Real-Time Battery Health Monitoring Breakthrough for EVs
As the electric vehicle (EV) industry accelerates toward mainstream adoption, one of the most pressing challenges remains the accurate, real-time assessment of battery health. The lithium-ion battery pack is the heart of any EV, and its performance directly influences driving range, safety, and longevity. Yet, monitoring the State of Health (SOH) of these batteries during daily operation has long been a technical hurdle—especially when relying on incomplete or fragmentary charging data. Now, a new algorithm developed by researchers at Harbin University of Science and Technology is poised to transform how battery health is estimated in real-world conditions.
In a study published in the Transactions of China Electrotechnical Society, Dr. Xianhua Song and her graduate student Quanzheng Yao have introduced a novel method capable of estimating battery SOH with unprecedented accuracy using only partial charging cycles—data commonly recorded during everyday EV use. This advancement could significantly enhance battery management systems (BMS), enabling more reliable range predictions, optimized charging strategies, and extended battery life.
The research addresses a critical limitation in current battery diagnostics: most SOH estimation techniques require full charge-discharge cycles to generate reliable data. In practice, however, EV owners rarely deplete their batteries completely or charge them from empty to full. Instead, they engage in “segment charging”—topping up from 30% to 80%, or 50% to 100%, depending on daily needs. This behavior creates a gap in usable data, making traditional SOH models less effective in real-time applications.
To overcome this challenge, Song and Yao developed a hybrid machine learning and filtering framework named DEKF-WNN-WLSTM. The acronym stands for Dual Extended Kalman Filter-Wavelet Neural Network-Wavelet Long Short-Term Memory Neural Network. While the name may sound complex, the underlying principle is elegantly designed to extract maximum insight from minimal input.
At its core, the model leverages the strengths of multiple advanced computational techniques. It begins with the Extended Kalman Filter (EKF), a well-established algorithm used in nonlinear system estimation. The EKF is particularly effective in tracking dynamic systems like batteries, where internal states change over time due to aging, temperature, and usage patterns. However, traditional EKF-based methods often suffer from error accumulation and require frequent recalibration using full-cycle data.
Song and Yao’s innovation lies in enhancing the EKF with two specialized neural networks: a Wavelet Neural Network (WNN) and a Wavelet Long Short-Term Memory (WLSTM) network. These networks are not standard deep learning models; they are engineered to handle the unique characteristics of battery data, which often contains noise, nonlinearity, and time-dependent degradation patterns.
The WNN brings robustness and high-resolution feature extraction. By replacing conventional activation functions with wavelet functions—mathematical tools adept at analyzing signals across different scales—the network can detect subtle changes in voltage and current profiles that indicate battery aging. This is particularly useful during partial charging, where small variations in the charging curve can reveal significant information about capacity loss.
The WLSTM component, on the other hand, excels at capturing long-term dependencies in sequential data. Unlike standard neural networks that may “forget” earlier inputs, the WLSTM retains critical historical information, allowing it to model the gradual degradation of a battery over hundreds of cycles. By integrating wavelet functions into the LSTM’s gates, the model achieves a finer-grained understanding of temporal dynamics, improving its predictive accuracy.
What sets this research apart is the dual-filter architecture. The first EKF estimates the full charging time based on fragmentary data, while the second EKF continuously corrects the estimation error by learning from past discrepancies. This self-correcting mechanism prevents the drift that typically plagues long-term predictions. By training the WNN-WLSTM networks on one complete charging cycle and three partial cycles, the model learns to generalize across different usage patterns without requiring repeated full charges.
The team tested their algorithm using real-world data from ternary lithium-ion batteries provided by Shenzhen Neware Electronic Co., Ltd. The dataset included over 180 charge-discharge cycles under constant current conditions, simulating typical EV usage. The results were striking: the average relative error in full charging time estimation was just 1.01%, significantly lower than the 1.76% achieved by a previously published EKF-GPR method. More importantly, the error remained stable over time, demonstrating the model’s ability to resist cumulative inaccuracies.
For EV manufacturers and fleet operators, such precision has far-reaching implications. Accurate SOH estimation enables better battery warranty management, more efficient second-life applications for retired EV batteries, and improved resale value assessments. It also enhances driver confidence by providing realistic range estimates that account for actual battery degradation, not just idealized models.
From a technological standpoint, the DEKF-WNN-WLSTM model represents a shift toward data-driven, adaptive battery diagnostics. Unlike physics-based models that rely on detailed electrochemical parameters—often difficult to measure and prone to variation—the new method learns directly from operational data. This makes it more scalable and easier to deploy across different battery chemistries and vehicle platforms.
Moreover, the model’s reliance on everyday charging fragments means it can operate continuously in the background, updating SOH estimates with every charge. This real-time capability is essential for advanced BMS features such as adaptive charging, where the system adjusts charging speed and voltage based on the battery’s current health to minimize stress and prolong lifespan.
The implications extend beyond individual vehicles. As smart grids and vehicle-to-grid (V2G) technologies gain traction, accurate battery health data becomes crucial for grid stability and energy market participation. Utilities and aggregators need reliable information about the available capacity and degradation state of EV batteries before enrolling them in demand response programs. A model like DEKF-WNN-WLSTM could provide that transparency, enabling safer and more efficient integration of EVs into the energy ecosystem.
Another advantage of the method is its computational efficiency. Despite its sophistication, the model is designed to run on embedded systems with limited processing power. The researchers used MATLAB and Python with TensorFlow/Keras to implement the neural networks, interfacing them with the Kalman filter routines. This modular approach allows for integration into existing BMS architectures without requiring a complete hardware overhaul.
Security and reliability are also enhanced. By reducing the need for periodic full discharge cycles—which can accelerate battery wear—the model promotes safer, more sustainable operation. It also minimizes user intervention; drivers no longer need to perform special calibration routines to maintain accurate SOH readings.
Looking ahead, the research opens several avenues for future development. One direction is the extension of the model to other battery types, such as lithium iron phosphate (LFP) or solid-state batteries, which have different degradation characteristics. Another is the incorporation of additional sensor data, such as temperature and internal resistance, to further refine the estimates.
There is also potential for cloud-based deployment, where anonymized battery data from thousands of vehicles could be aggregated to train even more robust models. Such a system could detect early signs of manufacturing defects or environmental stress factors, enabling proactive recalls or service campaigns.
From a policy perspective, accurate SOH monitoring supports the transition to a circular economy in the automotive sector. As governments push for higher EV adoption and stricter emissions regulations, ensuring battery longevity and recyclability becomes paramount. Transparent health tracking empowers consumers, facilitates battery leasing models, and strengthens the business case for battery swapping infrastructure.
The work of Song and Yao also highlights the growing importance of interdisciplinary research in advancing EV technology. Their approach combines signal processing, control theory, and deep learning—fields that are increasingly converging to solve complex engineering problems. It underscores the role of academic institutions in driving innovation that complements industry-led development.
For Harbin University of Science and Technology, this study reinforces its position as a leader in intelligent monitoring and machine learning applications. Dr. Song’s expertise in quantum computing and image security may seem distant from battery diagnostics, but her background in complex system modeling provides a unique perspective that bridges domains.
In the broader context of global decarbonization, every improvement in battery efficiency and lifespan contributes to reducing the environmental footprint of transportation. By extending the usable life of batteries and optimizing their performance, technologies like DEKF-WNN-WLSTM help maximize the return on the energy and resources invested in their production.
As EVs become more integrated into daily life, the demand for smarter, more responsive battery systems will only grow. The ability to understand a battery’s health not just in the lab, but in the real world, under real driving conditions, is no longer a luxury—it is a necessity. Song and Yao’s research delivers a powerful tool toward that goal, setting a new benchmark for real-time battery diagnostics.
The automotive industry stands at a crossroads, where software and artificial intelligence are becoming as critical as mechanical engineering. Battery management is no longer just about preventing overcharge or thermal runaway; it is about predictive intelligence, adaptive control, and lifelong optimization. The DEKF-WNN-WLSTM model exemplifies this shift, transforming raw charging data into actionable insights that benefit drivers, manufacturers, and the planet.
In conclusion, the breakthrough by Xianhua Song and Quanzheng Yao offers a practical, accurate, and scalable solution to one of the most persistent challenges in electric mobility. By harnessing the power of hybrid machine learning and advanced filtering, their method enables real-time SOH estimation using only the partial charging data generated in everyday use. This eliminates the need for disruptive full-cycle tests and reduces reliance on complex electrochemical models. The result is a more reliable, user-friendly, and sustainable approach to battery health monitoring—one that could soon become standard in next-generation electric vehicles.
Published in Transactions of China Electrotechnical Society, DOI: 10.19595/j.cnki.1000-6753.tces.222241