External Damage Prevention and Status Monitoring System for HVDC Submarine Cables Based on Distributed Optical Sensing

DAS systems are based on phase-sensitive optical time-domain reflectometry (Φ-OTDR) [9,10]. When short optical pulses are launched into a single-mode optical fiber, Rayleigh backscattering occurs owing to the intrinsic inhomogeneities within the fiber. When external vibrations or acoustic signals are applied around the fiber, they induce slight perturbations in the physical characteristics of the fiber, which in turn result in phase changes in the backscattered light. By analyzing these phase variations, information such as the location, amplitude, and frequency of vibrations can be obtained [11].

Unlike conventional OTDR, DAS enables dynamic sensing and is capable of detecting vibrations and acoustic signals in real-time. The spatial resolution is within approximately 10 m and monitoring is possible over distances of up to several tens of kilometers.

Conventional technologies for submarine cable monitoring can be broadly categorized into underwater inspection, electrical monitoring, OTDR-based static condition monitoring, and external surveillance systems. Underwater inspection involves the periodic assessment of cable conditions through direct examination using divers or Remotely Operated Vehicles (ROV). Electrical monitoring detects anomalies by calculating locations using current sensors, such as Rogowski coils, which are typically installed at substation terminals. OTDR-based static condition monitoring is a technique for detecting connection points, signal losses, and cable breakages in optical fibers. However, because it relies on one-time measurements, it has limitations in detecting dynamic changes. External surveillance systems monitor vessel movements around submarine cables using radar or cameras. However, they are limited because they cannot directly monitor the condition of the cables.

Compared with conventional HVDC cable-monitoring techniques, DAS offers the following advantages for submarine cable monitoring.

1. Real-time continuous monitoring: Enables immediate threat detection by continuously monitoring the entire length of cables in real-time.

2. Non-intrusive use of existing infrastructure: Reduces installation and maintenance costs by utilizing existing optical fiber cables as sensors without requiring additional equipment.

3. High-sensitivity and precise localization: Capable of detecting subtle vibrations and noise, and accurately identifying the event location with meter-level resolution.

4. Comprehensive data analysis: Generates vast amounts of data that can be analyzed using learning-based algorithms to improve the accuracy of threat detection.

The HVDC submarine cable system for external fault prevention and condition monitoring comprises the following hardware components (Fig. 1 (a)).

1. DAS (Interrogator) unit: Connected to the optical fiber to acquire and store distributed acoustic/vibration data.

2. NAS (Storage): Stores the raw data acquired by the DAS unit and supplies the data to the analysis module.

3. Analysis/Web server: Loads data from the NAS, performs analysis, stores event results in the database, and includes a web server to operate the user interface (Web UI).

The software primarily comprises two components (Fig. 1(b) and (c)).

Fig. 1.

System Overview (a) Hardware configuration, (b) DAS HMI, (c) Web-based User Interface.

1. DAS HMI(Human-Machine Interface): A software interface for configuring the measurement conditions of the DAS equipment, checking raw data, and monitoring system status. It provides a UI designed for technical personnel to operate equipment.

2. User interface (Web UI): This is a user interface that displays the monitored event information and visualizes detected events.

The system was developed through the following procedure:

1. Development of DAS-based data acquisition and classification technology

• •Long-term data collection and labeling for a 50 km section of the HVDC submarine cable.
• •Classification of normal/abnormal states and analysis of statistical characteristics.

2. Development and verification of data analysis algorithms

• •Development of data analysis algorithms for abnormal state diagnosis.
• •Verification of analysis algorithms using collected data.
• •Real-time data analysis and validation of their applicability.

3. Establishment of a real-time monitoring system

• •Construction of a real-time monitoring system for submarine cables between Jindo and Seojeju converter stations.
• •Integration with vessel information systems such as Vessel Pass (V-PASS), Automatic Identification System (AIS), and radar monitoring.
• •Improvement and verification of real-time monitoring accuracy.

A data collection system was established to acquire long-term data using a DAS system applied to the HVDC submarine cables. A management system was developed to systematically store and manage raw waterfall and fast Fourier transform (FFT) data, enabling long-term trend analysis without data loss.

Because DAS data collected from submarine cable monitoring exhibit significant variations in sensitivity and environmental noise depending on the location, a normalization process is essential for effective analysis [11,12]. In this study, various normalization techniques were reviewed, and the optimal method was selected and applied.

First, several applicable normalization methods, such as min-max, Z-score, and exponential normalization, divided by the average and median, were comparatively analyzed. Min-max normalization is a technique that scales data to a specific range (typically 0 to 1), which has the advantage of placing all features on the same scale. Z-score normalization transforms the data into a normal distribution with a mean of 0 and standard deviation of 1 by subtracting the mean and dividing by the standard deviation for each data point. In addition, other methods, such as exponential normalization and normalization by dividing by the mean or median, were examined.

Based on the results of the comparative analysis, the method using Z-score normalization followed by outlier detection based on a 3-sigma (three times the standard deviation) threshold was found to be the most effective (Fig. 2). This method enables the detection of outliers that are difficult to identify using simple thresholding by leveraging statistical occurrence probabilities, thereby allowing consistent anomaly detection across diverse environments. By applying normalization techniques, the data distribution was uniformly transformed, which significantly improved the outlier detection performance.

Fig. 2.

Z-score normalization that enables statistical probabilitybased detection of outliers beyond simple thresholding capabilities, where the x- and y-axes indicate time sample index and signal amplitude, respectively

Additionally, in this study, a signal extraction technique based on DbMN was developed and applied. This technique amplifies the variation in vibration intensity compared with the original data distribution by dividing the signal by its mean, thereby enhancing the detectability of anomalies. By applying the DbMN-based normalization and anomaly detection algorithm, vibration signals were binarized using a threshold value of 3, and the vibration signal density within each crop was evaluated based on a vibration ratio threshold of 0.01. The detected crops were then grouped into a single event using an algorithm that integrates them based on distance and time ranges.

To ensure optimal performance, comparative experiments were conducted using various threshold values (3, 5, and 10) (Fig. 3). The results demonstrated that Threshold 3 most effectively balanced false positives and false negatives while accurately detecting anomalous signals. This data preprocessing and normalization procedure serves as the input for the subsequent anomaly detection algorithm and plays a critical role in determining the overall detection performance of the system.

Fig. 3.

Signal extraction based on Divided by Mean Normalization (DbMN), where the x-axis is the distance from 0 to 51,353 m and the y-axis is Datetime [dd HH:MM].

To detect abnormal conditions, several algorithms were developed and applied.

1. Trendline-based anomaly detection: A trendline-detection method using a moving average was implemented. A technique was developed to detect anomalous signals based on deviations from the trend line, which cannot be easily identified using a simple threshold owing to the presence of continuous background signals (Fig. 4).

Fig. 4.

Trend line detection using the moving average technique, where the x-axis indicates the time sample index and y-axis is the signal amplitude.

2. Unsupervised anomaly detection algorithm: The isolation forest algorithm was applied utilizing the property that anomalies are isolated more quickly than normal instances through random partitioning. In addition, a robust random cut forest (RRCF) algorithm was applied for anomaly detection in the real-time streaming data (Fig. 5).

Fig. 5.

Anomaly detection algorithm based on isolation forest and RRCF, where the x-axes of the two left figures represent time sample index and y-axes are the signal amplitude and anomaly scores.

3. Autoencoder-based anomaly detection: An autoencoder-based technique that learns the pattern of normal data and identifies data with large reconstruction errors as anomalies was applied.

Excavator breaker signals generated during construction in the underground section were detected, the locations of the peak signal intensity were identified, and the corresponding data were collected. The collected data were analyzed in terms of the amplitude and frequency characteristics to identify the features of the excavator breaker signals. The signals were compiled into a database to enable their identification in the event of similar patterns.

Hammering was performed at arbitrary locations, and the detected signal data were collected. Transient hammering signals were confirmed to be insufficient for the effective mapping of the underground section, highlighting the need for a signal source capable of providing continuous and rapid signals. Vehicle signals were also measured at an intersection while driving straight. These signals were identified as originating from large or high-speed vehicles. Their locations were confirmed through comparison with the cable installation map.

By applying the Z-score normalization technique, the outliers that were difficult to distinguish using simple thresholds were effectively detected by calculating their statistical occurrence probabilities. Through the application of normalization techniques, the data distribution was transformed to become more uniform, thereby enhancing the outlier detection performance.

Applying the DbMN-based signal extraction method amplified the fluctuations in the vibration intensity and emphasized the outliers more effectively than in the original data distribution. This enabled the successful implementation of a method that binarized vibration signals using a threshold of three and determined the vibration signal density within a crop based on a vibration ratio of 0.01.

Fig. 7 illustrates the implementation results of the process from waterfall data acquisition → normalization and binarization for abnormal vibration detection → abnormal vibration crop detection and continuous pattern grouping, confirming the successful detection of persistent vibrations around 12:33 on July 11, 2024, at locations of 2 km, 5.6 km, 9 km, and 10 km. This confirmed the effectiveness of the proposed algorithm.

Fig. 7.

Example of practical application of signal processing techniques: (a) Waterfall data, (b) Crop & bin data, (c) Merged crop data, where the x-axis indicates distance from 0 to 50,000 m and y-axis is Datetime [dd HH:MM].

The monitoring system provides a detected event-list display function and event-location visualization features (side and top views). In addition, functions for removing events from the list upon confirmation and manually sending SMS alerts were implemented. A live/history-mode toggle function was designed to enable real-time monitoring and retrospective history retrieval.

An SMS notification service was implemented to automatically send alerts to onsite personnel when an event was detected. A protocol specification document for communication with the existing SMS-sending server was developed, and an integration module based on TCP/IP communication was implemented. The packet definitions and communication flow between the Server (Rx), Client (Tx), and Key Generator were designed and constructed to establish a reliable communication framework through the defined packet structures, including SMS_INFO, REF_INFO, RTN_INFO, KEY_INFO, and RST_INFO.

A system architecture was designed to collect vessel information from the ship information server and provide it to the DAS via TCP/IP. Depending on the presence of irregularities, data storage or deletion was managed, and a system was developed to monitor acoustic, vibration, and frequency data using abnormal detection algorithms, thereby generating alarms accordingly.

To verify the event-localization accuracy, the error between the detected event location and the actual signal injection location was calculated, and it was verified whether the error was within the target localization accuracy of ±10 m. As a result of performance testing based on arbitrary connection point locations, the average and maximum errors at Test Site 1 were 6.96 m and 9.31 m, respectively, whereas those at Test Site 2 were 5.25 m and 7.98 m, respectively, thereby achieving the target accuracy of within ±10 m.

To verify the event detection speed target of less than 15 s, the elapsed time between event occurrence and database entry notification to the user was analyzed and compared. Twelve repeated event tests were conducted, resulting in a minimum detection time of 1.743 s, a maximum of 10.739 s, and an average of 3.795 s, thereby meeting the target of < 15 s.

To verify an event classification accuracy target of > 90%, anomaly detection tests were conducted to identify abnormal patterns within the collected data. A confusion matrix was generated by comparing the classification results of the AI-based algorithm with the labeled event results. Fig. 8 shows the confusion matrix results of the event classification algorithm. The classified events included anchoring, dragging, earthquakes, and vessels. The labels for each event type are Drop (DP), Drag (DR), Earthquake (EQ) and Ship (SH).

Fig. 8.

Confusion matrix results of the event recognition and classification

algorithm.For validation, the data labels were compared with the classification results to generate a confusion matrix, and the accuracy was calculated by combining the number of correct predictions along the diagonal with the number of misclassifications in the off-diagonal elements. For this evaluation, accuracy was calculated by dividing the number of correct predictions by the total number of cases.

The validation results showed an average classification accuracy of 97.45%, exceeding the target of 90%. This indicates that the developed system can distinguish between normal and abnormal signals with high accuracy.

To verify that the event classification precision target was greater than 80%, the confusion matrix was converted into a True/False table based on each event and then evaluated. True/False conversion simplifies classification results by considering all other events as false, resulting in only true or false outcomes. Precision indicates the proportion of signals classified as anomalous by the system.

For verification, the precision of each label was calculated based on the confusion matrix using the formula TP/(TP + FP), where TP is the true positive, TN is the true negative, FP is the false positive, and FN is the false negative. The verification results showed that the precision of each label for abnormal signal detection exceeded 90.76%, achieving the target of 80% or higher. This indicates that the developed algorithm can effectively reduce the number of false positives.

To verify an event classification recall target of at least 80%, the confusion matrix was converted into a True/False table for evaluation. Recall indicates the proportion of actual abnormal signals that the system correctly identifies as abnormal.

The validation method involved calculating the recall for each label based on the confusion matrix, using the formula TP/(TP + FN). The validation results showed that the recall for each label of the anomaly detection was 92.48%, exceeding the target of 80%. This demonstrates the strong capability of detecting actual abnormal signals without missing occurrences.

Summarizing these performance evaluation results, it was confirmed that the developed system satisfied the target criteria for all assessment metrics, including event location accuracy, detection speed, classification accuracy, precision, and recall, thereby verifying its applicability in practical operational environments.