Time–Frequency Map-Based Abnormal Signal Detection

Abnormal signal detection plays a significant role in monitoring the state of running machinery. Most of the previous methods deem the abnormal signal detection an issue of signal analysis, but this strategy may be ineffective due to the high noise in the original data. Aiming to solve this problem, this study regards abnormal signal detection as an issue of image classification, such that advanced image noise removal and feature representation methods can be investigated with the aim of improving abnormality detection accuracy. Excellent deep learning-based image classification methods are investigated in this study, which can directly output classification results using original data and avoid the interference of the artificial factor. However, the deep image network may be unsuitable to the nature of the vibration signal, which significantly affects the generalization and accuracy of detection results. Aiming to solve this problem, a novel deep architecture is proposed in this study to improve abnormal signal detection in terms of correctness and efficiency. Consequently, a novel abnormal signal detection method is proposed by combining the time-frequency map and our deep learning architecture. Experimental results demonstrate that the time-frequency map can extract representative abnormality features, and our deep learning method can improve the classification capability. The detection correction of our method is observed to be consistently higher than 99.90% for different databases, and the time cost of our method is bearable at the same time. The main contribution of our method lies in that it investigates and optimizes advanced deep learning architectures for abnormal signal detection.