Data generated by sets of sensors can be used to perform predictive maintenance on industrial systems. However, these sensors may suffer faults that corrupt the data. Because the knowledge of sensor faults is usually not available for training, it is necessary to develop an agnostic method to learn and detect these faults. According to these industrial requirements, the contribution of this paper is twofold: 1) an unsupervised method based on the successive application of specialized anomaly detection methods; 2) an agnostic evaluation method using a supervised model, where the data labels come from the unsupervised process. This approach is demonstrated on two public datasets and on a real industrial dataset.