作為一個工科生，在全國人民抗擊病毒宅在家里的這段時期迎來自己最長的一個寒假。但是，學習科研不能因為寒假過長而落下，現(xiàn)在結(jié)合目前在看的一篇論文，給廣大在讀（ke)學(yan)生(gou)分享一下快速看完一篇英文論文的方法～
個人情況簡介
-上海交大研一在讀，一作EI論文1篇已錄用
-研究方向為復雜裝備的狀態(tài)檢測、可靠性
英文論文的結(jié)構(gòu)
在閱讀一篇論文時，首先要搞清楚英文論文的結(jié)構(gòu)，主要分為以下部分：
-Abstract
論文摘要。摘要部分出現(xiàn)在論文的最前面，因此也最為重要，需要重點閱讀。
-Introduction
介紹。介紹論文研究的背景、研究對象等信息，對不熟悉背景的同學有引導的作用。有時還包括研究方法論的系統(tǒng)介紹。
-Methodology
方法論。主要介紹研究采用的算法（研究方法）的原理、創(chuàng)新點。
-Experiment & Result
實驗和結(jié)果展示。這部分結(jié)合研究方法進行實驗（計算機實驗、力學實驗等等），對實驗數(shù)據(jù)、結(jié)果進行展示分析，通常以圖表的形式呈現(xiàn)。
-Conclusion
結(jié)論。通過對實驗結(jié)果的分析，對全文的結(jié)論進行總結(jié)和強調(diào)。
-Reference
參考文獻。一般來說參考文獻可以不看。但在判斷論文質(zhì)量的時候可以起到作用，參考文獻越多，文獻越權(quán)威，可以判斷該論文的質(zhì)量越好（當然期刊級別也可以判斷，所以盡量不要看水刊哦）
閱讀順序
對于閱讀一篇英文論文的順序，不同的同學會有不同的喜好。從個人的角度來說，我的順序通常時Abstract->（Introduction）->Conclusion->Experiment & Result->Methodology
采用這樣的順序原因如下：
-摘要部分展示了最主要的背景和結(jié)論，應該重點閱讀，當發(fā)現(xiàn)摘要展示的內(nèi)容和我們的期望/研究對象不相符，可以直接跳過該論文。
-對研究背景了解可以直接跳過Introduction部分，該部分主要是結(jié)合研究背景來強調(diào)研究的重要性（每個研究者當然都會說自己的研究重要）。當然不熟悉研究背景還是可以看一看～
-結(jié)論部分重申了重要的結(jié)論和重要的實驗結(jié)果，對于這些結(jié)論，我們會感到好奇，究竟是怎么得到的？如果論文得出的都是我們不感興趣的（與我們的研究對象不相關(guān)）的結(jié)論，可以直接跳過該論文。
-在結(jié)論部分我們發(fā)現(xiàn)了令我們感興趣的結(jié)論，于是我們可以回到實驗和結(jié)果展示部分，仔細研究實驗和結(jié)果展示的圖、表。作者為什么要繪制這些圖表？作者通過怎樣的實驗得到了這些圖和表？
-最后閱讀方法論部分。對于我而言，這樣可以省去被數(shù)學公式、物理原理難住的時間。如果確定論文的方法可行，適合自己的研究，再弄懂這一部分。此時最好再借助一些博客和工具書，詳細理解公式的推導過程。
總的來說就是：
1.通讀摘要、（背景介紹）、結(jié)論，確認是否為自己感興趣（相關(guān)）的方向、方法。
2.閱讀實驗和結(jié)果展示部分，了解實驗是如何進行的，數(shù)據(jù)進行了怎樣的處理和分析。
3.確定論文研究方法可借鑒，再閱讀方法論，徹底理解方法論的公式推導。如果并不想采用論文的研究方法，這一部分的閱讀也可省去。
實例
現(xiàn)在以目前正讀的一篇會議論文Wind Turbine Structural Health Monitoring: A Short
Investigation Based on SCADA Data來簡單介紹一下這個過程。
首先是摘要部分
*The use of offshore wind farms has been growing in recent years, as steadier and higher
wind speeds can be generally found over water compared to land. Moreover, as human activities tend to complicate the construction of land wind farms, offshore locations, which can be found more easily near densely populated areas, can be seen as an attractive choice. However, the cost of an offshore wind farm is relatively high, and therefore their reliability is crucial if they ever need to be fully integrated into the energy arena. As wind turbines have become more complex, efficient, and expensive structures, they require more sophisticated monitoring systems, especially in offshore sites where the financial losses due to failure could be substantial. *This paper presents the preliminary analysis of supervisor control and data acquisition (SCADA) extracts from the Lillgrund wind farm for the purposes of structural health monitoring. A machine learning approach is applied in order to produce individual power curves, and then predict measurements of the power produced of each wind turbine from the measurements of the other wind turbines in the farm. A comparison between neural network and Gaussian process regression is also made.
可以看出，斜體部分是介紹整篇論文的背景。（另外，如果有同學摘要閱讀困難的話建議谷歌翻譯哦）介紹了海上風場的日益增長，然而海上風場的運維費用昂貴，因此，海上風機的狀態(tài)監(jiān)測具有重要性。粗體部分介紹的是論文采用的方法：首先用機器學習方法繪制出每個風機的功率曲線，再用其他風機的功率值去預測某個風機的功率值，并且對比了神經(jīng)網(wǎng)絡(luò)和高斯過程兩種方法。
接著看結(jié)論部分
由于摘要中的背景比較詳細了，我們可以直接跳過Introduction部分直接看Conclusion。
This paper presented a preliminary exploration of the suitability of SCADA extracts from the Lillgrund wind farm for the purposes of SHM. Artificial neural networks and Gaussian processes were used to build a reference power curve (wind speed versus power produced) for each of the 48 turbines existing in the farm. Then, each reference model was used to predict the power produced in the rest of the turbines available, creating thus a confusion matrix of the MSE errors for all combinations.
The results showed that nearly all models were very robust with the highest MSE error to be 4.8291, and this was happening when the model trained in turbine 4 was predicting power from turbine 3. Both turbines 3 and 4 are located in the outside row of the wind farm. It was shown that when wind speed data which did not come from time instances where the error status was ‘0’ (meaning healthy data), were used as an input to the trained neural networks, the MSE error was significantly larger. Although, it was seen that in some cases the very large MSE was due to emergency stops or manual stops, and it is currently not known whether there was scheduled maintenance, this result, still shows the potential for novelty detection in the turbines.
In this spirit, the confusion matrices that were presented earlier can form the baseline for thresholds for a population-based SHM of the whole farm. It is anticipated that the power curve, and possibly other similar features, will be adequate to be used in future work in the construction of control charts for the monitoring of the whole wind farm and of the potential interaction or influence of the turbines with one another during their normal operation.Future work will also focus on the full analysis of the error statuses that were presented during the recorded time. In the comparison of the regression between neural networks and Gaussian processes, it was shown that there were no significant differences, with the networks performing with slightly lower MSE error.
結(jié)論部分是對摘要部分未說明完全的部分更詳細的補充。斜體進一步擴充了摘要的內(nèi)容：采用了神經(jīng)網(wǎng)絡(luò)和高斯過程對48個風機進行了功率曲線建模，接著采用每個風機的模型對其他風機的功率進行預測，得到了均方誤差（MSE errors）的混淆矩陣（confusion matrix）。粗體部分詳細分析了結(jié)果：所有的模型很穩(wěn)健，最高的均方誤差為4.8291，接著分析了最高均方根誤差產(chǎn)生的原因（采用風機4的模型預測風機3）。強調(diào)了產(chǎn)生的混淆矩陣可以作為風機異常狀態(tài)監(jiān)測的基線值，可以結(jié)合控制圖來監(jiān)測整個風場的運行狀態(tài)。最后的斜體字強調(diào)了一下未來可能的研究工作。
通過結(jié)論，我們大概知道了論文的研究方法。接著可以去看論文的實驗和結(jié)果展示部分。
再看實驗和結(jié)果展示部分
從結(jié)論我們看出，最重要的結(jié)果是神經(jīng)網(wǎng)絡(luò)和高斯過程均方誤差的混淆矩陣，所以這里只找這兩個圖就好；果不其然，在文章的正中間找到了。
我們再來看看作者給出對應的分析(神經(jīng)網(wǎng)絡(luò)）：
From the results it is clear that almost all the trained networks are very robust and the maximum MSE error is around 5, which mainly occurs in turbines 3 and 4 which are located in the outside row of the wind farm.
和結(jié)論部分類似，說明了模型穩(wěn)健，最大的均方誤差在5左右，發(fā)生在位于風場最外排的風機3和風機4之間。
Subsequent scanning of the data revealed that the majority of the instances where the regression error becomes high (in turbine 4) happened when the turbine was not working, either from emergency stops or manual stops...Essentially, Figure 2 shows a map of potential thresholds, which can be used for the monitoring (in a novelty detection scheme) of the turbines individually or as a population.
這里說明了當風機未正常工作時，MSE的值非常高。混淆矩陣圖展示了一個風機正常工作的潛在閾值，可以用來進行單個風機或整個風場的狀態(tài)監(jiān)測。
這部分可以說就是論文結(jié)論部分的拓展。
最后——
當看完這些之后，我們再思考，這個方法/實驗可以用到我們的問題中去嗎？如何實現(xiàn)？這時就可以進一步結(jié)合其他的參考資料了解一下神經(jīng)網(wǎng)絡(luò)回歸、高斯過程回歸、MSE的計算和混淆矩陣的生成這些問題啦！
到此，一篇論文讀完。
參考文獻
Papatheou E, Dervilis N, Maguire E, et al. Wind turbine structural health monitoring: a short investigation based on SCADA data[C]. 2014.
原文鏈接
這篇文章如果對你有幫助的話請點個贊哦～
