Equipment Health Condition Monitoring And Prognostics Techniques Using Echo State Networks

Equipment security is an important part of social public security and with the systematicand complicated increasingly of modern industrial equipment, if a fault occurs, is bound tohave some effect on industrial production, or even result in casualties. Therefore, it can notonly prevent the happening of accidents but also eliminate a hidden danger and reduce loss totimely monitor and predict the healthy status of the equipments. An equipment healthmonitoring and prediction method based on echo state networks has been studied in this paper.The main contents are as follows:The fundamental algorithm and optimization algorithm of the echo state network (ESN)have been studied. The apply of ESN in terms of time series forecasting and time signalclassification was discussed through numerical simulation experiment. Experimental resultsindicated that: the size of ESN network’s dynamic memory (DR), the initial states of DR andsignal noise can all effect the ESN’s performance of classification and prediction. The biggerDR, the smaller error of the ESN’s prediction. Because input weights, DR weights andfeedback weights matrix were all randomly generated, in the training process, abandon someinitial output prediction of ESN can improve ESN prediction performance, reduce predictionerror and increase the accuracy of the long-term forecasts; in addition, the network robustnesscould been improved by adding appropriate noise in the ESN training.The extraction method of the features of rolling bearing vibration’s time domain andfrequency domain has been studied in this paper. Experiments showed that when damageappeared,the peak and kurtosis etc. time domain parameters would increase and differenttypes of injury time domain index parameters have significant difference between the twogroups; in addition, different types of damage’s vibration signal was decomposed by waveletpacket transform and the energy distribution also show different characteristics; Therefore,extraction the vibration signal’s time and frequency characteristics could describe fault stateand reduce the dimension of characteristic parameters.Equipment fault diagnosis method based on ESN has been studied. A diagnosis modelbased on the ESN was set up using extracted characteristic parameter as the input vector, and this model could effectively identification rolling bearings’ different damage types.Experiments showed that, for the same ESN diagnosis model, with wavelet packet energydistribution as ESN input of the network could achieve a higher diagnosis accuracy, and timedomain parameters as ESN network input had good real-time, but classification accuracy wasslightly inferior to the former.Rolling bearings health condition monitoring and forecasting method based on waveletpacket entropy was proposed in this paper. wavelet packet entropy is a given time interval willsignal wavelet packet decomposition, with its energy distribution characteristic parameters asthe information entropy. wavelet packet entropy is volatility clustering and time-varying. Thevolatility clustering means using only one parameter, entropy to describe a given intervalsignal characteristics. Time-varying can characterize the interval of time signal different statevariable change. The time-varying of wavelet packet entropy can effective monitor equipmentstate’s change, and keep effective surveillance on the equipment’s health. Forecasting thechanges of wavelet packet entropy by ESN could predict the change trend of the equipments.