Sensorless control technology for induction motors has been widely applied in various industrial scenarios due to its advantages of low cost and high reliability. However, significant instability can occur in the low-speed generation region. Traditional signal injection methods are not suitable for induction motors with weak anisotropy and may cause torque ripple issues. The virtual voltage injection method can fundamentally resolve the instability in sensorless control of induction motors in the low synchronous speed region without causing torque ripple problems. Nevertheless, this approach also increases the required d-axis current reference for motor control, leading to higher no-load and light-load losses. To address this issue, this paper integrates artificial intelligence algorithms to construct and train a torque estimation network for induction motor sensorless control under virtual voltage injection. Based on the estimated torque from the network, the d-axis current reference is compensated accordingly, allowing it to adapt to load variations, thereby reducing no-load and light-load losses.