Abstract: This study investigates minor disruptions in the Keda Torus eXperiment (KTX) and presents a deep learning-based approach for disruption prediction. Plasma disruptions in magnetic confinement devices are critical to operational safety, making their prediction essential. Machine learning methods provide a foundation for predicting disruptions and informing mitigation strategies. Although the stabilizing shell in reversed-field-pinch (RFP) devices typically suppresses disruptions, minor disruptions occur frequently in KTX and are closely linked to plasma current center-of-gravity displacement. This work demonstrates, for the first time, the accurate and effective prediction of minor disruptions in an RFP configuration. The KTX recurrent neural network (KRNN) prediction framework, which integrates two-dimensional (2D) closed-boundary eddy-current diagnostic arrays, additional diagnostic methods, and long/short-term memory networks (LSTMs), enables early warning of minor disruptions and supports active feedback control. Notably, the 2D closed-boundary electromagnetic probe array in KTX, which captures more spatial distribution information than one-dimensional (1D) arrays, extends the early warning time for minor disruptions by an average of 21.1%. The KRNN achieves a maximum prediction accuracy of 98%.