Battery life estimation using machine learning techniques


Battery life estimation or state of health (SOH) estimation deals with accurately assessing the percentage value of the State of Charge (SoC), which is generally used as a metric to quantify the energy left in a battery compared to the energy it had when it was originally full.

When SoC equals 100%, it means the battery is now fully charged; and 0% SoC represents an empty battery. SoC also indicates that how long a battery will continue to perform before it needs recharging. State of Power (SoP) is the ratio of peak power, i.e., the maximum power that a battery can persistently provide.

An accurate SoC estimation plays a critical role in a reliable battery management system, especially in hybrid electric vehicles (HEVs) and electric vehicles (EVs) like autonomous cars and drones. Unlike the fuel level in traditional combustion engine vehicles, the SoC cannot be directly measured in these applications.

The batteries’ performance deteriorates with time and usage due to their electrochemical constituents’ degradation, resulting in capacity and power fade. This is called battery aging and is a consequence of multiple coupled aging mechanisms influenced by different factors such as battery chemistry and manufacturing and environmental and operating conditions.

When the battery fails to meet the energy or power requirement for its application, it is commonly defined as the end of life (EOL). To ensure batteries’ safety and reliability despite aging, it requires battery life or SoC estimation techniques and tools.

The key factors considerably affecting the battery life are as follows:

  • High Temperatures: Extremely high temperatures may trigger an ultimate threat named “thermal runaway,” accelerating SEI layer growth rates on the anode, faster LLI and cell resistance increase, metal dissolution from the cathode, and electrolyte decomposition.
  • Low Temperatures: Extremely low temperatures slow down the transport of Li ions in both electrodes and in the electrolyte. Attempts of fast charging at low temperatures can also create crowding of Li ions.
  • Over-charge/discharge: When a cell is overcharged, the cathode is over-delithiated (no active lithium available), and the anode is over-lithiated (no more ‘room’ for lithium). The cathode material suffers from irreversible structural change when overdelithiated, followed by the dissolution of transition metal ions (such as Mn2+) and active material decomposition.
  • High currents: Excessive charge and discharge currents can cause localized overcharge and discharge, leading to the degradation reactions as generalized overcharge and over-discharge. High currents come with more heat waste, raising the cell temperature and concomitantly the rates of aging processes.
  • Mechanical stresses: Batteries are subjected to stress from sources, such as manufacturing, electrode material expansion during operation, gas evolution in mechanically constrained cells, and external loading during service. When stress exceeds a certain limit, the electrode experiences material failure associated with cracking or fracture. This results in significant degradation of cell performance and capacity fade.

Battery life or SoC estimation techniques

SoC estimation techniques are used to track the actual performance of batteries in operation. The estimated value reflects the current capability of a battery to store and supply energy/power relative to that at the beginning of its life. Incremental capacity/differential voltage (IC/DV) analysis, differential thermal voltammetry (DTV), and differential mechanical parameter (DMP) analysis are the most frequently mentioned techniques.

1. IC/DV analysis

IC/DV analysis provides a non-destructive means of characterizing cells and has been widely used for aging mechanism identification. IC is calculated by
differentiating the battery capacity change from the terminal voltage change for a sufficient small-time interval. At the same time, DV is defined as the inverse of IC. The differentiation
transforms voltage plateaus in charge/discharge curves into clearly identifiable peaks in IC curves and valleys in DV curves.


  • Easy to monitor, only needs two parameters (voltage and capacity)
  • It can be applied to batteries with different types, sizes, and chemistries
  • Works for partial charging/discharging conditions
  • Easy to be implemented in BMS for online applications.


  • Limited to low current rates (< 1 C)
  • Sensitive to measurement noise – requires smoothed curves
  • Influenced by the operation temperature
  • Computing dV for chemistries with large voltage plateaus (e.g., LFP cells) might yield infinite solutions.

2. DTV analysis

DTV analysis is a complementary tool to existing SOH diagnostic techniques, which combines the concept of IC analysis with temperature measurements
to infer thermodynamic information about the electrode materials. The DTV technique probes the cell surface temperature during galvanostatic (dis)charges, and it is obtained by differentiating the temperature (T) concerning voltage (dT/dV) and plotting against cell voltage.


  • Easy, only needs two parameters (voltage and temperature)
  • It can be used for monitoring cells in parallel
  • Applicable for partial charging/discharging conditions
  • Easy for BMS implementation


  • Needs additional and calibrated temperature sensors
  • Sensitive to testing temperature variations
  • Challenges in overcoming noise in the temperature measurement.

3. DMP analysis

This method calculates the battery SOH estimation by understanding and modeling batteries’ mechanical behavior, such as the variation of cell level strain (ε) and stress. The intercalation/de-intercalation of Li-ions in/from the electrode active materials are associated with the volume change, expanding, and contracting in repeatable patterns. Mechanical stress in a cell evolves due to electrode expansion against a constraint normal to the plane of electrodes. It can be measured by load sensors on the cell surface.


  • It can be applied for cells with a high initial SOC
  • Not limited to low and constant current rates
  • Applicable to high current rates.


  • Needs additional equipment for the mechanical parameter measurement
  • Not applicable to cells constrained with hardcovers
  • Difficult for online application.

Machine learning (ML) methods

Machine Learning (ML) is a method of data analysis that automates analytical model building. It is based on the idea that systems can learn from data, identify patterns and make decisions or predictions with minimal human intervention. The first step in the machine learning-based SOH estimation is data collection. Measurable battery parameters, such as temperature, current, and voltage data, are recorded during the operation and used to train the model.

However, not all data are relevant to cell aging. A second step is to extract the features representative of the aging process. The third step is to train a machine learning model to describe the relationship between the battery SOH and the extracted features. Once the model is trained, the last step is implementing it in an application. Feature extraction is a critical step and significantly affects the SOH estimation performance. More meaningful and accurate input data will produce more relevant and accurate predictions.

The common process to analyze the battery by machine learning is composited by three steps: (a) collecting data and selecting suitable features, (b) constructing machine learning models, and (c) analyzing the target systems. The conventional machine learning methods used for SoC estimation methods are as follows:

  • Feedforward neural network (FNN) – Based on capacity loss during charging and internal resistance extracted from voltage variation.
  • Recurrent neural network (RNN) – Based on capacity loss during charging and capacity and resistance estimation.
  • Radial basis function (RBF) neural network
  • Hamming networks (HNN) – Based on ECM internal resistance estimation.
  • Support vector machine (SVM) – based on capacity and resistance variation
  • Bayesian network (BN) – Based on resistance.
  • Extreme learning machine (ELM) – Based on voltage variation.
  • Dynamically driven recurrent network (DDRN) – Based capacity loss and number cycles.
  • Sparse bayesian predictive modeling (SBPM) – Based capacity loss and sample entropy