BP神經(jīng)網(wǎng)絡(luò)的異常點(diǎn)檢測(cè)應(yīng)用可行性研究 本科畢業(yè)設(shè)計(jì)（論文）

摘  要

異常點(diǎn)數(shù)據(jù)是指數(shù)據(jù)集中與眾不同數(shù)據(jù)。這部分?jǐn)?shù)據(jù)的量小，但是對(duì)于我們的日常生產(chǎn)生活的影響極大。因此，異常點(diǎn)檢測(cè)被廣泛應(yīng)用于網(wǎng)絡(luò)入侵檢測(cè)，金融保險(xiǎn)，天氣預(yù)報(bào)以及新藥研制等領(lǐng)域。相對(duì)于大量的正常數(shù)據(jù)挖掘而言，異常點(diǎn)檢測(cè)被稱作小模式數(shù)據(jù)挖掘。BP算法是一種常用的數(shù)據(jù)挖掘算法。但是BP算法進(jìn)行實(shí)際數(shù)據(jù)的異常點(diǎn)數(shù)據(jù)挖掘過程中存在：實(shí)際數(shù)據(jù)的維數(shù)較高，存在冗余特征的干擾，以及在高維特征下，數(shù)據(jù)量不充分的問題。因此，本文分析BP神經(jīng)網(wǎng)絡(luò)處理各種數(shù)據(jù)的情況，并得到以下結(jié)果。（1）BP神經(jīng)網(wǎng)絡(luò)能夠較好的分離特征單一的仿真數(shù)據(jù)；但是（2）特征相似性較大的數(shù)據(jù)集，難以分離判斷；（3）正常數(shù)據(jù)不充分或者不具有代表性，因此正常數(shù)據(jù)類學(xué)習(xí)不充分，從而導(dǎo)致異常無法判斷。針對(duì)以上問題，本文提出了以下的改進(jìn)措施：（1）BP算法前進(jìn)行特征約簡(jiǎn)（映射）從中選取有益于異常檢測(cè)的特征（2）多神經(jīng)網(wǎng)絡(luò)融合，不同神經(jīng)網(wǎng)絡(luò)識(shí)別不同的特征，相互取長(zhǎng)補(bǔ)短，融合后得到最終的結(jié)果。

關(guān)鍵字：異常，BP，異常點(diǎn)檢測(cè)，神經(jīng)網(wǎng)絡(luò)

注：本設(shè)計(jì)（論文）題目來源于教師的國(guó)家級(jí)（或部級(jí)、省級(jí)、廳級(jí)、市級(jí)、校級(jí)、企業(yè)）科研項(xiàng)目，項(xiàng)目編號(hào)為：          。

 
Abstract

Outlier data is the data set different data. This part of the small amount of data, but for our daily production and life of great. Therefore, the anomaly detection is widely used in network intrusion detection, finance, insurance, weather, and new drug development and other fields. Relative to the large number of normal data mining, the anomaly detection model is called data mining small. BP algorithm is a commonly used data mining algorithm. But the BP algorithm to real data outliers exist in the data mining process: the higher the dimension of the actual data, there are redundant features of the interference, and high-dimensional feature, the issue of inadequate data. Therefore, this paper analyzes a variety of BP neural network processing of data, and to get the following results. (1) BP neural network can better separation characteristics of a single simulation data; but (2) the characteristics of similar large data sets, separation is difficult to judge; (3) normal data is not sufficient or not representative, so the normal data class learning is not sufficient, leading to abnormal can not judge. To solve the above problem, this paper proposes the following improvements: (1) BP algorithm before feature reduction (map) benefit from anomaly detection features selected (2) integration of multiple neural networks, different neural network to recognize the different characteristics of each each other, the final fusion result.

Key Words：Outliers-Data，BP，Algorithms，Neural Networks
 
目  錄
1引言 1
1.1背景 1
1.2 傳統(tǒng)已有異常點(diǎn)算法介紹 1
1.2.1基于統(tǒng)計(jì)學(xué)的異常點(diǎn)檢測(cè)算法 1
1.2.2基于距離的異常點(diǎn)檢測(cè)算法 2
1.2.3基于密度的算法 3
1.2.4基于偏差的異常點(diǎn)檢測(cè) 5
1.2.5基于聚類的異常點(diǎn)檢測(cè)算法 6
2基于屬性特征在異常點(diǎn)檢測(cè)中的研究 7
3 BP神經(jīng)網(wǎng)絡(luò)介紹 9
3.1模型簡(jiǎn)介 9
3.2計(jì)算各層節(jié)點(diǎn)輸出 9
3.3 修正權(quán)值 10
4 異常檢測(cè)中BP神經(jīng)網(wǎng)絡(luò)的設(shè)計(jì) 13
4.1可微閾值單元 13
4.2單個(gè)BP網(wǎng)絡(luò)結(jié)構(gòu)設(shè)計(jì) 13
4.3BP神經(jīng)網(wǎng)絡(luò)學(xué)習(xí)過程的基本步驟 14
5實(shí)驗(yàn)研究 17
5.1研究使用的數(shù)據(jù)庫(kù)介紹 17
5.2訓(xùn)練方案一實(shí)驗(yàn)：把bp神經(jīng)網(wǎng)絡(luò)相似性代替距離算法相似度量 17
5.3訓(xùn)練方案二實(shí)驗(yàn)：用單個(gè)神經(jīng)網(wǎng)絡(luò)對(duì)訓(xùn)練數(shù)據(jù)庫(kù)整體特性進(jìn)行學(xué)習(xí) 18
5.4訓(xùn)練方案三實(shí)驗(yàn)：多神經(jīng)網(wǎng)絡(luò)各種形式訓(xùn)練及其決策 19
5.4.1實(shí)驗(yàn)設(shè)計(jì)思路 19
5.4.2實(shí)驗(yàn)方案及步驟 20
5.4.3實(shí)驗(yàn)分析 22
5.4.4實(shí)驗(yàn)失敗原因分析 23
5.5BP調(diào)參實(shí)驗(yàn) 25
5.5.1對(duì)實(shí)驗(yàn)一調(diào)整隱層實(shí)驗(yàn) 25
5.5.2對(duì)實(shí)驗(yàn)二調(diào)整隱層實(shí)驗(yàn) 26
5.5.3對(duì)實(shí)驗(yàn)三調(diào)整隱層實(shí)驗(yàn) 29
5.6數(shù)據(jù)仿真實(shí)驗(yàn) 31
5.6.1實(shí)驗(yàn)思路 31
5.6.2實(shí)驗(yàn)步驟 31
5.6.3實(shí)驗(yàn)結(jié)果 32
5.6.4結(jié)果分析 33
5.7實(shí)驗(yàn)整體分析 33
總結(jié)與展望 35
致謝 39