This study aimed to develop a method for identifying tactile sensations based on the interactions between fingertips and material surfaces using fingertip acceleration and angular velocity measurement data, and to relate the obtained data to subjective sensory ratings of tactile sensations, using machine learning. A 6-axis Inertial Measurement Unit (IMU) was attached to the fingertip of each participant, and the acceleration and angular velocity were measured while the participant traced 20 different materials. After the tracing task, the participants rated the tactile sensations on a 7-point semantic differential (SD) scale for six basic sensory items that constitute the tactile sensation: roughness, warmness, softness, wetness, friction, and comfort. The measured acceleration data were transformed into spectrograms using short-time Fourier transform (STFT). Using the spectrograms as input data, convolutional neural network (CNN) model was used for machine learning to classify each of the six sensory items into seven levels. When tested on data from new participants, the CNN model achieved classification accuracies of approximately 40-60% for all six sensory evaluation items. However, further evaluation revealed that the accuracy values for “wetness” and “warmness” largely reflected the central tendency bias in the sensory rating, resulting in reduced estimation reliability. In contrast, “friction,” “roughness,” “softness,” and “comfort” showed relatively higher estimation reliability, suggesting that the IMU data more effectively captured the characteristics of these specific sensations.