Research On Anomaly Detection Technique Of Reciprocating Compressors Based On The Transformer Model

With the rapid development of deep learning techniques,it has become a research hotspot to introduce those techniques to industrial production for achieving intelligent transformation.Reciprocating compressors are the key equipments in the field of oil and natural gas.Developing the anomaly detection technology is of great significance to maintain stable equipment operation and ensure production safety.The existing anomaly detection technologies of reciprocating compressors mostly deal with the single time signal,and the effect relies heavily on human experiences.This thesis uses the linear transformation and one-dimensional convolution calculation to perform feature processing on multi-dimensional time series data,and uses the linear layer as the decoder to obtain an improved Transformer model.On the basis of the improved Transformer model,we establish a kind of anomaly detection method for reciprocating compressors by combining the deep learning and dynamic threshold theory.Furthermore,we select the practical production data of the reciprocating compressor unit to analyze the actual application effect of anomaly detection.Through the experimental comparative analysis of the improved Transformer model in the data embedding expressions,optimization algorithm and network depth,we obtain the optimal parameters of the model in the experimental scenarios,and perform anomaly detection on the signals of two types of reciprocating compressors,respectively.The experimental results show that the optimal F1 score is 0.80 for the anomaly detection of vibration signal data,while the optimal F1 score is 0.81 for the anomaly detection of thermal performance signal data.Those results imply that the improved model in this work is effective in dealing with the anomaly detection of reciprocating compressors.