توصيفگر ها :
يادگيري ماشين , خوشه بندي , نابالانسي , ناهم محوري , يادگيري عميق
چكيده فارسي :
با پيشرفت روزافزون علم، روش¬هاي نويني جهت پايش وضعيت سيستم¬ها جايگزين روش¬هاي سنتي شده¬اند. اين روش¬ها باعث افزايش طول عمر سيستم شده و همين¬طور از خسارات جبران¬ناپذير جلوگيري مي¬شود. بنابراين هم نيروي انساني كاهش مي¬يابد و هم از لحاظ صرفه¬ي اقتصادي باعث صرفه¬جويي زياد در هزينه¬ها مي¬شود. براي تشخيص عيوب در سيستم¬هاي دوار، از روش¬هاي يادگيري ماشين مي¬توان استفاده كرد. يادگيري ماشين به دو شكل نظارت¬شده و غير نظارت¬شده است. استفاده از يادگيري ماشين غير نظارت¬شده باعث هوشمندي بيشتر در پيش¬بيني عيب مي¬شود. براي رسيدن به اين هدف از خوشه بندي خودكار استفاده شده است. در اين روش دو شاخص اعتبار جهت خوشه بندي به صورت خودكار و همين¬طور تركيب خطي آن¬ها به عنوان توابع هزينه يا هدف تعريف شده است. ويژگي¬هاي كورتوسيس، اسكيونس و پارامتر كرست از سري زماني داده¬ها گرفته شده است. بعد از استخراج ويژگي¬ها، آن¬ها به صورت يك آرايه به عنوان ورودي به الگوريتم خوشه¬بندي خودكار داده مي¬شود. در ادامه با استفاده از دو الگوريتم¬ فراابتكاري ازدحام ذرات و كلوني زنبور عسل به بهينه¬سازي توابع هدف پرداخته شده است. در اين پژوهش از دو الگوريتم جهت مقايسه¬ي نحوه¬ي خوشه¬بندي و مقدار تابع هدف استفاده شده است. ¬با استفاده از اين روش مجموعه¬اي از خطاهاي يك سيستم را مي¬توان در نظر گرفت و پايش وضعيت بر روي سيستم انجام داد. دو خطاي رينگ داخلي و رينگ خارجي در ياتاقان¬هاي غلتشي در نظر گرفته شده است وسپس ويژگي¬هاي مورد نظر استخراج گرديده تا به عنوان ورودي به الگوي تشخيص عيب داده شود. بعد از بهينه¬سازي انجام شده در اين روش توسط الگوريتم¬هاي فراابتكاري، خوشه¬بندي خطاها به صورت خودكار و مجزا انجام گرديده است. از ديگر عيوب موجود در سيستم¬هاي دوار مي¬توان به عيوب نابالانسي و ناهم¬محوري اشاره كرد. براي تشخيص اين عيوب هم به استخراج ويژگي نيازمند هستيم كه در يك آرايه جمع¬آوري مي¬شوند. اين آرايه به عنوان ورودي به يك شبكه¬ي يادگيري عميق جهت طبقه¬بندي داده¬ها داده مي¬شود. بعد از آموزش شبكه، با استفاده از داده¬هاي تست به آزمايش آن پرداخته شده است. ماتريس درهم¬ريختگي و منحني مشخصه عملكرد سيستم نيز براي داده¬هاي تست و آموزش به عنوان خروجي از شبكه در نظرگرفته¬شده است.
چكيده انگليسي :
With the increasing progress of science, new methods to monitor the status of systems have replaced traditional methods. These methods increase the lifespan of the system and prevent irreparable damage. Therefore, this work will reduce both the manpower and, in terms of economic efficiency, it will save a lot of costs. By using vibration analysis on the systems, it is possible to find out the existence of defects in them. After collecting data in the time domain, we need to check the frequency domain. One of the most basic methods for fault diagnosis is to display the spectrum of vibrations in the frequency domain or frequency spectrum. The frequency spectrum includes the engagement frequency and its side bands. There are always side frequency bands around the engagement frequency and its harmonics, which are caused by amplitude or frequency modulation or both simultaneously. With the growth of defects, the amplitude of vibrations in the sideband frequencies also grows. But the frequency spectrum usually consists of many different harmonics. These harmonics have sidebands around them. The more complex the components of the system, the more the number of sidebands, which makes the frequency spectrum more complex. Therefore, the use of Cepstrum's method separates different harmonics from each other. One of the important advantages of the cepstrum compared to the spectrum is that with a small change in the position of the measured point, the overall shape of the spectrum changes a lot, while the cepstrum is less sensitive to this issue. In addition, the envelope can also be used due to its low sensitivity to random fluctuations or noise. But after detecting the defect using vibration analysis, machine learning methods can be used to increase the power in detecting defects. There are two forms of machine learning, supervised and unsupervised. The use of unsupervised machine learning leads to greater intelligence in fault prediction. To achieve this goal, automatic clustering has been used. In this method, two validity indices for automatic clustering and their linear combination are defined as objective functions. The characteristics of kurtosis, skewness ,and crest factor are taken from the time series of the data. After extracting the features, they are given as an array as input to the automatic clustering algorithm. In the following, two meta-heuristic algorithms of particle swarm and artificial bee colony have been used to optimize the objective functions. By minimizing the objective functions, clustering is done automatically. In this research, two algorithms have been used to compare the clustering method and the value of the objective function. By using this method, a set of system errors can be considered and the status of the system can be monitored. Two faults of the inner race and outer race in rolling bearings have been considered and then the desired features have been extracted to be given as input to the fault diagnosis model. After the optimization done in this method by meta-heuristic algorithms, the clustering of errors has been done automatically and separately. Other defects in rotating systems include imbalance and misalignment. To detect these defects, we need to extract features that are collected in an array. This array is given as input to a deep-learning network for data classification. After training the network, it has been tested using test data. The confusion matrix and the characteristic curve of the system performance are also considered for the test and training data as the output of the network.
