Deeppipe: an unsupervised model for abnormal condition detection of multi-product pipelines

As one of the main transportation methods for multi-products, pipeline transportation has the characteristics of convenience, efficiency, and safety. Nevertheless, because of the complicacy of the scheduling schedule, the on-site frequent operations lead to a variety of conditions. Only relying on the manual monitoring of the pipeline state cannot accurately distinguish between unsteady conditions and abnormal conditions. At present, due to the limited samples, many detection models can only recognize limited types of abnormal conditions, and cannot judge the unknown anomalies. To overcome the above problems, in this study, an unsupervised model is developed to detect the abnormal conditions of multi-product pipelines with high accuracy. First, to solve the difficulty of limited sample data, a simulated SPS model is established by collecting field pipeline and equipment parameters, generating a large number of simulated unsteady and abnormal samples. Second, to explore the temporal dependencies within the sample data, an unsupervised model based on the LSTM (Long Short Term Memory) with an attention mechanism is built. Third, taking simulated and real pipelines as research objects, the performance verification of the customized unsupervised model is carried out. The results show that the customized unsupervised algorithm can effectively guide pipeline safety management when an abnormal condition occurs.