Research On Fault Early Warning And Diagnosis Technic Of Reciprocating Compressor Based On Machine Statistical Learning

The Reciprocating compressors are the core equipment of process industries such as refining and chemical industry.Once the equipment breaks down,it will affect the production benefits of the enterprises,and seriously threaten the personal safety.Therefore,it is of great significance to establish a reliable reciprocating compressor anomaly detection and fault diagnosis system.In the field of anomaly detection of the reciprocating compressors,the effect of the existing fault early warning methods are difficult to be further improved because they cannot accurately identify the changes in the excitation signals of the reciprocating compressors.In the field of fault diagnosis of the reciprocating compressors,due to the fewer types and sources of the existing fault samples,the diagnostic accuracy of the fault diagnosis methods based on supervised learning pattern recognition for the faults of different units is often low.And because this kind of methods can not identify other kinds of faults besides the training samples,their application in fault diagnosis of the reciprocating compressors has great limitations.Aiming at the above problems,the feature space is constructed by using the high-dimensional features of vibration signals of the reciprocating compressors and the Infinite-t Mixture Model is used to fit the feature space in this paper.The fault early warning is realized by analyzing the change of the model.If there is an abnormality,the contribution rate of each feature to the current state is calculated.The accurate fault diagnosis is conducted by analyzing the differences between the feature contribution rate of this state and various fault states.The specific research contents are as follows:(1)Because there are many kinds of faults in the reciprocating compressors and the mechanisms of some faults are not clear,the simple combination of the features extracted by traditional feature extraction methods is often difficult to fully describe the running status of the equipment.Aiming at this problem,a denoising auto-encoder is used for unsupervised learning of features of the vibration signals.The time domain signals under the normal working conditions of the reciprocating compressors are taken as training samples.After determining the network structure of the denoising auto-encoder,the model is trained with these samples.The encoder part of the trained model is reserved for extracting the robustness sensitive features of the reciprocating compressors.(2)A fault early warning method based on Infinite-t Mixture Model for the reciprocating compressors is proposed.The feature space is constructed by using the high-dimensional features of vibration signals of the reciprocating compressors,and the Infinite-t Mixture Model is used to fit the feature space.On the basis of accurately characterizing the distribution characteristics of the vibration signals,the sensitivity of the model to the variation of equipment running status is improved.The normal working condition model is used as benchmark model,and the distance between the benchmark model and the real-time working condition model is accurately measured by the matching based-KL divergence approximation method.The fault early warning is carried out by comparing the distance with the self-learned alarm threshold.The effectiveness of the proposed method is verified by using the actual fault case data of the reciprocating compressors.The results show that the proposed method not only improves the accuracy of the fault early warning,but also greatly advances the time point of the alarm.The effect of the early warning has been significantly improved.(3)An intelligent fault diagnosis method based on feature contribution rate for the reciprocating compressors is proposed.The contribution rate of each feature to the real-time running status is calculated by using the real-time woking condition samples and the benchmark model constructed in the fault early warning process,and the contribution rates of each feature to various faults status are calculated by using each type of the fault training samples and the benchmark model according to the Bayesian contribution rate.The preliminary classification of the fault is conducted by comparing the average distances between the feature contribution rate of real-time running status and the diagnosis models.whether the real-time working condition samples belong to the preliminary matching fault or other types of faults besides the training samples is accurately determined by using the proposed discrimination criterion.On this basis,a self-learning method of the diagnosis model is proposed.The effectiveness of the proposed method is verified by using actual fault case data of the reciprocating compressors.The proposed method has a high diagnostic accuracy for the faults of different reciprocating compressors,and can accurately identify other types of faults besides the training samples.In addition,the proposed method,by virtue of its self-learning ability of the diagnostic model,can automatically modify the diagnosis model by using the samples of erroneous diagnosis,it effectively improves the accuracy of subsequent fault diagnosis.