| As an important part of automatic production equipment,once the hydraulic system fails,it will seriously affect the safe and reliable operation of the system.When the deep feature information of system state data mining is used for state recognition,it can significantly shorten the troubleshooting time of hidden faults and improve the monitoring ability of equipment state.For the high-dimensional and complex state data of hydraulic system,the paper extracts the low dimensional features of hydraulic system unsupervised,and classifies them to realize the function of state recognition and fault monitoring.Firstly,a feature selection method based on feature weight is proposed for phased data with obvious state transition or state data with obvious vibration component.After calculating the entropy weight of features,the weight of Hotelling test is used for adjustment test,and a small number of features are selected and reconstructed to realize state recognition.Secondly,in order to make full use of the advantages of multi-channel data and avoid information loss caused by data vectorization,this paper proposes an iterative optimization method based on Tucker decomposition technology of tensor structure data,which improves the stability of core features of tensor data.And the solution of a support tensor machine for tensor data classification is fully discussed.Then,in the face of the difficult to identify problems such as the abnormal working state of the spring coefficient declining of the overflow valve,this paper proposes an unsupervised feature extraction method based on deep learning,applying a two-stage stacked noise reduction coding network(SDAE).This network solves the problem that deep coding networks are difficult to converge,and improves the ability to extract features from high-dimensional complex data.And through a network based on spatial structure maintaining constraints(LP＿SDAE),the relative position of the data in the reconstruction space is restricted from spreading or drifting to improve the stability of the feature.At the same time,by changing the training method of LP＿SDAE,the network is applied to online monitoring of abnormal data,and good monitoring results have been achieved.Finally,in order to verify the effectiveness of the diagnosis method,a hydraulic forging machine experiment was carried out.The method in this paper was applied to the experimental data for feature extraction,and several unsupervised feature extraction methods were compared.The experimental diagnosis results show that the method in this paper has a higher diagnostic accuracy in processing complex data of the hydraulic system. |
