Unfortunately, I don't know if the Akida technology is being investigated here, but it sounds very similar to me as a layman.
......
"The model's input is the capacity time series until the present time point. No other features or feature selection process is required. The model then outputs the future capacity time series of the cell until the EOL. The capacity of the cells can be measured directly with capacity tests in the training dataset, or estimated with any standard SOH estimation models [[51], [52], [53]] onboard a battery management system. Fig. 2b shows the choice of input, as well as the sampling process of the input data that was undertaken before feeding to the model. The input approach is an increasing window of the remaining capacity array, from the beginning of life (BOL) of the cell, to the present time point. This input series increases in size as more historical data becomes available, and the model uses the increased information to generate improved predictions.
Since both the input and output series lengths are generally constant in LSTM networks, a normal LSTM is insufficient to perform this task. However, RNN-based sequence-to-sequence (S2S) models [54] are specially designed for sequence prediction problems. The S2S approach has a better performance compared with other RNN architectures and performs especially well in predicting sequences of variable length [54], which is the target use case of this work since battery degradation history can be viewed as a variable-length sequence. Fig. 2c highlights the general architecture of the S2S model, which requires two multi-layer LSTM architectures working together. The encoder encodes the input sequence into a static embedding vector, which is then fed to the decoder which decodes the embedding vector and processes it to provide the output sequence desired. The model performance was validated with processor-in-the-loop tests with both normal and noisy capacity data, as shown in Fig. 3. The mean absolute percentage error (MAPE) during curve prediction and absolute cycle number errors for the knee-point and the two EOLs are used to evaluate the prediction performance of the models, as defined in the ‘Evaluation criteria’ section.
"
Finally, the model was compared with a typical LSTM-RNN iterative prediction model (see Methods for details). The results in Fig. 6 show the S2S model as the better performer in predictive ability in both the best- and worst-cases in all metrics, as well as faster than the LSTM model. This is expected since the complexity of the S2S model is almost double of that of the LSTM model, which results in the iterative model taking less training time. However, the iterative nature of the LSTM model requires it to run many times during implementation to provide the full curve output, whereas the S2S model can provide the entire curve in one shot, thereby significantly reducing the computation time,
up to almost 15 times in average. The detailed evaluation and comparison can be found in Fig. S7 and Table S2.
"
"
4.2. Applications and outlook
The main applications of the model are degradation analysis and failure predictions for improved maintenance, ensuring safe and reliable operation of battery systems. The model can also be used to analyze battery packs with various topologies by identifying inter-pack cell degradation variations and pointing out the strongest and weakest cells within each battery branch. The noise handling capability of the model reaffirms the application on real-world data, where the capacity estimation may have uncertainties that need to be considered.
The ‘cell-passport’ also adds value to first- and second-life applications from the collection of usage information during the whole life of each cell, introducing exciting developments for battery warranties and insurance. Furthermore, digital certificates can be provided along with the cell-passports for various second-life applications.
"
May this be related to Akida IP?
"
5. Conclusions
In this paper, a one-shot battery degradation trajectory prediction model was proposed for batteries under real-world operations, coupled with a cloud-based prognostics framework. The model is based on sequence-to-sequence learning and is trained and validated on a 48 NMC/graphite cell dataset. Model training requires only measured operational data without the need for additional parameterization or feature engineering. The model provides a ‘one-shot’ prediction of the entire future degradation trajectory, which decreases the computational burden almost 15 times compared with current iterative prediction approaches. An early prediction capability from as few as 100 cycles is possible not only for degradation prediction but also for the endpoints of both first- and second-life and the degradation knee-point. The performance of the model and its robust noise handling capability is validated by processor-in-the-loop tests with 1.8% mean error in the best-case and 7.8% maximum error in the worst-case. This work primarily serves as a proof of concept for the use of modern deep learning-based architectures in the domain of battery prognostics.
In general, the health prognostics framework in this work can not only be applied in batteries with different materials but also further in other energy storage systems, e.g. fuel cells and super-capacitors.
"
This excerpt from the research project reminds me of the characteristics attributed to Akida.
"
Processor-in-the-loop test
Validation was done by serving the trained model into an embedded system, Jetson Nano [59].
This small embedded hardware validation of the model is significant mainly to demonstrate the standalone ‘prediction on edge’ capability where the proposed model, after being fitted into a local vehicle can operate perfectly, generate and store results even without constant connectivity or remote access to external servers. The Nano functions as a mini-sized onboard computer with an
NVidia CUDA-enabled GPU, which enables it to perform deep learning tasks. The board runs a Linux-based operating system that supports a full-sized TensorFlow environment; therefore, almost all functionalities of architectures designed in other such systems can be ported. In the ‘processor-in-the-loop’ test [60], data arrays are fed to the Nano's storage, from which it can run the model to validate the various validation scenarios taking the sensor data sequentially following a real operation usage style. The validation in the Nano is done mainly to demonstrate the model accuracy, computing capability and the viability of using the device in future BMSs.
"
Any Idea?