0 Introduction
The timing and intensity of rumination activities in cows are crucial metrics for assessing their daily behavioral patterns [ 1]. Continuous and real-time monitoring of rumination activities is beneficial for maximizing animal welfare and farm productivity [ 2]. Currently, the primary methods for monitoring rumination activities in cow are contactless, utilizing machine vision and wireless sensor technology [ 3]. Wearable sensor-based monitoring systems have gained increasing popularity due to their cost-effectiveness and ease of integration with wireless networks. The most commonly used sensors include sound sensors, pressure sensors, and velocity sensors. Sound monitoring technology identifies cow rumination activity by analyzing the sounds produced during the rumination process. However, sound recognition has a restricted detection range and is vulnerable to interference from noisy environments, which can compromise system efficacy. The use of hydraulic tubes in pressure sensors to collect cow information can affect their comfort and is susceptible to damage, potentially risking leakage of fluids. Velocity sensor can overcome the limitations of sound sensors and pressure sensors. Tani et al. [ 4] utilized a monitoring system equipped with a single-axis acceleration sensor. This system extracts feature patterns from feeding and rumination to distinguish jaw movements and then matches similar feature patterns in unanalyzed activities. However, the accuracy of recording chewing signals is affected by the attachment position of the sensor. Vázquez Diosdado et al. [ 5] developed a decision tree algorithm that utilizes three-axis acceleration data collected from sensors mounted on the cow's neck to distinguish behaviors of lying, standing, and eating. Benaissa et al. [ 6] also fixed three-axis acceleration sensors on the cow's neck to gather data and devised a simple decision tree algorithm to identify eating and rumination behaviors. Shen et al. [ 7] conducted further research on identifying cow rumination behavior using data obtained from three-axis acceleration sensors. Hou [ 8] proposed a deep learning model based on cow activity data to recognize cow rumination behavior, building on machine learning techniques. In all the studies mentioned above, the raw data collected by sensors need to be transmitted to a backend system for processing, making it challenging to achieve real-time monitoring of cow rumination behavior. Additionally, the transmission of a large volume of data results in higher energy consumption and shorter battery life for the sensors.
The progression of edge computing is accelerating the trend of shifting from cloud computing to the edge. Edge computing has become a solution that moves cloud services closer to the network edge, closer to data sources, and Internet of Things (IoT) devices [ 9- 11]. Edge intelligence, utilizing both edge computing and artificial intelligence technologies, enables the deployment of artificial intelligence (AI) algorithms at the network's edge, that analysis and aggregation occur near the data capture points [ 12]. Edge intelligence primarily consists of federated edge and split learning models. The federated edge intelligence model is achieved by deploying federated learning in wireless edge networks [ 13]. This model consists of multiple terminal devices and edge servers, collaboratively training an AI model across multiple nodes. Models trained locally on terminal devices are aggregated at the edge server, enabling terminal devices and edge servers to share the AI model [ 14]. Split edge intelligence is based on split learning techniques, dividing deep learning models into sub-models and performing distributed training at the edge. It splits the AI model into two parts, with one part deployed on terminal devices near the data input layer, and the other part trained on edge servers [ 15].
With the increasing popularity and development of IoT devices in the context of smart farming, cattle farms are now producing a wealth of data. The capability to process data in real-time at the terminal nodes is crucial for these farms. However, terminal devices have limited battery life and computational capabilities, making it a challenge to independently manage tasks that are both energy- and computation-intensive [ 16]. Edge computing is an emerging computing model that allows for computation to be executed at the network edge. It supports compute-intensive real-time monitoring applications in resource-constrained cattle farm environments. This approach significantly alleviates the load on network bandwidth and cloud data centers, reducing latency and energy consumptio in computation [ 17, 18]. Devices upload data to edge servers physically located nearby, offloading compute-intensive and energy-intensive tasks to edge servers. This effectively reduces energy consumption on terminal devices and enables real-time processing of tasks [ 19]. Bu and Wang [ 20] introduced a smart agricultural land system based on deep reinforcement learning, incorporating the concept of edge computing. This system comprises agricultural data acquisition, edge computing, agricultural data transmission, and cloud computing layers. Shen et al. [ 21] fused three-axis accelerometer into edge computing equipment to collect cow rumination information.
To enable the computational tasks to be completed in real-time at the edge, near the cow data source, thereby reducing data transmission volume and network latency, leveraging the advantages of wearable devices and edge computing, wearable six-axis sensor devices was employed in this research to collect information on cow rumination activities and deep learning algorithm was combined to realize the recognition of rumination behavior.
1 Materials and methods
1.1 Experimental data collection
The experiment was conducted at the Acheng Experimental base of Northeast Agricultural University from May 20 to June 20, 2022. The experimental subjects were 10 healthy Holstein cows in the non-lactating period. Six-axis sensors were used as the terminal devices for six-axis acceleration data collection, with a sampling frequency of 5 Hz. The total duration of the dataset is over 180 hours. The terminal devices were installed on the collars worn by the cows as shown in Fig. 1. The cows were separated in two separate barns, with each cow occupying a fenced enclosure made of iron bars. The experiment involved feeding the cows with a ratio of 3∶7 of concentrate feed and ryegrass hay twice a day, and providing them with an adequate water supply. Each cowshed was equipped with an infrared night vision camera installed 1.5 m in front and 1.7 m above the ground, totaling 10 cameras to serve as a verification system. These infrared night vision cameras were synchronized with their respective terminal devices in terms of clock time, enabling the continuous recording of cow activities throughout the experiment, which were then stored in the cloud. Through monitoring video footage, it was observed manually that cows exhibited continuous chewing and swallowing actions while at rest, which was considered as the occurrence of rumination behavior.
Fig. 1 Cow rumination behavior monitoring experimental field and sensor wearing position |
The perception module of the terminal device consists of the MPU6050, which is a 6-axis sensor that integrates a 3-axis MEMS gyroscope and a 3-axis MEMS accelerometer. It can be connected to third-party digital sensors through the I2C interface and communicate with all registers of the device. The wireless transmission module also utilizes Narrow Band-Internet of Things (NB-IoT) technology to reduce power consumption of the terminal device by transmitting data through network aggregation. The edge server is located at the edge layer and is responsible for handling service requests through the rational deployment and allocation of computing and storage capabilities at the network edge. The Nvidia Jetson AGX Xavier is chosen as the edge server.
Tensorflow 2.3 was chosen as the deep learning framework for the edge server, and Python 3.6 was employed as the development language. The edge server underwent a firmware flashing process using JetPack 4.2 initially. Subsequently, Miniforge was installed, and environment variables were configured. Finally, the deep learning framework Tensorflow 2.3 was installed.
1.2 Data processing
1.2.1 Pose analysis and calculation
By obtaining the cow's posture information at the current moment, a better understanding of the cow's rumination behavior can be achieved. Pose analysis and estimation were performed using three-axis acceleration and three-axis angular velocity data.
1) Selection of the posture coordinate system and definition of the posture angles. The calculation of the posture angles for cows required a coordinate system transformation, commonly using the body-fixed coordinate system (b-frame) and the navigation coordinate system (n-frame). In this study, local Cartesian coordinates coordinate system was chosen as the navigation coordinate system. The data collection from the terminal device was performed in the body-fixed coordinate system, while the posture calculation was done in the navigation coordinate system. Therefore, a coordinate system transformation was required. The coordinate transformation diagram is shown in Fig. 2. Since the body-fixed coordinate system is fixed to the cow, it changes with the cow's neck posture, thus, the cow's posture changes can be expressed by the coordinate transformation matrix from the navigation coordinate system to the body-fixed coordinate system, as shown in Equation (1) [ 22].
Fig. 2 Coordinate schematic diagram of dairy cow posture transitions |
Where is the coordinate transformation matrix from the navigation coordinate system to the body-fixed coordinate system; is the initial state of the body in the navigation coordinate system represented as the attitude vector; is the attitude vector of the body after the posture change in the body-fixed coordinate system.
In this study, the update of cow posture transformation matrices was accomplished using quaternion methods. This choice was made because quaternion methods could offer advantages over Euler angle methods when dealing with the rotation of the cow's posture, particularly in pitch angle induced a deviation of . Unlike Euler angles, quaternion methods do not encounter issues with singularities and gimbal lock, and they provide a comprehensive representation of posture information in all directions. Additionally, quaternion posture calculations involve lower computational complexity and enable real-time updates of posture angles.
2) Fusion attitude estimation based on the Kalman filtering. Using a single sensor, either an accelerometer or a gyroscope to estimate attitude angles can lead to errors [ 23]. Therefore, a Kalman filter algorithm was employed for data fusion to achieve attitude estimation and mutual complementation of multiple sensors. This approach aimed to address noise interference and obtain the optimal estimation of attitude angles.
The corresponding attitude angle information was calculated using the three-axis acceleration measured by the accelerometer, and the observation update equation of the system was established. The angular velocity update equation was then established using the information from the three-axis angular velocity and serves as the state update equation in the Kalman filtering (KF) process. The state update equation is given in Equation (2) .
Where is the optimal estimation of attitude angle at time k; is the optimal estimation of attitude angle at the previous time step; is the measurement value of the gyroscope, is the priori estimate of gyroscope error at time k; is the optimal estimation of gyroscope output error; and are the noises of the attitude system. The observation update equation of the system is established using the corresponding attitude angle information calculated from the three-axis acceleration measured by the accelerometer, and it is given in Equation (3) .
Where is the attitude angle information at time k; is the measurement noise of the attitude system.
The structure diagram of the attitude angle estimation based on the Kalman filter algorithm is shown in Fig. 3. By iterative computation, the data fusion of the attitude detection system is performed, and the optimal estimation of the attitude angles is ultimately obtained. By collecting three-axis acceleration data, the three-axis angular velocity data was calibrated and compensated to calculate the optimal pitch angle and roll angle .
Fig. 3 Block diagram of attitude Angle solution using the Kalman filter algorithm |
Fig. 4 shows the pitch angle and roll angle of a cow during rumination, comparing the experimental results obtained with and without the using of Kalman filtering for attitude estimation. Through comparative analysis, it can be observed that Kalman filtering reduces the fluctuations in the estimated attitude angles. The fusion of multisensor data based on Kalman filtering yields better results for attitude estimation compared to using a single sensor alone.
Fig. 4 Comparison diagram of attitude estimation experiments with and without the Kalman filter |
1.2.2 Feature extraction and selection
To comprehensively explore the statistical characteristics of the data, both time-domain and frequency-domain features were extracted from the six-axis data. SMV(Support Vector Machine) and attitude angles undergoed Fourier transformation to obtain frequency spectrum energy, frequency domain entropy, and the DC component as frequency-domain features. Frequency spectrum energy and frequency domain entropy reflected the energy consumption of rumination behaviors [ 24]. The DC component represents the magnitude of the first component obtained after performing a fast Fourier transform. It is related to the reverse gravity acceleration and the components in the x, y, and z axes, thus reflecting the attitude of the cow's neck.
For the six-axis data, a data feature set consisting of 96 dimensions was extracted, which included both time-domain and frequency-domain features. The feature information is presented in Table 1. The six-axis data was segmented, and a continuous and non-overlapping sequence of 288 frames was selected as the minimum processing unit. Feature extraction was performed within groups using a sliding window of length 16 frames. The sliding window has a window length of 3, resulting in a dataset of 96 frames. After feature extraction, the data set has an input dimension of 96×96×1.
Table 1 Time_domain and frequency_domain feature extraction information for six_axis data |
| | | |
|---|---|---|
| | 9 | |
| | 9 | |
| | 9 | |
| | 9 | |
| | 9 | |
| | 9 | |
| | 9 | |
| | 9 | |
| | 9 | |
| | 6 | |
| | 3 | |
| | 3 | |
| | 3 | |
1.3 Federated edge intelligence model
1.3.1 Improved MobileNet v3 with Co-attention mechanism
The deployment of deep neural network models typically requires high-performance computing hardware support. However, edge devices have limited network resources compared to cloud servers, necessitating the selection of lightweight neural networks as the base network for research on cow rumination behavior recognition methods. In 2017, the Google team proposed the MobileNet network, a network specifically designed for lightweight convolutional neural networks, aimed at enabling neural networks to be deployed on edge devices such as mobile devices and embedded systems. MobileNet constructs lightweight neural networks through depthwise separable convolutions, significantly reducing model parameters and computational requirements with a minor decrease in accuracy compared to traditional convolutional neural networks. MobileNet has evolved into three versions: v1, v2, and v3, becoming an essential tool for neural network applications on mobile devices.
MobileNet v3 incorporates the Squeeze and Excitation SE attention in some bneck blocks to improve model accuracy by increasing the weights of salient feature channels. However, SE attention ignores positional information and only considers internal channel information. In this study, the CA attention was introduced after the depth wise convolution, which followed the inverted residual module within the bneck structure, thereby forming the CA-bneck. As shown in Fig. 5, the CA-bneck structure embedds positional information into channel attention, thereby reduces additional computational costs and further enhances MobileNet v3's focus on key features.
Fig. 5 Structure diagram of CA-bneck |
The constructed dataset of size 96×96×1 was used as the input for the CA-MobileNet v3 network. The structure of the improved fusion cooperative attention mechanism in the CA-MobileNet v3 network is shown in Table 2. The CA-MobileNet v3 network performed downsampling through convolutional stride operations without pooling operations.
Table 2 Structure of CA-MobileNet v3 |
| Input | | Exp size | Output | CA | NL | |
|---|---|---|---|---|---|---|
| 96×96×1 | conv2d，3×3 | — | 16 | — | HS | 2 |
| 48×48×16 | bneck，3×3 | 16 | 16 | √ | RE | 2 |
| 24×24×16 | bneck，3×3 | 72 | 24 | — | RE | 2 |
| 12×12×24 | bneck，3×3 | 88 | 24 | — | RE | 1 |
| 12×12×24 | bneck，3×3 | 96 | 40 | √ | HS | 2 |
| 6×6×40 | bneck，5×5 | 240 | 40 | √ | HS | 1 |
| 6×6×40 | bneck，5×5 | 240 | 40 | √ | HS | 1 |
| 6×6×40 | bneck，5×5 | 120 | 48 | √ | HS | 1 |
| 6×6×48 | bneck，5×5 | 144 | 48 | √ | HS | 1 |
| 6×6×48 | bneck，5×5 | 288 | 96 | √ | HS | 2 |
| 3×3×96 | bneck，5×5 | 576 | 96 | √ | HS | 1 |
| 3×3×96 | bneck，5×5 | 576 | 96 | √ | HS | 1 |
| 3×3×96 | conv2d，1×1 | — | 576 | √ | HS | 1 |
| 3×3×576 | pool，3×3 | — | — | — | — | 1 |
| 1×1×576 | conv2d 1×1，NBN | — | 1 024 | — | HS | 1 |
| 1×1×1 024 | conv2d 1×1，NBN | — | 1 | — | — | 1 |
|
1.3.2 Federated edge intelligence model based on CA-MobileNet v3
The construction of the federated edge intelligence model involved two main entities: terminal devices and edge servers. Terminal deviced collaborate to train and share a lightweight neural model, while the edge server was responsible for collecting local model parameters sent by the terminal devices and performing aggregation. In this study, the terminal deviced and edge server shared the fused cooperative attention mechanism of CA-MobileNet v3 and perform collaborative training using distributed cow data samples available on the terminal devices.
The federated edge intelligence system in this study consisted of M terminal deviced and a base station equipped with an edge server. The base station was connected to the terminal devices through a wireless channel. The training process of the federated edge intelligence model based on CA-MobileNet v3 follows the flowchart shown in Fig. 6. The learning process of the federated edge intelligence model typically involved four steps:
Fig. 6 Flow chart of training of federated edge intelligent model |
1) The edge server initialized the relevant parameters, and the terminal devices download the global model as their initial local models;
2) The terminal deviced train their local models using the collected real-time local data and upload the model parameters;
3) The edge server collected the local model parameters from each terminal device and performs aggregation to update the global model;
4) The above three steps were repeated until the global model converges, and the edge server deploys the updated model parameters to each terminal device.
Through interactions with cattle ranch users, terminal devices obtain labeled training samples. These samples are used as the local dataset as Equation (4) .
where represents the features of the i-th training sample of the m-th terminal device, represents the corresponding sample label indicating whether it is rumination or not, and denotes the number of training samples owned by the m-th terminal device. The training objective of federated learning is to minimize the global loss function L( w). When the CA-MobileNet v3 with the fused cooperative attention mechanism is deployed on the terminal devices, the local loss function of the m-th terminal device on its local dataset is defined as shown in Equation (5) , and the global loss function is defined as shown in Equation (6) .
In federated edge intelligence systems, edge servers collect local model parameters from terminal devices and to refine the global model through aggregation. In order to improve the performance of federated learning, the strategy of averaging the parameters was adopted, as implemented in the FedAvg algorithm proposed by Mc Mahan et al [ 25]. , as shown in Equation (7) .
Where represents the data size of the local dataset on terminal device m; represents the total data size; represents the local model parameters of terminal device m in the r-th communication round; represents the global model parameters at this time.
1.4 Split edge intelligence model
The convolutional recurrent neural network (CRNN) combines the strengths of convolutional neural networks (CNN) and bidirectional long short-term memory (Bi-LSTM) networks to handle tasks that involve sequential data [ 26, 27]. The MobileNet-RNN model based on the idea of CRNN was proposed. The model consists of three parts: a convolutional neural network module, a recurrent neural network module, and a fully connected module. The CNN module was composed of a lightweight CNN, CA-MobileNet v3, which incorporates a fusion collaborative attention mechanism. Firstly, CA-MobileNet v3 was used to extract features from the six-axis data of cows, reducing computational complexity. Secondly, the extracted features were fed into a Bi-LSTM layer, followed by a fully connected module to recognize the rumination behavior of cows. The fully connected layer consisted of one fully connected layer and one Softmax classification layer. To avoid overfitting, a batch normalization (BN) layer was introduced after the Bi-LSTM output and the fully connected layer. Finally, the Softmax classification layer was used to recognize whether cows were engaged in rumination behavior. The computational process is illustrated in Equations ( 8 ) and ( 9 ).
Where R={ represents the output results of the Bi-LSTM; denotes the results after batch normalization processing.
To implement split edge intelligence, the MobileNet-LSTM model was proposed, and the overall network structure is depicted in Fig. 7. The MobileNet-LSTM network comprised a CNN module, a RNN module, and a Fully Connected module. The CNN module is deployed on the edge server. CA-MobileNet v3 extracts features from the input dataset, ultimately obtaining a sequence of cow behavior features through feature dimension reduction. This sequence of behavior features was then fed into the RNN module, which consists of a Bi-LSTM network. The Bi-LSTM network captured temporal correlations in cow behavior data. The fully connected module included two fully connected layers and a Softmax classification layer. BN was introduced after each fully connected layer to prevent overfitting. Finally, the softmax classification layer was employed to recognize cow rumination behavior.
Fig. 7 Structure diagram of MobileNet-LSTM network |
To build a distributed edge intelligence model, the MobileNet-LSTM was divided into two parts using split learning techniques. A splitting layer was introduced between the convolutional neural network module and the recurrent neural network module. The shallow network and deep network were trained separately. The shallow network, which consists of the convolutional neural network module, was deployed on the terminal devices for feature extraction. It extracts useful features from the raw input data. The deep network, including the recurrent neural network module and the fully connected module, was deployed on the edge server. The deep network performed fusion learning among features to extract more complex higher-order features. As shown in Fig. 8, the training process of the split-based edge intelligence model based on MobileNet-LSTM typically involves five steps.
Fig. 8 Training flow chart of split edge intelligence model |
1) Initialization: Each terminal deviced and the edge server initialize their respective network models;
2) Data collection and forward propagation: Terminal deviced collect data and perform forward propagation until the splitting layer, obtaining the output features of the splitting layer. These features were then uploaded to the edge server;
3) Forward and backward propagation at the edge server: The edge server receives the feature data, performs forward and backward propagation, and obtains the gradients of the splitting layer. It sent these gradients to all terminal devices;
4) Backward propagation at terminal devices: Terminal devices used the gradients of the splitting layer to perform backward propagation;
5) Iteration: Terminal deviced and the edge server iteratively execute the above steps until the model converges.
1.5 Performance evaluation metrics
The performance evaluation metrics calculated based on the recognition results were effective for measuring the uncertainty between the predicted class and the true class, aiming to evaluate the classifier's performance. In real-time cow rumination behavior recognition tasks, a test data input was assigned to one and only one predefined class, allowing for clear definitions of true positive (TP), false positive (FP), true negative (TN), and false negative (FN). Therefore, performance can be evaluated using metrics such as Precision, Recall, F 1-Score, Specificity, and Accuracy. The formulas are shown in Equations ( 10 )~( 14 ).
2 Results and analysis
2.1 Test results of federated edge intelligence model
In the experiment, a preprocessed dataset that conforms to the input requirements of the neural network model was selected for training. The dataset was divided into training, validation, and testing set in a ratio of 6:2:2. The dataset includes approximately 5 million preprocessed six-axis acceleration data points, with around 3 million data points in the training set and approximately 1 million data points each in the validation and test sets. The CA-MobileNet v3 network was utilized to construct the real-time cow rumination behavior recognition model. During the model training, the cow rumination recognition model was trained on the edge server using the training set. The validation set was used for parameter adjustment and further training until the model converges. In the model test, the cow rumination recognition model obtained during the training phase based on CA-MobileNet v3 was deployed on both the terminal devices and the edge server. The testing set was then used to evaluate the rumination recognition results.
To validate the role of the collaborative attention module in improving rumination recognition accuracy, a comparison was made between the MobileNet v3 and CA-MobileNet v3 model. The performance evaluation metrics of both models are presented in Table 3. From the data in the table, it can be observed that CA-MobileNet v3 exhibits a relative increase of 5.3% in precision and 2.7% in accuracy compared to MobileNet v3. This suggested that the integration of collaborative attention enhances the model's performance in cow rumination recognition tasks.
Table 3 Performance comparison of MoblieNet v3 and CA-MoblieNet v3 model of federated edge intelligence model |
| | MoblieNet v3 | CA-MoblieNet v3 |
|---|---|---|
| Precision/% | 90.5 | 95.8 |
| Recall/% | 91.6 | 93.6 |
| F 1-Score/% | 91.1 | 94.7 |
| Specificity/% | 94.5 | 97.7 |
| Accuracy/% | 93.5 | 96.2 |
The federated edge intelligence model leveraged data to complement each other, effectively increasing the effective data volume. It achieved this without the need to centralize cow data samples, enabling efficient feature extraction and utilization. During the training process of the federated edge intelligence model, the following settings were applied: Training Batch-size 16; number of Epochs 125; Initial Learning Rate 0.001; number of Terminal Devices Participating in Federated Learning System 10.
In the federated learning process, each terminal device uploaded the weight parameters of its local model to the edge server after receiving the new global model. The federated center server was hosted on the edge server, equipped with an NVIDIA GTX 1080Ti GPU, an Intel Core i7-7800X CPU, 64GB of RAM, and 11GB of VRAM.
Comparative experiments were conducted by setting different values for I (the number of iterations for local training on terminal devices) and analyzing the accuracy and loss curves shown in Fig. 9. From the loss curve, it can be observed that as I increase, the model's convergence speed becomes faster. The accuracy curve revealed that as I increase from 1 to 6, the accuracy of the converged model gradually improves. However, when I increase from 6 to 10, the accuracy of the converged model shows a declining trend. This is because increasing the value of I results in more local iterations of gradient descent training on terminal devices, leading to faster convergence and improved accuracy. However, an excessively large I value leads to too many local training iterations between two model aggregations, causing overfitting to local data and weakening the effectiveness of the FedAvg federated aggregation algorithm.
Fig. 9 Accuracy curve and Loss curve of training models with different I values of CA-MobileNet v3a. Accuracy curves b. Loss curves |
In summary, setting I = 6 achieves the best recognition performance, and the corresponding performance evaluation metrics are shown in Table 4. From Table 4, it can be seen that the federated edge intelligence model based on CA-MobileNet v3, after undergoing federated learning, exhibits improvements in performance metrics. Specifically, it shows a 4.6% increase in recall and a 2.4% increase in accuracy. These experimental results indicate that the CA-MobileNet v3 model based on federated learning can effectively enhance the recognition of cow rumination behavior.
Table 4 Performance comparison results of federated learning |
| Performance metrics | non-federated learning model | federated learning model |
|---|---|---|
| | 95.0 | 97.1 |
| | 93.3 | 97.9 |
| | 94.1 | 97.5 |
| | 97.2 | 98.3 |
| | 95.8 | 98.2 |
2.2 Test results of split edge intelligence model
The split edge intelligence model was trained in a supervised manner. To validate the effectiveness of the proposed method, ablation experiments were conducted between CA-MobileNet v3 and MobileNet-LSTM. Both models were configured with the same training parameters: a batch size of 16 125 epochs, and an initial learning rate of 0.001. Ablation experiments on CA-MobileNet v3 and MobileNet-LSTM resulted in loss and accuracy curves for both training and testing, as shown in Fig. 10.
Fig. 10 Comparison of training curves between CA-MobileNet v3 and MobileNet-LSTM of split edge intelligence modela. Accuracy curves b. Loss curves |
As seen in Fig. 10a, the MobileNet-LSTM, based on the split edge intelligent model, achieves a final recognition accuracy of 96.2%. This was an improvement compared to the 95.8% accuracy achieved by CA-MobileNet v3. The higher accuracy of MobileNet-LSTM was attributed to its fusion of contextual information related to cow behavior, which enhances the recognition accuracy. From Fig. 10b, it can be observed that MobileNet-LSTM converges faster compared to CA-MobileNet v3. This is because MobileNet-LSTM incorporates a BN layer, which speeds up the training of the network.
2.3 Experimental contrastive analysis
The federated edge intelligence model based on CA-MobileNet v3 and the split edge intelligence model based on MobileNet-LSTM were selected for comparative experiments. The performance indicators of the experimental results are shown in Table 5.
Table 5 Comparison of the performance of federated and split edge intelligence models |
| | Federated edge intelligence model | |
|---|---|---|
| Precision | 97.1 | 95.8 |
| Recall | 97.9 | 93.7 |
| F 1-Score | 97.5 | 94.8 |
| Specificity | 98.3 | 97.7 |
| Accuracy | 98.2 | 96.2 |
The federated edge intelligent model based on CA-MobileNet v3 outperforms the split edge intelligent model based on MobileNet-LSTM, leading to further improvements in performance metrics. This is because the federated learning-based model utilizes local training and uploads model parameters, aggregates the parameters to update the global model until convergence, and can extract deeper cow rumination behavior features. On the other hand, the split edge intelligent model required substantial intermediate data transmission between the terminal devices and the edge layer, which may result in data loss and a decrease in recognition accuracy.
Fig. 11 presents the experimental results in the form of box plots for MobileNet v3, CA-MobileNet v3, the federated edge intelligent model, and the split edge intelligent model. Through experimental comparisons, it can be observed that the federated learning-based CA-MobileNet v3 network, i.e., the federated edge intelligent model, not only improves the recognition accuracy of rumination behavior but also led to a more concentrated data distribution. Compared to the split edge intelligent model based on MobileNet-LSTM, the federated edge intelligent model based on federated learning and CA-MobileNet v3 exhibited a more stable and concentrated prediction distribution, with fewer outliers. This indicates that the model has greater reliability and effectiveness in recognizing cow rumination behavior. Although the method proposed by Shen et al. [ 7] achieved a recall of 94.3%, which was better than that of this research, it was at the cost of transmitting large amounts of data and increasing equipment energy consumption. In contrast, although the performance indices of this study have declined slightly, this method have reduced the amount of data transmission and cloud computing to achieve real-time cow rumination recognition in an environment with low network performance.
Fig. 11 Boxplot of experimental results distribution for MobileNet v3， CA-MobileNet， Federated edge intelligence model， and Split edge intelligence model |
3 Conclusions
In this study, based on the concept of edge computing, proposed the use of edge devices to capture and process real-time six-axis acceleration signals of cows. By integrating a cooperative attention mechanism into the MobileNet v3 network, the CA-MobileNet v3 network was introduced. The federated edge intelligent model was subsequently constructed utilizing the CA-MobileNet v3 network in conjunction with the FedAvg model aggregation algorithm.
Experimental findings reveal that the proposed CA-MobileNet v3 network enhances precision by 5.3% compared to the MobileNet v3 network, while the FedAvg federated aggregation algorithm boosts the recall rate by 4.3% within the federated edge intelligent model, underscoring the efficacy of the proposed federated edge intelligent model. Furthermore, leveraging the CA-MobileNet v3 network and the Bi-LSTM network, a split edge intelligent model based on MobileNet-LSTM was designed, and comparative experiments were conducted between the federated edge and split edge intelligent models. The experimental results show that the federated edge intelligent model achieves the best recognition performance, with average Precision, Recall, F 1-Score, Specificity, and Accuracy for cow rumination behavior reaching 97.1%, 97.9%, 97.5%, 98.3%, and 98.2%, respectively.
The propoesd edge intelligence model not only effectively expands the dimensionality of cow data samples but also improves the accuracy of cow rumination behavior recognition.
