As China′s oil and gas exploration and development continues to move towards unconventional oil and gas reservoirs, it is vital to develop core equipment and key technologies for deep resource exploration. Acoustic logging while drilling tool is a key equipment for deep earth exploration, while its core technology is blocked by foreign countries. The quality of acoustic signals received by domestically developed instruments has declined due to the influence of various noise sources such as drilling operations, tool eccentricity, mud circulation, and circuit transmission noise. This article introduces an overall design of an acoustic logging while drilling instrument, including a transmitting circuit, transducer, and array receiving module. A convolutional neural network algorithm based on encoder-decoder architecture is adopted to reduce the noise of acoustic data in the time-frequency domain, which enhances the signal-to-noise ratio of the instrument′s received signals. The time-frequency features of the noise-containing signal after a short-time Fourier transform are used as the inputs to the network. The neural network with a U-shape architecture learns the sparse representations of the signal and noise in the data and generates the time-frequency masks at the same time. In this way, the separation of the signal and noise is realized. To address the lack of open data sets, theoretical modeling is implemented for the instrument addressed in this article, and many theoretical simulations are carried out to obtain the signal database for different model parameters, and high-quality noise data are collected to establish the noise database. After training, the neural network is able to intelligently reduce the complex multi-source downhole noise, and this algorithm can achieve a good noise reduction effect on the test data and the acoustic signals collected from the instrument site, which greatly improves the quality of the received signals of the instrument under the interference of multi-source noise, such as low-frequency noise, circuit overshooting, complex oscillations, and mutant noise.