Xi'an Jiaotong University uses AI for battery modeling and health management, su

Recently, the team led by Assistant Professor Zhao Zhibin from Xi'an Jiaotong University has combined physical models with deep neural networks to propose an AI algorithm known as "Physics-informed neural network" (PINN) for battery modeling and health status prediction. This algorithm has been applied to both battery modeling and the prediction of battery health status.

The research team has also made the relevant dataset and complete code open source ().

In addition, by integrating three other datasets from different battery manufacturers, they validated their method on data from 387 batteries across 310,705 samples, achieving an average absolute percentage error of 0.87%.

To verify this method, the team conducted battery degradation experiments, creating a comprehensive dataset composed of 55 nickel-cobalt-manganese batteries and simulating six different working conditions.

The experimental results show that the proposed Physics-informed neural network is applicable to datasets of batteries with different chemical compositions.

Capable of adapting to different types of batteries and various usage scenarios, this approach is expected to facilitate the development of battery health management systems.Translate the following text into English:

Advertisement

Specific applications may include:

1. For electric vehicles.

- Enhance the accuracy of battery life prediction and battery life management systems, optimizing the battery usage strategies and maintenance strategies for electric vehicles.

2. For aerospace.

- Improve the battery management systems for satellites and drones, ensuring the reliability and safety of missions, and reducing mission failures due to battery malfunctions.Thirdly, for portable electronic devices.

Optimizing battery management for devices such as smartphones and laptops to enhance user experience and extend device lifespan.

Fourthly, for energy storage systems.

In large-scale energy storage systems, optimizing the monitoring and management of battery health status to ensure the energy system operates stably and efficiently."AI+ Batteries" Research, How to "Add Value to Excellence"?

In recent years, the use of lithium-ion batteries has grown at an astonishing rate, permeating almost every aspect of social life.

Lithium-ion batteries, with their high energy density, low self-discharge rate, and long service life, have become the main energy storage devices in various fields such as portable electronic devices, electric vehicles, aerospace, and more.

However, the large-scale application of lithium batteries has also brought a series of new challenges and issues, especially in terms of safety, reliability, and environmental protection. Therefore, the degradation modeling and health management of lithium batteries have become particularly important.

In the field of aerospace, as an important energy supply for high-tech equipment such as satellites and drones, lithium-ion batteries play a crucial role.These devices have extremely high requirements for the reliability and stability of batteries. Once there is a problem with the battery, it may lead to mission failure, and even cause huge economic losses and serious safety hazards.

By conducting fine degradation modeling and health management of lithium batteries, potential faults can be detected in advance to ensure the normal operation of equipment in complex environments.

At the same time, the research group found that the research on physical information neural networks in battery modeling and battery assessment has begun to show promise.

However, despite the fact that a large number of papers on battery health management have been published in some journals in recent years. However, the health management methods designed in these papers are mainly aimed at specific datasets.

Once a different dataset is used, the methods proposed in the papers may become invalid. That is to say, most of the published papers are still at the stage of using physical knowledge to preprocess data, and have not achieved a deep integration of physical models and neural networks.On the other hand, they found that the "state-of-health" (SOH) estimation method based on traditional deep learning has significant fluctuations in predictive error, which leads to the model's accuracy largely depending on the quality of the data.

"One teacher, one student," starting from scratch.

In recent years, Zhao Zhi Bin's team has mainly focused on intelligent maintenance and health management of major equipment, with the main research subjects being various large mechanical equipment.

In fact, it was also by chance that Zhao Zhi Bin was exposed to the field of battery health management.

"To elaborate, the germination of this research is as follows: Previously, the larger team I was in carried out a project on the health status of a certain type of satellite battery. Later, we came up with the idea of researching the health status of lithium batteries and assigned this task to doctoral student Wang Fujin." said Zhao Zhi Bin.At that time, Wang Fujin was still a newly enrolled master's student, and under the condition that Zhao Zhibin's team had no relevant research foundation, Wang Fujin also felt at a loss for this topic.

Moreover, at that time, the team only had a "one teacher and one student" personnel model like Zhao Zhibin and Wang Fujin, and everything had to start from scratch.

Later, through methods such as extensive literature review, research and learning at other universities, and communication with enterprises, as well as with the gradual development of the "satellite battery health status project," the research team gradually had the conditions for researching lithium battery health management.

Subsequently, they first collected a large amount of battery degradation data.

High-quality data is the foundation for model training and model verification. To this end, they obtained data through various means, including laboratory testing, calling on public data sets, and cooperating with enterprises to obtain actual usage data, etc.After analyzing the data, universal statistical characteristics were extracted and the impact of different features on the prediction of battery health status was explored.

To design a more universal set of features, the team analyzed a large amount of battery degradation data and ultimately summarized a set of common statistical characteristics.

After fully understanding the characteristics of the data, combining the rigor of physical models with the flexibility of neural networks, they proposed the Physical Information Neural Network, an AI algorithm.

In terms of model design and development, multiple factors need to be considered: including the selection of physical models, the design of neural network architecture, the optimization of loss functions, and the tuning of hyperparameters during training.

Through a large number of experiments, the model structure gradually became more refined, and the prediction accuracy and stability have also improved.To verify the effectiveness of the physical information neural network model, the research team conducted new experiments on multiple battery datasets that include different chemical compositions and various operating conditions.

The experimental results show that in the prediction of battery health status, the physical information neural network model not only performs exceptionally well but also significantly outperforms traditional deep learning methods and purely physical models.

Especially with the aid of transfer learning, the model demonstrates strong generality and robustness even across different datasets.

Considering the complexity of electrochemical equations, the team proposed an empirical model of battery degradation from the perspective of battery degradation experience equations and state-space, and used the physical information neural network to capture the degradation dynamics of the battery.

To address different battery types, usage scenarios, and charging and discharging protocols, the research team designed a universal feature extraction method to extract features from a short period of data before the battery is fully charged.After analyzing the feature data, they found that the correlation coefficient between the extracted features and the battery health status is mainly related to the chemical composition of the battery, and is less affected by the charging and discharging schemes.

Future new direction: Research on AI-empowered intelligent maintenance and health management of major equipment.

Subsequently, they began to write the paper and submit it. Recently, the related paper was published in Nature Communications with the title "Physics-informed neural network for lithium-ion battery degradation stable modeling and prognosis" [1].

Xi'an Jiaotong University doctoral student Wang Fujin and Associate Researcher Zhai Zhi are the co-first authors, and Assistant Professor Zhao Zhibin and Professor Chen Xuefeng from Xi'an Jiaotong University are the co-corresponding authors.

At present, Zhao Zhibin has set the research direction of his group as: AI-empowered intelligent maintenance and health management of major equipment.To this day, the team has already included two PhD students and three master's students researching in this direction, as well as undergraduate members who have accumulated a certain amount of scientific research experience in this field.

"At present, they have already produced a number of high-level papers," said Zhao Zhi Bin.

In the future, the research group will continue to follow the dual drive direction of "physics + data" to carry out research on battery health management, and will mainly explore the health management of lithium batteries on spacecraft and vertical take-off and landing aircraft.