A long-term operation and maintenance monitoring method and system for submarine cables

Abstract

The present application belongs to the technical field of cable operation and maintenance monitoring, and particularly relates to a long-term operation and maintenance monitoring method and system for submarine cables, which constructs a multi-parameter submarine cable global original data acquisition system through the combined arrangement of multiple sensing hardware, global time sequence synchronization and hardware anti-interference packaging; adopts a double-layer differentiated noise reduction architecture combined with feature extraction and database comparison technology to filter out marine random noise and artificial equipment fixed interference in layers, complete the extraction and identification of weak partial discharge signals; fuses multi-source monitoring data and introduces dynamic correction parameters and deep learning models to adapt to marine working conditions; fuses frequency domain detection technology and traditional time domain monitoring means to realize full-type submarine cable fault monitoring verification through photoelectric bidirectional verification; relies on edge cloud collaborative architecture to complete adaptive adjustment of monitoring threshold working conditions, quantifies insulation aging trend, predicts fault risk in advance, and constructs a submarine cable intelligent monitoring and operation and maintenance closed-loop system.

Description

Technical Field

[0001] This invention belongs to the field of cable operation and maintenance monitoring technology, and in particular relates to a method and system for long-term operation and maintenance monitoring of submarine cables. Background Technology

[0002] In marine power transmission projects, submarine cables are the core carriers for cross-sea power transmission and grid connection of offshore new energy sources. Lay in complex underwater environments for extended periods, they are susceptible to multiple factors such as seawater corrosion, tidal erosion, seabed friction, and marine electromagnetic interference, gradually leading to potential problems like insulation aging, partial discharge, and mechanical damage. Traditional submarine cable monitoring methods often employ single-time-domain signal acquisition, limiting monitoring dimensions and making it difficult to distinguish between environmental noise and actual fault signals. They also lack the ability to identify slowly changing defects such as insulation degradation and latent discharge. Furthermore, the commonly used fixed threshold values ​​are unsuitable for adapting to time-varying marine conditions, resulting in significant false alarms and missed alarms, hindering accurate monitoring of all types of faults across all time periods.

[0003] Current submarine cable operation and maintenance models mostly remain at the stage of reactive response after a fault occurs, lacking the ability to deeply mine monitoring data, fuse features, and predict trends. Various types of sensing data and operational data are independent of each other, failing to form a multi-source information complementary verification mechanism. In addition, existing monitoring algorithm models rely on historical fixed samples, which are prone to model degradation and operational condition adaptation deviations after long-term operation. Furthermore, remote monitoring cannot fully match the actual underwater physical conditions, and there is a lack of on-site calibration and model iterative optimization, making it difficult to support long-term, intelligent operation and maintenance management of submarine cables throughout their entire life cycle. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, an embodiment of the present invention provides a method for long-term operation and maintenance monitoring of submarine cables, the method comprising: By combining and deploying multiple types of sensing hardware, implementing full-domain time synchronization and hardware anti-interference encapsulation, a multi-parameter submarine cable full-domain raw data acquisition system is constructed. A dual-layer differentiated noise reduction architecture is adopted, which combines feature extraction and database comparison technology to filter out random ocean noise and fixed interference from artificial equipment in layers, thereby completing the extraction and identification of weak partial discharge signals. By integrating multi-source monitoring data and introducing dynamic correction parameters and deep learning models, it can adaptively adapt to marine conditions. By integrating frequency domain detection technology with traditional time domain monitoring methods, all types of submarine cable fault monitoring and verification can be achieved through photoelectric bidirectional verification. Based on the edge-cloud collaborative architecture, the monitoring threshold and working conditions are adaptively adjusted to quantify the insulation aging trend, predict fault risks in advance, and build a closed-loop system for intelligent monitoring and maintenance of submarine cables. Establish a dual mechanism of data iteration and underwater field calibration to continuously optimize the algorithm model and correct operating condition deviations.

[0005] Furthermore, embodiments of the present invention also provide a long-term operation and maintenance monitoring system for submarine cables, characterized in that it includes: A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the above-described long-term operation and maintenance monitoring method for submarine cables by executing the machine-executable instructions.

[0006] In another aspect, embodiments of the present invention also provide a computer program product, the computer program product including machine-executable instructions, the machine-executable instructions being stored in a computer-readable storage medium, the processor of a computer device reading the machine-executable instructions from the computer-readable storage medium, the processor executing the machine-executable instructions, causing the computer device to execute the above-described long-term operation and maintenance monitoring method for submarine cables.

[0007] Based on the above, a multi-parameter, full-domain data acquisition system was constructed. Combined with dual-layer noise reduction and multi-scale feature extraction techniques, it effectively filters out random marine noise and artificial electromagnetic interference, achieving accurate extraction of weak partial discharge signals. Through time-frequency fusion monitoring and photoelectric bidirectional cross-verification, it addresses both transient explicit faults and latent slow-changing faults, significantly reducing the probability of false fault identification. Relying on a deep learning feature fusion model, it completes in-depth mining of multi-source heterogeneous data, fully capturing the implicit correlations and coupling patterns between data, significantly improving the accuracy and comprehensiveness of submarine cable fault identification under complex marine conditions.

[0008] This invention leverages an edge-cloud collaborative architecture to achieve dynamic adaptive adjustment of monitoring thresholds. Combined with time-series modeling, it quantifies insulation aging trends and predicts fault risks in advance. This, along with a tiered and categorized maintenance strategy, forms a complete intelligent maintenance closed loop. Simultaneously, it establishes a dual optimization mechanism of cloud data iteration and underwater field calibration to continuously correct algorithm deviations and update model parameters, effectively avoiding model degradation and perception blind spots. The entire solution can stably adapt to the ever-changing marine environment, enabling submarine cables to shift from passive fault handling to proactive risk prevention and long-term intelligent maintenance throughout their entire lifecycle. Attached Figure Description

[0009] Figure 1 This is a schematic diagram of the execution flow of the long-term operation and maintenance monitoring method for submarine cables provided in this embodiment of the invention.

[0010] Figure 2 This is a schematic diagram of exemplary hardware and software components of the long-term operation and maintenance monitoring system for submarine cables provided in this embodiment of the invention. Detailed Implementation

[0011] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1This is a flowchart illustrating a long-term operation and maintenance monitoring method for submarine cables provided in one embodiment of the present invention. The following is a detailed description of this long-term operation and maintenance monitoring method for submarine cables.

[0012] Step S110: Construct a multi-parameter submarine cable full-domain raw data acquisition system by combining and deploying multiple types of sensing hardware, implementing full-domain time synchronization and hardware anti-interference encapsulation; In its specific implementation, this step addresses the industry challenges of single-dimensional submarine cable monitoring data, poor synchronization, severe interference, and low reliability of raw data under complex marine conditions. It establishes a comprehensive architecture combining multi-type sensing hardware deployment, unified time-series calibration across the entire domain, and hardware-level anti-interference protection to build a multi-parameter, full-coverage, and high-purity raw data acquisition system for submarine cables. During implementation, the deployment logic of various sensing hardware types—mechanical, electrical, electromagnetic, and environmental—is coordinated. Hardware networking achieves full coverage based on differences in submarine cable laying environments, abandoning the traditional fragmented acquisition mode of independent single-sensor acquisition, inconsistent timing, and weak anti-interference capabilities. Simultaneously, a holistic system design is implemented from multiple dimensions, including hardware deployment, time synchronization, electromagnetic shielding, and corrosion protection and sealing. This ensures the synchronous acquisition and complete recording of submarine cable operation status data, marine condition data, and environmental interference data from the data source. This provides a high-quality raw data foundation with time-aligned data, complete dimensions, and controllable noise for subsequent signal noise reduction, feature extraction, and fault identification model training, thereby improving the integrity and reliability of submarine cable monitoring data from the source.

[0013] Step S111: Based on the laying mileage, laying water depth, seabed geological conditions and historical fault distribution characteristics of the submarine cable, a multi-sensor hardware fusion networking method is adopted to implement differentiated segmented deployment of the entire submarine cable. Integrated comprehensive sensing units are uniformly distributed in the conventional straight submarine cable sections, and sensing nodes are densely deployed in weak and high-risk sections such as the submarine cable landing section, underwater intermediate joint, bend pipe laying section, reef friction section, and shallow sea tidal erosion section. In its implementation, this step abandons the traditional approach of uniformly deploying a single network along the entire submarine cable without considering operational risks. Instead, it implements differentiated and targeted network deployment of sensing nodes based on the actual laying conditions of the submarine cable. First, it comprehensively reviews the cable's mileage, water depth distribution, seabed geological structure characteristics, and historical high-incidence fault locations and defect distribution patterns, dividing the entire cable line into conventional, stable sections and high-risk, vulnerable sections. For conventional, straight sections with stable hydrology and no significant friction or erosion, an integrated sensing unit is deployed using a uniformly spaced, distributed deployment method to achieve basic full-area coverage. For high-risk, vulnerable sections such as cable landing sections, underwater intermediate joint locations, bends in the cable laying area, reef friction zones, and areas with severe shallow-sea tidal erosion, a denser deployment method is adopted to increase the density of sensing nodes, enhance the data acquisition accuracy and detail capture capabilities at vulnerable locations, and achieve a graded, zoned, and differentiated intelligent network across the entire line, balancing overall coverage with precise monitoring capabilities for key sections.

[0014] Step S112: Construct a multimodal integrated sensing hardware acquisition array, integrating distributed optical fiber sensing units, high-frequency current traveling wave acquisition terminals, and micro quantum magnetometers in a collaborative network. The distributed optical fiber sensing units are used to collect data on submarine cable temperature, cable deformation, axial strain, and vibration disturbances throughout the cable laying process. The high-frequency current traveling wave acquisition terminals installed close to the submarine cable's metal sheath and grounding end are used to capture line current mutations and fault traveling wave pulse signals at high frequency. The micro quantum magnetometers deployed close to the outer wall of the submarine cable are used to collect high-precision electric and magnetic field disturbances around the submarine cable and magnetic distortion signals induced by stray currents in seawater. In the specific implementation of this step, a multimodal integrated collaborative sensing hardware array is constructed to achieve synchronous acquisition of multi-dimensional information on the submarine cable's mechanics, electrical, and electromagnetic aspects. Distributed fiber optic sensing units are laid integrally with the submarine cable throughout its operation, closely fitting the cable body to continuously collect data on temperature changes, cable deformation, axial strain, and external vibration disturbances during cable operation, comprehensively capturing the cable's mechanical damage and stress changes. High-frequency current traveling wave acquisition terminals are installed at the cable's metal sheath and grounding end to capture real-time line current surge signals and fault traveling wave pulse signals using high-frequency sampling mode, accurately recording the transient characteristics of electrical faults. Simultaneously, miniature quantum magnetometers are deployed close to the outer wall of the cable, leveraging their high-precision sensing capabilities to collect real-time data on dynamic disturbances in the electric and magnetic fields around the cable, as well as magnetic distortion information induced by stray currents in seawater. Through the collaborative networking of these three types of hardware, comprehensive synchronous acquisition of multiple parameters (mechanical, electrical, and electromagnetic) is achieved, constructing a multi-dimensional integrated sensing and acquisition system.

[0015] Step S113: Establish a full-domain nanosecond-level timing synchronization system. Configure a high-precision satellite timing module at the land-based main control terminal to output a unified reference time signal. Synchronize the time to all underwater sensing terminals through the submarine cable communication core line. Each underwater terminal has a built-in clock calibration chip for real-time dynamic time synchronization. Control the time synchronization error of all nodes to the nanosecond level. Configure a transmission delay compensation algorithm to correct the timing offset in real time based on the node laying distance and signal transmission rate. In the specific implementation of this step, a high-precision time synchronization system at the nanosecond level across the entire area is established to completely solve the data timing deviation problems caused by time misalignment and transmission delay in traditional underwater terminal acquisition. A high-precision satellite timing module is deployed on the land-based main control platform to output a unified global reference time signal as the time reference for all sensing devices along the entire line. A dedicated timing link is built based on the communication core wires embedded in the submarine cable to synchronously transmit the reference time signal to all underwater sensing terminals. Each underwater terminal has a built-in high-precision clock calibration chip to perform real-time dynamic time calibration of its local clock, strictly controlling the time synchronization error of all acquisition nodes along the entire line to the nanosecond level. At the same time, a dedicated transmission delay compensation algorithm is configured for terminal devices with different laying distances. Based on the terminal laying mileage and signal transmission rate, the timing offset caused by signal transmission is calculated and corrected in real time to ensure that the timing of all multi-source heterogeneous monitoring data along the entire line is completely aligned.

[0016] Step S114: Implement full-dimensional hardware protection and anti-interference treatment for all underwater sensing terminals. Use a titanium alloy anti-corrosion shell with insulating and waterproof sealant for overall encapsulation. Block electromagnetic radiation interference from offshore wind power, ships, and platforms, as well as seawater stray current coupling interference through multiple layers of high permeability magnetic shielding and insulation isolation layers on the outside of the terminal. At the same time, adopt an integrated sealed shielded connector design for the equipment wiring ports and signal interfaces. In its implementation, this step addresses the challenges of complex marine electromagnetic interference, seawater corrosion, and stray current coupling interference by providing comprehensive hardware protection and anti-interference measures for all underwater sensing terminals. All underwater terminals utilize a high-strength titanium alloy corrosion-resistant shell, combined with insulating and waterproof sealant for integrated encapsulation. This effectively resists long-term seawater corrosion, underwater high-pressure penetration, and damage from marine organism attachment, ensuring stable long-term underwater operation. Multiple layers of high-permeability magnetic shielding and insulation are added externally to the terminals to physically block electromagnetic radiation interference from offshore wind power equipment, ship electrical systems, and offshore platform equipment, while also isolating electric field coupling interference from stray seawater currents. All device wiring ports and signal transmission interfaces employ an integrated sealed shielded connector design, eliminating electromagnetic leakage and signal crosstalk at interface locations. This comprehensive hardware-level suppression of external interference significantly improves the purity and stability of the original acquired signals.

[0017] Step S115: After completing the hardware network, power on and debug each sensing node, conduct signal sampling tests and sensitivity calibration, unify the sampling frequency, sampling accuracy, data storage format and transmission protocol of each terminal, and establish a real-time self-check and abnormal alarm mechanism for terminal operating voltage, temperature, signal strength and operating status. Finally, a full-coverage, multi-parameter, high-purity and high-synchronization raw data acquisition system for the entire submarine cable domain is constructed.

[0018] In the specific implementation of this step, after completing the deployment and anti-interference encapsulation of the entire sensing hardware network, full-point equipment debugging and standardized calibration are carried out. Power-on testing, signal sampling experiments, and equipment sensitivity calibration are performed on all sensing nodes along the entire submarine cable. The sampling frequency, sampling accuracy, data storage format, and communication transmission protocol of all terminals are standardized to eliminate data deviations caused by inconsistencies in the acquisition parameters of different hardware devices. Simultaneously, a terminal self-testing mechanism is established to monitor the operating voltage, operating temperature, output signal strength, and online operating status of each sensing terminal in real time. In case of abnormal voltage, signal attenuation, or equipment offline, an abnormal alarm is automatically triggered. Through comprehensive debugging, standardized calibration, and a real-time self-testing alarm mechanism, a standardized data acquisition system covering the entire submarine cable, synchronously acquiring multiple parameters, ensuring high signal purity, and providing stable and reliable operation is ultimately constructed.

[0019] Step S120: A dual-layer differentiated noise reduction architecture is adopted, combined with feature extraction and database comparison technology, to filter out marine random noise and fixed interference from artificial equipment in layers, thereby completing the extraction and identification of weak partial discharge signals; In the specific implementation of this step, addressing the challenges of weak partial discharge signals, complex noise interference, and difficulty in distinguishing between true and false signals under complex marine conditions, a signal purification and identification system combining two-layer differentiated noise reduction and intelligent feature comparison is constructed. Marine interference is categorized into two types: broadband random noise caused by ocean current temperature variations and narrowband fixed interference caused by artificial electrical equipment. A targeted layered noise reduction strategy is designed to avoid over-filtering of weak partial discharge signals using a single noise reduction method. First, an adaptive dynamic noise reduction algorithm is used to filter out irregular marine random noise. Then, a second-layer intelligent notch filtering mechanism precisely removes fixed narrowband interference from artificial equipment, achieving precise layered suppression of both types of interference. Based on the high-purity signal, refined feature extraction is performed, and intelligent comparison and identification are conducted using a pre-set massive sample database to accurately distinguish between real partial discharge signals and various types of interference pseudo-signals. This effectively solves the technical problems of weak partial discharge signals being easily submerged by noise and the difficulty in identifying latent defects using traditional methods, achieving accurate extraction and reliable identification of early-stage weak partial discharge signals.

[0020] Step S121: Based on massive submarine cable monitoring data, classify and define marine interference signals in advance. Divide marine interference signals into two types: broadband random Gaussian noise generated by ocean current disturbances and temperature fluctuations, and fixed-frequency narrowband electromagnetic interference generated by offshore wind power, ship electrical systems, and offshore platform power supply equipment. Set the partial discharge pulse signal of submarine cable as the target effective signal. Perform pre-amplification and normalization preprocessing on the high-frequency raw monitoring signals collected on site to unify the signal amplitude range and sampling scale. In this step, based on massive historical submarine cable monitoring data, precise classification and definition of marine interference types are completed, and a standardized signal differentiation benchmark is established. Marine interference signals are clearly divided into two categories: one is broadband random Gaussian noise generated by natural environmental factors such as ocean current disturbances and seawater temperature fluctuations, characterized by dispersed frequency bands, diffuse energy, and no fixed pattern; the other is narrowband fixed-frequency electromagnetic interference generated by offshore wind power, ships, and offshore platform power supply equipment, characterized by fixed frequency and concentrated energy. Simultaneously, short-time pulse signals of partial discharge in submarine cable insulation are set as the target valid signals. High-frequency raw monitoring signals acquired in real-time on-site undergo unified pre-amplification and amplitude normalization pre-processing to standardize the amplitude range and sampling scale of all signals, eliminating analytical errors caused by uneven raw signal strength and inconsistent scales.

[0021] Step S122: An improved empirical wavelet transform noise reduction mechanism is adopted. In view of the broadband and irregular random noise characteristics of the marine environment, adaptive dynamic frequency band segmentation and multi-mode decomposition are implemented on the preprocessed high-frequency monitoring signal. Based on the real-time frequency domain distribution characteristics, noise frequency bands and effective signal frequency bands are autonomously distinguished. Through an iterative screening mechanism, marine random noise components without fixed shape are accurately removed, while effective characteristic signals such as partial discharge exclusive short-time pulses and sudden attenuation are adaptively retained. In this specific implementation step, an improved empirical wavelet transform denoising mechanism is employed to specifically suppress broadband random noise in the ocean, adapting to the irregular, wide-band dispersion, and unstable energy characteristics of ocean noise. Instead of a fixed frequency band segmentation mode, the preprocessed standardized high-frequency monitoring signal utilizes real-time frequency domain characteristics to achieve adaptive frequency band division and multi-mode decomposition. Through sequential mode feature comparison, the system accurately distinguishes between chaotic and irregular ocean random noise modes and effective partial discharge signal modes with abrupt attenuation characteristics. Simultaneously, an iterative screening and purification logic is introduced to repeatedly decompose and re-examine the mixed mode components in multiple rounds, gradually removing residual scattered noise components while preserving the waveform details and amplitude characteristics of weak partial discharge pulses throughout. Finally, the filtered effective mode components are reconstructed to restore the complete time-series signal, thoroughly filtering out broadband random noise in the ocean while preserving the weak effective partial discharge signal to the greatest extent possible, achieving the first layer of precise signal purification.

[0022] Step S1221: Read the high-frequency original signal after pre-amplification and normalization, perform a global fast Fourier transform on the entire signal, solve for the signal spectrum, energy spectrum and power spectrum, statistically analyze the energy ratio and amplitude fluctuation range of different frequency intervals, capture the distribution characteristics of broadband dispersion and energy diffusion of ocean random noise, and mark the frequency domain intervals where partial discharge pulse signals are concentrated, and establish the frequency domain characteristic judgment benchmark of signal and noise. In this step, the high-frequency raw monitoring signal, after pre-amplification and normalization, is first read. A full-domain Fast Fourier Transform is then performed on the entire time-series signal to fully determine its spectral, energy, and power distributions. By statistically analyzing the signal energy proportions, amplitude fluctuation ranges, and energy concentration characteristics across different frequency ranges, the typical distribution characteristics of broadband dispersion, energy diffusion, and the absence of obvious peaks in marine random noise are accurately captured. Simultaneously, through frequency domain peak localization, the specific frequency domain intervals where the partial discharge pulse signal energy is concentrated are marked, clarifying the differentiated characteristics between effective signals and environmental noise in the frequency domain, and establishing a stable and reusable frequency domain judgment benchmark for signals and noise.

[0023] Step S1222: Run the improved empirical wavelet transform algorithm to perform multi-mode decomposition on the signal from low frequency to high frequency. Based on the real-time spectrum energy inflection point and amplitude change point, dynamically and adaptively complete the frequency band segmentation so that the frequency range and bandwidth of each sub-band are adapted to the current signal frequency domain characteristics, and generate multi-order adaptive mode components. In this specific implementation step, an improved empirical wavelet transform algorithm is run based on the acquired real-time spectral characteristics to achieve adaptive multi-level frequency band segmentation and mode decomposition of the signal. The algorithm traverses and analyzes segment by segment from the low-frequency range to the high-frequency range, automatically identifying the signal's frequency domain structure based on the energy inflection points and amplitude abrupt change points of the real-time spectral curve, and dynamically and adaptively completing the full-band segmentation. This ensures that the bandwidth range and frequency interval of each sub-band perfectly match the actual frequency domain distribution characteristics of the current signal, without the need for manually preset fixed parameters. Through adaptive frequency band segmentation, multi-order refined modal components are generated, corresponding to the components of different frequencies and energy characteristics of the signal, achieving refined layered decomposition of the original mixed signal.

[0024] Step S1223: Extract four types of characteristic parameters for each modal component: time-domain waveform, energy concentration, waveform continuity, and amplitude change rate. Combine these with preset judgment rules to complete modal attribute identification. Modes with dispersed energy, messy waveforms, and no obvious abrupt changes are identified as ocean random noise components. Modes with short-term amplitude jumps and sudden and attenuated waveform characteristics are identified as effective partial discharge signal components. In this step, multi-dimensional feature parameter extraction and attribute identification are performed on each of the decomposed modal components. Four core feature parameters are extracted for each mode: temporal waveform regularity, energy spatial concentration, waveform continuity, and amplitude change rate, to construct a basis for modal attribute discrimination. Modes with dispersed energy, disordered waveforms, no obvious amplitude abrupt changes, and no concentrated energy range are uniformly classified as marine random noise components. Modes with short-term amplitude jump characteristics, obvious sudden waveforms and exponential decay patterns, and localized energy concentration are classified as effective partial discharge signal components. By replacing single threshold judgment with a multi-feature joint discrimination method, the ability to distinguish between weak partial discharge modes and random noise modes under complex sea conditions is effectively improved, avoiding the misjudgment and rejection of weak effective signals and ensuring the accuracy of signal identification.

[0025] Step S1224: Start the iterative screening and purification process, remove the pure noise mode components identified in the first round, repeatedly carry out multi-mode decomposition and feature re-examination on the remaining mixed mode components, iteratively investigate and remove residual scattered random noise components, lock the weak pulse waveform characteristics throughout the process, and ensure the effective signal amplitude and waveform integrity. In the specific implementation of this step, an iterative screening and purification process is initiated to achieve complete noise removal and preservation of effective signal fidelity. First, the pure noise mode components accurately identified in the first round are eliminated, retaining the remaining mixed mode components containing effective signals. Adaptive multi-layer frequency band decomposition and mode feature re-examination are repeatedly performed on the mixed mode components, iteratively identifying scattered random noise components hidden in the details of weak signals, and gradually removing residual noise layer by layer. Throughout the process, typical waveform characteristics of short-duration partial discharge pulses and sudden attenuation are locked, prioritizing the retention of low-energy, small-abrupt latent partial discharge details to avoid losing weak effective features during the iteration process. Through multiple rounds of iterative screening, irregular random noise in the ocean is completely eliminated, while preserving the waveform integrity and feature authenticity of weak partial discharge signals to the greatest extent possible, achieving deep purification of complex random noise.

[0026] Step S1225: Reconstruct all valid modal components retained after multiple rounds of iterative screening to restore a clean electrical signal with complete timing and waveform characteristics, accurately suppress ocean broadband random noise, and output the preliminarily purified signal to the subsequent signal processing stage.

[0027] In this step, all effective modal components retained after multiple rounds of iterative screening are reconstructed and synthesized to recover a clean monitoring electrical signal with continuous timing, complete waveform, and faithful characteristics. During reconstruction, each effective mode is superimposed strictly according to the original signal's timing logic, without altering the amplitude scale, abrupt change timing, attenuation pattern, or timing distribution characteristics of the effective signal; only random noise interference is removed. This processing method thoroughly solves the problem of irregular noise in the ocean broadband masking weak partial discharge signals, achieving precise suppression of the first layer of environmental random noise and outputting a preliminary purified signal with a high signal-to-noise ratio.

[0028] Step S123: Based on the parameterized intelligent notch filter noise reduction architecture, for fixed artificial electromagnetic narrowband interference generated by offshore wind power, ships, and offshore platforms, the main frequency characteristics of the interference are dynamically captured through real-time spectrum monitoring, and the bandwidth, attenuation coefficient and suppression depth of the notch filter are adaptively generated and iteratively optimized in real time. In the specific implementation of this step, a parameter-adaptive intelligent notch filter noise reduction architecture is constructed to address the fixed narrowband artificial electromagnetic interference generated by the operation of electrical equipment in offshore wind power, ships, and offshore platforms, achieving targeted and precise interference suppression. Traditional fixed-parameter notch filters cannot adapt to complex scenarios with slight frequency drift and multiple frequency superpositions of interference in the field. This technology dynamically captures the dominant frequency, energy intensity, and frequency band range of artificial interference through real-time spectrum tracking technology, and adaptively iteratively updates the notch filter's center frequency, suppression bandwidth, attenuation coefficient, and suppression depth. It adaptively widens the suppression range for interference superimposed by multiple devices and adaptively reduces the suppression intensity for stable single interference, accurately matching the field interference characteristics throughout the process. While completely eliminating artificial narrowband interference, it fully preserves the effective signal frequency band and waveform details of partial discharge pulses, achieving a second layer of differentiated and precise noise reduction, further improving the purity of submarine cable monitoring signals.

[0029] Step S1231: Real-time spectrum acquisition and continuous spectrum tracking are carried out on the purified signal after the first layer of marine random noise reduction. The signal is analyzed frame by frame through short-time Fourier transform. The peak point, main frequency position and peak energy distribution characteristics of the signal frequency domain are extracted in real time. The narrowband fixed frequency interference components generated by the operation of offshore wind power, ships and offshore platform electrical equipment are continuously monitored. The set of dynamic artificial interference main frequencies under the on-site working conditions is locked, and the stable power frequency interference and the slightly drifting equipment operation interference frequency are distinguished. In this step, the purified signal after the first layer of random noise reduction undergoes frame-by-frame continuous spectrum tracking and analysis. Short-time Fourier transform is used to perform sliding-window frame-by-frame spectrum analysis on the time-series signal, extracting the frequency domain peak points, dominant frequency positions, peak energy magnitude, and frequency band distribution range of each frame in real time. Narrowband fixed-frequency interference components generated by offshore wind power, ships, and offshore platform power supply equipment are continuously and dynamically monitored, and the dynamic set of artificial interference dominant frequencies is summarized in real time. Simultaneously, stable power frequency interference is distinguished from slightly drifting equipment operation interference frequencies, accurately identifying the dynamic characteristics of artificial narrowband interference and grasping the superposition state and fluctuation patterns of interference in real time.

[0030] Step S1232: Based on the real-time captured interference main frequency characteristics, start the intelligent notch filter parameter adaptive generation logic. According to the peak energy, frequency concentration, and interference superposition intensity of the interference main frequency, automatically initialize the notch filter center frequency, basic bandwidth, initial attenuation coefficient and suppression depth parameters to generate an initial parameterized notch filter model that adapts to the current field interference characteristics. In the specific implementation of this step, the traditional passive notch filter suppression mode with fixed frequency and fixed bandwidth is abandoned. Instead, an intelligent parameter adaptive generation logic based on real-time interference characteristics is established. Based on the real-time captured peak energy of the main frequency of artificial interference, frequency concentration, multi-signal superposition strength, and interference fluctuation state, the notch filter's center frequency, basic bandwidth, attenuation coefficient, and depth suppression parameters are automatically matched and initialized to generate a parameterized notch filter model that is fully adapted to the current field interference conditions. For high-energy concentrated narrowband interference, the suppression depth is automatically increased; for multi-frequency superposition interference, the initial suppression bandwidth is automatically widened; and for weak interference conditions, the attenuation intensity is automatically reduced. This achieves precise adaptation of the notch filter model parameters to the field interference state, ensuring targeted and efficient interference suppression.

[0031] Step S1233: Construct a dynamic fine-tuning mechanism for notch filter parameters in conjunction with operating conditions, continuously collect subsequent spectrum data in real time, track the slight drift of the main interference frequency, the changes of multiple frequency superposition and the fluctuation of interference energy, dynamically correct the notch filter parameters frame by frame, adaptively widen the suppression bandwidth for multi-device superposition interference scenarios, and adaptively reduce the suppression depth for weak interference high stable frequency scenarios. In this step, a dynamic fine-tuning and iterative mechanism for notch filter parameters is constructed to achieve real-time optimization of the notch filter effect. Based on the initial notch filter model, subsequent real-time spectrum data is continuously collected to dynamically track minute frequency drifts of the main frequency of artificial interference, changes in multi-band superposition, fluctuations in interference energy strength, and the intermittent occurrence of interference. The parameters of the notch filter are dynamically corrected frame by frame based on real-time spectrum changes. When interference from multiple devices occurs, the suppression bandwidth is adaptively widened; when the interference frequency is stable and the amplitude is weak, the suppression depth is adaptively reduced; and when the interference disappears instantaneously, the basic parameters are automatically restored. Through continuous dynamic iterative parameter tuning, the notch filter strategy is ensured to always adapt to real-time operating condition changes, avoiding interference residues or excessive attenuation of the effective signal caused by fixed parameters.

[0032] Step S1234: Targeted frequency domain suppression is implemented through the intelligent notch filter after dynamic parameter matching. Fixed-point depth attenuation and cancellation are performed on the narrowband interference frequency band corresponding to artificial equipment in the signal frequency domain. Non-interference frequency band and effective signal frequency band of partial discharge pulse are retained throughout the process. Under the premise of completely preserving the weak partial discharge pulse signal, artificial narrowband interference is stripped away, and a high-purity signal is output that simultaneously eliminates ocean broadband random noise and fixed narrowband interference of artificial equipment.

[0033] In this step, a smart adaptive notch filter with dynamically matched parameters is used to perform precise frequency-domain targeted suppression. This involves targeted, deep attenuation and cancellation of narrowband interference frequencies associated with artificial equipment within the signal frequency domain. During interference suppression, non-interference frequency band signals are strictly isolated, fully preserving the effective frequency band information and time-domain waveform characteristics corresponding to partial discharge pulses, without damaging the amplitude details, abrupt changes, and attenuation patterns of weak partial discharge signals. This noise reduction process completely eliminates artificial narrowband electromagnetic interference from offshore wind power, ships, and platform equipment, ultimately outputting a high signal-to-noise ratio, high-purity monitoring signal that simultaneously filters out both broadband random noise from the ocean and artificial fixed narrowband interference.

[0034] Step S124: Based on the Morse wavelet-TT joint feature extraction architecture, the Morse wavelet's multi-scale high-resolution fine decomposition capability is used to mine the weak pulse detail features, and the TT time-series trajectory analysis method is combined to characterize the temporal evolution law of the partial discharge signal, and multi-dimensional extraction and quantization of partial discharge feature parameters are performed. In this specific implementation step, addressing the issues of low amplitude, easy loss of details, and poor feature recognition of weak, latent partial discharge pulses under marine conditions, a joint feature extraction architecture combining Morse wavelet and TT time-series trajectory is constructed to achieve multi-dimensional fusion extraction of microscopic waveform features and macroscopic time-series evolution features. Leveraging the advantages of Morse wavelet multi-scale high-resolution fine decomposition, the high-purity signal after double-layer noise reduction is refined layer by layer to capture subtle abrupt changes in ultra-low-energy weak pulses and remove residual trace noise. Simultaneously, multi-dimensional quantization parameters such as pulse rise edge, attenuation characteristics, pulse width, amplitude distribution, repetition frequency, and time-domain interval are extracted. The TT time-series trajectory analysis method is introduced to reconstruct discrete pulse features into continuous time-series evolution trajectories, characterizing the dynamic evolution process of insulation degradation. By fusing microscopic waveforms and macroscopic time-series features, a standardized and highly discriminative set of partial discharge feature parameters is formed, thereby making the latent weak partial discharge features explicit.

[0035] Step S1241: Obtain the high-purity submarine cable monitoring signal with high signal-to-noise ratio after double-layer noise reduction processing as the feature mining input. In view of the characteristics of weak, latent partial discharge pulses with low amplitude, easy loss of details and low feature recognition under marine conditions, start Morse wavelet multi-scale fine decomposition processing. Adaptively match Morse wavelet waveform parameters and scale factors according to the frequency domain characteristics of submarine cable partial discharge pulses. Perform layered analysis on weak partial discharge pulse signals through layer-by-layer multi-scale refinement subdivision, peel off the subtle noise interference left by the signal layer by layer, completely preserve the subtle waveform details of ultra-low energy partial discharge pulses, and capture the information of minute pulse change. In this specific implementation step, the high signal-to-noise ratio submarine cable monitoring signal, after double-layer noise reduction and purification, is selected as the input for feature mining. The focus is on refining the decomposition of the weak, ambiguous, and easily residual minute noise issues inherent in marine partial discharge signals. Based on the inherent frequency domain distribution characteristics of submarine cable partial discharge pulses, the optimal waveform parameters of the Morse wavelet and multi-scale decomposition factors are adaptively matched. The weak pulse signal is then analyzed layer by layer through a coarse-to-fine, multi-scale partitioning method. During the decomposition process, the subtle residual noise components within the signal are gradually removed layer by layer, fully preserving the subtle waveform structure and minute pulse abrupt changes of the ultra-low energy partial discharge pulse, avoiding the loss of detail and feature ambiguity caused by the fixed scale of traditional decomposition methods.

[0036] Step S1242: Based on the Morse wavelet multi-scale refinement decomposition results, quantitative extraction of multidimensional features of partial discharge is carried out. The refined pulse waveform data is analyzed segment by segment to extract six quantitative features: pulse rise edge characteristics, exponential decay characteristics, effective pulse width, amplitude distribution law, pulse repetition frequency and time interval between adjacent pulses. This covers the microscopic morphological features of single pulses and the pulse temporal distribution features, and constructs a complete dataset of original features of partial discharge. In this step, a systematic extraction of multi-dimensional quantization features of partial discharge (PD) is carried out based on the results of Morse wavelet multi-scale refined decomposition. The refined pulse waveform data is analyzed segment by segment to extract six core quantization parameters that comprehensively characterize PD features, covering pulse rise steepness, exponential decay characteristics, effective pulse duration, amplitude probability distribution, pulse repetition frequency, and time interval between adjacent pulses. The extracted features include both microscopic morphological details of single pulses and pulse temporal distribution patterns, enabling a comprehensive characterization of PD features with different degrees of degradation and defect types. Through multi-dimensional feature quantization, a complete, multi-dimensional, and highly discriminative raw PD feature dataset is constructed.

[0037] Step S1243: Introduce TT time-series trajectory analysis technology to reconstruct the trajectory of the partial discharge pulse time-series characteristic parameters extracted in a continuous time period. Convert the discrete single-point pulse characteristics into a continuous time-series evolution trajectory. By analyzing the fluctuation trend of partial discharge pulse amplitude, the change law of pulse occurrence interval, and the discrete characteristics of pulse aggregation, the dynamic evolution process of submarine cable insulation from slight deterioration and latent discharge to intermittent discharge can be characterized, and the latent defect time-series correlation characteristics that cannot be characterized by a single waveform feature can be explored. In the specific implementation of this step, TT time-series trajectory analysis technology is introduced to mine the macroscopic evolution characteristics of partial discharge, overcoming the deficiency that single static waveform features cannot characterize the dynamic process of insulation degradation. The discrete partial discharge pulse characteristic parameters extracted successively within a continuous monitoring period are reconstructed into a time-series trajectory, forming a continuously traceable partial discharge evolution trajectory curve. By analyzing the trajectory amplitude fluctuation trend, pulse interval variation law, and pulse aggregation and discrete distribution characteristics, the entire dynamic evolution process of submarine cable insulation from slight latent degradation and sporadic weak discharge to intermittent and continuous discharge is fully depicted. This effectively uncovers the temporal correlation laws of latent defects that cannot be identified by single waveform features, achieving in-depth characterization of progressive and latent insulation defects.

[0038] Step S1244: Integrate the pulse micro-waveform features extracted by Morse wavelet with the macro-evolution features represented by TT time-series trajectory, perform normalization and quantization processing, redundant feature removal and dimension regularization optimization on all feature parameters, and construct a standardized, quantifiable, distinguishable and intelligently comparable set of submarine cable partial discharge feature parameters to achieve explicit transformation of weak, latent partial discharge signal features with strong concealment and ambiguous features.

[0039] In this step, the microscopic waveform features and macroscopic time-series trajectory features are fused and normalized to construct a standardized partial discharge (PD) feature parameter set. The multi-dimensional single-pulse microscopic features extracted by Morse wavelet decomposition are fused and summarized with the macroscopic evolution features represented by the TT time-series trajectory, forming a comprehensive, multi-level PD feature system. All fused feature parameters undergo unified normalization and quantization to eliminate dimensional differences. Redundant features are simultaneously filtered and eliminated, and feature dimensions are normalized and optimized, retaining highly correlated and discriminative core PD features. Ultimately, a standardized, quantifiable, comparable, and trainable set of submarine cable-specific PD feature parameters is formed, enabling the efficient conversion of weak, fuzzy, and highly concealed PD signals from implicit, ambiguous features to explicit, quantifiable features.

[0040] Step S125: Based on the pre-constructed marine environmental noise sample library and the submarine cable insulation defect partial discharge feature library, intelligent comparison and identification are carried out. The dual databases contain a large number of environmental interference waveforms, artificial electromagnetic interference waveforms, and partial discharge feature waveforms of different sea conditions and different operating scenarios, as well as different aging degrees and different defect types. The multi-dimensional feature parameters of the signals extracted in real time are intelligently matched and compared with the database samples. The feature similarity threshold judgment logic automatically distinguishes marine random noise, artificial equipment electromagnetic interference and real insulation partial discharge signals, filters early weak partial discharge signals, identifies minor insulation degradation and hidden defects of submarine cables, and completes the extraction and reliable identification of submarine cable partial discharge signals under marine electromagnetic environment.

[0041] In the specific implementation of this step, a pre-built dual-database intelligent comparison and identification system is used to accurately distinguish between interference signals and real partial discharge signals. A marine environmental noise sample library and a submarine cable insulation defect partial discharge feature library covering all scenarios are pre-built, containing a massive amount of environmental interference waveforms and various artificial electromagnetic interference waveforms under different sea conditions, seasons, and offshore operation scenarios. It also includes partial discharge feature waveforms with different insulation aging degrees, different defect types, and different fault stages. The real-time extracted multi-dimensional quantitative partial discharge feature parameters are compared with standard samples from the dual databases using high-precision intelligent matching. Based on a preset feature similarity threshold judgment logic, the system automatically distinguishes between marine random noise, artificial electromagnetic interference, and real insulation partial discharge signals, accurately screening early weak partial discharge features. This effectively identifies potential minor insulation degradation hazards and hidden defects in submarine cables, achieving stable, reliable, and accurate extraction and identification of submarine cable partial discharge signals in complex marine electromagnetic environments.

[0042] Step S130: Integrate multi-source monitoring data and introduce dynamic correction parameters and deep learning models to adaptively adapt to marine conditions; In this implementation process, addressing the shortcomings of traditional submarine cable monitoring technologies—such as single data dimensions, poor adaptability to operating conditions, weak model generalization ability, and insufficient accuracy in identifying time-varying marine scenes—an integrated adaptive monitoring mechanism combining multi-source data fusion, dynamic correction of operating conditions, and deep learning-based intelligent discrimination is established. By integrating heterogeneous monitoring data on submarine cable mechanics, electrical, and electromagnetic parameters with marine environmental operating condition data, the limitations of single-sensor information representation capabilities are overcome, achieving complementary verification of multi-dimensional state information. Simultaneously, multiple types of dynamic correction parameters for marine operating conditions are introduced to solve data distortion and feature shift problems caused by seawater environment, tidal disturbances, electromagnetic interference, and long-distance transmission. Relying on a hierarchical deep learning feature fusion model to automatically learn the implicit coupling rules of multi-source data, the model's feature extraction, fusion weights, and discrimination logic can dynamically and adaptively adjust according to real-time marine operating conditions. This overcomes the shortcomings of traditional fixed-parameter models that cannot adapt to complex time-varying marine scenarios, achieving full-scene adaptive adaptation for intelligent perception of submarine cable operating status, effectively improving the stability and accuracy of fault identification under complex sea conditions.

[0043] Step S131: By fully integrating heterogeneous monitoring data of submarine cable mechanics, electrical, and electromagnetic data, as well as marine working environment data, the time series benchmark and format of multi-source data are unified. Data cleaning, gross error removal, and normalization preprocessing are completed to eliminate data deviation caused by differences in data collection from multiple devices and to construct a time-ordered and dimensionally complementary multi-source fusion monitoring dataset for submarine cables. In this step, comprehensive data collection is conducted on various heterogeneous monitoring data, including those related to submarine cable mechanical deformation, temperature strain, electrical traveling waves, magnetic field distortion, stray currents, and the marine environment. A full-domain data fusion preprocessing operation is then performed. Addressing the issue of inconsistent acquisition frequencies, data formats, and timing references among different sensing hardware, unified timing alignment rules and standardized data standards are established. This normalizes heterogeneous data from multiple terminals across the entire line to the same timing coordinate system, completely eliminating timing misalignment caused by equipment differences. Simultaneously, batch data cleaning, gross error removal, abnormal noise filtering, and global normalization are carried out to correct data system deviations caused by sensor zero-point drift, transmission attenuation, and equipment differences, ensuring uniformity of dimensions, regularity of amplitude, and reliable quality of all types of monitoring data. Ultimately, a standardized multi-source fusion monitoring dataset for submarine cables is constructed, characterized by highly regular timing, complementary feature dimensions, and effective error suppression.

[0044] Step S132: Construct a dynamic parameter correction system for operating conditions that adapts to time-varying marine scenarios. Introduce multiple dynamic correction parameters, such as seawater deep pressure attenuation correction factor, seawater temperature transmission loss correction factor, tidal scour operating condition offset correction factor, offshore wind power and ship electromagnetic interference intensity correction factor, and long-distance submarine cable signal transmission delay correction factor. Dynamically iterate and update each correction factor parameter based on real-time collected marine environmental operating condition data, and implement point-by-point dynamic compensation correction for multi-source monitoring raw data. In the specific implementation of this step, a multi-factor dynamic parameter correction system adapted to time-varying ocean scenarios is constructed to specifically solve the problems of monitoring data distortion, feature offset and amplitude attenuation caused by complex marine environments. Combined with the underwater operation characteristics of submarine cables, five types of key dynamic correction factors are introduced, including deep seawater pressure attenuation, seawater temperature transmission loss, tidal scouring working condition offset, electromagnetic interference from offshore wind power and ships, and long-distance signal transmission delay, which fully cover various influencing factors of marine environment, external working conditions, electromagnetic interference and transmission links. Relying on on-site working condition information collected in real time such as water depth, water temperature, tidal flow velocity, electromagnetic interference intensity and transmission distance, the compensation parameters of each correction factor are iteratively updated in real time to realize the linkage change of correction rules and real-time sea conditions. Dynamic compensation and deviation calibration are carried out point by point and frame by frame on the original multi-source monitoring data, which adaptively cancels the data distortion caused by time-varying ocean interference, ensures that the monitoring data is real and consistent under different sea conditions, different time periods and different transmission distances, and greatly improves the working condition adaptability and accuracy of the original data.

[0045] Step S133: Construct a deep learning feature fusion model for submarine cable fault identification, take the dynamically corrected standardized multi-source monitoring data as input, build a multi-layer feature fusion neural network structure, respectively complete the low-dimensional representation extraction of optical fiber mechanical features, electrical traveling wave features and magnetic field distortion features through shallow networks, realize cross-fusion and complementary enhancement of multi-dimensional heterogeneous features relying on the middle-layer feature fusion layer, automatically mine the implicit correlation features and coupling change rules between multi-source data, complete the abstract condensation and redundant feature elimination of high-dimensional fusion features through deep networks, and screen out fusion features strongly associated with submarine cable insulation deterioration, partial discharge, mechanical damage and external interference; In the specific implementation process of this step, a deep learning layered feature fusion model adapted to multi-source heterogeneous data of submarine cables is constructed, the standardized monitoring data after multi-factor dynamic correction is taken as the model input, and a three-level neural network architecture of independent extraction in shallow layers, cross-fusion in middle layers and condensation screening in deep layers is constructed. Independent low-dimensional feature extraction of three types of heterogeneous data including mechanical, electrical and electromagnetic data is realized in the shallow layer, avoiding the loss of details caused by mutual interference of different types of features. Through the middle-layer cross-modal fusion mechanism, deep complementary coupling of multi-source features is realized, and the implicit correlation and time-varying rules that cannot be captured by traditional artificial features are actively mined. Relying on the deep convolutional residual structure to complete the abstract representation of high-dimensional features, filter invalid noise and redundant information layer by layer, and accurately retain the core fusion features highly related to submarine cable insulation deterioration, partial discharge, mechanical damage and external electromagnetic interference. Through the layered progressive feature learning mode, the intelligent, refined and adaptive conversion from original monitoring data to core fault features is realized, which significantly improves the discrimination and representation ability of submarine cable fault features under complex working conditions.

[0046] Step S1331: Perform dimensional splitting and channelization processing on the standardized multi-source monitoring data after dynamic correction, divide the fiber optic mechanical data, electrical traveling wave data, and magnetic field distortion data into corresponding independent input channels, and unify the data length, number of sampling points, and feature dimensions of each channel; In this step, to address the challenges of different dimensions, significant differences in feature distribution, and difficulties in unified modeling of the three heterogeneous data types (mechanical, electrical, and electromagnetic) of submarine cables, multi-source data dimensional decomposition and independent channel reconstruction were implemented. The complete monitoring dataset, after dynamic correction and standardization preprocessing, was precisely decomposed into three independent input channels according to data type: fiber optic mechanical data channel, electrical traveling wave data channel, and magnetic field distortion data channel, achieving partitioned and isolated input of data with different physical attributes. Data length, number of sampling points, temporal resolution, and feature dimensions were standardized for each channel to eliminate dimensionality mismatches caused by differences in sampling mechanisms of different sensing devices. Through channel normalization, each data type retains its unique physical characteristics and signal details while meeting the structured input requirements of neural networks, avoiding feature confusion, detail overload, and learning bias caused by heterogeneous data aliasing.

[0047] Step S1332: Construct a hierarchical deep learning feature fusion network. Set up a branched shallow feature extraction subnet in the first layer of the network. Configure independent one-dimensional convolution and pooling structures for three types of heterogeneous data: mechanical, electrical, and electromagnetic. Extract waveform features and abrupt change features corresponding to temperature, deformation, and strain through the mechanical subnet. Extract pulse amplitude, rise time, and decay characteristics through the electrical subnet. Extract magnetic field distortion amplitude, fluctuation frequency, and stray current features through the electromagnetic subnet. Complete the extraction of low-dimensional basic features of various single-source data. In this step, a branched shallow feature extraction subnet architecture is built. Independent convolution and pooling structures are configured differently for the signal characteristics of the three types of heterogeneous data, enabling targeted and accurate extraction of specific features. The mechanical branch subnet adapts to the slow-changing, cumulative, and deformation-type signal features of fiber optic sensing, focusing on extracting the time-domain waveform features and instantaneous abrupt changes corresponding to temperature drift, deformation fluctuations, strain abrupt changes, and vibration disturbances in submarine cables, fully preserving details of mechanical damage and stress anomalies. The electrical branch subnet adapts to high-frequency transient pulse signals, accurately capturing typical electrical fault features such as the amplitude of traveling wave pulses, the steepness of the rise edge, and the waveform attenuation rate. The electromagnetic branch subnet extracts the amplitude of magnetic field fluctuations, oscillation frequency, and stray current disturbance features for weak magnetic field distortion signals. Through a multi-branch independent extraction mode, the differences in the physical mechanisms of different signals are adapted, ensuring that various weak and specific low-level features are not lost or confused.

[0048] Step S1333: Construct a mid-level cross-modal adaptive feature fusion layer. A fusion strategy combining dimensional splicing fusion and dynamic weighted fusion is adopted. The feature vectors output by the three shallow subnets are spliced ​​dimensionally. At the same time, an adaptive feature weight allocation mechanism is introduced to dynamically adjust the fusion weights of mechanical, electrical, and electromagnetic features according to the real-time marine working conditions and signal strength. This strengthens the weights of high-contribution effective features and weakens the interference of low-correlation redundant features, thereby completing the deep cross-fusion of multi-dimensional heterogeneous features and automatically mining the implicit correlation characteristics and time-varying coupling change patterns between multi-source monitoring data. In this step, a mid-level cross-modal adaptive dynamic weighted fusion layer is constructed to address the shortcomings of traditional fixed-weight fusion methods, which cannot adapt to time-varying ocean conditions and dynamic changes in feature contribution. First, the mechanical, electrical, and electromagnetic feature vectors output from the three shallow subnets are dimensionally spliced ​​and fused to achieve complete aggregation of multi-dimensional feature information. Simultaneously, an adaptive weight allocation mechanism linked to ocean conditions is introduced. Based on real-time ocean tidal states, water temperature fluctuations, electromagnetic interference strength, and signal-to-noise ratio, the real-time contribution of each modal feature is dynamically calculated, automatically increasing the fusion weight of highly effective and highly correlated features under the current condition, and weakening the weight of redundant features with low correlation and dominated by interference. Through dynamic weighted cross-fusion, deep complementary enhancement of multi-source heterogeneous features is achieved, actively uncovering implicit correlations, coupling patterns, and time-varying evolution characteristics among various types of data under different sea conditions, significantly improving the scenario adaptability and fault characterization accuracy of the fused features.

[0049] Step S1334: Construct a deep high-dimensional feature condensation network containing multi-layer convolutional structures, residual modules and global pooling structures. Perform layer-by-layer abstraction, feature compression and deep representation learning on the fusion features output by the middle layer. Automatically filter out noise features, redundant features and weakly correlated features that are not related to the operating status of the submarine cable. Layer-by-layer condensation yields high-dimensional fusion features that are highly correlated with submarine cable insulation degradation, partial discharge, mechanical damage and external electromagnetic interference. In this step, a deep, high-dimensional feature condensation network integrating multi-layer convolution, residual modules, and global pooling is constructed to perform deep abstraction and intelligent filtering of the combined features after mid-layer fusion. Multi-layer convolutional structures are used to progressively increase the level of feature abstraction, refining the feature association representation capability. Residual modules are used to solve the gradient vanishing problem during deep network training, ensuring the stability and effectiveness of deep feature learning. During the layer-by-layer feature extrapolation process, residual noise features, redundant features, and weakly correlated interference features unrelated to the actual operating state of the submarine cable are automatically identified and filtered, eliminating invalid interference information caused by fluctuations in marine conditions. Through progressive compression, condensation, and purification, high-dimensional core fusion features highly coupled with insulation aging, partial discharge, mechanical damage, and external electromagnetic interference are ultimately retained, achieving precise focusing of fault features and significantly improved feature discrimination.

[0050] Step S1335: The refined high-dimensional fusion features are normalized and ranked by contribution through the fully connected layer and feature saliency screening module. Adaptive screening is performed based on the contribution of each feature to the classification and status determination of submarine cable faults, retaining highly significant fusion features and eliminating invalid and redundant features.

[0051] In this specific implementation step, the high-dimensional fusion features are finalized and adaptively optimized using the fully connected layer and feature saliency screening module. First, the high-dimensional condensed features output by the deep network are globally normalized to unify feature dimensions and numerical distribution, eliminating discrimination bias caused by differences in feature scale. Then, through a feature contribution ranking mechanism, the influence weights of each feature dimension on cable fault classification, defect identification, and state determination are enumerated one by one, quantifying the saliency level of each feature. Based on contribution thresholds, feature selection and dimensional simplification are adaptively completed, systematically retaining core fusion features with high correlation, high sensitivity, and high discriminative power, while completely eliminating invalid, redundant, and low-contribution interfering feature dimensions. Finally, a cable-specific core feature set that is dimensionally simplified, highly representative, robust, and adaptable to complex marine conditions is output, effectively improving model inference efficiency and state determination accuracy.

[0052] Step S134: Construct a sample dataset of submarine cable working conditions covering the entire scenario. The sample dataset includes normal working conditions, interference working conditions, latent defect working conditions, and explicit fault working conditions corresponding to different sea conditions, different interference intensities, different fault types, and different insulation aging degrees. The deep learning feature fusion model is iteratively trained using supervised learning. The model weight parameters and bias parameters are continuously optimized through backpropagation of the loss function to correct the model feature extraction and feature fusion logic. In this step, a finely labeled sample dataset of submarine cables covering all operating conditions is constructed, providing a complete supervised training foundation for the deep learning model. The sample data extensively includes normal operation samples, environmental interference samples, latent insulation degradation samples, intermittent partial discharge samples, and various overt fault samples under different seawater temperatures, tidal conditions, electromagnetic interference intensity, seabed conditions, and service life, achieving full coverage of all operating conditions, defects, and fault types. All samples have undergone accurate operating condition labeling and fault type hierarchical labeling, constructing a supervised learning dataset with a one-to-one correspondence between sample features and fault states. A supervised learning model is used to iteratively train the feature fusion model. The model prediction bias is calculated in real time through a loss function, and backpropagation continuously updates network weights and bias parameters. The multi-branch feature extraction logic and cross-modal fusion strategy are continuously optimized to correct model discrimination bias and continuously improve the model's ability to identify and adapt to complex operating conditions and subtle latent faults.

[0053] Step S135: Deploy the iteratively optimized deep learning model to the real-time monitoring inference stage. Perform continuous feature fusion and intelligent discrimination on the multi-source monitoring data that is accessed in real time and has been dynamically corrected. The model dynamically and adaptively adjusts the feature weights and discrimination thresholds according to the real-time marine conditions. It can adapt to the time-varying marine scenarios of tidal changes, electromagnetic interference fluctuations, and seabed environment changes without manual parameter debugging. It outputs the submarine cable operation status discrimination results, defect types, and hazard levels in real time.

[0054] In this specific implementation step, the deep learning feature fusion model, after training convergence and parameter optimization, is deployed to the submarine cable real-time monitoring inference process to achieve uninterrupted intelligent judgment of the submarine cable's operating status. Multi-source monitoring data collected in real-time on-site, after preprocessing and dynamic condition correction, is continuously input into the deployed inference model. The model automatically completes multi-channel feature extraction, cross-modal adaptive fusion, high-dimensional feature condensation, and salient feature selection. During inference, the model can adaptively adjust the feature fusion weights and internal discrimination thresholds of each modality according to real-time ocean condition fluctuations, without manual intervention or parameter tuning, automatically adapting to time-varying complex scenarios such as tidal changes, electromagnetic interference drift, and seabed erosion. The model outputs real-time status judgment results for submarine cables, including normal operation, external interference, latent insulation degradation, partial discharge, and mechanical damage, simultaneously outputting the specific type of defect and the risk level of potential hazards, achieving all-time, adaptive, and high-precision intelligent monitoring and fault assessment of submarine cables under complex ocean conditions.

[0055] Step S140: Integrate frequency domain detection technology with traditional time domain monitoring methods, and realize the monitoring and verification of all types of submarine cable faults through photoelectric bidirectional verification; In the specific implementation of this step, addressing the industry pain points of traditional submarine cable monitoring—relying solely on a single monitoring method, incomplete fault identification types, difficulty in distinguishing between true and false faults, and a high rate of missed detection of latent faults—this innovative approach integrates frequency domain detection technology with traditional time domain monitoring methods, combining fiber optic physical monitoring to construct a multi-dimensional fault verification system. Traditional time domain monitoring can only capture sudden, obvious faults, failing to identify slow, latent insulation degradation and electromagnetic dispersion anomalies. Conversely, pure frequency domain monitoring lacks physical state corroboration and is easily affected by environmental interference. This technology, by establishing a time-frequency dual-channel synchronous monitoring architecture, achieves complementary acquisition and analysis of transient and steady-state characteristics in the time domain, while simultaneously relying on fiber optic monitoring data to complete physical state corroboration, forming a new monitoring mode of photoelectric bidirectional verification and joint time-frequency analysis. Through multi-dimensional technology integration, it achieves comprehensive monitoring of both obvious sudden faults and latent slow-changing faults, and eliminates false anomalies caused by interference through multiple cross-verifications, comprehensively improving the completeness, accuracy, and reliability of submarine cable fault monitoring.

[0056] Step S141: Establish a dual-channel synchronous monitoring link in the time and frequency domains. Relying on high-frequency traveling waves and distributed optical fiber sensing devices, synchronously collect various time-domain operating characteristics of submarine cables. Combine global fast Fourier transform and short-time Fourier transform to complete the frame-by-frame frequency domain feature analysis of the signal. Perform time-series alignment and unified labeling on the heterogeneous time-frequency monitoring data to construct a time-series regular and dimensionally complementary submarine cable time-frequency synchronous monitoring dataset. In its implementation, this step establishes a fully synchronous dual-channel monitoring link in the time and frequency domains to address the issues of asynchronous time-frequency monitoring, data misalignment, and limited dimensionality. Utilizing parallel operation of on-site high-frequency traveling wave acquisition equipment and distributed fiber optic sensing devices, it continuously and synchronously acquires full-dimensional time-domain operational characteristic data, including submarine cable current traveling waves, pulse mutations, deformation vibrations, and temperature strain. Simultaneously, it employs global fast Fourier transform and short-time Fourier transform algorithms to perform frame-by-frame frequency domain analysis on the synchronously acquired raw monitoring signals, extracting frequency domain features such as spectral distribution, energy accumulation characteristics, frequency point offset, and harmonic distortion in real time. Addressing the inconsistencies in sampling rate, data format, and timing scale between the two heterogeneous data types (time and frequency), it uniformly completes precise timing alignment, data format standardization, and feature synchronization labeling, eliminating multi-dimensional data timing deviations and dimensional barriers. Ultimately, it constructs a highly standardized, complementary, and spatiotemporally synchronized integrated time-frequency monitoring dataset for submarine cables.

[0057] Step S142: Establish a two-way independent detection mechanism in the time domain and frequency domain. Based on time domain monitoring data, capture transient time domain features such as pulse mutation, waveform distortion, traveling wave reflection, strain anomaly, and temperature mutation to identify abnormal time domain conditions such as submarine cable short circuit, line break, mechanical damage, sudden partial discharge, and external force disturbance. Relying on frequency domain data, capture abnormal frequency domain features such as abnormal spectral energy shift, characteristic frequency band mutation, harmonic component surge, and electromagnetic distortion frequency point anomaly to identify insulation degradation, latent continuous discharge, marine electromagnetic interference, equipment noise interference, and long-distance transmission dispersion anomaly latent conditions that are difficult to identify in the time domain. This forms a differentiated detection mode for transient explicit faults monitored in the time domain and slow-changing implicit faults monitored in the frequency domain. In the specific implementation of this step, a differentiated fault detection mechanism is established that is independent yet complementary in the time and frequency domains to achieve accurate classification and identification of different types of faults. The time-domain monitoring link focuses on identifying transient and sudden operating conditions, relying on real-time time-domain waveform data to capture typical time-domain characteristics such as pulse mutations, waveform distortion, traveling wave reflections, abnormal mechanical strain, and instantaneous temperature changes, quickly determining explicit sudden faults such as submarine cable short circuits, breaks, mechanical damage, sudden partial discharges, and external disturbances. The frequency-domain monitoring link focuses on identifying slow-changing and latent operating conditions, continuously analyzing the signal spectrum to capture subtle frequency-domain changes such as spectral energy shifts, characteristic frequency band mutations, harmonic surges, and magnetic field distortion frequency shifts, accurately identifying latent anomalies that cannot be captured in the time domain, such as insulation aging, continuous latent discharge, marine electromagnetic interference, equipment noise interference, and long-distance transmission dispersion. Through time-frequency division of labor and complementary coverage, a differentiated and complete detection system is formed, enabling time-domain identification of explicit faults and frequency-domain capture of latent faults.

[0058] Step S143: Construct a photoelectric bidirectional cross-verification mechanism for fiber optic monitoring and electrical time-frequency monitoring. The physical deformation, vibration disturbance, and temperature anomaly point information collected by fiber optic sensing are compared and verified with the electrical fault characteristics obtained by time-domain and frequency-domain analysis in three dimensions: space, time sequence, and characteristics. When anomalies are simultaneously detected in the time domain and frequency domain and the optical sensing point status changes synchronously, it is determined to be a real submarine cable fault. When anomalies occur only in a single time domain or frequency domain and there are no corresponding disturbance characteristics in the optical sensing data, it is determined to be a false fault caused by marine random noise, equipment electromagnetic interference, or data jitter. This step involves constructing a two-way photoelectric cross-verification mechanism combining fiber optic and electrical time-frequency monitoring to completely resolve the issues of false alarms and misdiagnosis of faults under complex marine conditions. Real-time data collected by fiber optic sensors on submarine cable physical deformation, vibration disturbances, temperature anomalies, and anomaly locations are used as the physical truth, and are cross-compared in multiple dimensions with electrical anomaly characteristics obtained from time-domain and frequency-domain analysis. Spatially, the anomaly locations are checked for matching; temporally, the occurrence of anomalies is verified for synchronization; and characteristically, the electrical anomalies correspond to physical disturbances. When anomalies are simultaneously detected in the time and frequency domains, and corresponding abrupt changes in fiber optic sensing data occur synchronously, a genuine fault in the submarine cable is confirmed. When anomalies occur only in a single time or frequency domain, and there is no matching physical disturbance in the fiber optic cable, it is determined to be a false fault caused by marine random noise, electromagnetic interference, or data jitter. This triple photoelectric verification significantly improves the accuracy of fault identification.

[0059] Step S144: Construct a time-frequency joint judgment model adapted to different fault types. For physical faults such as mechanical damage and external dragging, combine the time-domain deformation mutation characteristics of optical fiber with the frequency-domain high-frequency disturbance energy concentration characteristics to complete the joint judgment. For faults such as partial discharge and insulation aging and deterioration, rely on the time-domain weak pulse accumulation characteristics and the frequency-domain continuous distortion band characteristics to achieve accurate identification. For electrical faults such as short circuit, grounding, and open circuit, use the time-domain traveling wave mutation characteristics and the frequency-domain fundamental wave offset and harmonic surge characteristics to complete the fault differentiation, comprehensively covering the types of obvious sudden faults and hidden continuous faults in submarine cables. In its implementation, this step involves constructing a dedicated time-frequency joint criterion model tailored to the mechanistic characteristics of different faults, enabling refined and differentiated accurate judgment of various fault types. For physical faults such as mechanical damage and external dragging, the model combines the abrupt changes in optical fiber time-domain deformation with the concentrated energy characteristics of high-frequency disturbances in the frequency domain to accurately distinguish between external disturbances and natural environmental fluctuations. For progressive degradation faults such as partial discharge and insulation aging, the model relies on the cumulative change characteristics of weak pulses in the time domain and the characteristics of continuous distortion bands in the frequency domain to capture the implicit evolutionary patterns of slow degradation, achieving early insulation defect identification. For sudden electrical faults such as short circuits, grounding, and open circuits, the model utilizes the instantaneous abrupt changes in traveling waves in the time domain combined with the fundamental wave offset and harmonic surge characteristics in the frequency domain to complete fault type differentiation and fault location. Through multiple types of dedicated joint criteria, the model comprehensively covers physical faults, insulation degradation faults, and sudden electrical faults in submarine cables, achieving accurate identification and classification of all fault types.

[0060] Step S145: Integrate the multi-dimensional verification conclusions of time-domain transient features, frequency-domain latent features, and optical-physical features, output the final operating status of the submarine cable, and simultaneously determine and output the fault type, fault location, fault occurrence time, fault degradation degree, and interference source information.

[0061] In the specific implementation of this step, three core judgment results—time-domain transient electrical characteristics, frequency-domain latent distortion characteristics, and fiber optic physical characteristics—are integrated to complete the final comprehensive judgment of the submarine cable's operating status. Based on multiple cross-verifications and a dedicated fault criterion model, the judgment conclusions from all dimensions are merged and summarized, eliminating the shortcomings of single-dimensional, one-sided judgments and relying on the complementary advantages of multiple features to output the final operating status of the submarine cable. Simultaneously, by combining the anomaly occurrence sequence, spatial location, characteristic amplitude, and degree of degradation evolution, the specific fault type, precise fault location, fault occurrence time, and severity of insulation degradation are accurately determined. Furthermore, frequency-domain interference characteristics are used to trace the source of natural marine interference and artificial interference from equipment. This achieves intelligent output across all dimensions of submarine cable faults, from presence / absence determination, type differentiation, location identification, degradation rating, and interference tracing.

[0062] Step S150: Based on the edge-cloud collaborative architecture, complete the adaptive adjustment of monitoring threshold conditions, quantify the insulation aging trend, predict fault risks in advance, and build a closed-loop system for intelligent monitoring and maintenance of submarine cables; In its specific implementation, this step leverages an edge-cloud collaborative layered architecture to address the technical shortcomings of traditional submarine cable monitoring, such as fixed thresholds, poor adaptability to operating conditions, delayed fault prediction, and passive maintenance. It constructs an adaptive, predictable, and iterative intelligent operation and maintenance closed-loop system. By parsing monitoring data locally in real time at the edge, real-time monitoring is ensured. A condition threshold mapping model is built using the advantages of cloud-based big data training, enabling dynamic adaptive adjustment of judgment thresholds according to time-varying marine conditions, completely eliminating the serious false alarms and missed alarms associated with traditional fixed thresholds. Based on long-term time-series monitoring data, insulation aging trends are modeled and quantified to proactively identify progressive fault risks and achieve early fault risk prediction. A tiered operation and maintenance strategy is matched according to the real-time operating status, degradation level, and risk level of the submarine cable, forming a differentiated intelligent decision-making mechanism. Simultaneously, continuous iterative optimization of the model is achieved through two-way data interaction between the edge and cloud, constructing a full-process intelligent operation and maintenance closed loop integrating monitoring, analysis, early warning, decision-making, operation and maintenance, and iterative optimization, comprehensively improving the intelligence level of submarine cable operation and maintenance.

[0063] Step S151: Construct an edge-cloud hierarchical collaborative monitoring architecture, relying on edge terminals to complete real-time analysis, feature extraction and preliminary judgment of multi-source monitoring data locally, and upload effective features and operating condition data to the cloud as needed; In its implementation, this step establishes a layered and collaborative intelligent monitoring architecture for submarine cables, integrating edge and cloud-based systems to achieve a seamless combination of real-time assessment and in-depth analysis. Edge terminals are deployed at the cable site to access multi-source monitoring data and marine condition data from across the entire cable line. This allows for on-site data preprocessing, real-time feature extraction, preliminary anomaly identification, and real-time threshold updates. Leveraging localized edge computing power significantly reduces data transmission latency, ensuring real-time monitoring and assessment. Simultaneously, the traditional full-data upload model is abandoned. Field data is filtered and only valid feature data, anomaly sample data, and condition change data are uploaded to the cloud platform on demand via encrypted links, effectively avoiding channel congestion, data redundancy, and transmission delays associated with full-data transmission. The cloud platform utilizes the massive amounts of uploaded data for in-depth modeling, model training, and trend analysis, forming a layered collaborative working mode of real-time edge processing and in-depth cloud iteration.

[0064] Step S152: Establish a correlation mapping model between marine operating conditions and monitoring thresholds. Based on the full operating condition samples in the cloud, obtain dynamic mapping rules and optimize various fault discrimination thresholds adaptively and iteratively by the edge terminal in combination with real-time sea conditions. In its specific implementation, this step involves constructing an intelligent mapping system that establishes a one-to-one correspondence between marine operating conditions and monitoring thresholds. This addresses the core challenge of traditional submarine cable monitoring, where fixed thresholds cannot adapt to time-varying marine conditions. Traditional monitoring thresholds, once set and remaining unchanged for a long period, cannot adapt to dynamic changes in tides, water temperature, electromagnetic interference, and seabed disturbances, easily leading to false alarms and missed alarms. This technology relies on a massive cloud-based full-condition sample training intelligent mapping model to establish a quantitative coupling relationship between multi-dimensional marine operating condition parameters and various fault discrimination thresholds. Edge terminals collect real-time sea state data and call upon cloud-based mapping rules to iteratively optimize various criterion parameters such as partial discharge threshold, deformation threshold, current surge threshold, and spectral distortion threshold. This enables the discrimination criteria to dynamically and adaptively adjust with real-time sea conditions, ensuring that monitoring criteria align with actual marine operating conditions. This fundamentally improves fault discrimination accuracy under complex sea conditions and reduces the probability of false alarms and missed alarms.

[0065] Step S1521: Collect a large number of historical monitoring samples of submarine cables and corresponding operating condition label data, and screen full-scenario sample resources covering different seawater temperatures, tidal depths, seabed disturbance states, offshore wind power electromagnetic interference intensity, ship stray electromagnetic interference, sea conditions in different seasons, different insulation aging degrees, different years of operation, and different fault strength levels. During this step, a comprehensive sample resource library of submarine cable operating conditions covering all scenarios is compiled and constructed to provide complete data support for threshold mapping model training. A vast amount of historical monitoring samples and corresponding operating condition label data from long-term submarine cable operation are extensively collected. These samples cover different seawater temperature ranges, different tidal depths, different seabed disturbance intensities, different levels of electromagnetic interference from offshore wind power and ships, different seasonal sea states, different insulation aging levels, different equipment operating years, and different fault strength levels. Through multi-scenario, multi-dimensional sample collection, the library comprehensively covers routine marine operating conditions, extreme interference conditions, different stages of submarine cable degradation, and fault states. This completely solves the problems of traditional models having single samples, incomplete scenario coverage, and weak generalization ability, ensuring that subsequent model training can learn the true coupling law between thresholds and sea states under all operating conditions, thus improving the model's universality.

[0066] Step S1522: Clean, deduplicate, remove outliers and standardize the collected raw sample data to construct a high-quality supervised training dataset with one-to-one correspondence between operating parameters, monitoring features and fault labels, and complete the full coverage of typical marine interference scenarios and all dimensions of deteriorated working conditions of submarine cables. In this step, standardized preprocessing is performed on the massive amount of raw samples collected to construct a high-quality model training dataset. Addressing issues such as redundancy, gross errors, anomalous noise, and mislabeling in the raw samples, data cleaning, duplicate sample removal, anomalous error filtering, and data standardization are performed sequentially to unify the data scale and feature dimensions of all samples. A strict correlation is established between operating parameters, monitoring features, and fault labels to form precise supervisory sample pairs with one-to-one correspondence, ensuring that each set of samples has clear operating attribute, monitoring feature, and fault status label. Through refined preprocessing, invalid and inferior samples are thoroughly eliminated, retaining high-value, highly representative, high-quality samples across all operating conditions, achieving full coverage of marine interference scenarios and submersible cable deterioration conditions.

[0067] Step S1523: Construct a correlation mapping network between multi-dimensional marine working condition characteristics and submarine cable monitoring discrimination thresholds. Set seawater temperature, real-time water depth and pressure, tidal current velocity, electromagnetic interference intensity, background noise energy level, and seabed vibration amplitude as model input features, and set partial discharge discrimination threshold, deformation anomaly threshold, current mutation threshold, and spectral distortion threshold as model output targets. Build a multi-input multi-output nonlinear regression mapping model adapted to submarine cable monitoring scenarios. In the specific implementation of this step, a multi-input, multi-output nonlinear regression mapping network adapted to the submarine cable monitoring scenario is constructed to achieve precise quantitative correlation between operating conditions and thresholds. Six core marine operating condition parameters—seawater temperature, real-time water depth and pressure, tidal current velocity, electromagnetic interference intensity, background noise energy level, and seabed vibration amplitude—are selected as model input features to comprehensively characterize the complex real-time marine operating environment. Four core monitoring criteria—partial discharge discrimination threshold, deformation anomaly threshold, current mutation threshold, and spectral distortion threshold—are selected as model output targets, covering the main fault monitoring dimensions of submarine cables. A dedicated nonlinear regression mapping model is constructed to adapt to the complex nonlinear coupling relationship between marine operating conditions and monitoring thresholds, overcoming the shortcomings of insufficient accuracy in traditional linear fitting.

[0068] Step S1524: Based on the constructed full-condition training dataset, supervised training is carried out on the mapping model to continuously fit the inherent coupling law between the dynamic fluctuation of marine conditions and the offset of various monitoring thresholds, automatically learn the optimal offset compensation amount of monitoring thresholds under different sea conditions, generate a standardized dynamic mapping rule library of conditions-thresholds, and establish a quantifiable, iterative, and real-time callable adaptive association mechanism for monitoring thresholds. In this step, supervised training of the mapping model is conducted using a high-quality training dataset covering all operating conditions to uncover the inherent coupling between operating conditions and thresholds. During model training, the model continuously fits the correspondence between dynamic fluctuations in marine operating conditions and various monitoring threshold offsets, autonomously learning the optimal offset compensation for various fault identification thresholds under different sea conditions. This allows for automatic adaptation to complex operating condition changes without human intervention. Through multiple rounds of iterative training, the feature association logic is continuously optimized, forming a standardized, reusable, and real-time callable dynamic mapping rule library for operating conditions and thresholds. This establishes a quantifiable, iterative, and highly adaptable adaptive association mechanism for monitoring thresholds. This effectively solves the drawback of traditional fixed thresholds that cannot dynamically adjust with operating conditions, ensuring that threshold adjustments perfectly match real-world interference and submarine cable operating conditions, significantly improving fault identification adaptability.

[0069] Step S1525: Continuously optimize the network weight parameters of the correlation mapping model through backpropagation of the loss function, and continuously reduce the fitting error between the marine working condition parameters and the monitoring threshold output results. After the model training accuracy converges, batch solidify the multi-dimensional working condition threshold mapping relationship, parameter compensation logic and dynamic correction coefficient into a lightweight inference rule package, and distribute it to each submarine cable field edge terminal, so that the edge terminal has local independent real-time threshold inference capability. In this step, the model weight parameters are continuously optimized through backpropagation of the loss function, constantly reducing the fitting error between the working condition input and the threshold output until the model training accuracy fully converges and the fitting effect reaches its optimal level. After the model training is completed, the mature multi-dimensional working condition threshold mapping relationship, parameter dynamic compensation logic, and working condition correction coefficients are encapsulated and solidified into a lightweight, fast-inference rule package. The lightweight inference rule package is batch-distributed and deployed to various field edge terminals, enabling the edge terminals to store the complete mapping rules locally and have independent real-time threshold inference capabilities without frequent interaction with the cloud to request parameters, significantly reducing communication latency and cloud computing power pressure. This achieves a layered working mode of cloud-based training and modeling and edge-based real-time inference, balancing model accuracy and field real-time requirements.

[0070] Step S1526: Collect on-site dynamic marine operating condition parameters, background interference parameters, real-time sea state data, environmental noise data, electromagnetic base interference data, and real-time operating baseline characteristics of submarine cables in real time through the edge terminal. Input the real-time operating condition characteristics into the locally fixed mapping inference rule package, dynamically calculate the optimal fault discrimination threshold adapted to the current marine operating conditions frame by frame, perform dynamic offset compensation, transmission attenuation correction, and operating baseline calibration on the original default fixed threshold of the system, and iteratively update various submarine cable monitoring criteria parameters in real time. In the specific implementation of this step, the monitoring thresholds are dynamically and adaptively updated in real time using edge terminals. The edge terminals continuously collect real-time data on dynamic marine operating conditions, environmental background interference parameters, electromagnetic base interference data, and the real-time operating baseline characteristics of the submarine cable, comprehensively sensing real-time sea condition changes. The collected real-time operating condition characteristics are input into a locally fixed mapping inference rule package, and the optimal fault discrimination threshold adapted to the current sea condition is dynamically calculated frame by frame. Based on the calculation results, dynamic offset compensation, transmission attenuation correction, and operating baseline calibration are performed on the system's original default fixed thresholds, and various monitoring criteria parameters such as partial discharge, deformation, current, and spectrum are iteratively updated in real time. This ensures that the monitoring thresholds at each moment are adapted to the current marine operating conditions and submarine cable operating status, completely solving the technical problems of rigid discrimination and high false alarm / missed alarm rates of fixed thresholds in complex, time-varying marine scenarios.

[0071] Step S1527: Construct a threshold smoothing update and boundary constraint mechanism, perform amplitude limiting constraint and smoothing filtering on each threshold iteration update process to suppress monitoring and judgment disorder caused by instantaneous threshold jumps, and configure reasonable upper and lower limit constraint ranges adapted to extreme marine conditions for various monitoring thresholds.

[0072] In this step, a dual constraint mechanism of smooth threshold update and boundary constraints is constructed to ensure the stability and reliability of the dynamic threshold adjustment process. To address the issue of instantaneous jumps and erratic fluctuations that can easily occur during frame-by-frame iterative threshold updates, amplitude limiting and smoothing filtering are applied to each threshold update data to delay abrupt fluctuations, ensuring continuous and stable threshold changes and avoiding monitoring judgment errors and erroneous state switching caused by drastic threshold jumps. Simultaneously, based on massive historical operational data, reasonable upper and lower limit constraint ranges are configured for each type of monitoring threshold, adaptable to both normal and extreme sea conditions, preventing excessive threshold offset, threshold failure, and monitoring out of control under extreme sea conditions. Through the dual mechanism of smooth optimization and boundary constraints, the long-term stable operation of the monitoring system is ensured while retaining the threshold's adaptive adjustment capability, balancing adaptability and stability.

[0073] Step S153: Based on long-term time-series monitoring data of submarine cables, extract multi-dimensional insulation degradation indicators and carry out time-series trend modeling, quantify the insulation aging and decay law, construct a fault risk prediction model, explore hidden and gradual degradation characteristics, and complete the preliminary prediction of early defects and potential faults of submarine cables. In its implementation, this step leverages long-term time-series monitoring data of submarine cables to model insulation degradation trends and predict faults in advance. It continuously gathers time-series monitoring data throughout the entire lifecycle of the submarine cable, tracking and statistically analyzing core insulation degradation indicators such as partial discharge frequency, signal distortion degree, cumulative mechanical deformation, and electromagnetic anomaly amplitude to comprehensively characterize the aging and performance degradation status of the cable insulation. A professional time-series fitting algorithm is used to model trends for multi-dimensional degradation indicators, accurately quantifying the insulation aging rate, cumulative degradation degree, and overall performance degradation pattern of the submarine cable. Based on the time-series degradation evolution characteristics, a dedicated fault risk prediction model is constructed to deeply explore the slow evolution patterns of latent insulation degradation and progressive damage, accurately identifying early minute defects and latent fault hazards that are difficult to detect manually. This achieves a technological upgrade for submarine cable faults, moving from traditional post-incident handling to pre-incident prediction and early prevention.

[0074] Step S154: Based on the real-time monitoring status of submarine cables, adaptive threshold parameters, insulation degradation levels and risk prediction results, establish a hierarchical and classified intelligent operation and maintenance decision-making mechanism, match differentiated operation and maintenance strategies for hidden dangers with different degradation levels and risk levels, and complete the accurate output and hierarchical handling of submarine cable fault early warning and operation and maintenance instructions. In the specific implementation of this step, a multi-dimensional, interconnected, hierarchical, and categorized intelligent operation and maintenance decision-making mechanism is constructed to achieve precise and differentiated operation and maintenance handling. Combining real-time monitoring of the submarine cable's operating status, adaptive threshold correction parameters, insulation aging quantification levels, and fault risk prediction probability—multiple core factors—a comprehensive assessment of the submarine cable's potential hazard risk level and degradation grade is conducted. For minor insulation degradation and low-risk hazards, operation and maintenance suggestions with routine tracking and monitoring and key data marking are automatically generated to achieve routine management. For moderate degradation and intermittent anomalies, refined operation and maintenance plans are generated, including fixed-point retesting, continuous status tracking, and specialized operational condition investigations. For severe degradation and high-risk fault hazards, timely and accurate early warning information, fault locations, and hazard levels are pushed out, along with practical operation and maintenance instructions for inspection location, on-site investigation, and equipment maintenance, completely solving the problems of traditional operation and maintenance's one-size-fits-all approach, extensive handling, and lack of specificity.

[0075] Step S155: Establish an edge-cloud intelligent operation and maintenance closed-loop system that integrates monitoring, analysis, early warning, decision-making, and operation and maintenance iteration. Through two-way data interaction between the edge and cloud, continuously iterate and optimize the threshold mapping model and risk prediction model to continuously improve the fault identification accuracy and working condition adaptability under complex marine conditions, and complete the intelligent, adaptive, and self-optimizing operation and maintenance management of submarine cables throughout the entire process.

[0076] During the implementation of this step, an edge-cloud intelligent operation and maintenance closed-loop system is established, integrating monitoring, analysis, early warning, decision-making, operation and maintenance, and iterative optimization to achieve intelligent self-optimizing management of the entire submarine cable process. Through high-speed two-way data interaction between the edge and cloud, real-time monitoring data, threshold adjustment data, and operation and maintenance execution feedback data are continuously transmitted back to the cloud platform. The cloud platform continuously iterates and optimizes the operating condition threshold mapping model and the fault risk prediction model based on the newly added data, constantly improving the model's adaptability to operating conditions and the accuracy of fault identification. Based on the continuous iterative optimization of the model, it empowers the on-site edge monitoring analysis and operation and maintenance decision-making, forming a positive cycle of data update, model optimization, monitoring upgrade, and operation and maintenance optimization. This completely solves the shortcomings of traditional submarine cable operation and maintenance monitoring, such as lag, model rigidity, and inability to adapt to changes in operating conditions, and realizes intelligent, adaptive, and long-term closed-loop management of submarine cable operation and maintenance under complex marine conditions.

[0077] Step S160: Establish a dual mechanism of data iterative update and underwater field calibration to continuously optimize the algorithm model and correct operating condition deviations.

[0078] In the specific implementation of this step, to address the issues of model degradation, operational condition adaptation deviations, and blind spots in remote sensing that are prone to occur during long-term operation of submarine cable monitoring algorithms, a dual optimization mechanism of online data iteration and underwater field calibration is constructed to achieve long-term accurate operation of the algorithm model. Traditional monitoring algorithm models have fixed training samples, which cannot adapt to the changing marine conditions and aging state of submarine cables year by year. The long-term operation results in a continuous decline in discrimination accuracy, and remote monitoring cannot fully match the actual underwater physical state, resulting in inherent extrapolation biases. This technology achieves continuous model adaptation to new operational conditions through dynamic data iteration in the cloud, and corrects algorithm extrapolation biases based on underwater field calibration, forming a dual guarantee system of online adaptive updates and offline accurate error correction. Through bidirectional continuous optimization, monitoring errors are eliminated in the long term, model degradation is suppressed, and the accuracy, stability, and scenario generalization ability of the submarine cable monitoring algorithm are continuously improved.

[0079] Step S161: Establish a cloud-based dynamic data iteration and update mechanism, construct a dynamic sample pool for the entire life cycle of submarine cables, continuously access daily new submarine cable normal operation data, various interference condition data, latent defect evolution data, fault sample data and corresponding real-time marine condition data, implement real-time filtering, automatic cleaning, feature alignment and label verification processing on the new data stream, remove invalid duplicate data and abnormal noise data, and continuously iterate the high-quality new samples after processing into the original model training sample set; In this step, a cloud-based dynamic iterative update mechanism is established to construct a dynamically expandable sample pool for the entire lifecycle of submarine cables. The cloud platform continuously receives daily updates of normal operation data, various marine interference condition data, insulation latent defect evolution data, fault sample data, and corresponding real-time marine condition label data, enabling continuous incremental expansion of sample resources. Automated filtering, intelligent cleaning, feature alignment, and label verification are performed on the daily new real-time data streams, automatically removing duplicate and invalid data, abnormal noise data, and incorrectly labeled samples, retaining high-quality and valid samples. These processed high-quality new samples are continuously iteratively incorporated into the original model training sample set, constantly enriching the operating scenarios, supplementing degraded samples, and updating operating feature characteristics, thus addressing the problems of traditional model sample solidification, incomplete coverage of new scenarios, and long-term degradation of generalization ability.

[0080] Step S162: Establish the algorithm model periodic iterative optimization logic, and adopt a two-level iterative update strategy that combines daily incremental fine-tuning and monthly full retraining. Every day, the model is lightly fine-tuned based on newly added incremental samples to adapt to changes in new sea conditions and equipment operating status in real time. Every month, the fault identification model, threshold mapping model, and risk prediction model are fully retrained based on the accumulated full dynamic sample set. The model weight parameters and bias parameters are updated through backpropagation of the loss function to continuously correct the model feature extraction deviation, working condition adaptation deviation, and fault discrimination deviation, and steadily improve the model's adaptability to complex and changing marine working conditions and fault identification accuracy. In the specific implementation of this step, a two-tiered model iterative optimization strategy of daily fine-tuning and monthly retraining is established to balance real-time adaptability and overall accuracy. Daily, based on newly added incremental sample data, lightweight parameter fine-tuning is performed on the threshold mapping model, fault identification model, and risk prediction model to quickly adapt to new sea conditions, minor changes in equipment operating status, and new interference scenarios, ensuring the model's real-time adaptability. Monthly, based on the full-cycle dynamic sample set accumulated in the cloud, comprehensive retraining is carried out on the three core models. The model weights and bias parameters are globally updated through backpropagation of the loss function, systematically correcting model feature extraction bias, operating condition adaptation bias, and fault identification bias. Through this iterative mode combining long and short cycles, model performance is continuously optimized, steadily improving the model's adaptability to complex and changing marine conditions and fault identification accuracy.

[0081] Step S163: Construct an underwater multi-mode on-site calibration operation system. Relying on underwater inspection robots, underwater sensor monitoring terminals, and manual fixed-point verification, conduct periodic underwater on-site inspections and parameter calibrations for key laying sections, high-risk sections, and historically abnormal sections of submarine cables. Collect physical truth data such as the actual deformation state of the submarine cable, seabed scouring state, cable sheath aging state, and actual underwater electromagnetic environment conditions on-site. Construct a one-to-one correspondence between cloud algorithm judgment results and underwater on-site truth data. This step involves constructing a multi-faceted underwater field calibration system to obtain real underwater physical data to correct algorithmic deviations. Utilizing a multi-faceted verification approach combining underwater inspection robots, dedicated underwater sensing terminals, and manual on-site verification, periodic field inspections are conducted, focusing on key cable laying sections, high-risk and vulnerable sections, and sections with a history of frequent anomalies. Accurate on-site data is collected on the cable's actual deformation, seabed erosion status, cable sheath aging and damage status, and the underwater electromagnetic environment and stray current conditions to accurately recreate the actual underwater operating conditions. The results of the cloud-based algorithm are matched one-to-one with the underwater field data to construct a complete pair of algorithm predictions and field data, providing a reliable offline calibration basis for subsequent deviation quantification and model correction.

[0082] Step S164: Implement targeted correction of algorithm operating condition deviations. Compare the judgment results such as fault location, degradation level, abnormal amplitude, and risk probability output by the cloud model with the true value data of underwater field survey point by point. Quantitatively solve the algorithm prediction deviation, feature offset deviation, and threshold adaptation deviation. For operating conditions, fault types, and sea state intervals with deviations, target and correct the model feature fusion weight, operating condition compensation coefficient, and threshold mapping rules to accurately compensate for the underwater environmental perception blind spots and inference errors in the remote monitoring algorithm. In its implementation, this step employs a precise algorithm deviation correction mechanism to compensate for the blind spots and extrapolation errors inherent in remote monitoring. The intelligent assessment results output by the cloud model, including fault location, insulation degradation level, abnormal fluctuation amplitude, and fault risk probability, are precisely compared point-by-point and scenario-by-scenario with the true physical data obtained from underwater field surveys. This quantitatively calculates model prediction bias, feature offset bias, and threshold adaptation bias. It accurately identifies the algorithm's assessment shortcomings under different sea conditions, fault types, and degradation stages, and specifically corrects the model's feature fusion weights, marine condition compensation coefficients, and threshold mapping rules. This precisely compensates for the limitations of remote monitoring in perceiving subtle underwater physical changes and the inherent errors in algorithm extrapolation, significantly improving the model's assessment accuracy in real, complex underwater scenarios.

[0083] Step S165: Establish a deviation archiving and boundary scenario completion mechanism. Archive the deviation data, correction parameters, and real working condition samples generated from each field calibration into a dedicated deviation correction sample library. Supplement the model with difficult sample scenarios such as extreme sea conditions, niche interference, and hidden slow degradation of the model, and improve the model boundary scenario processing logic. In the specific implementation of this step, a deviation archiving and boundary scenario completion mechanism is established to continuously improve the model's boundary processing capabilities. All deviation data, model correction parameters, and real-world operating condition samples obtained from each underwater field calibration are uniformly archived to build a dedicated submarine cable deviation correction sample library, enabling traceability and reusability of deviation issues. The focus is on collecting challenging scenario data that are scarce in traditional sample libraries, such as extreme sea conditions, niche and special interferences, and long-term, latent, slow degradation, to specifically supplement model boundary scenario samples. By continuously supplementing boundary challenge samples and archiving correction patterns, the model's processing logic for extreme operating conditions, special interferences, and weak latent degradation scenarios is continuously improved, effectively avoiding the recurrence of deviations under similar operating conditions and continuously addressing the algorithm's adaptation shortcomings and recognition blind spots in complex underwater real-world scenarios.

[0084] Step S166: Construct a two-way closed-loop optimization mechanism of online data iteration and underwater field calibration. The model is dynamically adapted to time-varying ocean conditions through online massive real-time data iteration, and the inherent bias is deduced through offline underwater field true value calibration to correct the algorithm. The complementary synergy between online iterative optimization and offline field correction is relied upon.

[0085] This step involves constructing a two-way closed-loop optimization system of online data iteration and underwater field calibration to achieve long-term self-optimization and upgrading of the algorithm model. Relying on continuous incremental data iteration in the online cloud, the model dynamically adapts to time-varying marine conditions and the long-term aging changes of submarine cables, ensuring real-time scenario adaptability. Based on offline underwater field calibration, algorithm inference biases are continuously corrected, perception blind spots are filled, and boundary scene handling capabilities are improved, addressing the deficiency of purely data-driven models lacking physical truth constraints. Through the complementary synergy of online adaptive updates and offline precise error correction, the system effectively suppresses judgment biases caused by marine environmental interference, equipment transmission errors, and insufficient model generalization, continuously improving the model's fault identification accuracy, operational condition adaptability, and operational stability, achieving closed-loop optimization of the submarine cable monitoring algorithm throughout its entire lifecycle.

[0086] Based on the same inventive concept, please refer to Figure 2 The diagram shows a schematic block diagram of a submarine cable long-term operation and maintenance monitoring system 100 provided in this application embodiment for performing the above-described submarine cable long-term operation and maintenance monitoring method. The submarine cable long-term operation and maintenance monitoring system 100 may include a communication unit 110, a machine-readable storage medium 120, and a processor 130.

[0087] In this embodiment, both the machine-readable storage medium 120 and the processor 130 are located within the long-term operation and maintenance monitoring system 100 for submarine cables and are separately configured. However, it should be understood that the machine-readable storage medium 120 may also be independent of the long-term operation and maintenance monitoring system 100 for submarine cables and may be accessed by the processor 130 via a bus interface. Alternatively, the machine-readable storage medium 120 may also be integrated into the processor 130 and may communicate and interact with external systems through the communication unit 110.

[0088] The processor 130 is the control center of the long-term operation and maintenance monitoring system 100 for submarine cables. It connects various parts of the system via various interfaces and lines, and performs various functions and processes data by running or executing software programs and / or modules stored in the machine-readable storage medium 120, and by calling data stored in the machine-readable storage medium 120, thereby providing overall monitoring of the system. Optionally, the processor 130 may include one or more processing cores; for example, the processor 130 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may not be integrated into the processor. The machine-readable storage medium 120 stores machine-executable instructions for executing the scheme of this application, and the processor 130 executes the machine-executable instructions stored in the machine-readable storage medium 120 to implement the long-term operation and maintenance monitoring method for submarine cables provided in the aforementioned method embodiments.

[0089] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

[0090] The embodiments of this application have been described above with reference to the accompanying drawings. Unless otherwise specified, the embodiments and features in the embodiments of this application can be combined with each other. This application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for long-term operation and maintenance monitoring of submarine cables, characterized in that: Includes the following steps: By combining and deploying multiple types of sensing hardware, implementing full-domain time synchronization and hardware anti-interference encapsulation, a multi-parameter submarine cable full-domain raw data acquisition system is constructed. A dual-layer differentiated noise reduction architecture is adopted, combined with feature extraction and database comparison technology, to filter out marine random noise and fixed interference from artificial equipment in layers, thereby completing the extraction and identification of weak partial discharge signals; By integrating multi-source monitoring data and introducing dynamic correction parameters and deep learning models, it can adaptively adapt to marine conditions. By integrating frequency domain detection technology with traditional time domain monitoring methods, all types of submarine cable fault monitoring and verification can be achieved through photoelectric bidirectional verification. Based on the edge-cloud collaborative architecture, the monitoring threshold and working conditions are adaptively adjusted to quantify the insulation aging trend, predict fault risks in advance, and build a closed-loop system for intelligent monitoring and maintenance of submarine cables. Establish a dual mechanism of data iteration and underwater field calibration to continuously optimize the algorithm model and correct operating condition deviations.

2. The long-term operation and maintenance monitoring method for submarine cables according to claim 1, characterized in that: The system constructs a multi-parameter submarine cable full-domain raw data acquisition system through the combined deployment of multiple types of sensing hardware, full-domain time synchronization, and hardware anti-interference encapsulation, including: Based on the laying mileage, laying water depth, seabed geological conditions and historical fault distribution characteristics of the submarine cable, a multi-sensor hardware fusion networking method is adopted to implement differentiated segmented deployment of the entire submarine cable. Integrated comprehensive sensing units are uniformly distributed in the conventional straight submarine cable sections, and sensing nodes are densely deployed in weak and high-risk sections such as the submarine cable landing section, underwater intermediate joint, bend pipe laying section, reef friction section, and shallow sea tidal erosion section. A multimodal integrated sensing hardware acquisition array was constructed, integrating distributed optical fiber sensing units, high-frequency current traveling wave acquisition terminals, and micro quantum magnetometers in a collaborative network. Distributed optical fiber sensing units were used to collect data on submarine cable temperature, cable deformation, axial strain, and vibration disturbances throughout the cable laying process. High-frequency current traveling wave acquisition terminals installed close to the submarine cable's metal sheath and grounding end were used to capture line current mutations and fault traveling wave pulse signals at high frequency. Micro quantum magnetometers deployed close to the outer wall of the submarine cable were used to collect high-precision electric and magnetic field disturbances around the submarine cable and magnetic distortion signals induced by stray currents in seawater. A full-domain nanosecond-level time synchronization system is established. A high-precision satellite time synchronization module is configured at the land-based main control terminal to output a unified reference time signal. The time is synchronized to all underwater sensing terminals through the submarine cable communication core line. Each underwater terminal has a built-in clock calibration chip for real-time dynamic time synchronization, which controls the time synchronization error of all nodes to the nanosecond level. A transmission delay compensation algorithm is configured to correct the time offset in real time based on the node laying distance and signal transmission rate. All underwater sensing terminals are protected against interference by hardware in all dimensions. They are encapsulated with titanium alloy anti-corrosion shell and insulating waterproof sealant. The terminal is blocked by multiple layers of high magnetic permeability shielding and insulating isolation layer to block electromagnetic radiation interference from offshore wind power, ships and platforms and stray current coupling interference from seawater. At the same time, the wiring ports and signal interfaces of the equipment are designed with integrated sealed shielded connectors. After completing the hardware network, each sensing node was powered on and debugged, signal sampling was tested and sensitivity was calibrated. The sampling frequency, sampling accuracy, data storage format and transmission protocol of each terminal were unified. A real-time self-checking and abnormal alarm mechanism for terminal operating voltage, temperature, signal strength and operating status was established. Finally, a full-coverage, multi-parameter, high-purity and high-synchronization raw data acquisition system for submarine cables was built.

3. The long-term operation and maintenance monitoring method for submarine cables according to claim 1, characterized in that: The method employs a dual-layer differentiated noise reduction architecture combined with feature extraction and database comparison techniques to filter out random ocean noise and fixed interference from artificial equipment in layers, thereby completing the extraction and identification of weak partial discharge signals, including: Based on massive submarine cable monitoring data, marine interference signals are classified and defined in advance. The marine interference signals are divided into two types: broadband random Gaussian noise generated by ocean current disturbances and temperature fluctuations, and fixed frequency narrowband electromagnetic interference generated by offshore wind power, ship electrical systems, and offshore platform power supply equipment. The partial discharge pulse signal of submarine cable is set as the target effective signal. The high-frequency raw monitoring signals collected on site are pre-amplified and normalized to unify the signal amplitude range and sampling scale. An improved empirical wavelet transform noise reduction mechanism is adopted. In response to the broadband and irregular random noise characteristics of the marine environment, adaptive dynamic frequency band segmentation and multi-mode decomposition are implemented on the preprocessed high-frequency monitoring signal. Based on the real-time frequency domain distribution characteristics, the noise frequency band and the effective signal frequency band are autonomously distinguished. Through an iterative screening mechanism, the random noise components of the ocean without fixed shape are accurately removed, while the effective characteristic signals of partial discharge exclusive short-time pulse and sudden attenuation are adaptively retained. Based on the parameterized intelligent notch filter noise reduction architecture, for fixed artificial electromagnetic narrowband interference generated by offshore wind power, ships and offshore platforms, the main frequency characteristics of the interference are dynamically captured by real-time spectrum monitoring, and the bandwidth, attenuation coefficient and suppression depth of the notch filter are adaptively generated and iteratively optimized in real time. Based on the Morse wavelet-TT joint feature extraction architecture, the Morse wavelet's multi-scale high-resolution fine decomposition capability is used to mine the detailed features of weak pulses. Combined with the TT time-series trajectory analysis method, the temporal evolution law of partial discharge signal is characterized, and multi-dimensional partial discharge feature parameters are extracted and quantized. Intelligent comparison and identification are carried out based on a pre-constructed marine environmental noise sample library and a partial discharge feature library of submarine cable insulation defects. The dual databases contain a massive amount of environmental interference waveforms, artificial electromagnetic interference waveforms, and partial discharge feature waveforms of different aging degrees and different defect types under different sea conditions and operating scenarios. The multi-dimensional feature parameters of the signals extracted in real time are intelligently matched and compared with the database samples. The feature similarity threshold judgment logic automatically distinguishes marine random noise, electromagnetic interference from artificial equipment, and real insulation partial discharge signals, filters early weak partial discharge signals, identifies minor insulation degradation and latent defects of submarine cables, and completes the extraction and reliable identification of submarine cable partial discharge signals under marine electromagnetic environment.

4. The long-term operation and maintenance monitoring method for submarine cables according to claim 3, characterized in that: The improved empirical wavelet transform noise reduction mechanism, targeting the broadband and irregular random noise characteristics of the marine environment, performs adaptive dynamic frequency band segmentation and multi-mode decomposition on the preprocessed high-frequency monitoring signal. Based on real-time frequency domain distribution characteristics, it autonomously distinguishes noise frequency bands from effective signal frequency bands. Through an iterative screening mechanism, it accurately removes irregular marine random noise components and adaptively retains effective characteristic signals such as short-time pulses and sudden attenuation signals specific to partial discharges, including: Read the high-frequency raw signal after pre-amplification and normalization, perform a global fast Fourier transform on the whole signal, and solve to obtain the signal spectrum, energy spectrum and power spectrum. Statistically analyze the energy ratio and amplitude fluctuation range of different frequency intervals, capture the distribution characteristics of broadband dispersion and energy diffusion of marine random noise, and mark the frequency domain intervals where partial discharge pulse signals are concentrated to establish a frequency domain characteristic judgment benchmark for signals and noise. The improved empirical wavelet transform algorithm is used to perform multi-mode decomposition on the signal from low frequency to high frequency. Based on the real-time spectral energy inflection point and amplitude change point, the frequency band is dynamically and adaptively divided so that the frequency range and bandwidth of each sub-band are adapted to the current signal frequency domain characteristics, and multi-order adaptive mode components are generated. Four types of characteristic parameters are extracted for each modal component: time-domain waveform, energy concentration, waveform continuity, and amplitude change rate. Modal attribute identification is completed by combining the preset judgment rules. Modes with dispersed energy, messy waveforms, and no obvious abrupt change characteristics are judged as ocean random noise components, while modes with short-term amplitude jumps and sudden and attenuated waveform characteristics are judged as effective partial discharge signal components. An iterative screening and purification process is initiated to remove the pure noise mode components identified in the first round. Multi-mode decomposition and feature re-examination are repeatedly carried out on the remaining mixed mode components. The process iteratively investigates and removes residual scattered random noise components, and locks the weak pulse waveform characteristics throughout the process to ensure the effective signal amplitude and waveform integrity. After multiple rounds of iterative screening, all effective modal components are reconstructed to restore a clean electrical signal with complete timing and waveform characteristics, thereby achieving precise suppression of irregular random noise in the ocean broadband and outputting the preliminarily purified signal to the subsequent signal processing stage.

5. The method for long-term operation and maintenance monitoring of submarine cables according to claim 3, characterized in that: The aforementioned parameterized intelligent notch filter noise reduction architecture addresses fixed artificial electromagnetic narrowband interference generated by offshore wind power, ships, and offshore platforms. It dynamically captures the dominant frequency characteristics of the interference through real-time spectrum monitoring, adaptively generates and iteratively optimizes the notch filter bandwidth, attenuation coefficient, and suppression depth in real time. This includes: Real-time spectrum acquisition and continuous spectrum tracking are carried out on the purified signal after the first layer of marine random noise reduction. The signal is analyzed frame by frame by short-time Fourier transform to extract the peak point, main frequency position and peak energy distribution characteristics of the signal in real time. The narrowband fixed frequency interference components generated by the operation of offshore wind power, ships and offshore platform electrical equipment are continuously monitored to lock the set of dynamic artificial interference main frequencies under the field conditions and distinguish between stable power frequency interference and equipment operation interference frequencies with small drift. Based on the real-time captured interference main frequency characteristics, the intelligent notch filter parameter adaptive generation logic is started. According to the peak energy, frequency concentration, and interference superposition intensity of the interference main frequency, the notch filter center frequency, basic bandwidth, initial attenuation coefficient and suppression depth parameters are automatically initialized to generate an initial parameterized notch filter model that adapts to the current field interference characteristics. A dynamic fine-tuning mechanism for notch filter parameters is constructed to link operating conditions. Subsequent spectrum data is continuously collected in real time to track the slight drift of the main interference frequency, the changes of multiple frequency superposition, and the fluctuation of interference energy strength. The notch filter parameters are dynamically corrected frame by frame. The suppression bandwidth is adaptively widened for multi-device superposition interference scenarios, and the suppression depth is adaptively reduced for weak interference high stable frequency scenarios. By implementing targeted frequency domain suppression through a smart notch filter with dynamic parameter matching, the narrowband interference frequency band corresponding to artificial equipment in the signal frequency domain is attenuated and canceled at a fixed depth. The non-interference frequency band and the effective signal frequency band of partial discharge pulse are preserved throughout the process. Under the premise of completely preserving the weak partial discharge pulse signal, the artificial narrowband interference is stripped away, and a high-purity signal is output that simultaneously eliminates ocean broadband random noise and fixed narrowband interference from artificial equipment.

6. The method for long-term operation and maintenance monitoring of submarine cables according to claim 3, characterized in that: The aforementioned Morse wavelet-TT joint feature extraction architecture utilizes the multi-scale, high-resolution, fine-decomposition capability of the Morse wavelet to mine the detailed features of weak pulses, and combines it with the TT time-series trajectory analysis method to characterize the temporal evolution of partial discharge signals, extracting and quantizing partial discharge feature parameters in multiple dimensions, including: High-purity submarine cable monitoring signals with high signal-to-noise ratios after double-layer noise reduction are obtained as input for feature mining. In view of the characteristics of weak, latent partial discharge pulses with low amplitude, easy loss of details and low feature recognition under marine conditions, Morse wavelet multi-scale fine decomposition processing is initiated. Based on the frequency domain characteristics of submarine cable partial discharge pulses, Morse wavelet waveform parameters and scale factors are adaptively matched. The weak partial discharge pulse signal is analyzed layer by layer through a multi-scale refinement subdivision method. The subtle noise interference remaining in the signal is removed layer by layer, and the subtle waveform details of ultra-low energy partial discharge pulses are completely preserved, and the information of minute pulse change is captured. Based on the results of Morse wavelet multi-scale refinement decomposition, quantitative extraction of multidimensional features of partial discharge is carried out. The refined pulse waveform data is analyzed segment by segment, and six quantitative features are extracted: pulse rise edge characteristics, exponential decay characteristics, effective pulse width, amplitude distribution law, pulse repetition frequency and time interval between adjacent pulses. This covers the microscopic morphological features of single pulses and the pulse temporal distribution features, and a complete dataset of original features of partial discharge is constructed. By introducing TT timing trajectory analysis technology, the timing characteristic parameters of partial discharge pulses extracted in continuous time periods are reconstructed to transform discrete single-point pulse characteristics into continuous timing evolution trajectories. By analyzing the fluctuation trend of partial discharge pulse amplitude, the variation law of pulse occurrence interval, and the discrete characteristics of pulse aggregation, the dynamic evolution process of submarine cable insulation from slight deterioration and latent discharge to intermittent discharge is depicted, and the timing correlation characteristics of latent defects that cannot be characterized by single waveform features are discovered. By integrating the pulse micro-waveform features extracted by Morse wavelet with the macro-evolution features represented by TT time-series trajectory, normalization and quantization processing, redundant feature removal and dimension regularization optimization are carried out on all feature parameters to construct a standardized, quantifiable, distinguishable and intelligently comparable set of submarine cable partial discharge feature parameters, thereby realizing the explicit transformation of weak, latent partial discharge signal features with strong concealment and ambiguous features.

7. The long-term operation and maintenance monitoring method for submarine cables according to claim 1, characterized in that: The method of fusing multi-source monitoring data and introducing dynamic correction parameters and deep learning models to adaptively adapt to marine conditions includes: By fully integrating heterogeneous monitoring data of submarine cables in mechanical, electrical, and electromagnetic fields, as well as marine working conditions and environmental data, the time series benchmark and format of multi-source data are unified. Data cleaning, gross error removal, and normalization preprocessing are completed to eliminate data deviations caused by differences in data collection from multiple devices, and a multi-source fusion monitoring dataset of submarine cables with regular time series and complementary dimensions is constructed. A dynamic parameter correction system for operating conditions adapted to time-varying marine scenarios is constructed. Multiple dynamic correction parameters are introduced, including seawater deep pressure attenuation correction factor, seawater temperature transmission loss correction factor, tidal scouring operating condition offset correction factor, offshore wind power and ship electromagnetic interference intensity correction factor, and long-distance submarine cable signal transmission delay correction factor. The parameters of each correction factor are dynamically updated iteratively based on real-time collected marine environmental operating condition data, and point-by-point dynamic compensation correction is implemented for multi-source monitoring raw data. A deep learning feature fusion model for submarine cable fault identification was constructed. Using dynamically corrected standardized multi-source monitoring data as input, a multi-layer feature fusion neural network structure was built. The shallow network was used to extract low-dimensional representations of optical fiber mechanical features, electrical traveling wave features, and magnetic field distortion features. The middle feature fusion layer was used to achieve cross-fusion and complementary enhancement of multi-dimensional heterogeneous features, automatically mining the implicit correlation features and coupling change patterns between multi-source data. The deep network was used to abstract and condense high-dimensional fusion features and remove redundant features, and screen out fusion features that are strongly correlated with submarine cable insulation degradation, partial discharge, mechanical damage, and external interference. A sample dataset of submarine cable operating conditions covering all scenarios is constructed. The sample dataset contains data on normal operating conditions, interference operating conditions, latent defect operating conditions, and explicit fault operating conditions corresponding to different sea conditions, different interference intensities, different fault types, and different insulation aging degrees. The deep learning feature fusion model is iteratively trained using supervised learning. The model weight parameters and bias parameters are continuously optimized through backpropagation of the loss function to correct the model feature extraction and feature fusion logic. The iteratively optimized deep learning model is deployed to the real-time monitoring inference stage. It performs continuous feature fusion and intelligent discrimination on multi-source monitoring data that is accessed in real time and dynamically corrected. The model dynamically and adaptively adjusts the feature weights and discrimination thresholds according to the real-time marine conditions. It can adapt to time-varying marine scenarios such as tidal changes, electromagnetic interference fluctuations, and seabed environment changes without manual parameter debugging. It outputs the submarine cable operation status discrimination results, defect types, and hazard levels in real time.

8. The long-term operation and maintenance monitoring method for submarine cables according to claim 7, characterized in that: The aforementioned deep learning feature fusion model for submarine cable fault identification uses dynamically corrected standardized multi-source monitoring data as input. A multi-layer feature fusion neural network structure is constructed. The shallow network extracts low-dimensional representations of fiber optic mechanical features, electrical traveling wave features, and magnetic field distortion features. The middle-layer feature fusion layer achieves cross-fusion and complementary enhancement of multi-dimensional heterogeneous features, automatically mining implicit correlations and coupling patterns among multi-source data. The deep network abstracts and refines high-dimensional fusion features and removes redundant features, selecting fusion features strongly correlated with submarine cable insulation degradation, partial discharge, mechanical damage, and external interference, including: The standardized multi-source monitoring data after dynamic correction is split into dimensions and channelized. The fiber optic mechanical data, electrical traveling wave data, and magnetic field distortion data are divided into corresponding independent input channels, and the data length, number of sampling points, and feature dimensions of each channel are unified. A hierarchical deep learning feature fusion network was constructed. A branched shallow feature extraction subnetwork was set in the first layer of the network. Independent one-dimensional convolution and pooling structures were configured for three types of heterogeneous data: mechanical, electrical, and electromagnetic. The mechanical subnetwork was used to extract waveform features and abrupt change features corresponding to temperature, deformation, and strain. The electrical subnetwork was used to extract pulse amplitude, rise time, and decay characteristics. The electromagnetic subnetwork was used to extract magnetic field distortion amplitude, fluctuation frequency, and stray current features, thus completing the extraction of low-dimensional basic features of various single-source data. A mid-level cross-modal adaptive feature fusion layer is constructed, and a fusion strategy combining dimensional splicing fusion and dynamic weighted fusion is adopted. The feature vectors output by the three shallow subnets are spliced ​​dimensionally. At the same time, an adaptive feature weight allocation mechanism is introduced to dynamically adjust the fusion weights of mechanical, electrical and electromagnetic features according to the real-time marine conditions and signal strength. This strengthens the weights of high-contribution effective features and weakens the interference of low-correlation redundant features, thereby completing the deep cross-fusion of multi-dimensional heterogeneous features and automatically mining the implicit correlation characteristics and time-varying coupling change patterns between multi-source monitoring data. A deep high-dimensional feature condensation network containing multi-layer convolutional structure, residual module and global pooling structure is constructed. The fusion features output by the middle layer are abstracted layer by layer, feature compression and deep representation learning are performed. Noise features, redundant features and weakly correlated features that are not related to the operation status of submarine cable are automatically filtered out. High-dimensional fusion features that are highly correlated with submarine cable insulation degradation, partial discharge, mechanical damage and external electromagnetic interference are obtained layer by layer. The high-dimensional fusion features are normalized and ranked by contribution through a fully connected layer and a feature saliency screening module. Adaptive screening is performed based on the contribution of each feature to the classification and status determination of submarine cable faults, retaining highly significant fusion features and eliminating invalid and redundant features.

9. The method for long-term operation and maintenance monitoring of submarine cables according to claim 1, characterized in that: The fusion of frequency domain detection technology and traditional time domain monitoring methods enables the monitoring and verification of all types of submarine cable faults through photoelectric bidirectional verification, including: A dual-channel synchronous monitoring link in the time and frequency domains was established. High-frequency traveling waves and distributed optical fiber sensing devices were used to synchronously collect various time-domain operating characteristics of submarine cables. The full-domain fast Fourier transform and short-time Fourier transform were combined to complete the frame-by-frame frequency domain feature analysis of the signal. Time-series alignment and unified labeling were implemented for the time-frequency heterogeneous monitoring data to construct a time-series regular and dimensionally complementary submarine cable time-frequency synchronous monitoring dataset. Establish a two-way independent detection mechanism in the time domain and frequency domain. Based on time domain monitoring data, capture transient time domain characteristics such as pulse mutation, waveform distortion, traveling wave reflection, strain anomaly, and temperature mutation to identify abnormal time domain conditions such as submarine cable short circuit, line break, mechanical damage, sudden partial discharge, and external force disturbance. Relying on frequency domain data, capture abnormal frequency domain characteristics such as abnormal spectral energy shift, characteristic frequency band mutation, harmonic component surge, and electromagnetic distortion frequency point anomaly to identify insulation degradation, latent continuous discharge, marine electromagnetic interference, equipment noise interference, and long-distance transmission dispersion anomaly latent conditions that are difficult to identify in the time domain. This forms a differentiated detection mode for transient explicit faults monitored in the time domain and slow-changing implicit faults monitored in the frequency domain. A photoelectric bidirectional cross-verification mechanism for fiber optic monitoring and electrical time-frequency monitoring is constructed. The physical deformation, vibration disturbance, and temperature anomaly point information collected by fiber optic sensing are compared and verified with the electrical fault characteristics obtained by time-domain and frequency-domain analysis in three dimensions: space, time sequence, and characteristics. When anomalies are detected simultaneously in the time domain and frequency domain and the optical sensing point status changes synchronously, it is determined to be a real submarine cable fault. When anomalies occur only in a single time domain or frequency domain and there are no corresponding disturbance characteristics in the optical sensing data, it is determined to be a false fault caused by marine random noise, equipment electromagnetic interference, or data jitter. A time-frequency joint judgment model adapted to different fault types is constructed. For physical faults such as mechanical damage and external dragging, the model combines the time-domain deformation mutation characteristics of optical fibers with the frequency-domain high-frequency disturbance energy concentration characteristics to complete the joint judgment. For faults such as partial discharge and insulation aging and deterioration, the model relies on the time-domain weak pulse accumulation characteristics and the frequency-domain continuous distortion band characteristics to achieve accurate identification. For electrical faults such as short circuit, grounding, and open circuit, the model uses the time-domain traveling wave mutation characteristics and the frequency-domain fundamental wave offset and harmonic surge characteristics to complete the fault differentiation, comprehensively covering the types of obvious sudden faults and latent continuous faults in submarine cables. By integrating multi-dimensional verification conclusions from transient features in the time domain, latent features in the frequency domain, and optical and physical features, the final operating status of the submarine cable is output, and the fault type, fault location, fault occurrence time, fault degradation degree, and interference source information are determined and output simultaneously.

10. A long-term operation and maintenance monitoring system for submarine cables, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the long-term operation and maintenance monitoring method for submarine cables according to any one of claims 1 to 9 by executing the machine-executable instructions.

Images