To address the limitations of conventional active noise control algorithms, such as slow convergence speed, poor robustness, and insufficient dynamic noise handling capability, an improved algorithm is proposed by integrating a one-dimensional convolutional neural network with fixed-coefficient filters and adaptive algorithms. The methodology involves extracting noise characteristics and dynamically selecting pre-trained fixed filters, while a Sigmoid function and quantization error compensation mechanism are introduced to optimize adaptive step-size parameters and enhance algorithmic stability. The algorithm is implemented on an STM32H750 high-performance embedded platform, constructing a real-time noise control system. Simulation results

demonstrate significant effectiveness in suppressing low-frequency noise, achieving an average noise reduction of 20-30 dB. Both temporal amplitude and spectral energy distribution of residual noise under dynamic environments are effectively suppressed, showing superior performance compared to traditional adaptive algorithms. Hardware experiments confirm that the results meet expected objectives. The effectiveness of combining deep learning with embedded hardware for active noise control applications is verified, providing an innovative and practical solution for real-time control in complex dynamic noise scenarios. This integration demonstrates considerable theoretical significance and practical application value through its successful

coordination of intelligent learning mechanisms with adaptive control frameworks.

active noise control(ANC), convolutional neural networks(CNN), STM32H750 microcontroller, filtered-x least mean square(FxLMS) algorithms