High-Degree Connectivity Sensor Networks: Applications in Pastured Cow Herd Monitoring

Abstract

This paper explores the application of mobile sensor networks in cow herds, focusing on the challenge of achieving local communication under minimal computational constraints such as restricted locality, limited memory, and implicit coordination. To address this, we propose a high connectivity based sensor network scheme that enables individual sensors to self-organize and dynamically adapt to topological variations caused by cow movements. In this scheme, each sensor acquires local distribution data from neighboring sensors, identifies those with high connectivity, and forms a local network with a star topology. The overlap of these local networks results in a globally interconnected mesh topology. Furthermore, information exchanged through broadcasting and overhearing allows each sensor to incrementally update and adapt to dynamic changes in its local network. To validate the proposed scheme, a custom wireless sensor tag was developed and mounted on the necks of individual cows for experimental testing. Furthermore, large-scale simulations were performed to evaluate performance in herd environments. Both experimental and simulation results confirmed that the scheme effectively maintains network coverage and connectivity under dynamic herd conditions.

1. Introduction

Recently, the adoption of ICT and IoT technologies [1,2,3] in the livestock industry has been promoted to improve operational efficiency and reduce labor demands. Commercial devices, including estrus detectors [4,5], birth monitors [6], and feeding robots [7,8], are increasingly employed in beef cattle production, replacing traditional manual tasks. Early detection of illness or stress in cows is considered feasible through behavioral analysis aimed at identifying abnormal patterns.

Various approaches have been explored to monitor and identify abnormal behaviors in cows [9,10]. Behavior analysis using a single sensor has been explored through several approaches such as attaching triaxial accelerometers and magnetometers to cows to monitor movements [11,12,13] or installing communication sensors in cowsheds to infer behavior from radio signal strength [14,15,16]. Other approaches include using depth cameras to monitor calving [17,18] and microcontrollers to track heart rate and body temperature [19,20]. Additionally, management systems have been proposed for large-scale environments, such as pastures, by implementing dedicated communication infrastructures. Examples include systems that detect estrus cycles using communication sensors to form behavior-monitoring networks [21,22], as well as methods for managing pasture areas by analyzing the movement patterns of GPS-equipped grazing cows [23,24]. Based on these studies, communication-enabled sensors can be deployed to implement management systems in large-scale environments such as pastures. Despite their innovative contributions, most existing monitoring approaches have primarily focused on individual cow behavior, with limited exploration of interactions and social relationships within the herd. Furthermore, the specific types of information necessary for mechanistic visualization of interactions among the cows remain poorly defined. To facilitate understanding of these approaches,

Table 1. Comparison of representative cow monitoring approaches.

Approach	Cites	Purpose	Limitations
accelerometers	[11,12,13]	individual movement	focus on individuals;
& magnetometers		detection	limited social analysis
communication sensors	[14,15,16]	behavior inference	signal instability
(RSSI)		via signal strength
depth cameras	[17,18]	calving/	indoor use only;
		posture monitoring	infrastructural facility
microcontrollers including	[19,20]	physiological health	no social-context
heart rate or temperature		monitoring	information
communication networks	[21,22]	estrus and activity	limited social-
(estrus detection)		monitoring	behavior data
GPS movement	[23,24]	pasture movement	low resolution for
analysis		analysis	social interactions

provides a summary. Unfortunately, few studies have investigated stress assessment based on social interactions [25,26], such as mutual licking, or examined behavior at the group level.

To address these limitations, this study proposes a communicative interaction scheme for visualizing interactions among cows in a freely grazing herd as illustrated in Figure 1. Specifically, we devise a sensor network based on mutual communication among cows using mounted wireless sensors, leveraging radio signal strength to enable both individual identification and group behavior analysis in grazing environments. To achieve this, we address two main challenges: an engineering challenge, involving the configuration of a sensor-based communication network, and an agricultural challenge, involving the analysis of social relationships among cows [25,26]. To tackle the engineering challenge, sensors are attached to each cow, and they are identified based on high-degree connectivity with nearby cows, enabling the generation of local networks. To cope with the agricultural challenge, a friend network is established by limiting the communication range to emphasize relationships between cows that interact more frequently. The overall network is subsequently formed by overlaying the local and friend networks. This facilitated the development of a behavioral model capable of monitoring dominance hierarchies and health conditions within the herd, enabling the detection of social dynamics and abnormal behaviors. The proposed scheme can help reduce labor demands and enhance overall productivity. Finally, to implement this concept in cows, we developed a neck-belt sensor tag and conducted evaluation experiments. Additionally, to assess the effectiveness of the proposed method in large-herd scenarios, we developed an in-house simulator and performed corresponding simulations.

This paper is structured as follows. Section 2 outlines our research problem and its definitions. Section 3 provides an explanation of the proposed solution. Section 4 details a custom neck-belt sensor tag. Section 5 shows the results from extensive experiments performed to evaluate the effectiveness and the performance of the proposed approach. Finally, Section 6 presents our conclusions.

2. Problem Statement

Figure 2a shows a herd of n cows equipped with a wireless sensor-tag device. First, the communication range of each cow $c_i$ is defined as $CR$, and it is assumed that the range of $CR$ is constant for $c_i$. Moreover, the inter-communications within $CR$ are used to exchange information and to obtain the radio signal strength between the sensor-tag devices. Specifically, $c_i$ sends a message to a neighboring cow $c_j$ in $CR$, and $c_j$ receives information including the radio signal strength $rs{s}_{ij}$ (Figure 2b). A specific range acquired $rs{s}_{ij}$ is defined as $FR$. The range of $FR$ is assumed to be constant for $c_i$ as in $CR$, namely $FR$ is contained within $CR$$(CR\supset FR)$. Although $c_i$ has its own $ID$ initially, no specific roles such as leader, source, sink, and gateway are assigned. $c_i$ executes the same algorithm asynchronously but has no long-term state information.

We define $1Hop$ as the state where $c_i$ can directly communicate with $c_j$ within $CR$. The set of cows within $1Hop$ of $c_i$ is expressed as $H_i^1$. $2Hop$ refers to communication with cows outside $CR$ via an intermediate $1Hop$ cows. For example, in Figure 2c, if $c_m$ lies outside $CR$ but is within $1Hop$ of $c_j\in H_i^1$, then $c_m$ is a $2Hop$’s cow of $c_i$. The set of such cows is symbolized by $H_i^2$. From $H_i^1$, $c_i$ selects neighbors $n_{ij}$, forming the neighbor set $N_i$. Furthermore, $c_i$ selects friends $f_{ij}$ from $H_i^1$ based on $rs{s}_{ij}$, with the set defined as $F_i$. The computation procedures for selecting $n_{ij}$ and $f_{ij}$ are described in the following section.

Figure 3 illustrates three message types used in the proposed scheme. These definitions are as follows: (1) the existence message $e{m}_i$, which is periodically broadcast by $c_i$ within $CR$ and acknowledged by $a{m}_j$; (2) the output message $o{m}_i$, which is sent to $c_j$ and acknowledged by $a{m}_j$; and (3) the request message $r{m}_i$, which is sent to $c_j$ and elicits a response with $o{m}_j$. For simplicity, any message from $c_j\in H_i^1$ is represented by $m_j$. Next, a wait time $w_t$ is defined to evaluate connectivity. When $c_i$ broadcasts a message to $c_j$, it measures the response interval $t_{ij}$. If $c_i$ receives a reply from $c_j$ within $w_t$, then $c_j$ is considered to be within $CR$ and is included in $H_i^1$. Conversely, if no reply is received within $w_t$, the connection is regarded as broken, indicating that $c_j$ has moved outside the $CR$ range. Meanwhile, $c_i$ can also receive and/or overhear broadcast information from neighboring nodes.

The local network associated with $c_i$ is defined through its $1Hop$ connections to $c_j$. This yields a local graph $G_i=\{V_i,E_i\}$, where $V_i=\{v_1,v_2,···,v_n\}$ is the set of vertices and $E_i=\left\{(v_i,v_j)\right|v_i,v_j\in V_i\}$ is the set of undirected edges, excluding self-loops. The global herd network is then defined as $G={\bigcup }_{(i=1)}^n{G}_i$. Next, based on $G_i$, as illustrated in Figure 4, herd behavior is classified into four states: $SA$ (mutual friends within $FR$), $SS$ (split into multiple local networks), $US$ (unified single network), and $IS$ (isolated cow disconnected from the herd).

Based on the definitions and assumptions above, this paper addresses the monitoring of herd behavior through local communication interactions in a cow herd. To solve this problem, we propose a monitoring algorithm that visualizes herd behaviors and inter-cow relationships using local network configurations and virtual distances. Specifically, $c_i$ selects $N_i$ and $F_i$ while forming $G_i$ through intercommunication within $CR$ and $rs{s}_{ij}$. Herd behaviors and relationships are then inferred from $G_i$, and by combining all $G_i$, G is self-organized. Since G changes dynamically with herd movement, it requires partial modification, updating, and repair in response to behavioral changes. Details on the monitoring scheme are provided in Section 3.5.

3. Monitoring Algorithm Based on ${\mathit{G}}_{\mathit{i}}$ Configuration

This section describes the monitoring algorithm for generating and maintaining $G_i$ based on herd behaviors. As shown in Figure 5, the algorithm consists of four functions: $N_i$ selection, $F_i$ selection, $G_i$ generation, and $G_i$ maintenance. Four of these functions are explained in this section. Based on this algorithm, the identification of herd behaviors is described in Section 3.5.

The algorithm takes $m_j$, $t_{ij}$, $w_t$ and $N_i$ as inputs, and outputs $H_i^1$, $H_i^2$, $F_i$, and $N_i$. First, it checks whether $N_i$ is empty. If this is satisfied, $N_i$ and $F_i$ are selected, and $G_i$ is generated. If $N_i$ is not empty, the network maintenance is performed to accommodate changes in $G_i$ due to cow movements. Once $G_i$ has been generated or updated, a server system determines the herd behavior based on the resulting connection state. Finally, the outputs are broadcast to $c_j$, and this process is repeated continuously.

Next, we define the terms used in the algorithm. The union of all $H_j^1$ transmitted by $c_j$ is ${\bigcup }_{(c_j\in H_i^1)}{H}_j^1$. The correspondence between $c_i$ and $H_i^1$ via intercommunication is $T_{i,j}$. $f{c}_{ij}$ denotes a candidate friend in $F_i$ computed from $H_i^1$. Information from $c_i$ to $H_i^2$ through $H_i^1$ is expressed as $rs{s}_{i,12}$, and $min.rss$ represents the minimum received signal strength. Information held by $c_i$ at time t is $H_i^1\left(t\right)$, with the previous time step defined as $H_i^1(t-1)$; for simplicity, the time indices are omitted in subsequent explanations.

3.1. $N_i$ Selection

The first computation function selects $N_i$ and consists of two steps: information gathering from $c_j$ and neighbor searching.

3.1.1. Acquisition of $c_j$ Information

In the first step, $c_i$ receives $m_j$ from $c_j$. It first checks whether $m_j$ is sent by $c_j\in H_j^1$ and whether the response from $c_j$ occurs within $w_t$. If both conditions are satisfied, $c_j$ is retained in $H_i^1$ and sends an updated $H_j^1$. Using $H_i^1$ and $H_j^1$ within $CR$, $c_i$ calculates the state of $H_i^2$ beyond $CR$. For example, in Figure 6, $c_j$, $c_k$, and $c_l$ are within $CR$, and $c_m$, $c_n$, and $c_o$ are beyond $CR$. When $c_i$ connects to $c_j$, $c_k$, and $c_l$, and obtains $H_j^1$, a correspondence table $T_{i,j}$ is organized as mapping elements of $H_i^1$ (right side of Figure 6). From $H_i^1$ and ${\bigcup }_{(j\in H_i^1)}{H}_j^1$, $H_i^2$ is calculated:

$$H_i^2=\left(\bigcup _{c_j\in H_i^1}{H}_j^1\right)-H_i^1-\left\{c_i\right\}.$$

(1)

By calculating $H_i^2$, $c_i$ estimates information about cows beyond its $CR$. The intercommunication correspondence between $H_i^1$ and $H_i^2$ is defined as $I_{i,12}$ and $I_{i,21}$, respectively. Here $I_{i,12}$ represents connections from elements of $H_i^1$ to elements of $H_i^2$ as shown on the left in Figure 7a.

$$I_{i,12}:H_i^1\to H_i^2.$$

(2)

$I_{i,12}$, also represented by $T_{i,12}$, is shown on the right in Figure 7a. Conversely, $I_{i,21}$ represents connections from elements of $H_i^2$ to elements of $H_i^1$ as shown on the left in Figure 7b.

$$I_{i,21}:H_i^2\to H_i^1.$$

(3)

Similarly, $I_{i,21}$ is represented as $T_{i,21}$ as shown on the right in Figure 7b. The combination of $I_{i,12}$ and $I_{i,21}$ defines the connections from $H_i^1$ to $H_i^2$ via $H_j^1$ and is expressed:

$$I_{i,21}\circ I_{i,12}.$$

(4)

By using the composition of $I_{i,12}$ and $I_{i,21}$, $c_i$ can estimate the connection state of $G_i$.

3.1.2. $N_i$ Calculation

The second step selects $N_i$ using $H_i^1$ and $H_i^2$ obtained from $c_j$. First, $c_i$ identifies the element of $H_i^1$ with the highest number of $E_i$ to cows in $H_i^2$, defined as $MAX\left(E_i\right)$, and selects it as the first neighbor $n_{ij}$. The corresponding elements in $H_i^1$ and $H_i^2$ are then removed. This process is repeated to select the second neighbor and subsequent neighbors, each time removing the corresponding elements, until $H_i^2$ is empty (the right side in Figure 8).

3.2. $F_i$ Selection

The second computation function, called friend sorting, selects and determines $F_i$.

3.2.1. $f{c}_{ij}$ Calculation

The first step is to obtain $f{c}_{ij}$ from the information on $c_j$ as mentioned in Section 3.1.1. If both $rss\left(H_i^1\right)$ and $rss\left(H_j^1\right)$ are within the $FR$ range (Section 2), the corresponding element of $H_i^1$ is selected as $f{c}_{ij}$; otherwise, $F_i$ is set to empty. Here, $rss\left(H_i^1\right)$ denotes $f{c}_{ij}$ of elements in $H_i^1$. For example, if $c_j$ and $c_k$ are in $c_i$’s $FR$ (see the left side of Figure 9), the step checks whether $rss\left(H_i^1\right)$ from $c_i$ to $c_j$, and from both $c_j$ and $c_k$ to $c_i$ are mutually within $FR$. Elements satisfying this criterion ($f{c}_{ij}$, $f{c}_{ik}$) are selected as $f{c}_{ij}$ (see the right of Figure 9).

3.2.2. $F_i$ Calculation

The second step is to extract $F_i$ from $f{c}_{ij}$ obtained in the previous step. To begin, the procedure checks whether $H_i^2$ is non-empty. If it is satisfied, $rs{s}_{i,12}$ is calculated under the condition $f{c}_{ij}<rs{s}_{jn}$.

$$rs{s}_{i,12}=f{c}_{ij}+rs{s}_{jn}.$$

(5)

Next, if the calculated $rs{s}_{i,12}$ is the minimum value, it is assigned to $min.rss$, and $f_{ij}$ is computed accordingly.

$$f_{ij}=min.rss-rs{s}_{jn}.$$

(6)

Finally, the computed $f_{ij}$, as shown in Figure 10, is designated as the primary member of $F_i$.

If $H_i^2=⌀$, it assigns the minimum $f{c}_{ij}$ to $min.rss$, and computes $f_{ij}$ as the primary element of $F_i$.

$$f_{ij}=min.rss.$$

(7)

If multiple $min.rss$ values are equal, the corresponding $f_{ij}$ are all designated as main members of $F_i$.

3.3. $G_i$ Generation

The third computation function generates $G_i$. The input of $c_i$ is $N_i$, and the output is $G_i=(V_i,E_i)$. Vertices consist of $c_i$ and its selected $n_{ij}$, expressed as $v_i,v_{i,1},···v_{i,k}$, forming the set $V_i$. Each edge between $v_i$ and $v_{i,k}$ is defined as $e_{ik}=(v_i,v_{i,k})$, and $E_i=\left\{e_{ik}\right|1\le k\le m\}$ denotes the connection of $G_i$. When $c_j$ is selected as $F_i$, the corresponding local network of $c_i$ is generated. Specifically, since $c_i$ is connected to $c_k$ in $N_i$ through a $v_i$-to-$v_k$ connection, as shown in Figure 11, this configuration can be regarded as a star network topology. Aggregating all $G_i$ enables complete self-organization of G. When individual $G_i$ overlap, the overall network forms a partially connected mesh topology (see Figure 12). After constructing $G_i$, $c_i$ broadcasts $H_i^1$ and $F_i$ and exchanges this information with $c_j$.

3.4. $G_i$ Maintenance

The fourth computation function in Figure 5, network maintenance, addresses changes caused by cow movements. These changes occur in three states: (1) only $F_i$ changes (Figure 13a), (2) only $H_i^1$ changes (Figure 13b), or (3) only $H_i^2$ changes (Figure 13c). This function consists of five steps: information acquisition, $f{c}_{ij}$ calculation, $F_i$ calculation, $N_i$ calculation, and $G_i$ reconfiguration.

3.4.1. Information Acquisition

As the first step, the same information acquisition procedures described in Section 3.1.1 are carried out, with the following updates and corresponding assessments.

3.4.2. $f{c}_{ij}$ Calculation

In the second step, $f{c}_{ij}$ is recomputed as explained in Section 3.2.1. If both $rs{s}_{ij}$ and the $H_j^1$ elements are within $FR$, the updated $rs{s}_{ij}$ is selected as $f{c}_{ij}$; otherwise, $F_i$ remains unchanged.

3.4.3. $F_i$ Calculation

In the third step, using $F_i$ calculated in Section 3.2.2, when $H_i^2$ for the newly computed $f{c}_{ij}$ is non-empty, the elements of $f{c}_{ij}$ and $H_i^2$ are compared. Next, $rs{s}_{i,12}$ is calculated using Equation (5) when $f{c}_{ij}<rss\left(H_i^2\right)$. If the resulting $f{c}_{ij}$ is minimal, it is assigned to $min.rss$. The new $min.rss$ is then compared with the previous value; when it remains minimal, a new $f{c}_{ij}$ is computed using Equation (6) and becomes the member of $F_i$. If $H_i^2$ is empty, the minimal $f{c}_{ij}$ is assigned to $min.rss$ after compared with the previous value, and a new $min.rss$ is used to update $F_i$ via Equation (7).

3.4.4. $N_i$ Calculation

The fourth step investigates $N_i$ described in Section 3.1.2, using $H_i^1$ and $H_i^2$ computed in Section 3.4.1. $c_i$ first selects the element of $H_i^1$ with the highest number of edges $E_i$ to cows in $H_i^2$ (Equation (3)) as the first neighbor. The corresponding elements in $H_i^1$ and $H_i^2$ are then removed. This process is repeated to select subsequent neighbors until $H_i^2$ is empty.

3.4.5. $G_i$ Reconfiguration

In the fifth step, $G_i$ is partially reconstructed using updated information, following the same procedure as the $G_i$ generation function. First, the cow’s state is determined based on its movement. When $F_i$ changes, the vertices of $c_i$ and newly selected $n_i$ are updated in $V_i$. If $H_i^1$ changes, the connections between the new $H_i^1$ and $H_i^2$ members are recalculated using Equation (4). Similarly, variations in $H_i^2$ immediately trigger reconfiguration of its connections with $H_i^1$. This partial reconfiguration maintains $G_i$ without requiring a full rebuild.

3.5. Determination of Herd Behavior Pattern

For the monitoring of herd behaviors as explained in Section 2 (Figure 4), the determination function for four types of group patterns is described. We first check whether $H_i^1$ is an empty set. If $H_i^1=⌀$, the cow is classified as $IS$, indicating it is isolated from the herd. If $H_i^1\ne ⌀$, the herd is evaluated as a single herd. If all cows belong to a single network, the herd is classified as $US$. If the network is disconnected, with multiple sub-networks, the herd is classified as $SS$. Finally, if cows within the herd are mutually within each other’s $FR$ and maintain a state of mutual friendship, the herd is classified as $SA$.

4. Development and Operational Experiments of Wireless Sensor Tag Device

To identify herd patterns in cows, this section introduces a wireless sensor tag device (hereafter referred to as the “sensor tag”), which implements the monitoring algorithm described in Section 3.

4.1. Design Concept of Wireless Sensor Tag Device

The development process of the sensor tag is divided into three parts: the components, the frame of attaching the tag to each cow, and the monitoring system through a server receiver. These elements are required for the execution of the monitoring algorithm. First, the tag should support mutual communication with $RSS$ acquisition. Second, the tag should execute computations using the algorithm. Third, the tag should temporally store data received from other tags.

Next, from an installation standpoint, two key requirements should be met: battery-powered operation and protective housing. Battery operation is essential since external power cannot be used when the tag is attached to cows. The tag-housing frame protects the tag, and allows for attachment to a wearable neck belt. In addition, monitoring requires a server connection to confirm tag status and collect data.

4.2. Development of Wireless Sensor Tag Device

To realize the algorithm, as illustrated in Figure 14, we developed a server tag and a sensor tag consisting of two parts: an electronic module and a frame. Details are presented in the following subsections.

4.2.1. Electronic Modules

The electronic module comprises a communication unit and an arithmetic unit. First, for the communication unit, we use the ESP32-WROOM-32 microcontrollers (Espressif Systems, Shanghai, China). This module combines Wi-Fi and Bluetooth in a single unit and is suitable for wireless communication. Furthermore, the module contains a 32 kHz oscillator and a 2.4 GHz band antenna. The wireless standard used is Wi-Fi: 802.11 b/g/n. Figure 15 shows the structure of the packets used for communication. As shown in Figure 3, for communication, two packet types are established: length-fixed and length-varying, depending on the type of messages. The packet format shown in Figure 15a is used for $e{m}_i$, $a{m}_j$, and $r{m}_i$. On the other hand, Figure 15b illustrates the packet format used for broadcasting $o{m}_j$. Specifically, in the packet, data length refers to the length of the data, the command defines the action for each message, source $ID$ is the sender’s $ID$, target $ID$ is the recipient, and data carries the information for the command.

Second, the arithmetic unit uses an ESP32 microcontroller, a low-cost, low-power SoC with an Xtensa dual-core 32-bit LX6 processor (240 MHz, <600 DMIPS) and 520 KB SRAM. A micro SD card slot and a Real-Time Clock (RTC) module (SparkFun Electronics, Niwot, CO, USA) store the algorithm’s data along with transmission and reception timestamps.

Finally, $rs{s}_{ij}$ is obtained via the tag’s Wi-Fi module. The module first scans the network to identify the number of tags within $CR$, after which an $RSS{I}_{ID}$ is assigned to each detected tag. The scanned $RSS{I}_{ID}$ values, including $rs{s}_{ij}$, are then logged to the micro-SD card.

4.2.2. Frame Unit

The frame unit, powered by three AA batteries (4.5 V), houses the electronic module in a PLA resin case with a waterproof lid and neck-belt buckle. The module and batteries are arranged in a two-layer structure (Figure 16). Module size: 55 × 45 × 20 mm; frame size: 75 × 100 × 55 mm (including buckle). Figure 17 shows the assembled tag.

4.2.3. Monitoring Unit: Server System

As shown on the left-hand side of Figure 14, the monitoring device consists of a computer and an electronic module. The LG gram computer is an Intel(R) Core Ultra 7155H CPU with Microsoft Windows 11 Home. The Microsoft Visual Studio C++ 2022 environment of Visual Studio Code 1.107.0 was used for the algorithm. A module is attached to the computer to receive the tag information.

4.3. Preliminary Experiments for RSS Measurement and Network Generation

The signal strength of the developed tag decreases with distance. To determine the effective ranges of $CR$ and $FR$ based on distance-dependent signal attenuation, preliminary experiments were conducted using ten sensor tags and one server tag. The server tag was fixed in place, while the sensor tags were positioned at 10 cm intervals from 0 to 1 m and at 1 m intervals up to 50 m, respectively. As shown in Figure 18, two preliminary experiments were conducted: one in an indoor classroom, where measurements were taken at 10 cm intervals, and another in an outdoor training field, where measurements were taken at 1 m intervals.

Figure 19 presents the measurement data where the left side depicts short-distance measurements (1 m range, 10 cm intervals), while the right side shows long-distance measurements (50 m range, 1 m intervals). At the short distance cases, signal strength decreased proportionally with distance from the server, whereas at the long-distance cases, it exhibited exponential attenuation. In particular, considering the body size of cows and their affiliative behaviors, a radius of 2 m was defined as the $FR$ range. The corresponding $RSSI$ measured at 2 m was $-60$ dBm. In addition, although the $RSSI$ gradually decreased up to 30 m, no further change was observed beyond 35 m; thus, this distance was defined as the radius of the $CR$ range, with the measured $RSSI$ at that point being $-90$ dBm.

Additional experiments employing the developed tags were performed to examine intercommunication among cows and assess the connectivity between the cow-mounted sensors and the server system. As shown in Figure 20, the tags were attached to the necks of five cows housed in a free-barn with paddocks at the Miyazaki Prefectural Livestock Experiment Station. Under this experimental environment, we examined how a fixed server tag adapted its topology when five sensors were in motion. Individual left sides in Figure 21 indicate the distributions of five cows and the right sides are the states of the server’s local network configuration as time went on. The experimental results confirmed that the proposed algorithm effectively constructed a server-centered local network and enabled stable connections among the sensor tags attached to each cow. In addition, the proposed setup successfully regenerated the server-centered local network despite the movements of the tags attached to each cow.

5. Evaluation Results and Discussion

5.1. Experimental Results in Free-Barn with Paddocks

To verify whether the network configuration changes in response to cow movements, we conducted evaluation experiments and compared a portion of their movements with the corresponding network configuration. The experiments in this subsection were carried out in an open paddock with an attached free barn (20 × 15 m), which does not represent an enclosed metal structure. Although metal posts are present around the paddock area, their impact on radio propagation is considerably smaller than that of an indoor barn with reflective walls. This is since the paddock is open on all sides, the roof is elevated with large openings, and no continuous metal walls are present that could generate strong multi-path reflections.

As the first experiment, to examine whether the network configuration responds to the movements of a specific cow, we conducted tests using six cows equipped with tags mounted on their necks. In Figure 22, the green-circled cow passes the yellow cow and moves toward the red cow. Figure 23 shows the corresponding network configuration and the variation in received signal strength during the herd behavior, taking the cow enclosed by the green circle as the reference. The rectangles highlighted in red indicate that the cows are in close proximity, which is defined as an RSSI value between −30 and −60 dBm. The results confirmed that the network configuration can be generated and dynamically adjusted according to the movements of specific cows. Furthermore, by comparing a portion of the herd behavior with the corresponding network configuration, we verified that the network structure changes in accordance with herd movements when using a particular cow as the reference point.

Next, a tag was installed as a server at predetermined locations within the paddock, while individual tags were mounted on the necks of seven cows to conduct two types of verification experiments. In the first setup, the server tag was placed in the feeder, and white lines were drawn on the ground at 1, 2, 3, 5, and 7 m intervals in a fan-shaped layout. Figure 24 and Figure 25 show the cow behaviors around the feeder and the variations in $rs{s}_{ij}$ depending on distances, respectively. The results indicated that the cow wearing the green tag closest to the feeder exhibited the strongest signal. In other words, signal strength is inversely proportional to the distance from the feeder, as shown in Figure 25, where the green-tagged cow exhibits the strongest signal, confirming its closest proximity to the server tag.

In the second setup, the server tag was placed at the cowshed entrance to measure changes in $rs{s}_{ij}$ as cows entered. Figure 26 shows the scenes in which cows approached the server, and Figure 27 presents the corresponding variations in signal strength with respect to distance from the entrance. As expected, each sensor tag exhibited a higher RSSI value when passing by the server, enabling the identification of entry order. From these results, we conclude that the algorithm implemented on the developed tags can be used to monitor the frequency of visits to feeding and drinking areas and to infer dominance relationships within the herd.

5.2. Experimental Results in Grazing Field

Field experiments were conducted in a grazing environment with tags attached to 10 cows in one $ha$ pasture. Depending on hierarchy and signs of estrus, the cows sometimes clustered near specific individuals or split into subgroups. Figure 28 illustrates the network configuration during herd integration and division. In Figure 28a, when multiple herds merged, “friends” were selected from cows within the effective communication range defined in the basic experiments. In Figure 28b, when the herd split, the connection with $H_i^1$ was temporarily lost and then re-established, demonstrating that the network could dynamically generated and reconfigured connectivity.

Figure 29 shows $rs{s}_{ij}$ received by a tag attached to one cow when two cows separated from the herd. The signals from the cow moving with the herd were highlighted in red, with strong signals observed in specific periods. The periods marked by the brown arrow correspond to intervals of a licking behavior, an affiliative interaction. These results indicate that the network accurately reflected inter-cow distances both during affinity behaviors and when cows temporarily split from the herd.

5.3. Simulation Results for Larger Number of Cow

In the pasture-field experiments, 10 cows were employed. To investigate the effectiveness of the proposed algorithm for the larger herds of cows, verification evaluations were conducted using our in-house network simulator. The simulator implemented in the Microsoft Visual C++ environment was used to model herd movements and depict neighbor–friend relationships graphically. In the simulation settings, $c_i$ has its own identifier, but no predefined roles such as leader, source, sink, or gateway are assigned. All $c_i$ independently execute the same algorithm without maintaining long-lived states and operate asynchronously with respect to other $c_j$. Specifically, $c_i$ can broadcast its information to adjacent $c_j$ within its $CR$ using an initially assigned unit in the simulation. Conversely, $c_i$ can receive and/or overhear broadcast information from neighboring nodes. Finally, it is assumed that communication occurs without noise or interference.

Simulation results for herds of 50 cows are shown in Figure 30. Light blue and yellow circles indicate the $CR$ range of each cow, while black lines represent edges from $c_i$ to $c_j$, namely $n_{ij}$. The thick red lines show neighbors selected as friends after network generation. $G_i$, comprising many fixed cows and five moving cows, undergoes partial updates since the red cows move arbitrarily. In detail, five red cows move arbitrarily and simultaneously, causing topological changes in their generated $G_i$, while the other cows remain stationary. Under the proposed algorithm, the cows partially updated $H_i^1$, $H_i^2$, $F_i$, and $N_i$ through overhearing, and modified only the portions of $G_i$ corresponding to these changes, rather than regenerating the entire $G_i$ for all cows. The findings demonstrate that $G_i$ responds to local variations in relative positions through partial updates, rather than full reconfiguration. Consequently, we confirmed the self-organizing capability of the proposed algorithm in adapting to network dynamics.

To evaluate the advantages of $N_i$ selection under the proposed algorithm, simulations of $G_i$ configuration were conducted while varying $CR$. The simulations employed 30 distinct initial distributions of 50 cows, with $CR$ configured to 50, 75, and 100 units. After $G_i$ configuration in individual simulations, the numbers of cows in $H_i^1$ and $N_i$ for each cow were recorded and summed. Additionally, one cow was selected to compare its behavior with that of the entire 50-cow herd. The statistical results are shown in Figure 31, where error bars indicate 95% confidence intervals and boxes represent the 25–75% interquartile range. As expected from $1Hop$ communication properties, the number of cows in $H_i^1$ increased with the expansion of $CR$. In contrast, the size of $N_i$ remained largely stable regardless of $CR$. Similarly, the $F_i$ selection exhibited a consistent trend regardless of $CR$. Consequently, these findings demonstrate that $N_i$ selection is robust to variations in $CR$, as each tag identifies neighbors with higher connectivity using local distribution information.

5.4. Discussion

Based on the experimental findings, we confirmed that a local network can be formed using individual communication states, including the received signal strength from sensor tags attached to individual cows. By clarifying the relationship between these variables and signal strength, the network reconfiguration process could be characterized, linking the splitting and merging of grazing cows to corresponding network changes. This enabled the development of a behavioral model for monitoring dominance hierarchies and health conditions within the herd, allowing the detection of hierarchical dynamics, isolation due to injury, and other behavioral changes in grazing environments. The proposed approach has the potential to reduce labor burdens on farmers and improve overall productivity.

In practice, quickly identifying estrus, injuries, or other abnormalities in grazing cattle remains challenging. Missed estrus signs, in particular, can lead to failed artificial insemination and reduced productivity. As future work, we will analyze herd behavior in greater depth based on habitual patterns and hierarchical relationships, with the aim of visualizing inter-individual interactions and overall social structure.

We plan to develop algorithms capable of detecting support-required conditions such as illness or parturition using behavioral data. Incorporating not only basic behaviors (standing, lying, walking) but also mounting behavior, a key indicator of estrus, is expected to further enhance detection accuracy. The insights gained from this study may also be applicable to other herd-forming livestock species, contributing to more advanced behavioral monitoring and herd management technologies across diverse production systems.

6. Conclusions

In this study, we focused on the monitoring of herd patterns and inter-cow relationships by using relative-communication distances and neighbor-centered networks. We proposed a local network generation algorithm using a two-layer communication range, and validated it through extensive experiments with cows equipped with wireless sensor tags. The proposed algorithm enabled the configuration of star networks among selected neighbors, and by combining local networks, a global partially connected mesh network self-organized. By incorporating signal strength relative to distances, the network adapted to changes caused by cow movements. Field experiments demonstrated that the friend selection function reflects both affiliative interactions and social isolation in cows, whereas simulations confirmed the feasibility of generating and reconfiguring networks within large herds. The proposed algorithm allows monitoring of an affinity behavior and isolation status in grazing environments, potentially reducing labor for livestock management and improving productivity.