Traditional spiral zone plates (SZPs) have been widely used to generate optical vortices, but this structure suffers from multiple focuses. To eliminate high-order foci, the current method is to design a binary structure that has a sinusoidal transmittance function along the radial direction. With the rapid development of artificial neural networks, they can provide alternative methods to design novel SZPs with a single focus. In this paper, we first propose the concept of generalized binary spiral zone plates (GBSZPs), and train a feedforward neural network (FNN) to obtain the mapping relationship between the relative intensity of each focus and the structural parameters of GBSZPs. Then the structural parameters of GBSZPs with a single focus were predicted by the trained FNN. It is found by simulations and experiments that the intensities of high-order foci can be as low as 0.2% of the required first order. By analyzing the radial transmittance function, it is found that this structure has a different distribution function from the previous radial sinusoidal function, which reveals that the imperfect radial sinusoidal form also can guide the design of binary zone plates to eliminate high-order foci diffraction. These findings are expected to direct new avenue towards improving the performance of optical image processing and quantum computation.