Fault Warning Of Reciprocating Compressor Based On Spatiotemporal Multivariate Information Fusio

Reciprocating compressors are important equipment for offshore petrochemical platforms.Foreseeing whether the compressor fails in advance and ensuring its smooth operation is of great significance for the high efficiency of industrial production and the personal safety of maintenance platform personnel.Based on the actual collected data of three reciprocating compressors on an oil platform,this paper analyzes the multidimensional data and takes the time and space characteristics of the data as the overall research direction to realize the prediction and fault diagnosis of the compressor’s overall operating status.The specific research content is as follows:The multi-dimensional time series data of reciprocating compressors has rich time and space information,so obtaining the time and space characteristics of the data is an important means to realize prediction and diagnosis.To solve this problem,firstly,a recursive self-attention mechanism,which is more suitable for time series processing,is proposed to realize the spatial feature extraction of single column data.The model is completely based on the information of the data itself.The data is segmented in time sequence and then local self-attention is calculated.The result is transmitted to the data information of the next time period through recursion,and finally the feature fusion is realized through the fusion gate.Experiments show that this method performs well in single-column prediction and is suitable for local spatial feature extraction.In addition to obtaining the original internal features of the data,the combination of compressor multi-dimensional data features and advanced model features can achieve deeper spatiotemporal feature extraction instead of relying only on recursive operations.Therefore,the Attn-LSTM model is proposed on the basis of classical attention.This model combines the attention mechanism with the LSTM model.After spatial features are obtained through spatial attention calculation,the spatial features of different periods are combined through temporal attention,and finally the final predicted value is obtained by decoding and computing.Experiments designed the health curve of reciprocating compressor to reflect the compressor’s operating conditions and use it as the predicted output of the model.Finally,integrating recursive self-attention and Attn-LSTM models,multisegment attention-LSTMs(MA-LSTMs)are designed as the final prediction and diagnosis model.The model first divides the data into three parts: motor,valve,and crankshaft,and uses recursive self-attention mechanism to acquire their internal data features to achieve spatial local attention calculation,and then uses spatial global attention and time in Attn-LSTM Attention operation and decoding calculation realize the prediction of compressor operating state,and output three local state curves at the same time to realize local health monitoring.Experimental results show that the model performs better than popular models in all aspects,and it can predict whether the fault will occur and determine where it occurs.