基于音頻特征和模糊神經網絡的禽流感病雞檢測

張鐵民，黃俊端

?

基于音頻特征和模糊神經網絡的禽流感病雞檢測

張鐵民1,2，黃俊端1

（1. 華南農業大學工程學院，廣州 510642； 2. 華南農業大學國家生豬種業工程技術研究中心，廣州 510642）

為了能在早期發現禽流感并進行預防，該文提出了一種基于音頻特征和模糊神經網絡的禽流感病雞檢測方法。依據獲取的家禽音頻和環境及其他噪聲的譜熵差別大的特點，在復雜環境中分析并提取出雞聲，丟棄非雞聲段，對提取的雞聲進行分析及處理，計算短時過零率、短時能量以及短時過零率與短時能量混合特征，用作判別患禽流感的病雞和健康雞的依據。利用T-S模糊神經網絡，對提取出來的家禽音頻特征進行訓練和識別，試驗表明隸屬度函數為鐘形函數、隸屬度個數為2時模糊神經網絡對試驗提取的3個雞聲特征組成的3組測試集的敏感性分別為75.47%、80.39%和76.92%，特異性分別為80.85%、79.59%和72.92%，正確識別率分別為78%、80%和75%。該研究為規?；仪蒺B殖場及大型家禽流通市場的禽流感病禽識別提供一套快速、高效檢測方法。

神經網絡；識別；提??；譜熵；短時過零率；短時能量；雞病檢測

0 引 言

隨著現代規模化養雞業的發展，禽流感對經濟、食品安全和人類健康有著重要的影響[1]，因此在規?；B殖環境中及時對患有禽流感的病雞進行快速、準確的識別不僅直接關系到養雞業的經濟效益，同時對預防禽流感交叉傳播具有重要意義。傳統的雞病診斷主要依靠獸醫對雞的姿態、雞冠、聲音以及糞便進行觀察[2-4]，費時費力，尤其對于規?；B殖場，效率低下。然而隨著現代電子計算機技術的發展，許多新的、更高效的雞病檢測方法被提出來。

目前多采用視頻角度,許多基于機器視覺對其他動物行為監測的方法被陸續提出[5-8]。而對于運用計算機視覺識別病雞、監測雞的行為，畢敏娜等[9-11]通過對家禽姿態識別進行病雞識別，其支持向量機模型在測試集的識別率達到99.469%。王琳等[12]用數值積分方法提取出雞的深度圖像特征，結合神經網絡，實現群體肉雞的質量估計。勞風丹等[13]用機器視覺實現對單只蛋雞的行為識別，監測其生產和健康狀況。而從音頻角度看，動物的發聲包含豐富的信息，能夠在一定程度上反饋動物的健康情況[14-16]，目前有許多基于音頻分析的方法應用于其他動物的行為和健康監測[17-22]。而對于分析雞的音頻研究，Banakar Ahmad等[23]用數據挖掘方法和Dempster-Shafer證據理論，結合支持向量機作為識別工具，識別和分類幾種常見的雞病。國內曹晏飛[24]針對棲架飼養模式下蛋雞發出的聲音，提出了基于功率譜密度特征的分類識別方法，該方法的平均識別率達到95%。余禮根等[25]以海蘭褐蛋雞為例，收集了其在小規模（5只）飼養條件下的叫聲信息，提取其包括持續時間、基音頻率、頻譜質心、共振峰及其衍生的發聲特征參數，構建出蛋雞發聲音頻數據庫，分析蛋雞發聲和其他行為的聯系，得出分析蛋雞發聲特征有助于了解其行為特性、機體狀態以及種群間的信息傳遞。

本文立足于應用音頻分析技術，通過分析籠養雞中的健康和感染禽流感病毒的雞叫聲，從音頻角度，提出一種在環境噪聲背景中提取出雞聲的有效音頻特征的方法，并用模糊神經網絡作為分類器，識別禽流感病雞的叫聲和健康雞的叫聲，以期為家禽養殖業提供一種非接觸式的自動化檢測禽流感的方法。

1 健康和患禽流感病的雞聲獲取方法

雞聲獲取方法：第一天：12:00將試驗用的14只五周齡的無特定病原（specific pathogen-free，SPF）白來航雞放入雞隔離器中，讓其熟悉生存環境，減少應激反應；第二天，16:30將錄音筆放入置于雞隔離器中，記錄雞聲，如圖1所示，錄音筆為T&F-91加強版32G數字高清錄音筆，采樣頻率48 000 Hz。錄音筆能長時間連續不間斷錄音120 h；第三天16:00對雞進行感染，感染后的雞呈現出眼瞼水腫，精神呆滯，聲音嘶啞，羽毛蓬松等特征。錄音筆持續記錄雞的叫聲直到第六天感染禽流感病毒雞全部死亡。取感染禽流感病毒前一天的雞的叫聲作為健康雞的叫聲（試驗的第二天），取感染禽流感病毒后的雞的叫聲作為禽流感病雞的叫聲（試驗的第四天）。

感染方法：采用的H7N9亞型禽流感病毒由華南農業大學獸醫學院禽病研究室分離鑒定，在華南農業大學動物生物安全三級（animal biosafety level 3，ABSL-3）實驗室中進行，所有操作均按照ABSL-3相關標準步驟及相關生物安全標準進行。利用Reed-Muench法計算病毒半數胚胎感染劑量（50% Embryo Infective Dose，EID50），用含10 000/mL青霉素和鏈霉素的無菌PBS將H7N9亞型禽流感病毒均稀釋到106EID50/0.1 mL，每只雞經滴鼻點眼感染106EID50/0.1 mL病毒稀釋液0.1 mL。

圖1 音頻采集環境

2 雞聲信號的時域分析與處理

2.1 預加重

由于聲音信號低頻信噪比將大，而高頻信噪比不足，需對輸入的數字雞聲信號的高頻部分進行預加重處理，采用高通濾波器進行預加重，以提高雞聲的高頻分辨率，高通濾波器的傳遞函數如下[26-27]。

其中為預加重系數，0.9<<1.0。設時刻語音采樣值為()，經過預加重處理后的結果為()=()?(?1)，這里取0.98。

圖2為原始含噪雞聲信號和預加重后的頻率幅值曲線，經過高通濾波器后，雞聲的低頻部分被削弱，高頻部分被增強。原始含噪雞聲信號和預加重后的含噪雞聲信號都經過歸一化處理。

圖2 預加重效果

2.2 雞聲信號的加窗與分幀

雞聲信號是時變信號，但在一個短時間內（10～30 ms），其特性基本保持不變，可以將其看做準穩態過程，其信號具有短時平穩性[27-30]。為了使截取的雞聲信號波形緩慢降為零，減小雞聲幀的截斷效應[24]，選hamming窗對雞聲信號分幀，取21.3 ms為一幀，hamming窗數學表達式如下[26-27]，表示幀長量。

2.3 基于譜熵法的雞聲端點檢測

熵表示信息的有序程度，由Shannon引用到信息理論中來，信號以信息熵來作為信息選擇和不確定性的度量[31]。Shen等[32]在試驗中發現語音的熵和噪聲的熵存在較大的差異，首次提出基于熵的語音端點檢測方法。雞叫聲的熵跟環境噪聲的熵有明顯不同，提出一種基于譜熵法的雞聲端點檢測，從一段含噪雞聲中準確地找出雞聲信號起始點和結束點，使有效的雞聲信號和無用的噪聲信號得以分離。

2.3.1 基于譜熵法的雞聲端點檢測算法

1）對含噪雞聲進行加窗分幀處理，計算每一幀的譜的能量。含噪雞聲信號定義為()，加hamming分幀處理后得到的第幀含噪雞聲信號為x()，其快速傅里葉變換(Fast Fourier transform，FFT)表示為X()，其轉置矩陣表示為X()，下標表示為第幀，表示第條譜線，表示FFT的點數，取=1 024。每一幀聲音信號在頻域中的短時能量E為

2）計算每一幀中每個樣本點的概率密度函數。定義某一譜線的能量譜為Y()=X()X()，則每個頻率分量的歸一化譜概率密度p()為

3）計算每一幀的譜熵值H。

4）設定判決門限進行端點檢測。本文選擇判決門限為含噪雞聲所有幀的譜熵值的平均值減去所有幀中譜熵值的最小值。

2.3.2 基于譜熵法的雞聲端點檢測效果

基于譜熵法的雞聲端點檢測流程如圖3所示。

圖3 基于譜熵法的雞聲端點檢測流程

本文選擇最小雞聲長度為3幀?；谧韵嚓P函數的雞聲端點檢測效果如圖4所示，圖中紅色實線表示雞聲起始幀位置，藍色虛線表示雞聲結束幀位置。

注：紅色實線表示端點檢測中雞聲起始幀的位置，藍色虛線表示端點檢測中雞聲結束幀的位置。

3 雞聲短時過零率和短時能量

3.1 雞聲短時過零率計算

短時過零率（Short-time zero-crossing rate，STZ）表示聲音信號波形穿過橫軸的次數，對于離散信號如果相鄰的取樣值改變符號則稱為過零，一幀雞聲信號x()的短時過零率Z的計算為

式中下標表示第幀雞聲信號。sgn[ ]是求符號函數，即

為了消除錄音器隨機微弱電流噪聲的影響，引入一個去噪變量，本文選擇去噪變量為0.000 1。則雞聲信號的短時過零率Z的計算為

3.2　雞聲短時能量計算

短時能量（short-time energy，STE）用來度量音頻信號的幅度值變化，雞在感染禽流感后叫聲發生改變，聲音能量也發生改變，一幀雞聲x()短時能量E的計算為

3.3 雞聲短時過零率與短時能量混合特征

雞聲短時過零率包含雞聲信號的符號信息，雞聲短時能量包含雞聲信號的幅度信息，將雞聲短時過零率與雞聲短時能量的數值相乘，作為同時包含雞聲信號的符號信息和雞聲信號的幅度信息的雞聲短時過零率與短時能量混合特征，一幀雞聲信號x()的短時過零率與短時能量混合特征K的計算為

4 基于模糊神經網絡的雞聲識別

神經網絡具有并行計算，分布式信息存儲，容錯能量強及自適應學習能力等優勢，模糊邏輯是一種處理不確定性和非線性的強有力的工具，模糊神經網絡將神經網絡與模糊邏輯結合起來，具備兩者的長處，性能比單純的神經網絡或者單純的模糊邏輯更強[33]。

4.1 T-S模糊模型

模糊模型主要有2種，一種是模糊規則的后件是輸出量的某一模糊集合，稱為模糊系統的標準模型，另一種是模糊規則的后件輸入是輸入語言變量函數，由Takagi等提出[34-36]，稱為T-S模糊模型。模糊系統的標準模型雖然符合人的思維和語言表達習慣，但存在計算復雜、不利于數學分析等缺點。本文選用T-S模糊模型。

T-S模糊模型用如下的if-then規則形式定義，在規則R的情況下，模糊推理為

根據模糊計算結果計算模糊模型的歸一化輸出值y

4.2 T-S模糊神經網絡

5 結果與分析

以雞聲短時過零率、短時能量和短時過零率與短時能量混合特征作為識別特征，取健康雞的識別特征和禽流感病雞的識別特征各450個，組成一個行數為900，列數為3的識別特征矩陣。構造行數為900，列數為1的矩陣作為標志矩陣，在標志矩陣中，禽流感病雞聲用1表示，健康雞聲用0表示，將和組合在一起作為數據集（，），訓練樣本為（x，y），x表示第個雞聲的識別特征，y表示第個雞聲是否患病的標志（=1,2，…900），隨機打亂樣本的順序，取前600個樣本作為訓練集，后300個樣本作為3組測試集，每組測試集100個樣本，各數據集中禽流感病雞聲特征和健康雞聲特征數量如表1所示。

表1 訓練集和測試集的禽流感病雞聲特征和健康雞聲特征

圖5 模糊神經網絡對測試集的識別率與訓練次數的關系曲線

由圖5可知，模糊神經網絡在訓練次數為16時對測試集的識別率達到穩定，隸屬度函數為鐘形函數的識別率最高。選擇隸屬度函數為鐘形函數，訓練次數為16次，分別計算該模糊神經網絡對3組測試集識別結果的8個統計值：正確的正例、正確的反例、錯誤的正例、錯誤的反例、敏感性、特異性、正確識別率和錯誤識別率。本文中，敏感性為某一測試集中被正確診斷為禽流感病雞聲的個數與該測試集中所有禽流感病雞聲的數量之比，特異性為某一測試集被正確診斷為健康雞聲的個數與該測試集中所有健康雞聲的數量之比，敏感性和特異性的計算公式如式（18），（19）所示。數據統計結果如表2所示。

由表2可知，隸屬度函數為鐘形函數，隸屬度個數為2時，模糊神經網絡對本試驗提取的3個雞聲特征組成的3組測試集的敏感性分別為75.47%、80.39%和76.92%，特異性分別為80.85%、79.59%和72.92%，正確識別率分別為78%、80%和75%。由表2可知模糊神經網絡對測試集的識別率最高達80%，識別率在75%到80%之間。

6 結 論

本文通過在ABSL-3實驗室中對5周齡的SPF雞做禽流感病毒感染試驗，收集了健康雞的叫聲和禽流感病雞的叫聲。以收集到的聲音數據為分析對象，首先對聲音信息進行預處理、加窗和分幀，接著分析含噪雞聲的譜熵特征，提出了一種基于譜熵法的雞聲端點檢測方法，在含有噪聲的雞聲錄音中截取出雞的叫聲，舍棄環境噪聲。

通過計算有效雞聲幀的短時過零率、短時能量以及短時過零率與短時能量混合特征，使用基于T-S模糊模型的模糊神經網絡做健康雞和禽流感病雞的雞聲特征識別，試驗表明隸屬度函數為鐘形函數、隸屬度個數為2時模糊神經網絡對本試驗提取的3個雞聲特征組成的3組測試集的敏感性分別為75.47%、80.39%和76.92%，特異性分別為80.85%、79.59%和72.92%，正確識別率分別為78%、80%和75%。

本文所提出的雞聲特征提取和識別方法對家禽養殖場的家禽疫病非接觸式、快速和自動識別具有重要意義。

[1] Brown Mac, Moore Leslie, Mcmahon Benjamin, et al. Constructing rigorous and broad biosurveillance networks for detecting emerging zoonotic outbreaks[J]. Plos One, 2015, 10(5): e124037.

[2] Mdbazlurr Mollah, Mda Hasan, Mda Salam, et al. Digital image analysis to estimate the live weight of broiler[J]. Computers & Electronics in Agriculture, 2010, 72(1): 48－52.

[3] Aydin A, Bahr C, Viazzi S, et al. A novel method to automatically measure the feed intake of broiler chickens by sound technology[J]. Computers & Electronics in Agriculture, 2014, 101(1): 17－23.

[4] Aydin A, Bahr C, Berckmans D. A real-time monitoring tool to automatically measure the feed intakes of multiple broiler chickens by sound analysis[J]. Computers and Electronics in Agriculture, 2015, 114:1－6.

[5] 浦雪峰，朱偉興，陸晨芳. 基于對稱像素塊識別的病豬行為監測系統[J]. 計算機工程，2009，35(21)：250－252.

Pu Xuefeng, Zhu Weixing, Lu Chenfang. Sick pig behavior monitor system based on symmetrical pixel block recognization[J]. Computer Engineering, 2009, 35(21): 250－252. (in Chinese with English abstract)

[6] Shao J, Xin H, Harmon J D. Comparison of image feature extraction for classification of swine thermal comfort behavior[J]. Computers and Electronics in Agriculture, 1998, 19(3): 223－232.

[7] Shao B, Xin H. A real-time computer vision assessment and control of thermal comfort for group-housed pigs[J]. Computers and Electronics in Agriculture, 2008, 62(1): 15－21.

[8] 劉龍申，沈明霞，柏廣宇，等. 基于機器視覺的母豬分娩檢測方法研究[J]. 農業機械學報，2014，45(3)：237－242.

Liu Longshen, Shen Mingxia, Bo Guangyu, et al. Sows parturition detection method based on machine vision[J]. Transactions of the Chinese Society for Agricultural Machinery, 2014, 45(3): 237－242. (in Chinese with English abstract)

[9] 畢敏娜，張鐵民，莊曉霖，等. 基于色差信息多色彩模型的黃羽雞快速分割方法[J]. 農業機械學報，2016，47(12)：293－298.

Bi Minna, Zhang Tiemin, Zhuang Xiaolin, et al. Fast segmentation method of yellow feather chicken based on difference of color information in different color models[J]. Transactions of the Chinese Society for Agricultural Machinery, 2016, 47(12): 293－298. (in Chinese with English abstract)

[10] 畢敏娜，張鐵民，莊曉霖，等. 基于雞頭特征的病雞識別方法研究[J]. 農業機械學報，2018，49(1)：51－57.

Bi Minna, Zhang Tiemin, Zhuang Xiaolin, et al. Recognition method of yellow feather chicken based on head features[J]. Transactions of the Chinese Society for Agricultural Machinery, 2018, 49(1): 51－57. (in Chinese with English abstract)

[11] Zhuang Xiaolin, Bi Minna, Guo Jilei, et al. Development of an early warning algorithm to detect sick broilers[J]. Computers & Electronics in Agriculture, 2018, 144: 102－113.

[12] 王琳，孫傳恒，李文勇，等. 基于深度圖像和BP神經網絡的肉雞體質量估測模型[J]. 農業工程學報，2017，33(13)：199－205.

Wang Lin, Sun Chuanheng, Li Wenyong, et al. Establishment of broiler quality estimation model based on depth image and BP neural network[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2017, 33(13): 199－205. (in Chinese with English abstract)

[13] 勞鳳丹，滕光輝，李軍，等. 機器視覺識別單只蛋雞行為的方法[J]. 農業工程學報，2012，28(24)：157－163.

Lao Fengdan, Teng Guanghui, Li Jun, et al. Behavior recognition method for individual laying hen based on computer vision[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2012, 28(24): 157－163. (in Chinese with English abstract)

[14] 劉恒，吳迪，蘇家儀，等. 運用高斯混合模型識別動物聲音情緒[J]. 國外電子測量技術，2016，35(11)：82－87.

Liu Heng, Wu Di, Su Jiayi, et al. Recognition of animal sound’s emotion based on gaussian mixture model[J]. Foreign Electronic Measurement Technology, 2016, 35(11): 82－87. (in Chinese with English abstract)

[15] 汪開英，趙曉洋，何勇. 畜禽行為及生理信息的無損監測技術研究進展[J]. 農業工程學報，2017，33(20)：197－209.

Wang Kaiying, Zhao Xiaoyang, He Yong. Review on no moninavasive monitoring technology of poultry behavior and physiological information[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2017, 33(20): 197－209. (in Chinese with English abstract)

[16] 于天福. 基于聲音特征的動物行為識別系統研究[D]. 哈爾濱：東北林業大學，2010.

Yu Tianfu. The Research of Animal Behavior Recognition System Based on Sound Features[D]. Harbin: Northeast Forestry University, 2010. (in Chinese with English abstract)

[17] 王丹聰. 基于多傳感器融合的豬只行為辨別[D]. 西安：陜西科技大學，2018.

Wang Dancong. Pig Behavior Identification Based on Multisensor Fusion[D]. Xi’an: Shaanxi University of Science and Technology, 2018. (in Chinese with English abstract)

[18] Moura D J, Silva W T, Naas I A, et al. Real time computer stress monitoring of piglets using vocalization analysis[J]. Computers and Electronics in Agriculture, 2008, 64(1): 11－18.

[19] Moura D J, Silva W T, Naas I A, et al. Real time computer stress monitoring of piglets using vocalization analysis[J]. Computers and Electronics in Agriculture, 2008, 64(1): 11－18.

[20] Guarino M, Jans P, Costa A, et al. Field test of algorithm for automatic cough detection in pig houses[J]. Computers and Electronics in Agriculture, 2008, 62(1): 22－28.

[21] Ferrari S, Piccinini R, Silva M, et al. Cough sound description in relation to respiratory diseases in dairy calves[J]. Preventive Veterinary Medicine, 2010, 96(3/4): 276－280.

[22] Milone D H, Rufiner H L, Galli J R, et al. Computational method for segmentation and classification of ingestive sounds in sheep[J]. Computers and Electronics in Agriculture, 2009, 65(2): 228－237.

[23] Banakar Ahmad, Sadeghi Mohammad, Shushtari Abdolhamid. An intelligent device for diagnosing avian diseases: Newcastle, infectious bronchitis, avian influenza[J]. Computers & Electronics in Agriculture, 2016, 127: 744－753.

[24] 曹晏飛. 復雜背景下蛋雞聲音分類提取方法研究[D]. 北京：中國農業大學，2015.

Cao Yanfei. Research on Methods for Classification and Extraction of Laying Hens' Vocalizations in Complex Environment[D]. Beijing: China Agricultural University, 2015. (in Chinese with English abstract)

[25] 余禮根，滕光輝，李保明，等. 蛋雞發聲音頻數據庫的構建與應用[J]. 農業工程學報，2012，28(24)：150－156.

Yu Ligen, Teng Guanghui, Li Baoming, et al. Development and application of audio database for laying hens[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2012, 28(24): 150－156. (in Chinese with English abstract)

[26] Oppenheim A V, Schafer R W, Buck W J R. Discrete-time signal processing[J]. Electronics & Power, 2009, 23(2):157.

[27] Allen J B, Rabiner L. A unified approach to short-time Fourier analysis and synthesis[J]. Proc IEEE, 1977, 65(11): 1558－1564.

[28] Portnoff M. Short-time Fourier analysis of sampled speech[J]. IEEE Transactions on Acoustics Speech & Signal Processing, 1981, 29(3): 364－373.

[29] Immerseel L V, Peeters S. Digital implementation of linear gammatone filters: Comparison of design methods[J]. Acoustics Research Letters Online, 2003, 4(3):59－64.

[30] Quatieri Thomas. Discrete-time Speech Signal Processing: Principles and Practice[M]. Prentice Hall, 2002.

[31] Shannon Claude E, Weaver Warren. The mathematical theory of communication[J]. Computers in Medical Practice, 1950, 3(9): 31－32.

[32] Shen J L, Hung J W, Lee L S. "Robust entropy-based endpoint detection for speech recognition in noisy environments"[C]// The 5th International Conference on Spoken Language Processing, Incorporating the 7th Australian International Speech Science and Technology Conference, Sydney Convention Centre, Sydney, Australia, 1998.

[33] Sun Zengqi, Xu Hongbin. Fuzzy neural network based on T S model[J]. Journal of Tsinghua University, 1997, 37(3):76－80.

[34] Takagi T, Sugeno M. Fuzzy identification of systems and its applications to modeling and control[J]. Readings in Fuzzy Sets for Intelligent Systems, 1993, 15(1): 387－403.

[35] Jin Yaochu, Jiang J, Zhu J.Neural network based fuzzy identification and its application to modeling and control of complex systems[J]. IEEE Transactions on Systems, Man and Cybernetics, 1995, 25(6):990－997.

[36] Takagi Tomohiro, Sugeno Michio. Fuzzy identification of systems and its applications to modeling and control[J]. IEEE Transactions on Systems Man & Cybernetics, 1985, 15(1): 116－132.

Detection of chicken infected with avian influenza based on audio features and fuzzy neural network

Zhang Tiemin1,2,Huang Junduan1

(1.,,510642,; 2.,,510642,)

Avian influenza influences the economy, food safety and human health. A rapid and accurate detection of chicken infected with avian influenza in farming not only directly benefits the chicken farming, but also prevents the cross propagation of avian influenza. This paper proposes a non-invasive disease poultry detection method based on voice analysis, which is designed to achieve the identification of the voice of chickens infected with avian influenza and that of the healthy ones. First, 14 white leghorn chickens of 5 weeks of age with specific pathogen free (SPF) were put into the isolated cage in the animal biosafety level 3 (ABSL 3) laboratory to record their voice. The voice samples of healthy chickens were collected by a T&F-91 enhanced 32G digital HD recording pen, and then the chickens were inoculated with the H7N9 avian influenza virus in the ABSL-3 laboratory. The H7N9 subtype avian influenza virus was diluted to 106EID50/0.1 mL with 10 000/mL penicillin and streptomycin free phosphate-buffered saline (PBS), which was then used to inoculate the chickens, each with 0.1 mL virus diluent. After that, the samples of infected chickens’ voice were collected. Secondly, in light of the fact that the frequency of chickens’ voice signal was higher than the ambient noise, the recorded voice signal was processed with pre-emphasis. The high pass filter was used, so as to weaken the signal of the noise and improve that of chickens’ voice. Thirdly, the processed chicken voice signal was further treated with the hamming window, and then it was divided into smaller segment, 21.3 ms per frames, which could be regarded as quasi steady state process. Fourthly, because the spectral entropy values of the obtained chickens’ voice and the noise were significantly distinguishing, the values of each frame were calculated out. Based on these values, the end point detection method was put forward, so that the chickens’ voice fragments were extracted from the complex ambient noise-containing record, while the non-chicken voice was discarded. Fifthly, the extracted chickens’ voice fragments were treated with time domain analysis, and 3 attributes (short time zero crossing rate, short time energy and the combination of them) were figured out as the characteristics of the healthy chickens and chickens infected with avian influenza. The 450 sampling voice of the healthy chickens and 450 of chicken infected with avian influenza were marked before their order being randomly disrupted. The marked samples were divided into 4 groups: 1 training set (600 samples) and 3 testing sets (100 samples in each group). Finally, the training set was trained by 3 Takagi-Sugeno (T-S) fuzzy neural networks (each with different types of the membership function: π function, Gaussian function and Bell function). It was revealed from the training result that the network with the bell function had the highest recognition rate. So the network with bell shape function was applied to the 3 testing sets and results were obtained respectively: the sensitivity was 75.47%, 80.39% and 76.92%, the specificity was 80.85%, 79.59% and 72.92%, and the true recognition rate was 78%, 80% and 75%. Therefore, this kind of detection method might provide a set of non-invasive, rapid and efficient methods for avian influenza infected chickens detection or identification in poultry farms and poultry circulation market.

neural network; recognition; extraction; spectral entropy;short time zero crossing rate; short time energy; chicken infected with avian influenza detection

10.11975/j.issn.1002-6819.2019.02.022

TP3-05

A

1002-6819(2019)-02-0168-07

2018-07-04

2018-12-30

國家重點研發計劃項目資助（2018YFD0500705）

張鐵民，教授，博士，主要從事智能檢測與控制研究。Email：tm-zhang@163.com

張鐵民，黃俊端. 基于音頻特征和模糊神經網絡的禽流感病雞檢測[J]. 農業工程學報，2019，35(2)：168－174 doi：10.11975/j.issn.1002-6819.2019.02.022 http://www.tcsae.org

Zhang Tiemin, Huang Junduan. Detection of chicken infected with avian influenza based on audio features and fuzzy neural network[J]. Transactions of the Chinese Society of Agricultural Engineering (Transactions of the CSAE), 2019, 35(2): 168－174. (in Chinese with English abstract) doi：10.11975/j.issn.1002-6819.2019.02.022 http://www.tcsae.org