The haptic perception technology for embodied agents is primarily approached through two strategies: the thorough extraction of deep features from data and the design of novel sensors to expand data modalities. Tactile data acquisition has encompassed diverse aspects such as image texture, temperature, electrical properties, force, and vibration frequency. While many methods achieve high recognition accuracy by designing complex sensor structures tailored to specific recognition features, these approaches often lack practicality for convenient and scalable implementation using industrially available general-purpose sensors. Proximity sensors offer a distinct advantage by generating differentiated sequential data during object approach, simultaneously addressing the requirements for sensor versatility and deep data utilization. Leveraging this characteristic, a bidirectional LSTM sequence feature extraction model supported by single-point proximity sensor data is proposed. Hierarchical loss for classification and piecewise loss for force value range estimation are designed within the model, effectively enhancing its ability to distinguish between similar categories and improving the accuracy of force value regression. On the constructed dataset, a classification accuracy of 98.86% is achieved by the model. For end-to-end force magnitude estimation, a minimum force estimation success rate of 76.19% is attained across all categories. These results verify the effectiveness of the proposed methodology.