Sparse Coding Based Machine Condition Recognition And Its Application In The Condition Monitoring Of A Heavy Roller Grinder

The safety, reliability and maintainability of equipment during their service lifeare issues of critical importance to be solved by the equipment manufacturing industryof China. Condition monitoring and the resulting condition based maintenance (CBM)are effective approaches in achieving those goals, so that it has become a value addedto be ?ourished by the equipment manufacturing companies, apart from the qualityof products in design and manufacturing. The condition monitoring of some basicmanufacturing equipment, such as the heavy roller grinder, is especially important,in which more challenges also exist: the disturbances in the monitored signals undermachining process, the problem of feature extraction from long sampling signals andsome inherent disadvantages in the feature based machine condition recognition. Witha goal of solving the problems induced by the special mechanical structure and work-ing condition of the heavy roller grinder, this dissertation does an in-depth study in therelated theoretical methods, experiments, and technical applications, which includethe following detailed items:1. In order to effectively extract the information carried by the monitoring data,an information processing principle existing in the biological sensory system,“redundancy reduction”, is taken for reference. Sparse coding is introduced foranalyzing vibration signals. Taking advantage of its good computing properties,an adaptive feature extraction framework is proposed. Following this frame-work, a redundant dictionary is firstly built by fusing several sub-dictionariesrespectively learned from different classes of monitoring data, and then the s-parse representations of monitoring data are to be solved with the redundantdictionary, which further leads to the extraction of sparse features from them.To build the redundant dictionary is in essence to accumulate prior knowledgeabout different machine conditions, and the extracted sparse feature becomesan index quantizing the existence of a certain machine condition. In the ex-perimental verification with the standard bearing vibration data, the extracted sparse features exhibit good discriminability, and the built redundant dictionaryof bearing condition can also adapt to the vibration signals under varied workingloads.2. The feature based machine condition recognition also has some inherent disad-vantages, such as the“inevitable misdiagnosis”problem,“lacking of historicalmonitoring data”, etc. For the“inevitable misdiagnosis”problem, by combingthe sparse features and the Self-Organizing Map (SOM), a model for visual-ly recognizing the machine condition is proposed. The main idea is to build amachine condition map by taking advantage of SOM's unsupervised clusteringability and its visualizability to represent the machine condition space built withsparse features. According to the mapped position of the monitored signals onthe machine condition map, a maintainer is able to judge the machine condi-tion by himself, and also able to discriminate the ambiguous mapping results.Further, with help of the novelty detection ability of SOM, a mechanism forupdating the proposed model is founded, so as to improve its diagnostic abilitywith the monitored data.3. For the problem of how to overcome the disturbances under the manufacturingprocess, the blind separation technique based on sparse component analysis isstudied, and a single-channel blind separation technique for vibration signal isproposed based on matching pursuit algorithm. This technique follows a basicprinciple in separating the vibration signals, which lessens the ambitions of blindsource separation to more realistic but useful blind component separation: pe-riodic, random nonstationary and random stationary. By projecting a vibrationsignal into an overcomplete dictionary composed by Gabor atoms and waveletpacket atoms, such three components can be successfully separated. It is veri-fied by experiments that the proposed technique can not only capture the weaktransient vibration signatures during the early stage of a developing fault, butcan also further separate the random nonstationary component by clustering theactivations of the wavelet packet atoms.4. With regard to the condition monitoring needs of the heavy roller grinder, the system architecture for it is proposed, and further the proposed methods areapplied. First, with special attention on the machine condition information pro-cessing and fusing, the system architecture is proposed by composing on-sitediagnostic equipment and a remote monitoring and control center. The sens-ing and data acquisition scheme are proposed and implemented, so that a dataacquisition system is set up. The vibration data under no-load condition anddifferent machining conditions are acquired. Three typical vibration signaturesare captured from the vibration data: gear meshing, apex sliding and grindingchattering. After analyzing these vibration signatures, a condition recognitionmodel is proposed based on the studied signal processing and condition recog-nition methods. Technically, the data interchanging and application invokinginterfaces in the condition monitoring system are standardized so as to ensureits extendibility, and the updating mechanism of the condition recognition mod-els are also founded so as to enable the improvability of their diagnostic abilityduring the grinder condition monitoring process.A prototype condition monitoring system for heavy roller grinder is achievedbased on the above theoretical and technical researches. During its monitoring appli-cation, a misalignment between the apexes of the header stock and the tail stock issuccessfully detected. And, the detection of chattering in grinding enables the workerto adjust grinding parameters to make sure the grinding quality.