Research On Data Optimization Strategy And Intelligent Diagnosis Method For Equipment Health Management

Mechanical equipment is an indispensable core production element in industries,crucial for production safety,enterprise profitability,and industrial development.It is also a vital lever for enhancing industrial modernization.As China’s industrial sector continues to upgrade,greater demands are placed on the stability,reliability,and safety of industrial equipment operation.In response to the uncertain risks of equipment maintenance failures,intelligent maintenance has emerged,providing essential support for national industrial upgrading and enhancing industrial core competitiveness by prognostics and health management(PHM).However,challenges such as data redundancy,loss,and imbalance in the growing big data of condition monitoring,which PHM relies on,pose obstacles in data storage,transmission,processing,and equipment health assessment,diagnosis,and prediction.Therefore,research on data optimization strategies and intelligent diagnostic algorithms for equipment health management is of great significance.Issues of data redundancy,loss,and imbalance in equipment PHM stem from the data sampling process in condition monitoring.Addressing these issues,this thesis takes rolling bearings,a common critical component in mechanical equipment,as the research subject.Starting from the analysis of big data generated by vibration data in PHM,the study categorizes the challenges into sample redundancy and sample loss caused by the sampling interval,as well as sampling redundancy resulting from high-frequency sampling and sample length.Correspondingly,predictive dynamic sampling strategies and high-frequency data compression models based on condition sequence transformation are constructed to optimize the sampling process and sampling results,alleviating problems in monitoring data.On this basis,during the healthy phase,anomaly detection models incorporate quantitative assessment of time-varying health conditions within safe regions,reducing safety risks and enhancing production efficiency.During the anomaly phase,the combination of condition sequences with spiking neural networks(SNN)is used to improve the extraction of common features between simulation and real data,constructing an SNN-based virtual-real transfer diagnostic model to address fault diagnosis issues under conditions of scarce real data.The main contributions of this thesis include:(1)Adaptive dynamic adjustment method of sample interval for the slow degradation process.To address the insufficient condition adaptability of data sampling methods,we summarize and refine key issues in existing research on condition-based irregular interval sampling strategies,focusing on sample intervals.Subsequently,a condition-adaptive monitoring and sampling framework tailored for slow degradation processes is introduced.Based on this framework,a dynamic sampling adjustment method is devised for degradation condition monitoring.To address the absence of sampling adjustment objectives,sampling target hyperparameters are introduced.To improve irregular time series prediction,piecewise cubic Hermite interpolation is employed.The issue of ignoring the trend of condition changes in sampling adjustment is resolved through the application of differential autoregressive moving average.Moreover,a variable swapping technique is devised to overcome the difficulty of directly predicting the sample interval.Through targeted improvements to existing problems,the proposed method optimizes or even eliminates the problems of sample redundancy and sample loss in sampling.(2)Nyquist sampling space-based vibration data compression method.To address the issue of data redundancy in vibration data after sampling,a binarization approach is introduced in fault mechanism-based methods to achieve compression of sample point size dimensions based on condition sequence transformation.Subsequently,by integrating different types of fault mechanism-based compression methods in the Nyquist space,a three-dimensional compression method for vibration data is proposed,resulting in a significant improvement in data compression ratio.Furthermore,a new diagnostic performance evaluation method is constructed by considering the relative performance of different compression methods,facilitating effective evaluation of diagnostic-oriented data compression methods’ diagnostic performance.By mitigating the mutual influence between different compression methods,the proposed method significantly reduces the space occupation of data while retaining critical fault information in the original data.(3)Vibration data compression method across sampling spaces.Addressing the inefficiency and inadequate security of the three-dimensional compression method for vibration data,as well as the difficulty in effectively expanding in sub-Nyquist space,a rapid condition sequence transformation is achieved through a clustering-based binarization approach.This is combined with discrete wavelet transform filtering and secondary binarization to accomplish down-sampling based on anti-aliasing filtering,while maintaining the data properties of the condition sequence in compression sensing by constructing a partial permutation orthogonal matrix.The proposed vibration data compression method across sampling spaces broadens the scope of data compression and further improves compression ratio.Additionally,by eliminating amplitude information and hiding spectral information,dual encryption in both time and frequency domains is achieved to attain faster compression speed,higher compression ratio,and enhanced security.(4)Anomaly detection method integrated with the condition evaluation of safe region.To address sudden anomalies caused by the steadiness of sampled data during the healthy phase,auxiliary information reflecting the time-varying degradation characteristics within the safe region is introduced.By reconstructing the solution process of the safe region based on support vector data description,the principle of solving the support vector data description model under the condition of determining the center of the hypersphere is derived.In model construction,the monotonicity and distinctiveness of health metrics are simultaneously considered.Anomaly detection and assessment of healthy conditions within the safe region are unified under the same framework,achieving dual-task learning based on the univariate support vector data description model.This alleviates the increase in maintenance costs and safety risks caused by sudden anomalies in anomaly detection.(5)Condition sequence-driven virtual-real transfer fault diagnosis method.Addressing the issue of fault diagnosis under conditions of scarce real data,a simulation model based on fault mechanisms is employed to construct a dataset comprising abundant simulation data and limited real data.To mitigate information loss during the encoding process,the concept of condition sequence transformation,rooted in spectral clustering,is employed.Additionally,an integrated discharge leakage model is introduced to optimize the feature learning of compressed data.Subsequently,the combination of condition sequences with a transfer diagnostic model based on SNN is employed to improve the extraction of common features between simulation and real data.The proposed method,by incorporating the idea of virtual-real transfer diagnosis,aids in overcoming the challenge of learning fault characteristics under conditions of scarce real data,thereby enhancing the overall performance in addressing tasks involving virtual-real transfer diagnosis.