Case in Point | How oorja Heat App + NXP eIQ Auto can enable efficient battery thermal management & better battery insights

November 15, 2023 battery management system, EV battery 11 min read

Introduction

The temperature of a Li-Ion battery cell plays a crucial role in its performance and degradation profile. Cell temperature is a key parameter in determining metrics such as SoH (state of health), internal impedance, and RUL (remaining useful life). At the same time, limitations with traditional BMS (battery management system) hardware mean OEMs are constrained to the use of a handful of physical temperature sensors. The primary limit is the number of auxiliary voltage channels supported by the BMS AFE (analog front end). Typically, a maximum of 6 to 9 temperature sensors can be monitored based on the AFE used. Adding more temperature sensors, even if possible, adds to the cost and complexity due to the wiring harness needed between the sensors and the BMS hardware. The net result of this constraint is that while the OEM may have good visibility into the temperature for a few of the cells corresponding to where the sensors are placed, for many other cells, the visibility into the cell temperature may not be as good. Based on the current state-of-the-art scenario, it is shown conceptually in Fig. 1.

Figure 1: Traditional BMS HW limitations mean OEMs have limited visibility into individual cell temperatures.

An actual 48V Li-Ion battery pack for use in electric 2W or 3W may have as many as 100 to 200 cells connected in series and parallel with still the same number of physical temperature sensors (e.g., 4-6). This means that the OEM has low-confidence cell temperature data for most of the cells in the battery pack. The consequences are significant:

OEMs often oversize their thermal management system to ensure battery safety.
OEMs cannot extract the actual state of health, limiting the value of a second life.

Machine learning offers a software-based approach for predicting individual cell temperatures based on existing HW temperature sensors and pack characteristics such as filler material, cell chemistry, and pack current profiles. However, the BMS that would incorporate machine learning for predicting cell temperatures is a critical system that needs to operate reliably & safely in the vehicle. As such, there are two crucial challenges with deploying ML in the real world for BMS:

Running edge ML with automotive quality software for inference and in real-time.
Generating accurate real-life training data covering the range of operating conditions and ensuring region-specific training data generation.

Currently, the only ways to generate data that can be used to train ML algorithms are real-life testing or physics-based simulations. Real-life testing at the design stage is highly time-consuming since the range of conditions to which the battery pack will be exposed must be captured, and real-life data gathering at this scale is impossible. Similarly, using physics-based models to generate synthetic data involves solving large 3D models for various real-life scenarios, which is highly time-consuming and requires intensive computational resources.

In this concept paper, NXP and oorja present a collaborative approach to solving the problem.

NXP eIQ Auto brings, for the first time, an automotive quality inference engine that can run in real-time on S32K3, the latest-generation automotive microcontroller platform. This enables running edge ML for electric 2W & 3W powertrain applications such as BMS & traction motor control.

oorja HEAT enables the quick generation of real-life high-fidelity battery response data at the battery pack design stage using a first-of-its-kind hybrid (physics model + ML-based) approach for quickly generating large amounts of high-fidelity training data for a given battery pack design. oorja’s validated approach is up to seven times faster than conventional physics-based modeling and up to 98% accurate.

This disruptive, new approach can be visualized conceptually with the help of Fig.2.

Figure 2: Conceptual framework showing the coming together of oorja HEAT + eIQ Auto for edge ML on the S32K3 platform.

Workflow

The workflow (shown in Fig.3) starts with a battery pack design in the oorja Design App.

Figure 3: oorja HEAT enables accurate temperature prediction for ML training data without needing expensive test equipment.

The oorja app uses a combination of high-fidelity physics-based models: electrochemical, thermal, and flow. It then uses information from cell-level HPPC data to combine the two to create a predictive model for the entire pack’s thermal behavior and performance. This information can then be used to optimize pack design and encode the BMS algorithms to ensure optimized pack usage.

In this case, we have chosen a typical 14S9P battery pack using NMC cells, with a nominal capacity of 4.8Ah and a power of 2.23kW. The pack design, shown in Fig. 4, is representative of an electric scooter battery pack.

Figure 4: Details of the battery pack modeled using oorja app.

Next, the pack designer selects the location of physical temperature sensors, e.g., based on thermal hotspot simulations. Once the pack design and temperature sensor locations are finalized, the next step is to run simulations with the oorja Heat App over multiple operating conditions. The oorja Heat App can predict individual cell temperatures with high confidence, as seen in Figure 5.

Figure 5: Cell Temperatures for the MIDC Discharge and 0.3C Charging Cycle- oorja HEAT App

The output of the simulations, comprising individual cell temperatures, pack voltage, and pack current, is then used as training and test data for the edge ML. The training data can be visualized using Fig. 6.

Figure 6: Visualizing the training dataset generated by oorja HEAT.

The next step is to train the ML model – e.g., a neural network to predict the temperatures of cells that are not monitored by a physical temperature sensor. This training is performed offline. Once the ML model is trained, the next step is to deploy the model on the S32K3 target platform. The eIQ Auto ML training and deployment process can be visualized in Fig. 7.

Figure 7: NXP eIQ Auto workflow that enables seamless progress from training to deployment.

Finally, the baseline results comprising predicted individual cell temperatures from the oorja Heat App are compared with those from the edge ML inference engine for evaluating the accuracy of cell temperature prediction. This combined workflow is shown in Fig. 8.

Figure 8: Combined workflow across oorja HEAT & NXP eIQ Auto

Multiple real-world driving profiles can be loaded into oorja HEAT to generate high-quality data for training the ML model. The scenarios used for training and testing are summarized in Fig. 9 below.

Figure 9: Multiple MIDC and US06 driving profiles were used for training the edge ML model.

An LSTM neural network was trained on six features summarized in Fig. 10 below.

Figure 10: Summary of features – four temperature sensors were selected as features in addition to pack voltage & current.

The neural network itself was modeled as a convLSTM neural network, visualized in Fig-11 below.

Figure 11: Visualization of the edge ML model based on ConvLSTM. The inference engine executes on S32K3 automotive uC.

Results

Table 1 below summarizes prediction accuracy results, averaged over all predicted cell temperatures. The MAE (Mean Absolute Error) is less than one deg C, while the R^2 coefficient is > 0.9, indicating a good fit with patterns seen in cell temperatures. All of this is with an edge ML model that can execute inference in real time on the S32K3 microcontroller. Overall, the results appear promising. Further optimizations can improve accuracy or reduce model footprint depending on the priority. Furthermore, the same model could be extended for applications such as predicting battery degradation and the state of charge/health parameters, maximizing the benefit of edge ML using eIQ Auto.

Table 1: Results for prediction accuracy and edge ML performance

Figure 12: Edge ML predicted vs True temperature for a typical cell.

Indicative results for proof-of-concept. Results may improve further with further ML training & model optimization.

Conclusion

Combining the ability to deploy an automotive quality ML inference engine in real-time on the edge with the capability to generate high-fidelity training data using simulation software such as oorja Heat App can unlock new opportunities for OEMs – e.g., to generate better battery insights for enhancing battery residual value & optimize the battery thermal management systems to save overall system-level costs. The eIQ Auto framework would also help you easily port the algorithm to other current and future NXP S32 microcontroller devices. While the current proof-of-concept demonstrates the use case of predicting individual cell temperatures, the beauty of the edge ML model is that the same model can potentially be extended to generate various other vital insights – as shown in Fig. 13. In conclusion, the ability to deploy automotive quality edge ML seamlessly can be a game-changer for unlocking the full potential of different business models associated with Li-Ion batteries.