Open Access
Add to BibSonomy
Abstract
This research addressed the drawbacks of the conventional hybrid structure and processing technique by presenting a novel distributed fiber optic sensor based on a hybrid Michelson and Mach-Zehnder interferometer. The sensor can achieve blind spot free positioning and has a wide response frequency, additionally its structure is not complex. It can obtain two phase signals from each of the two interferometers by using a demodulation method that uses a 3 × 3 optical coupler. To determine the position of the disturbance, we computed cross-correlations on the two signals following basic mathematical techniques. Markov Transition Field was used to transform the phase signals—which had been filtered by a band pass filter—into two-dimensional images. Tagged photos built a dataset, which is then fed into a neural network to identify patterns. Experiments have shown that the frequency response capacity of the structure was verified, and it was able to achieve location within 0-30 km with location errors of ±85 m. In a six-category pattern recognition, the testing set accuracy was 98.74%.
© 2024 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Distributed fiber-optical sensors have numerous advantages such as large-scale monitoring, high sensitivity, and high electromagnetic resistance [1]. They are widely used in pipeline monitoring, perimeter security, and bridge engineering [2–4]. Distributed fiber optic sensors mainly can be divided into two categories, one is optical time domain reflectometry (OTDR), which already has mature product offerings [5], these products have some advantages such as higher sensitivity and response speed, but these products often suffer from the following limitations: 1. Complex system structure, the structure also leads to high costs. 2. Resolution limitation, The resolution of an OTDR is constrained by the pulse width and sampling rate, thereby impeding precise detection and localization of closely spaced events [6].
The other is mainly based on interferometric methods which offers advantages such as low cost, and simple structure. The widely used interferometric structures for establishing distributed fiber-optical sensor monitoring systems include Mach-Zehnder interferometer (MZI), Michaelson interferometer (MI) and Signac interferometer (SI). Over the past decades, various types of sensing structure has been proposed, such as dual-MI [7], dual-MZI [8] or dual-SI [9,10]. Among these, research on hybrid structures have gained popularity, but these structures have some limitations. The MI-SI and dual-SI structures cannot detect interference signals in the center of the of the optical path, and sensing system is insensitive to the low frequency signals [11–14], both dual-SI and dual-MI structures have frequency responses that do not exhibit a no-blind zone [15]. However, conventional structures need two minus processing algorithm that may influence the efficiency of the localization. Distributed fiber-optic sensor based on merged MI and MZI has the possibility to solve the problem.
The phase signal obtained after subtraction is independent of the disturbance position and solely related to the disturbance signal itself. Additionally, the hybrid MI-MZI structure has consistent frequency response to the disturbance signal at any position. Machine learning-based pattern recognition for one-dimensional signals has seen advancements in recent years [16,17], but it is very difficult to build predictive models for one-dimensional time series signals for many reasons including One-dimensional signals often lack contextual information, which can hinder accurate understanding and interpretation of the signals in certain scenarios. What is more, it may lose other relevant information within them [18]. Markov Transition Field (MTF) addresses this challenge by transforming one-dimensional time series into two-dimensional images. Combining MTF with a convolutional neural network (CNN) offers a novel approach to pattern recognition [19].
In this paper, we achieved location accuracy of ±85 m using this structure and algorithm. The algorithm employed only requires one subtraction process, simplifying the localization process. To facilitate pattern recognition, we collected different types of disturbance signals at various locations along the sensing fiber. These one-dimensional signals were then transformed into 2D images using the MTF. By constructing a 2D convolutional neural network (CNN) and training the model using these 2D images, we achieved a recognition accuracy of 98.74% for 6 categories. These results demonstrate the potential for realizing high-precision pattern recognition applications using our approach.
2. Experiment setup and principle
A. System configuration and principle of operation
The MI-MZI structure used in this study is illustrated in Fig. 1. The structure consists of the following components: Two distributed feedback (DFB) lasers with different center wavelengths, serving as the light sources (S1 and S2). Four wavelength division multiplexers (WDM1, WDM2, WDM3, WDM4) for the coupling and separation of light with different wavelengths. Two 3 × 3 fiber couplers (C1 and C2) with a beam splitting ratio of 1:1:1. Four photodetectors (PD1, PD2, PD3, PD4) for photoelectric conversion. Faraday rotating mirrors (FRM1, FRM2) located at the end of the arms, which reflect the light beams from S1. The light from the two sources is input into the two 3 × 3 couplers separately, and they travel along their respective arms after beam splitting. The light beams from S1 are reflected by the Faraday rotating mirrors and return along the original paths, eventually meeting at C1. The light beams from S2 travel along their paths and meet at C1. After the formation of stable interference through the 3 × 3 couplers (C1), the interference signals are collected by the four photodetectors. These analog signals are then converted to digital signals by using a data acquisition card (DAQ). Finally, the signals are processed using a computer for further data analysis.
Fig. 1. Schematic diagram of MI-MZI interference and transmission paths of light with different wavelength (orange lines, λ2; blue lines λ1).
According to Fig. 1, the transmission paths of the MI are obtained by the following two paths:
Path1(λ2):
S1 → C1 →WDM3→ FRM1 → C1 → WDM1 → PD1
Path2(λ2):
S1 → C1 →WDM4→ FRM2 → C1 → WDM2 → PD3
The interference signal of MZI is obtained by the following two paths:
Path3(λ1):
S2 → C2 → WDM3 → C1 → WDM1 → PD2
Path4(λ1):
S2 → C2 → WDM4 → C1 → WDM2 → PD4
Both interfering lights are combined at C1 in the MI-MZI structure. Unlike conventional 2 × 2 couplers, where the initial phase of the two interfering lights is fixed at 0 and π, the initial operating state of the sensor is highly insensitive. As a result, an artificial bias signal is typically required to achieve proper interference. However, in this new structure, the ±$\frac{{2\pi }}{3}$ phase difference introduced by C1 provides the necessary passive bias for interferometry without the need for additional phase or frequency modulation in the optical path. The 3 × 3 coupler (C1) also plays a crucial role in the subsequent demodulation algorithm, enabling accurate detection and extraction of the interference signals [20].
B. Algorithm of location based on demodulated optical phase
For the MI, the optical signal detected by the two detectors can be characterized as:
where ${\varphi _{01}}$ and ${\varphi _{02}}$ are the phase difference (normally $\frac{{2\pi }}{3}$) that come with the C1 and C2 [21], ${\varphi _{MI}}(t )$ is the phase different due to the photo elastic effect and length change resulting from the disturbance, A1 and A2 are the amplitude factor determined by the light intensity and detector does not affect the demodulation. Similarly, for the MZI, the optical signal detected can be characterized as:We use the variable $\varphi (t )$ to represent the phase change caused by external disturbances. To demodulate and analyze the phase signals, we employ the inverse triangle-based demodulation algorithm [22]. we can get the phase difference under two structures as:
The phase changes ${\varphi _{MI}}(t )$ ($\mathrm{\Delta }{\varphi _1}(t )$) and ${\varphi _{MZI}}(t )$ ($\mathrm{\Delta }{\varphi _2}(t )$) generated by the same disturbance are processed to obtain $\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta }{\phi _2}(t )$ with a fixed time delay, and the algorithm only need once minus process. The principle of the signal processing algorithm is shown in Fig. 2. Signals $\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta }{\phi _2}(t )$ can be obtained from Eq. (5) and (6) as follows:
$\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta }{\phi _2}(t )$ contain $2{\tau _0} - 2{\tau _x}$ with embedded time information, it can be demonstrated as:
where Lx is the location of the disturbance point. The cross-correlation function of $\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta }{\phi _2}(t )$ can be calculated by the following equation:According to the cross-correlation theory, the cross-correlation function $R(\tau )$ maximizes when and only when $\tau = \varDelta \tau $. Therefore, by finding the value of the cross-correlation $R(\tau )$, the location of disturbance point Lx can thus be calculated by Eq. (11):
$\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta }{\phi _2}(t )$ is not affected by the time delay itself; ${\tau _0}$ is a fixed value, τx is calculable value. This characteristic implies that the phase signal effectively captures the complete information of the disturbance signal, independent of its position along the sensing fiber. our study is influenced by two main factors: the sampling rate of the acquisition card and the level of noise present in the data. To estimate the correlation between two discrete time series, we use the cross-correlation algorithm, which provides a discrete cross-correlation sequence. The maximum correlation coefficient and its deviation from the central position indicate the extent of correlation. The resolution at a sampling rate of 500 k samples per second and a sampling interval of 2 microseconds is approximately 200 m without using a fitting algorithm. Theoretically, by using quadratic term interpolation fitting, we can further increase the resolution to 2 m [23]. However, achieving this level of accuracy is challenging in practice due to noise interference. To mitigate these effects, band-pass filtering was applied to enhance the accuracy of location estimation.
C. Frequency response
The frequency response is a critical factor in evaluating a sensing system. Disturbance signals often comprise a wide range of frequencies, including both high-frequency and low-frequency components. It is crucial for a sensor system to have the capability to accurately detect and measure these complex frequencies. If a sensor system lacks the ability to capture certain frequencies within the disturbance signal, it can lead to a loss of sensing accuracy. This limitation occurs when the system fails to effectively detect and capture the desired frequency components, resulting in incomplete or inaccurate measurements. Therefore, a comprehensive frequency response that encompasses a broad range of frequencies is essential for maintaining high sensing accuracy.
We assume that the modulation signal modulates the phase as A$cos ({\omega t} )$. According to Eq. (5) and (6), the phase difference signals for MI and MZI can be obtained as:
According to the above equation, the descriptive equation of the frequency response of MZI and MI can be obtained as:
where L0 is the length of the sensing arm, Lx is the location of the disturbance point, it can be any value, according the Eq. (14), we can find some $\omega $ such that $cos \frac{{\omega n({{L_0} - {L_x}} )}}{c}$ is 0. For example, when the frequency is about 2000Hz, and $\mathrm{\Delta Location\;\ }$(${L_0} - {L_x}$) is 25 km, the value of the ${\phi _{MI}}(t )$ will be almost zero. Therefore, if you only use MI, it will have blind zone location. For MZI, its frequency response is independent of the $\omega $ and L0 and Lx. Therefore, MZI has the capability of Full Frequency Response. Figure 3(a) and (b) shows the frequency response of MI and MZI.Fig. 3. Schematic diagram of frequency response theory of different structure: (a) frequency of MZI and (b) frequency response of MI; (c) frequency response of MI-MZI. We suppose A = 1.
According to Eq. (7) and (8), the frequency response of the MI-MZI structure can be deduced as:
Its theoretical frequency response is shown in Fig. 3(c).
Based on the characteristics of the MI-MZI structure, it can be observed that the frequency response is independent of the position and exhibits stable response for different frequencies. This unique feature not only broadens the frequency response range but also eliminates the issue of insensitivity to specific frequencies at certain positions. As mentioned earlier, the limitations of dual-SI, conventional MI-SI, and dual-MI structures in achieving accurate location at small Lx values can be attributed to the poor frequency response of the MI and SI components. In contrast, the wide and flat frequency response of the sensor system used in this study provides theoretical support for addressing this problem. The sensor system's ability to maintain a consistent and reliable frequency response across different positions enhances its capacity to accurately detect and measure disturbances at various frequencies.
D. Markov transition field.
MTF can transform a 1D time series into a 2D image to take advantage of computer vision. Take a 1D time series as an example:
First, we divide the time series X (t) into Q quantile boxes (labeled as 1, 2…, Q, with the same amount of data in each quantile box). For example, if we use the quartile, we will put all the values in the quantile bucket to which they belong, 25%, 50%, 75%, 100%. This is somewhat similar to the bin value in a histogram. We can imagine each bucket as a state in a Markov model. Then Change each data in the time series to its corresponding quantile box serial number.
Second, we construct a Markov state transition matrix:
Remember that ${\omega _{ij}}$ here represents the transition probability from state i to state j. If we use Q quantile, then this matrix is Q.
In most cases, the maximum likelihood method is employed to estimate the transition probability. In this method, ${\omega _{ij}}$ can be normalized by dividing the count from state i to j by the total count from state i or by the count matrix. However, it is worth noting that the Markov state transition matrix derived from the original time series data is not very sensitive to the distribution of the original data and tends to lose temporal information. This can be seen as a limitation, as retaining temporal information is crucial for accurate analysis. To address this limitation, the Markov Transition Field (MTF) technique was developed.
MTF, also known as Markov transformation field, denoted as M, is an N × N matrix, where N is the timing length.
MTF represents the temporal relationship between data at different time points, providing insights into state adjacency. It offers several advantages. Firstly, MTF describes the dynamic evolution process of complex systems, capturing both static characteristics and dynamic laws. This makes it advantageous for studying complex dynamic systems. Secondly, MTF improves prediction accuracy by transforming time series data into probability distributions and utilizing their properties for prediction. Lastly, MTF enables the conversion of one-dimensional series into 2D images, which can contain more informative patterns crucial for pattern recognition. In contrast, the original MI-MZI signal only represents light intensity, lacking a linear relationship with disturbance signals. To address this challenge, phase signals ($\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta }{\phi _2}(t )$) are utilized for MTF analysis due to their close relation with disturbance signal. Bandpass filtering is applied to the phase-restored signal to mitigate signal drift and reduce noise-induced unreliability of disturbance signals.
E. Convolutional neutral network.
CNN is a deep learning architecture specifically designed for image analysis and pattern recognition tasks. One of the advantages of CNN is its ability to automatically learn hierarchical representations from raw image data, along with its local receptive fields and weight sharing mechanisms. It comprises multiple layers, including convolutional layers, pooling layers, and fully connected layers. CNN has been highly successful in various applications, particularly in computer vision tasks such as image classification.
While it is theoretically known that the CNN's performance improves with increasing network depth, when using MTF-generated images as input, the complexity of the extracted information is not high, therefore, a relatively simple CNN structure can still achieve a satisfactory level of accuracy. Figure 4 illustrates the CNN structure constructed in this paper for our study.
Figure 5 shows the flow of pattern recognition. The overall process involves obtaining the phase signal for localization, followed by noise reduction to obtain the MTF image for pattern recognition.
From the previous section, it can be seen that MI-MZI has some advantages for pattern recognition. The advantages can be reflected in the following three point: 1) From Eq. (7), $\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta }{\phi _2}(t )$ contain the information of disturbance signal, They can directly represent the disturbance signals. 2) The MI-MZI structure can obtain phase signals with wide and stable frequency response, therefore, the phase signal does not lose critical information. 3) The direct phase modulation captured by fiber optic vibration demonstrates a robust correlation with physical phenomena, which is shown in Fig. 12.
3. Experiment of no-blind zone location
As shown in Fig. 1, an experiment was set up with a fiber length of 30 km, utilizing a 2 mW DFB (S1) light source with a central wavelength of 1550 nm and a 2 mW DFB (S2) light source with a central wavelength of 1548 nm. The light generated by the S1 can pass completely through the passport of the WDM with a central wavelength of 1550.12 nm (passband ±0.37 nm) and the light generated by the S2 can pass completely through the passport of the WDM with a central wavelength of 1548.36 (passband ±0.37 nm). The A/D conversion is achieved by a 32-bit DAQ with a data sampling rate of 500 KS/s. The length difference between of the two sensing arms is within ±10 cm. The data processing of the interference signals is conducted on LabVIEW.
A. Frequency experiments
To simulate disturbance, PZT was introduced at Lx = 200 m and modulated using a signal generator. The modulator voltage was set to 100 mv and 300 mv, and the frequency response in the range of 10∼10000 Hz was measured at 100 Hz intervals. The average of 100 measurements was taken, and the same test was conducted at 10 km. The resulting frequency response curves are shown in Fig. 6. Despite some fluctuations at the same location due to external environmental noise, these can be considered minor changes. Increasing the voltage of the signal generator led to an increase in the frequency response. After applying bandpass filtering, as shown in Fig. 7, the frequency response exhibited a sine wave shape consistent with the input signal from the generator, and the filtered signal remained stable. At an applied voltage of 100 mv, the frequency response at 200 m and 10 km displayed a sine wave with an amplitude of 2. Increasing the voltage to 300 mv resulted in the frequency response increasing to three times its original value. The experimental results confirmed the independence of the frequency response on distance and frequency.
Fig. 6. Frequency response at 200 m and 10000 m. (a) Frequency response at 200 m and modulator voltage is 100 mv. (b) Frequency response at 10000 m and modulator voltage is 100 mv. (c) Frequency response at 200 m and modulator voltage is 300 mv. (d) Frequency response at 10000 m and modulator voltage is 300 mv.
Fig. 7. Filtering frequency response at 200 m and 10000 m: (a) 200 m-100 mv, (b) 10000 m-100 mv, (c) 200 m-300 mv, (d) 10000 m-300 mv.
The phase signals $\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta }{\phi _2}(t )$ can be obtained from the interferometric signals. By applying a band-pass filter in the frequency range of 50-200 Hz, the filtered signals exhibit a certain time delay. The delay time Δτ can be determined using the cross-correlation function R(τ), as shown in Fig. 8.
Fig. 8. Experiment results of PZT at 10km: (a) $\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta }{\phi _2}(t )$; (b) $\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta}{\phi _2}(t)$ after amplifying; (c) cross-correlation function R(τ) of $\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta}{\phi _2}(t)$; (d) R(τ) after amplifying.
B. Contact disturbance source location
To evaluate the no-blind zone location effect of the sensing structure, the fiber was deliberately disturbed at various positions. Figure 9 illustrates the signals received at the photodetectors for MI and MZI when perturbed at a distance of 10 km. Visually, MI and MZI exhibit distinct signal patterns due to their structural differences, but the signals are synchronized in time.
The signals ${\varphi _{MI}}(t )$ and ${\varphi _{MZI}}(t )$ are obtained by using the demodulated algorithm. According to the Eq. (7) and (8), we can obtain $\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta }{\phi _2}(t )\; $ respectively. The signals are filtered using a band-pass filter of 50-2000Hz and time delay $\tau $ is obtained by a correlation algorithm. The results are shown in Fig. 10. Then we can get the distance by ${L_x} = L - \frac{{c \cdot {\tau _m}}}{{2n}}$. The result is 9888 m, The disturbance signal is applied at 10 km, and the experimental results are consistent with the actual disturbance point.
Fig. 10. Experiment results of disturbance at 10 km: (a) ${\varphi _{MI}}(t)\;\textrm{and}\; {\varphi _{MZI}}(t)$; (b) $\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta }{\phi _2}(t )$; (c) $\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta}{\phi _2}(t)$ after filtering; (d) cross-correlation function R(τ) of $\mathrm{\Delta }{\phi _1}(t )$ and $\mathrm{\Delta }{\phi _2}(t )$.
The results of the positioning experiment for various distances in this structure were obtained by measuring at identical distance points on two sensing arms. Figure 11 shows the location errors at 0 m, 4 km, 10 km, 20 km, 30 km positions for 35 times. From the figure, it can be seen that the results at 0 m and 10 km have the low location errors within ± 20 m and ±15 m, but as the distance increases, the location errors at 20 km and 30 km increase to ±85 m, proving that the structure is equally accurate for long distance location. There are two main reasons for data separation at 20 km. First, one of the sensing arms is fused, the optical path state of this arm has more insertion loss and contact reflection, and the noise introduced by light reflection and scattering is greater, the results of the fused arm correspond to the negative error part. Second, the other sensing arm is close to the cool fan, which introduces large noise, the results of this arm correspond to the positive error part, causing data separation.
After analysis, we believe that the main sources of the error come from two aspects. 1) The thirty kilometers long fiber is wound in a small box; could lead to issues such as optical power loss and refractive index change. 2) There is a lot of noise in the experimental environment and structure, such as the Rayleigh scattering noise which causes the reduce of SNR.
4. Pattern recognition
A. Dataset
We added a fence fiber optic cable with a length of 10250.7 m for disturbance signal acquisition at different locations on the sensing arm of the original optical path, and collected signals of six intrusion events, include sawing, knocking, impacting, hammering, lashing, kicking, and extracted the detected signals. For strong disturbances such as hammering. We used a threshold judgment and sliding window extraction method. hammering is a type of short-term high-energy signals, when the peak-to-peak value of the data exceeds the threshold and the signal mean value is greater than a certain threshold, the time-series signal is extracted. A total of 12 milliseconds of phase information was collected before and after the onset of the disturbance. For disturbances with long duration and consistent signal amplitude changes, such as sawing, the time-series signal was extracted using a window-dividing method to extract 12-millisecond signals. Fig. 12 shows the signals captured by different events.
The phase change resulting from kicking exhibits a small initial vibration in the signal, which is due to the kicking acts on both the optical fiber cable and the guardrail at the same time. Therefore, it can be seen that at the beginning of the disturbance, the amplitude will first undergo a small change, and then suddenly increase and remain in an oscillatory state, with an overall amplitude of about 100 rad. The impact of the ball causes a significant sudden change, with the amplitude becoming very large, reaching about 400 rad.The hand-knocking has similar characteristics in the time domain to the kicking, but the amplitude difference is relatively large, with the hand-knocking amplitude being about 60 rad.The lashing mainly causes the optical fiber to slight oscillate, with an overall amplitude of 30-40 rad. Overall the observed differences in amplitude highlight distinct characteristics for each type of disturbance.
Fig. 13 shows the images obtained by MTF. We can observe that the diagonal of the transformed images reflects the intensity changes of the original 1D sequence, while the vertical and horizontal variations depict the correlation between different time points. Taking the first chart of sawing as an example, the overall symmetry along the diagonal line is consistent with the basic symmetrical characteristics of time-series signals. The color depth of the nodes can represent the intensity or importance of the state, and the color depth of the edges can represent the size of the transition probability between states. The characteristic of sawing is that there are intensity points and probability of occurrence at any time point, which is consistent with the characteristics of time-series signals. Intuitively, the images generated by the MTF algorithm for the same type of interference have both differences and similarities. After data collection, we created a dataset containing 3150 images and 6 categories, with each image size of 729*730 and saved in jpg format. The specific information of the dataset is shown in Table 1.
Table 1. Dataset details
B. Training results
Due to the large size of the original image, in order to accelerate the overall training speed, we converted the data image to a standardized image format with a size of 224*224. The dataset was randomly split into a training set and a testing set with a ratio of 5:1, and then input into the CNN network for training. The loss function was set to “CrossEntropyLoss”, which is mainly used to calculate the cross-entropy loss between the model prediction and the actual label. The optimizer was set to “Adam”, which optimizes the parameters of the model by adaptively adjusting the learning rate. This optimizer is adaptive to sparse gradients, without the need to manually adjust hyperparameters. In the hyperparameter setting, the learning rate was set to 0.001, the batch size was 32, and Adam was the default value. The model was trained for 200 times. The BatchNorm2d layer helps reduce the sensitivity of the model to the distribution of input data, improve the generalization ability of the model, and accelerate the convergence speed of the model. The convolutional layer was used to extract the features of the input image, which is beneficial for extracting local features. The structure has three convolutional layers, with the number of output channels starting from 32 and doubling sequentially. The RGB image has three channels. The convolution kernel size is 5*5. The activation function type is “Relu”, which introduces nonlinearity and alleviates the problem of gradient vanishing. Sparse activation reduces the risk of overfitting. The maximum pooling type of 2*2 was selected for feature reduction, preserving the main features and discarding unimportant details; the output layer consists of three fully connected layers, and the final output is 6 classifications. The neural network constructed in this article is based on python and torch1.12.1. The experiment was conducted on the Kaggle platform T4, but the running speed is 200 rounds per two hours. All parameters of the CNN are shown in Table 2.
Table 2. CNN details
After 200 training epochs, stable results are depicted in Fig. 14. The training set achieves 100% accuracy after 17 epochs, with a minimum loss of 5.5329197E-09. The testing set achieves a maximum accuracy of 99.32% and experiences slight fluctuations, with a loss function of 0.0193761. The model converges rapidly.
The testing set is sent to the training model for identification, and the heat map obtained is shown in Fig. 15. From the heat map, it can be seen that the accuracy is higher for different types of disturbance behaviors, the lowest is Hammering within only 98.0%, which is explainable, as hammering is stronger than acts such as lashing, but slightly weaker than impacting. Hammering itself contains signals like knocking and other behaviors, and the similarity of the signals is more like hand knocking, so the incorrect cases are mostly misidentified as knocking. At the same time, when collecting data, there are a few data that do not have enough hammering strength and do not meet the hammering standard (max amplitude is about 135-150 rad). Since the purpose of classification is to distinguish threatening disturbance behaviors, hammering behaviors that achieve certain effects are classified into one category. The accuracy of Impacting is as high as 100.0%. because the amplitude of this signal is 400 rad Far higher than other signals, in general the classification results are satisfactory. Some sawing signals are predicted to be kicking signals. There are two main reasons. First, some dirty data appears in the sawing dataset, mainly because the timing signal without cutting action at the beginning was added to the dataset, Second, too many rounds of model training result in overfitting, which leads to inaccurate model predictions.
5. Conclusion
This article provides a comprehensive verification of the localization algorithm based on the MI-MZI structure, offering both theoretical and practical evidence of its effectiveness. Through analysis of the localization algorithm and frequency response analysis of the sensing structure, it has been theoretically proven that this structure can overcome the limitations of traditional structures that have frequency response defects. Compared with other sensors, this structure’s algorithm and structure is not complex. Experimental measurements of the frequency response curve of the MI-MZI structure were conducted at sensing lengths of 10 km and 200 m. The results affirm that the frequency response capability of this structure aligns with the theoretical predictions, further validating its performance. Localization tests were conducted at different locations. The maximum localization error was ±85 m, mainly due to environmental noise and other factors.
To achieve pattern recognition, a process was proposed based on MTF and convolutional neural networks. The signal was converted into a two-dimensional image as a processing dataset, and six-class pattern recognition was achieved. The training model achieved an accuracy of 98.74%, with the Impacting achieving 100% accuracy.
In terms of positioning, the main factors influencing accuracy are: 1) The simplicity of the sensing structure in this study does not consider polarization attenuation. By utilizing polarization controllers and optimizing polarization characteristics, the negative effects of scattered light and polarization-induced challenges can be mitigated. 2) Environmental factors may have an impact. Due to the high sensitivity of the MZI, low-frequency environmental factors can become interference signals for the sensing optical cable. Although bandwidth filters were used for filtering, the influence of external air conditioning and personnel walking in the laboratory environment cannot be completely avoided. 3) The winding of the sensing optical fiber on the optical disc during the experiment could lead to issues such as optical power loss and refractive index change. To optimize pattern recognition, more signal data is necessary for neural network optimization, and the entire program requires optimization for practical application.
Funding
Shanghai Optical Fiber Monitoring Engineering Technology Research Center (17DZ2280600).
Disclosures
The authors declare no conflicts of interest.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from authors upon reasonable request.
References
1. K. K. K. Annamdas and V. G. M. Annamdas, “Review on developments in fiber optical sensors and applications,” in Fiber Optic Sensors and Applications VII (SPIE, 2010), Vol. 7677, pp. 205–216.
2. C. Xu, S. Du, P. Gong, et al., “An Improved Method for Pipeline Leakage Localization With a Single Sensor Based on Modal Acoustic Emission and Empirical Mode Decomposition With Hilbert Transform,” IEEE Sensors J. 20(10), 5480–5491 (2020). [CrossRef]
3. G. A. Cranch, L. Johnson, M. Algren, et al., “Crack detection in riveted lap joints using fiber laser acoustic emission sensors,” Opt. Express 25(16), 19457–19467 (2017). [CrossRef]
4. T. Kumagai, S. Sato, W. Ohnuki, et al., “Fiber-optic intrusion detection sensor for physical security system,” in 21st International Conference on Optical Fiber Sensors (SPIE, 2011), Vol. 7753, pp. 468–471.
5. Y. Lu, T. Zhu, L. Chen, et al., “Distributed Vibration Sensor Based on Coherent Detection of Phase-OTDR,” J. Lightwave Technol. 28, 3243–3249 (2010). [CrossRef]
6. J. C. Juarez, E. W. Maier, K. N. Choi, et al., “Distributed fiber-optic intrusion sensor system,” J. Lightwave Technol. 23(6), 2081–2087 (2005). [CrossRef]
7. X. Hong, J. Wu, C. Zuo, et al., “Dual Michelson interferometers for distributed vibration detection,” Appl. Opt. 50(22), 4333–4338 (2011). [CrossRef]
8. C. Ma, T. Liu, K. Liu, et al., “Long-Range Distributed Fiber Vibration Sensor Using an Asymmetric Dual Mach–Zehnder Interferometers,” J. Lightwave Technol. 34(9), 2235–2239 (2016). [CrossRef]
9. W. Yuan, B. Pang, J. Bo, et al., “Fiber-Optic Sensor for Acoustic Localization,” J. Lightwave Technol. 32(10), 1892–1898 (2014). [CrossRef]
10. S. J. Russell, K. R. C. Brady, and J. P. Dakin, “Real-time location of multiple time-varying strain disturbances, acting over a 40-km fiber section, using a novel dual-Sagnac interferometer,” J. Lightwave Technol. 19(2), 205–213 (2001). [CrossRef]
11. J. P. Dakin, Distributed Optical Fiber Sensor Systems (1989), Vol. 2, pp. 575–598.
12. E. Udd, “Distributed fiber optic strain sensor based on the Sagnac and Michelson interferometers,” in Smart Structures and Materials 1996: Smart Systems for Bridges, Structures, and Highways (SPIE, 1996), Vol. 2719, pp. 210–212.
13. S. J. Spammer, P. L. Swart, and A. A. Chtcherbakov, “Merged Sagnac-Michelson interferometer for distributed disturbance detection,” J. Lightwave Technol. 15(6), 972–976 (1997). [CrossRef]
14. B. Mohanan, P. J. Shaija, and S. Varghese, “Studies on merged Sagnac-Michelson interferometer for detecting phase sensitive events on fiber optic cables,” in 2013 International Mutli-Conference on Automation, Computing, Communication, Control and Compressed Sensing (2013), pp. 84–89.
15. X. Lai, H. Yu, Y. Ma, et al., “Disturbance location and pattern recognition of a distributed optical fiber sensor based on dual-Michelson interferometers,” Appl. Opt. 61(1), 241–248 (2022). [CrossRef]
16. H. Wu, J. Chen, X. Liu, et al., “One-Dimensional CNN-Based Intelligent Recognition of Vibrations in Pipeline Monitoring With DAS,” J. Lightwave Technol. 37(17), 4359–4366 (2019). [CrossRef]
17. B. Wu, S. Yuan, P. Li, et al., “Radar Emitter Signal Recognition Based on One-Dimensional Convolutional Neural Network with Attention Mechanism,” Sensors 20(21), 6350 (2020). [CrossRef]
18. K. He, X. Zhang, S. Ren, et al., “Deep Residual Learning for Image Recognition,” in (2016), pp. 770–778.
19. R. Zhang, F. Zheng, and W. Min, “Sequential Behavioral Data Processing Using Deep Learning and the Markov Transition Field in Online Fraud Detection,” arXiv, arXiv:1808.05329 (2020). [CrossRef]
20. S. K. Sheem, “Optical fiber interferometers with [3×3] directional couplers: Analysis,” J. Appl. Phys. 52(6), 3865–3872 (1981). [CrossRef]
21. R. Priest, “Analysis of fiber interferometer utilizing 3 × 3 fiber coupler,” IEEE J. Quantum Electron. 18(10), 1601–1603 (1982). [CrossRef]
22. H. Wu, Y. Feng, H. Xu, et al., “A new demodulation method to improve the sensitivity and dynamic range of fiber optic interferometric system,” in 9th International Conference on Optical Communications and Networks (2010), pp. 70–72.
23. Y. Ma, Y. Song, Q. Song, et al., “MI-SI Based Distributed Optical Fiber Sensor for No-Blind Zone Location and Pattern Recognition,” J. Lightwave Technol. 40(9), 3022–3030 (2022). [CrossRef]
