To address the issue of the Mach-Zehnder modulator transmission characteristic curve drifting due to environmental influences, a novel deep learning based bias control method utilizing a synergy between a neural network and an improved particle swarm algorithm is proposed. The proposed method enables the locking control of any working point by constructing multi-dimensional features as inputs to the neural network and establishing a specific mapping relationship with the bias voltage. Compared to other dither-based methods using proportional integral derivative controller, our method does not use second and higher harmonic that are hardly measurable, which is less costly. The experiments show that the proposed method is robust and invariant to changes in temperature and has good long time stability at different operating points. In the temperature perturbation test, the maximum fluctuation of the output optical power is 8.7 μW, which is only 6.3% of the maximum fluctuation for the conventional control algorithm.