The construction of three-dimensional complex shapes is one of the difficulties for application of meshless particle methods on internal flow simulation of fluid machinery. Based on the moving particle semi-implicit method (MPS), a generic discretizing approach was proposed to discretize arbitrary complex 3D geometry in fluid machines by combining the generic smooth wall model (GSW) and non-surface detection technique (NSD). The CAD models were discretized with shell meshes, and position of the nodes was extracted to generate single-layer wall particles. Initial fluid distribution was obtained by inputting fluid particles to the wall model based on the adaptability of fluid particles. Considering the main characteristic of the GSW model that the scale of wall particles has no influence on simulation, a coupled refining method including the local and nested refinement was developed. Surfaces with large curvature were locally refined, and the nesting refinement technique was applied to the surfaces that suffered from disturbance of wall particles on different surfaces. With this technique, the accuracy of normal vectors can be guaranteed and fewer particles were used to represent high-accuracy wall shape. A discretizing accuracy evaluation method was proposed, in which the normal smooth angle (NSA) was regarded as the criterion. Taking the actual centrifugal pump model as the example, the geometric accuracy of the discrete model was quantitatively described, and the recommended value of the NSA value of the characteristic model was analyzed and given. The whole procedure of constructing high-accuracy wall particle models was demonstrated, and the stability and accuracy of the proposed models were validated by two 3D hydrostatic cases with complex geometries, the dam-break impacting on an obstacle case and the inducer rotating in water case.