An eight-element 3D microphone array is designed for source separation and noise cancellation applications in noisy and reverberant environments with multiple sound sources. In the first phase the non-negative matrix factorization is applied to each channel of the array to isolate the target signal from the mixture. In the second phase a machine learning approach is applied for designing a beamformer by the means of deep learning techniques to learn and reconstruct the target signal coefficients. The matrix factorization and machine-learnt beamforming are shown effective tools for speech and music analysis in this contribution they are adapted to a novel context of non-stationary industrial signals. It is also shown that the proposed 3D array is a more effective tool for capturing the acoustic scene compared with the 2D rectangular sub-array (only the four microphones in the front panel) in terms of noise suppression and signal quality. A comparison made the proposed machine-learnt beamforming method and the baseline analytical method suggests superior performance of the machine-learnt approach.