As one of the most widely used energy storage technologies, lithium-ion batteries have been successfully applied in modern society, particularly in electric vehicles (EVs) and energy storage stations. However, the thermal runaway problem of lithium-ion batteries may lead to fires and explosions of large energy storage systems, seriously affecting people's safety and restricting its further application and development. Unfortunately, due to the limited development of electrode materials and battery structure, it is difficult to eliminate thermal runaway in the short term. Therefore, an accurate and effective early warning method can improve the safety performance of the equipment. At present, the early warning methods for lithium-ion batteries mainly include: knowledge-based, model-based, and data-driven. However, knowledge-based methods are difficult to acquire knowledge and establish rules; model-based methods require high accuracy and large calculation; data-driven methods can be well applied to early warning, but require a large amount of data, difficult threshold training, and results that are not easy to interpret. In order to protect life and property safety, it is necessary to develop a battery pack thermal runaway early warning method that does not require a large amount of data training and low calculation. In this study, we propose method based on state representation methodology (SRM) for early warning of thermal runaway in lithium-ion batteries. This method involves the use of dynamic thresholds and graded warnings to continuously monitor the battery's state in real time and identify and locate faulty batteries. The results show that the method can accurately identify early failures of the battery by amplifying the small changes in voltage, preventing thermal runaway. Additionally, the dynamic threshold processing achieved through the sliding window solves the problem of threshold requiring a large amount of data training, improving the ability to identify faulty cells. The use of graded alarms makes it easy to process alarm information and improves the accuracy of the warning, providing good guidance for practical application. For the obtained thermal runaway vehicle data, the warning time of the method was 27 minutes and 12 minutes ahead of the alarm time of the two thermal runaway vehicle battery management systems (BMS). Change the length of the observation window from 4 to 30, and advance the warning time by at least 7 minutes and 5 minutes 12 seconds. Balancing computing and time costs, a length of 4 observation window is the most suitable. In addition, the method has demonstrated good feasibility, robustness, and stability in the validation.