To study the safety and reliability of electric vehicle battery packs under mechanical abuse conditions, and prevent a series of safety accidents such as short circuits, thermal runaway, even fires and explosions that may occur when battery packs are subjected to collisions, this paper builds a mechanical abuse conditions of lithium battery thermal safety testing experimental system. Then safety experiments on lithium-ion batteries were designed and implemented under 4 mechanical abuse conditions: flat plate compression, cylindrical compression, spherical compression, and nail penetration test. A lithium battery monomer homogenization model was built through extensive data from mechanical abuse experiments. A complete lithium battery pack data set was established combined with the experimental data and lithium battery monomer homogenization model. Building on this foundation, a fusion mathematical model comprising support vector machine (SVM), sparrow search algorithm (SSA), and back propagation (BP) neural network was designed to predict the failure states of lithium battery packs under different parameter conditions, as well as the displacement and maximum load following failure under various parameter conditions. The experimental results show that SVM-SSA-BP can accurately predict the failure state, failure displacement and maximum load of the battery pack system. A comparative validation was conducted against four other competing models, and the generalization ability of SVM-SSA-BP model was evaluated by adding 20% Gaussian noise to the dataset. The results show that the predictive performance of the SVM-SSA-BP fusion mathematical model is superior to four other competing models, and it demonstrates good accuracy and stability. The model proposed in this article can accurately predict the safety performance of lithium battery packs under mechanical abuse conditions.