詢問式機器學習與深度學習之馬達故障診斷與預診斷

Abstract

故障診斷與預診斷技術存在許多挑戰，例如:(1) 深度學習與機器學習模型應用在故障診斷與預診斷的技術雖可實現，但仍須改進該技術以提高模型的準確性和適用性(2)對於故障之類型的辨識，特別是如何提前診斷運行中發生的故障，選擇方法的標準尚未制定(3)欠缺可視化的工具來輔助優化輸出故障診斷資訊及提供決策。因此本論文運用網頁架構實現對故障診斷資訊之可視化輸出及決策展示，將多種機器學習與深度學習模型結合詢問式學習來辨識故障類型，以提高模型準確性。筆者曾在2019 年參與思納捷科技公司「馬達監測系統」的開發，這套系統的故障診斷與預診斷方法是運用機率大小讓廠商了解設備運行狀態。為了讓故障診斷與預診斷技術更精進，且能輔助人快速辨識複雜馬達系統的多故障類型，本論文利用詢問式機器學習和深度學習可自主學習、自適應和處理複雜模式的特性，輔助判斷故障類型與狀態，以實現節約維修成本的目標。實驗一開始從真實馬達數據中取樣，再將取樣後的資料依據馬達常見故障類型作分類，以形成故障診斷資料集。在故障預診斷方面，運用浴缸曲線與取樣資料做搭配，形成故障預診斷資料集，再將各資料集視為感測訊號並輸入改良後的人工智慧演算法以呈現真實情境，實驗結果顯示在一維多特徵馬達頻譜訊號和詢問式監督式機器學習與深度學習演算法搭配下，馬達故障狀態檢測系統可以更快速且更精準辨識馬達的故障類型，該系統可彌補人對於訊號細微變化難以察覺、反應速度較電腦慢和無法全天候監控與故障診斷的缺陷。此外實驗結果也體現詢問式學習可提升大多數監督式機器學習與深度學習演算法的抗雜訊能力，這一特性對於當感測器數據常收集到噪聲、錯誤與冗餘信號時特別有幫助。換言之，即便該數據資料因人為或機械因素而參雜錯誤訊號，詢問式學習仍可以運用數據資料做更具可靠性的決策與辨識。實驗結果也顯示詢問式學習可以在不改變深度學習類神經網路模型的前提下，能提升該網路模型的分類效果。實驗結果也透過對模型參數變化製圖、可視化資料集內容以及人工智慧檢測報告的功能，來達到可解釋人工智慧的目的。也希冀本論文開發的故障診斷與預診斷技術可作為學者往後選擇方法的參考標準。

There are many challenges in fault diagnostics and pre-diagnostics techniques, such as:(1) Although the application of deep learning and machine learning models in fault diagnosis and pre-diagnosis technology can be developed, this technology still needs to be improved to increase the model accuracy and applicability. (2) For the identification of fault types, especially how to diagnose operational faults in advance, the standard of selecting the method has not yet been established. (3) There is a lack of visualization tools to assist in optimizing the output of fault diagnosis information and providing decision-making. Therefore, this thesis applies web page structure to realize the visual output and decision-making display of fault diagnosis information and combines multiple machine learning and deep learning models with query-based learning in order to identify fault types, which can help improve the model accuracy. The author participated in the development of the 'Motor Monitoring System' of InSynerger Technology Corporation in 2019. In order to make the fault diagnosis and pre-diagnosis technology more refined, and to help people quickly identify the multiple fault types of complex motor systems, this thesis utilizes main characteristics in query-based machine learning and deep learning, which are self-learning, self-adaptation, and processing of complex patterns to assist in judging faults type and status. Moreover, this technique helps achieve the goal of saving maintenance costs. At the beginning of the experiment, samples were taken from the real data, and then the sampled data were classified according to the common fault types of the motors to form a fault diagnosis data set. In terms of fault pre-diagnosis, the bathtub curve and sampling data are used to form a fault pre-diagnosis data set, and then each data set is regarded as a sensing signal and input to an improved artificial intelligence algorithm in order to present the real situation. Experimental results show that under the combination of one-dimensional multi-feature motor spectrum signals and query-based supervised machine learning and deep learning algorithms, Motor Fault Diagnosis System can identify motor fault types more quickly and accurately. This system can make up for the shortcomings that humans are hard to detect subtle changes in signals, humans have slower response times than computers', and humans are unable to monitor and diagnose faults 24/7. In addition, the experimental results also show that query-based learning can improve the anti-noise ability of most supervised machine learning and deep learning algorithms. This feature is especially helpful since the sensor data often collects noise, errors, and redundant signals. In other words, even if the data is mixed with error signals due to human or mechanical factors, query-based learning can still use the data to make more reliable decisions and identifications. The experimental results also show that query-based learning has the ability to improve the classification effect of the deep learning neural network model without changing the network model. The experimental results also achieve the purpose of explainable artificial intelligence through the functions of mapping model parameter changes, visualizing the contents of the data set, and artificial intelligence detection reports. It is also hoped that the fault diagnosis and pre-diagnosis technology developed in this thesis can be used as a reference standard for scholars to choose methods in the future.