The existing diagnostic techniques for detecting inter-turn short circuits (ITSCs) in induction motors face two primary challenges. Firstly, they suffer from reduced sensitivity, often failing to detect ITSCs when only a few turns are short-circuited. Secondly, their reliability are compromised by load fluctuations, leading to false alarms even in the absence of actual faults. To address these issues, a novel intelligent approach to diagnose ITSC fault is proposed. Indeed, this method encompasses three core components: a novel multi-sensor fusion technique, a knowledge map, and enhanced Convolutional Neural Networks (CNNs). First, the raw data collected from multiple sensors undergoes a transformation into 2D data using a novel image transformation based on Hilbert transform (HT) and variational mode decomposition (VMD), which is concatenate to a novel information map including frequency fault information and rotational speed. Then, this 3D multi information image is used as input to an improvement CNN model that apply a transfer learning for an enhanced version of SqueezNet with incorporating a novel attention mechanism module to precisely identify fault features. Experimental results and performance comparisons demonstrate that the proposed model attains high performance surpassing other Deep Learning (DL) methods in terms of accuracy.