A method for predicting cable life based on multi-stress acceleration
By employing a multi-stress accelerated cable life prediction method, semantic slicing and neural network modeling are performed using multi-source stress flow data and a knowledge base of cable material aging mechanisms. This solves the problem of distinguishing the semantics of stress combination working conditions in traditional methods, and achieves high-precision cable life prediction and risk area location.
Patent Information
- Application Number
- CN202611255296.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-18
- Publication Date
- 2026-11-10
AI Technical Summary
Traditional cable life prediction methods cannot effectively distinguish the semantics of different stress combinations and their historical evolution on the specific impact of insulation damage paths, lack the ability to respond to stress time evolution, and are difficult to achieve high-confidence full life cycle health management.
By acquiring multi-source stress flow data of cables in service environment, semantic slicing is performed using a cable material aging mechanism knowledge base to identify working condition combination features and generate stress path semantic segments. Supervised learning mapping is performed using a pre-trained stress path encoder to construct a cable degradation trajectory graph neural network. End-to-end reasoning is then performed to obtain the regional damage evolution probability distribution, and finally, the predicted value of the remaining cable life in high-risk areas is generated.
It achieves high-precision, spatially self-aware cable life prediction, which can accurately predict the remaining life of cables and locate high-risk areas without relying on empirical formulas or statistical distribution assumptions, thus improving the accuracy and interpretability of the prediction.
Smart Images

Figure CN123070180A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of cable health status monitoring and life prediction technology, and in particular to a cable life prediction method based on multi-stress acceleration. Background Technology
[0002] Traditional cable life prediction methods typically rely on accelerated aging experiments and material parameter identification, combined with classic statistical life models such as the Arrhenius equation, Weibull probability distribution, or Cox regression to fit degradation patterns. With the widespread adoption of sensors and online monitoring systems, some technical solutions have attempted to directly input data such as temperature, humidity, and vibration into data-driven neural networks or regression analysis models using multivariate stress models or simple multidimensional feature splicing methods. Representative existing technologies often employ static stress parameters or averaging methods to superimpose multi-source environmental stress data into feature vectors, which are then combined with statistically based life degradation models for prediction. Some studies have also attempted to introduce several stress coupling terms to describe the synergistic effects of multiple stresses, but these are limited to low-order linear or quadratic interaction terms, resulting in finite structures and weak responsiveness to stress time evolution.
[0003] However, the aforementioned existing technologies still have the following shortcomings: First, the stress field changes during cable service are highly dynamic and have complex temporal correlations. Traditional static models cannot distinguish the semantics of different stress combinations and their historical evolution's specific impact on insulation damage paths, resulting in limited accuracy in lifespan estimation and difficulty in maintaining robustness under sensor noise or data loss. Second, the causal mechanism between stress and material degradation is extremely complex. Models relying on high-order explicit functions or manually designed interaction terms cannot cover the real damage state distribution under varying real-world conditions and lack sufficient generalization ability. Third, many models rely too heavily on empirical formulas or global average degradation trends during design, lacking fine-grained real-time mapping of the spatial heterogeneity of damage and risk clustering evolution in different structural regions of the cable, making it difficult to achieve high-confidence full life-cycle health management. Summary of the Invention
[0004] This application provides a cable life prediction method based on multi-stress acceleration, which aims to solve one of the problems or issues of the prior art mentioned in the background.
[0005] This application provides a cable life prediction method based on multi-stress acceleration, which specifically includes:
[0006] S1: Acquire the stress flow input to be processed synchronously collected by the cable in the service environment.
[0007] S2: Based on the cable material aging mechanism knowledge base, perform semantic slicing processing on the stress flow input to be processed, identify the working condition combination features in the continuous stress data and generate stress path semantic segments with unique path labels.
[0008] S3: Perform supervised learning mapping operation on the stress path semantic segment using a pre-trained stress path encoder to convert the stress path semantic segment with a unique path label into a stress path embedding vector.
[0009] S4: Divide the physical region nodes according to the cable cross-sectional topology and initialize the intrinsic material parameters. At the same time, use the attention module to dynamically calculate the stress transmission weight based on the stress path embedding vector to construct a cable degradation trajectory neural network.
[0010] S5: The stress path embedding vector sequence within the sliding window is input into the cable degradation trajectory neural network for end-to-end inference to obtain the regional damage evolution probability distribution.
[0011] S6: Perform a weighted aggregation operation based on the regional damage evolution probability distribution to generate a predicted cable remaining life value that includes a heat map of high-risk areas.
[0012] This application provides a cable life prediction method based on multi-stress acceleration, which has the following advantages:
[0013] This application employs step S1 to simultaneously collect multi-source stress flow data including temperature, vibration, and humidity, providing a complete dynamic stress input for subsequent semantic modeling. Step S2 uses a cable material aging mechanism knowledge base to perform semantic slicing of the stress flow input, identifying the combined characteristics of operating conditions in continuous stress data and generating stress path semantic fragments with unique path labels. This correlates macroscopic monitoring signals with microscopic failure modes, solving the problem that traditional static models cannot distinguish the semantics of different stress combinations and their specific historical evolution paths. Step S3 utilizes a pre-trained stress path encoder to map semantic fragments into embedding vectors strongly correlated with microscopic degradation mechanisms, enabling the model to reconstruct the microscopic damage-driving mechanism from macroscopic signals, overcoming the mechanism misjudgment caused by neglecting operating condition semantics in traditional methods. Step S4 divides physical region nodes according to the cable cross-sectional topology and initializes the intrinsic material parameters. Simultaneously, an attention module dynamically calculates stress transmission weights based on the embedding vectors to construct a cable degradation trajectory graph neural network, allowing edge weights to dynamically change with real-time operating conditions. This solves the problem that models relying on high-order explicit functions or manual interaction terms cannot cover the true damage state distribution. Step S5 inputs the embedded vector sequence within the sliding window into the graph neural network for end-to-end inference to obtain the regional damage evolution probability distribution. Step S6 performs weighted aggregation based on this probability distribution to generate remaining lifetime prediction values that include heat maps of high-risk areas. This ensures that the prediction results simultaneously include a global lifetime scale and spatial risk location, solving the problem of existing technologies lacking fine-grained characterization of spatial heterogeneity and risk cluster evolution.
[0014] The above steps sequentially link multi-source stress sensing, stress path semantic slicing, encoder embedding, graph neural network spatial topology modeling, regional damage extrapolation, and global lifetime aggregation into a complete decision-making chain, enabling the system to achieve high-precision, interpretable, and spatially self-aware cable lifetime extrapolation without relying on empirical formulas or statistical distribution assumptions. Attached Figure Description
[0015] Figure 1 This is the main flowchart of a cable life prediction method based on multi-stress acceleration;
[0016] Figure 2 This is a sub-flowchart of a cable life prediction method based on multi-stress acceleration;
[0017] Figure 3 This is another sub-flowchart of a cable life prediction method based on multi-stress acceleration. Detailed Implementation
[0018] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.
[0019] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.
[0020] like Figure 1 As shown, this application provides a cable life prediction method based on multi-stress acceleration, specifically including:
[0021] S1: Acquire the stress flow input to be processed synchronously collected by the cable in the service environment.
[0022] S2: Based on the cable material aging mechanism knowledge base, perform semantic slicing processing on the stress flow input to be processed, identify the working condition combination features in the continuous stress data and generate stress path semantic segments with unique path labels.
[0023] S3: Perform supervised learning mapping operation on the stress path semantic segment using a pre-trained stress path encoder to convert the stress path semantic segment with a unique path label into a stress path embedding vector.
[0024] S4: Divide the physical region nodes according to the cable cross-sectional topology and initialize the intrinsic material parameters. At the same time, use the attention module to dynamically calculate the stress transmission weight based on the stress path embedding vector to construct a cable degradation trajectory neural network.
[0025] S5: The stress path embedding vector sequence within the sliding window is input into the cable degradation trajectory neural network for end-to-end inference to obtain the regional damage evolution probability distribution.
[0026] S6: Perform a weighted aggregation operation based on the regional damage evolution probability distribution to generate a predicted cable remaining life value that includes a heat map of high-risk areas.
[0027] S7: Update the stress path embedding vector based on the real-time online monitored stress flow data and trigger the time-step state iteration of the cable degradation trajectory neural network, outputting the corrected cable remaining life prediction value to complete the life extrapolation closed loop under dynamic multi-stress path.
[0028] Step S1: Acquire the stress flow input to be processed, synchronously collected by the cable in its service environment. The stress flow input to be processed includes one or more combinations of temperature fluctuation sequences, mechanical vibration spectrum sequences, and humidity time-series data. Specifically, it includes:
[0029] S1.1: Perform clock synchronization calibration on the distributed temperature sensor array, triaxial vibration accelerometer and capacitive humidity sensing unit deployed at key monitoring points of the cable to eliminate timestamp deviations between multiple source sensors and generate a set of original sensing signals with a unified time reference.
[0030] S1.2: Based on the original sensor signal set with a unified time reference, perform sliding window truncation and resampling operations to match the sampling frequency differences of different sensors and generate time-aligned multi-channel discrete sampling data frames.
[0031] The system receives a set of multi-source raw sensor signals after clock synchronization calibration. This set of multi-source raw sensor signals includes discrete time-series data from three channels: temperature, vibration, and humidity.
[0032] To address the issue of inconsistent sampling frequencies across channels, a target uniform sampling rate is determined. Typically, the highest sampling frequency or the least common multiple of the Nyquist sampling theorem is selected as the reference resampling frequency.
[0033] For the low-frequency sampled temperature and humidity channel data, cubic spline interpolation is used for upsampling to construct a continuous and smooth signal fitting curve. Interpolated data is then extracted at the target time point to fill the data gaps on the high-frequency time axis.
[0034] For high-frequency sampled vibration acceleration signals, if the sampling rate is higher than the target reference frequency, an anti-aliasing low-pass filter is designed to preprocess the original signal, and then a decimation operation is performed to reduce the data dimensionality, ensuring that the spectral information does not undergo aliasing distortion during downsampling.
[0035] Define a sliding time window of fixed length, set the window size and step parameters, and simultaneously capture data segments from three channels on a unified time axis to form a multidimensional data matrix.
[0036] Zero-mean normalization is performed on the extracted multidimensional data matrix to eliminate the numerical magnitude difference between different physical dimensions, so that the data of each channel are on the same order of magnitude. The resulting standardized multidimensional data matrix is the time-aligned multi-channel discrete sampling data frame, which facilitates subsequent feature fusion.
[0037] Through the above resampling and alignment processing, the original signals with a unified time base but different frequencies in the previous step are transformed into multi-channel discrete sampling data frames with strict time alignment and standardized dimensions. This achieves accurate matching of multi-source heterogeneous data in the time domain, providing a high-quality data foundation for subsequent frequency domain decomposition and outlier cleaning.
[0038] S1.3: The vibration channel data in the time-aligned multi-channel discrete sampling data frame is decomposed in the frequency domain using the Fast Fourier Transform algorithm to extract the energy distribution characteristics in a specific frequency band and generate a mechanical vibration spectrum sequence containing the main frequency component and harmonic components.
[0039] It receives multi-channel discrete sampled data frames that have undergone time alignment and normalization, and extracts the vibration acceleration signal channel data as the target for frequency domain analysis.
[0040] The extracted time-domain vibration signal is windowed using the Hanning window function to suppress spectral leakage, ensure a smooth transition of boundary data, and generate a windowed time-domain signal sequence.
[0041] The Fast Fourier Transform (FFT) algorithm is applied to the windowed signal sequence to convert the time-domain discrete signal into a complex frequency-domain signal, thus achieving a mapping from the time dimension to the frequency dimension.
[0042] The power spectral density distribution is obtained by calculating the square of the magnitude of the frequency domain signal. Based on the Nyquist sampling theorem, the effective frequency range is determined to be from zero to half the sampling frequency.
[0043] Within the effective frequency range, search for the global maximum point of the power spectral density, and identify the corresponding frequency as the main frequency component, representing the energy concentration band of mechanical vibration within the current window.
[0044] Local maxima in the power spectral density, excluding the dominant frequency, are identified. Peak frequencies with amplitudes exceeding a preset noise threshold are selected and marked as harmonic components, reflecting the nonlinear distortion characteristics of the vibration signal.
[0045] The frequency values and corresponding amplitudes of the dominant frequency component and each harmonic component are extracted and arranged in order of frequency from low to high to construct a structured mechanical vibration spectrum sequence containing dominant frequency and harmonic information.
[0046] By using the frequency domain decomposition process described above, the time-domain vibration data from the previous step is transformed into a mechanical vibration spectrum sequence containing the dominant frequency component and harmonic components, thereby realizing the quantitative extraction of vibration energy distribution characteristics and providing frequency domain feature input for subsequent stress path semantic slicing.
[0047] For example, the vibration signal sampling rate is set to 1000Hz, and the sliding window length is 1 second. A Hanning window is applied to the vibration data of 1000 sampling points, and a 1024-point FFT operation is performed. In the 0-500Hz frequency band, the dominant frequency is detected to be 50Hz with an amplitude of 2.5g; the harmonic components include 100Hz (amplitude 0.8g) and 150Hz (amplitude 0.3g). The frequencies and amplitudes are encapsulated into a sequence structure of [(50, 2.5), (100, 0.8), (150, 0.3)]. This spectral sequence accurately reflects the fundamental frequency vibration and its harmonic harmonics caused by motor operation, significantly improving the characterization ability of periodic mechanical stress characteristics.
[0048] S1.4: Perform outlier removal and linear interpolation repair processing on the temperature channel data and humidity channel data in the time-aligned multi-channel discrete sampling data frame to eliminate noise interference caused by sensor transient failures and generate cleaned temperature fluctuation sequence and humidity time series data.
[0049] Receive time-aligned multi-channel discrete sampling data frames and extract the original sequences of the temperature channel and humidity channel as the objects to be cleaned.
[0050] For temperature channel data, the local mean and standard deviation within the sliding window are calculated, a dynamic threshold interval based on the 3σ criterion is constructed, and data points exceeding this interval are identified and marked as outliers.
[0051] Outliers are removed by marking them, creating data gaps on the timeline while retaining the time-series indexes and values of valid data points.
[0052] Linear interpolation is used to handle data gaps in the temperature channel. The two most recent valid data points before and after the gap are selected, and a linear equation is constructed to calculate the temperature estimate at the time of the gap.
[0053] The interpolation result is calculated using the following formula:
[0054]
[0055] in, for Interpolated temperature at time, , These are the temperature values of adjacent valid points, respectively. , This is the corresponding timestamp.
[0056] The calculated interpolated temperature values are filled into the original missing positions to generate a preliminary cleaned temperature fluctuation sequence.
[0057] The same outlier detection process is performed on the humidity channel data, using a moving median filter to suppress sudden spike noise and identify outliers that deviate from the median by more than a preset multiple.
[0058] Outliers in the humidity channel are removed, and data breakpoints are repaired by linear combination of effective humidity values before and after, ensuring the physical continuity of humidity changes and obtaining time-series humidity data after cleaning.
[0059] By removing outliers and performing linear interpolation repair, the multi-channel data containing noise interference in the previous step is transformed into smooth, continuous, and missing temperature fluctuation sequences and humidity time-series data, thereby achieving quality purification of multi-source stress data and providing a high-confidence input basis for subsequent semantic slicing.
[0060] S1.5: The generated mechanical vibration spectrum sequence, the temperature fluctuation sequence after cleaning, and the humidity time series data after cleaning are structured and encapsulated and labeled with metadata to form a standardized multi-source heterogeneous data stream object, and the stress flow input to be processed is generated.
[0061] Step S2: Based on the cable material aging mechanism knowledge base, perform semantic slicing processing on the stress flow input to be processed, identify the working condition combination features in the continuous stress data and generate stress path semantic segments with unique path labels.
[0062] The cable material aging mechanism knowledge base is a pre-built structured dataset used to store the criteria for determining the aging failure of cable insulation materials under multi-stress coupling. Specifically, the knowledge base includes the following: a failure mode tag set covering typical microscopic damage types such as electrical tree initiation, water tree growth, interface debonding, thermal oxidation degradation, and mechanical fatigue fracture; a threshold rule set storing the critical thresholds for three stress components: temperature change rate, vibration energy density, and humidity gradient, where the critical thresholds are taken from the stress values that cause initial deterioration of the insulation material in accelerated life tests; a working condition combination identification logic rule set defining multi-dimensional stress threshold intervals and their coupling mapping relationships, used to match standardized stress feature vectors to corresponding working condition type identifiers; and a mapping index table of failure modes and stress paths, establishing the correspondence between semantic tags and failure mode categories. This knowledge base is constructed based on historical accelerated aging test data and cable material physical property parameters through offline statistical analysis and is iteratively updated in practical applications as new test data accumulates.
[0063] Specifically, it includes: S2.1: Acquire the temperature fluctuation sequence, mechanical vibration spectrum sequence and humidity time series data in the stress flow input to be processed, and use a sliding time window to synchronously extract and process the multi-source heterogeneous raw data to generate a standard time window data block containing the multi-dimensional stress state at the same time, which serves as the basic execution object for subsequent feature extraction.
[0064] The system receives a pre-processed stress flow input, which includes a time-aligned sequence of mechanical vibration spectra, a temperature fluctuation sequence after cleaning, and humidity time-series data.
[0065] Set the length and step parameters of the sliding time window, where the window length is determined based on the least common multiple of the thermal response time constant of the cable insulation material and the period of the dominant frequency of mechanical vibration, so as to ensure that the complete stress coupling cycle is captured.
[0066] On a unified timeline, data segments from three channels—temperature, vibration, and humidity—are captured synchronously at set intervals to form a multidimensional data matrix.
[0067] For each extracted data segment, its start and end timestamps are extracted to construct a standard time window data block containing multi-source stress states.
[0068] By constructing standardized time window data blocks, continuous multi-source heterogeneous data streams are discretized into analysis units with independent physical meaning, realizing the basic data structuring for subsequent working condition feature identification, and ensuring that the semantic slicing process can accurately reflect the multi-stress coupling state within a specific time period.
[0069] S2.2: Based on the standard time window data block, the threshold rule set in the cable material aging mechanism knowledge base is called to perform multi-dimensional feature quantization calculation on the temperature change rate, vibration energy density and humidity gradient, and output a standardized stress feature vector that characterizes the intensity of each stress component in the current window.
[0070] Receive the standard time window data block generated by S2.1, which contains synchronous time segments of three dimensions: temperature, vibration spectrum, and humidity.
[0071] The threshold rule set in the cable material aging mechanism knowledge base is called to extract the temperature change rate calculation logic, vibration energy density integral interval and humidity gradient judgment criteria.
[0072] Perform differential operations on the temperature channel data to calculate the ratio of temperature difference to time interval between adjacent sampling points, and obtain the instantaneous temperature change rate sequence.
[0073] The instantaneous temperature change rate sequence is smoothed by using a moving average filter to eliminate high-frequency measurement noise, and the maximum temperature change rate within the window is extracted as an indicator of the severity of thermal stress.
[0074] For the vibration spectrum sequence, the amplitude spectrum line of the corresponding frequency band is extracted according to the mechanical resonance frequency band range defined by the knowledge base.
[0075] Calculate the sum of squares of the amplitudes of each frequency component within the cut-off frequency band, and integrate over the time axis to obtain the vibration energy density value within the window.
[0076] For humidity time series data, calculate the humidity difference between the start and end times of the window, divide it by the window duration, and obtain the average humidity gradient.
[0077] The maximum temperature change rate, vibration energy density, and average humidity gradient are divided by their corresponding normalized reference values to achieve dimensional uniformity.
[0078] The normalized baseline value is taken from the critical stress threshold that causes the insulation material to undergo initial deterioration in the accelerated life test.
[0079] The standardized stress eigenvector is constructed, and its mathematical expression is as follows:
[0080]
[0081] in, To standardize the stress eigenvector, For the maximum rate of temperature change, The threshold for the rate of temperature change. For vibrational energy density, The vibration energy density threshold, The average humidity gradient, This is the humidity gradient threshold.
[0082] Through the above multi-dimensional feature quantization and normalization processing, the time window data block of the previous step is transformed into a standardized stress feature vector that characterizes the intensity of each stress component in the current window, thereby realizing the physical meaning alignment of multi-source heterogeneous stress data and providing a unified quantitative input basis for subsequent working condition mode matching.
[0083] S2.3: Using predefined working condition combination identification logic, the standardized stress feature vector is processed by pattern matching. Feature combinations that conform to specific coupling relationships are mapped to specific working condition type identifiers, generating preliminary working condition classification labels that reflect the thermal-vibration-humidity synergistic action mode.
[0084] The standardized stress feature vector is received, which contains quantized values of three dimensions: temperature change rate, vibration energy density, and humidity gradient.
[0085] The system invokes a set of operating condition combination identification logic rules pre-stored in the cable material aging mechanism knowledge base. The set of operating condition combination identification logic rules defines multidimensional stress threshold ranges and their coupling mapping relationships.
[0086] For each component in the standardized stress feature vector, interval matching judgment is performed. Based on whether the temperature change rate exceeds the preset thermal shock threshold, whether the vibration energy density falls into the mechanical fatigue frequency band, and whether the humidity gradient reaches the critical value of moisture absorption and expansion, a binary status identifier is generated respectively.
[0087] Based on Boolean logic operations, the status flag bits are combined and analyzed to identify the working conditions of single stress-dominated or multi-stress-coupled types.
[0088] For the identified operating conditions, the deviation ratio of each stress component from its threshold is calculated, and a weighted coupling strength index is constructed, as shown in the following formula:
[0089]
[0090] in, The coupling strength index is... , , These are characteristic values for temperature, vibration, and humidity, respectively. , , For the corresponding threshold, , , These are weighting coefficients determined based on aging mechanisms.
[0091] Based on the magnitude of the coupling strength index and the dominant stress type, a matching semantic tag is retrieved from a predefined working condition type identifier library, such as "thermal-vibration synergistic activation" or "damp heat aging dominant type".
[0092] Assign the retrieved operating condition type identifiers to the current time window to generate preliminary operating condition classification labels.
[0093] Through the above pattern matching and quantization processing, the standardized stress feature vector from the previous step is transformed into a preliminary working condition classification label with clear physical meaning, realizing the mapping of multi-source heterogeneous stress data to semantic working condition description, and providing a discretization basis for subsequent time series smoothing.
[0094] S2.4: Based on the preliminary working condition classification labels and combined with the time continuity constraint, perform temporal smoothing and boundary correction operations on the identification results of adjacent windows to eliminate instantaneous noise interference and generate a continuous stress working condition sequence with physical coherence.
[0095] Receive the preliminary operating condition classification label sequence generated by S2.3, which contains the operating condition type identifier corresponding to the discrete window on the time axis.
[0096] The constraint is that the working condition type of adjacent time windows should not frequently change. Temporal smoothing is applied to the preliminary working condition classification label sequence: a state transition penalty coefficient is set, and an energy penalty is applied to the identification results of working condition jumps between adjacent windows. The larger the jump amplitude, the higher the penalty. The energy penalty and the local identification confidence of each window are used as optimization objectives. A dynamic programming algorithm is used to solve for the globally optimal label sequence that minimizes the total cost.
[0097] Boundary correction is performed on the globally optimal label sequence: isolated operating condition segments with a duration lower than a preset thermal relaxation time threshold are detected and merged into adjacent dominant operating condition categories to eliminate misclassification caused by transient noise.
[0098] Through the above-mentioned time-series smoothing and boundary correction processes, the discrete preliminary labels from the previous step are transformed into a continuous operating condition sequence that conforms to the physical laws of cable aging, thereby achieving the expected technical effect of eliminating instantaneous noise interference and improving the robustness of operating condition identification.
[0099] S2.5: Perform label assignment processing on each complete working condition evolution process according to the unique hash encoding rule of the continuous stress working condition sequence, so as to finally generate stress path semantic fragments with unique path labels and corresponding to the micro-degradation mechanism.
[0100] Receive a sequence of continuous stress conditions after time-smoothing and boundary correction. This sequence consists of a series of physically coherent condition type identifiers.
[0101] Define unique hash encoding rules based on the SHA-256 algorithm, set the hash digest length and collision detection mechanism to ensure that different stress paths are mapped to unique identifiers.
[0102] For each complete working condition evolution process in the continuous stress working condition sequence, serialization splicing is performed, and the working condition type identifiers of adjacent time windows are combined into a string form of the original path descriptor in chronological order.
[0103] Add static attribute metadata such as cable model, insulation material type and service life to the prefix of the original path descriptor to construct a complete path input string containing spatiotemporal context information, so as to distinguish the aging benchmark differences of different individual cables.
[0104] The SHA-256 hash function is used to perform a one-way hash operation on the complete path input string to generate a fixed-length hexadecimal hash value as a preliminary path fingerprint, ensuring that the same stress evolution history will inevitably produce the same fingerprint identifier.
[0105] A micro-degradation mechanism association table is introduced to query the typical failure mode weight coefficients corresponding to the initial path fingerprint. If the current path fingerprint already exists in the historical database and is marked with a specific failure tendency (such as electrical treeing dominance), the fingerprint is retained and a failure mode label is attached.
[0106] If the current path fingerprint is newly generated, calculate its edit distance with the known standard aging path. When the edit distance is less than a preset threshold, classify it into a similar path cluster and inherit the main failure features within the cluster. Otherwise, mark it as a new stress path and assign it an independent new fingerprint identifier.
[0107] The generated unique path fingerprint and the corresponding micro-degradation mechanism label are structurally encapsulated to form a stress path fragment object with semantic information, wherein the fingerprint is used as the primary key index and the label is used as the attribute field to obtain the stress path semantic fragment.
[0108] Through the hash encoding and mechanism association processing methods described above, the continuous working condition sequence of the previous step is transformed into a stress path semantic segment with a unique identifier and containing microscopic damage mechanism information, realizing the digital mapping from macroscopic stress state to microscopic degradation trajectory, and providing a physically interpretable discretized input unit for the subsequent encoder.
[0109] like Figure 2 As shown, step S3: A supervised learning mapping operation is performed on the stress path semantic segment using a pre-trained stress path encoder, converting the stress path semantic segment with a unique path label into a fixed-dimensional stress path embedding vector strongly correlated with the microscopic degradation mechanism. Specifically, this includes:
[0110] S3.1: Based on the failure mode label set defined in the cable material aging mechanism knowledge base, perform one-hot encoding on the input stress path semantic fragment with unique path label to generate an initial working condition feature vector characterizing the specific working condition combination features.
[0111] Receive the stress path semantic fragment with a unique path label generated in step S2.5, which represents the coupled operating condition mode of temperature, vibration and humidity within a specific time period.
[0112] The system calls upon a predefined set of failure mode labels from the cable material aging mechanism knowledge base. This set of failure mode labels covers typical microscopic damage types such as electrical tree initiation, interface debonding, and polymer chain breakage.
[0113] Establish a mapping index table between semantic tags and failure modes to determine the specific failure mode category identifier corresponding to the current stress path semantic segment.
[0114] One-hot encoding is performed on the identified failure mode category identifiers to construct a zero vector with a dimension equal to the total number of failure modes. In the zero vector, the index position corresponding to the current failure mode category is set to 1, and the remaining positions are kept as 0, generating the initial binary feature vector.
[0115] The binarized feature vector is output as the initial working condition feature vector to characterize the attribution attribute of the current stress path at the micro-degradation mechanism level.
[0116] By using one-hot encoding, the discrete semantic labels from the previous step are transformed into sparse numerical vectors that can be processed by a computer, achieving a standardized transformation from qualitative working condition description to quantitative feature input, and ensuring that the subsequent neural network can accurately identify the micro-damage tendency corresponding to different stress paths.
[0117] For example, suppose the cable material aging mechanism knowledge base defines five main failure modes: electrical tree initiation, water tree growth, interface debonding, thermal oxidative degradation, and mechanical fatigue fracture. The current stress path semantic segment's label recognition result is "thermal-vibration-humidity co-activation condition." A lookup in the mapping index table shows that this condition mainly induces the "interface debonding" failure mode, with an index of 2 in the failure mode list. A zero vector of dimension 5 is constructed [0, 0, 0, 0, 0]. The value at index 2 is set to 1, resulting in the initial condition feature vector [0, 0, 1, 0, 0]. This vector clearly indicates the main damage direction of the current stress path, eliminating unstructured differences in the label text and providing a unified numerical input basis for subsequent extraction of high-dimensional nonlinear coupling features by the multilayer perceptron, significantly improving the model's response sensitivity to specific microscopic degradation mechanisms.
[0118] S3.2: The stress path encoder containing a multilayer perceptron structure is used to perform nonlinear transformation processing on the initial working condition feature vector to extract the high-dimensional nonlinear stress coupling features hidden in the semantic labels and output the intermediate layer stress coupling feature tensor.
[0119] The initial working condition feature vector generated by S3.1 is received. The stress path encoder adopts a multilayer perceptron structure, which consists of a first input layer, a first fully connected hidden layer, a second fully connected hidden layer, a third fully connected hidden layer, and a first output layer connected sequentially.
[0120] The dimension of the first input layer is the same as the dimension of the initial working condition feature vector; each fully connected hidden layer is connected in sequence, and all of them use the ReLU activation function for nonlinear mapping; the first output layer maps the output of the third fully connected hidden layer to the dimension of the intermediate layer stress coupling feature tensor.
[0121] The first input layer receives the initial working condition feature vector and passes it to the first fully connected hidden layer. The first, second, and third fully connected hidden layers sequentially receive the output of the previous layer, perform linear weighting and bias superposition operations, and then pass it to the next layer after element-wise nonlinear mapping by the activation function, until all three fully connected hidden layers have completed processing. The first output layer receives the feature vector output by the third fully connected hidden layer, performs linear weighting and bias superposition operations, and outputs an intermediate layer stress coupling feature tensor containing deep semantic information. Through the layer-by-layer nonlinear transformation processing of the multilayer perceptron described above, the discrete one-hot encoded vector of the previous step is transformed into a continuous high-dimensional stress coupling feature tensor.
[0122] For example, if the cable material aging mechanism knowledge base defines 5 main failure modes, then the dimension N of the initial operating condition feature vector is 5. The multilayer perceptron contains 3 hidden layers with 128, 64, and 32 nodes respectively. The first input layer receives an N-dimensional initial working condition feature vector. The first fully connected hidden layer maps this vector to 128 dimensions, performs linear weighting and bias superposition, and then performs non-linear mapping through the ReLU activation function, outputting a 128-dimensional feature vector to the second fully connected hidden layer. The second fully connected hidden layer maps the 128 dimensions to 64 dimensions, performs linear weighting and bias superposition, and then performs non-linear mapping through the ReLU activation function, outputting a 64-dimensional feature vector to the third fully connected hidden layer. The third fully connected hidden layer maps the 64 dimensions to 32 dimensions, performs linear weighting and bias superposition, and then performs non-linear mapping through the ReLU activation function, outputting a 32-dimensional feature vector to the first output layer. The first output layer maps the 32 dimensions to the target dimension (32 dimensions in this embodiment) of the intermediate layer stress coupling feature tensor, outputs the intermediate layer stress coupling feature tensor, and passes it to step S3.3 to perform supervised learning optimization.
[0123] S3.3: Construct a comparative loss function based on the stress path sample library labeled with actual failure modes in accelerated testing, perform supervised learning optimization operation on the intermediate layer stress coupling feature tensor, adjust the parameters of the stress path encoder, and generate an optimized stress coupling feature tensor with the ability to discriminate micro-degradation mechanisms.
[0124] Receive the intermediate layer stress coupling characteristic tensor output by S3.2, which contains the implicit nonlinear coupling information of the current stress path.
[0125] A supervised learning optimization environment based on an accelerated test sample library is constructed. The accelerated test sample library contains stress path labels for known failure modes (such as electrical tree initiation and interface debonding) and their corresponding microscopic degradation characterization data. A contrastive loss function is defined to quantify the distribution distance between similar and dissimilar operating conditions in the feature space, ensuring that stress paths with the same failure mechanism cluster in the embedding space, while paths with different mechanisms are separated from each other.
[0126] Using the triplet loss function as the optimization objective, the loss value is calculated using the following formula:
[0127]
[0128] in, The feature vector of the anchor sample. For positive samples (same failure mode) feature vectors, The feature vector of negative samples (different failure modes) For boundary spacing parameters, This is the loss value.
[0129] Based on the loss value, the gradient of the loss function with respect to the parameters of the stress path encoder is calculated using the backpropagation algorithm, and parameter biases that cause the dispersion of similar samples or the confusion of dissimilar samples are identified.
[0130] Repeat the above loss calculation and parameter update process until the loss value converges to a preset threshold, causing the feature vectors of similar stress paths to cluster in the feature space, while the feature vectors of dissimilar paths to separate from each other. After the above supervised learning optimization, the resulting intermediate layer stress coupling feature tensor is the optimized stress coupling feature tensor with the ability to discriminate microscopic degradation mechanisms.
[0131] By comparing the supervised learning optimization of the loss function, the intermediate layer stress coupling feature tensor of the previous step is transformed into an optimized stress coupling feature tensor with a clear ability to discriminate microscopic degradation mechanisms. This achieves a strong correlation mapping between stress semantics and physical failure mechanism, significantly improving the modeling accuracy of the subsequent graph neural network for unsteady coupling effects.
[0132] For example, setting the boundary spacing parameter The distance was set to 0.5. Samples exhibiting interface delamination due to "thermal-vibration synergy" were selected as anchor points and positive samples, while samples exhibiting water tree growth due to "humid heat dominance" were selected as negative samples. The squared distance between the anchor point and the positive sample was calculated to be 0.2, and the squared distance between the anchor point and the negative sample was 0.8. Substituting these values into the formula yielded a loss value of 0, indicating good feature discrimination. If the distances were 0.6 and 0.7 respectively, the loss value would be 0.4, triggering a parameter update. After 500 iterations of training, the distance within the same sample cluster decreased to less than 0.1, while the distance between dissimilar samples increased to more than 1.5, significantly improving the mechanistic discrimination ability of the feature tensor.
[0133] S3.4: Global average pooling is used to perform dimensionality reduction and compression on the optimized stress coupling feature tensor to eliminate redundant spatial dimensional information and generate a compressed stress state feature vector with uniform dimension.
[0134] The optimized stress coupling feature tensor with the ability to discriminate micro-degradation mechanisms is received. This tensor retains the spatial distribution information of high-dimensional nonlinear stress interaction but has dimensional redundancy.
[0135] A global average pooling operation is performed on the optimized stress coupling feature tensor, traversing all channel data points of the tensor in the spatial dimension.
[0136] For each feature channel, the arithmetic mean of all spatial values within that channel is calculated to aggregate local fluctuation information and extract global statistical features. The pooling output value of the k-th channel is calculated using the following formula:
[0137]
[0138] in, This is the pooled output value of the k-th channel. and These represent the spatial height and width dimensions of the feature tensor, respectively. Let be the feature value of the k-th channel in the i-th row and j-th column.
[0139] This calculation process eliminates the sensitivity of the feature map to specific spatial locations, retaining only scalar indices reflecting the overall stress coupling strength. The pooling outputs of all channels are arranged in channel index order to construct a compressed stress state feature vector with a unified dimension.
[0140] The L2 norm normalization process is performed on the compressed stress state feature vector of the unified dimension to eliminate the dimensional differences between different batches of data and ensure the input stability of the subsequent linear projection layer.
[0141] By using global average pooling and normalization, the high-dimensional redundant feature tensor from the previous step is transformed into a compressed stress state feature vector with a unified dimension and physical consistency, thus achieving the expected technical effect of data dimensionality reduction and information condensation.
[0142] S3.5: Perform a normalization mapping operation on the compressed stress state feature vector through linear projection to output a final stress path embedding vector that is strongly correlated with the micro-degradation mechanism and has a fixed dimension.
[0143] like Figure 3 As shown, step S4: Divide the physical region nodes according to the cable cross-sectional topology and initialize the intrinsic material parameters. At the same time, use the attention module to dynamically calculate the stress transmission weight based on the stress path embedding vector to construct a cable degradation trajectory graph neural network that includes spatial adjacency relationships and energy migration intensity.
[0144] The cable cross-sectional topology is a spatial structure model established based on the radially layered geometry of the cable, consisting of a set of physical region nodes and the adjacency relationships between nodes. The set of physical region nodes is obtained by spatially discretizing the cable insulation system, including conductor shielding layer nodes, main insulation layer nodes, and outer semiconducting layer nodes. The adjacency relationships between nodes are determined based on their direct adjacency in the radial direction of the cable cross-section, i.e., the shared interface relationship between adjacent sub-regions within the same material layer or between adjacent material layers. This topology provides the physical boundary conditions for the subsequent construction of a graph neural network, including node spatial location constraints and damage propagation.
[0145] Specifically, this includes: S4.1: Spatial discretization of the insulation system based on the cable cross-sectional topology, dividing the conductor shielding layer, main insulation layer and outer semiconducting layer into several independent physical region nodes, and assigning a unique node identifier to each physical region node to construct an initial node set.
[0146] Read the cable cross-sectional geometric design drawings and insulation material distribution parameters to determine the radial thickness and concentric circle topology of the conductor shielding layer, main insulation layer and outer semiconductive layer. The conductor shielding layer, main insulation layer and outer semiconductive layer together constitute the cable insulation system.
[0147] Based on the distribution law of electric field intensity gradient and the characteristics of heat conduction path, the continuous medium insulation system is divided into N discrete physical region nodes along the radial direction, where N is an integer greater than or equal to 3.
[0148] The interface between the conductor shielding layer and the main insulation layer, the key potential points inside the main insulation layer, and the interface between the main insulation layer and the outer semiconducting layer are divided into grids to generate structured grid data containing the spatial coordinates of the nodes.
[0149] Assign a globally unique node identifier (ID) to each partitioned physical region node, establish a mapping index table between node identifiers and physical spatial locations, and construct an initial node set V={v1, v2, ..., v N Based on the direct adjacency relationship of each physical region node in the radial topology of the cable cross section, i.e. the relationship of adjacent sub-regions within the same material layer or the shared interface between adjacent material layers, a spatial adjacency matrix is generated.
[0150] By discretizing the continuous insulating medium from the previous step, a set of computational nodes with independent attribute definitions is transformed, thus establishing the spatial topological foundation for the cable micro-damage evolution model.
[0151] For example, for a cross-linked polyethylene cable with a rated voltage of 110kV, the thickness of its main insulation layer is 16mm. The main insulation layer is radially divided into eight sub-region nodes with a thickness of 2mm. Including the conductor shielding layer nodes and the outer semiconducting layer nodes, a total of 10 physical region nodes are constructed. Node v1 corresponds to the conductor shielding layer, and nodes v2 to v9 correspond to the eight sub-layers of the main insulation layer from the inside out. 10 Corresponding to the outer semiconductive layer. Assign an ID to each node: ID(v1)=1, ID(v2)=2, ..., ID(v... 10 =10. Establish a mapping table {(1, r=0-5mm), (2, r=5-7mm), ..., (10, r=21-23mm)}. Based on the direct radial adjacency of each node (e.g., v1 is adjacent to v2, v2 is adjacent to v1 and v3, v9 is adjacent to v8 and v...), ...10 Adjacent, v 10 (Adjacent to v9), generating a spatial adjacency matrix. In this spatial adjacency matrix, the element corresponding to directly adjacent node pairs has a value of 1, and the element corresponding to non-adjacent node pairs has a value of 0. This serves as the physical boundary constraint for damage propagation in the subsequent construction of the graph neural network.
[0152] S4.2: Obtain the intrinsic material parameter values corresponding to each physical region node from the cable material intrinsic parameter database, and assign the intrinsic material parameter values as initial node feature vectors to the corresponding physical region nodes to complete the node attribute initialization. The intrinsic material parameter values include one or more combinations of dielectric constant, glass transition temperature, and crystallinity values.
[0153] The intrinsic parameter database for cable materials is a pre-stored collection of material physical property data, used to provide intrinsic parameter values for each physical region node. This database includes at least three parameter dimensions: dielectric constant, glass transition temperature, and crystallinity. Each parameter value is formed based on the factory-issued physical properties of the cable insulation material and offline test measurement data, and is structured and stored using material type and temperature range as a joint index. This ensures that nodes at different spatial locations can quickly retrieve the corresponding intrinsic parameter values based on their material layer type and current temperature state.
[0154] The intrinsic parameter database of cable materials is accessed to retrieve material property records that match the node identifiers of the physical regions described in S4.1. The dielectric constant values of the corresponding regions of the conductor shielding layer, main insulation layer, and outer semiconducting layer are extracted. These dielectric constant values characterize the polarization response capability of the material under the action of an electric field.
[0155] The glass transition temperature values of materials in each region are obtained. These glass transition temperature values define the critical thermodynamic threshold for the transition of polymer chain segments from the glassy state to the elastic state.
[0156] The crystallinity values of materials in each region are read to reflect the degree of order in the arrangement of polymer chains and their hindering effect on mechanical stress and chemical diffusion.
[0157] The extracted dielectric constant, glass transition temperature, and crystallinity values are arranged in a preset feature dimension order to construct an initial node feature vector, where each component corresponds to a specific intrinsic physical quantity of the material.
[0158] Z-score normalization is performed on the initial node feature vectors to eliminate the order-of-magnitude differences between different physical dimensions, ensuring gradient convergence stability during the subsequent training of the cable degradation trajectory neural network. The normalized feature vectors are then assigned to the corresponding physical region nodes, completing the initial configuration of node attributes.
[0159] Through the above data processing flow, discrete material physical properties are transformed into structured node feature vectors, realizing the digital mapping of cable micromaterial properties in the graph topology, and providing physically meaningful baseline state data for the subsequent calculation of dynamic stress transmission weights.
[0160] S4.3: Utilize the attention module to receive the stress path embedding vector sequence, take the initial node feature vector of each physical region node as the key vector, calculate the correlation coefficient matrix of different physical region nodes under the current working condition, and transform the correlation coefficient matrix into dynamic stress transmission weights that characterize the energy migration intensity.
[0161] The attention module adopts a standard multi-head attention architecture, which consists of a query linear transformation layer, a key linear transformation layer, a value linear transformation layer, a scaling dot product attention calculation unit, and an output linear transformation layer connected in sequence.
[0162] The system receives the stress path embedding vector sequence as query input, and obtains a query vector after mapping through the query linear transformation layer. It also receives the intrinsic material parameters of each physical region node in the cable cross-section topology as key-value pair input, and obtains a key vector and a value vector after mapping through the key linear transformation layer and the value linear transformation layer, respectively. The query vector and the key vector are input to the scaled dot product attention calculation unit, where the dot product similarity is calculated and normalized using the Softmax activation function to obtain the attention weights. These weighted sums are then calculated with the value vectors, and after mapping through the output linear transformation layer, a correlation coefficient matrix is output. Each element in the correlation coefficient matrix represents the dynamic stress transmission weight, calculated using the following formula:
[0163]
[0164] in, Indicates from node To the node Dynamic stress transmission weight, For stress path embedding vectors, For nodes With nodes The feature interaction vector.
[0165] After the above attention calculation, the generated correlation coefficient matrix is the dynamic stress transmission weight matrix.
[0166] By combining static material properties with dynamic stress semantics through the above attention calculation processing method, dynamic weight data reflecting the migration path of stress energy inside the cable under real-time operating conditions is transformed into dynamic weight data, realizing refined modeling of unsteady multi-stress coupling effects and significantly improving the graph neural network's ability to capture complex aging mechanisms.
[0167] S4.4: Based on the spatial adjacency relationship of the initial node set and the dynamic stress transmission weight, construct a directed weighted edge connection, and fuse the static connection representing the spatial adjacency relationship with the dynamic stress transmission weight representing the energy migration intensity to generate a cable degradation trajectory graph neural network.
[0168] Obtain the spatial adjacency matrix and dynamic stress transmission weight matrix of the initial node set as the underlying data foundation for constructing the graph structure.
[0169] Based on the physical contact relationship of the cable insulation layer, a static adjacency matrix is defined, where if node i and node j are directly adjacent in space, the element value is 1, otherwise it is 0, thus establishing the physical boundary constraints for damage propagation.
[0170] The dynamic stress transmission weight matrix output by the attention module is invoked. This dynamic stress transmission weight matrix characterizes the energy coupling strength and nonlinear interaction coefficient between nodes in different physical regions under the action of the current specific stress path embedding vector.
[0171] The static topology and dynamic weights are fused, mapping the static adjacency relationships to a dynamic weighted adjacency matrix using either the Hadamard product or a weighted summation. The final edge weights are calculated using the following formula:
[0172]
[0173] in, This is the fused directed weighted adjacency matrix. It is a static adjacency matrix. This is the dynamic stress transmission weight matrix. The Sigmoid activation function is used to normalize the weight range. This indicates element-wise multiplication.
[0174] The generated fusion matrix is sparsified by removing weak connections below a preset threshold to reduce the computational complexity of subsequent graph neural network inference and suppress noise interference.
[0175] The processed node feature vectors and the fused directed weighted adjacency matrix are encapsulated into a standard graph data structure object to form a cable degradation trajectory graph neural network.
[0176] Specifically, the cable degradation trajectory graph neural network is a temporal graph neural network structure, consisting of a second input layer, a graph convolutional layer, a temporal evolution layer, and a second output layer connected sequentially. The second input layer receives the stress path embedding vector sequence within a sliding window, and generates a dynamic stress feature tensor after temporal recombination. The graph convolutional layer performs spatial aggregation of the features of each physical region node based on the directed weighted adjacency matrix to generate a hidden state representation of the node. The temporal evolution layer uses gated recurrent units to iteratively update the hidden state of the node in multiple rounds to generate deep evolution features of the node. The second output layer consists of two fully connected hidden layers, mapping the deep evolution features of the node to a regional damage evolution probability distribution.
[0177] Through the above fusion processing method, the static topology and dynamic weights of the previous step are transformed into graph structure data with physical constraints and working condition sensitivity, thus achieving the expected technical effect of accurate modeling of damage propagation path under unsteady multi-stress coupling.
[0178] Step S5: Input the stress path embedding vector sequence within the sliding window into the cable degradation trajectory graph neural network for end-to-end inference to obtain the regional damage evolution probability distribution. The regional damage evolution probability distribution is used to characterize the damage evolution state of each physical region node in the next time step. Specifically, it includes:
[0179] S5.1: The stress path embedding vector sequence within the sliding window is obtained through the input layer of the cable degradation trajectory neural network as the initial input data. The stress path embedding vector sequence is then subjected to temporal reordering processing based on the time step alignment mechanism to generate a dynamic stress feature tensor with continuous time dependence, ensuring the continuity and integrity of the input data in the time dimension.
[0180] Based on the preset time step interval parameter, an index alignment check is performed on the stress path embedding vector sequence to remove missing or duplicate timestamp segments caused by sensor packet loss, ensuring the continuity of the time dimension.
[0181] Linear interpolation is used to fill the small time gaps between adjacent time steps, generating intermediate state embedding vectors with uniform temporal distribution, thus eliminating the risk of spectral aliasing caused by sampling non-uniformity.
[0182] The discrete embedded vectors, after alignment and interpolation, are stacked along the third dimension in chronological order to construct a three-dimensional dynamic stress feature tensor, where the first dimension is the batch size, the second dimension is the time step, and the third dimension is the feature dimension.
[0183] To enable the model to perceive the temporal dependence of stress evolution, the positions of each time step of the three-dimensional dynamic stress feature tensor are encoded: the position index of each time step in the sequence is extracted, the even-numbered dimensions are encoded using a sine function and the odd-numbered dimensions are encoded using a cosine function, and a position encoding vector corresponding to each time step position is generated. The position encoding vector is then superimposed on the feature dimension of the corresponding time step to obtain the dynamic stress feature tensor containing temporal position information.
[0184] By using the above temporal reorganization processing method, the discrete stress path embedding vector of the previous step is transformed into a dynamic stress feature tensor with continuous time dependence, thereby realizing the continuity and integrity of the input data in the time dimension and providing high-quality temporal state input for subsequent graph convolution propagation.
[0185] S5.2: The dynamic stress feature tensor is received by the graph convolutional layer of the cable degradation trajectory graph neural network and the cable cross-section topology is loaded. Feature aggregation operation is performed on the physical region nodes containing the intrinsic parameter properties of the material to generate a node hidden layer state representation that integrates spatial adjacency information and the current stress field distribution, thereby realizing the initial mapping of stress energy on the cable cross-section.
[0186] It receives a dynamic stress characteristic tensor and a predefined cable cross-section topology, which includes a set of nodes and an initialized vector of intrinsic material properties.
[0187] Extract the node feature matrix at the current time step. This matrix is formed by concatenating the node hidden state representation from the previous time step with the stress path embedding vector at the current time step.
[0188] Based on the physical adjacency relationships of the cable insulation layer, starting from the spatial adjacency matrix generated in S4.1, the features of neighboring nodes of each physical region node are weighted and aggregated, and the feature information of the neighboring nodes is passed to the central node. Then, a linear transformation is performed on the aggregated neighborhood features to extract local spatial context information. Finally, the transformed neighborhood aggregated features are added to the central node's own features and nonlinear activation is performed to update the hidden state representation of the current node. The following formula is used to calculate the node feature aggregation and update of the graph convolutional layer:
[0189]
[0190] in, This is the node feature matrix of the current layer. It is an adjacency matrix. For degree matrix, For learnable weight matrix, It is a non-linear activation function. This is the updated node feature matrix.
[0191] For the node features processed by the graph convolutional layer, a residual connection operation is performed, which adds the input features and output features of the current graph convolutional layer element by element to preserve the original material intrinsic information and alleviate the gradient vanishing problem. Subsequently, batch normalization is performed on the residual connection features to accelerate model convergence and stabilize the training process.
[0192] After the above multi-layer graph convolution aggregation and residual connection processing, the resulting node feature matrix is the hidden layer state representation of the node.
[0193] Through the above processing method, the dynamic stress characteristics of the previous step are transformed into a node hidden state representation that integrates spatial adjacency information and the current stress field distribution, thereby realizing the initial mapping of stress energy on the cable cross section.
[0194] For example, for a 10kV cross-linked polyethylene cable, the cross-section is divided into four nodes: the conductor shielding layer, the inner layer of the main insulation layer, the outer layer of the main insulation layer, and the outer semiconducting layer. The graph convolution layer is set to two layers, with a hidden layer feature dimension of 64. The input node feature dimension is 128 (64-dimensional material parameters + 64-dimensional stress embedding). The adjacency matrix A is constructed based on physical contact relationships, and the diagonal elements of the degree matrix D represent the number of neighbors for each node. After the first layer of graph convolution, the inner layer nodes of the main insulation layer aggregate the hot spot stress characteristics of the conductor shielding layer, and the component representing thermal damage in its hidden layer state vector increases from 0.3 to 0.45. The second layer of convolution further integrates the outer layer heat dissipation conditions, and the final output node hidden layer state accurately reflects the non-uniform influence of the radial temperature gradient on insulation aging. This node hidden layer state is the node hidden layer state representation that integrates spatial adjacency information and the current stress field distribution.
[0195] S5.3: Through the temporal evolution layer of the cable degradation trajectory graph neural network, based on the hidden state representation of the nodes and combined with the stress transmission weights generated by the attention module, the edge dynamic update operation is performed to reconstruct the connection strength of the cable degradation trajectory graph neural network, and a dynamic weighted adjacency matrix reflecting the energy migration path under the non-steady-state coupling effect is generated, which accurately characterizes the transmission direction and intensity of stress inside the insulation layer under different working conditions.
[0196] The node hidden state representation matrix output by S5.2 and the dynamic stress transmission weights generated by S4.3 are used as input data for the cable degradation trajectory neural network.
[0197] Extract the stress path embedding vector of the current time step as the query vector, and at the same time, take the hidden state vector of each node in the node hidden state representation matrix output by S5.2 as the key vector, input it into the attention mechanism module constructed in S4.3, calculate the dot product similarity between the query vector and each of the key vectors, and generate an unnormalized attention score matrix.
[0198] The attention score matrix is subjected to Softmax normalization to eliminate dimensional differences and transform it into a dynamic stress transmission coefficient in probability distribution form, which characterizes the instantaneous energy coupling strength between nodes in different physical regions. The specific formula is as follows:
[0199]
[0200] in, For nodes To the node The dynamic stress transmission coefficient, For unnormalized attention scores, For nodes The set of adjacent nodes.
[0201] The dynamic stress transmission coefficient is multiplied by the Hadamard product with the static adjacency matrix of the cable cross-section topology to generate a dynamic weighted adjacency matrix that reflects the unsteady coupling effect.
[0202] The dynamic weighted adjacency matrix is used to update the edge connection strength of the cable degradation trajectory neural network in real time, blocking the propagation of invalid information in low-relevance areas and strengthening the feature transmission of high-stress coupling paths.
[0203] By constructing a dynamic weighted adjacency matrix, the static topology from the previous step is transformed into a dynamic graph structure that adapts to real-time operating conditions, thereby enabling accurate characterization of energy migration paths under unsteady multi-stress conditions.
[0204] S5.4: Through the temporal evolution layer of the cable degradation trajectory graph neural network, the message passing mechanism of the graph neural network is driven by the dynamic weighted adjacency matrix to perform multiple rounds of iterative update processing on the hidden state representation of each physical region node, generating deep evolution features of nodes that characterize the cumulative effect of micro-damage, and simulating the spatiotemporal diffusion process of micro-defects such as electrical tree germination or interface debonding under complex stress paths.
[0205] It receives the dynamically weighted adjacency matrix generated by S5.3 and the hidden state representation of the nodes output by S5.2, which serve as the input data source for the message passing mechanism of the graph neural network.
[0206] Based on the edge weight coefficients in the dynamic weighted adjacency matrix, a weighted aggregation operation is performed on the hidden features of adjacent physical region nodes to obtain aggregated neighborhood features. The aggregated neighborhood features and the hidden state representation of the current node are input into a gated recurrent unit. This gated recurrent unit, as a temporal evolution module, is used to fuse the node state from the previous time step with the current neighborhood aggregation information. The gated recurrent unit determines the fusion ratio between the information from the previous time step and the current neighborhood information, and updates the hidden evolution vector of the node at the current time step.
[0207] After each iteration, the input and output features of the current iteration are added element by element. At the same time, the initial material intrinsic parameters assigned in S4.2 are added as bias terms to the deep evolution features to alleviate the gradient vanishing problem in deep network training and preserve the original material intrinsic information. In addition, during the training phase, some node connections are randomly masked with a preset probability to enhance the model's robustness to sensor noise and local data loss.
[0208] Repeating the above aggregation and update process allows node features to gradually absorb damage diffusion information from multi-hop neighborhoods, simulating the spatiotemporal path of the electrical tree expanding from its germination point into the depths of the insulation layer. After the above multiple rounds of iterative update processing, the resulting node hidden layer evolution vector is the node deep layer evolution feature characterizing the microscopic damage accumulation effect.
[0209] S5.5: Through the output layer of the cable degradation trajectory neural network, the deep evolution features of the nodes are extracted and nonlinear transformation operations are performed through the fully connected mapping layer. The high-dimensional evolution features are projected onto the predefined damage mode space to generate a regional damage evolution probability distribution containing information on the interface debonding risk level and the local conductivity growth rate, thus completing the end-to-end inference closed loop from stress input to microscopic damage state output.
[0210] Extract the node deep evolution feature tensor output by S5.4. This tensor contains the microscopic damage accumulation state information of each physical region node after multiple rounds of graph message passing.
[0211] The output layer of the cable degradation trajectory neural network consists of a fourth fully connected hidden layer and a fifth fully connected hidden layer connected sequentially: the fourth fully connected hidden layer receives the deep evolution feature tensor of the node, performs linear weighting and bias superposition operation, and after element-wise nonlinear mapping by the ReLU activation function, outputs the fourth hidden layer feature vector to the fifth fully connected hidden layer; the fifth fully connected hidden layer receives the fourth hidden layer feature vector, performs linear weighting and bias superposition operation, and normalizes the output by the Softmax activation function to ensure that the sum of the probability distributions of each damage mode is 1, and outputs the fifth hidden layer feature vector.
[0212] For the two key damage indicators, interface debonding risk level and local conductivity growth rate, independent output branches are set up respectively. The feature vector of the fifth hidden layer is input into the two independent output branches: the first output branch maps the feature vector of the fifth hidden layer to probability values of three discrete levels: low risk, medium risk, and high risk, for the interface debonding risk level; the second output branch maps the feature vector of the fifth hidden layer to the predicted growth rate relative to the initial intrinsic conductivity for the local conductivity growth rate.
[0213] During the training phase, the mean squared error loss function is used to constrain the deviation between the predicted conductivity value and the actual aging test data, and the cross-entropy loss function is used to constrain the accuracy of debonding risk classification. The weight parameters and bias terms of the fully connected layer are jointly optimized through the backpropagation algorithm, so that the model can accurately capture the nonlinear relationship between stress path semantics and micro-damage.
[0214] After the above-mentioned fully connected mapping and nonlinear transformation of the output layer, the output information, which includes the probability of interface debonding risk level and the local conductivity growth rate, is the regional damage evolution probability distribution.
[0215] Step S6: Based on the regional damage evolution probability distribution, perform a weighted aggregation operation to fuse local anomaly propagation information into a global system-level functional degradation index and generate a predicted cable remaining life value including a heat map of high-risk areas. Specifically, this includes:
[0216] S6.1: Obtain the regional damage evolution probability distribution output from the previous main step, and perform local anomaly propagation calculation on the regional damage evolution probability distribution based on the predefined node spatial adjacency relationship in the cable cross-section topology, generating a dynamic stress transmission weight matrix characterizing the migration intensity of damage between adjacent physical region nodes.
[0217] The system receives the regional-level damage evolution probability distribution vector output from step S5. This vector contains the probability values of the microscopic damage states of each physical region node in the cable cross-section at the current time step. Based on a predefined cable cross-section topology diagram, the spatial adjacency matrix between nodes is extracted to clarify the physical connection relationships between nodes in the conductor shielding layer, main insulation layer, and outer semiconducting layer.
[0218] Based on the spatial adjacency matrix, the gradient difference of damage probability between adjacent nodes is calculated to quantify the diffusion trend of local damage anomalies in the spatial domain. A damage migration activation function is introduced to map the gradient difference into a non-negative migration intensity coefficient, characterizing the probability of damage propagating from high-risk nodes to low-risk nodes.
[0219] Combining the dynamic stress transmission weights generated in step S5.3, the migration intensity coefficients are weighted and corrected to eliminate invalid propagation paths in non-stress-dominant directions. Through normalization, the sum of the weights of the out-degrees of each node is ensured to be 1, thus constructing a dynamic stress transmission weight matrix that reflects the spatiotemporal evolution of damage.
[0220] Through the above chain-like derivation process, discrete node damage probabilities are transformed into a dynamic stress transmission weight matrix with spatial correlation, realizing the construction of intermediate state data for the transformation of local damage anomalies into global degradation indicators, and providing accurate propagation path basis for subsequent weighted aggregation.
[0221] S6.2: Receive the dynamic stress transmission weight matrix and the initialized material intrinsic parameter properties, and use the graph attention mechanism to perform nonlinear transformation processing on the dynamic stress transmission weight matrix to construct a set of regional damage evolution weighted coefficients that reflect the energy migration path under the current working condition.
[0222] Receive the dynamic stress transmission weight matrix generated in S6.1 and the material intrinsic parameter properties initialized in S4.2, and construct the graph attention mechanism calculation module.
[0223] The dynamic stress transmission weight matrix is dimension aligned by mapping it to an adjacency tensor format that matches the number of nodes in the physical region of the cable cross-section, ensuring data dimensionality consistency between the spatial topology and the damage migration intensity.
[0224] A single-layer feedforward neural network is used to perform nonlinear feature transformation on each element in the adjacency tensor. The LeakyReLU activation function is introduced to preserve the stress suppression effect in the negative value region, generating a preliminary attention coefficient matrix.
[0225] Based on the intrinsic parameters of the material, a node importance bias vector is constructed. Physical properties such as dielectric constant and glass transition temperature are converted into scalar bias terms through a linear projection layer and superimposed on the preliminary attention coefficient matrix to reflect the constraint effect of the material's inherent properties on damage evolution.
[0226] The attention coefficient matrix after superimposed bias is normalized row by row using the Softmax function to ensure that the sum of the weights of the outgoing edges of each physical region node is 1, thereby eliminating the difference in dimensions and constructing a set of regional damage evolution weighted coefficients in the form of a probability distribution.
[0227] Through the nonlinear transformation and normalization of the graph attention mechanism described above, the dynamic stress transmission weights of the previous step are transformed into a set of regional damage evolution weighting coefficients with physical interpretability. This enables the quantification of damage contribution from simple stress energy transfer intensity to the combined influence of material properties and working conditions, significantly improving the physical realism and prediction accuracy of global lifetime aggregation.
[0228] S6.3: Based on the set of weighted coefficients for regional damage evolution, perform a node-by-node weighted summation operation on the probability distribution of regional damage evolution to integrate the dispersed local damage states into a global system-level functional degradation index that characterizes the overall insulation performance degradation of the cable.
[0229] The set of regional damage evolution weighted coefficients output by S6.2 and the regional damage evolution probability distribution generated by S5.5 are received as input data sources for calculating the global degradation index.
[0230] Traverse each physical region node in the cable cross-section topology and extract the damage evolution probability vector and dynamic weighting coefficient scalar corresponding to the current node.
[0231] For a single physical region node, an element-wise multiplication operation is performed to multiply the local damage probability value of the node with its corresponding energy migration weight, generating the contribution component of the node to the overall degradation of the system.
[0232] Based on the physical mechanism of cable insulation failure, an intrinsic material sensitivity factor is introduced as a correction term to nonlinearly weight the contribution components, so as to reflect the differences in stress response of different material layers (such as the main insulation layer and the semiconductive layer).
[0233] The global degradation contribution of the i-th node is calculated using the following formula:
[0234]
[0235] in, Let i be the global degradation contribution value of the i-th node. This is a weighted coefficient for regional damage evolution. This represents the probability of regional damage evolution. This is the intrinsic sensitivity factor of the material.
[0236] The global degradation contribution values of all physical region nodes are summed to eliminate local biases caused by spatial discretization and obtain scalar values that characterize the overall insulation performance degradation of the cable.
[0237] The accumulated results are normalized and mapped to the [0,1] interval to generate a standardized global system-level functional degradation index, which directly reflects the degree of deviation of the cable's current health status from its brand-new state.
[0238] By using a node-by-node weighted summation and normalization process, the dispersed local damage states are transformed into a unified global system-level functional degradation index, achieving a precise mapping from micro-regional damage to macro-system lifetime state, and providing a quantitative benchmark for remaining lifetime inversion.
[0239] S6.4: Based on the mapping relationship between the global system-level functional degradation index and the preset failure threshold, and combined with the time step cumulative effect, perform the remaining lifetime inversion calculation and output a quantified cable remaining lifetime prediction value.
[0240] Receive the global system-level functional degradation index sequence output by S6.3, which characterizes the cumulative degradation trajectory of cable insulation performance over time.
[0241] Extract the current global system-level functional degradation index value and use it as the baseline state quantity for remaining lifetime inversion calculation.
[0242] Call the predefined failure threshold parameter library to obtain the critical failure threshold for the corresponding cable model and operating condition. This threshold is determined based on the statistical analysis of insulation breakdown or dielectric loss mutation points in historical accelerated aging tests.
[0243] A remaining lifetime time mapping function is constructed, and a nonlinear regression model between the global system-level functional degradation index and the time step is established by fitting linear interpolation or exponential decay.
[0244] The predicted remaining lifetime time span is calculated using the following formula:
[0245]
[0246] in, For the predicted remaining lifespan, The preset critical failure threshold, This is the current global system-level functional degradation indicator. The average degradation rate is calculated based on the most recent N time steps. The physical duration of a single time step.
[0247] The calculated remaining lifetime time span is corrected for confidence intervals, and combined with the standard deviation of historical prediction error distribution, a quantitative lifetime prediction interval including upper and lower bounds is generated.
[0248] The quantified life expectancy prediction range is converted into a standardized remaining life expectancy prediction value, and the current health status level is labeled.
[0249] By using the above-mentioned remaining life inversion calculation method, the global system-level functional degradation index of the previous step is transformed into a quantified cable remaining life prediction value, realizing an accurate mapping from micro-damage state to macro-operation and maintenance decision indicators, and significantly improving the interpretability and engineering applicability of the life prediction results.
[0250] S6.5: Utilizing the location information of high-risk nodes in the regional damage evolution probability distribution and the numerical level of the global system-level functional degradation index, a visualized heat map of high-risk areas is generated through spatial interpolation, and the final encapsulation and output of the predicted cable remaining life value is completed.
[0251] The damage risk values and corresponding spatial coordinate indices of each physical region node in the regional damage evolution probability distribution are obtained, and a discretized node damage state matrix is constructed. Based on the node adjacency relationships in the cable cross-sectional topology, the boundary conditions and mesh density of the region to be interpolated are determined, and a mapping model from discrete nodes to continuous medium is established.
[0252] Radial basis function interpolation is used, with the spatial coordinates of each physical region node as input variables and the damage risk value as the output response, to calculate the damage estimate at non-node locations. A Gaussian kernel function is used to define the spatial weight decay characteristic, ensuring that the impact of neighboring high-risk nodes on the surrounding area decreases exponentially, consistent with the physical law of local damage diffusion.
[0253] The interpolated continuous damage field data is mapped to a two-dimensional polar or rectangular coordinate system and color-coded according to a preset risk level threshold range to generate a visualized heat map of high-risk areas. This heat map intuitively displays the specific location and severity of damage concentrations within the insulation layer, assisting maintenance personnel in locating potential fault points.
[0254] The system reads global system-level functional degradation index values, combines them with the current time step and historical cumulative degradation, and performs the final inversion calculation of remaining lifetime. The calculated remaining lifetime prediction value, high-risk area heatmap data, and confidence intervals are packaged into a standard JSON format data packet and transmitted as the final output to the monitoring terminal.
[0255] Through spatial interpolation and visualization encoding, the discrete probability distribution of the previous step is transformed into a continuous spatial damage heatmap and a quantified lifetime prediction value, thereby realizing the macroscopic visualization of the microscopic damage state and accurate lifetime extrapolation.
[0256] Step S7: Update the stress path embedding vector based on the real-time online monitored stress flow data and trigger the time-step state iteration of the cable degradation trajectory neural network, outputting the corrected cable remaining life prediction value to complete the life extrapolation closed loop under dynamic multi-stress paths. Specifically, this includes:
[0257] S7.1: Obtain the real-time online monitoring temperature fluctuation sequence, mechanical vibration spectrum sequence, and humidity time series data as the raw stress flow input, and use a sliding time window to extract continuous segments from the raw stress flow input to generate a real-time stress data window sequence to be updated.
[0258] Specifically, time-series data segments of temperature, vibration spectrum, and humidity at the current moment are read from the online monitoring data stream; the three-channel data are synchronously truncated with a preset sliding window length and step size, and the window length is consistent with the setting in S2.1; the truncated data segments are subjected to timestamp alignment verification to ensure that the three-channel data strictly correspond in the time dimension; the verified data segments are standardized according to the normalization parameters in S1.2 to generate a real-time stress data window sequence to be updated.
[0259] S7.2: Based on the real-time stress data window sequence, call the pre-trained stress path encoder to perform feature mapping operation, and convert the real-time stress data window sequence into a real-time updated stress path embedding vector that is strongly correlated with the current micro-degradation mechanism.
[0260] Receive the real-time stress data window sequence to be updated generated by S7.1. This sequence contains time-aligned multidimensional sensing data of temperature, vibration spectrum and humidity.
[0261] The feature extraction subnet in the pre-trained stress path encoder is invoked to perform tensor quantization reconstruction on the real-time stress data window sequence, mapping the multi-source heterogeneous time series data into a high-dimensional initial feature tensor.
[0262] By utilizing the multilayer perceptron structure inside the encoder to perform a nonlinear transformation on the high-dimensional initial feature tensor, high-order coupling features hidden in dynamic stress fluctuations are extracted, and intermediate layer stress coupling representations are generated.
[0263] Based on the constraint of the contrastive loss function, the similarity matching calculation between the stress coupling characterization of the intermediate layer and the failure mode prototype vector in the knowledge base of cable material aging mechanism is performed to strengthen the semantic association between feature vector and micro-degradation mechanism.
[0264] Global average pooling is used to compress the spatial dimension of the enhanced feature tensor, eliminating redundant noise and integrating local stress fluctuation information to generate a fixed-dimensional compressed stress state vector.
[0265] The compressive stress state vector is normalized by a linear projection layer, and the output is a real-time updated stress path embedding vector that is strongly correlated with the current microscopic degradation mechanism and has uniform dimensions.
[0266] Through the aforementioned feature mapping and dimensionality reduction processing, the real-time stress data window sequence from the previous step is transformed into a real-time updated stress path embedding vector, achieving a precise mapping from dynamic working conditions to the microscopic degradation semantic space, and providing a physically interpretable input benchmark for subsequent graph neural network weight updates.
[0267] S7.3: The attention module receives the real-time updated stress path embedding vector and performs dynamic weight calculation processing. Based on the calculation results, the edge connection stress transmission weights in the cable degradation trajectory graph neural network are updated to generate a dynamic topology graph structure with instantaneous energy transfer intensity.
[0268] The system receives the real-time updated stress path embedding vector output from S7.2 and inputs it as the query vector into the linear projection layer of the attention module. A fully connected traversal is performed on predefined physical region node pairs in the cable degradation trajectory graph neural network to obtain the hidden state representation of each node pair at the previous time step. The similarity score between the query vector and the key vector is calculated using a dot product attention mechanism to generate the original attention score matrix characterizing the stress coupling strength between nodes under the current operating condition.
[0269] A scaling factor is applied to the original attention score matrix to stabilize gradient propagation and prevent the softmax function from entering the saturation region, which would lead to gradient vanishing. The scaled scores are then normalized using the softmax activation function, converting the non-standardized similarity scores into dynamic stress transmission weight coefficients in probability distribution form. The specific formula is as follows:
[0270]
[0271] in, Let be the dynamic stress transmission weight from node i to node j. for The stress path embedding vector at time step, Let be the key vector of the node pair. For feature dimension, Let i be the set of adjacent nodes of node i.
[0272] The generated dynamic stress transmission weight coefficients are mapped to the edge attribute tensor of the graph neural network, replacing the original static adjacency matrix weight values. Based on the updated edge attribute tensor, the topological connection structure of the cable degradation trajectory graph is reconstructed, forming a dynamic topological graph structure with instantaneous energy transfer intensity.
[0273] By using dynamic weight calculation and topology reconstruction through attention mechanism, the static stress path embedding vector is transformed into dynamic graph structure data that reflects the coupling effect of real-time operating conditions, thereby achieving the expected technical effect of adaptive updating of stress energy migration path inside cable insulation layer.
[0274] S7.4: Input the dynamic topology graph structure and the regional damage evolution probability distribution of the previous time step into the cable degradation trajectory graph neural network to perform a single-step forward inference operation, and output the instantaneous regional damage evolution probability distribution representing the damage evolution state of each physical region node at the current moment.
[0275] The system receives a dynamic topology graph structure with instantaneous energy migration intensity generated by S7.3. This structure includes an updated node feature matrix and a dynamically weighted adjacency matrix. It reads the region-level damage evolution probability distribution vector output at the previous time step t-1 and uses it as the initial hidden state input for the graph neural network message passing mechanism. It performs neighbor node feature aggregation operations using graph convolutional layers, and weights and fuses the damage states of adjacent physical regions according to the edge weights defined by the dynamically weighted adjacency matrix. The hidden state representation of the i-th node at time t is calculated using the following formula:
[0276]
[0277] in, For nodes exist The hidden state at all times For learnable weight matrix, For node degree, For dynamic stress transmission weight, For nodes The neighborhood group, It is a non-linear activation function.
[0278] The spatiotemporal diffusion patterns of microscopic damage under unsteady coupled stress are captured by iteratively updating node states through a multi-layer graph attention network. The updated deep evolution features of the nodes are input into a fully connected mapping layer, and a nonlinear projection transformation is performed to map them to a predefined damage mode space. The output is an instantaneous regional-level damage evolution probability distribution that includes the interface debonding risk level, the local conductivity growth rate, and the microcrack propagation probability.
[0279] By using the message passing and state update mechanism of graph neural networks, the dynamic topology and historical damage state are transformed into a refined damage probability distribution of each physical region at the current moment, thus realizing the real-time and accurate characterization of the microscopic degradation process of cable insulation.
[0280] S7.5: Perform a global weighted aggregation operation based on the instantaneous regional damage evolution probability distribution to fuse local anomaly propagation information into a corrected global system-level functional degradation index, and generate a corrected cable remaining life prediction value containing a heat map of high-risk areas to complete the simulation closed loop.
[0281] Specifically, the instantaneous regional-level damage evolution probability distribution output in S7.4 is obtained, and the intrinsic material parameters initialized in S4.2 are read. Based on the spatial adjacency relationship of nodes in the cable cross-section topology, a weighted aggregation operation is performed on the instantaneous regional-level damage evolution probability distribution: traversing each physical region node, multiplying the damage probability value of each node by the corresponding intrinsic material sensitivity factor to obtain the global degradation contribution value of each node; summing the global degradation contribution values of all nodes to obtain the corrected global system-level functional degradation index; comparing the corrected global system-level functional degradation index with a preset failure threshold, and calculating the corrected remaining lifetime prediction value in combination with the current time step; simultaneously, using the high-risk node location information in the instantaneous regional-level damage evolution probability distribution, a high-risk region heat map is generated through radial basis function interpolation; the corrected remaining lifetime prediction value and the high-risk region heat map are encapsulated and output as the corrected cable remaining lifetime prediction value.
[0282] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.
[0283] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” “third,” and similar terms used in this patent application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms “comprising” or “including” and similar terms mean that the elements or objects preceding “comprising” or “including” encompass the elements or objects listed following “comprising” or “including” and their equivalents, and do not exclude other elements or objects. The “multiple” mentioned in the embodiments of this application refers to two or more. A and / or B indicate three possibilities: A; B; and A and B.
[0284] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for predicting cable life based on multi-stress acceleration, characterized in that, include: S1: Acquire the stress flow input to be processed synchronously collected by the cable in the service environment; S2: Based on the cable material aging mechanism knowledge base, perform semantic slicing processing on the stress flow input to be processed, identify the working condition combination features in the continuous stress data and generate stress path semantic segments with unique path labels. S3: Perform supervised learning mapping operation on the stress path semantic segment using a pre-trained stress path encoder to convert the stress path semantic segment with a unique path label into a stress path embedding vector. S4: Divide the physical region nodes according to the cable cross-sectional topology and initialize the intrinsic material parameter properties. At the same time, use the attention module to dynamically calculate the stress transmission weight based on the stress path embedding vector to construct a cable degradation trajectory neural network. S5: The stress path embedding vector sequence within the sliding window is input into the cable degradation trajectory diagram neural network for end-to-end inference to obtain the regional damage evolution probability distribution; S6: Perform a weighted aggregation operation based on the regional damage evolution probability distribution to generate a predicted value of the cable's remaining life that includes a heat map of high-risk areas.
2. The cable life prediction method based on multi-stress acceleration according to claim 1, characterized in that, Also includes: S7: Update the stress path embedding vector based on the real-time online monitored stress flow data and trigger the time-step state iteration of the cable degradation trajectory neural network, outputting the corrected predicted value of the cable's remaining life.
3. The cable life prediction method based on multi-stress acceleration according to claim 1, characterized in that, Step S3 specifically includes: Based on the failure mode label set defined in the cable material aging mechanism knowledge base, the input stress path semantic segment with unique path label is subjected to one-hot encoding to generate an initial working condition feature vector. The initial working condition feature vector is subjected to nonlinear transformation processing using a stress path encoder to output the intermediate layer stress coupling feature tensor. Based on the stress path sample library labeled with actual failure modes in accelerated testing, a comparative loss function is constructed. Supervised learning optimization is performed on the intermediate layer stress coupling feature tensor to adjust the parameters of the stress path encoder and generate the optimized stress coupling feature tensor. The optimized stress coupling feature tensor is subjected to dimensionality reduction and compression to generate a compressed stress state feature vector. Perform a normalization mapping operation on the compressed stress state feature vector to output the final stress path embedding vector.
4. The cable life prediction method based on multi-stress acceleration according to claim 3, characterized in that, The stress path encoder is composed of a first input layer, a first fully connected hidden layer, a second fully connected hidden layer, a third fully connected hidden layer, and a first output layer connected in sequence: the dimension of the first input layer is the same as the dimension of the initial working condition feature vector, the first fully connected hidden layer, the second fully connected hidden layer, and the third fully connected hidden layer are connected in sequence, and the first output layer maps the output of the third fully connected hidden layer to the dimension of the intermediate layer stress coupling feature tensor. The step of performing nonlinear transformation processing on the initial working condition feature vector using a stress path encoder to output an intermediate layer stress coupling feature tensor includes: The first input layer receives the initial working condition feature vector and passes it to the first fully connected hidden layer; The first fully connected hidden layer, the second fully connected hidden layer, and the third fully connected hidden layer receive the output of the previous layer in sequence, perform linear weighted and bias superposition operations, and pass them to the next layer after element-wise nonlinear mapping by the activation function, until the three fully connected hidden layers have finished processing. The first output layer receives the feature vector output by the third fully connected hidden layer, performs linear weighting and bias superposition operations, and outputs the intermediate layer stress coupling feature tensor.
5. The cable life prediction method based on multi-stress acceleration according to claim 1, characterized in that, Step S4 specifically includes: Based on the cable cross-sectional topology, the insulation system is spatially discretized, dividing the conductor shielding layer, main insulation layer and outer semiconducting layer into several independent physical region nodes, and assigning a unique node identifier to each physical region node to construct an initial node set. The intrinsic material parameter values corresponding to each physical region node are obtained from the intrinsic parameter database of cable material, and the intrinsic material parameter values are used as initial node feature vectors and assigned to the corresponding physical region nodes to complete the node attribute initialization. The stress path embedding vector sequence is received using the attention module. The initial node feature vector of each physical region node is used as the key vector. The correlation coefficient matrix of different physical region nodes under the current working condition is calculated, and the correlation coefficient matrix is converted into dynamic stress transmission weights. Based on the spatial adjacency relationship of the initial node set and the dynamic stress transmission weight, a directed weighted edge connection is constructed to generate a cable degradation trajectory graph neural network.
6. The cable life prediction method based on multi-stress acceleration according to claim 5, characterized in that, The cable degradation trajectory graph neural network is a temporal graph neural network structure, consisting of a second input layer, a graph convolutional layer, a temporal evolution layer, and a second output layer connected sequentially.
7. The cable life prediction method based on multi-stress acceleration according to claim 5, characterized in that, The intrinsic parameters of the material include one or more combinations of dielectric constant, glass transition temperature, and crystallinity.
8. The cable life prediction method based on multi-stress acceleration according to claim 5, characterized in that, Step S5 specifically includes: The stress path embedding vector sequence within the sliding window is obtained as the initial input data, and the stress path embedding vector sequence is subjected to temporal reordering processing to generate a dynamic stress feature tensor. The dynamic stress feature tensor is received and the cable cross-section topology is loaded. Feature aggregation operation is performed on the physical region nodes containing the intrinsic parameter properties of the material to generate a hidden state representation of the nodes. Based on the node hidden state representation and combined with the stress transmission weights generated by the attention module, a dynamic edge update operation is performed to reconstruct the connection strength of the cable degradation trajectory graph neural network and generate a dynamic weighted adjacency matrix. The message passing mechanism of the graph neural network is driven by the dynamic weighted adjacency matrix to perform multiple rounds of iterative update processing on the hidden state representation of each physical region node, thereby generating deep evolution features of the nodes. The deep evolution features of the nodes are extracted and a nonlinear transformation operation is performed to generate the regional damage evolution probability distribution.
9. The cable life prediction method based on multi-stress acceleration according to claim 5, characterized in that, The regional damage evolution probability distribution includes information on the risk level of interface debonding and the rate of increase in local conductivity.
10. The cable life prediction method based on multi-stress acceleration according to claim 1, characterized in that, The stress flow input to be processed includes one or more combinations of temperature fluctuation sequences, mechanical vibration spectrum sequences, and humidity time series data.
