Abstract: As a neural network model that simulates the operational mechanisms of biological neural systems, Spiking Neural Networks (SNNs) demonstrate significant potential for low-power computing and efficient hardware deployment. To enable on-chip learning for SNNs, it is essential to effectively map their training process onto hardware platforms. The current mainstream approach, the surrogate gradient training algorithm, can achieve satisfactory performance in software. However, it relies on the chain rule-based backpropagation process, which presents challenges in hardware implementation, including complex gradient computation and high resource consumption. To address this issue, this paper proposes a gradient-free SNN training method based on sparse direct feedback. This method employs Direct Feedback Alignment (DFA) to propagate output errors, thereby bypassing the chain rule differentiation in backpropagation and effectively reducing computational complexity and hardware resource requirements. To further optimize hardware implementation, a sparse structure is introduced into the feedback matrix, significantly reducing data transmission volume. Additionally, nonlinear functions are used to replace the differentiable approximations of spiking neuron derivatives, eliminating explicit gradient calculations. This approach ensures both biological plausibility and strong network performance. Experimental results on the MNIST dataset show that compared to the conventional DFA method, the proposed algorithm improves network training accuracy while reducing the number of weight updates. This study provides an efficient training implementation solution for on-chip learning of SNNs, contributing to the practical application of SNNs on low-power hardware platforms.