Introduction
The global population is expected to have increased to 9.7 billion people by 2050, representing more than 34% of the current global population (Michalk, et al. 2018; FAO, 2018). The demand for consumer products, as a consequence, has continued to rise in tandem with the increase in the population (Vranken and Berckmans, 2017). Aside from that, the problem is not just offering food, but also taking into consideration health circumstances such as malnutrition, disease propagation, and so on (Henchion et al., 2021). Livestock production is a major source of food supply for a large portion of the world's population. Around 33% of protein consumed by humans come from animal products such as meat and milk (Smith et al., 2013). In livestock products, pork is the second most consumed meat because due to its affordability and lower price (OECD-FAO, 2019). Pork meat accounted for approximately 22.6% of total meat output worldwide, and the volume of pig (Sus domesticus) production varies by location (Popescu, 2016). Pig farming is affected by several variables, including the number of pigs raised, the growth technique, the quality of the bio-environment, diseases control, and food consumption rate (Woonwong, et al., 2020; Popescu, 2016).
Traditional pig farming in the 1960s was characterized by the production of a small number of animals and the need for human labor to maintain the farm conditions (Thornton, 2010). In addition, efforts were made to check their health conditions. In the late 20th century, housing practices and management were changed, and the growth of major companies resulted in the commercialization of farm size and output (Guyomard et al., 2013). As the scale of large enterprises continues to expand, a system called "automation" has been established in terms of the production management process and welfare of the pig farms currently in operation (Wang et al., 2021). Accurate and continuous monitoring of the individuality of livestock requires highly reliable technology. This is called Precision Livestock Farming (PLF). PLF offers farmers with the option to create better and more precise quality products while also filling the need for animal products for human consumption.
In order to accomplish the goals for pig farming, professionals must be on-site to assess the condition of the animals and take necessary action continuously. It is also vital to monitor all separate entities fairly and take into consideration their unique qualities. However, these conditions will not only impair efficiency, but they will also cause a significant amount of financial and physical damage to the farmer's management.
With the increase in agricultural density, the effect of the breeding habitat of animals on production is notable. According to OCED-FAO (2019), pig raised over a wide range of conditions are capable of active self-expression, whereas pigs raised intensively develop in size and reach maturity more rapidly than the pigs raised over a wide range. Compared to pigs raised in a wide range of environments, they showed increased aggression, abnormal behavior and asocial characteristics such as navel stabbing, tail and ear biting behavior (Peden et al., 2018). Additionally, diseases and disease convictions account for 25% of pig deaths in the world (Kurian et al., 2021). Keeping pigs healthy is critical to being successful in production. It is important to be aware of any infections that may arise on the farm. Early detection of health abnormalities in pigs is of critical concern, and diseased pigs should be recognized with the symptoms of disease and separated from others.
Pigs often communicate their present state of health through vocalizations and yells. The sound of sick pigs may suggest respiratory infections, which are a leading cause of death and productivity loss in intensive pig farming (Chung et al., 2013). To mitigate the harm caused by numerous respiratory illnesses, equipment for collecting and evaluating livestock data must be developed. This information is quite valuable to farmers since it allows them to detect illnesses early and identify the severity of the diseases on their farms, which is extremely beneficial.
Sound analysis is critical since they aid in the classification and quantification of different respiratory diseases and crushing symptoms. Moreover, sound can be detected readily at a distance without disturbing the usual pig living conditions. Sound sensors provide considerable benefits over other types of sensors such as imaging sensors, thermal sensors, laser sensors and motion sensors. Sound sensors are inexpensive, require no direct contact with the object, and allow for simultaneous observation of a large number of animals on the farm (Neethirajan et al., 2021). In the last few years, sound analysis has become more important as a way to understand animal behavior, health, and well-being. Therefore, the purpose of this review article was to summarize current research on sound based techniques to monitor pig diseases and crushing symptoms in pig farms, as well as features of detection methods for improving detecting technical parameters.
Signs of sound-related pig disease and crushing symptoms
Sound is an effective tool to monitor health condition and cough sound is a major key factor in detecting common respiratory diseases. It is one of the body's defense mechanisms against respiratory infections, and it can be a sign of an infection or disorder of the respiratory system. Coughing is the most common symptom of a variety of disorders that damage the airways and lungs. Pig noises are often non-infectious include coughing, screaming, footsteps, and grunting. Among these, the regular coughing sound was mostly generated by several environmental irritants present on closed farms, such as dust, and ammonia. The most common noises associated with infections include sneezing, snuffling, and coughing.
Daily observation of pigs will help early identification of diseased or injured pigs. In order to classify sick pigs, it is important to know the characteristics of normal pigs. The presence of a sick pig becomes clearly vulnerable and may even be attacked by other pigs. The most prominent instance of visual change occurs when pigs have abnormality in the appearance. From mild skin diseases to acute bacterial sepsis, the respiratory rate may increase and the depth of breathing may vary. Because not all causes of cough are the same, the explanation may be different. The respiratory disease depend on the pigs characteristics and condition such as age, size, and type of infection. The speed of the coughing sound changes with time, and these factors are of great help in recognizing the health status of pigs. The sound of swine influenza A virus (IAVS), the most important cause of contagiousness in pigs, is almost similar to that of a goose (Vincent et al., 2014). This is because the virus that touches the nerve endings induces an explosive cough with painful itchiness. In contrast, the porcine reproductive and respiratory syndrome (PRRS) virus has a moist, high-toned cough sound (Sinn et al., 2016). The cough of pigs differs according to the type of virus, and if the sound can be recognized separately, it can reduce 30% of pigs or cattle dying from respiratory diseases (Oba et al., 2020).
Digestive problem due to lack of water supply causes constipation in young pigs. Digestive disease is the most important cause of sudden infertility in pigs, causing economic loss as well as fatal damage to pigs’ health. When one pig suffers from a contagious digestive disease, symptoms quickly appear in other pigs as well.
Piglet crushing is a regular occurrence in pig farms and often happens during the first few days after birth (Muns et al., 2016). Mother pigs crush piglets for two primary reasons. First, if the mother pigs’ health deteriorates, it becomes hard to care for the piglets. Piglets are born weighing 1.2 kg, which is about 1/208 the weight of a 250 kg mother pig (Jankowiak et al., 2020). Therefore, the ignorance of mother pigs towards the piglets due to sickness or other reason, the piglets are squeezed and overturned, and eventually die. When the mother pig lays on the piglets, the piglets creates a high-pitched tearing sound, which is distinct from the regular piglet sound, allowing us to readily identify the crushing from the sound (Chapel et al., 2018). Second, inadequate separation of mother and piglets, lengthy straw bedding to prevent piglets from escaping, insufficient rail or cage design, and low temperature are all factors in crushing in the pig farm (Nicolaisen et al., 2019) . Moreover, pigs with skin diseases such as cracks or open wounds or skin-related diseases cause pigs to bite their tails or sniff their noses (Statham et al., 2009). The skin-related diseases in pigs appear from the lack of air quality, lower temperature, and wind blows in which they live.
Cough sound analysis has the potential to not only monitor pig health, but also prevent widespread damage by early detection of respiratory disorders (Neethirajan et al., 2017). Respiratory and digestive diseases in pig reduce growth rates, weight losses and in severe cases even lead to death (Cohen et al., 2020). Pigs show strong synchronization, when one pig screams, the surrounding pigs remain quiet, allowing more precise sound data to be acquired. Interference from other noises can interfere with the detection of pig sounds. Automatic sound processing methods that uses spectrum to classify based on the structure, force, and frequency have been used. However, sound labeling helps to overcome the errors such as coughing to hear the sound. Table 1 shows different disease conditions detected by the sound in pig farms.
Sound sensors and systems for pig health monitoring
Automation is a tool that can detect the changes in health condition and behavior in pigs in order to improve their wellbeing (Matthews et al., 2016). The identification of the sound of a pig in the pig farm need to be detected accurately and automatically. This is accomplished via the use of technology for the collection and analysis of livestock data. When collecting data on pigs, it is essential to choose an approach that minimizes their stress. Usual sensor setting in the pig farm are shown in Fig. 1.
Sound-based monitoring shows a potential applications in PLF. PLF may be described as the management of individual animals by continuous real-time monitoring of their health, welfare, productivity and environmental effects (Berckmans, 2017). The sound analysis sensing platform is far easier to set up than conventional sensors, since the sensor is just an audio recorder (Neethirajan et al., 2021). Any sound based analysis begins with sound recording, which may be described as the process of obtaining audio signal information using microphones. The sensor is permanently mounted in one spot and captures ambient noise. As a result, this technology enables the use of a single sensor to monitor a large number of animals (Bishop et al., 2019). However, audio signal often suffers from significant hindrance in particular frequency range and the audio signal spectral properties (Cho et al., 2015). This issue renders the sound system susceptible to background noise, impairing transmission performance. A wide range of sensors can be used to gather information about the livestock animals using sound sensors. Therefore, research on wireless sensor network technologies has increased significantly in recent years. Zeng et al. (2021) introduces a three layer wireless sensor network using Zig-bee module to monitor pig house environment in real time. Chen and Liu (2019) introduced a fuzzy control based wireless sensor network to monitor pig breeding environment in the farm. To identify and track pigs in order to monitor their health environment condition, Ma et al. (2011) proposed the use of radio frequency identification (RFID) and wireless sensor networks. Table 2 shows different types of sound sensors usually used in pig farms.
Sound analysis methods for pig disease and crush monitoring
Pigs used sound to express their health and well-being, particularly when anything was happening with their bodies, such as coughing for respiratory symptoms or crushing of newborn piglets. The sensors are capable of detecting their sound or motion, but there are inaccuracies as well. Pigs, for example, will scream when they are in a stressful environment. At this point, the impression of the sound may be distorted as a result of the ambient background noise. Therefore, studies are continuing to collect data from the noise in a more accurate and timely manner.
Audio-based sensor are more effective in detecting clinical signs of respiratory disorders. The microphones utilized in such technologies, on the other hand, have spatial restrictions once it comes to detecting sound. These diseases are not unique to humans, but also to animals. The descriptive features of current research focusing on sound-based pig abnormality detection are represented in Table 3. Pig vocalizations are directly related to their pain and physical conditions. According to the researchers (Marx et al., 2003; Diana et al., 2019), the study of the sound components of particular vocal responses indicate that these are the synopsis to abnormalities and are closely connected to wounded or sick pigs. A sensor system may be classified according on the kind of sound it recognizes. Gutierrez et al. (2008) aimed to classify pig diseases through an acoustic analysis highlighting the difference in cough acoustic interval. The sensor was useful in supporting early detection methods based on online cough counter algorithms for early diagnosis of pig respiratory diseases. In fact, some experiments were performed in the laboratory and an algorithm was introduced to detect cough sounds and classify abnormal conditions in pigs (Van Hirtum el al., 2002).
Guarino et al., (2008) used an easy-to-use and simple algorithm to continuously record sounds and monitor them online using cheap microphones. The algorithm demonstrated the capability of recognizing several coughing instances, with an accuracy of 86.2% under field circumstances. Exadaktylos et al. (2008) suggested a real-time cough detection method for detecting sick pigs. The evaluation and identification were performed using a frequency domain properties of the sound signal, and an enhanced approach for extracting it. This approach employed a fuzzy c-means clustering on selected segments of the training signals and generates a frequency information reference that closely matches to the sick pig cough features. The approach was 85% successful in classifying the sound and 82% effective in detecting sick coughs. Silve et al. (2009) investigated the dynamics of pig cough noises in order to correlate them with diseased states of the pig respiratory system. Two categories of pigs were evaluated using the approach, and the findings reflect prior research demonstrating that diseased circumstances have an effect on cough sound length. This approach shed more lights on the influence of pig health condition changes based on the cough sound length.
Chung et al. (2013) developed a data mining approach for identifying abnormalities (respiratory illnesses) in pigs by analyzing acoustic data from an audio monitoring system. A total of 36 sick pigs were evaluated and cough sounds from infected pigs were captured for 30 minutes using a digital camcorder and sound analysis software utilizing a standard sound card. The sound was digitized on a computer and labeled using the sound collection procedure. Each cough-specific illness was categorized as post-weaning multi-systemic wasting syndrome (PMWS), PRRS, or Malignant Hyperthermia (MH), and utilized as reference data. Ferrari et al. (2007) emphasized the importance of the cough sound in identifying respiratory illnesses. They demonstrated that the cough sounds of sick pigs had a lower peak frequency (200 to 1100 Hz) than the cough sounds of healthy pigs (750 to 1800 Hz).
Polson et al. (2019) investigated the ideal location and design of a continuous audio monitoring system over a vast region in order to achieve high sensitivity for identification of respiratory diseases in pig. Sound sensors purchased from a commercial source were utilized to detect the pig sound, and each sensor represented an 18-20 m audio detecting zone. Von Borell et al. (2009) developed a tool based on the classification of three different classes of piglet vocalizations (i.e., grunting, squealing and screaming). They found that vocalization analysis in pigs can help identifying both pain and behavioral changes.
In recent years, several investigations in monitoring, data analysis, and transfer have intended to explain challenges associated to animal farms using artificial intelligent (AI) models including machine learning (ML), deep learning (DL), and artificial neural networks (ANN) (Bao and Xie, 2022). Yin et al. (2021) performed research to develop a highly accurate technique for detecting pig coughs in order to improve the alert system. They developed a classification system based on the fine-tuned AlexNet model and the spectrogram characteristics. The approach was shown to be more than 95% accurate. The method performance was affected by the background disturbances such as metallic sounds in the farm and a poor signal-to-noise ratio, which drowned out the spectrograms. Zhao et al. (2020) introduced a new approach for continuous pig cough sound detection based on Deep Neural Networks-Hidden Markov Models (DNN-HMM). The Wiener technique was used to remove noise from the continuous pig sounds and mel-frequency cepstral coefficients (MFCCs) were derived as feature vectors. The word error rate (WER) of each group was found to be 3.45% lower on average in DNN-HMM than in the other approaches.
A deep learning-based system for the early diagnosis of respiratory illness in developing pigs, was presented by Cowton et al. (2018). They used environmental sensors data to build and test their methodology, which consisted of two recurrent neural networks (RNNs), each of which included gated recurrent units (GRUs), in order to produce an auto-encoder (GRU-AE). The environmental data were gathered in the auto-encoder and processed in order to identify the abnormalities. If a sickness occurs, the system was capable of detecting and alerting within 1-7 days. According to Lee et al. (2015), an approach for assessing the sound features of pig vocalizations in order to diagnose pig wasting diseases was addressed. They employed pattern recognition to examine the differences between the sounds made by sick pigs and normal pigs. When support vector machine (SVM) was utilized as a detector, the results of the studies revealed the average detection accuracy of 98.4%.
Cordeiro et al. (2018) attempted to study the variations in pig vocalization across sex, age, and stress conditions. They were also suggested a method for recognizing pig stress situations. 40 pigs vocal sound were captured and the vocal pitch were found different for males (194.5 Hz) and females (218.2 Hz) as well as varied based on the age. Using the machine-learning approach, a decision-tree was constructed for identifying the stressed status of pigs with an efficiency of 81.92%. Wang et al. (2019) demonstrated a substantial variation in cough noises produced by animals under various air quality situations using power spectral density (PSD). Additionally, they created recognition system, which included MFCCs, principal component analysis (PCA), and support vector machine (SVM), which performed an average identification rate of 95% for cough sound data gathered from several pig farms.
Manteuffel et al. (2017) studied piglet crushing events by a mother pig. This technique attempts to identify crushing occurrences in real time using distress-specific vocalization features of piglets. In this case, by adding context-based event filters, the specificity was raised to 95%. The precision was about 30%, and the sensitivity was approximately 70%. Chen et al. (2021) presented "PigTalk," a piglet crushing monitoring system based on AI and the Internet of Things (IoT). PigTalk recognized piglet screams and immediately engaged the alert system to ensure the quick handling of the crushing incidents. The identification of piglet screams using machine learning (ML) revealed detection accuracy of up to 99.4%. Vandermeulen et al. (2015) created an automated classifier for defining and subtracting pig screams from all the farm noises. The sound that occurred and lasted longer than 0.4 seconds was classified as a scream. The classifier was adaptable and demonstrated a sensitivity of 71.83%, a specificity of 91.43%, and an accuracy of 83.61%.
A software based pig scream detection system in stress conditions was investigated by da Silva et al. (2019). The intensity of the vocal sounds of 40 pigs was considered in this study with various stress conditions such as temperature, and pain. According to the findings, the forecast for the pain was shown to be the most accurate (93.0%). The intensity and length of the vocalizations caused by pain were both quite high. Automatic scream classifier (Hemeryck and Berckmans, 2015) to monitor pigs in the big pig farm with the reference data is necessary. Under a variety of environmental situations, Domun et al. (2019) created three dynamic models and a neural network to identify change in behavior such as tail-biting, diarrhea, and littering. Li et al. (2019) studied the conditions of piglets under different musical environments and significantly affected different behavior patterns such as walking, lying, standing, and exploring.
A variety of image and video-based pig monitoring tools were utilized to provide real-time monitoring. The practical obstacles such as lighting, shadows, and a massy background, as well as a damp floor condition produced by urine or manure in the pig farm, resulted in erroneous image or video capture in the pig farming environment (Kim et al., 2017). Furthermore, insects and flies caused black spots to appear on the images or videos, resulting in inadequate information about pigs. Even after all this time, these challenges in farmers' contexts have remained a difficult problem to solve. At night, the majority of the pig farms were dark, which presents an additional obstacle to the imaging sensors. Several infrared cameras (Costa et al., 2013; Khoramshahi et al., 2014) have been claimed to provide positive results at night, but the cluttered background continues to present complications. Thermal imaging sensors (Cook et al, 2018; Caldara et al., 2014) were also employed to overcome the issue of a cluttered background, which was previously encountered, but these types of high-end sensors are costly.
To tackle the difficulty of a cluttered background, stereo sensors and depth sensors (Shi et al., 2016; Kim et al., 2017) were also utilized. Researchers are investigating the accuracy of these various sorts of sensors. A sound-based monitoring device, particularly in low-light settings, might help to alleviate this predicament. Furthermore, sound analysis is the sole way to detect the symptoms of various diseases in pigs that do not show any visible signs of illness. Real-time monitoring is also affected by processing speed, which has a substantial impact on image-based real-time monitoring, which is still another major issue. Sound-based monitoring can also avoid these issues as the data size is smaller compared to an image-based system.
Other agricultural animals, such as broilers (Sadeghi et al., 2015; Banakar et al, 2015), cows (Chung et al., 2013; Volkmann et al., 2019), and sheep (Bishop et al., 2017; Xuan et al., 2016) were also subjected to sound-based disease detection techniques. Sneezing noise from broilers aged 15-45days were detected using an automated monitoring system developed by Carpentier at al. (2019). The algorithm was developed for the laboratory environment while data were collected using SOMO+ device attached with a plug-and-play microphone to record the audio. The suggested method achieved an accuracy of 88.4% and a sensitivity of 66.7%. Using a combination of time domain and frequency domain feature extraction, Huang et al. (2019) studied sound analysis of healthy and avian influenza infected chickens. MFCCs were calculated as an input of SVM. The method was evaluated under laboratory condition and infected chicken detection accuracy rate was 84%-90%. More studies were observed (Lee et al., 2020) for the broiler farm to estimate diseased condition using sound devices. Carpentier et al. (2018) developed an algorithm to separate cough noises from other sounds in commercial calf raising facilities. A novel method for monitoring that eliminates the necessity for calibrated reference labels were used. The algorithm achieved a sensitivity of 41.4% and a precision of 94.2%. When the number of animals increases, the algorithm performs poorly. Weaned calves with respiratory problems were studied by Ferrari et al. (2010) to see if cough sounds could be differentiated from other noises. Developing an automated on-line monitoring technique for indicators to indicate respiratory illness activity, such as cough identification and early disease detection could be beneficial for cow farms.
The use of sound analysis in pig farms may aid in the detection of both pain and behavioral changes, suggesting the possibility of a better PLF system in the present day. Consequently, a PLF instrument that analyzes the vocal patterns of pigs and automatically identifies abnormalities might possibly be developed, giving the farmer with an early warning signal and/or assisting in the development of a timely remedy prior to the dispute surge.
Challenges and perspectives of the sound signal processing
In pig farms, if more attention is paid to each pig, even a small abnormality in the pig can be detected immediately, and disease can be prevented at an early stages. This leads to the investigation of how this approach may be more accurate than general eye inspection in studies of sound detection analysis and how to eliminate observational subjectivity (Maselyne, 2021). Hong et al. (2020) proposed several limitations with respect to sound detection in pig farms to detect the abnormalities in pigs.
Most of the research has a limited experimental environment, and the experiment should be conducted in the environment most similar to that of a pig farm.
It is very rare that the sound wave applied to the automatic detection and localization of event noise in pigs is not manually edited or shaded. Many studies are manually classifying data by using a microphone or a sensor to detect sound and then visually converting it into data.
Although there have been cases that have explained the effect of noise on the cough sound detection performance, there are few studies that have detected and accurately suggested an abnormal situation for noise.
Cost effectiveness of the methods and systems that used in farms for real time implementation and low cost analysis.
Besides, in the pig farm, several piglets often emit simultaneous noises, and the sound signals produced by the pigs are susceptible to distortions. The handling of the distorted pig sound signals is challenging to deal with. It is a problem to conduct signal separation, feature parameter extraction, and signal separation after distortions on live pig sound data, all of which need complex algorithms. With time, as the pig's maturity develops, the sound signal will also vary in proportion, which will undoubtedly have an impact on the precision with which the sound signal can be recognized. Over the course of the pig's entire life, the features of pig sound signals at different stages are investigated and evaluated to figure out how pig sound signals change with age. Individual pig farm cough sound data for abnormal pig sound detection and classification may prove challenging to obtain an effective results due to the expansion of the monitoring range and the improvement of the application effectiveness of monitoring equipment.
The predicted demand for agricultural production will increase by as much as 100% by 2050. This leads to the need for sustainable development of agriculture in the future. For a better environment, the conservation and resource management should be optimized and compatible. However, in order to maintain continuous development and optimization, it is limited by technology and techniques (Gebbers and Adamchuk, 2010). Small and medium-scale pig farms have specific requirements for pig abnormality detection systems, which include lower cost for the system, 24-hour monitoring, good precision of abnormality detection with a low number of false detection, a noise robust system as there are various sounds in pig farms, and the ability to easily and cost-effectively replace the sensors. Usual data collection process can be hampered by various environmental conditions such as dust, ammonia and methane, as well as poor air quality.
Although many studies are being conducted and efforts are being made to optimize the resources, further progress is being made in that it cannot operate with the low capacity and high efficiency level of the sensors and there is a limit to the performance that can be obtained at a low price, which could be a key factor in the agricultural sector.
Conclusion
Recent technology advancements enable the PLF industry to build more efficient, cost-effective, and environmentally friendly systems while simultaneously improving animal care and the quantity and quality of the ultimate product. This review study summarized current research on sound based approaches for monitoring pig diseases and crush symptoms in pig farms, as well as the features of detection methods that may be utilized to optimize detection parameters. A number of sensors may be used to gather data on farm animals, including sound sensors. Detection of acute respiratory illness symptoms may be improved by the use of sensors with an auditory component, and research into wireless sensor network technology has seen a significant increase in the usage of diverse sensors. Signal processing and artificial intelligence (AI) models have been used extensively in pig farm monitoring and data analysis in order to better grasp the underlying difficulties. Additionally, certain problems regarding sound detection in pig farms were addressed. The implementation of sound sensing technologies based on sound processing in the pig industry will be a possible medium for the development of new pig farm technologies.
Acknowledgement
This work was supported by the Korea Institute of Planning and Evaluation for Technology in Food, Agriculture and Forestry (IPET) through Smart Farm Innovation Technology Development Program, funded by Ministry of Agriculture, Food and Rural Affairs (MAFRA) (Project No. 421044-04), Republic of Korea.
