27602

تخمين طول عمر ياتاقان گردشي با استفاده از دادههاي ارتعاشي با يادگيري عميق

كارشناسي ارشد

مهندسي برق گرايش كنترل

1398

1401/06/22

سعيد عبادالهي

مهندسي برق

با ورود دانش علوم داده در صنعت، روش تعميرات و نگهداري نيز تغيير پذيرفته است. دستگاههايي كه پيشازاين با فلسفۀ تعميرات پيشگيرانه نگهداري ميشدند حالا به كمك علوم داده ميتوانند با فلسفۀ تعميرات پيشبين نگهداري شوند كه افزايش بهرهوري دستگاهها با كاهش هزينهها را به ارمغان ميآورد. تخمين طول عمر دستگاه از زيرمجموعههاي تعمير و نگهداري پيشبين است كه به كمك دانش يادگيري عميق در حال بررسي است. در اين پروژه، هدف بررسي تخمين طول عمر دستگاه به روش يادگيري عميق براي ياتاقان گردشي است كه از جمله قطعات مهم در دستگاههاي دوار در خط توليد است. دادههاي ارتعاشي ياتاقان براي انجام تخمين طول عمر حائز اهميت است و در اينجا به كمك دادههاي بانك اطلاعات مركز سيستمهاي هوشمند نگهداري دانشگاه ميشيگان اين تحقيق انجام شده است. دادههاي ارتعاشي پس از كرتبسازي در محيط پايتون شبيهسازي شدهاند. در اين تحقيق باتكيهبر روشهاي مطالعه شده در منابع ديگر ابتدا تبديل از زمان به فركانس صورت ميگيرد كه روشهاي تبديل فوريه زمان كوتاه و تبديل موجك پاكت شدة گسسته بررسي شدهاند. سپس بهمنظور ارائۀ استخراج ويژگي مناسب جهت پيادهسازي شاخص سلامت كه براي تخمين طول عمر استفاده ميشود، روشهاي آماري ميانگين مجذور مربعات و كشيدگي )كورتوسيس( مورد بررسي واقع شده و براي ادامۀ تحقيق از كشيدگي استفاده شده است. در ادامه براي تخمين بهتر از سه روش هموارسازي داده بر پايۀ رگرسيون استفاده شده كه اين. روشها رگرسيون وزني محلي، مدل فرايند تأخيري گاوسي و گاوسي معين هستند و پس از آن روشهاي يادگيري عميق حافظۀ كوتاه بلندمدت، تركيبي از اين روش با لايۀ كانولوشن، تركيب با لايۀ حافظۀ كوتاه بلندمدت سلولي و نيورالپرافت مورد بررسي قرار گرفتهاند. در انتها براي تحليل روشهاي تخمين از نمودار سهبعدي كاركرد دستگاه، طول عمر دستگاه و شاخص سلامت استفاده شده است تا نتايج بهدرستي با يكديگر مقايسه شوند. نتيجهگيري در اين تحقيق نشان ميدهد كه روشهاي يادگيري عميق متداول نيازمند به ميزانسازي خاص بهمنظور جلوگيري از بيش برازش هستند؛ اما با روش نيورالپرافت كه بهتازگي ارائه شده است ميتوان با يك ميزانسازي ساده تخمين درستي از عمر دستگاه ارائه كرد.

1401/09/29

Remaining Useful Life estimation of Rolling Element Bearing using vibration data via Deep Learning

9/13/2023 12:00:00 AM

In this day and age, the constant usage of equipment has resulted in lower efficiency and on time maintenance of the asset could increase the performance effectively. Preventive maintenance is introduced to plan the on-schedule maintenance of every equipment in the industry however, the implementation of this technique has been a costly approach in maintenance. With the help of data that is produced through the state-of-art hardware as well as top notch infrastructure, the cost of maintenance could be reduced. Predictive maintenance is the strategy that relies on data to plan the most optimal strategy for the maintenance of the equipment. Condition monitoring with the help of vibration sensors is one of the main sources of predictive maintenance strategy and more specifically, the estimation of the Remaining Useful Life of a machinery. In Remaining Useful Life of a system, with the help of Run-to-Failure data the health of the equipment can be monitored. Rolling Element Bearing is an equipment which is widely used in the industry and in order to estimate the Remaining Useful Life of a Rolling Element Bearing, a Run-to-Failure experiment is required in order to collect a proper data for estimation. This experiment could be done using a proper Data Acquisition card and vibration sensors attached to the bearing itself. After the acquisition of the proper data, this data needs to be processed using techniques in order to prepare it for proper investigation. This can be done using a programming language which helps with the cleaning and conversion of the data to a proper format. In this project, the format used for all of the data is .csv. Furthermore, the process of preparing the the data continues with converting the records into a numeric representation which helps with investigating the data on a bigger scale. Since there are many recordings logged in one sample file, a numeric representation is required to generalize the Run-to-Failure experimental data. With the help of statistics, a number of approaches could be taken such as Kurtosis or Root Mean Square. Either technique have shown promising performance in outputting a sound result thus Kurtosis was chosen for this project. Vibration data is actually a signal therefore in order to estimate the behavior of the signal and reach to any conclusion about it, signal processing techniques are required to convert the signal from time domain to frequency domain. For this conversion, Discrete Wavelet Packet Transform or Short Time Fourier Transform could be used but through the literature review the former was chosen for the investigation. Now data is preprocessed for the estimation and it requires further generalization to create a model. Smoothing is a technique used in time-series analysis to generalize the series of data for prediction. Many regression techniques could be implemented here and with the introduction of estimation of sample using Gaussian distribution, in this project Exact Gaussian Regression as a method that has never been used in Bearing data was chosen. The output of the smoothed data is called the health index of the bearing. Health index is the indicator of where the machinery has shown the signs of failure and break. Remaining Useful Life estimation requires Deep Learning methods which could be used for time-series analysis and there are plenty of hybrid and conventional techniques available to estimation the data. In this project the main focus was to find a new way to estimate the Remaining Useful Life of Rolling Element Bearing and after further investigation through the latest papers, NeuralProphet was the technique that has never been implemented in any investigation. This technique is an easy-to-use Deep Learning method which is not prone to overfitting which is a common problem with the other Deep Learning techniques.

تعمير و نگهداري پيشبين , علوم داده , يادگيري عميق , رگرسيون

Predictive Maintenance , Data Science , Deep Learning , Regression

Armin Alipour

Saeed Ebadollahi