Recent advances in sensors for real-time continuous glucose monitoring (CGM) and continuous subcutaneous insulin infusion pumps have been the treatment of type 1 diabetes mellitus with good results for the pediatric and adult population. However, on occasion these devices can fail, which can expose patients to serious risks, especially at night, such as inadequate insulin administration. This work proposes a method for real-time detection of faults in glucose sensors using deep neural networks. The neural classifier receives as input the error generated by time windows of the glucose sensor, with this information the failure classification is generated to generate an alert that can be used as an alarm to mitigate possible risk scenarios for the patient. The method used is tested with real data obtained by glucose sensors for continuous monitoring, the failures were added synthetically. Data aggregated as failures were generated randomly and within normal ranges of glucose samples. Three deep neural network models were used, the Long Short-Term Memory (LSTM), a convolutional neural network (CNN), and a Long Short Term Memory Fully Convolutional Network (FCN-LSTM). The results obtained by the neural networks are good, but the FCN-LSTM neural network shows a classification accuracy of up to 99.8 \(\%\) on real data, indicating that it can significantly improve the safety of patients with type diabetes 1, especially at night, by effectively detecting glucose sensor failures.