<p>Aiming at the existing robotic hand’s reliance on visual guidance for object recognition during the grasping process, which is limited by environmental lighting and object occlusion, this paper is committed to proposing a method for object recognition and posture determination based on contact force feedback. The paper first constructs a contact surface deformation model for classic button operations and grasping operations in power systems, based on Hertz contact and elastoplastic unloading theories, laying the theoretical foundation for the vector mechanics decomposition of the contact surface. It then proposes a dimensionality reduction analysis method for vector force arrays, identifying the boundary force characteristics of the contact surface to decompose the 3D contact surface into separate force planes and reducing the angle dimension in the vector array to a two-dimensional scalar array that only includes the magnitude of the force, thus laying the data foundation for introducing convolutional neural networks (CNNs). Through training and learning the mechanical features of the contact surface, the paper achieves perception and posture determination of objects during pressing and grasping processes. Practical experiments demonstrate that the proposed method for rapid analysis and extended posture determination of the grasping surface based on CNNs can achieve millimeter-level precision in grasping position determination and a two-dimensional posture angle of less than 1 degree for the grasped object, generally meeting the requirements of power system business scenarios. The blind-sense technology explored in this paper can be widely applied to non-visual perception at the end of humanoid robots and has significant engineering guidance value.</p>