Assessing Signal Distortion in Seismic Sensor Security Systems: A Machine Learning Approach to Target Identification

The ability to effectively detect targets across various domains has gained significant importance in the modern era. This ability is necessary in the military and civilian sectors, which have unique requirements and challenges. The critical nature of target detection has motivated numerous researchers to explore and innovate across a wide range of methods that aim to enhance detection capability, particularly in areas that face various types of illegal intrusions. When armed incursions suddenly occur, significant economic and human catastrophes may result; hence, urgent action is necessary to prevent such incursions from escalating [1]. In addition, unauthorized migrants who frequently gather at the borders of some countries may seize the opportunity to enter these countries in groups, imposing significant economic effects on these countries [2]. Meanwhile, poachers can infiltrate environmental reserves and natural parks, causing considerable harm by hunting and killing endangered animals, especially those on the brink of extinction [3].

Considering the aforementioned challenges, robust research is necessary to create effective security measures for protection and continuous monitoring. Initially, electric fences and electronic balloons have been installed in some areas to provide enhanced security. Furthermore, the integration of different types of cameras in surveillance systems. Within this approach, efforts have focused on improving the processing of pixels, leading to increased efficiency and accurate target discrimination. This capability enables continuous visual monitoring and recording, which are crucial for security and investigation, are realized [4]. In addition, robots equipped with advanced sensors and autonomous navigation systems have been employed to conduct comprehensive investigations in areas considered too dangerous or inaccessible for humans [5]. Regardless of whether they are used independently or with robots, improving the performance of algorithms is also a significant area of focus; this process aims to achieve precise detection of the intended targets in open and closed environments [6].

However, the theories mentioned above and their applications are inherently financially demanding and require substantial data storage capacity. This challenge is particularly significant when these theories and applications are deployed over regions where vast coverage is essential, such as extensive border areas between countries or large military zones. Considering the financial burden and logistical challenges associated with these technologies, many researchers have searched for alternative solutions. Accordingly, different types of sensors have gained popularity because of their low cost and manageable data storage requirements. Among them, vibration or seismic sensors have been proven to be valuable. These sensors have been widely utilized in various security applications because of their ability to detect friction between objects or targets and the ground [7] which produces slight surface distortions. Vibration sensors can accurately detect these distortions by converting them into electrical signals; hence, the presence of objects can be sensed effectively [8]. This capability highlights the effectiveness of vibration sensors in detecting a wide range of targets, including people, vehicles, and animals [9, 10].

Seismic sensors exhibit numerous advantages, such as being cost-effective, requiring minimal operational power, and generating data that are easier to store and manage than conventional monitoring systems. Their cost-effectiveness and energy efficiency make them particularly suitable for large-scale deployments. In addition, the sensors' capacity to produce manageable datasets simplifies the processing and analysis of collected information. This feature further enhances the practicality of these sensors. Furthermore, vibration sensors can be buried discreetly under the soil, providing a significant advantage for concealment. Such hidden deployment helps avoid detection by targets and ensures that sensors are protected from environmental factors and tampering. The use of sensors in various security scenarios ranging from border monitoring to wildlife conservation emphasizes their importance as practical and effective alternatives in target detection. Given their versatility and efficiency, vibration sensors have become invaluable tools in security and surveillance. The deployment of these sensors across diverse environments and their ability to reliably detect targets highlight their crucial role in modern detection systems [11].

Despite their advantages, the use of signal-based detection methods also poses some challenges. Firstly, the signals generated by seismic sensors can be influenced by various factors, complicating their deployment and effectiveness. The localization of sensors is a significant concern because they are frequently used in diverse areas, such as forests, urban environments with large buildings, or other complex terrain. Some regions may not be fully covered, leading to gaps in detection if sensors are not placed optimally or if the mapping of targeted protected areas is inadequate [12]. The characteristics of the environment in which these sensors will be placed are also critical in determining the speed and amplitude of the waves that will be detected. For example, the characteristics of the ground surface, i.e., whether it is flexible or rigid, significantly affect wave amplitude. A rigid surface may reflect waves differently than a flexible one, affecting the accuracy of the detected signals. Given such variability, understanding how different surfaces influence signal propagation is essential to ensure reliable detection [13, 14]. In addition, the activity of the target plays a significant role in the characteristics of the signals. For example, vehicle load, wheel type, and the extent of the target’s friction with the ground surface contribute to the resulting distortions regardless of whether the target is human or vehicle. Subsequently, these distortions are converted into electrical signals; however, the characteristics of these signals can vary widely depending on the aforementioned factors [10, 14].

Another critical factor that affects the accuracy of signals generated by seismic sensors is environmental noise. Proximity to factories, train tracks, aircraft runways, or other sources of constant noise can interfere with the ability of sensors to detect targets accurately. In the study by Parihar et al. [15], where elephant noise was problematic, the researchers had to use a low-pass passive RC filter with a cutoff of 110 Hz to eliminate the sounds of elephants, thereby improving the accuracy of the signal. The above factors can result in potential inaccuracies in detection results, including false positives or false negatives. Therefore, studying and understanding the behavior of these signals and the external factors that can affect them is crucial. This research area has become increasingly important, and numerous studies have focused on mitigating environmental noise and optimizing sensor deployment strategies to reduce interference and increase detection accuracy. However, previous research has not fully addressed the practical aspects of understanding signal behavior in certain key areas. To fill this gap, our study aimed to thoroughly investigate the behavior of different types of targets and their interactions with various environments. To achieve this objective, we conducted experiments involving various targets, including humans, animals, cars, and motorbikes, to achieve this objective. Each target demonstrated different activities and behaviors.

In our experiments, we observed humans and animals while running and walking states. For humans, we analyzed the distinct signal patterns generated during brisk running compared with those produced during steady walking. Similarly, we studied moving animals and noted variations in signals produced by different gaits and speeds. Moreover, we tested cars and motorbikes at speeds ranging from slow movement to rapid acceleration as they approached and entered restricted areas. This set of activities enabled us to capture a broad spectrum of signal behavior associated with different types of movement. We conducted experiments in environments with varying surface types to understand how different types of terrain affect signal amplitude. We selected surfaces, such as mud, asphalt, grass, and soil, to represent various real-world conditions. We observed how soft and pliable surfaces affected the amplitude and speed of signals by conducting experiments in muddy areas.

Meanwhile, experiments performed on asphalt provided insights into the influence of hard and rigid surfaces on signal reflection and transmission. Moreover, grass and soil surfaces provided additional variations, highlighting the effects of vegetation and lose ground on signal behavior. By focusing on diverse target activities and environmental conditions, we expect that the presence and movement of the targets will directly affect the signals detected by sensors. Understanding these effects is crucial for improving the accuracy and reliability of detection systems. For example, friction is produced by different types of movement on various surfaces, generating distinct signal patterns that can be used to differentiate among targets.

Furthermore, we investigated the specific characteristics of different environments, such as urban, rural, or natural reserves, which require protection. These environments pose unique challenges and are likely to produce significant variations in sensor signals.

By thoroughly analyzing and examining signals, we aim to increase our understanding of key factors that influence signal generation and propagation. Our goal is to comprehensively understand how different targets and environmental conditions affect signal behavior.

Accurate classification of various target movements will enable us to improve the deployment and effectiveness of vibration sensors in diverse scenarios. This comprehensive analysis will help identify optimal sensor placement strategies and refine signal processing techniques to minimize false positives and negatives.

In summary, our study aims to provide a comprehensive understanding of the interactions between targets and their environments. By addressing these practical issues, we intend to develop robust and reliable detection systems that can be used effectively in various applications ranging from border security to wildlife conservation. In this work, we extensively discuss the points above to elucidate key factors that influence the signals generated by vibration sensors in real-time and during actual events. Our objective is to improve the practical applications of these sensors in target detection by ensuring their reliable and accurate performance across diverse scenarios. We endeavor to improve our understanding and classification of target movements by investigating the behavior of various targets and the effect of different environments on signal propagation. This comprehensive approach will enable us to refine sensor deployment strategies and signal processing techniques, ultimately minimizing false negatives and ensuring the effectiveness of vibration sensors in real-world settings.

1.1 Contribution for paper

This paper makes several key contributions to address the existing gaps in seismic sensor research. While prior studies have explored various sensor types, target detection methods, and environmental conditions, they have often overlooked the simultaneous effects of multiple environmental and target factors on signal behavior. Our research offers a more comprehensive analysis by focusing on the critical factors that influence signal variation: the target type, the target's activity, and the surface on which the target moves.

We propose a novel experimental setup that includes the deployment of eight vertical SM-24 geophones across different surface types, representing both urban and rural environments. This approach enables us to capture the diversity of signal behavior caused by the variation in surface flexibility and hardness, which directly impact wave propagation. Using a range of targets—including humans, animals, motorbikes, and cars—we ensure that the collected data reflects real-world scenarios in which these targets interact with the environment in diverse ways.

Additionally, our month-long data collection period offers a more dynamic perspective on how changes in surface conditions over time due to weather or other environmental factors can affect seismic sensor signals. This extended timeframe allowed us to account for seasonal shifts, ensuring our analysis is robust and applicable across different conditions.

Our key contributions can be summarized as follows:

(1) A comprehensive study of seismic sensor signal variation across multiple surface types and target activities, examining the influence of target behaviors (e.g., walking, running, and moving at different speeds) on the characteristics and morphology of the seismic waveforms generated upon reaching a restricted area.

(2) Deployment of eight vertical SM-24 geophones for real-time data collection under diverse conditions.

(3) Examination of the impact of environmental conditions on seismic signal shape during propagation, identifying and characterizing various conditions (e.g., asphalt, grass, muddy, and soil areas) and how they affect signal characteristics.

(4) Extended monthly data collection to factor in environmental and temporal changes affecting signal behavior.

(5) We employed machine learning for target classification to distinguish between different types of targets based on their seismic signals. These models demonstrated high accuracy in recognizing various targets and activities, underscoring the effectiveness of data-driven approaches for enhancing detection and classification in diverse environments.

The paper is structured as follows: Section 2 reviews related work on seismic sensor target detection. Section 3 details the experimental setup, including signal processing and classification. Section 4 presents results and analysis, while Section 5 provides a discussion. Section 6 concludes with key findings.

In this work, we thoroughly investigated factors that influence the signals produced by seismic sensors. We focused specifically on the target type, target activity, and the surface on which the target moves. We recognized the importance of these variables; therefore, we designed a comprehensive experiment that could fully capture their effects on the generated waveforms. Our methodology involved deploying eight vertical SM-24 geophone sensors across different surface types, namely, soil, asphalt, grass, and mud, within a farm environment to simulate urban, rural, and natural conditions. We selected various targets, including humans, animals, motorcycles, and vehicles. Each target was engaged in multiple activities to ensure a robust dataset.

001.png

Figure 1. Method used in this study

We continuously collected data over a month to enable us to account for temporal variations and environmental changes. In this manner, we could comprehensively understand how various factors affect seismic signals. Apart from data collection, we also focused on signal processing and feature extraction in the time domain to analyze the recorded waveforms. We aimed to capture the characteristics that differentiate various targets and activities by extracting relevant features from the time-domain signals. Then, the extracted features were used for classification. We employed the SVM and k-NN algorithms to classify signals based on the identified patterns shown in Figure 1. Through this approach, we could distinguish accurately among different types of targets and their activities, ultimately enhancing the reliability and accuracy of detection systems based on seismic sensors.

3.1 Description of study areas

We collected data from diverse geographical areas. Each area was strategically selected to capture various geological and environmental conditions. The study areas included rural, urban, and natural environments, as shown in Figure 2. These areas were chosen based on the specific objectives of the data collection project, with the objective of the data collection project to obtain a comprehensive understanding of the phenomena being investigated.

002.png

Figure 2. Environment of the study areas

Each geophone sensor was integrated into a node unit, which served as a data collection point. The nodes had data storage systems, communication interfaces, and power sources. Thes nodes played a pivotal role in real-time data transmission and storage. Meticulous planning and sensor deployment were crucial during data collection to ensure the accuracy and reliability of the collected information.

We strategically positioned two geophone sensor nodes (Nodes 1 and 2) in each chosen area, creating a well-structured network of data collection points. The sensors (S1, S2, S3, S4, S5, S6, S7 and S8) were carefully installed, with an interval of 2.5 m from one another, to ensure optimal area coverage. We also maintained a consistent 2.5 m gap between two parallel lines of geophone sensors to enhance our ability to detect and analyze events that may cross from one side of an area to another. Such configuration effectively divided the monitored region into segments, allowing us to capture comprehensive ground vibrations and movement data. The data collection area covered a distance of 17.5 m, as shown in Figure 3.

003.png

Figure 3. Data collection area

We chose a variety of targets, including humans, animals, motorbikes, and cars. Each target was engaged in multiple activities to ensure a robust dataset. We continuously collected data over a month, enabling us to account for temporal variations and environmental changes. This approach allowed us to comprehensively understand how the aforementioned factors affect seismic signals.

3.2 Nodes for data collection

The data collection nodes served as the backbone of the information-gathering system. They seamlessly bridged the physical aspect of geophone sensors with the digital aspect of data analysis. Each node was meticulously designed. It comprised several essential components, shown in Figure 4(a), to ensure robust data acquisition, storage, and transmission. The core of the data collection node is the Raspberry Pi 3 Model B. It is a versatile and powerful single-board computer that serves as a central processing unit. Given its compact size and energy efficiency, the Raspberry Pi 3 is ideal for field deployments, facilitating data acquisition, storage, and transmission. It works in tandem with the other node components. We integrated four ADS115 ADCs into each node to handle the analog signals from the geophone sensors.

0041.png

(a) Nodes for data collection

0042.png

(b) Structure of a node

Figure 4. Design of nodes

These high-resolution ADCs convert the analog output of the geophone sensors into digital data, which can be processed and analyzed efficiently. Through this setup, vibrations and ground movements detected by the sensors are accurately captured and digitized for further analysis. Each node has an Xbee module for wireless communication and data transfer. This module provides robust and reliable connectivity; thus, it is crucial for transmitting data from remote field locations to a central processing unit. The Xbee module is kept safely inside a protective shell to ensure durability and weather resistance. These characteristics are essential for outdoor installations. A carefully constructed wiring configuration establishes the connection between the Raspberry Pi, Xbee module, and geophone sensors. This wiring configuration enables faultless data exchange between the geophones and the Raspberry Pi, ensuring that every vibration or ground movement is captured accurately. Each node also includes resistors to protect the geophone sensors from potential damage due to electrical surges or other issues, as shown in Figure 4(b).

3.3 Multi-node geophone signal acquisition system

The data acquisition unit consists of three main components: the sensor module, the AD acquisition module, and the computer. The sensor module comprises two nodes equipped with four SM-24 geophone vibration sensors. These geophones have a natural frequency of 10±2.5% Hz, a bandwidth of 10 Hz to 240 Hz, and a sensitivity of 28.8 V/m/s. The sensor module connects to the AD acquisition module via wires.

The AD acquisition module includes three breadboards that house four ADS1115 converters, which are connected to a Raspberry Pi 3 Model B. Data is transmitted to the computer through two XBee-S2 modules: one attached to the Raspberry Pi (sender) and the other to the laptop (receiver). A laptop with a Core i7 processor running Windows 11 is used to configure the Raspberry Pi using the Imager program and the XBee-S2 modules via XCTU software.

Each node is powered by a 20,000 mAh power bank, which provides the required electrical current. Signals from the sensors are collected and synchronized by assigning unique channel addresses (48, 49, 4A, 4B) to each sensor within the node, as depicted in Figure 5. The received analog signals are converted to digital using a 16-bit ADC with a voltage range of ±0.256 V, processed with a gain of 16, and sampled at a rate of 128 samples per second. The digitized signals are then transmitted to the computer via the XBee-S2 modules, with the Raspberry Pi 3 serving as an intermediary.

005.png

Figure 5. Addresses for channels

Data were collected over one month at different times across various environments, including mud, asphalt, grass, and soil, at a distance of 17.5 meters, as shown in Figure 3. Eight sensors were deployed, with four sensors assigned to each node, spaced 2.5 meters apart, and each node collected signals simultaneously. Data collection occurred over 10 seconds each time, starting when a target entered the experimental area and continuing until it exited. Each target's activities were recorded 15 times in each environment, resulting in 960 signals per target. The experiment included four targets, each performing different activities across the four environments.

Table 2. Activities of the targets

Data collection was conducted weekly to account for environmental variations, such as changes in mud consistency, soil moisture, asphalt temperature, and water content in grass. In the first week, a person performed two activities—walking and running—while traversing different paths in the experimental area to evaluate detection variability. Signals were recorded as the person moved from the first to the eighth sensor in all four environments. During the second week, signal samples were collected from a car traveling at two speeds, 10 km/h and 20 km/h, in the four environments. In the third week, data were recorded from a motorcycle moving at two speeds, 5 km/h and 10 km/h, in the same environments. In the final week, signals were collected from an animal performing two activities—walking and running—across all environments. In total, 3840 signals were recorded, encompassing all targets, activities, and environments within the one-month experimental period shown in Table 2.

3.4 Signal processing and algorithm

By effectively distinguishing the desired signals from the surrounding noise, signal processing plays a crucial role in enhancing the accuracy of seismic sensor readings. This process involves three key steps: the removal of the direct current component, noise reduction, and normalization [11]. These steps are selected depending on the specific requirements or environmental factors that affect the sensor-generated signals. For example, normalization standardizes measurements and eliminates direct current (DC) components; hence, the accuracy of signal interpretation is improved [24]. Meanwhile, the DC component removal process addresses the inherent baseline bias in signal acquisition instruments to ensure cleaner data. In scenarios where noise originates from the target and the surrounding environment [8], such as mechanical sounds and ambient disturbances, a band-pass filter is used to mitigate these interferences effectively. To further refine signals for analysis, samples from each observation window undergo [23] column-wise normalization using an ADC, producing row vectors with a mean of zero. With this comprehensive approach to signal processing, the reliability and precision of seismic data analysis are optimized.

Our signal processing methodology, implemented in MATLAB, adapts to the varying noise characteristics of the four different environments (asphalt, mud, soil, and grass), as shown in Figure 6, which was used in the experiment.

The key focus was adjusting the amplitude threshold for noise reduction based on the specific environment, while the passband frequency of the high-pass Butterworth filter remained constant across all environments.

The high-pass Butterworth filter was designed with a fixed passband frequency of 10 Hz and a sampling frequency of 30 Hz. The normalized cutoff frequency was calculated as $f_c=$ $F_{\text {pass }} /\left(\frac{F_s}{2}\right)$, where $F_s=30 \mathrm{~Hz}$. Filtering was performed in MATLAB using the filtfilt function to avoid phase distortion, ensuring the temporal integrity of the signals. The amplitude thresholds for detecting the start and end of the signals were set according to the characteristics of each environment. For example, environments with higher noise levels (such as grass and mud) required higher amplitude thresholds to ensure that only meaningful signal data was included in the analysis. These thresholds were defined as $T_{\text {start }}$ and $T_{\text {end }}$, and the start and end indices of the detected signal were determined as $i_{\text {start }}=\min \left\{i \| x[i] \mid>T_{\text {start }}\right\}$ and $i_{\text {end }}=\max \{i \| x[i] \mid>$ $\left.T_{\text {end}}\right\}$.

006.png

Figure 6. Noise environment without targets

A binary mask $m[i]$ was then created to distinguish the signal of interest from the surrounding noise. This mask was applied to the signal, and the noise estimate was calculated by applying the high-pass filter to the masked signal, expressed as $n[i]=H_{(s)} .(x[i] . m[i])$, where $H_{(s)}$ represents the highpass filter response. Finally, the cleaned signal was derived by subtracting the noise estimate from the original signal: $m[i]=$ $x[i]-n[i]$.

Adjusting the amplitude threshold according to the environmental conditions while maintaining a constant passband frequency effectively mitigates noise without altering the detected signals within the detection range. The processed signals were then saved for further analysis.

We chose time-domain features because they directly capture the essential characteristics of the geophone signals, reflecting both amplitude and statistical properties without the complexity of transformations. This approach is ideal for distinguishing different activities and environments. The seven features—mean, standard deviation, minimum, maximum, median, skewness, and kurtosis—were selected to describe the signal comprehensively. The mean and standard deviation capture the average and variability of the signal, while the minimum and maximum values highlight extreme points. The median provides a robust central measure, and skewness and kurtosis describe the shape of the signal’s distribution, revealing asymmetry and the presence of outliers, as shown in Eq. (1).

Normalization was applied using Standard Scaler to ensure all features were on the same scale, preventing larger values from dominating the analysis. This step improves the performance of machine learning models by making all features comparable and ensuring balanced contributions from each feature.

$X=\left[\begin{array}{ccccccccccccc} { mean }_1 & { std }_1 & { min }_1 & { max }_1 & { median }_1 & { skew }_1 & { kurtosis }_1 & \ldots & { env1 }_1 & { env2 }_1 & \ldots & { act1 }_1 & { act2 }_1 \\ { mean }_2 & { std }_2 & { min }_2 & { max }_2 & { median }_2 & { skew }_2 & { kurtosis }_2 & \ldots & { env1 }_2 & { env2 } & \ldots & {act1}_2 & {act2}_2 \\ \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \vdots & \ddots & \vdots & \vdots & \ddots & \vdots & \vdots \\ { mean }_n & { std }_n & { min }_n & { max }_n & { median }_n & { skew }_n & { kurtosis }_n & \vdots & { env1 }_n & { env2 } & \vdots & { act1 }_n & { act2 }_n\end{array}\right]$       (1)

One-hot encoding was applied to the categorical variables ‘environment’ and ‘activity’, which were normalized, resulting in $\boldsymbol{X}_{ {norm }}$. Then, the dataset was divided into training and testing sets, with 80% of the data used to train the models and 20% used to test them. The SVM model was also trained and evaluated, and its mean accuracy was computed as Eq. (2):

$Accuracy_{s v c}=\frac{1}{\left|X_{\text {test }}\right|} \sum_{i=1}^{\left|X_{\text {test }}\right|} 1\left\{\hat{y}_{s v c, i}=y_{ {test }, i}\right\}$    （2）

where, $\left|\boldsymbol{X}_{\text {test }}\right|$ is the number of test samples, $\widehat{\boldsymbol{y}}_{s v c, \boldsymbol{i}}$ is the predicted labels and $\boldsymbol{y}_{\text {test }, i}$ is the actual labels. Similarly, the $\mathrm{k}-\mathrm{NN}$ model was trained and evaluated, and its mean accuracy was presented by Eq. (3):

$Accuracy_{k n n}=\frac{1}{\left|X_{\text {test }}\right|} \sum_{i=1}^{\left|X_{\text {test }}\right|} 1\left\{\hat{y}_{k n n, i}=y_{ {test }, i}\right\}$        （3）

Which measures how well the class labels of the test data are correctly predicted by the model. This variable is calculated by comparing the predicted labels with the actual labels. Finally, model performance is assessed across different data divisions by applying k-fold cross-validation. The mean accuracy for $\mathrm{k}-\mathrm{NN}$ over k folds is calculated using the following equation: Average Test Accuracy ${ }_{k n n, s v m}=$ $\frac{1}{k} \sum_j^k 1$ Accuracy $_{k n n, s v m}^j$. Accordingly, the reliability and generalization of the models are guaranteed.

Generally, researchers need to provide more detailed explanations of signals produced by seismic sensors for detecting targets. This hesitancy arises from the inherent variability in signals produced by targets during data collection. This variability can be significantly influenced by environmental factors or noise originating from the environment or the target, such as sounds or movements. Seismic sensors are susceptible to sudden events and other occurrences. This condition may considerably affect the generated signals because of their ability to detect any movements within their sensitivity range in a given environment. The residential or rural setting and the surface from which the signal is collected are crucial factors in determining signal characteristics. After acknowledging these challenges, we chose to undertake a comprehensive approach to signal collection and analysis. This approach involved collecting data from various environments, using different entities as targets, and engaging in diverse activities to fully capture the details of the signals as they are encountered. This approach aimed to provide a deep understanding of the capabilities and limitations of the sensors. Although the signals were visually similar, a meticulous analysis revealed precise detection of impact points within the required exploration time frame. The variations in signal amplitudes across different environments and detection categories could be ascribed to the flexible nature of surface materials.

In the study by Khaleghian and Taheri [14], a robot was observed across surfaces, including grass, concrete, soil, and asphalt, revealing varying radial acceleration signals, particularly below 20 Hz. The initial slip ratios also differed, with higher friction surfaces exhibiting lower ratios at equivalent speeds. Grass achieved the highest slip amplitude, followed by concrete, soil, and asphalt. In the study by DuPont et al. [13], the laser line striper, developed at Carnegie Mellon University, presented varying spatial frequency responses across surfaces mounted on a robot; grass obtained the highest magnitude, followed by sand, gravel, and asphalt. Dupont et al. [25] utilized robot vibrations to collect magnitude frequency responses by using proprioceptive sensors. Four simplifying assumptions were made: the robot’s center of gravity, body mass, motion limited to small angles, and tires that maintain ground contact. The results showed higher responses on asphalt, followed by dirt, mud, and grass.

During our experiment on human activities, running exhibited a higher signal amplitude than walking. In the study by Ghosh et al. [8], the seismic sensor detected a peak voltage of nearly two mV in human running, a significant increase compared with human walking, which was less than 5e-4 V. This finding indicated a fourfold increase in peak ground velocity during running. Furthermore, in some studies [26, 27], several gait parameters exhibited monotonic changes with escalating speed during walking and running. Such alterations included heightened step length, cycle duration, and reduced stance duration [28, 29]. This transition was marked by a sudden reduction of 35% in ground contact time and a surge of 50% in peak ground reaction force. Such abrupt changes highlighted the amplification of signals during running, as hypothesized previously.

Moreover, the difference in amplitude occurred at various car speeds. In previous research [30], realistic displacement amplitudes were obtained after applying an improved processing procedure to the processed pulse signal corresponding to a single axle load at speeds of 35 km/h and 70 km/h and vertical displacement amplitudes of 0.1, 0.3, and 1 mm. Notably, the signal amplitude at 35 km/h was higher than that at 70 km/h.

Analysis revealed some similarities between the signal patterns generated by animal and human targets, corroborating previous comparative studies [31, 32]. However, significant differences are also evident, particularly in the number of legs and the structure of feet. Animals have more feet in contact with the ground, and their hooves generate signals that differ from those produced by human feet. Although similarities exist, notable variations occur in the distances between successive detection during running, which are shorter in animals than in humans. In addition, the space between the legs is less, similar to that in walking dogs [33], although some differences occur due to the unique nature of animal hooves. This characteristic also affects the walking pattern of animals, where all four legs come in contact with the ground nearly simultaneously, with overlapping distances [34].

At a speed of 10 km/h, the wave amplitude of the car was consistently higher across all environments but was lower at approximately 20 km/h [10]. This phenomenon can potentially be explained by considering the vector amplitude of the car's speed. Our experiments verified that vector amplitude was smaller at higher speeds. This finding is aligned with our observations across various environments [30]. In some instances [35], however, the wave amplitude at the car's speed was greater than at a slower speed, contrasting with the general trend. Such inconsistency can be attributed to several factors, including the friction between the car's tires and the ground, the car's load during the experiment, and the nature of the ground. For example, the car may produce signals that differ from the expected norms when heavily loaded [14, 25]. In addition, the roughness of the tires and the speed at which the signal reflects off the ground are crucial for determining the observed wave amplitudes. Accordingly, this issue remains a subject of ongoing research. To increase the understanding of these complexities, researchers may conduct experiments under various environmental conditions using different types of tires with varying roughness levels and vehicle loads. This comprehensive approach will help clarify the interplay of factors that affect wave amplitudes in vehicle detection systems.

Moreover, such differences can be attributed to the weight disparity between cars and the size of their wheels compared with those of a motorbike. The car's larger wheels [36] resulted in more detection points, as reported by Ashhad et al. [37]. The discrepancy in detection signals between the car and the motorbike remained evident even though audio signals were employed to detect targets.

The findings presented in this study have significant practical implications for improving seismic sensor-based security systems. By capturing the variability of signals across different environments and target types, this work provides a foundation for enhancing the reliability and robustness of these systems in real-world deployments. For instance, understanding the sensitivity of seismic sensors to varying surface types and movement patterns can inform the design of algorithms that better distinguish between human, animal, and vehicle activities, thereby reducing false alarms in diverse settings. Additionally, the insights into amplitude variations with speed and surface flexibility can guide the development of adaptive thresholding techniques tailored to specific deployment environments. Despite these advancements, the study has limitations, such as the inherent dependency on environmental and surface conditions that may introduce signal noise or distortions. Future research could mitigate these challenges by employing advanced signal processing techniques and machine learning models to enhance feature extraction and classification accuracy. Moreover, conducting experiments under controlled conditions with a broader range of target types, surface materials, and environmental scenarios could provide a deeper understanding of sensor performance. This approach not only enables the development of standardized protocols for deploying seismic sensors in security applications but also ensures consistent performance across various scenarios. By establishing clear guidelines for sensor placement, signal processing, and interpretation, it enhances the scalability of these systems to cover larger areas or diverse environments effectively.