AI Based Digital Twin Model for Cattle Caring

The sensor’s raw data set consists of 11 columns with over 50 million rows; a part of the original data set is displayed in Fig. 2. It entails the treatment of 759 cattle of various breeds and genders in various methods. There are eight categories used to classify cattle’s status: Resting, Rumination, High Activity, Medium Activity, Panting (Heavy Breathing), Grazing and Walking. Each sensor takes a minute-by-minute reading of the cows’ real-time status, with each cow having 74,455 data points collected between AU_time 8:06 am on August 10 and 1:01 am on October 1, 2019. Five cattle breeds are represented in the data sets: Angus, Brahman, Brangus, Charolais and Crossbred. There are also 13 potential therapy combinations, which are depicted in Table I. This section focuses on systematically processing of these data, including data segmentation, data cleaning, and data calculating.

Acquiring the cattle’s state changes across the sample period needs to average one group’s data of cows due to large amount of discrete and lost data from a single cow. For example, averaging the resting time per hour of 14 Brahman females under Positive control can determine variations in the resting state of Brahman treated with Positive control during the sample period. The time series after data cleaning are different between each cattle’s data set, since invalid data collected by sensors in the farm’s IoT system is usually a random process. Therefore, the data processing in this step is to average the state data of the cattle with the same time serial number and obtain the state curve of the cattle in the whole cycle. The process flow chart of an average state time for several cattle can be found in Fig. 4.

The state diagram of cattle in the entire cycle can be obtained after the program has been executed. Fig. 5 shows the calculated hourly rest time of the cattle in the whole cycle (Brahman Female with Positive Control). The number on the abscissa corresponds to the corresponding day, which includes all 24 hours. The ordinate represents the rest time corresponding to this hour in minutes.

Fig. 6 is a detailed zoomed-in part of Fig. 5 and located between days 16 and 20. It is obvious that the rest time of cattle varies periodically with a cycle of one day. The peaks of the daily rest time can be found in both early morning and late-night while the valleys can usually be identified at forenoon and afternoon hours.

The averaged single rest cycle data result (which is 24 hours) of a single cattle is plotted in Fig. 7. The entire sampling cycle is approximately 52 days as shown in Fig. 5. The abscissa refers to the o’clock, i.e., from 0:00 to 23:59, and the ordinate relates to the rest period in minutes at this hour (Brahman Female with Positive control). The average period’s plot is flatter than a single period’s plot. However, the trend and structure of these two are nearly identical, and a single cycle has more individual points and noises.

Throughout the sample period, the cattle’s condition varies on daily basis. The plot of the entire activity cycle contains noise and outliers in Fig. 5. Therefore, denoising the sampled data is required.

FIR and IIR are two types of digital filters that are extensively employed. In theory, an IIR function’s filtering effect is superior to that of an FIR function of the same order, but divergence can occur. The IIR digital filter has a high precision for amplitude-frequency characteristics, with a non-linear phase, it is suited for audio signals that are insensitive to phase information. FIR digital filters have lesser amplitude-frequency precision than IIR digital filters. However, the phase is linear, meaning the time difference between signals of various frequency components remains unaltered after going through the FIR filter. In addition, the calculation time delay is relatively tiny, it is suited for real-time signal processing[52, 53]. Because the state of the cattle is a time-series data, it is critical to ensure that the filtered phase remains constant. Therefore, in this work, we use the FIR low-pass filter for denoising.

Cattle monitoring data are sampled once every 60 seconds in this study, resulting in a sampling frequency of around 0.0167Hz. Noise frequency is more extensive than sampling frequency, so the signal between 0 and 0.0167Hz is kept while the signal above 0.0167Hz is eliminated. In Fig. 14, the filter length is set to 5, and the filter’s shape corresponds to its frequency. The filtered result is depicted in Fig. 15, which uses the resting time of Brahman Female’s cow with Positive control as an example.

Fig. 16 is a local detailed version of Fig. 15, focusing on the comparison of before using FIR filtering and after using FIR filtering from the 16th to the 20th day. Data performance is optimized after introduction of the FIR filter for smooth signal processing, and the data trend can be clearly identified.

After going through the FIR filter, Fig. 17 provides an image of a single rest period (one day, Day 17). In comparison to Fig. 7, it exhibits the same trend, i.e., one day’s rest time after filtering is nearly the same as one day’s typical rest time. This feature demonstrates that the cattle’s condition changes on a regular basis. It also indicates that the FIR filtered signal is effective and precise. The FIR filter effectively minimizes noise, and eliminates outliers, and gross inaccuracy. As a result, the signal filtered by the FIR filter can be used for subsequent modeling and prediction.

The program flow chart for establishing the LSTM model is shown in Fig. 18. First, import the data previously filtered by the FIR filter, and divide it into a test set and a training set. Second, the LSTM model is created. Setting parameters: the number of input neurons, output neurons, hidden neurons, learning rate, batch size, epoch size (i.e., the number of training cycles) and the number of LSTM layers[25]. The loss error is chosen as the mean square error, and the LSTM neural network is trained using the Adam optimisation technique[25]. The cycle ends when the number of training times is reached, and the lowest loss error will be the output.

It is critical to determine the input, output, and time series before using the constructed LSTM model for cattle state prediction. The cattle’s state must be presented as the output, and the number of the independent variable hours must be seen as a time series, according to the characteristics of the data sets. As a result, determining input variables is a challenging aspect of this approach. Because the output variable must be data with periodic changes, the input must be a known fixed periodic function. Time series as a fixed periodic function can be used as input. To be more specific, given that the state cycle of cattle is one day, it is appropriate to determine the input variable as the number of hours on the clock each day. The input and output variables, as well as the time series, for the resting time of Brahman Female’s cattle with Positive control are as follows:
Input: The number of hours on the clock each day (24 hours).
Output: The resting time during this hour (e.g. The resting time at 7:00 means that the resting time during one hour from 7:00 to 7:59).
Time series $t$: The sequence number of this hour (e.g. 0:00am on the first day is the first hour, and $t$ is 1. So on, 0:00am on the second day is the 25th hour, and $t$ is 25).

• •

  Training:
  Both the input and output data are periodicity. The distinction is that the input in this cycle has a set value and trend, whereas the output in each cycle has a varied value. For example, the input is 0 at 0:00 am on Day 17th and 0:00 am on Day 24th, as shown by the two red lines in Fig. 19, but the output is different. In other words, the same input might result in multiple outcomes regardless of time. Although the input is the same, the input’s matching time series is not. As a result, when a single input correlates to numerous outputs in a time series, the LSTM model can successfully handle the problem.

  

• •

  Testing and prediction:
  90% of the data is used for training, and 10% for prediction and testing. For example, the input data sets for training are $input_{t_{1}}$ through $input_{t_{90}}$, while the data sets for testing are $input_{t_{91}}$ through $input_{t_{100}}$. The training outcomes are depicted in Fig. 21.

  

  

The predict and actual results are similar shown in Fig. 20. The training loss reduce during the training process, showing that model is converged and practical. However, the prediction results’ error is relatively significant, which indicates further requirements of the parameter optimization in the model.

For optimization and comparison purpose, the number of hidden units, LSTM layers, the batch size, and the epoch size were all modified.
Hidden units size: 4, 8, 16, 32, 64, 128, 256.
The number of LSTM layers: 1, 2, 3, 4, 5, 6, 7.
The batch size: 3, 6, 12, 24, 48, 96.
The epoch size: 100, 500, 1000, 2000, 5000, 10000, 20000.

• •

  Selection of the number of LSTM layers
  The number of hidden units is 16, the batch size is 24, and the epoch size is 2000, all of which are randomly chosen. Only the number of layers in the LSTM is modified with the other parameters fixed: 1, 2, 3, 4, 5, 6, 7. The box diagram for the mean square deviation in the model learning process is shown in Fig. 22.

  

  The top line and bottom line represent the edge’s maximum and minimum values respectively. The upper quartile is represented by the box’s upper edge, while the box’s lower edge represents the lower quartile. The orange line represents the median. When comparing the seven box charts, increasing the number of layers has a minor impact on the mean square error of model training. However, in terms of model performance, using more LSTM layers, the running speed will be slower and it becomes more complex, and the result of the model operation is affected[25]. As a result, two layers of LSTM are best for this model.

• •

  Selection of the hidden units size

  

  To determine the size of the hidden units, we keep the batch size and epoch size unchanged and run the LSTM model with different hidden units size, i.e., 4, 8, 16, 32, 64, 128, 256. The box diagram of the mean square is shown in Fig. 23. In terms of error size and ultimate training effect, the choice of 128 hidden units is the best for training the data, with the majority of the mean square error values falling below 25.

• •

  Selection of the epoch size
  Select two layers of LSTM with 128 hidden units while keeping the rest of the settings the same: the batch size is 24, but the epoch size can be any of 100, 500, 1000, 2000, 5000, 10000, or 20000. Fig. 24 shows a box diagram for the mean square deviation in the model learning process.

  

  The epoch size is the number of times the learning algorithm works in the entire training data set. An epoch means that each sample in the training data set has the opportunity to update internal model parameters. In theory, the more training sessions there are, the better the fit and the lower the error. In practice, however, overfitting occurs when the epoch size exceeds a specific threshold, causing the training outcomes to deteriorate[25]. The epoch size of 100, 500, 1000, 2000, 5000, 10000, and 20000 is chosen in Fig. 24. The inaccuracy rapidly decreases and approaches zero as the epoch size increases from 100 to 10000. When the epoch size increases to 20,000, the error is still tiny, but it is greater than when the epoch size is 10,000, indicating an overfitting occurrence. Therefore, the model with a 10000 epoch size has the best effect.

• •

  Selection of the the batch size

  

  The batch size, which can be 3, 6, 12, 24, 48, or 96, is altered when using two layers of LSTM with 128 hidden units and an epoch size of 10000. The box diagram is shown in Fig. 25. The batch size refers to the number of samples fed into the model at once and divides the original data set into batch size data sets for independent training. This method helps to speed up training while also consuming less memory. To some extent, batch size training can help to prevent the problem of overfitting[25]. As a result, when building the model, an acceptable batch size should be chosen. When the batch size is 24, the minimum value of the produced mean square deviation data set is the smallest in terms of minimum value and median.

To sum up, the best parameters for the LSTM model are 128 hidden neurons, 24 batch size, 10000 epoch size, 2 LSTM layers. Fig. 26 shows the loss value after optimizing parameters, while Fig. 27 shows the training and prediction outcomes after optimizing model parameters. The LSTM model has a good prediction on the resting state of cattle, which largely adheres to the periodic changes in cattle state and has a modest error. Loss-train: 0.00787; Loss-test: 7.823968.

Figs.28-31 depicts the LSTM model’s training and prediction on different sexes, breeds, combined treatments, and states, respectively.

The trend of the results predicted by this LSTM model are nearly identical to the actual data. The model for Brahman males performs relatively poorly, which can be attributed to their relatively random rest state, poor cycle regularity, and other external environmental factors. It is possible that increasing the size of the data collection may result in improved predictions. Overall, the LSTM-based model for cattle states cycle is accurate and effective, and it can accurately predict the dynamic trend of the next cattle state cycle.

• •

  The LSTM model is applied to the other gender of cattle.

  

• •

  The LSTM model is applied to the other breed of cattle.

  

• •

  The LSTM model is applied to the other combined treatment of cattle.

  

• •

  The LSTM model is applied to the other state of cattle.

  

This work primarily involves three different pain treatments for pain management: topical anesthetic, meloxicam, and a combination of the two. In animal practice, topical anesthetics are the most widely utilized pre-emptive analgesics[55]. Without causing loss of consciousness, these chemicals cause a reversible loss of sensation in the local area. Local anesthetics enter nerve cells and block open sodium channels, inhibiting nerve impulse production and transmission. However, topical anesthetic must work in an alkaline environment, and the quality of local anesthesia in diseased, ischemic, or damaged tissue is frequently poor because the more acidic environment inhibits the medication from being separated[43, 55]. Topical anesthetics typically take 2 to 5 minutes to take effect, and studies have shown that they reduce plasma cortisol concentrations following castration, but not the mean area under the cortisol effect curve[43]. This issue shows that using topical anesthetics alone can help reduce the immediate pain associated with castration[43]. Meloxicam, an oxicam-class NSAID, has been licensed for the adjuvant treatment of acute respiratory illnesses in the European Union. Meloxicam muscle injection and horn nerve block lowered serum cortisol response for longer than cattle with a topical anesthetic before cauterization. This phenomenon means that the use of meloxicam before horn removal is effective in reducing pain in cattle. However, meloxicam alone is not effective in reducing castration-related acute distress. As a result, using a combination of topical anesthetic and meloxicam to reduce discomfort during castration or dehorning is beneficial[43].

The construction of a smart digital twin model of the state of cattle is primarily achieved in this work. It is primarily built on a farm IoT system to collect the state data of cattle under various combined treatments, with data cleaning and calculating. The average data of 24 hours are fitted, and the data of the whole sampling period are de-noised. In addition, a deep learning-based LSTM model for cattle state dynamics is developed using the data after noise reduction, and the model can predict the state change of cattle in the next cycle. The model’s accuracy and effectiveness are demonstrated when the prediction results are compared to the actual results.

Furthermore, the association between various behaviour patterns and bovine pain, the relationship between pain degree and behaviour pattern of cattle under various pain treatments are investigated. It is fund that the cattle in pain spent less time walking and ate less food, according to the final data. In cattle undergoing castrating or dehorning, combining topical anesthetic with meloxicam can be observed to successfully reduce the cattle’s pain.

Because pain is regarded as an emotion with a high subjective aspect, and the cattle behaviour is significantly impacted by external influences, existing data and technology do not quantify pain in a fully effective fashion. As a result, simply connecting the state of a cow with pain may not be accurate or helpful enough. Quantifying pain and measuring the intensity of discomfort in proportion to the duration of each behavioural state is a critical challenge that will need to be addressed in the future. Furthermore, encapsulating the entire research into another system is a critical step towards commercializing digital twins in the future.