叫聲在生豬福利監測中的研究進展與挑戰

Pig Sound Analysis： A Measure of Welfare

JI Nan， YIN Yanling， SHEN Weizheng* ， KOU Shengli， DAI Baisheng， WANG Guowei

（College of Electrical and Information， Northeast Agricultural University， Harbin 150030， China ）

Abstract： Pig welfare is closely related to the economical production of pig farms. With regard to pig welfare assess‐ ment， pig sounds are significant indicators， which can reflect the quality of the barn environment， the physical condi‐ tion and the health of pigs. Therefore， pig sound analysis is of high priority and necessary. In this review， the rela‐ tionship between pig sound and welfare was analyzed. Three kinds of pig sounds are closely related to pig welfare， including coughs， screams， and grunts. Subsequently， both wearable and non-contact sensors were briefly described in two aspects of advantages and disadvantages. Based on the advantages and feasibility of microphone sensors in contactless way， the existing techniques for processing pig sounds were elaborated and evaluated for further in-depth research from three aspects： sound recording and labeling， feature extraction， and sound classification. Finally， the challenges and opportunities of pig sound research were discussed for the ultimate purpose of precision livestock farming （PLF） in four ways： concerning sound monitoring technologies， individual pig welfare monitoring， commer‐ cial applications and pig farmers. In summary， it was found that most of the current researches on pig sound recogni‐ tion tasks focused on the selection of classifiers and algorithm improvement， while fewer research was conducted on sound labeling and feature extraction. Meanwhile， pig sound recognition faces some challenging problems， involv‐ ing the difficulty in obtaining the audio data from different pig growth stages and verifying the developed algorithms in a variety of pig farms. Overall， it is suggested that technologies involved in the automatic identification process should be explored in depth. In the future， strengthen cooperation among cross-disciplinary experts to promote the development and application of PLF is also nessary.

Key words： pig sound classification; animal welfare; sound analysis; feature extraction; precision livestock farming; sound monitering

CLC number：S-1; S828"""""""""" Documents code：A""""""""""""""""""""""" Article ID：SA202204004

Citation：JI Nan， YIN Yanling， SHEN Weizheng， KOU Shengli， DAI Baisheng， WANG Guowei. Pig sound analysis： A measure of welfare[J]. Smart Agriculture， 2022， 4（2）：19-35.（in English with Chinese abstract）

紀楠， 尹艷玲， 沈維政， 寇勝利， 戴百生， 王國維.叫聲在生豬福利監測中的研究進展與挑戰[J].智慧農業（中英文）， 2022， 4（2）：19-35.

1 Introduction

Pork" consumption" holds" a" stable" increase" inthe worldwide. As an example， pork is the most con ‐sumed" meat" in" China. In 2021，" pork" productionreached 52.96 million tonnes， accounting for morethan half of the total output of pork， beef， mutton， and poultry[1，2]. The huge demand for pork has led to a growing trend toward of modern pig production. Production" intensification" and" specialization" are two typical characteristics of modern pig farming[3] which contribute to increased productivity of pigs， leading" to" the" economic" efficiency" of production. Simultaneously， animal welfare is being a concern with the farm mode transformation in pig herds.

In general， animal welfare includes three parts， i. e.， natural living， affective states， and basic health and functioning[4] in different behaviors and various conditions. Monitoring" animal" body" conditions" is beneficial for both animals and farmers. First of all， animal" health" could" be" improved，" which" decrease usage" of veterinary" drugs" and" reduced" mortality. Meanwhile， better animal health contributes to bet‐ ter animal emotion[5]. Secondly， less cost in veteri‐ nary bills leads to direct financial benefits for farm ‐ er and improved quality of pork[6]. Most important‐ ly， animal health may directly affect human health. It can reduce the risk of zoonoses to monitoring the physical conditions of animals. It is becoming even more" crucial" to" investigate" the" relationships" be‐ tween good welfare， good health， and disease resis‐ tance with" enormous commercial and" social bene‐ fits in setting higher standards.

Precision" livestock" farming （PLF） is to" assist farmers" in" making" appropriate" management" deci‐ sions to avoid certain risks by using real-time moni‐ toring" technologies. Some" reviews" related" to" pig welfare and relevant technologies have been report‐ ed recently. Mahfuz et al.[7] provided a general over‐ view and instruction on smart tools and applications in modern pig farming. Both non-invasive and inva‐ sive methods were involved and discussed. Tzanida‐ kis et al.[8] summarized three main categories of non- intrusive technologies， including camera-based， mi‐crophone， and communication information technolo ‐gy （CIT） sensors， and attempted to predict techno ‐logical" developments" in potential ways. Schillingset al.[9] affirmed the impact of sensor applications onanimal welfare from two aspects of health and emo ‐tion. Meanwhile， both benefits and underlying risksof PLF were explored and discussed. Furthermore，Racewicz et al.[10]analyzed different technologies toachieve effective monitoring of pig health. They em ‐phasized" that" pig" health" and" welfare" measuresshould be integrated with the data obtained to estab ‐lish reliable monitoring systems for pig productionassessment. As" a" diverse" range" of" smart" sensorsemerged from technology development，" their" com ‐mercial" generation" possibilities" and" contributionsto welfare were also investigated and evaluated[4].

The above reviews briefly described the rele‐vant technologies applied in pig farms， demonstrat‐ing the potential and value of PLF technology ex ‐plicitly. Moreover， it is worth remarking that previ‐ous studies have acknowledged sound as a potential‐ly" useful" indicator" for" inferring" animal" welfare.Sound" analysis" techniques" and" sensors" have" beendeveloped" rapidly" in" recent" years，" which" providethe" opportunities to" achieve" automated monitoringpigs and improve their welfare. The main purposeof this paper is to focus on the relationship betweenpig sounds and welfare based on microphone moni‐toring. The existing pig sound processing technolo ‐gies" were" summarized" and" discussed" in" detail.Shortcomings and outlooks in terms of the relatedtechnologies were also elaborated. Finally， the chal‐lenges" and perspectives" of sound monitoring withPLF in modern pig farms were proposed.

2 Pig sound and welfare

Sound carries emotional， physiological and in ‐dividual information[11， 12]. It could be considered aspotentially valuable indicators for discerning animalwelfare. Table 1 illustrates different welfare indica‐ tors related to pig sounds. It can be seen that studies on" pig" sounds" were" mainly" devoted" to" coughs， screams， and grunts. In these articles， cough sounds can reflect air quality in pigsties and become crucial indicators for health monitoring. Regarding screams and grunts， they provided the reactions to the physi‐ cal conditions， which are beneficial to improve the pig welfare.

2.1 Pig sound and environment

In an intensive pig farm， air quality and heat stress are two influential factors associated with the living" environment，" directly" affecting" pig" welfare and product" quality. Ferrari" et" al.[13]" assessed heat stress" by" analyzing" continuous" pig" screams" and grunts. It was reported that a peak frequency value higher than 750 Hz of both sounds was considered as an indicator of heat stress[13]. Amaral et al.[14] dem ‐ onstrated the relationship between the" sound pres‐ sure levels （SPL） produced by piglet and the ther‐ mal" environment" of the" pig" nursery. A" range" of 56.3～60.3 dB was regarded as a good indicator to assess" thermal" comfort. Meanwhile，" cough" sound was regarded as an objective and non-invasive bio ‐ marker for the respiratory state in studies of expo ‐ sure to air pollutants[15， 16]. Also， the results indicated that" cough" sound" analysis" could" provide" valuable and" qualitative" information" about" the" air" quality conditions" in" commercial" livestock" farming，" with 1737 Hz on average peak frequency of pig cough （472 Hz higher than the control group） under better indoor air quality[16]. Pessoa et al.[17] assessed a base‐ line of cough free of respiratory disease and investi‐ gated the relationship between environmental condi‐ tions and cough frequency of pigs. It revealed a pos‐ itive correlation between ammonia （NH3） concentra‐tion changes and continuous coughing of pig.

2.2 Pig sound and physical condition

Sound" analysis" has" been" applied" to" evaluatepig" physical" condition" such" as" body" temperaturechanges， pain， hunger and thirst. Moi et al.[18] identi‐fied" the" differences" in" swine" vocalization patternsaccording" to" different" stress" conditions （thirst （noaccess" to" water）， hunger （no" access" to" food），" andheat stress）. Pig was found to be thirsty when soundintensity ranged from 73.87 dB to 80.18 dB. With avalue" higher" than 80.18 dB，" it" indicated" that" thepigs were hungry or under heat stress. For furtherconfirming the pig's status， pitch frequency present‐ed a difference， with the hunger of 212.87～276.71Hz and heat stress of higher than 276.71 Hz[18]. Theabnormal conditions by analyzing grunt， frightenedscreams and feeding howl sounds were also detect‐ed[19， 20]. The results showed that the total sound rec‐ognition rate could achieve 95.5%[19]. Besides， vocal‐ization is a valuable tool for identifying" situationsof stress in pigs during the castration procedure[21， 22].Without" local" anaesthesia，" piglets" uttered" almosttwice screams during the experiment. Also， screamscharacteristics" are" significantly" different" fromgrunts[23]. Moreover，" different" acoustic" parameterswere beneficial for evaluating the level of pain inpiglet management. The results showed that the val‐ues" of" pitch，" intensity，" and" maximum" amplitudewere enhanced from pigs in normal status to castra‐tion[24]. Based on the researches of screams charac‐teristics，" representative" features" were" focused" onand taken into consideration to define pig screamsfor constructing a more accurate classifier. And pigscreams were defined when the pig sound durationwas" longer" than 0.4 s[25]. A" simple" voting" systemwas constructed to classify the screams with a preci‐sion" of 83%[25]. With" respect" to" specific" emotionanalysis of pigs， Riley et al.[26] and Moura et al.[27] proved that phonations increased with fear and dis‐tress piglets. Grunting was also found to be highlyvariable， with the lowest grunting for happy emotion..

2.3 Pig sound and health

Both screams and coughs could be considered as direct indicators for monitoring pig healthy con ‐ dition. For instance， the screams of healthy piglets and sick ones （affected by traumatic arthritis） were evaluated by the presence" of significant differenc‐ es[51]. Although artificial neural network was sensi‐ tive" to" more" errors" in" discriminating" between healthy" and" sick pigs， it was meaningful to prove the feasibility by using screams[51]. In existing stud‐ ies， cough analysis is predominant among all kinds of pig sounds， especially wasting diseases and respi‐ ratory diseases.

It could be found that early efforts focusing on pig" cough" detection" in pig herds were undertaken under laboratory conditions from a successive study by" Katholieke" Universiteit" Leuven" in" Bel‐ gium[28， 30， 31， 33， 36]. As demonstrated in the studies， the differences" generated" by" the" sound" analysis" con ‐ firmed the variability of sound parameters depend‐ ing" on" the" health" status" or" disease" of the" animal. Due to lesions of the respiratory system， infectious cough" sounds" differed" from" non-infectious" cough sounds. Also， it was proved feasible to complete the pig cough classification from various sounds during a" series" of trials. And then， the" experiments were transferred to a commercial pig farm for further test‐ ing[29， 32， 34]. As" expected，" the" most" intuitive" perfor‐ mance was the decrease in accuracy of cough detec‐ tion in a complex commercial farm compared to a controlled environment[34]. The main reason was that background noise was the biggest interference fac‐ tor in sound detection. Although classification per‐ formance was much lower， it could still be regarded as" an" indicator" of disease" in pig" farms. These re‐ searches made crucial contributions since the results targeted" different" frequency" ranges" of pig" sounds and served as a basis for the development of morecomplicated models.

Based on the large number of basic investiga‐tions" of pig" coughs，" Korea University" focused" onthe study of wasting disease. The experiments wereperformed on a commercial pig farm and aimed atclassifying" the" different" porcine" wasting" diseases，such" as" Postweaning" Multisystemic" Wasting" Syn ‐drome （PMWS）， Porcine Reproductive and Respira‐tory Syndrome （PRRS） virus， and Mycoplasma Hy ‐opneumoniae （MH）[35， 37， 38， 40]. The" research" wasshown" to" be" robust" against" noises" in" pig" herds.Moreover， a low-cost sound sensor system was sug ‐gested" to" be" applied" in" small" and" medium-sizefarms with limited budgets [40].

Previous" researches" have" contributed" signifi‐cantly to the study of pig cough sounds. While theresearches were conducted in the last decade in Chi‐na. In the early stage， research on sound recognitionof predelivery Meishan sow was tried in laboratoryconditions[43]. A team" of Nanjing Agricultural Uni‐versity transferred the experiment from the laborato ‐ry to a pig farm to collect sounds and conduct an in-depth study on the sound of Meishan sows[44]. Mean ‐while，" Taiyuan" University" of" Technology" carriedout a" series of experiments with machine learningon pig" cough" characteristics， localization， recogni‐tion" algorithms，" and" multi-sensors" co-monitor‐ing[45-48]. Based on the ongoing development of tech ‐nology，" Huazhong" Agricultural" University" applieddeep learning methods in pig cough recognition andachieved satisfying results[49， 50]. Besides the researchteams mentioned above， many researchers in Chinaare" focusing" on" the" performance" improvement" ofcough" sound recognition by" adopting" and" finetun ‐ing various algorithms. Breakthroughs in technolo ‐gies， including machine learning and deep learning，are boosting the development of PLF.

2.4 Summary

In summary， as an indicator of animal welfare， pig" sound has been well-tested and" some" findings have been generated in progress. However， it is evi‐ dent that the current researches on pig sounds main ‐ ly concentrate on coughs， screams and grunts. The meanings of other pig sounds and their relationship with welfare need to be further studied. Meanwhile， it is relatively" straightforward that the majority" of previous studies were targeted at improving a partic‐ ular type of pig welfare based on qualitative or quan ‐ titative" studies. Indeed，" a" specific" sound points to multiple" aspects" of animal" welfare. This" situation dramatically" increases" thenbsp; soundscape" complexity， especially in the commercial pig farm. As a whole， the relationship between pig sound and welfare is a combination" of" multiple" occurrences" rather" than changes caused by a single factor. This has become a major issue for animal welfare research based on pig sounds. Meanwhile， it deserves to be a" significant study. For example， it is possible to distinguish dif‐ ferent coughing sounds as further management indi‐ cators. Whether the coughing sound is triggered by diseases （pig wasting diseases and respiratory diseas‐ es） or" air pollutant" in the pigsty. Such refinement would have a profound effect on the global improve‐ ment of the PLF and positive welfare.

3 Sound analysis

Benefit from the development of sensor tech ‐ nology，" animal" welfare" could" be" monitored" in" di‐ verse" manners[7]. In" general，" monitoring" sensors available" in" pig" farms" could" be" divided" into" two types，" namely" invasive" sensors" and" non-invasive sensors. RFID and accelerometers are two common ‐ ly used invasive sensors[53]. The advantage of inva‐ sive" sensors" is" that" they" satisfy" the" identification and" tracking" requirements" of" individual" informa‐tion. In contrast， the disadvantages are also apparentin two aspects. Injury， pain， and stress are broughtto pigs when attaching a tag， which goes against ani‐mal welfare. Another limitation is that the devicesare not" easy to maintain. Non-invasive" equipmentfrequently" used" in" pig" farming" include" camera-based" sensors，" microphones，" and" infrared" thermalcameras[54]. The" advantage" is" relatively" easy" tocheck equipment in time and to reduce the pressureon" pigs. However，" researches" are" still" focused" onthe group monitoring level. Improving the accuracyof individual monitoring is one of the challenges ofnon-invasive equipment.

Among the non-invasive" equipment， a micro ‐phone has been adopted for pig sound recognitionand welfare assessment benefitting from its non-in ‐vasive and continuous monitoring merits[10]. Statisti‐cally， the number of studies based" on microphonetechnology is the fourth highest among the existingsmart" technologies[4]. Welfare" monitoring" researchare in high demand. Meanwhile， it shows relativelypotential for commercial implementation due to itslow cost of devices. It is an important component ofPLF by" combining" technological" advancements" inthe management process and animal behavior[55， 8].

In general， the process of sound analysis con ‐sists of four steps， namely， sound recording[56] ， indi‐vidual sounds labeling[57] ， sound feature extraction [58]and classification [59] ， as shown in Fig.1.

3.1 Sound recording and labeling

The first step is the collection of raw data re‐corded" by" microphone. Typically，" the" microphonewas placed around the pen in a pig room. Its posi‐tion" was" explored" in" field" conditions" in" terms" ofheight and the relative position from the walls anddisruptive sound sources such as ventilation. Also，the sampling rate could be adjusted. In most cases，44.100 kHz was used in the field condition. Another factor" to" consider" is" the" number" of microphones. Limited" by" experimental" conditions，" one" micro ‐ phone was adopted in most existing studies. Inevita‐ bly，" the" number" of microphones" could" also" affect the sound recording quality. Subsequently， the col‐ lected sound recordings need to be pre-processed to reduce background noises as much as possible for further processing" and" analysis. Pre-processing in ‐ cludes filtering， pr-emphasis， framing， and window ‐ ing，" etc. In" order" to" evaluate" the" performance" of classification algorithms， it is necessary to segment and" label the" continuous recording" data. The" seg ‐ mentation and labeling of individual" sounds could be completed automatically or manually. Currently， most of the studies on pig sounds were implement‐ ed in a manual labeled way. While few studies were focused" on" the" recording" labeling. The" double threshold endpoint detection method was frequently used in collecting individual sound segments from recordings[38， 45]. Besides， Li et al.[47] adopted the bi‐ rectional" long" short-term" memory-connectionist temporal" classification （BLSTM-CTC） model" to complete" the" continuous" cough" sound" recognition task， with a total accuracy of 93.77%.

Obviously， researches" on labeling" and" extrac‐tion of pig sound segments are lacking. Most stud‐ies" remained" at" the" manual-labeled" stage. This" ispartly" because" the" research" requires" the" involve‐ment of experts in animal research to give certaincriteria， but there is no unified standard yet. Anoth ‐er" reason" is" that" an" open-source" database" of pigsounds is scarce. Due to the different characteristicsof pig growth， it is necessary to capture informationof breed，" age， weight，" as well" as the" environmentand location where these sounds are produced. Allthese" information" would" enrich" the" diversity" andcomprehensiveness of pig sounds. A sufficient num ‐ber of sound samples are needed to extract valuableinformation and to eliminate distracting sounds. Al‐so， it is a key step to achieve high accuracy in thefollowing sound analysis.

3.2 Sound features extraction

After labeling individual sounds from continu ‐ous recordings， the" crucial" step" is to" extract valu ‐able" information" from" sound" signals，" which" istermed" as" audio" feature" extraction[60]. In" existingstudies，" various" features" were" extracted" from" fivedomains， including time， frequency， cepstral coeffi‐cients， time-frequency， and deep features， as shownin Fig.2.

Time-domain" features" are" fundamental" fea‐tures" which" represent" signal" variation" regardingtime. Among them， duration and amplitude are of‐ten" chosen" to" explore" the" basic" information" andproperties" contained" in" sound" itself. Duration" is" akind" of rhythm-based" feature，" which" represents" aregular" recurrence" of" patterns" over" time. It" wasproved that the average duration for infectious andhealthy coughs were 0.67 s and 0.43 s in the lengthof" a" single" cough，" representatively[31]. While" themaximum amplitude refers to the maximum ampli‐tude of the sound wave， which was used to estimatethe" level" of pain" of piglets[24]. The results" showed that maximum" amplitude was" growing" from pain- free" to" castration，" with" the" value" ranging" from 0.2683 Pa to 1 Pa [24]. Root mean square （RMS） is an energy based features that can be used to mea‐ sure the sound loudness. The RMS value of non-in ‐ fectious pig" coughs has been proved to be higher than that of infectious pig coughs[31]. Another energy based" feature" called" short-time" energy （STE） is commonly combined with zero crossing rate （ZCR） to detect voice activity. It was found that the STE of piglets grunts was highly variable in distress[26].

Features extracted from the frequency domain are effective ways in conducting signal processing of pig sound. Simply， a peak is an index of the maxi‐ mum power of a signal. The average peak frequen ‐ cies of infected and healthy pig coughs were calcu ‐ lated of 618 Hz and 1603 Hz[24] ， which means it can be used to effectively distinguish the coughs of in ‐ fected" and" healthy pig. Tonality based" features" in terms of fundamental frequency and pitch were uti‐ lized to monitor pig vocalization， especially for de‐ tecting whether the pigs were in normal. When the av ‐ erage value of formants was lower than 2671.99 Hz and duration of the signal was lower than 0.28 s， the

piglets" were" proven" to" be" in" normal" condition[61].

Otherwise， pigs could be considered in an abnormal

state. Moreover，" the" pitch" could" help" identifyingsex， age， and distress[52]. It indicated that the pitchvalue" of female" pigs （218.2 Hz） was" higher" thanmale pigs （194.5 Hz） when all the pigs were in thesame" conditions[52]. Also，" the" pigs" in" nursery" andgrowing stage held higher pitch， followed by the fin ‐ishing stage[52]. Spectrum shape based features werealso" used" in" recognizing" various" sounds" sufferedfrom" different" diseases，" including" spectral" flux，spectral" spread" and" spectral" centroid[62， 63]. In" addi‐tion to the features mentioned above， power spectraldensity （PSD） and" energy" envelope" were" anothertwo common representative features in the frequen ‐cy domain， which were often used to distinguish pigcoughs and non-cough sounds.

The cepstrum is obtained by applying a Fouri‐er inverse transform to the logarithm of the signalspectrum. Developed" by" Davis" and" Mermelstein，Mel frequency cepstrum" coefficients （MFCCs） arecommonly" utilized" in" human" speech" and" animalsound recognition[64]. Both the" original" coefficientsand their first-order or second-order coefficients areadded and combined as the acoustic features in theprocess of feature extraction. For instance， the first 20 coefficients" were" extracted" as" a" whole" feature vector" for" discriminating" infectious" coughs" in pigs[65]. To reflect both static and dynamic character‐ istics， 12-dimensional" original" and 12-dimensional first-order" delta" coefficients" were" calculated" from each" cough" sound" sample[16]. Furthermore， 39-di‐ mension" MFCCs，" combining 13-dimensional" MF ‐ CC and first-order as well as second-order differen ‐ tial" coefficients， were" obtained" for" continuous pig cough sound recognition[39]. In addition， linear pre‐ diction cepstral coefficient （LPCC） and its first-or‐ der differences were also utilized in detecting abnor‐ mal status of dry and wet cough sounds[45].

For" time-frequency" features，" one-dimensional audio" signals are" frequently transformed into two- dimensional time frequency representations. Among them， short-time Fourier transform （STFT） and Mel- STFT" spectrograms" are" frequently" used" in" pig sound recognition in two ways. One is that spectro ‐ grams are combined with deep learning models. In this way， the process of hand-crafted features is not required. For instance， STFT spectrograms were ap ‐ plied to Alexnet and MnasNet deep architectures， re‐ spectively[40， 41]. While Mel spectrograms were adopt‐ ed" to" convolutional" block" attention" module" with convolutional neural networks （CNN） for recogniz‐ ing abnormal pigs sounds[66]. The other way is ex ‐ tracting deep features based on deep learning mod‐ els， which are regarded as feature extractors. Lee et al.[67] extracted deep features from a 6-layers CNN and put into muti-layer perception for pig wasting diseases" classification[67]. A" MFCC-CNN" feature was extracted from a one-layer CNN and put into a support" vector" machine （SVM） classifier" for" pig cough recognition in Shen et al.[42]

Up to now， most studies aimed at the time or frequencynbsp; domain" features" of pig" sounds. Fewerstudies have been conducted on time-frequency do ‐main" as well as deep" features. Besides， other fea‐tures could be considered in the pig sound analysis.For instance， harmonicity is utilized to distinguishtonal" and" noises，" which" have" been" used" in" birdsound classification[68，69]. Spectrum shape based fea‐tures" have" been" used" in" music" and" animal" soundclassification，" including" spectral" centroid，" spectralroll" off，" spectral" flatness，" spectral" bandwidth[70-72].Moreover，" other" time-frequency" representationscould" be" investigated" in" time-frequency" featuresand deep features， in terms of MFCC[73] ， mel-scaledspectrograms[74] ， constant-Q transform （CQT）[75].

3.3 Sound classification

In previous researches， mathematical modelingof analyzing and classifying pig sounds can be di‐vided into three main categories： statistical analysis，machine learning and deep learning.

Statistical analysis was used to complete funda‐mental" research" on" the" pig" sound" characteristics.For instance， one-way analysis of variance （ANO ‐VA） was one of the most frequently used statisticalanalyses[76]. It was" shown that healthy" coughs hadmuch higher peak frequencies （750～1800 Hz） thaninfectious coughs （200～1100 Hz）[34]. Also， a signifi‐cant" difference （Plt;0.001） was" observed" betweennon-infectious coughs （a mean duration of 0.43 s）and infectious coughs （mean duration from 0.53 s to0.67 s）[34]. Thus， single cough duration could be re‐garded as an indicator to classify different kinds ofcough" sounds[77]. Subsequently，" ANOVA" has" beenfurther" used" to" distinguish" pig" wasting" disease[37].The results" indicated that no" differences" in" coughdurations between normal coughs and coughs withdiseases[37]. In addition， there is a significant differ‐ence between porcine circo virus type 2（PCV2） andother coughs （normal， porcine reproductive and re‐spiratory syndrome （PRRS） and Mycoplasma hyo ‐ pneumoniae （MH） cough sounds） in pitch， intensity， and" formants 1， 2， 3，" and 4[37]. Not" only" cough sounds but also grunts and screams were analyzed using" the" statistical" analysis" to" assess" heat" stress and evaluate the level of pain on pig farms[13，24]. The results" showed" the" differences" in" pig" grunts" and screams，" which" was" beneficial" for" pig" production management in a good welfare way[13，24].

Machine" learning" has" demonstrated" superior performance in many fields[78]. Fuzzy c-means clus‐ tering was used to form two clusters： cough and non- cough sounds [79] ， including in laboratory installation with nebulization of citric acid[33] and pathologic dis‐ ease and under aerial pollutant control design with ammonia， dust， and temperature[15]. The overall per‐ formance of identified sounds （chemically induced coughs，" sick" coughs，" and" other" sounds） achieved 85.5%[33]. Moreover，" average" cough" classification achieved 94% in different experiment designs with various aerial pollutants[15]. SVM is suitable for clas‐ sification in both linear and nonlinear ways[80]" and was used in a variety of fields among pig sounds， es‐ pecially in pig cough recognition. It was shown that the average detection accuracy of wasting disease ap ‐ proached 98.4%[81]. Subsequently， Wang et al.[16] pro ‐ vided an average recognition rate of 95% for cough sounds" in" different" air" qualities. Shen" et" al.[42] achieved a cough recognition of 97.72%. A decision tree was built for classifying diverse conditions in different status， such as thirst， hunger， and thermal stress[18] ， cold and pain[61] ， as well as distress condi‐ tions[52].

Deep learning is a popular tool in recent years， contributing to its strong ability in pattern recogni‐ tion[82-84]. Some deep learning models have been fine‐ tuned" to" be" applied" in" pig" sound" recognition" in terms" of" CNN" and" recurrent" neural" networks（RNN）. For" CNN" models，" Yin" et" al.[41]" finetunedAlexnet model to recognize the pig coughs， with anaccuracy of 96.8%. Although CNN was proved tobe effective in recognizing spectrograms， but CNNinevitably generated various redundant informationduring" the" process. For" this" reason，" an" attentionmechanism" named" convolutional" block" attentionmodule （CBAM） was" introduced" for" optimizingCNN[66]. The study provided a satisfying recognitionrate of abnormal pig sounds with 94.46%[66]. Sincedeep neural networks require greater computationalcapacity" and" higher" hardware" requirements，" theseconditions become" one" of the" factors" limiting pigsounds research" into practical" applications. To" ad‐dress this issue， researchers introduced lightweightmodels" to" pig" sound" classification. A" lightweightmodel" based" on" MnasNet" and" MobileNetV2 wasused to classify pig sounds with different pig diseas‐es， and got an F1-score of 94.7%[40] and a total accu ‐racy of 97.3%[19]. For RNN models， not only RNNbut" also" its" variant" models" including" long" short-term memory （LSTM）， BLSTM， CTC and gate re‐current unit （GRU） were applied in pig cough recog ‐nition[49， 50， 85]. The" results" proved" that" RNNs" wereable to be feasible and stable models for completingthe classification task[85].

3.4 Summary

Over" the" years，" many" bioacoustics" modelshave been developed to analyze and recognize pigsounds. Statistical analysis is the scientific methodsof" collecting，" exploring，" and" presenting" largeamounts of data to discover underlying patterns andtrends[86] ， and it is a classical method for calculatingvariations among variables. A limitation of empiri‐cal methods is that their applicability is often con ‐strained to the conditions during experimental test‐ing. When conditions exceed the scope of the inves‐tigation， they may not be applicable. For instance， one-way ANOVA requires that the dependent vari‐ able is normally distributed in each group and that the" within-group" variability" is" similar" across groups. However， it is better to ignore assuming the distribution" of the" data" and" employ" it" directly" for prediction" in" solving practical" sound" classification problems[87]. For" comparison，" machine" learning （ML） provides complementary data modeling tech ‐ niques and has become a more desirable approach to handling complex data sets. An advantage of ML is the flexibility， which means it contains a number of adjustable parameters. However， it also introduc‐ es certain complexity in the selection of parameters for better fitting the model. Meanwhile， the predic‐ tive" results" are" relevant" to" the" selected" features， which becomes one" of the" factors" sensitive to the choice of ML algorithm. To this end， deep learning models" dominate" great potential" in" addressing" the problem" of automatic" extraction" of abundant" fea‐ tures from original data. However， deep learning re‐ lies on training and modification of the model by a large amount of data， leading to more complex sim ‐ ulation and computation. Therefore， machine learn ‐ ing is still a useful and continuously researched ap ‐ proach. By studying the existing literature， it can be found that the selection of a classifier is still subject to a certain degree of randomness. The most" suit‐ able" classifier" should be" further validated" on run ‐ ning time for achieving the trade-off between accu ‐ racy and processing speed.

4 Challenges and perspectives

4.1 Sound monitoring

It" could" be" found" that" a" specific" production phase was commonly targeted in pig sound analysis in Table 1. Specifically， fattening pigs hold the high ‐ est percentage， followed by sows and piglets， whichis consistent with Gómez et al[4]. Measurement" andfurther validation of pig production stages are rea‐sonably necessary" and" still" lacking. On" one hand，the diversity of pig sounds is present at all growthstages. In" other words， monitoring typical" sounds，such" as" coughs" and" screams，" is" required" at" eachgrowth" stage. Therefore，" sound monitoring" can beenhanced by" expanding the range" of study" stages.On the other hand， due to the outbreak of Africanswine" fever virus （ASFV），" stricter management" isconducted" in" big" commercial" pig" farms" and" re‐searchers" are" temporarily" curtailed. As" a" result，" itleads" to" intermittent" experiments" in" pig" farms.This" situation" will" be" improved" as" the" epidemiceases. Meanwhile，" continuing" experiments" can" beconsidered" from" small" and" medium-sized" pigfarms in the future.

In addition， with regard to pig sound localiza‐tion， it is an important topic to be investigated in thefuture. On one hand， it locates the pig with healthyproblems， which is better to" enhance the manage‐ment of pig herds. On the other hand， the relevantstudies are fewer and still stand in the early explor‐atory stage in the field. Currently， time difference ofarrival （TDOA） between different microphones wasapplied" in" pig" cough" localization[30， 47， 88]. Althoughthe current positioning results are proved to be feasi‐ble in pig houses （mean error less than 1 m）， a chal‐lenging issue also comes to the surface， namely thetrade-off between the number of microphones" andtheir cost. These results are instructive and meaning ‐ful during the experimental phase for further studiesand the number of tested microphones is acceptable.However， when considered for application in a com ‐mercial context， the cost associated with each addi‐tional" microphone" is" undoubtedly" expensive. Thisalso motivates researchers to deepen the cough posi‐tioning research and continuously optimize the ex ‐periments. The aim is to achieve an optimal balance between the number of microphones， positioning ac‐ curacy and cost. Another problem is how to locate and track sick pigs in real-time， since the pig is a moving target. In addition， it is worthwhile to study how to solve the problem of multi-targeting local‐ ization when the number of targets with abnormal conditions is large. By far， sound recognition is also relatively difficult to locate from group recognition to individual recognition. Given that sound localiza‐ tion" studies" are" still" scarce， pig" sound" analysis" is suggested to be developed for sound localization in the future.

In" general， the" current researches were based on a laboratory or a specific pig farm with great dif‐ ferences" in" size，" environment，" and" individual" pig conditions. Despite modern technologies have been implemented to analyze the collected" sounds from the barn， expected results are" still not available at this" stage. This" is because the" combination" of the pig" housing" environment" and" the" individual" body condition affects the variability of the pig's vocaliza‐ tions. Therefore，" the" current" methods" are" insuffi‐ cient for the complex interaction between pigs and their complex environment. This has become a ma‐ jor factor that makes it difficult to popularize PLF at present[89]. To" address this" issue，" it" is" suggested that attention could be focused on interactions be‐ tween multiple monitoring modules rather than con ‐ centrating on individual processes in the future. It is possible" for us to" find the most" appropriate" inter‐ face between multiple modules by interpreting mul‐ tiple sets of inputs from a variety of biological re‐ sponses. Moreover， it could be a better way to han ‐ dle the whole PLF process.

Cough based identification monitoring technol‐ ogy is still at a highly technology dependent stage of development. No" integration" of animal welfarehas been considered to date. This could be due to alack" of well-developed" rating" system" between" thewelfare indicators and the pig sounds. Hence， it isessential and extremely" significant for applying topig production. To address the weakness， the partici‐pation" and" cooperation" of researchers" in" differentdisciplines should be strengthen in the future. Thelatest study by Silva et al.[90] focused on generatingtechnical" monitoring" indicators" by" monitoring" thefrequency of coughs in the pig house and the corre‐sponding disease diagnosis. Their preliminary find‐ings validated those dry and non-productive coughsindicated" the" presence" of" Mycoplasma" hyopneu ‐moniae. Although these researches are essential andsignificant，" they" have" not" yet" been" carried" out" inChina. It is recommended to refine the details of ex ‐periments in pig farms by combining sound analysisand pig diseases in the future.

4.2 Individual welfare

Although" microphone" sensor-based" sound" lo ‐calization techniques are constantly being upgraded，they" are" only" capable" of narrowing" the" range" ofsound monitoring as much as possible. It is still dif‐ficult to be precise about the individual pigs in thisway. However，" individual" identification" of pigs" isnecessary. As an example， when coughs are moni‐tored to" occur" in" a" certain pen，" it" is" important toidentify and separate the coughs pig from the herd.Therefore， it is promising to aggregate different sen ‐sors together to promote individual pig welfare inthe" future. For" instance，" it" could" be" attempted" toidentify unhealthy pigs by using a facial recognitionsystem based on camera-based technologies to over‐come the limitations of sound monitoring. In addi‐tion，" computer" vision" can" provide" information" onbehavioral interactions between individuals， includ‐ing the detection of aggressive events and mood ele‐vating" behaviors. Another" promising" approach" is the application of remote video monitoring technol‐ ogy. Individual pig behaviors can be monitored vi‐ sually in this way， including body condition， lame‐ ness， feed intake， and oestrus [8].

4.3 Commercial applications

Currently， most of the researches on pig vocal‐ ization monitoring are still in the developing stage， with few advanced commercial products in terms of SoundTalks[91]" and" STREMODO" system[92]. Sound‐ Talks is a cough monitoring developed by a Belgian company， which is used to measure cough sounds in an automated and continuous way. The STREMO ‐ DO system proposed by Germany company records and" assesses" stress" vocalization" in pig" group. Be‐ sides， Yingzi Technology established in China is a promising" and" new" technology" company[93] ，" which strives" to" create" a" multifaceted" platform" for" data- driven agriculture and to facilitate pig farms devel‐ opment involving livestock management and biose‐ curity. The microphone-based method requires" ex ‐ tensive consideration of various factors such as vari‐ ability and diversity in the barn environment of pig herds. However， it is not easy to meet the demand for" accuracy" within" technical" reliability" and" low cost within equipment maintenance.

4.4 Farmer concerns

The goal of PLF will lead researchers to up ‐ grade the technology into productization eventually. However， in addition to concentrating on technolo ‐ gy， concerns from pig practitioners" should also be taken" into" account. After" all，" the" practitioners" are the key to implement PLF in pig farming. Concerns from farmers are raised in a few aspects， such as in ‐ ternal constraints in pig farms， technical costs， and technician support. Specifically， the application fea‐ sibility is also being tested by whether the hardwareconditions in the farm meet the requirements. Withsuitable" farm" conditions，" it" is" another" concernwhether the emerging technology can match the ex ‐pected results. In" addition， the" cost" of applicationand maintenance is" also worth" considering. More‐over， complex new knowledge may reduce motiva‐tion to" learn" and prevent" farmers" from" embracingnew technology.

Regarding these" concerns， the" following" sug ‐gestions" to" consider" in" the" future" are" proposed.First，" further" validation" between" PLF" techniquesand research into positive welfare indicators shouldbe encouraged to enhance the confidence and trustof pig practitioners. Secondly， equipment should bedesigned and optimized from the view of farmers toprovide convenience and reduce product costs. Fi‐nally， the awareness of farmers should be raised bystrengthening" training" and" communications. Al‐though PLF has not yet reached an industry consen ‐sus， it is highly topical. Moreover， pig practitionersare expecting PLF to assist in pig farming.

Contactless sound analysis is a promising wayto assess pig welfare. It is a key element of PLF andit has been proved to be a feasible way to promotethe PLF. Apparently， it is ongoing progress in pigproduction" and" key" technology. Overall，" consider‐ing the demand for farm intensification and betteranimal welfare， PLF could be a trend for livestockmanagers" to" boost" their" productivity" and" animalcomfort.

References：

[1] Development" situation" of pig" industry" in 2021 andprospect in 2022[J]. China Animal Industry， 2022（3）：39-40.

[2] GAO G. Development prospect of China's meat indus‐try in 2022[J]. Meat Industry， 2022（2）：1-5.

[3] NORTON" T，" CHEN" C，" LARSEN" M" L V，" et" al. Re‐view： Precision" livestock" farming： Building 'digitalrepresentations' to" bring" the" animals" closer" to" thefarmer[J]. Animal， 2019， 13（12）：3009-3017.

[4] GóMEZ Y， STYGAR A H， BOUMANS I J M ， et al. Asystematic" review" on" validated" precision" livestock farming technologies for pig production and its poten‐ tial to assess animal welfare[J]. Frontiers in Veterinary Science， 2021， 8： ID 660565.

[5] INGVARTSEN" K" L，" NutritionMOYES" K. Immunefunction and health of dairy cattle[J]. Animal， 2013， 7：112-122.

[6] DAWKINS" M" S. Animal welfare" and" efficient" farm ‐ing： Is conflict inevitable？[J]. Animal Production Sci‐ ence， 2017， 57（2）：201-208.

[7] MAHFUZ S， MUN H S， DILAWAR M A， et al. Appli‐cations of smart technology as a sustainable strategy in modern" swine" farming[J]. Sustainability， 2022， 14（5）：ID 2607.

[8] TZANIDAKIS C， SIMITZIS P， ARVANITIS K， et al.An" overview" of the" current" trends" in" precision" pig farming technologies[J]. Livestock Science， 2021， 249：ID 104530.

[9] SCHILLINGS J， BENNETT R， ROSE D C. Exploringthe potential" of precision" livestock" farming technolo‐ gies to help address farm animal welfare[J]. Frontiers in Animal Science， 2021， 2： ID 639678.

[10] RACEWICZ P， LUDWICZAK A，" SKRZYPCZAK E，et" al. Welfare" health" and" productivity" in" commercial pig herds[J]. Animals， 2021， 11（4）： ID 1176.

[11] MANTEUFFEL G， PUPPE B， SCH?N P C. Vocaliza‐tion of farm animals as a measure of welfare[J]. Ap‐ plied" Animal" Behaviour" Science， 2004， 88（1-2）：163-182.

[12] MCLOUGHLIN" M" P，" STEWART R，" MCELLIGOTTA G. Automated bioacoustics： Methods in ecology and conservation" and" their" potential" for" animal" welfare monitoring[J]. Journal of The Royal Society Interface， 2019， 16（155）： ID 20190225.

[13] FERRARI S， COSTA A， GUARINO M. Heat stress as‐sessment by" swine related vocalizations[J]. Livestock Science， 2013， 151（1）：29-34.

[14] AMARAL P I S， CAMPOS A T， YANAGI JUNIOR T，et al. Using sounds produced by pigs to identify ther‐ moneutrality" zones" for" thermal" environment" assess‐ ment" ratios[J]. Engenharia" Agrícola， 2020， 40（3）：266-271.

[15] HIRTUM AVAN， BERCKMANS D. Objective recogni‐tion of cough sound as biomarker for aerial pollutants： Aerial pollutants and cough sound[J]. Indoor Air， 2004， 14（1）：10-15.

[16] WANG X， ZHAO X， HE Y， et al. Cough sound analy‐sis to assess air quality in commercial weaner barns[J]. Computers and Electronics in Agriculture， 2019， 160：8-13.

[17] PESSOA J， CAMP MONTORO J， PINA NUNES T， etal. Environmental risk factors influence the frequencyof coughing and sneezing episodes in finisher pigs on afarm free of respiratory disease[J]. Animals， 2022， 12（8）： ID 982.

[18] MOI M， N??S I DE A， CALDARA F R， et al. Vocal‐ization data mining for estimating swine stress condi‐tions[J]. Engenharia Agrícola， 2014， 34（3）：445-450.

[19] CANG" Y，" LUO" S，" QIAO" Y. Classification" of" pigsounds based on deep neural network[J]. Transactionsof the" Chinese" Society" of Agricultural" Engineering，2020， 36（9）：195-204.

[20] ZHANG Z. Pig anomaly detection based on audio anal‐ysis" technology[D]. Taiyuan： Taiyuan" University" ofTechnology， 2017.

[21] LEIDIG" M" S，" HERTRAMPF" B，" FAILING" K，" et" al.Pain and discomfort in male piglets during surgical cas‐tration" with" and" without" local" anaesthesia" as" deter‐mined by vocalisation" and" defence behaviour[J]. Ap‐plied" Animal" Behaviour" Science， 2009， 116（2-4）：174-178.

[22] CORDEIRO A F DA S， N??S I DE A， OLIVEIRA SR DE M， et al. Efficiency of distinct data mining algo‐rithms for classifying stress level in piglets from theirvocalization[J]. Engenharia" Agrícola， 2012， 32（2）：208-216.

[23] MARX G， HORN T， THIELEBEIN J， et al. Analysisof pain-related vocalization" in young pigs[J]. Journalof Sound and Vibration， 2003， 266（3）：687-698.

[24] CORDEIRO A F DA S， N??S I DE A， BARACHO MDOS" S， et al. The use of vocalization signals to esti‐mate the level of pain in piglets[J]. Engenharia Agríco‐la， 2018， 38（4）：486-490.

[25] VANDERMEULEN J， BAHR C， TULLO E， et al. Dis‐cerning" pig" screams" in" production" environments[J].PLoS One， 2015， 10（4）： ID e0123111.

[26] RILEY J L， RILEY W D， CARROLL L M. Frequencycharacteristics in animal species typically used in laryn‐geal research： An exploratory investigation[J]. Journalof Voice， 2016， 30（6）： e 17-e24.

[27] MOURA D J， SILVA W T， NAAS I A， et al. Real timecomputer" stress monitoring" of piglets using vocaliza‐tion analysis[J]. Computers and Electronics in Agricul‐ture， 2008， 64（1）：11-18.

[28] MOSHOU" D，" CHEDAD" A，" HIRTUM" AVAN，" etnbsp; al.Neural recognition system for swine cough[J]. Mathe‐matics" and" Computers" in" Simulation， 2001， 56（4-5）：475-487.

[29] FERRARI S， SILVA M， GUARINO M， et al. Analysisof cough sounds for diagnosis of respiratory infectionsin intensive pig farming[J]. Transactions of the ASA ‐BE， 2008， 51（3）：1051-1055.

[30] SILVA M， FERRARI S， COSTA A， et al. Cough local‐ization for the detection of respiratory diseases in pighouses[J]. Computers" and" Electronics" in Agriculture，2008， 64（2）：286-292.

[31] FERRARI" S，" SILVA M， GUARINO M， et al. Coughsound" analysis" to" identify" respiratory" infection" in pigs[J]. Computers" and" Electronics" in" Agriculture， 2008， 64（2）：318-325.

[32] EXADAKTYLOS" V，" SILVA" M，" FERRARI" S，" et" al.Time-series analysis for online recognition and local‐ ization of sick pig （Sus scrofa） cough" sounds[J]. The Journal" of the Acoustical" Society" of America， 2008， 124（6）：3803-3809.

[33] EXADAKTYLOS V， SILVA M， AERTS J M， et al. Re‐al-time recognition of sick pig cough sounds[J]. Com ‐ puters" and" Electronics" in" Agriculture， 2008， 63（2）：207-214.

[34] GUARINO M， JANS P， COSTA A， et al. Field" test" ofalgorithm" for" automatic" cough" detection" in" pig houses[J]. Computers" and" Electronics" in Agriculture， 2008， 62（1）：22-28.

[35] GUTIERREZ W M， KIM S， KIM D H， et al. Classifi‐cation of porcine wasting diseases using sound analy‐ sis[J]. Asian-Australasian Journal of Animal Sciences， 2010， 23（8）：1096-1104.

[36] FERRARI" S，" SILVA" M，" EXADAKTYLOS" V，" et" al.The" sound" makes" the" difference： The" utility" of real time sound analysis for health monitoring in pigs[M]// ALAND" A，" BANHAZI" T. Livestock" housing. The Netherlands： Wageningen Academic Publishers， 2013：407-418.

[37] CHUNG Y， OH" S， LEE J， et al. Automatic detectionand recognition of pig wasting diseases using sound da‐ ta in audio surveillance systems[J]. Sensors， 2013， 13（10）：12929-12942.

[38] KIM H. Automatic identification of a coughing animalusing audio and video data[C]// The fourth Internation‐ al Conference on Information Science and Cloud Com ‐ puting — PoS（ISCC2015）. Guangzhou，" China： Sissa Medialab， 2016.

[39] ZHAO" J，" LI" X，" LIU" W，" et" al. DNN-HMM" basedacoustic" model" for" continuous" pig" cough" sound" rec‐ ognition[J]. International" Journal" of Agricultural" and Biological Engineering， 2020， 13（3）：186-193.

[40] HONG M， AHN H， ATIF O， et al. Field-applicable piganomaly" detection" system using" vocalization" for" em ‐ bedded" board" implementations[J]. Applied" Sciences， 2020， 10（19）： ID 6991.

[41] YIN Y， TU D， SHEN W， et al. Recognition of sick pigcough" sounds based" on" convolutional neural network in" field" situations[J]. Information Processing" in Agri‐ culture， 2021， 8（3）：369-379.

[42] SHEN W， TU D， YIN Y， et al. A new" fusion featurebased on convolutional neural network for pig cough recognition in field situations[J]. Information Process‐ ing in Agriculture， 2021， 8（4）：573-580.

[43] XU， Y， SHEN M， YAN L， et al. Research of predeliv‐ery Meishan sow cough recognition algorithm[J]. Jour‐nal" of Nanjing" Agricultural" University， 2016， 39（4）：681-687.

[44] ZHANG" H. Design" and" implementation" of Meishanpig" continuous" cough" sound" monitoring" system[D].Nanjing： Nanjing Agricultural University， 2020.

[45] ZHANG Z， TIAN J， WANG F， et al. The study on char‐acteristic parameters extraction and recognition of pigcough sound[J]. Heilongjiang Animal Science and Vet‐erinary Medicine， 2017（23）：18-22.

[46] HAN L， TIAN J， ZHANG S， et al. Porcine abnormalsounds" recognition" using" decision-tree-based" supportvector" machine" and" fuzzy" inference[J]. Animal" Hus‐bandry amp; Veterinary Medicine， 2019， 51（3）：38-44.

[47] LI J， TIAN Y， ZHANG S. Research on recognition andlocalization" of porcine cough" sounds[J]. HeilongjiangAnimal" Science" and" Veterinary" Medicine， 2020（14）：36-41.

[48] ZHANG" S. Research and application on multi-sourcemonitoring and information fusion method for porcineabnormal" behaviors[D]. Taiyuan： Taiyuan" Universityof Technology， 2020.

[49] ZHAO J. Pig cough sounds recognition based on deeplearning[D]. Wuhan： Huazhong Agricultural Universi‐ty， 2019.

[50] LI X， ZHAO J， GAO Y， et al. Pig continuous coughsound recognition based on continuous speech recogni‐tion technology[J]. Transactions of the CSAE， 2019， 35（6）：174-180.

[51] RISI N， KéSIA OLIVEIRA SILVA， PAULO R F Z， etal. Use" of artificial" intelligence" to" identify" vocaliza‐tions emitted by sick and healthy piglets[C]// LivestockEnvironment VIII. St. Joseph， Michigan， USA： Ameri‐can" Society of Agricultural and Biological Engineers，2009.

[52] CORDEIRO" A" F" DA" S，" N??S" I" DE" A，" SILVALEIT?O" FDA，" et" al. Use" of vocalisation" to" identifysex， age， and distress in pig production[J]. BiosystemsEngineering， 2018， 173：57-63.

[53] HALACHMI" I，" GUARINO" M，" BEWLEY" J，" et" al.Smart animal agriculture： Application of real-time sen‐sors to improve" animal well-being" and production[J].Annual" Review" of Animal" Biosciences， 2019， 7（1）：403-425.

[54] BENJAMIN M， YIK S. Precision livestock farming inswine welfare： A review for swine practitioners[J]. Ani‐mals， 2019， 9（4）： ID 133.

[55] BERCKMANS" D. General" introduction" to" precisionlivestock" farming[J]. Animal" Frontiers， 2017， 7（1）：6-11.

[56] POLITIS" A，" MESAROS" A，" ADAVANNE" S，" et" al.Overview" and" evaluation" of sound" event" localizationand detection in DCASE 2019[J]. IEEE/ACM Transac‐ tions" on" Audio，" Speech，" and" Language" Processing， 2021， 29：684-698.

[57] CHANDRAKALA S， JAYALAKSHMI" S L. Environ‐mental audio scene and sound event recognition for au‐ tonomous surveillance： A survey and comparative stud‐ ies[J]. ACM Computing Surveys， 2020， 52（3）：1-34.

[58] ZEBARI R， ABDULAZEEZ A， ZEEBAREE D， et al.A" comprehensive" review" of" dimensionality" reduc‐ tion" techniques" for" feature" selection" and" feature" ex‐ traction[J]. Journal of Applied Science and Technolo‐ gy Trends， 2020， 1（2）：56-70.

[59] GARRETA R， MONCECCHI G. Learning scikit-learn：machine learning in python： Experience the benefits of machine learning techniques by applying them to real- world" problems" using" python" and" the" open" source scikit-learn" library[M]. Birmingham， UK： Packt" Pub‐ lishing Ltd， 2013.

[60] SHARMA G， UMAPATHY K， KRISHNAN S. Trendsin audio signal feature extraction methods[J]. Applied Acoustics， 2020， 158： ID 107020.

[61] SILVA" CORDEIRO ADA，" DE ALENCAR" N??S" I，OLIVEIRA S， et al. Understanding vocalization might help" to" assess" stressful" conditions" in" piglets[J]. Ani‐ mals， 2013， 3（3）：923-934.

[62] DIANA A，" CARPENTIER" L，" PIETTE" D，" et" al. Anethogram of biter and bitten pigs during an ear biting event： First" step" in" the" development" ofnbsp; a" Precision Livestock Farming tool[J]. Applied Animal Behaviour Science， 2019， 215：26-36.

[63] ZHANG" S， TIAN J， BANERJEE A， et al. Automaticrecognition of porcine abnormalities based on a sound detection" and" recognition" system[J]. Transactions" of the ASABE， 2019， 62（6）：1755-1765.

[64] DAVIS S， MERMELSTEIN P. Comparison of paramet‐ric representations for monosyllabic word recognition in" continuously" spoken" sentences[J]. IEEE" Transac‐ tions" on" Acoustics，" Speech，" and" Signal" Processing， 1980， 28（4）：357-366.

[65] GIESERT A L， BALKE W T， JAHNS G. Probabilisticanalysis of coughs in pigs to diagnose respiratory infec‐ tions[J]. Landbauforschung vTI Agriculture and Forest‐ ry Research， 2011， 3（61）：237-242.

[66] GENG Y， SONG P， LIN Y， et al. Voice recognition ofabnormal" state" of pigs" based" on" improved" CNN[J]. Transactions of the CSAE， 2021， 37（20）：187-193.

[67] LEE J，" CHOI Y， PARK D，" et" al. Sound noise-robustporcine" wasting" diseases" detection" and" classification system" using" convolutional" neural" network[J]. The Journal of Korean Institute of Information Technology， 2018， 16（5）：1-13.

[68] THAKUR A， THAPAR D， RAJAN P， et al. Deep met‐ric learning for bioacoustic classification： Overcomingtraining data scarcity using dynamic triplet loss[J]. TheJournal" of the Acoustical" Society" of America， 2019，146（1）：534-547.

[69] XIE J， HU K， ZHU M， et al. Investigation of differentCNN-based models for improved bird sound classifica‐tion[J]. IEEE Access， 2019， 7：175353-175361.

[70] FUZ， LUG， TINGK M， et al. A survey of audio-basedmusic classification" and" annotation[J]. IEEE Transac‐tions on Multimedia， 2011， 13（2）：303-319.

[71] HUANG C J， CHEN Y J， CHEN H M， et al. Intelligentfeature extraction and classification of anuran vocaliza‐tions[J]. Applied Soft Computing， 2014， 19：1-7.

[72] XIE J， HU K， ZHU M， et al. Bioacoustic signal classifi‐cation in continuous recordings： Syllable-segmentationvs" sliding-window[J]. Expert" Systems" withnbsp; Applica‐tions， 2020， 152： ID 113390.

[73] NGO D， HOANG H， NGUYEN A， et al. Sound con‐text" classification basing" on join" learning" model" andmulti-Spectrogram" features[J/OL]. arXiv：2005.12779[cs， eess]， 2020.

[74] HUZAIFAH M. Comparison of time-frequency repre‐sentations for environmental sound classification usingconvolutional""" neural""" networks[J/OL]."" arXiv：1706.07156[cs]， 2017.

[75] NGUYEN" T，" NGO" D，" PHAM" L，" et" al. A re-trainedmodel based on multi-kernel convolutional neural net‐work for acoustic scene classification[C]//2020 RIVFInternational Conference on Computing and Communi‐cation" Technologies （RIVF）. Piscataway，" New" York，USA： IEEE， 2020：1-5.

[76] DOUGLAS C E， MICHAEL F A. On distribution-freemultiple comparisons in the one-way analysis of vari‐ance[J]. Communications" in" Statistics-Theory" andMethods， 1991， 20（1）：127-139.

[77] FERRARI S， SILVA M， SALA V， et al. Bioacoustics： Atool" for" diagnosis" of" respiratory" pathologies" in" pigfarms[J]. Journal of Agricultural Engineering， 2009， 40（1）： ID 7.

[78] SANCHEZ-VAZQUEZ M J， NIELEN M， EDWARDSS A， et al. Identifying associations between pig patholo‐gies using a multi-dimensional machine learning meth‐odology[J]. BMC" Veterinary" Research， 2012， 8（1）：ID 151.

[79] PAL N R， PAL K， KELLER J M， et al. A possibilisticfuzzy c-means clustering algorithm[J]. IEEE Transac‐tions on Fuzzy Systems， 2005， 13（4）：517-530.

[80] RTAYLI N， ENNEYA N. Enhanced" credit card" frauddetection based on SVM-recursive feature eliminationand hyper-parameters optimization[J]. Journal of Infor‐mation Security andApplications， 2020， 55： ID 102596.

[81] LEE J， JIN L， PARK D， et al. Acoustic features for pigwasting" disease" detection[J]. International" Journal" ofIntellectual Property Management， 2015， 6（1）：37-46.

[82] HE K， ZHANG X， REN S， et al. Deep residual learn‐ing for image recognition[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition （CVPR）. Piscataway， New York， USA： IEEE， 2016：770-778.

[83] LECUN Y， BOTTOU L， BENGIO Y， et al. Gradient-based learning applied to document recognition[J]. Pro‐ ceedings of the IEEE， 1998， 86（11）：2278-2324.

[84] SIMONYAN K， ZISSERMAN A. Very deep convolu‐tional networks for large-scale image recognition[J/OL]. arXiv：1409.1556[cs]， 2015.

[85] XIONG Z. Design of pig cough monitoring system infattening" pig" houses[D]. Harbin： Harbin" Engineering University， 2021.

[86] H?RDLE W K， SIMAR L. Applied multivariate statis‐tical" analysis[M]. Berlin， Heidelberg： Springer Berlin Heidelberg， 2015.

[87] VALLETTA J J， TORNEY C， KINGS M， et al. Appli‐cations of machine learning in animal behaviour stud‐ ies[J]. Animal Behaviour， 2017， 124：203-220.

[88] EXADAKTYLOS" V，" SILVA" M，" FERRARI" S，" et" al.Sound localisation in practice： An application in locali‐ sation" of" sick" animals" in" commercial" piggeries[M]//STRUMILLO P. Advances in sound localization. Shen‐zhen： InTech， 2011.

[89] WANG R. How do pig practitioners consider artificialintelligence" in" pig" farming？[J]. Swine" Industry" Sci‐ence， 2019， 36（4）：46-48.

[90] SILVAA P S P， STORINOG Y， FERREYRAF S M， etal. Cough associated with the detection of Mycoplasmahyopneumoniae" DNA" in" clinical" and" environmentalspecimens" under" controlled" conditions[J]. PorcineHealth Management， 2022， 8（1）： ID 6.

[91] PESSOA J，" RODRIGUES" DA" COSTA M，" GARCíAMANZANILLA E， et al. Managing respiratory diseasein finisher pigs： Combining quantitative assessments ofclinical" signs" and" the" prevalence" of lung" lesions" atslaughter[J]. Preventive" Veterinary" Medicine， 2021，186： ID 105208.

[92] SCH?N P， PUPPE B， MANTEUFFEL G. Automatedrecording of stress vocalisations as a tool to documentimpaired welfare in pigs[J]. Animal Welfare， 2004， 13：105-110.

[93] SUN J. Technology changes pig farming[J]. China Ru‐ral Science amp; Technology， 2020（1）：36-39.

摘要：叫聲是評估生豬福利水平的重要方式之一。本文首先分析了生豬叫聲與福利之間的相互關系。其中，與生豬福利密切相關的三種生豬叫聲包括咳嗽聲、尖叫聲和呼嚕聲?；谶@三種聲音進一步分析聲音與環境，聲音與身體狀況，以及聲音與健康之間的關系。隨后，對當下的生豬福利監測所采用的傳感器，包括穿戴式與非接觸式兩大類進行分析，并簡述不同方式的優劣勢。基于非接觸式的優勢及麥克風傳感器技術的可行性，從聲音的獲取和標記、特征提取以及聲音分類三個方面對現有的生豬聲音處理技術進行了闡述和評估。最后，從聲音監測技術、生豬個體福利監測、商業應用以及養豬從業者四個角度討論了叫聲在生豬福利監測中面臨的研究困境以及發展趨勢。研究發現，目前關于生豬聲音分析的研究大多集中在分類器的選擇和識別算法的改進上，而對端點檢測和特征選擇的研究較少。同時，當下面臨的主要挑戰還包括不同生長階段的音頻數據獲取難度較高，缺乏公共的豬舍內音頻數據庫以及缺少完善的聲音指標與動物福利監測評價體系?？傮w來說，建議進一步對聲音識別過程中涉及的各部分技術進行深入探索，同時加強跨學科專家之間的合作，共同推動聲音監測在生豬實際生產中的應用，從而加快精準畜牧業的實現。

關鍵詞：生豬聲音識別;動物福利;聲音分析;特征提取;精準畜牧業;聲音監測