基於漸進式學習法之類神經網路於無骨雨刷適配化彈片幾何規格之擴展

Abstract

近年無骨雨刷因其結構簡潔美觀，市場需求大幅增長，對應之設計需求亦日益提升。前人針對無骨雨刷之關鍵零件—金屬彈片，提出結合數值模型與機器學習(machine learning)之彈片幾何適配化方法，可快速提供業者具備優良刮刷表現之適配化彈片弧形，然僅適用於單一彈片長度、彈片厚度、及承壓單位負荷，對於其餘彈片規格仍須仰賴基於有限元素分析(finite element analysis)之彈片幾何適配化流程，所需時間成本高昂。 本研究針對此限制，應用漸進式學習法(progressive learning)於前人提出之基於機器學習之彈片適配化方法，旨於擴展適配化流程之適用範圍。研究架構為透過文克爾模型(Winkler’s model)與基於有限元素分析之無骨雨刷刮刷模型，分別累積彈片接觸壓力與刮刷壓力分佈資料，並以兩類神經網路(artificial neural networks, ANN)分別進行訓練，藉以快速預測對應之結果。並應用漸進式學習於基於有限元素之無骨雨刷刮刷模型之類神經網路，透過少量資料擴展適用之彈片規格，提升此類神經網路於彈片規格參數之泛化能力，同時維持快速預測刮刷壓力分佈之效率，最終模型適用於業界常用之7種彈片長度l (500、525、550、600、625、650 mm)；2種單位承壓負荷p (1.4、1.6 g/mm)；3種彈片厚度t (0.9、1.0、1.1 mm)，共42種彈片規格。 實際應用則透過兩類神經網路模型串接，經由彈片與玻璃參數快速預測接觸壓力分佈，再進一步預測刮刷壓力分佈，並據以判斷刮刷表現。此外，本研究將適配化流程同時應用於多片玻璃，透過交叉比對刮刷結果，獲致同時適用多片玻璃之泛用型彈片幾何。最終，本研究可適用於多種彈片規格，並於非特定玻璃快速預測任意彈片幾何之刮刷壓力分佈，據以判斷實務刮刷表現，於業界設計適配化及泛用型彈片幾何上具備高度實用價值。

Demands of flat wipers have increased recently due to their simpler structure compared to conventional wipers. Previously a series of procedures based on the machine learning approach were introduced to effectively forecast wiping patterns of the flat wiper with a given metallic flexor, one of the major components dominating the wiping capability. However, aforementioned approach was merely suitable for the certain specification of flexor. In order to further predict wiping patterns of the flexor with various design parameters such as length, thickness, and applied load, the concept of progressive learning is then introduced in the present study. A discrete Winkler model and a finite element analysis (FEA) were used to obtain contact pressure distributions of a flexor being compressed against a windshield glass and to evaluate wiping patterns of a flat wiper swiping over the windshield, respectively. These numerical results were subsequently adopted as datasets to train two artificial neural networks (ANNs) for rapid assessments of the wiping performance. A progressive learning approach was applied to the artificial neural network of the FEA-based wiping model to efficiently expand various specifications of flexor with rather limited amount of additional data. Wiping patterns based on the artificial intelligence model are in good agreement with those based on the finite element analysis, and the wiping capability of wipers can be reasonably evaluated. The current proposed procedure demonstrates high practical values to provide the compatible flexor geometry for the specific windshield and simultaneously for different windshields within relatively short period.