Open Access
Add to BibSonomy
Abstract
The co-route optical fibers, comprising both co-cable and co-trench fibers, pose a significant potential risk to network service quality assurance by operators. They are incapable of achieving high-precision recognition and visual state management. In this study, we gathered both static and dynamic optical fiber data using a linewidth tunable light source (LTLS) and introduced a multimodal detection architecture that applies ensemble learning to the collected data. This constitutes what we believe to be the first field trial of concurrent recognition of optical fibers found both in co-cables and co-trenches. To identify co-cable fibers, we employed a double-layer cascaded Random Forest (DLC-RF) model based on the static features of fibers. For co-trench fiber, the dynamic characteristics of fiber vibrations are utilized in combination with multiple independent curve similarity contrast learners for classifying tasks. The proposed architecture is capable of automatically detecting the condition of the optical fiber and actively identifying the same routing segment within the network, eliminating the need for human intervention and enabling the visualization of passive optical fiber resources. Finally, after rigorous testing and validation across 11 sites in a typical urban area, including aggregation and backbone scenarios within the operator's live network environments, we have confirmed that the solution's ability to identify co-routes is accurate, exceeding 95%. This provides strong empirical evidence of its effectiveness.
© 2023 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Due to the passive nature of optical cables, effective management and visualization of their status have been unattainable for an extended period [1]. In urban settings, mismanagement of “passive resources” such as optical cables and fiber trenches can lead to the coexistence of primary and backup optical routes within the same cable conduit, resulting in the emergence of co-routing optical fibers. Co-routing fibers encompass both co-cable and co-trench fibers, with co-cable fibers referring to those encapsulated within the same optical cable and co-trench fibers indicating those buried within the same trench. It is important to note that all co-cable fibers are necessarily co-trench fibers, but the reverse is not always true. The disruption of the cable or trench housing co-routing fibers can consequently lead to the simultaneous failure of primary, backup, and associated services. If the interrupted location occurs in the co-trench and co-cable scenario, the corresponding service security is 0%. According to rough estimates cited by Reuters, profitable websites around the world lose $\$ 29$ million in revenue for every hour of disruption, which seriously affected user experience and brand reputation as well as industry development. Therefore, operators urgently need an effective and reliable way to identify the co-route fiber to improve network survivability.
Meeting the challenge of network faults and anomalies and improving network survivability has always been an important field in the research on network reliability [2–4]. To deal with the service outage caused by optical fiber interruption, the operators generally go with shared risk link groups (SRLGs), which contribute network planners to implement corresponding measures to enhance network reliability and performance, such as rational backup path planning, fault tolerance strategies, and resource management [5,6]. Similar to SRLGs, the co-route fiber is called shared risk optical fiber links (SROFL) in literature [7]. It should be pointed out that co-routing fibers encompass SROFL, and SROFL, in turn, subsume SRLG. When two fiber cables are routed in proximity (co-located or in the same pipe), their geographical proximity and similar laying methods lead to comparable fiber characteristics and aging features, the response to external vibration is very similar, or even the same. Consequently, these fibers exhibit inherent similarities, enabling the utilization of artificial intelligence techniques to discern whether they originate from the same cable or trench.
In our previous work [1], we proposed an artificial intelligence (AI) method for co-route optical fiber detection based on twin neural network and extraction of multimodal features, e.g. fiber static, dynamic, and site features. Although its detection accuracy can exceed 90%, it should be noted that most of the data used in these evaluations comes from artificial laboratory environments rather than actual network conditions. In order to discern SROFL, Ref. [7] harnessed neural networks to acquire the weight distribution of their proposed adaptive weighting algorithm, consequently enhancing the precision of co-cable recognition. Although this work conducted on-site testing, the dataset was drawn from 54 dark fibers across 14 optical cables, a diminutive sample size that poses challenges to ensure robust model generalization. Reference [8], on the other hand, employed ultrasonics to excite optical cables, achieving co-route identification based on vibrational characteristics. The work [8] has exhibited promising potential, yet it lacks field experimentation. Elevating network survivability necessitates ensuring high efficacy of co-route recognition models in real networks, a task heavily reliant on extensive field testing. The sudden nature of network failures demands network administrators to respond to network states in real-time. While the aforementioned endeavors have made commendable contributions, they fall short in terms of online recognition and robust generalization capabilities, making it difficult to verify that their work is suitable for large-scale deployment in networks.
To address the limitations of prior research efforts, we propose an architecture that enables online identification of the co-cable and co-trench fibers in an operator`s networks and is expected to be deployed on a large scale in real networks. This architecture can achieve high accuracy while ensuring the generalization ability of the model. This framework deploys Intelligent Sensing Units (ISUs) within the network, which gather both dynamic and static data from optical fibers in actual network environments. These ISUs collaborate with SDN controllers to achieve online recognition of co-route fibers. Leveraging the principles of ensemble learning, we bolster the model's generalization capabilities in intricate network scenarios. Specifically, for the co-cable optical fiber recognition task, a DLC-RF model is built using Decision Trees (DT) to accomplish the classification task. Moreover, for the co-trench optical cable identification problem, multiple independent base learners are developed. They assess whether pairs of optical fibers belong to the same trench based on the similarity of vibration event curves. The key contributions of this paper are summarized as follows:
- • We present, to our knowledge, for the first time, an architectural framework capable of concurrent online recognition of co-cable and co-trench optical fibers based on ensemble learning without human intervention throughout the entire process. Subsequently, rigorous testing was conducted within operator`s live network environments. In these field experiments, static data from optical fibers amounted to 4,998 pairs, while the dynamic data exceeded a magnitude of over 100,000.
- • We introduce, to our knowledge, for the first time, the application of a double-layer cascading ensemble learning paradigm to the realm of coaxial cable optical fiber recognition. We propose a DLC-RF classifier, which integrates the classification outcomes of all conceivable feature vector combinations for each fiber pair, furnishing a comprehensive classification profile for these pairs. In the field test, the accuracy of the co-route identification is more than 95%.
The rest of this paper is organized as follows. Section II summarizes the research status of optical network survivability and fiber identification. Section III introduces the system architecture and mathematical model of the co-route fiber recognition. Section IV presents the results of simulation experiments and field tests. Finally, conclusions are given in Section V.
2. Related work
While many studies have been conducted to improve network survivability [9–12], SRLGs focuses on the risk of link outages, which are components of links in optical networks that share common risk factors, such as geographic conditions, construction projects, and natural disasters. Two natural standard definitions have been proposed to capture the probabilistic implications associated with probabilistic SRLGs [13]. Furthermore, the treatment of related failures in the form of SRLGs has a long-standing history, as evidenced by prior works such as [14].
Historical data is often utilized in these approaches, with the assumption that the risk groups are part of the input. Through a self-constructed model, a pair of risk-disjoint paths is identified. This proximity can be considered as a simple form of geographically correlated failures. Other literature on path planning includes [15–20].
However, these endeavors have primarily focused on route planning from the perspective of fault probabilities, yet risks persist in deploying primary and backup services or associated services on the same optical fiber. Hence, it becomes imperative to identify fiber co-routing from the perspective of underlying resource management. The static characteristics of the fiber can be extracted by optical time-domain reflectometers (OTDRs), which have found extensive utility in testing, debugging, and maintaining fiber optical networks [21–25]. These tests aim to assess the reflection and loss conditions at various positions, such as fiber connections, splices, connectors, and fiber ends [1,26]. Furthermore, we employed a phase sensitive optical time-domain reflectometer (φ-OTDR) to capture the fiber's vibration data. φ-OTDR, is a distributed fiber optic sensing technology, which are widely recognized for their high sensitivity and ease of handling, making them applicable in various scenarios, including earthquake prediction and monitoring tasks such as underwater cable surveillance [27].
Q. Cui et al. employed differential interference contrast microscopy to capture fiber images and explored the automatic fiber identification technique using a genetic neural network algorithm, which combines neural networks with genetic algorithms [28]. The accuracy of their approach was validated. OTDR data encompass various fiber characteristics, and T. J. Xia et al. demonstrated, for the first time, that accurate determination of the geographical location of deployed fiber cables can be achieved by utilizing OTDR distance in distributed fiber sensing technology [29]. Their work focused on fiber localization and did not provide a method for co-cable identification. Y. Shi et al. accomplished accurate classification of fiber events through transfer learning. However, they did not analyze the relationship between fiber events and fiber cables [27,30]. Additionally, some studies have employed machine learning algorithms and OTDR data to accomplish network fault localization, yet the task of co-cable identification has been overlooked [31–34]. There are also many researches on the identification and classification of optical fiber vibration events by using machine learning methods, which are reviewed in detail in [27]. Similarly, these works ignore the similarity of event characteristics and the relevance of the problem of optical fiber co-routing identification.
3. System architecture and mathematical model of the co-route identification
In our solution, the co-route recognition module is placed within the SDN controller. We deploy ISUs at the source and aggregation nodes of the optical transport section. These ISUs are capable of simultaneously monitoring multiple optical fibers using both OTDR and φ-OTDR techniques. They collect data on scattering characteristics due to fiber degradation, fractures, and bending, as well as polarization features resulting from activities such as vehicular traffic, excavation, pedestrian movement, and running.
3.1 System architecture
In Fig. 1, we show the system scheme of detection of co-route optical fiber identification. Data acquisition and processing are mainly completed by ISU, which consists of transmitting module, receiving module and processing module, shown in Fig. 1. The transmitting modules are further divided into LTLS, optical modulators, EDFA and isolators. LTLS has two operational modes, namely, broadband mode (BM) and narrowband mode (NM), which correspond to BM and NM in Fig. 1. When LTLS operates in BM, as shown in Eq. (1), the slope of the detected Rayleigh signal curve is related to fiber loss, enabling the assessment of fiber quality in terms of attenuation, reflection, fusion splices, and aging, representing static characteristics.
Fiber Rayleigh backscattered signals contain a large amount of fiber characteristic information, such as connectors, joints, fusion points, and the fiber itself, which are all reflected in Rayleigh scattering signals as characteristic information, as shown in Fig. 2. When operating in NM, as revealed in Eq. (2), interference occurs in the Rayleigh scattering, and when there is an external signal intruding, as displayed in Fig. 2, the phase of the interference signal changes due to the external intrusion and the change is linear to the intrusion signal. By using this relationship, the optical fiber can sense the external vibration and identify the dynamic features of the fiber. The intelligent recognition unit (IRU) is deployed within the software defined networking (SDN) controller, as exhibited in Fig. 1, allowing the feature data of the optical fibers to be input into a pre-trained IRU for online co-route detection.
The IRU maintains communication with the ISU, collecting static and dynamic fiber feature data transmitted by ISU. Subsequently, the IRU analyzes the collected fiber feature data to identify segments of optical fibers belonging to the same cable or trench.
3.2 Mathematical model of random forest algorithm for co-cable recognition
The OTDR data of optical fibers encompass a multitude of characteristics. In this study, we categorize the fiber features into two parts: curve features and event features. Curve features are numerical values extracted from the OTDR files, represented as one-dimensional sequential signals. These curve features are input into a Gated Recurrent Unit (GRU), producing a sequence feature of length 64. Subsequently, the obtained sequence feature is fed into a fully connected layer, resulting in a curve feature vector of length 4, as shown in Fig. 3(a). For event features of an optical fiber, we select four feature values: event type, event occurrence location, echo loss, and splice loss, to construct an event feature vector (where event type is encoded as a specific number).
In Fig. 3(b), we concatenate the four-length curve feature vector with the four-length event feature vector to obtain a comprehensive eight-length fiber feature vector. If a fiber has i events, the size of the matrix storing its fiber comprehensive feature vectors would be i*8. This approach ensures that no event information is lost while maintaining a manageable length for the fiber feature vectors to avoid excessive search space and impact on classification performance [37,38].
The task of determining whether any two fibers belong to the same cable can be decomposed into a binary classification task. As depicted in Fig. 3(c), we gather all feature vector combinations of fiber pairs and assign labels to them. Fiber pairs within the same cable are labeled as 1, while pairs from different cables are labeled as 0.
In this section, we elaborate on the construction and training process of the RF model, which is employed for the classification task using the extracted features from OTDR data. RF is an ensemble learning method that combines multiple DTs to make predictions. It leverages the principle of bagging and incorporates randomness in both feature selection and sample selection.
The construction of the data set requires random sampling: Given a labeled dataset consisting of feature vectors from different fiber pairs, we randomly sample the data with replacement to create multiple subsets, known as bootstrap samples. For each bootstrap sample, we construct a DT. The tree's depth is set to six layers, enabling the model to capture complex relationships among the features effectively. At each layer of the DT, instead of considering all features, a random subset of features is chosen. This promotes diversity among the trees in the forest, enhancing the model's generalization ability. To determine the optimal splitting of nodes during tree construction, we employ the Gini impurity index, which quantifies the level of impurity in a node based on the class distribution [36]. The feature and corresponding threshold that minimize the Gini impurity are selected for node splitting. Mathematically, the Gini impurity index for a given node with respect to the class labels can be expressed as follows:
where ${p_i}$ denotes the probability of randomly selecting an element belonging to class i from the node.To compute ${p_i}$, we divide the number of elements in the node that belong to class i by the total number of elements in the node. In the context of a DT, the Gini impurity index is used to determine the optimal split for a given node by evaluating the impurity reduction achieved by potential splits. The split that minimizes the Gini impurity is chosen as the optimal split.
In training process, the DT grows recursively by iteratively splitting the nodes until a stopping criterion is met, such as reaching the maximum depth or a minimum number of samples required for further splitting. After constructing multiple DTs using different bootstrap samples and random feature subsets, they are combined to form the RF ensemble.
In the previous sections, we constructed the event feature vectors of fibers as vectors with a length of 4. However, in practical scenarios, fibers often have more than one event, indicating that the constructed fiber feature vectors actually form a group of vectors, with the number of vectors depending on the number of events in the fiber. In this study, the DLC-RF model treats the combination of feature vectors from a pair of fibers as the smallest processing unit. To achieve more accurate classification results and ensure no loss of fiber event information, we need to consider the classification results of all possible combinations of feature vectors from a pair of fibers and integrate them to provide the final classification result.
As shown in Fig. 4, Fiber a has p events, and Fiber b has q events. the feature vector matrix of this fiber pair will have $p\ast q$ rows, signifying that the DLC-RF model will entail $p\ast q$ fundamental processing units. In the process of recognition, let $p\ast q = T$, then the RF model will make classification decisions for these T fundamental processing units. Each classification outcome of each unit is denoted as ${y_t}$, where t designates the t-th fundamental processing unit. Each ${y_t}$ takes only two values, 1 and 0, representing ‘co-cable’ and ‘non-co-cable,’ respectively.
For the final verdict layer of the DLC-RF model, as illustrated in Eq. (4), the ultimate classification outcome, ${y_{final}}$, can be expressed as follows:
Here, I() represents the indicator function. It yields 1 if the condition within the parentheses is true, otherwise, it returns 0. The aforementioned formula signifies that the final classification result, ${y_{final}}$, corresponds to the majority class. If ‘co-cable’ (y = 1) occurs more frequently than ‘non-co-cable’ (y = 0), ${y_{final}}$ will be set to 1; otherwise, it will be set to 0.
In the event that an equal number of occurrences of 0 and 1 transpire in the classification, the final categorization is determined through a probabilistic output approach. In this case, the ultimate classification outcome will be the category with the higher probability value. This determination is accomplished by averaging the probabilities produced for each category by each DT. This mode of majority decision empowers the model to enhance its classification performance by aggregating the random forest decisions of multiple fundamental processing units.
3.3 Process and scheme of co-trench identification
Given the unwieldy volume of raw data, transmitting it across the network would lead to intolerable latency. Hence, we convey lightweight vibration event curve data to the corresponding processing units. The extracted vibration event curves are transmitted to the IRU within the SDN controller, equipped with multiple independent base learners to perform similarity comparison learning of the vibration event curves. The mechanism for determining whether any two optical fibers belong to the same trench follows the same principles of classification as RF. Each decision maker independently assesses and, in the end, the classification results are determined through the majority vote. This similarity comparison employs three curve similarity comparison algorithms: ED, DTW, and FD.
The ED is used to measure the straight-line distance between two points. The Euclidian distance between the two curves, Li and Lj, is the square root of the sum of the squares of the EDs of the points between them. The smaller the value of ED, the more similar the curves Li and Lj are. If Li and Lj are p-dimensional trajectory segments with length of n, the ED between Li and Lj can be expressed as:
Different from the ED algorithm, the FD algorithm takes more account of the connection between the points in the curve, and its calculation method involves the parameter ɛ. It means that two observers (on the curves Li and Lj, respectively) are traveling at a fixed speed and cannot be separated by a distance of ɛ. FD is calculated as follows:
The principle of DTW involves calculating a distance matrix between each point of two sequences and utilizing dynamic programming to compute the shortest path within the distance matrix. The sum of the shortest path computed is then considered as the curve similarity, denoted as D, which can be precisely described by the following equation:
Figure 5 illustrates the application of the DTW algorithm in computing the optimal path for similarity between two vibration curves. Each curve comprises 2000 sampling points, and the distances between corresponding values of these two curves are depicted through a heatmap. Following the determination of the optimal path, Equation eight is employed to calculate the mean distance, denoted as ‘D’, which serves as one of the criteria for subsequently discerning whether optical fibers share the same trench.
4. Experimental demonstration and performance evaluation
4.1 Evaluation measures
In this paper, we use accuracy to evaluate the classification effect. As shown in Table 1, to calculate the evaluation measures, we need to define a second-order confusion matrix.
Table 1. The definition of the confusion matrix.
True Positive (TP) represents co-route data that is correctly identified, False Negative (FN) represents co-route data is identified as non-co-route data. False Positive (FP) represents non-co-route data is identified as co-route data, and True Negative (TN) represents non-co-route data that is correctly identified.
Accuracy represents the amount of the data correctly classified as a percentage of the entire:
In order to more fully explain the classification accuracy of RF model, we also use F1-score as an evaluation index, which takes both precision rate and recall rate into account. The precision rate and recall rate can be expressed as:
The F1-score is considered a weighted average of accuracy and recall and can be calculated by the following formula:
4.2 Empirical results
Figure 6 shows the trend of some typical events across the OTDR curve of the fiber, such as splicing losses and reflection event. Positive splice loss refers to the loss of optical energy caused during the connection of optical fibers due to factors such as loose junctions or non-smooth optical fiber end faces. In contrast, negative splice loss pertains to the lack of synchronization in the optical fiber cores at the connection point, resulting in incomplete alignment of optical fiber end faces and the consequent loss of optical beam transmission. Both of them lead to the attenuation of optical signals.
Except for the event type, which is explicitly represented as textual information, all other parameters are numerical. For the sake of uniform measurement, we aim to avoid semantic segmentation in our experiments. Therefore, we map the event types to numerical values for ease of processing. The mapping relationship is shown in Table 2.
Table 2. Event types encode mappings.
As shown in Fig. 7(a), two fibers from the same cable have the same number of events, and the same type of events are detected at the same position of the fiber. Specifically, at a distance of 0.585 km from the OTDR detector, the event type “reflection” was detected by both fibers, which was coded as “1”. And the event type “loss/drop/gain”, with the location of 0.981 km, was coded as “4”.
Fig. 7. Examples of feature vectors of different fibers, distinguished by different colors. The curve feature vector is enclosed in square brackets, and the event feature vector is enclosed in parentheses. (a) Co-cable fibers of fiber x and fiber z. (b) Non-co-cable fibers of fiber y and fiber z.
We can also see that the reflection loss and splicing loss values of the two fibers in the same cable are also very close. From the two curves in Fig. 7(a), it can be seen that the OTDR curve trajectories of the two fibers are very similar, and this similarity is also reflected in the curve feature vectors of the fibers. Meanwhile, Fig. 7(b) shows the characteristic information of two fibers from different optical cables, and they have very large dissimilarity in both curve characteristics and event characteristics.
The dataset encompasses various scenarios in the live network, including metropolitan, aggregation, and backbone networks. The fiber lengths range from 0.6 km to 90 km, and the number of fiber events varies from 2 to 12. Additionally, subsequent verification revealed that approximately 20% of the measured fibers in the dataset experienced bending or compression, reflecting the actual conditions of the passive resources in the live network and providing rich environmental diversity to the dataset.
The dataset was divided into training and testing sets in an 8:2 ratio. In order to verify the performance of the proposed DLC-RF model, two algorithms of support vector machine (SVM) and naive Bayes classification (NBC) are selected for comparison. Figure 8(a) displays the performance of the three models. We conducted a hundred repetitions of the experiment and took the average. It is evident that, with the increase in data volume, the accuracy of DLC-RF gradually rises, eventually reaching over 95%. In contrast, the performance of the other two algorithms is less satisfactory, with SVM hovering around 60% accuracy and NBC even lower at approximately 44%.
Fig. 8. (a) Average accuracy of the three model in different size of data (fiber pairs). (b) K-fold verification of accuracy, where K = 10. (c) The F1-score of the three model in different size of data (fiber pairs).
To assess the model's generalization capability, we partitioned the dataset into training and testing sets using ten different methods, as depicted in Fig. 8(b). DLC-RF consistently outperforms the others, maintaining an accuracy range of 94% to 97%, while SVM and NBC's performance remains poor. To further evaluate the model's performance, Fig. 8(c) presents the F1-score, which takes into account both precision and recall. As observed, the DLC-RF model we propose continues to outshine the other two models, reaffirming its superiority in the problem of route recognition.
Figure 9 presents an overview of the on-site experimental architecture. The proposed ISUs are deployed within the OTN equipment, and the IRUs are embedded within the OTN controller. Within the communication site, the OTN devices is connected to the Optical Distribution Frame (ODF) via tail fibers. The ISUs extract features from the collected raw data, transmitting the fiber's feature vectors and vibration event curve data to the OTN controller. The recognition of co-route optical fibers is accomplished by the IRU.
Our objective is to deploy the proposed architecture at scale across the entire network. While preliminary simulations provided initial evidence of the model's superiority, it is essential to conduct validation tests in real network scenarios. Figure 10 presents the results of network validation, where we conducted recognition tests for same-cable fibers at a total of 11 sites, encompassing various primary network scenarios, including aggregation, metro, and backbone scenarios. The circles and arrows in the Fig. 10 show how the data corresponds to the vertical axis. The actual co-cable data is represented by the orange columns, while the recognized co-cable data is depicted by the blue columns. The values of these columns correspond to the left vertical axis. The curve represents the recognition accuracy, with its values corresponding to the right vertical axis.
The metro optical network scenario, being a typical setting for co-route occurrences, underwent extensive testing at Site 11. As illustrated in Fig. 10, a total of 4,297 fiber pairs were tested, out of which 4,122 pairs of co-route fibers were correctly identified, achieving an impressive recognition accuracy of 95.93%. At Sites 1, 2, 3, and 5, the recognition accuracy even reached 100%, however, at Sites 8-10, the recognition accuracy did not exceed 90%. Throughout the entire testing process, consisting of 4,932 fiber pairs to be recognized across all 11 sites, a total of 4,709 fiber pairs were correctly identified, resulting in an overall accuracy rate of 95.48%. Therefore, although individual sites displayed recognition accuracy below 90%, it remains acceptable as the overall recognition accuracy continues to exceed 95%.
In the context of the co-route issue involving co-trench scenarios, we employed multiple independent base learners to accomplish the recognition and learning of vibration event curves. To enhance the model's performance, a substantial dataset is leveraged to iteratively adjust similarity thresholds, indicative of co-trench conditions. Subsequently, we embarked on a 15-day testing and validation campaign within live network environments, with the results showcased in Fig. 11. Analogous to Fig. 10, the orange bars denote the actual number of co-trench fiber pairs, while the blue bars represent the correctly identified co-trench fiber pairs. Both of their values correspond to the vertical axis on the left. Unlike the static data of optical fibers, the identification of vibration data is a notably time-intensive endeavor. Figure 11 also documents the time consumed during each test days, which correspond to the right vertical axis.
The first day's testing was the briefest, lasting merely 48 minutes, while on the eighth day, it extended to a maximum of 17.2 hours. On the first and seventh days, only one pair of co-trench fiber pairs was tested, and they were both correctly identified. Conversely, on the eleventh day, two pairs of co-trench fiber pairs were not correctly identified. Over the course of the 15-day testing period, we examined a total of 206 pairs of same-duct fiber pairs. Out of these, 195 pairs were correctly identified, resulting in an overall recognition accuracy rate of 95%.
To provide a more illustrative depiction of our work, as depicted in Fig. 12, we have crafted a schematic representation of the trench routes from our on-site testing. The total length of the cables involved surpasses 4 kilometers between OTN site1 and OTN site2. In Fig. 12, the red and green line segments represent the routing schematics of two tested optical fibers within the actual operator`s network. During the common wiring segment, the two segments coincide and are represented by blue line segments.
In Fig. 12, we have demarcated four locations where vibration events were concentrated in the two fiber laying routes tested. Location A, B, and C are positioned along the routes of co-trench optical fibers, with respective distances of 700-800 meters, 1400 meters, and 1900 meters from the testing source. The non-co-trench segment commenced at a distance of 2 kilometers from the light source, at Location D with marking 4, a substantial volume of vibration events was recorded. Throughout this testing phase, the model reliably identified co-trench events at markings 1 to 3. Moreover, the model exhibited dependable discernment when faced with non-co-trench routing segment, at Location D in Fig. 12, no instances of misclassification as co-trench segments occurred.
5. Conclusion
This study proposed an integrated learning model corresponding to different data characteristics to complete the task of online co-route optical fiber recognition. The proposed DLC-RF model uses the static characteristics of optical fiber to show a strong performance advantage in co-cable fiber identification. For the issue of co-trench fiber identification, the dynamic characteristics of optical fiber are utilized to build a number of multiple independent base learners, and complete the classification task by the curve similarity of vibration events. It is the first field trial of online co-cable and co-trench optical fiber concurrent recognition without human intervention. It is tested and verified at 11 sites in three typical scenarios: metropolitan area, aggregation and backbone in the field network, and the co-route recognition accuracy rate was more than 95%, which makes the proposed model is expected to be applied in large-scale networks and laying a reliable foundation for the digital twin of optical network.
Funding
National Natural Science Foundation of China (U21B2005, 62201088, 62122015, 62271075); Key Technologies Research and Development Program (No. 2022YFB2903301); BUPT-China Mobile Research Institute Joint Innovation Center.
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 the authors upon reasonable request.
References
1. Y. Li, C. Li, Z. Liu, et al., “Research and Experiment on AI-based Co-cable and Co-trench Optical Fibre Detection,” in Proceedings of European Conference on Optical Communication (ECOC2022).
2. H. Yang, X. Zhao, Q. Yao, et al., “Accurate fault location using deep neural evolution network in cloud data center interconnection,” IEEE Trans. Cloud Comput. 10(2), 1402–1412 (2022). [CrossRef]
3. H. Yang, Y. Wan, Q. Yao, et al., “Anomaly prediction with hybrid supervised/unsupervised deep learning for elastic optical networks: a multi-index correlative approach,” J. Lightwave Technol. 40(14), 4502–4513 (2022). [CrossRef]
4. H. Yang, B. Wang, Q. Yao, et al., “Efficient Hybrid Multi-Faults Location Based on Hopfield Neural Network in 5 G Coexisting Radio and Optical Wireless Networks,” IEEE Trans. Cogn. Commun. Netw. 5(4), 1218–1228 (2019). [CrossRef]
5. B. Vass, E. R. Bérczi-Kovács, Á. Barabás, et al., “A Whirling Dervish: Polynomial-Time Algorithm for the Regional SRLG-Disjoint Paths Problem,” IEEE/ACM Trans. Networking 1, 1–12 (2023). [CrossRef]
6. L. Liu, B. Li, B. Qi, et al., “Optimization of communication capacity for load control considering shared risk link group in source-grid-load system,” Int. J. Elec. Power. 122, 106166 (2020). [CrossRef]
7. Z. Zhao, Z. Zhang, H. Shi, et al., “Field Trail of Shared Risk Optical Fiber Links Detection Based on OTDR and AI Algorithm,” 2022 Asia Communications and Photonics Conference (ACP2022). [CrossRef]
8. W. Zuo, H. Zhou, Y. Qiao, et al., “Investigation of co-cable identification based on ultrasonic sensing in coherent systems,” IEEE Photon. Technol. Lett. 35(21), 1155–1158 (2023). [CrossRef]
9. H. Zhao, H. Yang, Guo, et al., “Accurate Fault Location based on Deep Neural Evolution Network in Optical Networks for 5 G and Beyond,” in Optical Fiber Communication Conference (OFC2019). [CrossRef]
10. A. Yu, H. Yang, Q. Yao, et al., “Accurate Fault Location Using Deep Belief Network for Optical Fronthaul Networks in 5 G and Beyond,” in IEEE Access 7, 77932–77943 (2019). [CrossRef]
11. A. Yu, H. Yang, K. K. Nguyen, et al., “Burst Traffic Scheduling for Hybrid E/O Switching DCN: An Error Feedback Spiking Neural Network Approach,” IEEE Trans. Netw. Serv. Manage. 18(1), 882–893 (2021). [CrossRef]
12. T. Yu, H. Yang, Q. Yao, et al. “Multi visual GRU based survivable computing power scheduling in metro optical networks,” IEEE T. Netw. Serv. Man. (2023). [CrossRef]
13. B. Vass, J. Tapolcai, Z. Heszberger, et al., “Probabilistic shared risk link groups modeling correlated resource failures caused by disasters,” IEEE J. Select. Areas Commun. 39(9), 2672–2687 (2021). [CrossRef]
14. J. Tapolcai, L. Rónyai, B. Vass, et al., “List of shared risk link groups representing regional failures with limited size,” In IEEE INFOCOM 2017-IEEE Conference on Computer Communications,1–9 (2017).
15. Q. Yao, H. Yang, B. Bao, et al., “SNR re-verification-based routing, band, modulation, and spectrum assignment in hybrid C-C + L optical networks,” J. Lightwave Technol. 40(11), 3456–3469 (2022). [CrossRef]
16. H. Yang, Q. Yao, B. Bao, et al., “Multi-associated parameters aggregation-based routing and resources allocation in multi-core elastic optical networks,” IEEE/ACM Trans. Netw. 30(5), 2145–2157 (2022). [CrossRef]
17. H. Yang, J. Yuan, C. Li, et al., “BrainIoT: brain-like productive services provisioning with federated learning in industrial IoT,” IEEE Internet. Things. 9(3), 2014–2024 (2022). [CrossRef]
18. J. Zhang and Z. Jia, “Coherent Passive Optical Networks for 100 G/λ-and-Beyond Fiber Access: Recent Progress and Outlook,” IEEE Network 36(2), 116–123 (2022). [CrossRef]
19. J. S. Wey and J. Zhang, “Passive Optical Networks for 5 G Transport: Technology and Standards,” J. Lightwave Technol. 37(12), 2830–2837 (2019). [CrossRef]
20. Q. Zhuge, X. Liu, Y. Zhang, et al., “Building a digital twin for intelligent optical networks [Invited Tutorial],” J. Opt. Commun. Netw. 15(8), C242–C262 (2023). [CrossRef]
21. A. M. Rizzo, L. Magri, D. Rutigliano, et al., “Known and unknown event detection in OTDR traces by deep learning networks,” Neural Comput. Appl. 34(22), 19655–19673 (2022). [CrossRef]
22. Y. Mao, I. Ashry, B. Wang, et al., “Sensing within the OTDR dead-zone using a two-mode fiber,” Opt. Lett. 45(11), 2969–2972 (2020). [CrossRef]
23. M. Nakagawa, M. Ohzeki, K. Takenaga, et al., “Novel Inter-Core Crosstalk Measurement Method Using a Loopback and Bidirectional OTDR Technique,” J. Lightwave Technol. 41(12), 3842–3848 (2023). [CrossRef]
24. H. Li, Q. Sun, T. Liu, et al., “Ultra-high sensitive quasi-distributed acoustic sensor based on coherent OTDR and cylindrical transducer,” J. Lightwave Technol. 38(4), 929–938 (2020). [CrossRef]
25. M. Ohashi, A. Nakamura, T. Oda, et al., “Measurement of fiber parameters of pure silica core fibers based on the OTDR technique,” Opt. Express 29(10), 15078–15088 (2021). [CrossRef]
26. F. Yu, S. Liu, L. Shao, et al., “Ultra-low sampling resolution technique for heterodyne phase-OTDR based distributed acoustic sensing,” Opt. Lett. 47(14), 3379–3382 (2022). [CrossRef]
27. D.F. Kandamali, X. Cao, M. Tian, et al., “Machine learning methods for identification and classification of events in ϕ-OTDR systems: a review,” Appl. Opt. 61(11), 2975–2997 (2022). [CrossRef]
28. Q. Cui and S. Yin, “Automatic Identification Technology of Optical Fiber based on Genetic Neural Network Algorithm,” 2022 IEEE 6th Advanced Information Technology, Electronic and Automation Control Conference (IAEAC). IEEE, (2022). [CrossRef]
29. T. J. Xia, G. A. Wellbrock, M. F. Huang, et al., “First proof that geographic location on deployed fiber cable can be determined by using OTDR distance based on distributed fiber optical sensing technology,” Optical Fiber Communication Conference. (OFC2020). [CrossRef]
30. Y. Shi, Y. Li, Y. Zhang, et al., “An easy access method for event recognition of Φ-OTDR sensing system based on transfer learning,” J. Lightwave Technol. 39(13), 4548–4555 (2021). [CrossRef]
31. K. Abdelli, J. Y. Cho, F. Azendorf, et al., “Machine-learning-based anomaly detection in optical fiber monitoring,” J. Opt. Commun. Netw. 14(5), 365–375 (2022). [CrossRef]
32. K. Abdelli, H. Grießer, C. Tropschug, et al., “A BiLSTM-CNN based multitask learning approach for fiber fault diagnosis,” Optical Fiber Communication Conference. (OFC2021). [CrossRef]
33. K. Abdelli, H. Grießer, C. Tropschug, et al., “Optical fiber fault detection and localization in a noisy OTDR trace based on denoising convolutional autoencoder and bidirectional long short-term memory,” J. Lightwave Technol. 40(8), 2254–2264 (2022). [CrossRef]
34. C. Zhang, D. Wang, J. Jia, et al., “Potential failure cause identification for optical networks using deep learning with an attention mechanism,” J. Opt. Commun. Netw. 14(2), A122–A133 (2022). [CrossRef]
35. Margaret, “Basics of OTDR (Optical Time-Domain Reflectometer),” FS Community, (2018), https://community.fs.com/blog/understanding-otdr-dead-zone-specifications.html
36. W. Leo, Breiman Adele, Cutler Andy, et al., “RandomForest,” (2015).
37. Q. Yao, H. Yang, C. Li, et al. “Federated transfer learning framework for heterogeneous edge IoT networks,” China Commun. (2023).
38. A. Yu, H. Yang, C. Feng, et al. “Socially-aware traffic scheduling for edge-assisted metaverse by deep reinforcement learning,” IEEE Network (2023). [CrossRef]
