BU-907b: Advancements in Battery Testing

Battery testing is the product of electrochemical evidence and data analysis with artificial intelligence. A battery cannot be “measured,” only estimated by analyzing its symptoms. These symptoms change with state-of-charge (SoC), temperature, agitation, storage and age. As an example, a good battery with low charge performs similar to a weak battery that is fully charged. Modern test methods must identify these conditions. Figure 1 summarized current SoH test methods organized into five categories.

**Figure 1:** State-of-health (SoH) estimations

Pulse Method

**Figure 2:** Electrochemical Dynamic Response

Applying pulses measures the internal resistance (Ri) of a battery. Differences between pulse duration and depth of discharge reveal unique performance characteristics. However, capacity assessment is not possible as Ri does not correlate reliably with capacity.

Dynamic Electrochemical Response (DER) checks ion-flow in Li-ion batteries to examine the dynamic behavior relating to battery health (SoH). DER tests a broad range of batteries by profiling a model-specific battery.

Multi-model EIS

**Figure 3:** Spectro™ combines EIS with Artificial Intelligence

EIS applies a sinusoidal signal from 2KHz to 100mHz, followed by filtering and magnitude extraction. Modelling and statistical analysis occurs by fitting Nyquist plots with matrices derived from same-type batteries but different SoH. Data fusion correlates values of key parameters to measure capacity, Cold Cranking Amp (CCA), State-of-Charge (SoC) and other readings.

Scientists believe that future battery diagnostics lies in Electrochemical Impedance Spectroscopy (EIS). Cadex was first to test batteries with EIS under the Spectro™ trademark, incorporated in Spectro™ devices to test lead acid and Li-ion chemistries from 3V to 60V with capacities up to 250Ah. Ri must be above 500 micro-ohms Figure 3 illustrates the mechanics. See also see BU-904.

Several algorithmic methods can be used to measure battery SoH with EIS.

Fuzzy Logic

Fuzzy logic has been the most common algorithm in EIS battery analysis. Model-based parameters are fuzzified and passed through membership functions called matrices. The data is “defuzzified” to obtain battery SoH and other readings. Figure 4 illustrates deployed membership functions.

In some applications, fuzzy logic is superseded with other methods that provide solid results with low-level training data.

Artificial Neural Network (ANN)

ANN relies on big data to provide a classified output using artificial neural network. The connecting units with hidden layers are called neurons. They form a biological brain that simulates animal instinct. Correct prediction is mostly determined by the number of hidden layers, number of neurons in each hidden layer, activation function, optimizer learning rate and training epochs. One might say: “Garbage in; garbage out,” but the Data-driven method produces reliable results with known user-patterns. Figure 5 demonstrates a typical Artificial Neural Network.

Gaussian Process Regression (GPR)

Model-specific parameters are trained with machine-learning to measure battery SoH by the Bayesian approach. GPR offers superior results with fewer training data than ANN. Cadex labs achieves accuracies of 90% with lead acid batteries, findings that are being verified by scientists at UBC. The tests repeated with Li-ion systems are getting promising results. The Gaussian Process Regression as illustrated in Figure 6 is a promising variant.

Cadex labs study other methods to assess battery SoH with the results shown in Table 7. With a pool of 800 lead acid test batteries with various SoH, Gaussian Process, Neural Net and Fuzzy Logic are of interest. Other methods reflect lower accuracies when testing bad batteries.

**Table 7:** Percentage of correct prediction of 800 lead acid starter batteries. Neural Net and Gaussian Process have best results.

Definition: Good Batteries have a reading above 40% capacity; Poor Batteries are below 40%. 40% is the common pass/fail threshold of starter batteries.

Green: Good battery
Red: Bad battery

Cadex is experimenting with proprietary algorithms that will gain further improvements over the results shown in this paper using multi-model EIS to study the kinetic reactions of batteries at various SoH conditions. The goal is to create robust matrices with the least number of sample batteries serving as training.

Adaptive Filter

The parser estimates the usable battery capacity by reading state-of-charge with the Extended Kalman Filter (EKF) and counting coulombs to fill the available space to reach full charge. The usable capacity is the sum of the measured SoC plus the energy added. For best result, each battery should undergo a one-time calibration by cycling a good pack. In addition, the battery must be sufficiently discharged to attain accurate readings. A “charge runway” from 40% to 100% provides accurate value in assessing the usable capacity.

**Figure 8:** SoC estimation with the Kalman Filter

Not all batteries are fully charged when tested. Combining EIS with the Kalman Filter improves the accuracy when testing a battery at low SoC. Identifying a good battery at low charge is of special interest for car makers to reduce warranty claims. Figure 8 illustrates how large fluctuations are being levelled under controlled usage.

Last Updated: 26-Apr-2024

Previous Article Next Article

Batteries In A Portable World

The material on Battery University is based on the indispensable new 4th edition of "Batteries in a Portable World - A Handbook on Rechargeable Batteries for Non-Engineers" which is available for order through Amazon.com.

More Details Buy On Amazon