic features. A deep learning model is used to input the decoupled acoustic features into a pre-trained deep learning model and output the internal temperature of the battery. The deep learning model includes: The CNN feature extraction module is used to extract local transient features from the input sequence. The BiLSTM sequence modeling module is used to capture the long-term evolution trend and hysteresis effect of battery thermal behavior from both positive and negative time series directions. The attention mechanism module performs weighted focusing on the output sequence of the BiLSTM sequence modeling module to obtain the feature vector; A fully connected layer is used to integrate information and perform nonlinear mapping on the feature vectors to obtain the internal temperature of the battery; The decoupling process includes: Based on static calibration experimental data, an FNN decoupling model is constructed, and the multidimensional acoustic feature set is input into the FNN decoupling model. An estimation benchmark for the estimated flight time is used, and the residual between the measured value and the estimation benchmark is calculated as a new feature after decoupling. Set a SOC threshold. If the SOC threshold is exceeded, calculate the residual of the bottom wave phase shift. If the SOC threshold is not exceeded, retain the original signal. The unprocessed surface wave amplitude, the time-of-flight residual after full decoupling, and the bottom wave phase shift after conditional decoupling are combined to obtain the decoupled acoustic characteristics.

Images