Thermal runaway characteristics and hazards of lithium-ion batteries under low ambient pressure in-flight conditions are studied in a dynamic pressure chamber. The influence of ambient pressures (95 kPa and 20 kPa) and packaging forms (cylindrical and pouch commercial batteries) were especially investigated. The results show that the values of heat release, temperature, and $\mathrm{C}{\mathrm{O}}_2$ concentration decrease with the reduction in pressure from 95 kPa to 20 kPa, while the total hydrocarbon and $\mathrm{C}\mathrm{O}$ increase. Without violent fire, explosion, and huge jet flames, the thermal hazards of TR fire under 20 kPa are lower, but the amount of toxic/flammable gas emissions increases greatly. The amount of $\mathrm{C}\mathrm{O}$ and hydrocarbons varies inversely with the thermal hazards of fire. Under low-pressure environments of cruise altitude, the thermal hazards of TR fire for pouch cells and the toxic/potentially explosive hazards of gas emissions of cylindrical cells need more attention. The performance of TR hazards for two packaging types of battery is also different. Pouch cells have higher thermal hazards of fire and lower combustible/toxic emitted gases than cylindrical cells. The thermal runaway intensity of individual cells decreases under lower ambient pressure, but the burning intensity increases dramatically when thermal runaway occurs in a battery pack. The open time of a safety valve (rupture of the bag) is shortened, but the trigger time for a thermal runaway varies for different formats of batteries under 20 kPa. Those results may be helpful for the safety warning and hazard protection design of Li batteries under low-pressure conditions.

Accurate prediction of the State of Health (SOH) of lithium-ion batteries is essential for improving the safety and longevity of energy storage systems. This paper introduces ExpertMixer, a novel model based on a fused expert network for SOH estimation. By combining the strengths of state space models and recurrent neural networks, the model effectively handles the joint optimization of long-sequence dependency modeling and complex dynamic feature extraction. To improve temporal representation, ExpertMixer utilizes sampling time-based rotary position encoding (RoPE). It consists of two expert modules: a Mamba module designed to capture global degradation trends and an LSTM module focused on modeling local dynamic fluctuations. These are adaptively fused through a learnable gating mechanism that supports multi-scale feature integration. Experiments performed on the NASA PCoE dataset show that ExpertMixer achieves optimal performance on the NASA L subset, with an average MAE of 1.047 and RMSE of 1.603. It surpasses the traditional CNN BiGRU model, which had an MAE of 2.286, by 54.2%, and improves upon the advanced SambaMixer model, which had an MAE of 1.072, by 2.3%. Under low-temperature conditions using Battery 47, the model reduces the prediction error for nonlinear degradation to an MAE of 0.539, significantly exceeding all compared methods. Ablation studies verify the effectiveness of the dual-expert structure and fusion mechanism; removing the gating module results in an 18.7% decrease in performance. This research offers a new framework for lithium battery life prediction that demonstrates improved accuracy and generalization capability, suggesting potential practical value for intelligent energy storage management.

Existing deep learning methods for lithium-ion battery State of Health (SOH) prediction rely almost exclusively on discharge voltage–current curves, ignoring electrochemical impedance spectroscopy (EIS) data that directly reflects internal degradation mechanisms. Fusing these two modalities is non-trivial: discharge curves are high-dimensional temporal sequences residing on a continuous dynamical manifold, while impedance features are low-dimensional static snapshots with fundamentally different statistical distributions. However, naive concatenation introduces modal conflicts rather than complementary gains. We propose the Hybrid Sensing Synergy Architecture (HSSA), which combines a Mamba backbone ($O\left(L\right)$ complexity) for discharge curve modeling with a Q-former module that aligns impedance features into the temporal representation space via learnable query tokens and cross-attention. A prepend fusion strategy injects the aligned queries as prefix tokens, enabling the backbone to condition on internal electrochemical context from the first time step. On the NASA battery dataset, HSSA achieves MAE of 0.887 (large-scale, 11 batteries, a 9.8% improvement over unimodal Mamba), 1.457 (medium-scale, five batteries, a 28.0% improvement), and 2.705 (small-scale, four batteries, an 8.7% improvement), demonstrating consistent improvements across all data regimes. On out-of-sample battery B28, HSSA achieves 65.3% improvement. Ablation studies confirm that Q-former alignment is essential and prepend fusion significantly outperforms concatenation-based alternatives.