A data early warning method and system for marine cable fault

Abstract

The present application relates to the technical field of cable fault diagnosis, in particular to a data early warning method and system for marine cable fault, comprising obtaining load voltage, current, real-time torque, winding layer number and acoustic-electric signal of the drum cable in operation; using load current phase to analyze partial discharge distribution, combining mechanical wave cycle to synchronously align acoustic emission signal, constructing dynamic acoustic-electric-force composite feature matrix; then using Euler-Savari formula combined with layer number to solve radial extrusion force, constructing spatial topology graph and mapping electric parameters; obtaining theoretical insulation loss prediction value through coding layer and regression prediction layer; finally extracting loss deviation and discharge frequency deviation as double residual error, inputting Bayesian decision model to output fault warning level. The present application realizes accurate early warning under dynamic working condition through multi-physical field feature fusion.

Description

Technical Field

[0001] This invention relates to the field of cable fault diagnosis technology, specifically to a data-based early warning method and system for marine cable faults. Background Technology

[0002] Marine reel cables are a crucial medium for power transmission and signal exchange in mobile electrical equipment. According to relevant electrical testing standards, online monitoring of the cable's insulation loss angle and partial discharge characteristics is a core method for assessing its operational safety.

[0003] In dynamic operating environments, the cable reel is wound in multiple layers along the reel. Existing monitoring solutions typically use sensors to acquire high-frequency induced current or acoustic emission signals, and use the load current as a time reference to perform phase analysis and feature extraction of the signals.

[0004] However, due to the influence of the coil winding distribution characteristics, the electromagnetic coupling environment and the heat dissipation boundary of the medium differ significantly at different winding depths. This non-uniform distribution characteristic causes the microscopic distribution of the electric field inside the cable to drift, which in turn interferes with the sensor's capture of insulation degradation signals. When processing electrical variable measurements under such dynamic conditions, existing technologies cannot compensate for signal gain fluctuations and phase shifts caused by changes in winding position in real time. As a result, under multi-layer superimposed interference, the measured values ​​deviate from the true physical state, making it impossible to achieve accurate quantification of insulation loss and discharge frequency.

[0005] Therefore, a data-based early warning method and system for marine cable faults is proposed. Summary of the Invention

[0006] The purpose of this invention is to provide a data-driven early warning method and system for marine cable faults. Through multi-physics feature fusion, it achieves accurate early warning under dynamic operating conditions. This includes acquiring load voltage, current, real-time torque, number of winding layers, and acoustic-electric signals during the operation of the reel cable; using the load current phase to analyze the partial discharge distribution and combining it with the mechanical fluctuation period to synchronously align the acoustic emission signals, constructing a dynamic acoustic-electric-force composite feature matrix; subsequently, using the Euler-Savari formula combined with the number of layers to calculate the radial compressive force, constructing a spatial topology map and mapping electrical parameters; obtaining the theoretical insulation loss prediction value through an encoding layer and a regression prediction layer; and finally extracting the loss deviation and discharge frequency offset as dual residuals, inputting them into a Bayesian decision model to output the fault warning level.

[0007] To achieve the above objectives, the present invention provides the following technical solution: A data-driven early warning method for marine cable faults includes: The system acquires the load voltage, load current, real-time torque of the drum drive motor, current number of cable winding layers, partial discharge pulse sequence, and ultrasonic emission signal during the operation of the cable reel. Using the phase of the load current as a time reference, the phase distribution of the partial discharge pulse sequence is analyzed. Combined with the mechanical fluctuation period of the real-time torque, the ultrasonic acoustic emission signal is synchronously correlated and aligned to construct a dynamic acoustic-electrical-mechanical composite feature matrix. The radial compressive force distribution of the cable on the drum is calculated based on the number of winding layers. A spatial topology map is constructed by combining the drum's geometric parameters. The load voltage and load current are mapped onto the spatial topology map. The composite feature matrix and the spatial topology map are input into the encoding layer to obtain the spatiotemporal feature vector. The spatiotemporal feature vector, along with the load voltage, load current, and number of winding layers, are input into the regression prediction layer to obtain the theoretical insulation loss prediction value. The first residual is calculated between the measured insulation loss value and the theoretical insulation loss prediction value. The discharge frequency offset of the partial discharge pulse sequence within the load current cycle is extracted as the second residual. The first residual and the second residual are input into the Bayesian decision model, and the composite damage probability index is calculated by combining the abnormal offset of the real-time torque. The fault warning level is then output.

[0008] Preferably, the process of acquiring various data during the operation of the reel cable includes: acquiring the load voltage and load current through an electrical parameter acquisition unit deployed at the power output end of the reel, and capturing a partial discharge pulse sequence in the rising edge range of the load current using a high-frequency current transformer; reading the real-time torque in real time through the inverter communication interface of the reel drive motor, and calculating the current number of winding layers of the cable based on the linear mapping relationship between the total number of reel rotations and the cable arrangement step distance using an absolute encoder installed at the axial end of the reel; and acquiring the ultrasonic acoustic emission signal generated by the cable during the winding and unwinding compression process through a piezoelectric ceramic sensor attached to the inside of the reel flange.

[0009] Preferably, the process of constructing the dynamic acoustic-electric-force composite feature matrix includes: dividing the single cycle of the load current into several phase intervals, counting the number of discharges and the discharge amplitude of the partial discharge pulse sequence in each phase interval, and constructing a two-dimensional phase-discharge distribution feature vector; extracting the spectral features of the real-time torque to determine its mechanical rotation cycle data, and using a cross-correlation algorithm to match the envelope abrupt change point of the ultrasonic acoustic emission signal with the phase zero point of the load current within the mechanical rotation cycle data, thereby achieving synchronization of the acoustic signal and the electrical signal on the time axis; and performing tensor splicing on the phase-discharge distribution feature vector, the synchronized ultrasonic acoustic emission signal, and the amplitude fluctuation sequence of the real-time torque to generate an acoustic-electric-force composite feature matrix with multi-dimensional attributes in the phase domain, time domain, and energy domain.

[0010] Preferably, the process of constructing the spatial topology map includes: using the Euler-Savari formula in combination with the number of winding layers to calculate the cumulative radial tension coefficient of the cable at different stacking depths, and deriving the corresponding radial compressive force distribution gradient based on the elastic modulus of the cable cross section; constructing a three-dimensional mesh spatial topology map characterizing the physical constraint relationship between cable layers, with the axial arrangement position of the drum and the radial layer coordinates as vertices and the mechanical contact paths between adjacent layers and between adjacent turns in the same layer as edges; injecting the load voltage and the load current as node attribute vectors into the spatial topology map, and correcting the edge weights in the spatial topology map according to the radial compressive force distribution gradient in combination with the pressure-weight mapping table to quantify the electromagnetic coupling strength and thermal resistance characteristics between cable layers under different extrusion states.

[0011] Preferably, the process of obtaining the theoretical insulation loss prediction value includes: using the composite feature matrix as node input, extracting and aggregating data of adjacent cable nodes according to the node connection relationship determined by the spatial topology map to obtain an intermediate tensor that integrates spatial location attributes; performing sliding sampling operation on the intermediate tensor on the time axis using a convolution kernel to extract the changing trend of the cable in continuous winding and unwinding cycles, and outputting the spatiotemporal feature vector of fixed dimension; merging the spatiotemporal feature vector with the real-time collected load voltage, load current, and winding layer number to construct a combined input vector containing historical state features and current electrical parameter excitation; inputting the combined input vector into a fully connected layer for multi-layer linear transformation and nonlinear activation to compress and map the high-dimensional vector into a single-dimensional value, and outputting the theoretical insulation loss prediction value corresponding to the current moment.

[0012] Preferably, the process of obtaining the first residual and the second residual includes: acquiring the voltage and current phase difference at the end of the cable reel in real time, and calculating the measured insulation loss value reflecting the overall insulation state of the cable; performing point-by-point subtraction between the measured insulation loss value and the theoretical insulation loss prediction value with the same time step, extracting the abnormal loss component that exceeds the model prediction range, and obtaining the first residual; using the voltage zero point of the load current as the starting reference, mapping the partial discharge pulse sequence to a phase period interval of 0 to 360 degrees, and counting the number of pulses emitted in each interval; retrieving the standard discharge frequency distribution curve corresponding to the current number of winding layers and load power from the pre-stored cable status fingerprint database; subtracting the actual number of pulses emitted from the standard discharge frequency distribution curve, calculating the pulse number increment in each phase interval, and defining the total increment in the entire cycle as the second residual.

[0013] Preferably, the process of outputting the fault warning level includes: extracting the historical failure rate of the cable corresponding to the current winding layer as the initial prior distribution of the Bayesian decision model; calculating the deviation value of the real-time torque relative to the current load power, and adjusting the confidence of the initial prior distribution according to the deviation value to obtain the corrected prior probability; inputting the first residual and the second residual as the observation evidence set into the Bayesian decision model, and calculating the conditional probability likelihood values ​​of the first residual and the second residual under different damage states respectively; recursively updating the corrected prior probability and the conditional probability likelihood values ​​using the Bayesian formula to calculate the composite damage probability index characterizing the real-time deterioration degree of the cable insulation layer; performing interval matching between the composite damage probability index and multiple preset damage risk thresholds, and determining and outputting the corresponding fault warning level according to the matching result.

[0014] A data early warning system for marine cable faults includes: The feature sensing module acquires the load voltage, load current, real-time torque of the drum drive motor, current number of winding layers of the cable, partial discharge pulse sequence and ultrasonic emission signal during the operation of the drum cable. The correlation and fusion module uses the phase of the load current as a time reference to analyze the phase distribution of the partial discharge pulse sequence. Combined with the mechanical fluctuation period of the real-time torque, it performs synchronous correlation and alignment of the ultrasonic acoustic emission signal to construct a dynamic acoustic-electric-force composite feature matrix. The loss prediction module calculates the radial compressive force distribution of the cable on the drum based on the number of winding layers, constructs a spatial topology map based on the drum's geometric parameters, and maps the load voltage and load current to the spatial topology map. The composite feature matrix and the spatial topology map are input into the encoding layer to obtain the spatiotemporal feature vector. The spatiotemporal feature vector, along with the load voltage, load current, and number of winding layers, are input into the regression prediction layer to obtain the theoretical insulation loss prediction value. The integrated early warning module calculates the first residual between the measured insulation loss value and the theoretical predicted insulation loss value, and extracts the discharge frequency offset of the partial discharge pulse sequence within the load current cycle as the second residual. The first residual and the second residual are input into the Bayesian decision model, and the composite damage probability index is calculated by combining the abnormal offset of the real-time torque, and the fault early warning level is output.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. By zoning the load current phase and using a cross-correlation algorithm to match the envelope abrupt change point of the ultrasonic acoustic emission signal with the phase zero point of the electrical signal, precise synchronous correlation of three heterogeneous data types—acoustic, electrical, and mechanical—on the time axis was achieved. By tensor splicing of phase distribution characteristics, acoustic emission sequences, and torque amplitude fluctuations, a composite feature matrix with multidimensional attributes in the phase, time, and energy domains was constructed. This processing flow can identify mechanical compression events and discharge pulses highly correlated with the electrical phase from complex background noise. Physical causal logic between signals was established during the feature extraction stage, providing structured multidimensional feature support for subsequent identification of transient anomalies in the insulating medium during continuous loading and unloading cycles.

[0016] 2. By utilizing the Euler-Savari formula combined with the number of winding layers to calculate the cumulative radial tension of the cable at different stacking depths, and constructing a three-dimensional mesh space topology diagram characterizing the physical constraints between layers, a quantitative description of the non-uniform distribution characteristics of the reel cable is achieved. By introducing a pressure-weight mapping table to dynamically correct the edge weights in the topology diagram, the micromechanical environment of the cable is directly linked to the electromagnetic coupling strength and thermal resistance characteristics, thus recreating the true physical boundary conditions of the cable in a multi-layered winding state at the model level. This provides a clear physical basis for shielding signal fluctuations caused by changes in winding depth, enabling the monitoring system to identify and compensate for non-faulty signal drift caused by interlayer compression, ensuring the physical fit between the evaluation benchmark and actual operating conditions.

[0017] 3. By extracting the historical failure rate corresponding to the number of winding layers as the initial prior distribution and adjusting the confidence level of the prior probability using the deviation of real-time torque relative to load power, an evaluation architecture with dynamic environmental perception capability was established. By introducing the first and second residuals as observation evidence sets into the Bayesian decision model, a progressive correlation between macroscopic loss anomalies and microscopic discharge offsets at the probabilistic level was realized. This logical architecture can incorporate the instantaneous mechanical stress offset borne by the cable into the damage evaluation system. Through recursive update calculation, a composite damage probability index characterizing the degree of insulation layer deterioration is obtained, transforming isolated sensor data into risk quantification indicators with logical correlation, thereby providing maintenance personnel with early warning levels that are more consistent with the actual fatigue evolution law of cables. Attached Figure Description

[0018] Figure 1 This is a schematic flowchart of a data early warning method for marine cable faults according to the present invention; Figure 2 This is a schematic diagram of the structure of a data early warning system for marine cable faults according to the present invention; Figure 3 This is a schematic diagram of the process for constructing a spatial topology graph according to the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Please see Figures 1 to 3 This invention provides a data-based early warning method for marine cable faults, the technical solution of which is as follows: A data-driven early warning method for marine cable faults includes: The system acquires the load voltage, load current, real-time torque of the drum drive motor, current number of cable winding layers, partial discharge pulse sequence, and ultrasonic emission signal during the operation of the cable reel. Using the phase of the load current as a time reference, the phase distribution of the partial discharge pulse sequence is analyzed. Combined with the mechanical fluctuation period of the real-time torque, the ultrasonic acoustic emission signal is synchronously correlated and aligned to construct a dynamic acoustic-electrical-mechanical composite feature matrix. The radial compressive force distribution of the cable on the drum is calculated based on the number of winding layers. A spatial topology map is constructed by combining the drum's geometric parameters. The load voltage and load current are mapped onto the spatial topology map. The composite feature matrix and the spatial topology map are input into the encoding layer to obtain the spatiotemporal feature vector. The spatiotemporal feature vector, along with the load voltage, load current, and number of winding layers, are input into the regression prediction layer to obtain the theoretical insulation loss prediction value. The first residual is calculated between the measured insulation loss value and the theoretical insulation loss prediction value. The discharge frequency offset of the partial discharge pulse sequence within the load current cycle is extracted as the second residual. The first residual and the second residual are input into the Bayesian decision model, and the composite damage probability index is calculated by combining the abnormal offset of the real-time torque. The fault warning level is then output.

[0021] Furthermore, the process of acquiring various data during the operation of the drum cable includes: acquiring the load voltage and load current through an electrical parameter acquisition unit deployed at the power output end of the drum, and capturing a partial discharge pulse sequence in the rising edge range of the load current using a high-frequency current transformer; reading the real-time torque in real time through the inverter communication interface of the drum drive motor, and calculating the current winding layer number of the cable based on the linear mapping relationship between the total number of drum rotations and the cable arrangement step distance using an absolute encoder installed at the axial end of the drum; and maintaining the active status of the nodes in real time based on the displacement information fed back by the encoder during the dynamic process of continuous winding and unwinding of the drum. When the cable is wound into the drum, the system allocates a new unique node index for the newly added cable length in the memory space and marks it as active to participate in the calculation; when the cable is unwound from the drum, the system automatically identifies the node at the corresponding coordinates and marks it as inactive, no longer participating in feature aggregation calculation. This position-triggered dynamic allocation mechanism ensures that the topology map can be reconstructed in real time following the physical stacking state of the cable; the ultrasonic acoustic emission signal generated by the cable during the winding and unwinding compression process is collected by a piezoelectric ceramic sensor attached to the inside of the drum flange.

[0022] During the maintenance of node activity status, the historical features of nodes marked as inactive are retained in a circular buffer until the corresponding temporal sliding window completely slides out, ensuring data integrity during temporal convolution. Simultaneously, during real-time updates of the topology graph, spatial aggregation operations are performed only on the currently active node set to control computational complexity. This hierarchical management mechanism ensures both the real-time nature of the topology structure and the continuity of temporal feature extraction.

[0023] Inside the distribution cabinet at the power output end of the drum, high-precision voltage transformers and current transformers are deployed as electrical parameter acquisition units. The high-voltage electrical signal is proportionally reduced and then input as data through an analog-to-digital converter circuit. For the capture of partial discharge pulses, a high-frequency current transformer is installed on the grounding wire of the cable shield. The data monitors the phase characteristics of the load current in real time. When the load current is detected to enter the rising edge range, the high-frequency current transformer is triggered to start the supersonic sampling mode. The sampling frequency is set to the megahertz level or above to capture weak discharge pulse sequences with amplitudes in the picocoulomb range. This sequence is then timestamped with the current phase moment. When the voltage / current transformer captures the "voltage zero crossing point" or "current rising edge trigger point" of the load current, the hardware generates a microsecond-level hardware interrupt and immediately resets an internal high-speed counter. The timestamps of all subsequent captured picocoulomb-level pulses are essentially offset counts relative to this zero point.

[0024] Through the inverter's communication interface (such as Modbus TCP or CANopen protocol), data is periodically sent to the drive motor controller to query data. The drive motor returns the current real-time torque value based on the feedback from its internal current loop and speed loop. This value reflects the dynamic traction force borne by the cable during winding or unwinding. Simultaneously, the real-time frequency of the inverter is obtained through the communication link to help verify the running direction and speed of the drum, thereby ensuring the logical consistency between the torque data and the physical motion state. The high-speed acquired electrical and acoustic signals are first stored in a deep buffer with a fixed duration (such as 50 milliseconds) to wait for the torque data with lag transmitted back via Modbus or CANopen. When the torque data arrives, the corresponding acoustic and electrical signals are retrieved from the buffer and aligned and encapsulated according to the timestamp provided by the communication protocol.

[0025] An absolute encoder is installed at the rotation center of the axial end of the drum. The encoder shaft is rigidly connected to the drum spindle. When the drum rotates, the absolute encoder outputs a unique Gray code or digital signal representing the current rotation angle and total number of turns. The data is pre-set with the cable layout pitch and the drum's geometric diameter parameters. By accumulating the total number of rotations and combining the reciprocating motion law of the layout pitch in the axial direction, the number of rotations is converted into the cable's stacking depth in the drum's radial direction. Based on the preset maximum capacity of each layer, the current accumulated number of turns is rounded to accurately calculate the current winding layer of the cable, thereby locating the cable's layer position in the spatial topology. The system pre-sets the cable outer diameter, drum inner diameter, and interlayer coefficient considering extrusion deformation. The layer number is determined by dividing the "current accumulated rotation angle" by the "single turn angle" and then combining it with the "maximum number of turns per layer" for logical judgment. When the encoder value reaches the single-layer boundary value, the system automatically increments the layer number variable by one and updates the current radial radius parameter.

[0026] Multiple piezoelectric ceramic sensors are attached in a circular array to the inner wall of the flanges on both sides of the drum. These sensors are in close contact with the flanges via high-strength acoustic coupling adhesive to capture the weak acoustic emission signals generated when the cable layers are squeezed or slipped. The mechanical vibrations sensed by the sensors are converted into millivolt-level electrical signals, which are then processed by a preamplifier circuit and a bandpass filter circuit to filter out background mechanical noise, retaining the ultrasonic components with frequencies between 20kHz and 100kHz. When the drum is rotating without load, a segment of background mechanical noise is first collected, and its average amplitude is calculated as a "silent reference." Only when the signal amplitude collected by the piezoelectric ceramic sensors exceeds this reference value will the noise be silenced. Only when the absolute encoder position changes at a specific multiple and a trigger command is received simultaneously, will the ultrasonic sampling window from 20kHz to 100kHz be truly opened. During the rotation of the drum, the acoustic acquisition window is opened according to the trigger signal of the absolute encoder to acquire the ultrasonic acoustic emission signal synchronized with the mechanical extrusion action, thus completing the complete encapsulation of the acoustic-electrical-force multidimensional raw data. Before constructing the composite feature matrix, the raw pulse amplitude, real-time torque value, and acoustic emission envelope value are normalized to map all values ​​to a dimensionless range between 0 and 1, ensuring that the weights of each dimension are comparable in the subsequent neural network model calculation.

[0027] By deploying high-frequency current transformers, communication links, and absolute encoders at the shielding layer grounding wire, the inverter communication interface, and the end of the reel, the synchronous quantization of pico-coulomb pulses, dynamic torque, and the number of cable stacking layers was achieved. Combined with the capture of extrusion sound signals by piezoelectric ceramic sensors, a three-in-one monitoring closed loop of sound, electricity, and force was constructed, ensuring a tight correlation between the original information on the time axis and spatial location, and providing a data benchmark with high physical consistency for subsequent insulation degradation assessment.

[0028] Furthermore, the process of constructing the dynamic acoustic-electric-force composite feature matrix includes: dividing the single cycle of the load current into several phase intervals, statistically analyzing the discharge count and amplitude of the partial discharge pulse sequence in each phase interval, and constructing a two-dimensional phase-discharge distribution feature vector; extracting the spectral features of the real-time torque to determine its mechanical rotation cycle data, and within the mechanical rotation cycle data, using a cross-correlation algorithm to match the envelope abrupt change point of the ultrasonic acoustic emission signal with the phase zero point of the load current, thereby achieving synchronization of the acoustic signal and the electrical signal on the time axis; and tensor splicing the phase-discharge distribution feature vector, the synchronized ultrasonic acoustic emission signal, and the amplitude fluctuation sequence of the real-time torque to generate an acoustic-electric-force composite feature matrix with multi-dimensional attributes in the phase domain, time domain, and energy domain.

[0029] After receiving the periodic signal of the load current, the zero-crossing detection algorithm is used to identify the positive zero-crossing point of the voltage / current as the phase start reference. A single electrical cycle (e.g., 20ms) is divided into a preset number (e.g., 128 or 256) of phase intervals. For each interval, the corresponding partial discharge pulse timestamp is retrieved from the cache. The total number of pulse occurrences within the interval is counted, and the maximum and average values ​​of the pulse amplitude are recorded. Specifically, the number of discharges and the amplitude are scaled to the [0, 1] interval by a preset normalization function (maximum-minimum normalization in this embodiment) to eliminate amplitude differences caused by different range sensors and ensure the uniformity of the feature vector in terms of dimensions. Through this mapping method, the discrete pulse sequence is transformed into a fixed-dimensional two-dimensional numerical matrix, namely the phase-discharge distribution feature vector, thereby fixing the randomly occurring discharge events in the coordinate system of electromagnetic physical phase.

[0030] For the real-time torque sequence obtained from the inverter communication interface, a Fast Fourier Transform (FFT) or Discrete Wavelet Transform (DFT) is performed to separate the characteristic components related to the reel rotation frequency from the time-domain signal. By identifying the main energy peak in the spectrum, the current mechanical rotation cycle of the cable reel is locked. This cycle serves as the basic time window for subsequent acoustic signal processing, ensuring that the data interception length is completely synchronized with the physical process of one rotation of the reel, eliminating sampling deviations caused by reel speed fluctuations. To ensure the accuracy of spectrum identification, the sampling window duration for reading the real-time torque signal should be set to cover 3 to 5 complete reel rotation cycles, for example, continuously sampling for 10 to 20 seconds at the rated speed, so that the frequency resolution after the FFT transformation is sufficient to distinguish minute speed fluctuations at the 0.1 Hz level.

[0031] Within the defined mechanical rotation cycle, the acquired ultrasonic acoustic emission signal is first subjected to Hilbert transform or moving average calculation to extract the energy envelope reflecting the signal intensity change. Threshold detection identifies abrupt changes on the envelope (i.e., the starting point of acoustic events caused by cable compression or friction). Subsequently, using a cross-correlation algorithm, the acoustic envelope abrupt changes are used as matching features and subjected to moving correlation calculation with the phase zero-point sequence of the load current. By continuously adjusting the delay time of the acoustic signal relative to the electrical signal, the maximum value of the cross-correlation coefficient is searched, thereby locking the precise position of the acoustic signal on the electromagnetic phase axis and achieving synchronous alignment of the two heterogeneous data types: acoustic and electrical. The physical correlation is based on the fact that the electromagnetic torque pulsation of the drive motor induces weak mechanical vibrations of the drum at the same frequency, which modulates the ultrasonic envelope generated by cable compression. By searching for the maximum value of the cross-correlation coefficient, the signal transmission delay between the acoustic sensor and the electrical sampling point can be accurately compensated.

[0032] After timing alignment is completed, the feature synthesis stage begins. The two-dimensional phase-discharge distribution feature vector is used as the basic structure of the matrix. In its channel dimension, the envelope values ​​of the synchronized ultrasonic acoustic emission signal are embedded in phase order. At the same time, the amplitude fluctuation sequence of real-time torque is resampled to align its data length with the number of phase intervals and is incorporated into the matrix as a third feature dimension. Specifically, a cubic spline interpolation algorithm is used for resampling to uniformly map the torque data and acoustic emission envelope data transmitted asynchronously by the frequency converter into a sequence of 128 or 256 points of equal length, consistent with the number of phase intervals. This gives the final composite feature matrix a fixed tensor shape of 128×N×3 in the spatial dimension. Through this tensor splicing method, a composite acoustic-electrical-mechanical feature matrix with multi-dimensional attributes in the phase domain (reflecting electric field distribution), time domain (reflecting mechanical evolution), and energy domain (reflecting damage intensity) is finally generated, providing structured data support for the input of subsequent deep learning models.

[0033] In this embodiment, the acoustic-electric-force composite feature matrix is ​​preferably represented in three-dimensional tensor form, with its dimensions arranged in the order of "phase index dimension × time slice index dimension × feature channel dimension". The length of the phase index dimension is consistent with the preset number of phase intervals, typically 128 or 256; the time slice index dimension is used to record feature snapshots within multiple consecutive power frequency cycles, typically with a length of 10-20; the feature channel dimension includes at least four types of channels: normalized partial discharge number values, normalized partial discharge amplitude values, normalized acoustic emission envelope values, and normalized real-time torque amplitude values. Through the above dimension definitions, those skilled in the art can directly construct a composite feature tensor structure that meets the requirements of this invention based on the sampled data without creative effort.

[0034] By zoning the load current phase and statistically analyzing its discharge characteristics, and then aligning the ultrasonic signal with the current zero point using a cross-correlation algorithm, deep fusion of heterogeneous data in the phase, time, and energy domains was achieved. This solved the problem of feature discretization caused by differences in sampling rate and physical properties of multi-source signals. The composite feature matrix generated through tensor concatenation provides a high-dimensional input source with physical correlation information for subsequent models.

[0035] Furthermore, the process of constructing the spatial topology graph includes: using the Euler-Savari formula in conjunction with the number of winding layers to calculate the cumulative radial tension coefficient of the cable at different stacking depths, and deriving the corresponding radial compressive force distribution gradient based on the elastic modulus of the cable cross-section; using the axial arrangement position and radial layer coordinates of the drum as vertices, and the mechanical contact paths between adjacent layers and between adjacent turns in the same layer as edges, constructing a three-dimensional mesh spatial topology graph characterizing the physical constraint relationship between cable layers. Due to the large number of cable nodes, this embodiment uses a sparse storage structure (such as an adjacency list) to record the connection relationship between nodes, storing only the edge information where physical contact exists. To reduce memory usage, the system uses a coordinate mapping function to convert the axial position, layer depth, and phase information of the cable into a matrix index, thereby enabling rapid positioning of sensor data and topology nodes. During early warning calculations, the system only performs topology calculations on local areas with severe stress and active discharge signals, ensuring real-time early warning under large-scale node conditions. The load voltage and load current are injected into the spatial topology graph as node attribute vectors, and the edge weights in the spatial topology graph are corrected according to the radial extrusion pressure distribution gradient combined with a pressure-weight mapping table, quantifying the electromagnetic coupling strength and thermal resistance characteristics between cable layers under different extrusion conditions.

[0036] In the spatial topology construction phase, a three-dimensional coordinate system is first initialized in memory, and preset drum geometric parameters are called, including the drum's bottom radius, the cable's outer diameter, and the material's cross-sectional elastic modulus. The current number of winding layers calculated by the absolute encoder is read in real time and used as a boundary condition input into the discrete mechanics calculation model. Specifically, a discrete iterative method is used, dividing the cable on the drum into several small calculation segments according to the single-turn rotation angle. Based on the Euler-Savari principle, starting from the outermost cable, the centripetal constraint force component generated by each turn of cable on the lower layer is calculated layer by layer. These centripetal force components are accumulated in the radial direction to obtain the cumulative radial tension coefficient at different stacking depths. Then, the elastic modulus of the cable cross-section is used as a conversion coefficient to convert the abstract tension coefficient into a specific physical pressure value. Specifically, the equivalent radius of curvature of the cable after being subjected to interlayer compression deformation is calculated first using the Euler-Savari formula combined with the current number of winding layers. When a cable is wound in multiple layers, the bottom layer of cable experiences a slight cross-sectional distortion due to the accumulated pressure from the upper layers. This physical deformation alters the actual path curvature of the cable on the drum. The real-time tension value, calculated from the real-time torque, is divided by this dynamically generated equivalent radius of curvature to obtain the initial unit area pressure of the current layer of cable on the contact surface of the lower layer. Subsequently, a discrete iterative algorithm is used to accumulate the pressure increment generated by each layer radially inward, starting from the outermost layer of the cable, until the calculation reaches the bottom layer in contact with the drum core. This forms pressure gradient data covering the entire winding depth, providing an accurate physical reference for subsequent edge weight correction of the spatial topology graph. Finally, a continuous pressure distribution gradient sequence is generated in the radial dimension. This sequence characterizes the attenuation trend of the compressive pressure from the center of the drum to the outermost cable, providing a physical basis for subsequent topology modeling.

[0037] Next, the system enters the logical modeling program for the topology, allocating a three-dimensional mesh data structure in memory space. Using the axial arrangement position of the cable roll (determined by the cable routing step) and radial layer coordinates as spatial vertex coordinates, a unique node index is established for each cable micro-segment. To characterize the physical constraints between cables, connection edges are automatically established based on the cable's geometric contact logic: traversing all nodes, if two nodes are on the same layer and their axial distance equals the routing step, a lateral connection edge is established in the adjacency matrix; if two nodes are on adjacent layers and their radial projections coincide or are on the mechanical contact path, a longitudinal connection edge is established. Considering the helix angle of the cable during actual winding, a [missing information - likely a specific method or technique] is used when finding longitudinal connection edges. The neighborhood radius search algorithm, centered on the current node, defines a spherical search area with a radius of 1.1 to 1.3 times the cable diameter in the radially adjacent upper and lower layers. If multiple candidate nodes are matched within the search area, the system prioritizes establishing logical connection edges with the node with the largest physical contact area with the current node according to the Euclidean distance minimization principle to eliminate redundant connections. All adjacent layer nodes falling into this area are connected by edges. This connection method can accurately simulate the physical contact and electric field distribution path of the cable in a non-ideal alignment state. In this way, a three-dimensional mesh space topology graph containing the set of nodes and the set of edges is constructed, and each node carries its real-time spatial coordinate information in the drum.

[0038] Subsequently, the node attribute injection and edge weight dynamic correction process is executed. The load voltage and load current data, which are collected in real time and normalized, are converted into node attribute vectors and mapped to the corresponding spatial topology nodes. In the vectorization process, the original waveform is not directly mapped. Instead, a sliding window feature extraction is performed on the load current and voltage. Specifically, the effective value of voltage, the effective value of current, and the phase angle difference between the two in the current cycle are extracted and encapsulated into a three-dimensional feature vector. If it is a node with obvious partial discharge characteristics, the energy centroid frequency of the high-frequency component is also extracted as a fourth dimension attribute. In this way, each spatial topology node not only has coordinates, but also carries the local electromagnetic excitation characteristics at that physical location. To quantify the impact of non-uniform compression on the electrical signal transmission environment, a pre-stored "pressure-weight mapping table" is invoked. This table is essentially a discrete function mapping relationship describing the relationship between pressure and electromagnetic coupling coefficient. This mapping table was obtained through pre-experimental calibration. During the experimental phase, the shift in the equivalent dielectric constant of the cable insulation layer and the change in interlayer mutual inductance coefficient were recorded under different pressure gradients (e.g., each 5 kPa increase represents a step). During the operational phase, based on the pressure value of the current node, the corresponding correction factor is retrieved from the table. An increase in the weight value indicates that under that pressure, the electromagnetic coupling path between the cable layers is more... If the thermal resistance decreases due to the compression of the air gap, the radial extrusion pressure distribution gradient value calculated above is extracted, and the weight of each connecting edge in the topology diagram is adjusted in real time: for interlayer connecting edges with high extrusion pressure, the corresponding weight value is increased according to the mapping table to simulate enhanced electromagnetic coupling strength; for areas with low extrusion pressure or high thermal resistance, the weight parameters are adjusted accordingly. Through this dynamic update of matrix values, the originally uniform mesh structure is endowed with non-uniform physical property characteristics, thereby realistically simulating the electromagnetic environment and heat dissipation boundary of the cable in a multi-layer winding state at the data level.

[0039] By introducing the Euler-Savari formula to quantify the radial extrusion pressure distribution and constructing a three-dimensional mesh space topology map characterizing the interlayer physical constraints, accurate modeling of the non-uniform physical environment under cable lap conditions was achieved. Through dynamic correction of edge weights using a pressure-weight mapping table, the interference of different extrusion degrees on electromagnetic coupling strength and thermal resistance characteristics was effectively quantified, compensating for signal gain fluctuations caused by changes in winding position.

[0040] Furthermore, the process of obtaining the theoretical insulation loss prediction value includes: using the composite feature matrix as node input, extracting and aggregating data from adjacent cable nodes based on the node connection relationship determined by the spatial topology graph to obtain an intermediate tensor that incorporates spatial location attributes; performing sliding sampling operation on the intermediate tensor on the time axis using a convolution kernel to extract the changing trend of the cable in continuous winding and unwinding cycles, and outputting the spatiotemporal feature vector of fixed dimension; merging the spatiotemporal feature vector with the real-time acquired load voltage, load current, and winding layer number to construct a combined input vector containing historical state features and current electrical parameter excitation; inputting the combined input vector into a fully connected layer for multi-layer linear transformation and nonlinear activation to compress and map the high-dimensional vector into a single-dimensional value, and outputting the theoretical insulation loss prediction value corresponding to the current moment.

[0041] In the specific implementation process of obtaining the theoretical insulation loss prediction value, the constructed acoustic-electrical-mechanical composite feature matrix is ​​first mapped onto each node of the spatial topology graph as the initial feature input of the node. Then, according to the pre-set adjacency matrix and edge weight relationship of the spatial topology graph, the spatial aggregation operation of the node features is performed. Specifically, for each cable node, neighboring nodes connected to it through horizontal edges (adjacent rings in the same layer) and vertical edges (between adjacent layers) are retrieved. The acoustic envelope, discharge frequency, and torque fluctuation data of these neighboring nodes are extracted and combined with the previously calculated edge weights (reflecting the electromagnetic coupling strength under compression), and a weighted sum is performed to incorporate local physical environment interference into the node attributes, generating an intermediate tensor containing spatial topology attributes. Before performing the weighted summation, the edge weights of adjacent nodes are first normalized. Specifically, the sum of the weights of all connected edges of the current node is calculated, and the original weight of each edge is divided by the sum to obtain a percentage contribution coefficient. Then, the feature vectors of neighboring nodes are multiplied by the corresponding contribution coefficients and accumulated. This approach ensures that the aggregated intermediate tensor remains consistent in magnitude regardless of whether the cable is in the inner (high-voltage) or outer (low-voltage) layer of the reel, thus avoiding data bias caused by spatial location differences.

[0042] Next, temporal feature extraction is performed on the generated intermediate tensors. A time-series sliding window with a fixed length is maintained in memory. The intermediate tensors in multiple consecutive sampling periods are arranged in chronological order, and a one-dimensional convolution kernel is used to perform convolution operations along the time axis. The length of the time-series sliding window is usually set to cover at least one complete mechanical vibration cycle or electrical cycle in the cable winding and unwinding cycle (e.g., 20 milliseconds to 100 milliseconds). The size of the one-dimensional convolution kernel is set to one-third to one-fifth of the window length, and the stride is set to 1. Through this short window and small stride configuration, the phase drift trajectory of the partial discharge pulse in a very short time can be captured with high resolution, thereby accurately extracting the spatiotemporal feature vector reflecting the insulation degradation trend. After pooling, the output of the convolutional layer is flattened into a fixed-dimensional spatiotemporal feature vector, which highly condenses the physical constraints of the current spatial location of the cable and its historical operating trend.

[0043] Subsequently, the effective values ​​of the load voltage and load current, as well as the current number of winding layers determined by the absolute encoder, are extracted in real time. These raw electrical parameters are then merged with the previously generated spatiotemporal feature vector. Before merging, the effective values ​​of the real-time load voltage and load current are standardized and scaled. The preset cable rated parameters are retrieved, and the instantaneous acquired values ​​are mapped to a numerical range consistent with the spatiotemporal feature vector (e.g., between 0 and 1). In this way, the weights of the merged combined input vector in each channel are comparable, preventing the high-dimensional electrical parameters from masking weak insulation characteristic signals. In this step, the "historical state characteristics" reflecting long-term deterioration trends and the "current electrical parameter excitation" reflecting instantaneous working intensity are concatenated into a high-dimensional combined input vector, ensuring that the prediction model considers both the physical fatigue accumulation of the cable and the current electrical load changes.

[0044] Finally, the combined input vector is input into the fully connected neural network of the regression prediction layer. This fully connected network contains multiple hidden layers, which are processed through multi-layer linear weighted transformation and non-linear activation functions (such as the ReLU function) to abstract and compress the high-dimensional features layer by layer. The fully connected network adopts a funnel-shaped architecture with at least three hidden layers. The number of nodes in the first hidden layer is consistent with the dimension of the combined input vector. Subsequently, the number of nodes in each layer decreases proportionally (50% in this embodiment). The design logic of this structure is to extract complex electromagnetic coupling non-linear features through the first few layers and achieve semantic dimensionality reduction of the features through the later layers. After each linear transformation, the negative fluctuations are set to zero by the ReLU activation function, retaining the effective positive features representing insulation damage. Finally, it is mapped to a single-dimensional loss prediction value in the output layer. In the last layer, the multi-dimensional hidden features are compressed into a single-dimensional value using a linear mapping operator. This value represents the theoretical insulation loss prediction value that the cable should have under the current operating conditions. This prediction value serves as the benchmark for subsequent residual calculations and is used to shield non-fault loss fluctuations caused by changes in winding position and extrusion pressure.

[0045] By performing node data aggregation and time-axis sliding sampling within a spatial topology, deep extraction of the spatiotemporal evolution characteristics of cable insulation state was achieved. Intermediate tensors containing temporal patterns were merged with real-time electrical parameters, and a multi-layer fully connected network was used to compress high-dimensional environmental features into theoretical insulation loss predictions. This provides a dynamic and accurate physical reference for assessing insulation degradation and enhances the stability of the system in predicting cable performance evolution under continuous deceleration and retraction cycles.

[0046] Further, the process of obtaining the first residual and the second residual includes: acquiring the voltage and current phase difference at the end of the cable reel in real time, and calculating the measured insulation loss value reflecting the overall insulation state of the cable; performing point-by-point subtraction between the measured insulation loss value and the theoretical insulation loss prediction value with the same time step, extracting the abnormal loss component that exceeds the model prediction range, and obtaining the first residual; using the voltage zero point of the load current as the starting reference, mapping the partial discharge pulse sequence to a phase period interval of 0 to 360 degrees, and counting the number of pulses emitted in each interval; retrieving the standard discharge frequency distribution curve corresponding to the current number of winding layers and load power from the pre-stored cable status fingerprint database; subtracting the actual number of pulses emitted from the standard discharge frequency distribution curve, calculating the pulse number increment in each phase interval, and defining the total increment in the entire cycle as the second residual.

[0047] In the specific implementation of obtaining the first and second residuals, the overall insulation state of the cable is first quantitatively calculated. Through the electrical parameter monitoring unit deployed at the end of the cable reel, the voltage and current waveforms at the end are collected in real time and synchronously. The phase characteristics of the fundamental waves of the two are identified by the zero-point detection algorithm or fast Fourier transform, and the phase difference between the voltage and current is calculated. Based on the phase difference, the loss tangent value characterizing the dielectric loss is calculated and defined as the measured insulation loss value reflecting the degree of overall insulation degradation of the cable. During the acquisition, the time stamp deviation between the end monitoring unit and the center is controlled within microseconds by the synchronous trigger pulse at the end of the reel shaft or the clock synchronization protocol of industrial Ethernet (such as PTP). The acquired end waveform data first enters a circular buffer for preprocessing. Before performing point-by-point subtraction, the measured sample points that completely match the calculation time of the theoretical prediction value are retrieved from the buffer according to the timestamp mark attached to the data packet, thereby eliminating the false residual fluctuations caused by signal transmission delay.

[0048] Next, a dynamic difference operation between the measured and predicted values ​​is performed. The theoretical insulation loss prediction value corresponding to the current time step output by the regression prediction layer in the previous steps is retrieved from the memory buffer and aligned with the measured insulation loss value calculated in real time. The theoretical prediction value is subtracted from the measured value using point-by-point subtraction logic, thereby filtering out the normal physical loss component caused by changes in the number of winding layers and fluctuations in mechanical pressure. The difference after the subtraction operation is the abnormal loss component that exceeds the prediction range of the model. It is encapsulated and defined as the first residual, which is used to characterize the insulation anomaly caused by non-environmental factors.

[0049] In the process of obtaining the second residual, the phase normalization mapping of the partial discharge pulse is first performed. Taking the positive zero-crossing point of the load current voltage as the time starting reference, a complete electrical cycle (i.e., 0 to 360 degrees) is divided into a preset number of phase intervals (such as 128 or 256 intervals). All partial discharge pulse sequences captured in the current cycle are traversed. According to the time offset of each pulse relative to the voltage zero point, it is assigned to the corresponding phase interval. The cumulative number of pulses emitted in each interval is counted to form an actual phase-discharge frequency distribution histogram. Before the quantity statistics, amplitude threshold filtering is first performed. According to the electromagnetic noise background of the current environment, a threshold (such as 5 picocoulombs or a specific millivolt level voltage) is set. Only pulses with amplitudes exceeding the threshold are judged as valid discharge events. At the same time, a high-pass filtering algorithm is used to remove low-frequency mechanical vibration interference related to torque fluctuation frequency. The pure pulse sequence after screening is then assigned according to the phase angle to ensure that the second residual can accurately reflect the micro-breakdown frequency inside the insulating medium, rather than external environmental noise.

[0050] Subsequently, a fingerprint comparison based on environmental characteristics is performed. The current number of winding layers and load power values ​​are extracted as search indexes. The standard discharge frequency distribution curve matching the specific operating condition is retrieved from the pre-stored cable status fingerprint database. This standard curve represents the frequency distribution that the cable should have under the current compression state and voltage stress due to normal background noise or weak discharge. The actual number of pulses emitted is subtracted from the standard distribution curve for each phase interval to calculate the increment of the number of pulses exceeding the standard value in each interval. When calculating the increment, a one-way subtraction logic is performed for each phase interval. If the actual number of pulses is greater than the standard curve value, the difference is recorded as a positive increment; if the actual value is less than or equal to the standard value, the increment for that interval is recorded as zero. This processing method aims to focus on "newly added abnormal discharge points" and avoid the random reduction of normal background fluctuations from canceling out abnormal signals in other phases, ensuring... The final summed second residual maintains extremely high sensitivity to capture weak discharge signals caused by insulation degradation. Finally, the pulse increments of all intervals within the entire cycle are summed and the total increment is defined as the second residual, which is used to quantify the abnormal activity level of active discharge points inside the cable. The cable status fingerprint database is constructed through factory calibration or offline learning. Specifically, in the initial stage of cable operation (in a healthy state), the system controls the drum to perform full-stroke winding and unwinding under different power conditions such as no load, half load, and full load, recording the partial discharge characteristics of the cable at each winding depth. The discharge pulses under these normal operating conditions are statistically averaged according to the phase distribution to generate the corresponding standard frequency distribution curve and stored as a "health fingerprint". Each fingerprint entry in the database is uniquely determined by the "winding layer number - load power" dual index, thus providing a reliable statistical reference for anomaly judgment during the operation phase.

[0051] By point-by-point difference between measured insulation loss values ​​and theoretical predictions, combined with the calculation of discharge frequency offset based on a cable condition fingerprint database, precise isolation of insulation anomaly components is achieved. This dual-residual extraction mechanism can effectively filter background loss interference under normal dynamic loads, and accurately capture the pulse number increment within a specific phase interval through the second residual, providing quantitative evidence for distinguishing between environmental mechanical noise and actual insulation discharge defects.

[0052] Furthermore, the process of outputting the fault warning level includes: extracting the historical failure rate of the cable corresponding to the current winding layer number as the initial prior distribution of the Bayesian decision model; calculating the deviation value of the real-time torque relative to the current load power, and adjusting the confidence of the initial prior distribution according to the deviation value to obtain the corrected prior probability; inputting the first residual and the second residual as the observation evidence set into the Bayesian decision model, and calculating the conditional probability likelihood values ​​of the first residual and the second residual under different damage states respectively; recursively updating the corrected prior probability and the conditional probability likelihood values ​​using the Bayesian formula to calculate the composite damage probability index characterizing the real-time deterioration degree of the cable insulation layer; performing interval matching between the composite damage probability index and multiple preset damage risk thresholds, and determining and outputting the corresponding fault warning level according to the matching result.

[0053] In the specific implementation of outputting fault warning levels, the prior parameters of the Bayesian decision model are first initialized. The current winding layer number determined by the absolute encoder is extracted as an index, and the historical failure rate data corresponding to that layer number is retrieved from the pre-stored cable life cycle database. This historical failure rate is defined as the initial prior distribution when the model enters the calculation loop. That is, before acquiring the current sensor observation data, the initial probability of the cable being in different damage states (such as normal, slight degradation, severe degradation, and breakdown risk) is derived based on historical statistical laws. Specifically, the initial prior distribution is constructed as a four-dimensional discrete probability vector. The four elements in the vector correspond to the four physical states of normal, slight degradation, severe degradation, and breakdown risk, respectively. In the initialization stage, according to the historical failure rate retrieved from the database, the failure probability is proportionally allocated to the three damage state elements, and the remaining probability value is assigned to the normal state element, ensuring that the sum of the four probabilities is always equal to 1. This vectorized representation provides a unified data structure for subsequent matrix operations.

[0054] Next, a priori probability correction based on mechanical stress state is performed. Real-time torque values ​​and current load power data are extracted, the ratio of the two is calculated, and compared with the torque-power mapping relationship under standard operating conditions to obtain the torque deviation value. If the torque deviation exceeds the preset mechanical disturbance threshold, it indicates that the cable is under abnormal physical tension or compression. At this time, a confidence adjustment operator reduces the weight of "normal state" in the initial prior distribution and correspondingly increases the probability score of "deteriorated state" to obtain a corrected prior probability that reflects the current dynamic mechanical background. When performing the correction, the confidence adjustment operator adopts linear decay logic based on the deviation ratio. First, the ratio of the torque deviation value to the preset threshold is calculated. If the ratio is greater than 1, the probability value of "normal state" is multiplied by a decay coefficient less than 1 (this coefficient decreases as the deviation ratio increases). Then, the deducted probability increment is redistributed to the "severely deteriorated" and "breakdown risk" states according to the preset weight coefficient. This step ensures that the abnormality of mechanical stress can be reflected in the tendency of the prior probability in real time and quantitatively.

[0055] Subsequently, the conditional probability likelihood value is calculated using the observational evidence set. The first residual (abnormal insulation loss) and the second residual (discharge pulse increment) obtained in the previous steps are used as the core observational evidence input model. For each preset damage state, a set of pre-calibrated probability density functions are run internally to calculate the probability interval of the current first residual value falling into each damage state's probability interval, and the probability of the second residual value falling into the corresponding interval. The internally preset probability density function adopts a Gaussian distribution model. For each damage state, the corresponding residual mean and standard deviation are pre-stored in the library. The processor substitutes the real-time acquired first and second residual values ​​into the Gaussian function operator of the corresponding state to calculate the probability density score of the value under the distribution of that state. For example, when the first residual (abnormal insulation loss) is large, its score under the "severe degradation" distribution will be significantly higher than that under the "normal state," thereby realizing the statistical support of the damage state by the observational evidence, and thus obtaining the conditional probability likelihood values ​​for the two types of residuals. This process realizes the transformation of macroscopic loss offset and microscopic discharge offset into mathematical statistical evidence.

[0056] In a preferred embodiment, the mean and standard deviation of the residuals under each damage state can be statistically estimated using a large amount of historical operating data collected during the offline phase: during the equipment health phase, the corresponding state is regarded as a "normal state" sample, and the mean and standard deviation of the residual distribution are calculated; after a minor degradation, severe degradation, or breakdown risk event occurs, the residual data within the corresponding time window is divided into corresponding state samples and their statistical characteristics are calculated respectively.

[0057] During the probability synthesis stage, the Bayesian formula is recursively updated, and the corrected prior probability is multiplied by the conditional probability likelihood value corresponding to each residual. The result is then normalized to ensure that the sum of the probabilities of all potential states is 1. After this logical iteration, the system calculates a value between 0 and 1, which is the composite damage probability index characterizing the real-time deterioration of the cable insulation layer. This index integrates historical fault probabilities, current mechanical stress offset, and acoustic-electric residual evidence, eliminating false alarms that may be caused by a single index.

[0058] Finally, the fault level interval matching and output are performed, and a preset damage risk threshold table is retrieved. This table divides the interval from 0 to 1 into multiple level segments, such as: safe segment, attention segment, warning segment, and danger segment. The calculated composite damage probability index is compared with these thresholds level by level to determine the interval range to which it belongs. Based on the comparison results, the display interface or alarm unit is driven to output the corresponding fault warning level (such as green normal, yellow warning, orange alarm, or red shutdown command), thereby completing the closed loop from multi-source data perception to decision assessment. In the threshold matching table, the interval division follows the principle of non-equal increment. In specific implementation, 0 to 0.4 is defined as the safe segment (green), 0.4 to 0.7 is defined as the attention segment (yellow), 0.7 to 0.9 is defined as the warning segment (orange), and above 0.9 is defined as the danger segment (red). This division method takes into account the accelerating effect of insulation degradation in the later stage, that is, when the probability index is close to 1, the risk level escalation is more sensitive, thus allowing sufficient emergency response time for ship electrical personnel.

[0059] By constructing a Bayesian decision model, the historical failure rate and real-time torque deviation corresponding to the number of winding layers are transformed into a dynamically corrected prior distribution, thus realizing the scientific classification of fault warning levels. The double residual is used as the evidence set for recursive updating, and the calculated composite damage probability index integrates the likelihood values ​​of multi-source observation data, making the final output fault level more consistent with the actual deterioration trajectory of the cable.

[0060] By integrating and monitoring electrical parameters, mechanical torque, and acoustic emission signals during the operation of the cable reel, a phase reference was used to achieve rigorous spatiotemporal alignment of multi-source heterogeneous data. The radial extrusion pressure distribution under different winding depths was accurately quantified by combining the Euler-Savari formula. By constructing a spatial topology map to dynamically correct electromagnetic coupling and thermal resistance characteristics, signal gain fluctuations and phase shifts caused by changes in winding position were effectively compensated. By using regression prediction models and Bayesian decision logic, macroscopic loss residuals and microscopic discharge frequency shifts were deeply integrated, realizing the transformation from single threshold judgment to composite damage probability assessment. In complex dynamic environments, it can effectively filter mechanical interference and accurately capture weak insulation degradation signals, providing a scientific early warning basis with high physical consistency for the preventive maintenance of marine cables.

[0061] Example 2: First, during winch operation, the acquisition unit deployed at the power output end acquires the load voltage and current in real time, and uses a high-frequency current transformer to capture the partial discharge pulse sequence in the current rising edge interval. At the same time, it reads the real-time torque of the drive motor through the frequency converter interface, and combines it with the absolute encoder installed at the axial end of the drum to accurately calculate the current number of cable winding layers based on the mapping relationship between the number of rotations and the arrangement step distance. The piezoelectric ceramic sensor attached to the inside of the drum flange synchronously acquires the ultrasonic acoustic emission signal generated by the cable extrusion.

[0062] In the data preprocessing stage, the system uses the phase of the load current as a time reference benchmark, divides a single cycle into several phase intervals, counts the number of discharges and amplitudes in each interval, and constructs a two-dimensional phase distribution vector that reflects the micro-discharge characteristics. Subsequently, the mechanical fluctuation period is determined by performing spectrum analysis on the real-time torque, and the cross-correlation algorithm is used to match the envelope abrupt change point of the acoustic emission signal with the current phase zero point to eliminate the transmission delay between heterogeneous signals. Finally, the phase distribution characteristics, the aligned acoustic signal, and the torque fluctuation sequence are tensor-joined to generate an acoustic-electric-mechanical composite feature matrix with phase domain, time domain, and energy domain attributes.

[0063] To address the non-uniform physical environment caused by the multi-layered winding of cable reels, the system uses the Euler-Savari formula and combines it with the current number of winding layers to calculate the cumulative radial tension coefficient of the cable at different stacking depths. This allows for the deduction of the corresponding radial extrusion pressure distribution gradient. Based on the reel's geometric parameters, the system constructs a three-dimensional mesh space topology map characterizing the physical constraint relationship between layers. Load voltage and current are injected into the map as node attributes, and edge weights are corrected according to the extrusion pressure gradient. This quantifies the electromagnetic coupling strength and thermal resistance characteristics between cable layers under different extrusion conditions.

[0064] Subsequently, the system enters the loss prediction process, taking the composite feature matrix as the initial input, and performing data aggregation based on the node connection relationship determined by the spatial topology map to obtain an intermediate tensor that integrates spatial location attributes. Using a convolution kernel, sliding sampling is performed on the time axis to extract the changing trend of the cable in continuous winding and unwinding cycles, and outputs a spatiotemporal feature vector. This vector is merged with the real-time electrical parameters and input into the fully connected layer. Through nonlinear mapping calculation, the theoretical insulation loss prediction value corresponding to the current working condition is obtained. This value serves as the dynamic benchmark for subsequent determination of the insulation status.

[0065] During the early warning determination phase, the system acquires the voltage and current phase difference at the cable end in real time to calculate the measured insulation loss value, and subtracts it from the theoretical prediction value to extract the first residual that exceeds the model prediction range. At the same time, the actual discharge pulse is mapped to the phase period, and the standard discharge curves under the corresponding layer number and power are retrieved from the state fingerprint library for comparison. The pulse number increment is calculated as the second residual. Finally, the two types of residuals are input into the Bayesian decision model, and combined with the prior probability corrected by the historical failure rate and torque deviation, the composite damage probability index is recursively updated to obtain the composite damage probability index. The corresponding fault early warning level is output according to the risk threshold.

[0066] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A data-based early warning method for marine cable faults, characterized in that, include: The system acquires the load voltage, load current, real-time torque of the drum drive motor, current number of cable winding layers, partial discharge pulse sequence, and ultrasonic emission signal during the operation of the cable reel. Using the phase of the load current as a time reference, the phase distribution of the partial discharge pulse sequence is analyzed. Combined with the mechanical fluctuation period of the real-time torque, the ultrasonic acoustic emission signal is synchronously correlated and aligned to construct a dynamic acoustic-electrical-mechanical composite feature matrix. The radial compressive force distribution of the cable on the drum is calculated based on the number of winding layers. A spatial topology map is constructed by combining the drum's geometric parameters. The load voltage and load current are mapped onto the spatial topology map. The composite feature matrix and the spatial topology map are input into the encoding layer to obtain the spatiotemporal feature vector. The spatiotemporal feature vector, along with the load voltage, load current, and number of winding layers, are input into the regression prediction layer to obtain the theoretical insulation loss prediction value. The first residual is calculated between the measured insulation loss value and the theoretical insulation loss prediction value. The discharge frequency offset of the partial discharge pulse sequence within the load current cycle is extracted as the second residual. The first residual and the second residual are input into the Bayesian decision model, and the composite damage probability index is calculated by combining the abnormal offset of the real-time torque. The fault warning level is then output.

2. The data early warning method for marine cable faults according to claim 1, characterized in that, The process of acquiring various data during the operation of the cable reel includes: The load voltage and load current are acquired by an electrical parameter acquisition unit deployed at the power output end of the drum, and a partial discharge pulse sequence is captured in the rising edge range of the load current using a high-frequency current transformer; the real-time torque is read in real time through the inverter communication interface of the drum drive motor, and the current winding layer number of the cable is calculated based on the linear mapping relationship between the total number of drum rotations and the cable arrangement step distance using an absolute encoder installed at the axial end of the drum; the ultrasonic acoustic emission signal generated by the cable during the winding and unwinding compression process is acquired by a piezoelectric ceramic sensor attached to the inside of the drum flange.

3. The data early warning method for marine cable faults according to claim 1, characterized in that, The process of constructing the dynamic acoustic-electric-force composite feature matrix includes: dividing the single cycle of the load current into several phase intervals, counting the number of discharges and the discharge amplitude of the partial discharge pulse sequence in each phase interval, and constructing a two-dimensional phase-discharge distribution feature vector; extracting the spectral features of the real-time torque to determine its mechanical rotation cycle data, and using a cross-correlation algorithm to match the envelope abrupt change point of the ultrasonic acoustic emission signal with the phase zero point of the load current within the mechanical rotation cycle data, thereby achieving synchronization of the acoustic signal and the electrical signal on the time axis; and tensor splicing the phase-discharge distribution feature vector, the synchronized ultrasonic acoustic emission signal, and the amplitude fluctuation sequence of the real-time torque to generate an acoustic-electric-force composite feature matrix with multi-dimensional attributes in the phase domain, time domain, and energy domain.

4. The data early warning method for marine cable faults according to claim 1, characterized in that, The process of constructing the spatial topology map includes: using the Euler-Savari formula in combination with the number of winding layers to calculate the cumulative radial tension coefficient of the cable at different stacking depths, and deriving the corresponding radial extrusion pressure distribution gradient based on the elastic modulus of the cable cross section; constructing a three-dimensional mesh spatial topology map characterizing the physical constraint relationship between cable layers, with the axial arrangement position of the drum and the radial layer coordinates as vertices and the mechanical contact paths between adjacent layers and between adjacent turns in the same layer as edges; injecting the load voltage and the load current as node attribute vectors into the spatial topology map, and correcting the edge weights in the spatial topology map according to the radial extrusion pressure distribution gradient and the pressure-weight mapping table to quantify the electromagnetic coupling strength and thermal resistance characteristics between cable layers under different extrusion states.

5. The data early warning method for marine cable faults according to claim 1, characterized in that, The process of obtaining the theoretical insulation loss prediction value includes: using the composite feature matrix as node input, extracting and aggregating data from adjacent cable nodes based on the node connection relationship determined by the spatial topology graph to obtain an intermediate tensor that incorporates spatial location attributes; performing sliding sampling operation on the intermediate tensor on the time axis using a convolution kernel to extract the changing trend of the cable in continuous winding and unwinding cycles, and outputting the spatiotemporal feature vector of fixed dimension; merging the spatiotemporal feature vector with the real-time acquired load voltage, load current, and winding layer number to construct a combined input vector containing historical state features and current electrical parameter excitation; inputting the combined input vector into a fully connected layer for multi-layer linear transformation and nonlinear activation to compress and map the high-dimensional vector into a single-dimensional value, and outputting the theoretical insulation loss prediction value corresponding to the current moment.

6. The data early warning method for marine cable faults according to claim 1, characterized in that, The process of obtaining the first residual and the second residual includes: acquiring the voltage and current phase difference at the end of the cable reel in real time to obtain the measured insulation loss value reflecting the overall insulation state of the cable; performing point-by-point subtraction between the measured insulation loss value and the theoretical insulation loss prediction value with the same time step to extract the abnormal loss component that exceeds the model prediction range and obtain the first residual; using the voltage zero point of the load current as the starting reference, mapping the partial discharge pulse sequence to a phase period interval of 0 to 360 degrees and counting the number of pulses emitted in each interval; retrieving the standard discharge frequency distribution curve corresponding to the current number of winding layers and load power from the pre-stored cable status fingerprint database; subtracting the actual number of pulses emitted from the standard discharge frequency distribution curve to calculate the pulse number increment in each phase interval, and defining the total increment in the entire cycle as the second residual.

7. The data early warning method for marine cable faults according to claim 1, characterized in that, The process of outputting the fault warning level includes: extracting the historical failure rate of the cable corresponding to the current winding layer number as the initial prior distribution of the Bayesian decision model; calculating the deviation value of the real-time torque relative to the current load power, and adjusting the confidence level of the initial prior distribution according to the deviation value to obtain the corrected prior probability; inputting the first residual and the second residual as the observation evidence set into the Bayesian decision model, and calculating the conditional probability likelihood values ​​of the first residual and the second residual under different damage states respectively; recursively updating the corrected prior probability and the conditional probability likelihood values ​​using the Bayesian formula to calculate the composite damage probability index characterizing the real-time degradation degree of the cable insulation layer; performing interval matching between the composite damage probability index and multiple preset damage risk thresholds, and determining and outputting the corresponding fault warning level according to the matching result.

8. A data early warning system for marine cable faults, characterized in that, include: The feature sensing module acquires the load voltage, load current, real-time torque of the drum drive motor, current number of winding layers of the cable, partial discharge pulse sequence and ultrasonic emission signal during the operation of the drum cable. The correlation and fusion module uses the phase of the load current as a time reference to analyze the phase distribution of the partial discharge pulse sequence. Combined with the mechanical fluctuation period of the real-time torque, it performs synchronous correlation and alignment of the ultrasonic acoustic emission signal to construct a dynamic acoustic-electric-force composite feature matrix. The loss prediction module calculates the radial compressive force distribution of the cable on the drum based on the number of winding layers, constructs a spatial topology map based on the drum's geometric parameters, and maps the load voltage and load current to the spatial topology map. The composite feature matrix and the spatial topology map are input into the encoding layer to obtain the spatiotemporal feature vector. The spatiotemporal feature vector, along with the load voltage, load current, and number of winding layers, are input into the regression prediction layer to obtain the theoretical insulation loss prediction value. The integrated early warning module calculates the first residual between the measured insulation loss value and the theoretical predicted insulation loss value, and extracts the discharge frequency offset of the partial discharge pulse sequence within the load current cycle as the second residual. The first residual and the second residual are input into the Bayesian decision model, and the composite damage probability index is calculated by combining the abnormal offset of the real-time torque, and the fault early warning level is output.

Images