Development of intelligent algorithms for distributed fiber disturbance detection in complex environments

Abstract

The development of modern infrastructure has improved gradually with the growth of the economy. Various industries necessitate a lot of protection and require a lot of construction volume. The safety of people and property will be greatly threatened if a threat happens. Consequently, the Internet of Things (IoT) device and technology-based safety perimeter system are essential. Distributed optical fiber sensors (DOFSs) are a developing IoT infrastructure and offer a promising solution for continuous monitoring of extensive areas, but they face challenges in complex environments where ambient noise and disturbances can mask true intrusion signals. The objective of the study is to develop intelligent algorithms that enhance the detection capabilities of DOFS in complex environments, focusing on effectively separating disturbance from background noise. Data collection involved capturing six different types of disturbances such as mild touching, knocking, hitting, slapping, trampling, and occasional wagging. Then, pre-processing of input data is performed by applying z-score normalization. Feature extraction using discrete wavelet transform (DWT) for capturing time-frequency features from DOFS signals is implemented. The study proposed a novel dynamic barnacle mating algorithm-tuned long short-term memory (DBM-LSTM) method that effectively utilizes temporal pattern recognition to distinguish disturbance events. The results demonstrated accuracy in distinguishing intrusion events from ambient noise, with an improved accuracy rate of 99.73%. This indicates that the study highlights a major advancement toward the goal of superior safety system performance in complicated environments.

Keywords

distributed fiber disturbance complex environments dynamic barnacle mating algorithm-tuned long short-term memory (DBM-LSTM)fiber communication

Introduction

Perimeter security systems are considered one of the most essential technologies to ensure the safety of various sectors. Traditionally, cameras, motion detectors, and other sensors have been used for intrusion detection, which may produce limitations like the need for line of sight (LoS) and prone to false alerts. There are several methods to enhance perimeter security and one among them is the use of optical fiber-based sensors, specifically distributed acoustic sensing (DAS). Optical fiber sensors are potential for use in detecting disturbances because of their sensitivity capabilities.¹ Due to recent advancements, identification systems of intrusion based on DOFS perimeter security detection have become increasingly popular for use in various security monitoring applications. It achieves sensing signals throughout the fiber by using optical fiber as the sensing element, enabling real-time supervision of intrusions at any location and time.²

Due to the benefits of intervention resistance, excellent concealment, robust positioning accuracy, compact size, and extended detection range, DAS systems are extensively employed in applications like marine structure monitoring, railway intrusion monitoring, pipelines, traffic flow, vehicle tracking, seismic wave detection, and power cables.³ Since perimeter security is considered the first line of defense against potential attacks, it is essential to any organization’s security procedures. Accurate prediction of breaches while avoiding false alarms caused by external influences is a crucial challenge that perimeter intrusion detection systems (PIDSs) frequently confront. It is vital to employ the technical framework in conjunction with a dispersed network of fiber optic cables placed throughout the perimeter as sensors to identify vibrations and other disturbances.⁴ By measuring variations in the vibration amplitude along fiber optic cables, DAS can identify events or disruptions. Investigating the vibration strength captured by the DAS system allows one to differentiate between potentially malicious intrusions, harmless activity, and normal background noise.⁵ Different types of disturbances in DOFS are shown in Figure 1.

Figure 1.

Types of intrusions in optical sensing fibers.

Concurrently, the recent enhancements in technologies such as artificial intelligence (AI) methods enhance autonomous decision-making in real time. With the use of distributed optical fiber sensor (DOFS) technology, intrusions into the fiber infrastructure can be detected using sensing technology.⁶ It is possible to detect ambient environmental vibrations, fiber/cable strain, acoustic effects, and temperature by using the reflectance of coherent Raman, Rayleigh, and Brillouin light wave scattering methods.⁷ Certain weak, dangerous signals like those from manual digging are typically drowned out by loud background noises, especially in complex contexts where interference from these sources is inevitable. Unfortunately, these combined high interference effects may result in cases of prospective dangers being disregarded, leading to significant economic loss.⁸ In DOFS, deep learning (DL) anomaly detection plays a crucial role in vibration prediction with instantaneous and intelligent algorithms. It is a subset of deep neural network (DNN)-based ML approaches that are exposed to derive hierarchical illustration automatically from unprocessed data that is fed as input. The ability of DL algorithms to extract elements from challenging data enables more precise classification of intrusion signals.⁵

The process of analyzing the presence of anomalies and intrusions in the DOFS signals is the main motive of the present study. The study is categorized into three different methods, namely, data collection, feature extraction, and pattern recognition. Initially, the data was collected by capturing intrusions under six varying vibration disturbances. Feature extraction was performed by using discrete wavelet transform (DWT) and then pattern recognition by using the DBM-LSTM model to recognize and distinguish the intrusions effectively.

Related works

Research⁹ presents a new approach for the detection of intrusion events using recurrent neural networks (RNNs) combined with LSTM models. These methods were integrated to guide human analysts to predict possible intrusions in the network. Further, it provided enhancement in making decisions accordingly. Improved detection accuracy, generalization, and false positive rate were achieved by the model. A detection technique based on pattern separation to set up a perimeter security system in a complex environment was developed.¹⁰ Ambient and intrusion signals were separated by the method from mixed signals effectively. The ability to isolate and recover intrusion signals produced average security data retention (SDR) values and has demonstrated high performance and low response time.

A temporal 2-dimensional (2D) modeling technique used for DOFS data was implemented to capture the unique features present in the optical fiber signals.¹¹ Extensive investigations on spatiotemporal characteristics of the dispersed signals were made possible by the establishment of a representation for optical fiber signals. High-performance event classification based on optical fiber data was produced and provided improved clustering ability for varying scenarios. A DL technique that just learns typical data features from typical events was analyzed in a distributed optical fiber acoustic sensing (DAS) network.¹² Initially, DAS signal characteristics were extracted using a convolutional autoencoder and then a clustering method that finds the normal data’s feature center was also performed. After several evaluations, it was verified that the model procured improved efficacy in real-high-speed rail intrusion events. Finally, by using a shallow autoencoder, the model was able to reduce the parameters significantly.

An effective recognition framework called time-frequency features convolutional neural network (TFF-CNN) was presented to combine signal time-frequency information for a safety monitoring system.¹³ The model was based on the FFT co-generation matrix (FFT^T) and Gramian angular difference fields (GADFs) where fiber optic signal two-dimensional time-frequency image characteristics were manually extracted. Test findings showed a reduced false alarm rate, and detection response time with increased TFF-CNN accuracy. A feature-fusion-based circular dilated-convolutional block attention module-bi-directional LSTM (CDIL-CBAM-BiLSTM) network model was deployed in the study.¹⁴ It was used to deal with the issue of low prediction accuracy of signal data gathered by the DAS network. To enhance the model’s performance in the interim, a convolutional and attention block was included. Findings resulted in perimeter security settings and achieved improved identification accuracy and a shorter recognition time.

The goal of the study was to target drone intrusion in the fiber optic DAS system with a multi-source hazardous event identification approach.¹⁵ A dual-stage recognition approach was applied along with wavelet denoising to derive the improved signals from weak disturbances. Next, a hybrid model framework that effectively recognizes multi-source dangerous events targeting drone infiltration was performed and was built on CNN, LSTM, and the self-attention mechanism. An end-to-end 3-dimensional (3D) assisted CNN incorporated into the DAS system was used in the analysis performed by the author.¹⁶ To distinguish sensing targets more accurately, the approach aimed to automatically and concurrently extract three-dimensional input in terms of space, frequency, and time. As a result, the model was able to produce multi-dimensional collaboration with fast and precise recognition.

To decrease the distributional overlap between samples of known and unknown categories and increase classification resilience, the use of anchor point learning (APL) CNN was employed in the article.¹⁷ By creating an opponent with the APL, an adversarial regularization was introduced to improve the feature representation’s intra-class compactness. A method called data augmentation integration (DAI) was implemented to simultaneously deal with different typical constraints.¹⁸ The technique enhanced model generalization performance without losing recognition speed. By effectively simulating the variability of DAS signals, the model produced low events-mixing, input-shift, and noise ratio.

In research,¹⁹ a hybrid DL network based on Bayesian optimization (BO) that combined a BiLSTM with CNN was presented for multiclass identification and classification of patterns. To evaluate the efficiency, nine varying classes of sensing input were chosen as data samples and gathered from a dual Mach–Zehnder interferometer (DMZI), which was known as a perimeter security system using optical fiber. A pattern recognition that relies on Resnet 152 CNN, as well as short-time Fourier transform (STFT) was developed to deliver accurate pattern recognition and increase the stability of the optical fiber system.²⁰ Five different stages such as frequency extraction using median filter, conversion using STFT, removal of redundant information, and pattern recognition using Resnet 152 CNN model were performed. The improved performance of the study demonstrated the viability of using deep CNN (DCNN) to solve pattern recognition problems. A noise adaptive mask-masked autoencoders (NAM-MAE) method was provided, which was applied to intelligent event identification in DAS.²¹ It relied on the innovative mask mode of MAE. The outcomes showed increased convergence speed and training accuracy. To improve the recognition and performance of signal detection in real-time, a solution that relied on a generalized regression neural network (GRNN) and modified filter bank (FB) feature was developed.²² A gated perimeter sensing system experimentally obtained four types of sensing signals under three different weather situations. With an average recognition time of just 0.07 seconds, the accuracy reached up to 98.22%.

Research gaps

Various studies have explored single-source event detection, but there is a lack of comprehensive approaches that effectively integrate multi-source data in DOFS to improve detection accuracy and reduce false alarms. Despite advancements in model efficiency, many existing systems still struggle with the real-time processing of high-volume data. Research is needed to develop algorithms that can process signals in real-time without sacrificing accuracy or increasing response time, especially in dynamic environments. Investigating the adaptability of detection systems to diverse and complex environments with significant noise and interference is crucial for enhancing the robustness of DOFS technologies. While some models claim improved accuracy, the issue of high false positive rates remains prevalent. Exploring hybrid models that combine multiple learning techniques could further enhance detection reliability. With the ability to obtain an improved recognition rate, the proposed study implemented the DBM-LSTM model. This helps in enhancing the detection reliability.

Research methodology

This study analyzes DOFS signals for detecting intrusions in the network. The first step in gathering the data was recording intrusions in a range of environmental circumstances. To efficiently identify and reject the intrusions, feature extraction was carried out using WT, followed by pattern recognition using the DBM-LSTM model. The flow diagram of the proposed DBM-LSTM approach is provided in Figure 2 for better understanding.

Figure 2.

Methodology flow diagrams.

Data collection

To detect intruder occurrences, an experimental system was deployed around the building’s perimeter, and an optical fiber cable arranged. Then, using the perimeter security system, six different common events were simulated such as instances of mild touching, trampling, hitting, knocking, slapping, and occasional wagging are considered intrusions and are classified as known class events. The features of various occurrences alter in the temporal domain depending on the variation in disturbance shape and degree. The sensing signal will abruptly alter as soon as the light touches the object and then rapidly decay.

As the banging, impacting, and slapping against the fence occur, the signal’s amplitude will shift considerably. The three types of sensing signals then exhibit varying degrees of attenuation in addition to the fence’s vibration attenuation. The hitting and slapping signals attenuate the slowest, whereas the knocking signal attenuates the fastest. The sensor signal will vibrate vigorously for a long time following the trampling event, and then it will progressively diminish. Further, the sensing signal will exhibit periodic changes in the shape of a sine wave as a result of the periodic wagging event.

Pre-processing

Pre-processing is crucial before feeding input into the network to improve the robustness of the model. To convert data into a uniform scale, z-score normalization otherwise known as standardization is utilized in the pre-processing approach. DOFS data contains various features measured in different units such as amplitude, and frequency. In the presence of ambient noise and disturbances, normalizing the features helps in better distinguishing between background noise and actual intrusion signals. This can improve the effectiveness of algorithms like the proposed DBM-LSTM method. Initially, the average mean ( $μ$ ) and standard deviation (SD) ( $σ)$ values of each feature dataset are computed by using equations (1) and (2).

M e a n (μ) = \frac{1}{N} \sum_{i = 1}^{N} a_{i}

(1)

SD (σ) = \sqrt{\frac{1}{N} \sum_{i = 1}^{N} {(a_{i} - μ)}^{2}}

(2)

Here, $a_{i}$ denotes each data point, and $N$ represents the total number of observations. Further, each data points are transformed by applying z-score normalization by using the formula as shown in equation (3).

z_{i} = \frac{a_{i} - μ}{σ}

(3)

Each feature value in the input dataset is divided by SD values after the mean has been subtracted. A new dataset is created as a result, and all of the values are converted to z-scores so that the DBM-LSTM model can use them effectively. By checking, for every characteristic, the new mean should be almost 0 and the SD should be nearly one.

Feature extraction using discrete wavelet transform

Specifically, with optical fiber communications, the feature extraction stage is crucial to signal processing for several reasons. The large resources and dimensionality present in the dataset are reduced without losing any significant information, building models with less computational power and machine effort. Thus, it aids in speeding up training and learning, improving data visualization, and lowering the risk of over-fitting issues. The input is broken down into several frequency components by the DWT, which records low-frequency patterns like ambient noise as well as high-frequency transients like intrusions. Accurately identifying disturbance occurrences requires this dual capability, particularly in settings where noise can obscure intrusion signals. The current study makes use of DWT to evaluate and break down the incursion signal into its constituent characteristics. The mother function, as stated by equation (4), is the single function that DWT uses to deconstruct the signals into several functions.

φ (s) = \frac{1}{\sqrt{2}} φ (\frac{s - x}{y}) y, x \in T, y > 0

(4)

Here, the shifting and scaling parameters are represented by $x a n d y$ , with space of wavelet as $T$ . DWT examines the frequency and temporal domain properties of the signal. Thus, equation (5) deliberates the DWT.

P (y, x) = \frac{1}{\sqrt{y}} \int φ (\frac{(s - x)}{y}) d s

(5)

Here, to acquire the representation of the signal as detail ( $A$ ) and approximation ( $D$ ) coefficients, the filtered signal is decomposed into high and loss pass filters as shown in equation (6).

P (s) = \sum_{m = - \infty}^{m = + \infty} A_{k, n} ω (2^{- m} s - m) + \sum_{m = - \infty}^{m = + \infty} \sum_{m = - \infty}^{m = + \infty} 2^{\frac{- i}{2}} D_{i, n} φ (2^{- i} s - n)

(6)

Here, $ω$ is defined as a function of scale, and $m$ is the level. From the DWT coefficients, several signal and statistical characteristics are extracted. These features help in enhancing the classification and detection capabilities of intrusion events. The following features can be assessed by equations (7)–(10).

E n e r g y = \sqrt{\sum_{y = 0}^{k - 1} \sum_{x = 0}^{n - 1} t^{2} (y, x)}

(7)

S D = \sqrt{\frac{1}{N} \sum_{n = 1}^{N} {(T (n) - μ)}^{2}}

(8)

E n t r o p y = \sqrt{\sum_{y = 0}^{k - 1} \sum_{x = 0}^{n - 1} t (y, x) \log_{2} t (y, x)}

(9)

K u t o s i s (K U R) = \frac{E {(S (n) - μ)}^{4}}{σ^{4}}

(10)

Proposed hybrid DBM-LSTM model

The hybrid method integrates dynamic barnacle mating algorithm-tuned long short-term memory to enhance the performance of detecting intrusions in DOFS. This hybridization aims to leverage LSTM’s capability for capturing temporal dependencies in time-series data while utilizing DBM for optimizing LSTM’s hyperparameters. The LSTM model is used to analyze and classify intrusion signals from the DOFS. The architecture comprises input, forget, and output gates, which enable the model to maintain relevant information while discarding unnecessary data. The output of the LSTM is the predicted class label.

Hyper-parameter identification

It is crucial for LSTM performance and includes as follows. Learning rate that affects how quickly the model converges, number of units that determines the complexity of the LSTM layer, batch size influencing how many samples are processed before updating the model, total iterations over the entire training dataset, dropout rate preventing from over-fitting issues by randomly dropping a portion of neurons during training and the activation functions used within the LSTM gates. Then the DBM is employed to fine-tune the above-mentioned hyper-parameters dynamically by involving,

Initialization

The barnacle population represents hyperparameter combinations.

Selection

Barnacles are evaluated based on their fitness, which is typically defined by a validation metric derived from the LSTM model’s performance.

Reproduction

Successful barnacles reproduce to create new candidate solutions. This mimics the biological mating process, allowing the algorithm to explore and exploit effective hyperparameter values.

Iterative improvement

The DBM iteratively updates the hyper-parameters based on their performance until a stopping criterion is met. This dynamic approach enables the DBM-LSTM model to adaptively find the best parameters suited for the specific intrusion detection task.

Long short-term memory model (LSTM)

Unlike RNN, LSTM can remember data for extended periods and is meant to avoid long-term reliance issues. The LSTM design is more sophisticated, consisting of four layers, whereas the RNN architecture’s hidden layers have a straightforward structure. The cell state is the main building block of LSTM and the gates use the sigmoid function to add or delete data from the cell state. Figure 3 shows the model of the LSTM structure with three gates.

Figure 3.

LSTM model.

Input or update gate

The data that is stored in the cell state by the LSTM is chosen by this input gate. A sigmoid function is used by the input gate layer to select which data will be updated when a $\tan h$ layer applies a new vector to be added to the cell state. Using the newly determined vector values, the LSTM tends to update the cell state by disregarding previously chosen data. The input and cell state are recomputed by using equations (11) and (12).

I_{g} = σ (T_{I} \cdot [H_{g - 1}, z_{g}]) + a_{I}

(11)

\hat{C_{g}} = \tan h (T_{C} \cdot [H_{g - 1}, z_{g}]) + a_{C}

(12)

Forget gate

It uses a sigmoid function to choose which information from the cell state the LSTM will remove after examining the input data and the data from the preceding concealed layer. Here, 1 means maintains it, 0 means delete it. It is expressed by using equation (13).

F_{g} = σ (T_{F} \cdot [H_{g - 1}, z_{g}] + a_{F})

(13)

Output gate

It uses a sigmoid function to determine the output, and predicts which segment of cell LSTM will produce output. The result is further transferred using a layer called $\tan h$ , which outputs only the data that is preferred to send into the following neuron. It is evaluated by using equation (14).

O_{g} = σ (T_{O} \cdot [H_{g - 1}, z_{g}] + a_{O})

(14)

Here, $T_{I}$ , $T_{C}$ , $T_{F}$ , ${a n d T}_{O}$ , are represented as the weights of each layer and the biases are $a_{I}$ , $a_{C}$ , $a_{F}$ , and $a_{n d a O}$ . In this context of the classification of intrusions in DOFS, the input and output variables are given by (X) and (Y). LSTM model is applied, in which the mapping function is learned from input to the output, $Y = f (X)$ . The output variables are predicted when the new input data is applied and is given by equation (15).

{\begin{cases} A = [\begin{array}{l} A_{1} \\ A_{2} \\ A_{3} \\ \begin{array}{l} . \\ . \\ A_{i} \end{array} \end{array}] = [\begin{array}{l} a_{11} \\ a_{21} \\ a_{31} \\ \begin{array}{l} . \\ . \\ a_{i 1} \end{array} \end{array} \begin{array}{l} a_{12} \\ a_{22} \\ a_{32} \\ \begin{array}{l} . \\ . \\ a_{i 2} \end{array} \end{array} \begin{array}{l} a_{13} \\ a_{23} \\ a_{33} \\ \begin{array}{l} . \\ . \\ a_{i 3} \end{array} \end{array} \begin{array}{l} \dots \\ \dots \\ \dots \\ \begin{array}{l} . \\ . \\ \dots \end{array} \end{array} \begin{array}{l} a_{1 n} \\ a_{2 n} \\ a_{3 n} \\ \begin{array}{l} . \\ . \\ a_{i n} \end{array} \end{array}] \\ B = [\begin{array}{l} b_{1} \\ b_{2} \\ b_{3} \\ \begin{array}{l} . \\ . \\ b_{i} \end{array} \end{array}] \end{cases}

(15)

Here, the input sequence is given by $A = (A_{1}, A_{2}, A_{3}, \dots, A_{i})$ and $A_{j} = (a_{i 1}, a_{i 2}, a_{i 3}, . . a_{i n})$ , is denoted as the element of input variables and thus the output sequence $B = (b_{1}, b_{2}, b_{3}, \dots . b_{i})$ is obtained.

Dynamic barnacle mating (DBM) optimization algorithm

Microorganisms known as barnacles cling to objects that are present mostly in the water. Their most notable feature is their long penis. Every competitor and neighbor within reach of their penis is included in their mating group. The mating behavior of barnacles served as the model for the barnacle mating optimizer. The practical optimization problem is solved by simulating three processes, namely, initialization, selection, and reproduction processes. The process flow involved in the DBM algorithm is illustrated in Figure 4.

Figure 4.

Process flow diagram for the proposed DBM algorithm.

Primary individuals

Initially, it is assumed that barnacles are the candidate solution, and that equation (16) can be used to represent the population matrix.

Y = [\begin{array}{l} y_{1}^{1} & \dots & y_{1}^{n} \\ ⋮ & ⋱ & ⋮ \\ y_{N}^{1} & \dots & y_{N}^{n} \end{array}]

(16)

Here, n denotes the barnacle’s population, while the decision variable is $N$ with upper and lower limits as shown in equations (17) and (18):

L_{b} = [L_{b}^{1}, \dots, L_{b}^{j}]

(17)

U_{b} = [U_{b}^{1}, \dots, U_{b}^{j}]

(18)

Choice method

The selection of reproducing barnacles is contingent upon the penis’s length ( $p_{l})$ . Exploration and exploitation are both present in DBM. Sperm casting is still being investigated as a means of reproduction. The following formulas (19) and (20) were used to construct this method with mate parents $B a r n a c l e_{b}$ and $B a r n a c l e_{m}$ .

B a r n a c l e_{b} = r a n d (N)

(19)

B a r n a c l e_{p} = r a n d (N)

(20)

Mating

Based on the Hardy–Weinberg principle, the DBM algorithm considers the frequency produced by the parental genotypes present in the family as well as inherited qualities. Thus, the novel reproduction is acquired by taking into account the subsequent equation (21).

Y_{j}^{N_{n e w}} = m Y_{b a r n a c l e_{b}}^{N} + {n Y}_{b a r n a c l e_{p}}^{N}

(21)

Here, the pseudo-random on average spread value is provided by $m$ ranging between 0 and 1, where $n = (1 - m)$ . However, the male and female barnacle variables are defined by $Y_{b a r n a c l e_{b}}^{N}$ and $Y_{b a r n a c l e_{p}}^{N}$ . When the barnacle chooses to mate with some other than the $p_{l}$ value that is first determined, the sperm is directed as the algorithm exploration term to occur. It is acquired by applying equation (22).

Y_{j}^{N_{n e w}} = r a n d \times Y_{b a r n a c l e_{p}}^{n}

(22)

Here, the random number ranging between 0 and 1 is denoted by $r a n d$ . Thus, for the exploration phase, the new reproduction is developed by the female barnacle. The offspring are evaluated and merged with the parents to study the solution matrix based on the size of the population. Subsequently, the sorting procedure is used to eliminate the subpar outcomes and select the good solutions that supply the population size. The steps involved in the proposed DBM-LSTM model are provided in Algorithm 1.

Results and discussion

The study involves performance analysis and the results are delivered in this section. To ensure the effectiveness of the suggested model, existing approaches that involve disturbance detection for DOFS in complex environments are used for comparison. The results of the proposed model are compared and demonstrated with enhanced performance.

Experimental configuration

The entire experimentation was conducted by using a Windows 10-assisted computer powered by an Intel Core i7-7500 U CPU. It is operated at a base clock speed of 2.70 GHz, with 16 GB of RAM. The major programming language was Python which is used for conducting the test analysis.

Evaluation metrics

The study uses specific performance metrics that include precision, F1-score, recall, and accuracy in classifying vibration behaviors in DOFS. This supports in validating the efficiency of the DBM-LSTM model.

Accuracy

It is defined by representing the overall suitability in making predictions by the model. It assesses the model on how often the detection matches the actual class labels in the dataset and is given by equation (23).

A c c u r a c y = \frac{T P + T N}{T P + T N + F P + F N}

(23)

Here, the true positive and true negative are denoted by $T P a n d T N$ , while false positive and false negative are given by $F P a n d F N$ . A higher accuracy indicates a better-performing model.

Precision

It computes the quality of the positive predictions made by the model, which enumerates the number of instances classified as positive that are positive. This can be expressed by formula (24).

P r e c i s i o n = \frac{T P}{T P + F P}

(24)

High precision denotes that the suggested model determines an intrusion which is likely to be accurate.

Recall

It computes the capacity of the model to identify all appropriate instances. In addition, the proportion of definite positive situations that were correctly predicted is also determined denoted as shown in equation (25).

R e c a l l = \frac{T P}{T P + F N}

(25)

The high recall involves the successful prediction of the actual disturbances. This is important in this context where missing disturbances could damage security measures.

F1-score

It balances precision along with recall, thus delivering a single score. It is denoted by equation (26).

F 1 - s c o r e = 2 \times \frac{P r e c i s i o n \times R e c a l l}{P r e c i s i o n + R e c a l l}

(26)

When the F1-score is high, then it indicates that both precision and recall are reasonably high, making it a robust measure of the model’s performance in detecting disturbances.

Performance analysis

The performance of the proposed DBM-LSTM model is analyzed by implementing several performance metrics and the results are portrayed in Table 1 with Figures 5–7 illustrating the values procured by the model.

Table 1.

Recognition results of the proposed DBM-LSTM method.

Features	F1-score	Precision	Recall
Mild touching	97.24	98.55	98.12
Knocking	97.60	98.67	98.36
Hitting	98.15	99.11	98.74
Slapping	97.31	99.35	97.28
Trampling	98.52	97.12	98.60
Occasional wagging	98.63	98.30	97.24

Figure 5.

Performance values based on precision.

Figure 6.

Comparison based on F1-score values.

Figure 7.

Performance analysis based on recall values.

Six types of disturbances were used in the analysis such as mild touching, knocking, hitting, slapping, trampling, and occasional wagging. Based on these features, the F1-score, precision, accuracy, and recall were evaluated. From Table 1, it is understood that the proposed DBM-LSTM model demonstrates consistently high performance across all evaluated features, indicating the model’s robustness. However, it is essential to consider the implications of the observed high false positive rates for certain features, specifically slapping and trampling, which reached 97.31% and 98.52% precision, respectively. Although these rates suggest strong accuracy in many instances, high false positives can compromise the operational viability of the proposed system, particularly in security-sensitive applications where false alarms may lead to unnecessary responses, resource allocation, and potential operational fatigue. Therefore, optimizing the model to reduce false positive rates without sacrificing detection accuracy is crucial for practical implementations in real-world scenarios.

Precision, recall, and F1-scores are also notably high, indicating that the system is effective in accurately predicting and detecting various intrusion-related activities. Further, a comparative analysis of existing methods and proposed is performed to verify the efficacy of the present study. Conventional methods such as Multiscale CNN MSCNN,²³ Gramian angular summation field-CNN model ConvNeXt tiny network (GASF-ConvNeXt-TF),²⁴ recurrent plot (RP), and Inception-v3 network³ were used in the analysis. The accuracy rate of predicting the disturbances is tabulated in Table 2 Figure 8 shows the comparison based on prediction accuracy.

Table 2.

Accuracy values obtained by proposed and existing methods.

Approaches	Prediction accuracy (%)
MSCNN²³	98.67
GASF-ConvNeXt-TF²⁴	99.30
RP and Inception-v3³	99.70
DBM-LSTM [proposed]	99.73

Figure 8.

Comparison of accuracy values.

The MSCNN²³ model achieved 98.67% prediction accuracy. However, it shows that while the method is reliable, there is room for improvement compared to more advanced approaches. Another approach of GASF-ConvNeXt²⁴ acquired a prediction accuracy of 99.30%, this method demonstrates a significant improvement over MSCNN. The RP and Inception-v3³ network produced a prediction accuracy of 99.70% indicating a high performance in detecting intrusions, showing that this combination of techniques effectively captures critical features in the data. The proposed DBM-LSTM hybrid model learns the temporal dependencies and makes accurate predictions. It achieved a prediction accuracy of 99.73%, which is slightly higher than the RP and Inception-v3 method. This indicates that the proposed model not only matches but outperforms existing techniques in the context of disturbance detection in DOFS.

Discussion

The metrics presented for various disturbance detection approaches underscore significant advancements in the effectiveness of intelligent learning models applied to DOFS technologies. Nevertheless, it is critical to acknowledge the potential limitations of the DBM-LSTM model. Firstly, the model’s adaptability to diverse environmental conditions, such as extreme temperatures or heavy ambient noise, can affect its performance and reliability. Real-world deployment scenarios often present challenges that may not be fully simulated during training phases. Additionally, the integration of the proposed system with existing security frameworks poses considerations of compatibility and data interoperability. Future research should investigate these aspects to ensure a more comprehensive understanding of the practical utility and robustness of our approach in diverse and complex operational environments.

Six different disturbances are involved in the study analysis and performance is validated based on F1-score, recall, and precision. The slightly lower precision and recall values for some features like slapping and trampling suggest room for improvement, but overall the method performs well in distinguishing these different event types. A comparison of existing models with the proposed DBM-LSTM is performed. MSCNN²³ captures features at multiple scales, enhancing its ability to recognize patterns in complex data. GASF-ConvNeXt²⁴ involves the conversion of time-series data into two-dimensional images, which are then processed by the ConvNeXt architecture for improved performance. The RP³ model indulges the dimensionality reduction process with the Inception-v3 network in image classification tasks. The proposed DBM-LSTM approach integrates the strengths of DBM for unsupervised feature extraction and LSTM networks for sequence prediction. These findings not only emphasize the importance of security technologies but also suggest significant implications for their real-world applications.

Conclusion

This study has explored the potential of DOFS technology for real-time monitoring of intrusions in extensive areas. The research addresses the inherent challenges posed by complex environments, where ambient noise and disturbances can obscure true intrusion signals. Through the innovative application of techniques, the study facilitates the differentiation of intrusions from background noise in complex environments. Initially, the data were collected that comprises of capturing intrusions by considering six different vibration responses. Then, pre-processing of input data is performed by applying z-score normalization. The pre-processed data are then exposed to feature extraction using the DWT process. Furthermore, the recognition of intrusions is analyzed and predicted by using the proposed DBM-LSTM model. The findings of the study showed that the proposed model exhibited a 99.73% accuracy rate indicating that the DBM-LSTM method not only enhances the accuracy of intrusion detection but also significantly improves detection speed while reducing false positive rates. Though the study has produced better prediction results, it can be further enhanced by increasing the accuracy rates. Future research can build upon these findings to further refine detection capabilities and explore additional applications of IoT technologies in safety and security domains.

Statements and declarations

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This research was funded by the Science and Technology Project of State Grid Qinghai Electric Power Company (No. 522800230003).

ORCID iD

Linhui Yang

References

Vikram

Patel

Alsalman

. Measurement of optical fiber sensors for intrusion detection and warning systems fortified with intelligent false alarm suppression. Opt Quant Electron 2024; 56(6): 939.

Dejdar

Záviška

Valach

, et al. Image edge detection methods in perimeter security systems using distributed fiber optical sensing. Sensors 2022; 22(12): 4573.

Lyu

Jiang

, et al. Abnormal events detection based on RP and inception network using distributed optical fiber perimeter system. Opt Laser Eng 2021; 137: 106377.

Dong

Zhao

Huang

, et al. Distributed fiber optic acoustic sensing system intrusion full event recognition based on 1-D MFEWnet. Phys Scr 2024; 99(4): 045506.

Vikram

Patel

Chaturvedi

, et al. Detecting accurate parametric intrusions using optical fiber sensors for long-distance data communication system. Opt Fiber Technol 2023; 80: 103453.

Han

Huang

, et al. Deep learning-based intrusion detection and impulsive event classification for distributed acoustic sensing across telecom networks. J Lightwave Technol 2024; 42: 4167–4176.

Wellbrock

Xia

Huang

, et al. Explore benefits of distributed fiber optic sensing for optical network service providers. J Lightwave Technol 2023; 41(12): 3758–3766.

Zhang

, et al. An efficient separation and identification algorithm for mixed threatening events applied in fiber-optic distributed acoustic sensor. IEEE Sens J 2023; 23: 24763–24771.

Muthunambu

Prabakaran

PrabhuKavin

, et al. A novel eccentric intrusion detection model based on recurrent neural networks with leveraging LSTM. Comput Mater Continua (CMC) 2024; 78(3): 3089–3127.

10.

Lyu

Niu

Tian

, et al. Identification of intrusion events based on distributed optical fiber sensing in complex environment. IEEE Internet Things J 2022; 9(23): 24212–24220.

11.

Gong

Yang

Zhang

, et al. A pipeline intrusion detection method based on temporal modeling and hierarchical classification in optical fiber sensing. IEEE Sens J 2024; 24: 19327–19335.

12.

Xie

Wang

Zhong

, et al. Label-free anomaly detection using distributed optical fiber acoustic sensing. Sensors 2023; 23(8): 4094.

13.

Qiu

Karimi

, et al. TFF-CNN: distributed optical fiber sensing intrusion detection framework based on two-dimensional multi-features. Neurocomputing 2024; 564: 126959.

14.

Wang

Zhang

, et al. Feature fusion-based fiber-optic distributed acoustic sensing signal identification method. Meas Sci Technol 2023; 34(12): 125141.

15.

Zhou

, et al. Multi-source threatening event recognition scheme targeting drone intrusion in the fiber optic DAS system. IEEE Sens J 2024; 24: 32185–32195.

16.

Liu

Wang

, et al. Multi-dimensional information extraction and utilization in smart fiber-optic distributed acoustic sensor (sDAS). J Lightwave Technol 2024; 42(19): 6967–6980.

17.

Jiao

Gupta

, et al. Open set intrusion event recognition using anchor point learning for distributed optical fiber system. IEEE Trans Instrum Meas 2024; 73: 1–13.

18.

Zhang

, et al. Modified data augmentation integration method for robust intrusion events recognition with fiber optic DAS system. J Lightwave Technol 2023; 42: 453–462.

19.

Sun

Yang

Fang

, et al. On Bayesian optimization-based CNN-BiLSTM network for multi-class classification in distributed optical fiber vibration sensing systems. IEEE Trans Instrum Meas 2024; 73: 1–14.

20.

Jin

Liu

Jiang

, et al. Pattern recognition of distributed optical fiber vibration sensors based on resnet 152. IEEE Sens J 2023; 23: 19717–19725.

21.

Sun

Wang

, et al. Intelligent water perimeter security event recognition based on NAM-MAE and distributed optic fiber acoustic sensing system. Opt Express 2023; 31(22): 37058–37073.

22.

Fang

Wang

. Signal identification based on modified filter bank feature and generalized regression neural network for optical fiber perimeter sensing. Opt Fiber Technol 2022; 72: 102993.

23.

Lyu

Huo

Liu

, et al. Robust intrusion events recognition methodology for distributed optical fiber sensing perimeter security system. IEEE Trans Instrum Meas 2020; 70: 1–9.

24.

Wang

Zhuo

Liu

, et al. GASF-ConvNeXt-TF algorithm for perimeter security disturbance identification based on distributed optical fiber sensing system. IEEE Internet Things J 2024; 11: 17712–17726.