Home

The most widely used calving prediction method is a calculation method that subtracts 3 months from the final insemination month and adds 10 day to the final insemination day[1]. However, this calculation method shows a deviation of 15 days maximum according to cow. Also the actual calving date can differ according to the number and sex of fetuses, and calving season.

The calving period can also be predicted by observing the physical and behavioral changes in cows. Impending cows show swelling of the vulva and udder development[8]. About 6 hours prior to calving, behavioral changes such as elevation of the tail or frequent lowering of the head to see the stomach can be observed[9]. However, it is difficult to observe every cow in a large-scale farm with vast numbers of cattle. Also some calving occurs at night or dawn, making this prediction method ineffective. Another method that analyzes vaginal temperature, rumination time, lying time, and lying bouts to predict calving times[10] has complex measurement processes and a low accuracy rate of 74%.

There have been active researches on predicting calving times by analyzing body temperature changes before calving. Study[11] measured the rectal and vaginal temperatures of impending cows in regular intervals 48 hours prior to calving, however, the measurement process was difficult and the small drop size of body temperature made accurate prediction impossible. Research[12] concluded that if the rumen temperature drop was greater than 0.2℃ there was a high probability that calving would occur within 24 hours. Vaginal sensor is inserted to measure and analyze temperature to predict calving time[13]. The vaginal sensor transmits an alarm message to the farm manager when it is discharged with the ruptured ventral sac and detects a change in temperature. This method transmits an alarm at the actual time calving starts and thus makes it possible for an accurate prediction of parturition. However, it accompanies the danger of bacterial infection during the process of inserting the sensor in the vaginal area.

Existing calving prediction methods require regular data analysis by the user and so there have been continuous attempts to automate this process using AI. However, such studies have focused on externally installed acceleration sensors and cameras, and due to inaccurate data and sensor durability issues, practical use and case studies have started to be presented after 2016.

Study[14] proposed an algorithm predicting insemination outcomes of Irish dairy cows using 5 variables including milk genetic trait. The logistic regression algorithm was applied but the method showed low results with an F-Score of 52.03%. Study[15] monitored activity, condition, etc. of dairy cattle by using random forest algorithm with data collected from acceleration sensors installed on the necks of cattle. Also, Heat detectors was installed near the vulvar area and predicted cattle diseases using deep learning from color changes caused by temperature change[16]. This method has low accuracy as it indirectly measures temperature and has issues such as difficulties in measuring at nighttime. Study[17] proposed a disease prediction method by applying data such as deep body temperature, amount of activity, feed ration to deep learning(4-layer neural network). This method showed a performance rate of more than 80% and giving rise to the potential of deep learning application in the area of cattle condition detection.

In this paper, we propose a calving prediction method using ingestible sensors to measure deep body temperature and applying such collected data to deep learning to provide an accurate calving time for impending cows.

The proposed calving prediction process is outlined in Fig. 1. The ingestible sensor measures the rumen temperature and the measured data is transmitted to the remote server through the LoRa Gateway. The data transmitted by the sensor through the LoRa Gateway guarantees safe transmission with no data loss to the server. The data transmitted to the server is saved to each database according to entity, location, and farm. Such categorized data is transmitted to the analysis server for real-time analysis. This data is analyzed with deep learning and the final calving prediction date is calculated. Cows with impeding calving are notified to farm managers so proper assistance can be provided.

Fig. 1. 
Calving prediction process

Ingestible bio-sensors(LiveCare, uLikeKorea Co., South Korea) were used to measure the deep body temperature of cattle. Bolus guns were used to orally administer the sensors while preventing any saliva contamination. The orally administered sensors settle in the rumen as in Fig. 2. Ingestible sensors once settled in the rumen are free from any external influences and measure deep body temperature in 10-minute intervals. The bio-sensors can detect minute changes in core body temperature up to 0.1℃. The bio-sensor is installed in a non-toxic plastic capsule that protects it from gastric fluids and pressure. The capsule does not have any rough surfaces or sharp corners making it safe to install in cows[18].

Fig. 2. 
Ingestible sensor settled in rumen

The data measured and collected in this proposed method is time-series data. The purpose of this method is to analyze the collected data and provide accurate predictions of calving time(determining whether calving is imminent or not). This method can be defined with single input and binary classification problems.

LSTM(Long short-term memory) is frequently used in natural language processing, and it is especially useful in processing time-series data where data correlation is important.

This paper promoted higher performance by augmenting FCN(Fully convolutional networks) to LSTM as in Fig. 3[19]. The LSTM-FCN model has proved to offer outstanding performance for open data ECG200.

Fig. 3. 
LSTM-FCN model

The input time-series data is normalized by expressing the difference from the average value using the average and standard deviation. The normalized value is entered into LSTM and the one-dimension convolution layer according to each time frame. Both outputs are concatenated and processed through the softmax function for a probability value [probability of calving condition, probability of normal condition]. If the cow is definitely in calving condition the output value will be [1, 0], and for normal condition the output will be [0, 1]. The probability of calving condition = 1 – probability of normal condition.

The experiment was conducted from August 1, 2018 to October 31, 2018 on 4 pregnant cows over 24-months-old bred in a domestic cattle farm in Chungnam province, Korea. Information of the entities including breed, sex, age, and expected calving date are illustrated in Table 1. Entities A&D and B&C have similar ages and calving dates and thus will be compared and analyzed for body temperature patterns.

Ingestible sensors were orally administered to all 4 entities on July 31, 2018. The bio-sensors measured deep body temperatures of each cow in ten minute intervals and transmitted the data to the server. The collected data was analyzed for deep-learning based calving predictions.

AWS EC2(Amazon Elastic Compute Cloud) was used for deep learning training and testing. Specifications are as follows: Intel Xeon E5-2686 @ 2.3GHz CPU, 61GB RAM, NVIDIA Tesla K80 GPU, Windows Server 2016 OS. AWS EC2 is a web service that provides stable and scalable computing resources. Recently, the system has provided GPU instances which is optimal for deep learning computing, making it the preferable system for many researchers and developers.

The temperature data collected every 10 minutes are converted into one-hour unit data by calculating the average value of each hour. Each hour unit that accompanies calving symptoms is checked. If the number of temperature data with calving symptoms exceeds a certain number it is labeled as calving condition 1, and if not normal condition -1.

The deep learning structure was realized using Keras with Tensorlflow backend. A total of 369,154 parameters were used with 3 one-dimension convolution layers and 1 LSTM layer. In addition the early learning rate was 0.001, with a batch size of 128 and the epoch number for repeated learning of the total data set was 3000.

He initialization was used for each parameter initialization. Fig. 4 displays the training examples and history of each unit hour and temperature data with calving symptoms. Training wasconducted for a model that detects calving when temperature data with calving symptoms exceed 15 within a 72-hour unit. Also, we used 3 methods ike below to avoid the overfitting  problem.

1)　Splitted the data and performed cross validation
2)　Added a drop-out in each layer
3)　Applied L2 normalization

Fig. 4. 
Training examples and history

The early learning rate of 0.001 rose to 0.000125 in the final learning session. The accuracy rate for the training data set was 100%, and 97% for the data set. In addition, this model showed over 97% accuracy when applied to a separate experiment data set(Table 2).

During the experiment period, the deep body temperature of all 4 entities were measured in a stable manner without the bio-sensors being lost or entities showing any abnormal behavior. A total of 17,280 deep body temperature data was collected by the data collection device, and an average of 4,320 data was analyzed per entity. Table 3 is a summary of the experiment results and all 4 cows completed parturition during the experiment period.

The average body temperature of the entities during the experiment period was 39.25℃ and showed temperature changes within the normal body temperature range. There were some differences among entities, however most entities recorded an average body temperature of 39.42℃ from 10 days prior to 1 day prior to calving, which was about 0.2℃ higher than the overall average body temperature during the experiment period. In addition, most entities experienced fall in body temperature 24 hours prior to calving, recording an average of 38.79℃ which was about 0.5℃ lower than the overall average.

Fig. 5 shows the results of the calving probability rates of each entity that was predicted through the deep learning model based upon the hour-unit body temperature data.

Fig. 5. 
Calving probability graph by entity

The probability value is expressed from 0~1 and the higher the calving probability, the closer the value is to 1. The red arrow indicates the actual calving time. The unit hour is 72 hours and the entity was decided to be in the calving process when the number of calving condition body temperature exceeded 15. Experiment results show that with the exception of Entity C, the model accurately predicted the calving date.