Study On Fault Diagnosis And Health Status Prediction For Products With Degradation Process

Given that sudden product and system failures always result in significant losses for users and operators,managers urgently need a method to assess the degradation status of products and diagnose faults.Fault diagnosis is the assessment of whether a product has experienced a fault by measuring,recording,and monitoring the degree of deviation from normal operating conditions in real time.The essence of health state prediction is to evaluate the operational and degradation status of a product based on its historical information,and then predict the next stage of degradation for the product.In order to further analyze the fault and health management of products,this dissertation primarily focuses on analyzing two types of product failure modes: soft failures and hard failures.Soft failure refers to the process in which the functionality of a product gradually decreases over time under the combined influence of internal and external factors until it completely fails.Hard failure refers to the situation where a product is unable to perform its intended functions due to a failure.In severe cases,it can disrupt the entire system and result in the product being unable to operate.Meanwhile,based on whether the physical components of the product consist of identical elements,two types of degradation products are distinguished: those with multidimensional heterogeneous distributed characteristic data and those with multidimensional homogeneous distributed characteristic data.This dissertation proposes a fault diagnosis model and health state prediction model framework with failure modes based on products with different physical structures corresponding to different failure modes.The framework is modeled,analyzed,and validated using application bearings and lithium batteries to assess the reliability and generalizability of the proposed methods.The research focus and contributions of this dissertation are as follows:(1)Degradation products composed of components with different physical structures exhibit multidimensional data with heterogeneous distribution characteristics during operation.This dissertation proposes a fault diagnosis model based on these characteristics and the reliability and generalization of the model are validated through bearing experiments.Due to the coupling between bearing components and the high-dimensional data generated by these components,the improved modified ensemble empirical mode decomposition(MEEMD)and Adjusted Mahalanobis-Taguchi system(AMTS)are introduced to extract the main components of fault characteristic information.The modified health index(MHI)is constructed to functionally reflect the health status of bearings by incorporating the main fault characteristic information.Single fault mode experimental results demonstrate that this method overcomes the limitations of health index(HI),such as the generation of Na Ns values(Not a number)during degradation calculation,which prevents HI from accurately reflecting the health status of bearings under single fault mode.On the other hand,when facing the occurrence of a failure in a specific component within the bearing under multiple fault mode,a deep neural network(DNN)is incorporated as a classifier for the main fault characteristic information to trace the location of the faulty component.This model can guide maintenance personnel in promptly replacing the correct faulty component,ensuring the continued reliable operation of the product.The experiments demonstrate that this model can effectively reduce the cost of losses caused by component failures and provide real-time warnings for potential failures of the product at a micro level.(2)This dissertation proposes a health state prediction model for degradation products composed of components with different physical structures.It predicts the degradation status of bearings and evaluates their operational condition.The specific manifestation of hard failure in bearings is the inability to function properly.By employing MEEMD and different prediction strategies,characteristic data reflecting the degradation status are extracted.Combining the degradation characteristics of the bearings,the long short-term memory(LSTM)neural network is utilized to predict the next stage of degradation.On the other hand,when bearings experience soft failure,the operating data of each cycle is transformed into Mahalanobis distance(MD).Combined with the MHI method,this allows for differentiation of different operational states of the bearings.Then,the LSTM neural network is employed to predict the next stage of degradation for the bearings.Experimental results indicate that this method can effectively capture the sub-healthy states within the operational status of the bearings.Furthermore,the application of this method assists managers in timely detecting product degradation,thereby enabling early warnings and future estimation of the product’s degradation status.(3)When a product is composed of identical components,the functionality of multiple identical parts can be stacked to enhance the product’s capabilities,and the multidimensional data generated during the operation of the product exhibits homogeneous distribution characteristics.This dissertation designs a fault diagnosis model for lithium batteries based on the data characteristics mentioned above and conducts simulation analysis.The lithium battery used in new energy vehicles is a collective of multiple individual cells,forming a complex nonlinear and time-varying system with various inconsistencies.When batteries do not exhibit obvious abnormalities,it becomes challenging to detect early-stage internal failures.This dissertation proposes a fault diagnosis model based on voltage inconsistency,utilizing a hybrid approach of multiscale permutation entropy(MPE)and coefficient of variation(CV)rules.Firstly,the dataset of individual battery cells collected by sensors is cleansed,and their MPE values are calculated.Then,based on the 3-sigma approach,fault diagnosis thresholds for individual battery cells are designed.Finally,the CV rule is utilized to locate the faulty cell.The results indicate that this method can effectively reduce false alarms caused by data errors and enhance real-time fault warning efficiency for new energy vehicle companies.(4)This dissertation proposes a health state prediction model for degradation products composed of identical physical structure components,specifically for predicting the degradation status of lithium batteries.First,the raw data of lithium batteries is decomposed into intrinsic mode functions with different time scales using the MEEMD method.Then,the mean impact value(MIV)neural network method is applied to filter out noise.The denoised intrinsic mode functions are treated as the potential feature information for the operational state of the lithium-ion battery and are inputted into a bidirectional long Short-term memory(Bi-LSTM)neural network prediction model to forecast the degradation status.The results demonstrate that the proposed health state prediction model performs well in terms of accuracy and stability for different types of lithium batteries.On the other hand,the data collection process for lithium batteries can be complicated.When the sensors fail to collect complete battery capacity data,the measurement of voltage,current,and temperature can be utilized as influencing factors for predicting the degradation status and these three factors collectively contribute to approximately one-third of the battery’s health state.The results indicate that managers can achieve a balanced operational quality index for the product by monitoring other controllable factors,ensuring the stable operation of the product.