A multi-parameter perception submarine cable fault signal identification method

Abstract

The present application relates to the technical field of submarine cable fault monitoring and operation and maintenance, in particular to a multi-parameter sensing submarine cable fault signal identification method, which comprises multi-source signal active excitation, multi-parameter data acquisition, signal preprocessing and feature extraction, and fault identification and positioning; the multi-source signal active excitation generates a wide frequency voltage signal and an optical fiber laser signal of 0.1-100MHz, forming an integrated source assembly; the multi-parameter data acquisition synchronously collects wide frequency impedance, temperature, stress and magnetic field multi-dimensional response data of the submarine cable; the signal preprocessing and feature extraction use an adaptive noise suppression algorithm to process the original signal, extract the characteristic parameters and construct a fault characteristic matrix; the fault identification and positioning integrate the features based on a dynamic weight fusion strategy, input an improved deep belief network model, and complete fault classification and positioning; the present application can realize the transformation from passive repair to active early warning, significantly reduce the operation and maintenance cost of submarine cables, and ensure the stable operation of marine energy and information networks.

Description

Technical Field

[0001] This invention relates to the field of submarine cable fault monitoring and maintenance technology, specifically a method for identifying submarine cable fault signals using multi-parameter sensing. Background Technology

[0002] As the global energy transition accelerates, the scale of offshore wind farms and transoceanic power transmission projects continues to expand, and the laying length and operational pressure of submarine cables are also increasing in tandem.

[0003] Submarine cables operate in a high-pressure, high-humidity, and highly corrosive marine environment, making them extremely susceptible to damage from factors such as ship anchoring, tectonic activity, tidal erosion, and biological corrosion. This can lead to issues like insulation aging, mechanical damage, water ingress, and short circuits. Once a fault occurs, navigation must be closed for emergency repairs, which are very expensive and can cause a chain reaction, including shutdowns of offshore platforms, power outages in the region, and communication network failures, even threatening the safety of personnel working at sea.

[0004] Therefore, utilizing multi-dimensional sensing technology to achieve online monitoring of submarine cable operation, sensing the transient real-time changes in local stress on the submarine cable due to external forces such as anchor damage, and judging the impact on the submarine cable is of great value in preventing submarine cable damage. Therefore, a multi-parameter sensing submarine cable fault signal identification method is proposed to improve the above-mentioned problems. Summary of the Invention

[0005] To address the technical problems of traditional submarine cable fault identification methods that rely on a single signal source (such as OTDR, broadband impedance method, etc.), resulting in weak anti-interference capability, low fault location accuracy, and delayed identification of hidden defects, this invention provides a multi-parameter sensing submarine cable fault signal identification method to solve the above problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A method for identifying submarine cable fault signals using multi-parameter sensing includes the following steps: Step S1: Active excitation of multi-source signals: Generate a wideband voltage signal and fiber laser signal of 0.1-100MHz to form an integrated source component. Use frequency division multiplexing and wavelength division multiplexing technology to inject the signal into the submarine cable, and realize the coordinated control of optical and electrical signals through PID model.

[0007] Step S2: Multi-parameter data acquisition: Simultaneously acquire broadband impedance, temperature, stress, and magnetic field response data of the submarine cable.

[0008] Step S3: Signal preprocessing and feature extraction: The original signal is processed using an adaptive noise suppression algorithm to extract feature parameters and construct a fault feature matrix.

[0009] Step S4: Fault identification and localization: Based on the dynamic weight fusion strategy, integrate features and input them into the improved deep belief network model to complete fault classification and localization.

[0010] As a preferred embodiment of the present invention, the broadband voltage signal in step S1 includes three modes: continuous periodic sine wave signal, step voltage signal, and pseudo-random sequence wave; the minimum resolution for magnetic field acquisition is 1 nT, the resolution for fiber optic temperature acquisition is 0.1℃, and the expression of the PID model is:

[0011] in and These are the real-time control quantities for wideband voltage signals and laser signals, respectively.

[0012] In a preferred embodiment of the present invention, the excitation frequency band is adaptively selected based on the submarine cable length and voltage level in step S1: When the length of the submarine cable L ≤ 50km, select 50-100MHz.

[0013] When 50km < L ≤ 150km, select 10-50MHz.

[0014] When L > 150km, select 0.1-10MHz.

[0015] When the rated voltage is ≤110kV, the lower limit of the excitation frequency band is 0.1MHz.

[0016] When the voltage is between 110kV and 220kV, the lower limit is 1MHz.

[0017] The lower limit is 5MHz when the voltage is greater than 220kV.

[0018] As a preferred embodiment of the present invention, the adaptive noise suppression algorithm in step S3 employs an improved wavelet threshold denoising algorithm, specifically including: The original signal was decomposed into three levels of wavelet, and the db4 wavelet was selected as the basis wavelet.

[0019] Calculate the noise standard deviation of the high-frequency components of each layer. and adaptive threshold .

[0020] in, For the first j Adaptive threshold for high-frequency components in the layer. For the first j The noise standard deviation of the high-frequency components of the layer. This represents the number of signal sampling points.

[0021] The following threshold function is used for processing:

[0022] in, For the processed first j Layer k A high-frequency coefficient, To process the first j Layer k A high-frequency coefficient.

[0023] Finally, the signal is reconstructed using inverse wavelet transform.

[0024] As a preferred embodiment of the present invention, the feature parameters extracted in step S3 include: The five categories of wideband impedance are resonant frequency, resonant amplitude, dielectric loss factor, equivalent capacitance, and equivalent inductance.

[0025] The temperature monitoring points are categorized into four types: average temperature at monitoring points, temperature difference between adjacent monitoring points, temperature change rate, and peak temperature.

[0026] The stress is classified into four categories: peak stress, mean stress, stress variation range, and number of stress abrupt changes.

[0027] The magnetic field is classified into four categories: magnetic field amplitude, magnetic field distortion rate, magnetic field change frequency, and magnetic field peak value.

[0028] The fault feature matrix is ​​an M×17 dimensional matrix, where M is the number of monitoring points, 17 is the total number of feature parameters, and the matrix elements are normalized values.

[0029] As a preferred embodiment of the present invention, the dynamic weight fusion strategy in step S4 adopts a two-factor dynamic weight allocation based on fault type and signal-to-noise ratio, and the weight allocation function is:

[0030] in, For the first i Weights of class parameters This parameter is the matching coefficient between itself and the current fault type. For this parameter Signal-to-noise ratio adaptation coefficient.

[0031] As a preferred embodiment of the present invention, the improved deep belief network model in step S4 includes 5 RBM hidden layers and 1 Softmax output layer, adopts a combination structure of Gaussian-Bernoulli RBM and Bernoulli-Bernoulli RBM, and introduces Dropout layer and L2 regularization; the pre-training adopts contrastive divergence CD-1, and the fine-tuning adopts Adam optimizer and cross-entropy loss function.

[0032] As a preferred embodiment of the present invention, step S4 further includes a graded early warning system for latent defects, which is divided into three levels according to the degree of deterioration: mild, moderate, and severe. A mild warning is defined as a single key characteristic parameter deviating from the baseline by 10%-20% and remaining lifetime > 80%. A moderate warning is defined as a deviation of 20%-40% and a remaining lifespan of 40%-80%, or two or more types of parameter triggering anomalies. A severe warning is defined as a deviation > 40% and remaining lifespan < 40%, or two or more types of parameter anomalies with at least one type of deviation > 20%.

[0033] As a preferred embodiment of the present invention, the fiber laser source in step S4 is a narrow linewidth laser with a linewidth of less than 1 kHz, distributed in the C-band, a wavelength scanning range of 100 nm, an output power greater than 10 dBm, and a sweep frequency step of less than 2 MHz; the intensity, frequency, wavelength, polarization state, and phase parameters of the optical signal are extracted using coherent detection; and the temperature and stress information are demodulated by solving a system of simultaneous equations using a Raman scattering, Brillouin scattering, and Rayleigh scattering multiplexing mechanism.

[0034] Beneficial effects

[0035] This invention establishes a mathematical coupling model between the equivalent capacitance and resistance of the submarine cable busbar and the location of the aging section of the submarine cable. It quantifies the influence of converter characteristics on spectral characteristics, enabling online extraction of equivalent capacitance parameters under system operation. By tracking the evolution of the capacitor from initial aging to severe aging, the change process of submarine cable insulation aging is understood. This solution combines online implementation capability with high-precision positioning performance, providing a new technical approach for submarine cable condition monitoring and fault early warning. Specific beneficial effects are as follows: Multi-dimensional collaborative perception: Integrating four types of parameters—wideband impedance, temperature, stress, and magnetic field—covering fault characteristics across all dimensions—electrical, thermal, mechanical, and electromagnetic—and improving the detection rate of latent faults through multi-parameter cross-validation.

[0036] Wideband adaptive detection: The signal frequency covers 0.1-100MHz, adapting to the detection needs of submarine cables of different lengths and types, and improving the fault location accuracy of long-distance submarine cables through wideband modal response.

[0037] Strong anti-interference capability: Adopting an adaptive noise suppression and dynamic weight fusion algorithm, the recognition accuracy is ≥98% in complex marine environments, and the accuracy remains above 97% when SNR≥10dB, significantly reducing the false judgment rate.

[0038] Highly efficient and compatible operation: It does not rely on dedicated pre-embedded optical fibers, is suitable for all types of new and old submarine cables, improves efficiency compared to traditional methods, and reduces on-site deployment and maintenance costs.

[0039] Proactive early warning capability: Establish a baseline feature database for normal operation of submarine cables, classify and issue early warnings for defects such as insulation aging, water treeing, and minor damage, track the development trend of defects, predict the remaining lifespan, guide planned maintenance, and avoid sudden failures. Attached Figure Description

[0040] Figure 1 This is a design diagram of the multi-parameter sensing submarine cable fault signal identification system provided in this embodiment of the invention; Figure 2 This is a schematic diagram of the implementation principle of the multi-fusion injection source component provided in the embodiments of the invention; Figure 3 This is a schematic diagram of the multi-parameter fiber optic sensing multiplexing mechanism. Figure 4 It is a test circuit model based on a distributed multi-coil structure. Detailed Implementation

[0041] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0042] Example 1: Please refer to Figure 1-4 As shown, this invention proposes a multi-parameter sensing method for identifying submarine cable fault signals. By integrating broadband impedance response, temperature, stress, and magnetic field sensing parameters, it constructs a complete technical route of "active excitation - multi-source acquisition - data fusion - accurate identification." The specific implementation scheme is as follows: Figure 1 As shown, it mainly includes four core modules: multi-fusion injection source integrated device, integrated fiber sensing and processing, broadband impedance insulation aging sensing, and multi-parameter fault identification and diagnosis. The technical implementation methods of the four modules are described respectively.

[0043] I. Active excitation by multi-source signals (step S1) 1. Multi-fusion injection source section The fiber laser source and multi-frequency injection signal components are integrated into a single design, achieving both physical structure integration and circuit-level integration: in terms of physical structure, they share a common housing, heat dissipation module, and mounting bracket; in terms of circuit level, they share an FPGA main control unit. This saves installation space and allows for parallel fusion processing of the two injection signal sources, improving source processing efficiency.

[0044] Coherent detection is used to extract parameters (intensity, frequency, wavelength, polarization state, phase, etc.) from the sensing optical signal, enabling monitoring of the submarine cable's operating temperature and strain caused by external forces. A narrow-linewidth laser (less than 1 kHz linewidth, distributed in the C-band) is employed, with a wavelength scanning range of 100 nm, an output power greater than 10 dBm, and a sweep frequency step of less than 2 MHz. The laser emitted by the laser is split into two paths via coupler 1: The light pulse is modulated into an optical pulse by an electro-optic modulator, then the peak power of the pulse is increased by an optical amplifier, and finally it enters the optical fiber under test through a circulator.

[0045] Another path randomizes the polarization state of the light through a polarizer and enters one input of coupler 2, where it is used as the local oscillator light.

[0046] 2. Multi-frequency active excitation section A wideband voltage injection signal of 0.1-100MHz is generated using direct digital frequency synthesis technology, including three modes: continuous periodic sine wave, step voltage signal, and pseudo-random sequence wave. Signal parameters are adjusted via closed-loop feedback and injected into the submarine cable under test through a coupling channel. A series wave trap reduces power supply interference. Wideband impedance setting parameters include, but are not limited to, the acquired impedance value, dielectric loss factor, capacitance, and inductance parameters. The minimum resolution for magnetic field parameter acquisition is set to 1nT.

[0047] The excitation frequency band is adaptively selected based on the submarine cable length and voltage level. The specific mapping relationship is as follows: For L≤50km, select 50-100MHz; for 50km<L≤150km, select 10-50MHz; for L>150km, select 0.1-10MHz. The lower limit of the frequency band is 0.1MHz for voltage levels ≤110kV, 1MHz for 110kV<V≤220kV, and 5MHz for L>220kV. The specific implementation principle is as follows... Figure 2 As shown.

[0048] 3. Source-side integrated control module and PID coordinated control Develop a source-side integrated control module to achieve split injection of optical and electrical signals. The electrical signals are implemented using frequency division multiplexing, injecting signals of different frequency ranges into different transmission channels to ensure sufficient frequency spacing between signals to prevent mutual interference. The optical signals are implemented using wavelength division multiplexing, using optical signals of different wavelengths to transmit in optical fibers. Wavelength selection devices are used to split and inject signals of different wavelengths into the submarine cable, enabling simultaneous transmission of multiple signals.

[0049] Signals are routed according to their priority, with important control signals and urgent data given higher priority to ensure their stability and timeliness during transmission. High-quality transmission resources and channels are allocated preferentially to reduce signal delay and attenuation.

[0050] When fiber laser source signals and multi-frequency injected signals are transmitted in submarine cables, they are affected by factors such as fiber dispersion and loss, causing changes in the phase, amplitude, and frequency of the signals. These changes affect the coordinated operation between the two signals. Therefore, a PID mathematical control model is established to describe the changes and interrelationships of the two injected signals. A continuous time-domain measurement method is used, with two separate control outputs. State equations and output equations are established to describe the system's state changes and output response. The model expression is as follows:

[0051] in and These represent the real-time control quantities of the broadband injection voltage signal and the laser signal, respectively. and The ratio of the voltage signal to the laser signal determines the error response speed of the injected signal; and These are integral coefficients to eliminate steady-state errors; and These are differential coefficients, used to suppress overshoot and improve dynamic stability; and Two signals respectively t The basic principle of PID control algorithm is to adjust the control quantity by proportional, integral and derivative operations based on the error between the given value and the actual value, so that the actual value is as close as possible to the given value. In the integrated control of submarine cable source side, the given value is the ideal value of the signal strength, phase and frequency parameters set in advance, while the actual value is the current parameter value of the two source signals obtained by real-time monitoring by the sensor.

[0052] II. Multi-parameter data acquisition (step S2) 1. Integrated fiber optic modulation and demodulation process This invention designs a monitoring device based on multi-parameter distributed optical fiber sensing. It integrates multiple signals with high integration and parallel demodulation using fewer optical fiber resources, enabling real-time monitoring of temperature, strain, and disturbance. The phase OTDR component shares a laser source to achieve distributed vibration detection, and the BOTDR system module shares a light source to achieve distributed temperature detection. A sensor cross-sensitive parameter demodulation model is designed to achieve high-precision signal processing of multi-parameter signals of temperature, vibration, and disturbance, making full use of the physical characteristics of various nonlinear scattering effects in optical fibers.

[0053] In sensing mechanisms based on Raman scattering, the Raman anti-Stokes light is only sensitive to temperature, and the signal transduction formula is as follows:

[0054] in, The power of the Raman anti-Stokes scattering light. This is the Raman temperature coefficient, which is related to the optical fiber material. For the change in temperature, , For strain and temperature coefficients. The sensing mechanism based on the Brillouin and Raman scattering effects, by multiplexing the measured data and substituting them into the simultaneous equations above, allows for the separation of temperature and stress information at each scattering location along the fiber.

[0055] The temperature and stress sensing mechanism based on the Rayleigh scattering effect is sensitive to the frequency of the probe light pulse. Different probe frequencies correspond to different characteristic probe curves. Therefore, by scanning the probe frequency and using related algorithms, the characteristic probe curves corresponding to temperature and stress can be measured. The relationship between the shift of the characteristic frequency of the Rayleigh probe curve and temperature and stress is shown in the following formula:

[0056] By combining the Brillouin scattering effect with the Rayleigh scattering frequency sweep detection sensing mechanism, the temperature and stress values ​​in the simultaneous equations can be demodulated from the measurement results (the principle is as follows). Figure 3 (As shown).

[0057] 2. Broadband Impedance Insulation Aging Response Process The key to using broadband response technology for monitoring the insulation condition of submarine power cables lies in: (1) The broadband circuit model accurately represents various electrical parameters; (2) The broadband transfer function reflects the sensitivity of changes in insulation electrical parameters.

[0058] Broadband Circuit Model: The broadband impedance mode circuit model of submarine cable is represented in single-phase form. The alternating magnetic field in the cable is divided into M coils. The electrical parameters of the same type are equal in different coils. Taking into account the equivalent branches of the insulation between each layer, a model is formed as follows: Figure 4 The test submarine cable circuit model shown is based on a distributed multi-coil structure. Under wideband voltage excitation injection, this model has three types of electrical parameters: Coil parameters: The resistance and self-inductance parameters of the M coils in each layer of the winding on the same side are consistent, and they have a significant skin effect over a wide frequency range.

[0059] Equivalent branch parameters of each insulation component: The equivalent electrical parameters of the same type of insulation are different between different layers of windings, and they have frequency-varying characteristics.

[0060] Equivalent parameters of electromagnetic coupling between coils: There is mutual inductance between M coils in each layer of winding, and the size of the inductance matrix is ​​5M×5M.

[0061] The impedance mode circuit model of submarine cable is capacitive over a wide frequency range. The variation of the equivalent capacitance parameters of each insulating component has a dominant effect on the amplitude-frequency characteristics of the broadband common-mode impedance response. The impedance response in the resonant frequency band depends on the equivalent capacitance of the insulation and the common-mode mutual inductance between the coils.

[0062] Response analysis method: A numerical method is employed, based on a circuit model to describe the analytical relationship between port characteristic electrical quantities and the state of the monitoring unit. Multi-port state equations are constructed using the voltage of the monitoring test nodes and branch currents of the submarine cable. Based on the fitting of rational function expressions with characteristic quantities, the wave characteristics of the excitation voltage in the winding are described, and its poles characterize the parallel resonant points of the broadband response. Changes in capacitance parameters cause an overall shift in the broadband response test curve; neutral point grounding issues lead to changes in the high-frequency resonant points. Changes in inductive electrical parameters also affect the amplitude-frequency characteristics of the broadband response, requiring further differentiation and quantification.

[0063] 3. Complete Wideband Adaptive Detection Process To further improve the fault location accuracy of long-distance submarine cables, this invention adopts the following broadband adaptive detection steps: Based on the length and voltage level of the submarine cable, the excitation frequency band of 0.1-100MHz is adaptively selected (see step S1 for specific mapping relationship).

[0064] Three types of excitation waveforms—sine, step, and pseudo-random sequence—are generated using direct digital frequency synthesis technology.

[0065] Broadband signals are injected into the submarine cable via a coupling channel, and a series wave trap suppresses power supply interference.

[0066] Simultaneously collect impedance, dielectric loss, capacitance, and inductance response data at different frequencies.

[0067] A wideband impedance-fault location mapping model is established, and a signal attenuation compensation algorithm based on the least squares method is used to compensate for long-distance signal attenuation.

[0068] The location of local defects can be determined by the offset of the resonance point on the response curve.

[0069] It outputs a fault distribution map of the entire line, with a maximum detection distance of up to 300km.

[0070] The signal attenuation compensation algorithm based on the least squares method is as follows: Set the distance of the coastal cable The signal amplitude at that location is Satisfies the exponential decay model ,in The attenuation coefficient is used. By collecting signal amplitudes from multiple known fault points at different distances, an overdetermined system of equations is constructed, and the optimal attenuation coefficient is fitted using the least squares method. This allows for amplitude compensation of the long-distance detection signal, resulting in a compensated signal amplitude of [value missing]. This eliminates the impact of signal attenuation caused by long-distance transmission on the accuracy of fault location.

[0071] III. Multi-dimensional Collaborative Perception Methods To achieve multi-parameter cross-validation and improve the detection rate of latent faults, this invention further adopts the following multi-dimensional collaborative sensing process: Deploy distributed electromagnetic sensing units and fiber optic sensing units to simultaneously collect broadband impedance spectrum, temperature distribution, stress-strain, and magnetic field distortion data.

[0072] The four types of parameters were normalized, and the min-max normalization method was used to eliminate the differences in dimensions and amplitudes.

[0073] Establish parameter association rules to determine whether an anomaly in a single parameter is corroborated by other parameters.

[0074] When at least two types of parameters simultaneously trigger the abnormal threshold, it is determined to be a potential fault.

[0075] Extract multi-parameter feature fingerprints of insulation aging, moisture, mechanical damage, core breakage, short circuit and compound faults.

[0076] The features are input into the improved DBN model, and the fault type and confidence level are output.

[0077] Latent defects are classified and marked to form an early warning list.

[0078] IV. Signal Preprocessing and Feature Extraction (Step S3) 1. Adaptive noise suppression algorithm (improved wavelet threshold denoising) Algorithm principle: Based on the multi-scale decomposition characteristics of wavelet transform, the acquired original multi-parameter signal is decomposed into low-frequency approximate components (including fault features) and high-frequency detail components (including noise). An adaptive threshold function is designed to perform threshold processing on the high-frequency detail components, suppress noise components, retain effective fault feature signals, and then reconstruct the signal through inverse wavelet transform.

[0079] The specific implementation steps are as follows: The original multi-parameter signals (wideband impedance, temperature, stress, magnetic field) are decomposed into three layers of wavelets, and the db4 wavelet is selected as the base wavelet.

[0080] Calculate the noise standard deviation of the high-frequency components of each layer. and adaptive threshold , where N is the number of signal sampling points.

[0081] The following adaptive thresholding function is used to perform thresholding on the high-frequency components of each layer:

[0082] in, For the processed first j Layer k A high-frequency coefficient, To process the firstj Layer k A high-frequency coefficient.

[0083] The processed low-frequency and high-frequency components are subjected to inverse wavelet transform to obtain the denoised clean signal.

[0084] Repeat the above steps to complete the noise suppression processing for all multi-parameter signals.

[0085] 2. Feature Parameter Extraction and Fault Feature Matrix Construction The types of feature parameters extracted for each parameter: Wideband impedance: resonant frequency, resonant amplitude, dielectric loss factor, equivalent capacitance, and equivalent inductance (5 categories).

[0086] Temperature: Average temperature at monitoring points, temperature difference between adjacent monitoring points, temperature change rate, and peak temperature (4 categories).

[0087] Stress: peak stress, average stress, stress variation range, number of stress abrupt changes (4 categories).

[0088] Magnetic field: magnetic field amplitude, magnetic field distortion rate, magnetic field change frequency, and magnetic field peak value (4 categories).

[0089] The fault feature matrix is ​​M×N dimensional, where M rows correspond to the number of monitoring points on the submarine cable, and N = 5 + 4 + 4 + 4 = 17 columns. Matrix elements are normalized values ​​of the corresponding monitoring points and their corresponding feature parameters (using the min-max normalization method). Construction method: The 17 types of feature parameters for each monitoring point are arranged sequentially in the order of "wideband impedance characteristics → temperature characteristics → stress characteristics → magnetic field characteristics," forming the feature row vector corresponding to that monitoring point. The feature row vectors of all monitoring points are combined to form the fault feature matrix.

[0090] 3. Feature screening criteria for different fault types The mutual information method is used to screen key features for each fault type. A mutual information threshold of 0.6 is set, and feature parameters with a mutual information value ≥ 0.6 for the fault type are selected as key features for that fault. The specific correspondence is as follows: Insulation aging: offset of broadband impedance resonant point, dielectric loss factor, and temperature change rate.

[0091] Mechanical damage: peak stress, number of stress abrupt changes, and magnetic field distortion rate.

[0092] Moisture: equivalent capacitance, average temperature, magnetic field amplitude.

[0093] Core breakage: broadband impedance amplitude, magnetic field distortion rate, and number of stress abrupt changes.

[0094] Short circuit: equivalent inductance, frequency of magnetic field change, and peak temperature.

[0095] Composite faults: the intersection and supplementary features of the key features of two types of single faults.

[0096] V. Fault Identification and Location (Step S4) 1. Dynamic weight fusion strategy A two-factor dynamic weight allocation based on fault type and signal-to-noise ratio is adopted. The weight allocation function is as follows:

[0097] in, For the first i Class parameter ( i =1 represents a wideband impedance. i =2 represents temperature. i =3 represents stress. i =4 is the weight of the magnetic field. This parameter is the matching coefficient between itself and the current fault type (determined based on the mutual information value; the higher the mutual information value, the better). cross Large (values ​​range from 0.1 to 0.9). This is the signal-to-noise ratio (SNR) adaptation coefficient for this parameter (the higher the SNR, the better). The larger the value, the range is 0.2-1.0.

[0098] Specific allocation rules: For different fault types, the parameters are determined in advance through experiments. Reference values ​​(such as wideband impedance during insulation aging) ,temperature ,stress ,magnetic field ; Real-time acquisition of the signal-to-noise ratio (SNR) of various parameters, and calculation ( Let be the real-time signal-to-noise ratio of the i-th type of parameter. (The highest signal-to-noise ratio among the four types of parameters). Will and Substitute the above weight allocation function to calculate the real-time dynamic weights of each parameter. The feature parameters of each parameter are weighted and fused according to the weights to obtain a fused feature vector, which is then input into the improved DBN model.

[0099] 2. Improved Deep Belief Network (DBN) Model Design Broadband impedance response, temperature, stress, magnetic field, and other multi-dimensional sensing parameters are converged into a unified data analysis system. To improve the accuracy and type of fault identification, this invention designs an improved DBN algorithm model, the specific design process of which is described below: Configuration hierarchy: A 5-layer RBM hidden layer (H1, H2, H3, H4, H5 in sequence) + 1 Softmax output layer is designed. The input layer dimensions match the fault feature dimensions (such as temperature, stress, alternating induced magnetic field, etc., features of submarine cable monitoring). The number of nodes H1-H5 is designed according to the following sequence: "Incremental number of nodes in input layer → H1 → H2 → H3 → H4 → H5 simplified". For example, 64-dimensional input → H1 210 → H2 29 → H3 28 → H4 27 → H5 26 = 64, enhancing feature abstraction capabilities. The Softmax output layer has 8 nodes, which can represent at least 5 typical and 3 composite fault types. A Gaussian-Bernoulli RBM is used as the input layer H1 (adapted to continuous monitoring data). The intermediate training layer uses a Bernoulli-Bernoulli RBM (to enhance continuous feature extraction) and introduces a Dropout layer (dropout rate 0.2) to reduce overfitting. The RBM energy function is expressed as follows:

[0100] Energy value For visible layer nodes, For hidden layer nodes, As weight, This is the bias for the visible layer.

[0101] Training strategy: Pre-training employs an unsupervised, layer-by-layer training approach. Input layer H1 uses normalized fault features as input and is trained with contrastive divergence (CD-1) for 500 iterations at a learning rate of 0.01. H1-H2 / H2-H3 / H3-H4 / H4-H5 use the output of the previous RBM hidden layer as input and are also trained with CD-1, decreasing the number of iterations (400, 300, etc.) and gradually reducing the learning rate (0.01 → 0.005). The contrastive divergence formula is as follows:

[0102] Supervised backpropagation model: After pre-training, the RBM weights are frozen, and the H5 output is fed into the Softmax layer. Fault labels are used as supervision signals, and the Adam optimizer (learning rate 0.001) is employed. Fine-tune the entire network; use the cross-entropy loss function, iterate 300 times, and use a batch size of 32.

[0103] The SMOTE algorithm is used to augment composite fault samples to alleviate sample imbalance; Gaussian noise (signal-to-noise ratio 0.05) is added to enhance robustness. The composite sample description is as follows:

[0104] For samples of the same type, L2 regularization (weight decay coefficient 1e-4) is added during the pre-training stage, and an early stopping mechanism is set during the fine-tuning stage. If the accuracy of the validation set does not improve for 10 consecutive rounds, the training will stop.

[0105] 3. Hidden Defect Grading and Early Warning The degradation is classified into three levels: mild, moderate, and severe. The core basis for classification is the degree of deviation between the fault characteristic parameters and the normal baseline parameters, combined with the remaining service life for auxiliary judgment. Mild warning: A single key characteristic parameter deviates from the baseline by 10%-20%, the remaining lifetime is >80%, and only one type of parameter triggers the anomaly, with no other parameters providing corroboration; recommended measures include strengthening monitoring.

[0106] Moderate warning: A single key characteristic parameter deviates from the baseline by 20%-40%, the remaining lifetime is 40%-80%, or two or more parameters trigger anomalies, with deviations of 10%-20%; recommended measures: inspection within one month.

[0107] Severe warning: A single key characteristic parameter deviates from the baseline by more than 40%, the remaining lifetime is less than 40%, or two or more types of parameters trigger anomalies, and at least one type of parameter deviates by more than 20%; recommended measures: immediate emergency repair.

[0108] VI. Strong anti-interference and high-precision identification process This invention also provides a method for strong anti-interference and high-precision identification, the specific implementation steps of which are as follows: The original signal is collected and filtered out from 50Hz power frequency, ship electromagnetic, and tidal noise interference.

[0109] Time-frequency domain analysis is performed on the filtered signal to extract fault feature components.

[0110] The wideband impedance, temperature, stress, and magnetic field weights are adaptively allocated based on the fault type (using a dynamic weight fusion strategy).

[0111] Construct a fused feature vector and input it into the improved DBN model for classification.

[0112] Tested under different signal-to-noise ratio environments, it retains stable recognition capability when SNR≥10dB.

[0113] By comparing the results of single parameter identification, misjudgments and omissions can be corrected.

[0114] The final recognition result is output, with an overall accuracy of ≥98%.

[0115] VII. High-efficiency and compatible operation method This invention does not rely on dedicated pre-embedded optical fibers and is applicable to all types of submarine cables, both new and old. The specific implementation steps are as follows: It uses general-purpose electromagnetic and fiber optic sensing units, eliminating the need to modify the submarine cable structure.

[0116] It is compatible with both old submarine cables without pre-embedded optical fibers and new composite submarine cables.

[0117] One-click start for automatic detection of the entire line, eliminating the need for segmented manual operation.

[0118] The entire process of signal injection, acquisition, and analysis for a 10km submarine cable can be completed in ≤2 hours.

[0119] Automatically generate test reports and fault location records.

[0120] Compared to traditional OTDR and manual dragging methods, efficiency is improved by more than 80%.

[0121] It supports remote online monitoring, reducing the frequency of offshore inspections.

[0122] VIII. Proactive Early Warning and Maintenance Methods This invention achieves a shift from reactive emergency repairs to proactive early warning systems. The specific implementation steps are as follows: Establish a baseline feature database for normal operation of submarine cables and continuously collect multi-parameter data.

[0123] Real-time monitoring of parameter change trends to identify early signs of degradation.

[0124] Graded early warning for defects such as insulation aging, water treeing, and minor damage.

[0125] Push early warning information and suggested handling measures (strengthen monitoring for mild warnings, conduct inspections within one month for moderate warnings, and carry out emergency repairs immediately for severe warnings).

[0126] Track defect development trends and predict remaining lifespan.

[0127] Provide guidance for planned maintenance to avoid unexpected malfunctions.

[0128] Statistical analysis of failure rates and maintenance costs verifies the effectiveness of early warning systems.

[0129] The remaining lifetime prediction uses a reliability model based on the Weibull distribution, and the specific steps are as follows: Establish degeneration characteristic indicators: Reduce the dimensionality of the multi-parameter feature matrix through principal component analysis, and extract the first principal component as the health index. Normalized to the interval [0,1], where 1 represents a brand new state and 0 represents a critical fault state.

[0130] Degradation trajectory fitting: using an exponential degradation model ,in For runtime, For scale parameters, The shape parameter is determined using maximum likelihood estimation based on historical monitoring data (HI values ​​at least 6 time points). and .

[0131] Remaining life calculation: Let the failure threshold be... (Empirical value), then the remaining lifespan .

[0132] Uncertainty quantification: The Markov chain Monte Carlo method is used to give the 90% confidence interval of the remaining lifetime.

[0133] Update mechanism: Every 10 sets of new data are collected, the model parameters are re-estimated to achieve dynamic updates of remaining life prediction.

[0134] Example 2:

[0135] To verify the effectiveness of the method of the present invention, a submarine cable fault simulation and detection experimental platform was built, with the following specific configuration: The submarine cable under test is a 110kV cross-linked polyethylene (XLPE) submarine power cable with a length of 10km. It contains 3 main cores and 1 composite optical fiber, without any pre-embedded special sensing optical fiber, to simulate the scenario of old submarine cables.

[0136] Fault simulation module: Sets 5 typical faults (water tree aging simulation), mechanical damage (squeezing / scratching simulation), moisture ingress (end seal failure simulation), core breakage (single core breakage simulation), and short circuit (two cores conducting simulation) and 2 composite faults (mechanical damage + short circuit, moisture + core breakage).

[0137] Testing equipment: multi-frequency excitation unit (output frequency 0.1-100MHz, amplitude 0-10kV), distributed composite fiber optic measurement component (resolution 0.1℃, sampling interval 1m), high-precision magnetic field sensor (resolution 1nT, measurement range 0-10mT), data acquisition card (sampling rate 1MHz, bit depth 32bit).

[0138] Experimental steps: Baseline data acquisition: Collect multi-parameter data under fault-free conditions to establish a baseline feature library for the normal operation of submarine cables.

[0139] Fault simulation and data acquisition: Five typical faults and two composite faults are triggered sequentially. After each fault state has been running stably for 10 minutes, multi-parameter data are collected. Ten sets of samples are collected for each fault.

[0140] Comparative experimental design: The same fault was detected simultaneously using the traditional OTDR method, the single broadband impedance method, and the manual drag detection method. The fault identification rate, location error, and detection time of each method were recorded.

[0141] Data processing and analysis: The method of this invention is used to preprocess, feature fusion and intelligent recognition of the collected multi-parameter data, and to statistically analyze key indicators.

[0142] Experimental results: 1) Fault identification accuracy Experimental results show that the accuracy of the method of the present invention in identifying five typical and compound faults is shown in the table below:

[0143] As shown in the table, the average recognition accuracy of the method of the present invention reaches 98.9%, which is significantly higher than that of the traditional OTDR method (63.5%) and the single broadband impedance method (91.1%). In particular, it has obvious advantages in recognizing latent faults such as insulation aging and moisture, as well as complex faults.

[0144] 2) Detection efficiency test A comparative test was conducted on the efficiency of full-line inspection of a 10km submarine cable. The method of this invention requires only 2 hours for a single full-line inspection, including signal injection, data acquisition, analysis, and identification. The traditional OTDR method takes 3 hours and cannot identify non-fiber faults, while the manual towing method requires segmented inspection, with a total time of over 12 hours. The inspection efficiency of the method of this invention is more than 80% higher than that of traditional methods, significantly reducing operation and maintenance costs.

[0145] 3) Anti-interference performance test In an experimental environment simulating marine environmental interference (including 50Hz power frequency interference, ship electromagnetic interference, and tidal noise simulation), the fault identification accuracy under different signal-to-noise ratios (SNR) was tested. The results showed that when SNR ≥ 10dB, the identification accuracy of the method of the present invention remained above 97%, while the accuracy of the single broadband impedance method dropped to below 85% when SNR = 15dB, and the traditional OTDR method still had a false judgment rate of more than 10% when SNR = 20dB. This indicates that the method of the present invention has stronger anti-interference ability and is adaptable to complex marine environments.

[0146] 4) Overall performance comparison The overall performance comparison of the method of this invention with the traditional OTDR method and the single broadband impedance method on a 10km submarine cable is as follows:

Claims

1. A method for identifying submarine cable fault signals using multi-parameter sensing, characterized in that, Includes the following steps: Step S1: Active excitation of multi-source signals: Generate a wideband voltage signal and fiber laser signal of 0.1-100MHz to form an integrated source component. Use frequency division multiplexing and wavelength division multiplexing technology to inject the signal into the submarine cable, and realize the coordinated control of optical and electrical signals through PID model. Step S2: Multi-parameter data acquisition: Simultaneously acquire broadband impedance, temperature, stress, and magnetic field response data of the submarine cable. Step S3: Signal preprocessing and feature extraction: The original signal is processed using an adaptive noise suppression algorithm to extract feature parameters and construct a fault feature matrix; Step S4: Fault identification and localization: Based on the dynamic weight fusion strategy, integrate features and input them into the improved deep belief network model to complete fault classification and localization.

2. The method according to claim 1, characterized in that: In step S1, the broadband voltage signal includes three modes: continuous periodic sinusoidal signal, step voltage signal, and pseudo-random sequence wave; the minimum resolution for magnetic field acquisition is 1 nT, and the resolution for fiber optic temperature acquisition is 0.1℃. The expression for the PID model is: in u ( t )and f ( t) These are the real-time control quantities for wideband voltage signals and laser signals, respectively.

3. The method for identifying submarine cable fault signals using multi-parameter sensing according to claim 1, characterized in that... In step S1, the excitation frequency band is adaptively selected based on the cable length and voltage level. When the submarine cable length L≤50km, select 50-100MHz; When 50km < L ≤ 150km, select 10-50MHz; When L > 150km, select 0.1-10MHz; When the rated voltage is ≤110kV, the lower limit of the excitation frequency band is 0.1MHz. When the voltage is between 110kV and 220kV, the lower limit is 1MHz. The lower limit is 5MHz when the voltage is greater than 220kV.

4. The method for identifying submarine cable fault signals using multi-parameter sensing according to claim 1, characterized in that... The adaptive noise suppression algorithm in step S3 employs an improved wavelet threshold denoising algorithm, specifically including: The original signal was decomposed into three levels of wavelet, and the db4 wavelet was selected as the base wavelet. Calculate the noise standard deviation of the high-frequency components of each layer. and adaptive threshold , in, For the first j Adaptive threshold for high-frequency components in the layer. For the first j The noise standard deviation of the high-frequency components of the layer. This represents the number of signal sampling points. The following threshold function is used for processing: in, For the processed first j Layer k A high-frequency coefficient, To process the first j Layer k A high-frequency coefficient; Finally, the signal is reconstructed using inverse wavelet transform.

5. The method for identifying submarine cable fault signals using multi-parameter sensing according to claim 1, characterized in that: The feature parameters extracted in step S3 include: Five categories of wideband impedance: resonant frequency, resonant amplitude, dielectric loss factor, equivalent capacitance, and equivalent inductance; The temperature monitoring points are categorized into four types: average temperature at each monitoring point, temperature difference between adjacent monitoring points, temperature change rate, and peak temperature. The stress is classified into four categories: peak stress, mean stress, stress variation range, and number of stress abrupt changes. The magnetic field is classified into four categories: magnetic field amplitude, magnetic field distortion rate, magnetic field change frequency, and magnetic field peak value. The fault feature matrix is ​​an M×17 dimensional matrix, where M is the number of monitoring points, 17 is the total number of feature parameters, and the matrix elements are normalized values.

6. The method for identifying submarine cable fault signals using multi-parameter sensing according to claim 1, characterized in that: Place The dynamic weight fusion strategy in step S4 adopts a two-factor dynamic weight allocation based on fault type and signal-to-noise ratio. The weight allocation function is as follows: in, For the first i Weights of class parameters This parameter is the matching coefficient between itself and the current fault type. For this parameter Signal-to-noise ratio adaptation coefficient.

7. The method for identifying submarine cable fault signals using multi-parameter sensing according to claim 1, characterized in that: The improved deep belief network model in step S4 includes 5 RBM hidden layers and 1 Softmax output layer. It adopts a combination structure of Gaussian-Bernoulli RBM and Bernoulli-Bernoulli RBM, and introduces Dropout layer and L2 regularization. The pre-training uses contrastive divergence CD-1, and the fine-tuning uses Adam optimizer and cross-entropy loss function.

8. The method for identifying submarine cable fault signals using multi-parameter sensing according to claim 1, characterized in that: Step S4 also includes a graded early warning system for latent defects, which is divided into three levels according to the degree of deterioration: mild, moderate, and severe. A mild warning is defined as a single key characteristic parameter deviating from the baseline by 10%-20% and remaining lifetime > 80%. A moderate warning is defined as a deviation of 20%-40% and a remaining lifespan of 40%-80%, or two or more types of parameter triggering anomalies. A severe warning is defined as a deviation > 40% and remaining lifespan < 40%, or two or more types of parameter anomalies with at least one type of deviation > 20%.

9. The method for identifying submarine cable fault signals using multi-parameter sensing according to claim 1, characterized in that: The fiber laser source in step S4 is a narrow linewidth laser with a linewidth of less than 1 kHz, distributed in the C-band, a wavelength scanning range of 100 nm, an output power greater than 10 dBm, and a frequency sweep step of less than 2 MHz. The intensity, frequency, wavelength, polarization state, and phase parameters of the optical signal are extracted using coherent detection. A Raman scattering, Brillouin scattering, and Rayleigh scattering multiplexing mechanism is used to demodulate temperature and stress information by solving a system of simultaneous equations.

Images