Research On Key Technology Of Condition Monitoring And Fault Diagnosis And Its Application On Plunger Pump In Truck Crane

The safety and reliability of equipments during their service life hasarisen increasing attention in group in the industry and customer. Crane is animportant construction machinery to ensure the successful completion ofmany major infrastructural projects. Since the machinery always works in thesevere working condition, it is one of the special machines with the highestaccident rate. Once the key parts break down, the whole equipment may bedestroyed even whole production, the efficiency will be affected and even thecatastrophic accidents will occur. The real-time condition monitoring andon-time fault diagnosis of crane is an effective measure to avoid it unplanneddowntime and ensure safety in production.The study of condition monitoring and on line intelligence faultdiagnosis of construction machinery is one of the hot spot. However, it stillhas many challenges, such as the real signal recovery from the contaminatedoriginal signal, the effective features extraction, the establishment of real timefault diagnosis model, and the diagnosis model for the absence of faultcondition samples. This dissertation focuses on the key technology ofcondition monitoring and fault diagnosis of construction machinery, and itsapplication on plunger pump in large truck crane in the complicated workingconditions. The main contents are as follows:(1) Due to its good localization and multi-resolution features in thetime-frequency domain, the wavelet transform has been widely used, andvarious signals with different signal to noise ratio (SNR) have correspondingsuitable decomposition levels to obtain effective filtering results. This studydescribes a novel scheme of selecting the appropriate decomposition levelbased on block bootstrap and white noise test. The scheme consists of threemain steps. Firstly, decompose the vibration signal into wavelet domain, andthe correlations between the wavelet coefficients are measured by lag autocorrelations. Secondly, according to the intensity of correlation at eachlevel, either the block bootstrap or general bootstrap procedure is adopted toproduce new pseudo-samples from the original wavelet coefficients series.Finally, as actual signal and noise have different translating characters alongthe levels in wavelet domain, the suitable decomposition level is achievedthrough whitening test on the wavelet coefficients, and the accuracy of thetest is also obtained by the pseudo-samples. The simulation and experimentresults show that the proposed procedure can be used to determine the seemlydecomposition level adaptively and obtain the superior filtering capability forthe signal contaminated with white noise.(2) The threshold is another key factor for wavelet denoising. A novelapproach is proposed by using advanced false discovery rate procedure(AFDR). The main idea is based on controlling false discovery rate (FDR)through combination of all three stepwise procedures (step-up, step-down,step-up-down) and estimation of the number of true null hypotheses. TheAFDR procedure differs from the standard FDR procedure in two respects,i.e., enhancing the efficiency by reducing the number of tested hypotheses andimproving the power. The proposed procedure consists of two main steps:firstly, the signal is represented more parsimoniously in wavelet domain;secondly, a most appropriate stepwise FDR procedure is selected according tothe character of wavelet coefficients. Both the numerical simulation resultsand the experimental results show that the proposed approach is a competitiveshrinkage method compared with other popular techniques.(3) A novel method based on wavelet leaders multifractal features formachinery fault diagnosis is proposed. The multifractal features, combined byscaling exponents, multifractal spectrum, and log cumulants, are utilized toclassify various fault types and evaluate various fault conditions of rollingelement bearing, and the classification performance of each feature and theircombinations are evaluated. Eight wavelet packet energy features are alsointroduced together with multifractal features. Experiments on11fault datasets of bearing show that a promising classification performance is achieved. At the same time, the experimental results show that the classificationperformance of the classifier trained with eight wavelet packet energy featuresin tandem with multifractal features outperforms that of the classifier trainedonly with wavelet packet energy features, time domain features, ormultifractal features, and it is also superior than that of wavelet packet energyfeatures in tandem with time domain features, or multifractal featurescombined with time domain features. The feature selection method based ondistance evaluation technique is exploited to select the most relevant featuresand discard the redundant features, and therefore the reliability of thediagnosis performance is further improved.（4）Promptly and accurately dealing with the equipment breakdown isvery important in terms of reliability and downtime decreasing. A novel faultdiagnosis method based on particle swarm optimization (PSO) and relevancevector machine (RVM) is proposed. The particle swarm optimizationalgorithm is utilized to determine the kernel width parameter of the kernelfunction in RVM, and the two-class RVMs with binary tree architecture aretrained to recognize the condition of mechanism. For multi-fault modesamples, the experiments show that a very sparse diagnosis model is realizedand a promising classification performance is achieved, which is veryappropriate for on-line real-time application. In addition, at the initializationrunning period of the rotating mechanism, the collected samples are alwaysones in the normal condition, and the fault signals only appears after a certainrunning time, so the general fault diagnosis model can not be trainedeffectively. In this paper, a hybrid fault diagnosis scheme for rotatingmechanism is proposed based on ant colony SVDD (Ant-SVDD) and DaviesBouldin index (DBI) K-Cluster method. Firstly, the SVDD model isconstructed for the samples in the normal condition, and the ant colonyalgorithm is utilized to optimize the SVDD parameters. Secondly, when thenumber of rejected samples reaches the given threshold, the K-Cluster methodis employed to classify these samples and the labels is obtained. Furthermore,the number of the classification is determined in accordance with the Davies Bouldin index. Finally, the one class samples are trained with SVDDindividually, and the SVDD classifiers are combined to a complete diagnosismodel based on a binary tree structure. For the multi-fault mode samples ofrolling element bearing, experiments show that a promising classificationperformance is achieved.(5) The system architecture and the database system of the intelligentmaintenance are designed for truck crane, and with the proposed keytechnology of condition monitoring and fault diagnosis, the wavelet transformpreprocessing based condition analysis and PSO-RVM, ANT-SVDD clustermethod based intelligent diagnosis for pump in truck crane are realized. Theenvelope of the signal before and after denoising shows that the signal qualitygets a great improvement, which improves the accuracy of conditionmonitoring. The proposed diagnosis methods are employed in the diagnosis ofplunger pump in truck crane. The six states, including normal state, bearinginner race fault, bearing roller fault, plunger wear fault, thrust plate wear fault,and swash plate wear fault, are used to test the classification performance ofthe proposed PSO-RVM model, which compares with the classical models,such as back-propagation artificial neural network (BP ANN), ant colonyoptimization artificial neural network (ANT ANN), relevance vectormachines (RVM), and support vector machines with particle swarmoptimization (PSO-SVM), respectively. The experimental results show thatthe PSO-RVM is superior to the first three classical models, and has acomparable performance to the PSO-SVM, the diagnostic accuracy achievingas high as99.17%and99.58%, respectively. But the needed number ofrelevance vector is far fewer than that of support vector, and the former isabout1/12to1/3of the latter, which indicates that the proposed PSO-SVMmodel is more suitable for applications that requires low complexity andreal-time monitoring. With the same pump data, the ANT-SVDD model isbuilt for the normal state, and then the DBI K-cluster method is employed forthe novelty samples. The ANT-SVDD classifier for the new labeled samplesis constructed and combined into the total diagnosis model. This technology realizes the intelligent diagnosis for the new fault class.