The rapid rise of industrial automation has underscored the need for accurate detection to boost efficiency, ensure product quality, and reduce labor costs. Tasks like connector mating and grasp stability rely heavily on analyzing tactile force profile time series, which are difficult to model due to complex, non-stationary temporal dynamics, subtle signal variations, and the high cost of acquiring labeled data. Such factors limit the effectiveness of conventional supervised approaches, making it challenging to build reliable models with limited annotations. To address this, we propose Tactile-driven Multiple Instance Learning (Tacti-MIL), a lightweight deep learning framework that integrates a multi-scale convolutional backbone for temporal feature extraction with a Transformer-based MIL module. This design enables efficient aggregation of temporal patterns and robust performance even with small datasets. Extensive evaluations show that our proposed Tacti-MIL outperforms baseline models, offering a balance between detection accuracy and computational efficiency for industrial detection.