A Hypergraph Neural Network Evaluation Method for Vehicle-Mounted Cable Terminals Based on Differential Dynamic Features

Abstract

This invention discloses a hypergraph neural network evaluation method for vehicle-mounted cable terminals based on differential dynamic features, comprising the following steps: acquiring partial discharge pulse signals of EMU cable terminals; combining multi-order odd derivatives into multi-dimensional dynamic feature vectors; constructing a dynamic hypergraph neural network model; pre-training the dynamic hypergraph neural network model on unlabeled partial discharge pulse signal samples; fine-tuning the trained dynamic hypergraph neural network model; and inputting new partial discharge pulse signals into the fine-tuned dynamic hypergraph neural network model, which outputs the corresponding insulation state category or degradation degree through a classifier. This invention, through the synergistic design of "spectral odd-order differential dynamic features + dynamic feature hypergraph neural network + self-supervised dynamic hypergraph learning," achieves significant progress in dynamic representation of the entire insulation degradation process, high-order structure correlation modeling, and robust evaluation under limited sample conditions, possessing outstanding substantive features and significant engineering application value.

Description

Technical Field

[0001] This invention relates to a hypergraph neural network evaluation method for on-board cable terminals based on differential dynamic features, belonging to the technical field of insulation status evaluation for on-board cable terminals in trains. Background Technology

[0002] Onboard cable terminals are critical components of high-speed trains, and their insulation condition directly affects the safety and reliability of the train's power supply. Due to the multi-layered and complex interface structure of the insulation of onboard cable terminals, long-term operation is susceptible to defects such as air gaps, electrical trees, and carbonization channels at the interface between the main insulation and stress control tubes, which can induce partial discharge and lead to insulation degradation or even breakdown. Therefore, research on insulation condition assessment and early warning methods for onboard cable terminals is of great significance.

[0003] Currently, numerous studies attempt to combine HFCT signals with machine learning and deep learning methods to automatically identify partial discharge modes and assess insulation status of cables and cable terminals. While these methods have made some progress in automatic feature extraction and recognition accuracy, they still have shortcomings in assessing the insulation status of vehicle-mounted cable terminals. Most methods still rely on static or low-order features, either extracting statistics, energy distribution, and single-time-frequency plots from HFCT signals, or mapping the signal into two-dimensional images such as GAF and time-spectrum plots and directly feeding them into the model. This reflects more the average behavior or overall texture within a certain time window, making it difficult to characterize the dynamic evolution characteristics such as the rate of spectral change, abrupt inflection points, and high-frequency oscillations during insulation degradation. At the same time, most methods adopt a fully supervised learning paradigm, which heavily relies on a large number of balanced, labeled partial discharge samples. However, the cost of obtaining defect samples for vehicle-mounted cable terminals is high, real fault data is scarce, and there is a lack of self-supervised learning mechanisms oriented towards partial discharge structural features. As a result, the generalization and engineering practicality under small sample and complex operating conditions remain limited. Summary of the Invention

[0004] The purpose of this invention is to address the problems existing in the prior art by providing a hypergraph neural network evaluation method for vehicle-mounted cable terminals based on differential dynamic features.

[0005] The technical solution provided by this invention to solve the above-mentioned technical problems is: a hypergraph neural network evaluation method for vehicle-mounted cable terminals based on differential dynamic features, comprising the following steps:

[0006] Step S10: Obtain the partial discharge pulse signal of the EMU cable terminal;

[0007] Step S20: Normalize the partial discharge pulse signal and perform fast Fourier transform to obtain the pulse spectrum. Calculate the multi-order odd derivatives of its amplitude spectrum and combine the multi-order odd derivatives into a multi-dimensional dynamic feature vector.

[0008] Step S30: Divide the amplitude spectrum into multiple frequency bands, with the center frequency of each frequency band corresponding to a graph node. Take the average amplitude of the frequency band as the node feature, initialize the hyperedge based on the KNN algorithm, and construct a dynamic hypergraph structure.

[0009] Step S40: Construct a dynamic hypergraph neural network model, perform feature propagation through hypergraph convolution, and introduce a cluster-based hyperedge adaptive update mechanism to enable node features-hyperedge structure-propagation operator to form a closed-loop evolution.

[0010] Step S50: Construct a self-supervised dynamic hypergraph contrastive learning framework to learn the structural representation of cable terminal insulation degradation of EMU trains. Pre-train the dynamic hypergraph neural network model on unlabeled partial discharge pulse signal samples to obtain the trained dynamic hypergraph neural network model.

[0011] Step S60: Fine-tune the trained dynamic hypergraph neural network model using a small number of labeled partial discharge pulse signal samples;

[0012] Step S70: The new partial discharge pulse signal constructs a multidimensional dynamic feature vector and a hypergraph, and inputs it into the fine-tuned dynamic hypergraph neural network model. The corresponding insulation status category or degradation degree is output by the classifier to realize the automated insulation status assessment and early warning of the vehicle-mounted cable terminal.

[0013] A further technical solution is that, in step S10, a partial discharge test is conducted on the cable terminal of the EMU, and the partial discharge pulse signal is extracted from the original partial discharge data.

[0014] A further technical solution is that, in step S10, the time threshold dual-pulse method is used to extract the partial discharge pulse signal from the original partial discharge data.

[0015] A further technical solution is that the specific process of step S20 is as follows:

[0016] Step S21: Normalize the amplitude of the partial discharge pulse signal to obtain a preprocessed signal;

[0017] Step S22: Multiply the preprocessed signal by a window function and perform a fast Fourier transform to obtain the spectrum;

[0018] Step S23: Calculate the first, third, and fifth derivatives of its amplitude spectrum;

[0019] Step S24: Combine the first-order, third-order, and fifth-order derivatives into a three-dimensional dynamic feature vector.

[0020] A further technical solution is that the specific process of step S30 is as follows:

[0021] Step S31: Divide the entire frequency band of its amplitude spectrum into M frequency bands in an equal-width or non-equal-width manner;

[0022] Step S32: Each frequency band center frequency corresponds to a graph node, and the average amplitude of the frequency band is taken as the node feature;

[0023] Step S33: Perform K-nearest neighbor search with node features as input, and construct a hyperedge for each node that includes itself and its K nearest neighbors;

[0024] Step S34: Construct hyperedges based on nearest neighbor nodes and calculate the hyperedge correlation matrix;

[0025] Step S35: Simultaneously assign weights to each hyperedge to obtain the hypergraph.

[0026] A further technical solution is that the workflow of the cluster-based hyperedge adaptive update mechanism in step S40 is as follows:

[0027] Step S41: After training for several rounds, perform K-means clustering on the features of the current layer nodes to divide the nodes into several clusters, with each cluster corresponding to a new candidate hyperedge;

[0028] Step S42: Construct a new correlation matrix based on the new candidate hyperedges;

[0029] Step S43: Linearly merge the initial correlation matrix and the new correlation matrix according to the weight coefficients;

[0030] Step S44: Recalculate the node degree, hyperedge degree, and hypergraph propagation operator using the fused correlation matrix, and continue subsequent training.

[0031] A further technical solution is that the training process of the dynamic hypergraph neural network model in step S50 is as follows:

[0032] Step S51: Generate multiple views of the same partial discharge pulse signal and obtain two sets of different but semantically consistent hypergraphs;

[0033] Step 52: Input the two sets of hypergraphs into the parameter-shared dynamic hypergraph neural network model to obtain the corresponding sample-level embedding vectors;

[0034] Step S53: Take the sample-level embedding vector of the given sample as the positive sample pair, and the embeddings of other samples as the negative samples, and calculate the contrast loss between the two.

[0035] Step S54: On a large amount of unlabeled partial discharge pulse signal data, the dynamic hypergraph neural network model is pre-trained using the above-mentioned contrast loss to obtain a trained dynamic hypergraph neural network model.

[0036] A further technical solution is that in step S51, multiple views are generated by adding Gaussian noise, randomly perturbing some edge weights, or deleting a small number of nodes / superedges.

[0037] A further technical solution is that, in step S60, one or more fully connected classifiers are added to the top of the pre-trained dynamic hypergraph neural network model, and cross-entropy loss is used for supervised fine-tuning.

[0038] A further technical solution is that the supervision loss function in step S60 is:

[0039]

[0040] In the formula: For loss; Indexed by category; For the first One-hot vectors of the true labels of each sample; Output class probabilities for the model.

[0041] The present invention has the following beneficial effects: Compared with the prior art, the present invention has made significant progress in the dynamic representation of the entire insulation degradation process, the modeling of high-order structure associations, and the robust evaluation under few sample conditions through the collaborative design of "spectral odd-order differential dynamic features + dynamic feature hypergraph neural network + self-supervised dynamic hypergraph learning". It has outstanding substantive features and significant engineering application value. Attached Figure Description

[0042] Figure 1 This is a flowchart of the method of the present invention;

[0043] Figure 2 This is a structural diagram of the partial discharge test platform for the vehicle-mounted cable terminal of the present invention;

[0044] Figure 3 This is a structural diagram of the dynamic hypergraph neural network evaluation method of the present invention. Detailed Implementation

[0045] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0046] like Figure 1 As shown, the present invention provides a method for evaluating the differential dynamic features of a vehicle-mounted cable terminal using a hypergraph neural network, comprising the following steps:

[0047] Step S10: Conduct a partial discharge test on the EMU cable terminal and extract the partial discharge pulse signal from the original partial discharge data using the time threshold double pulse method.

[0048] The experimental platform structure for the partial discharge test is shown in the attached figure. Figure 2 As shown, it consists of an external power supply, a voltage regulator, a transformer, a current-limiting resistor, a voltage divider capacitor, an HFCT sensor, and a digital oscilloscope.

[0049] The HFCT sensor mainly consists of a Rogowski coil and an integrating resistor. The Rogowski coil comprises a toroidal magnetic core and multiple turns of winding around the core. For ease of installation and removal, the HFCT sensor typically features a separate, arc-shaped structure, enabling non-invasive detection while the electrical equipment is operating normally.

[0050] An HFCT sensor is connected in series on the grounding circuit of the vehicle cable terminal, with a bandwidth preferably of 3MHz to 30MHz to cover the high-frequency spectrum components of typical partial discharge.

[0051] The output voltage signal of the HFCT was acquired using a digital oscilloscope, with a sampling frequency of [missing information]. f s Preferably 100 MHz, single sampling time window T The preferred time window is 20 ms to ensure that multiple partial discharge pulses can be recorded completely within each time window.

[0052] Under laboratory conditions, cable terminal samples with different defect types (such as interface air gaps, electrical trees, carbonization channels, etc.) and different degrees of degradation were artificially prepared by adjusting the applied voltage to the working voltage; under field conditions, HFCT signals of the on-board cable terminals of running EMUs were directly collected as unlabeled samples.

[0053] The acquired raw discrete-time signal is denoted as:

[0054]

[0055] The specific process for extracting partial discharge pulse signals is as follows: setting a voltage amplitude threshold. and time interval threshold When the sampling point voltage Greater than And the time interval between the sampling point and the previous sampling point that meets the condition is less than At that time, they are identified as part of the same partial discharge pulse. Assume there are two adjacent sampling points that meet the threshold condition. and ,

[0056] like , ,and ( Sampling points (corresponding time), then arrive The sampling points between them are grouped into a single partial discharge pulse.

[0057] By traversing the original data using this method, partial discharge pulses can be extracted:

[0058]

[0059] In the formula: L The length of the partial discharge pulse sequence.

[0060] Step S20: Normalize the partial discharge pulse signal and perform fast Fourier transform to obtain the pulse spectrum. Calculate the multi-order odd derivatives of its amplitude spectrum and combine the multi-order odd derivatives into a multi-dimensional dynamic feature vector.

[0061] Step S21: Normalize the amplitude of the partial discharge pulse signal to obtain a preprocessed signal;

[0062]

[0063] In the formula: The normalized pulse signal; The mean of the sampled sequence, Standard deviation; The original pulse signal;

[0064] Step S22: Multiply the preprocessed signal by a window function and perform a fast Fourier transform to obtain the spectrum;

[0065]

[0066] Frequency resolution is The corresponding frequency is .

[0067] Its amplitude spectrum is as follows:

[0068] or

[0069] Step S23: Calculate the first, third, and fifth derivatives of its amplitude spectrum;

[0070] Taking the central difference of the spectrum along the frequency direction, the first derivative is approximately:

[0071]

[0072] This quantity reflects the "slope" of the spectrum as the frequency changes, and is used to describe the rate of change and migration trend of energy distribution between different frequency bands.

[0073] The third derivative uses the central third-order difference acceleration operator, and its correct form is:

[0074]

[0075] Its corresponding difference coefficient vector is:

[0076]

[0077] This operator can produce a stronger response to the inflection points, acceleration changes and structural abrupt changes in the spectrum, and is suitable for revealing the stage inflection point characteristics of the spectrum morphology when insulation evolves from early micro-defects to middle and late-stage degradation.

[0078] The fifth derivative can be implemented using the central higher-order difference operator commonly used in this field, and its discrete form can be expressed as:

[0079]

[0080] in The pre-selected fifth-order central difference operator (e.g., a finite difference form based on a 5-point or 7-point template) is mainly used to emphasize subtle high-frequency oscillations and local rapid fluctuations in the spectrum, reflecting the changes in spectral details caused by the enhanced local discharge activity in the later stages of degradation.

[0081] Step S24: Combine the first-order, third-order, and fifth-order derivatives into a three-dimensional dynamic feature vector;

[0082] Since the numerical magnitude and dynamic range of derivatives of different orders differ significantly, this invention addresses the differences in the numerical magnitude and dynamic range of derivatives of different orders. The derivatives are normalized according to the maximum absolute value within the window:

[0083]

[0084] In the formula: Frequency index before normalization place First derivative, Normalized frequency index place Derivative order; To prevent division by zero of tiny positive numbers (such as...) ).

[0085] The normalized derivative retains the sensitivity of different orders to the rate of change, inflection point and high-frequency oscillation, and ensures numerical stability, which is convenient for subsequent network training and structural modeling.

[0086] Based on this, the multi-order derivatives at the same frequency index k are synthesized into a three-dimensional dynamic feature vector:

[0087]

[0088] This results in a three-channel spectral sequence, providing high-quality structural input for subsequent graph and hypergraph construction based on frequency band nodes.

[0089] Step S30: Divide the amplitude spectrum into multiple frequency bands, with the center frequency of each frequency band corresponding to a graph node. Take the average amplitude of the frequency band as the node feature, initialize the hyperedge based on the KNN algorithm, and construct a dynamic hypergraph structure.

[0090] Step S31: Measure the full frequency band of its amplitude spectrum. Divide into M frequency bands using either equal-width or non-equal-width methods;

[0091] Step S32, Center frequency of each frequency band Corresponding to a graph node The average amplitude of the frequency band is taken as the node feature;

[0092] No. i The feature vector of a frequency band is defined as all the features within that frequency band. k superior Average value:

[0093]

[0094] In the formula: For the first i Node characteristics of each frequency band; For the first i Each frequency band contains a set of frequency indexes;

[0095] Step S33: Perform K-nearest neighbor search with node features as input, and construct a hyperedge for each node that includes itself and its K nearest neighbors;

[0096] For a set of nodes, based on node characteristics Perform a K-nearest neighbor search on the input, for each node Construct a superedge that includes itself and its K nearest neighbors:

[0097]

[0098] In the formula: For the first One super edge, This represents the K-nearest neighbor search algorithm;

[0099] Step S34: Construct hyperedges based on nearest neighbor nodes and calculate the hyperedge correlation matrix;

[0100] Using the correlation matrix Indicate the relationship between a node and its hyperedge:

[0101]

[0102] Step S35: Simultaneously assign a weight to each superedge. Obtain the hypergraph ;

[0103] Step S40: Construct a dynamic hypergraph neural network model, perform feature propagation through hypergraph convolution, and introduce a cluster-based hyperedge adaptive update mechanism to enable node features-hyperedge structure-propagation operator to form a closed-loop evolution.

[0104] Its dynamic hypergraph neural network model structure is as follows: Figure 3 As shown;

[0105] 1) Node degree matrix and hypermarginality matrix They are defined as follows:

[0106]

[0107]

[0108] in, For the correlation matrix Element;

[0109] 2) Construct the hypergraph propagation operator;

[0110]

[0111] in, For propagation operators;

[0112] 3) Hypergraph convolutional layer;

[0113] Let the feature matrix of the input nodes be... The output of the hypergraph convolutional layer is:

[0114]

[0115] In the formula: The input is the node feature matrix. This is the node feature matrix output after convolution. It is a non-linear activation function. Let the node degree matrix be... It is the hypermarginality matrix. For the first Layer weight matrix.

[0116] By stacking multiple layers of hypergraph convolution, node feature representations that gradually aggregate larger neighborhood and higher-order structural information can be obtained;

[0117] 4) Graph-level / sample-level embedding vectors;

[0118] For a certain segment of HFCT signal, the hypergraph features within the segment are aggregated through time pooling, attention weighting, or simple averaging to obtain a sample-level embedding vector z as the insulation state representation of the signal segment;

[0119] 5) Cluster-based adaptive hyperedge update mechanism;

[0120] To overcome the problem that the hypergraph structure is fixed and cannot reflect the evolution of features, this invention introduces a clustering-based adaptive hyperedge update mechanism during the training process, enabling a closed-loop evolution of "node features – hyperedge structure – propagation operator". This mechanism is specifically as follows:

[0121] 5.1. Clustering based on current features;

[0122] After several rounds of training, the features of the current layer nodes are... Perform K-means clustering, and Each node is divided into There are 3 clusters, each corresponding to a new candidate superedge. ;

[0123] 5.2. Construction of a new association matrix;

[0124] Construct a new correlation matrix :

[0125]

[0126] 5.3. Hyperedge fusion and propagation operator update

[0127] The initial Knn hyperedge correlation matrix Constructing a new correlation matrix Linear fusion based on weighting coefficients:

[0128]

[0129] in, The fused correlation matrix, The initial correlation matrix based on Knn. To construct a new correlation matrix, This represents the correlation matrix fusion algorithm. For the weights of the correlation matrix, ;

[0130] Then use Recalculate node degree, hyperedge degree, and hypergraph propagation operator. The process of "feature update – clustering – structure update" is repeated continuously, enabling the hypergraph structure to adaptively evolve as the insulation degradation features are learned, thus better capturing multi-stage, high-order association patterns.

[0131] Step S50: Construct a self-supervised dynamic hypergraph contrastive learning framework to learn the structural representation of cable terminal insulation degradation of EMU trains. Pre-train the dynamic hypergraph neural network model on unlabeled partial discharge pulse signal samples to obtain the trained dynamic hypergraph neural network model.

[0132] To fully utilize the large amount of unlabeled HFCT data, this invention constructs a self-supervised dynamic hypergraph learning framework. This framework learns the intrinsic structural representation of insulation degradation under unlabeled conditions through contrastive learning, specifically including the following steps:

[0133] Step S51: For the same original HFCT sequence, generate multiple "views" by adding Gaussian noise, randomly perturbing some edge weights, or deleting a small number of nodes / hyperedges, thereby obtaining two different but semantically consistent hypergraphs. , ;

[0134] Step S52: Input the two sets of hypergraphs into the parameter-shared dynamic hypergraph neural network model to obtain the corresponding sample-level embedding vectors. , ;

[0135] Step S53: For a given sample, its two-view embedding , As positive sample pairs, the embeddings of other samples are used as negative samples. The contrastive loss is defined as:

[0136]

[0137] In the formula: To compare the losses, For temperature parameters, Represents the embedding vector and Similarity between them Iterate through all samples in the batch. Minimize... This causes multi-view structural representations under the same insulation state to be close to each other in the embedding space, while representations under different states are far apart from each other;

[0138] The similarity between embedded vectors is calculated using cosine similarity:

[0139]

[0140] Step S54: On a large amount of unlabeled HFCT data, the dynamic hypergraph neural network model is pre-trained using the above contrast loss to obtain initial parameters that can reflect the inherent structural laws of insulation degradation, that is, the trained dynamic hypergraph neural network model is obtained.

[0141] Step S60: Add one or more fully connected classifiers on top of the pre-trained HGNN-DF encoder and use cross-entropy loss for supervised fine-tuning.

[0142] Let the first The one-hot vector of the true label of each sample is The model outputs the class probability as The supervised loss function is:

[0143]

[0144] In the formula: For loss; Indexed by category; For the first One-hot vectors of the true labels of each sample; Output class probabilities for the model;

[0145] Step S70: The new partial discharge pulse signal is used to construct a multidimensional dynamic feature vector and a hypergraph according to the above process, and is input into the fine-tuned dynamic hypergraph neural network model. The corresponding insulation status category or degradation degree is output by the classifier to realize the automated insulation status assessment and early warning of the vehicle-mounted cable terminal.

[0146] This invention constructs a dynamic evolution feature map on the frequency-time plane by calculating the first, third, and fifth odd derivatives of the HFCT spectrum: the first derivative depicts the migration trend of energy between different frequency bands, the third derivative highlights the spectral inflection points during the transition of insulation degradation stages, and the fifth derivative emphasizes the high-frequency oscillations and subtle oscillation structures in the later stages of degradation. Furthermore, a dynamic map is constructed using frequency bands as nodes and the similarity of odd derivative features as edges, allowing the evolution trajectory of the cable terminal from healthy to multi-stage degradation to be explicitly represented in structural space. This fundamentally overcomes the shortcomings of traditional detection and general static feature methods in depicting the dynamic laws of the entire insulation degradation process.

[0147] Building upon this foundation, this invention constructs a dynamic feature hypergraph and introduces an adaptive hyperedge update mechanism: first, locally correlated hyperedges are formed using K-nearest neighbors; then, during training, node features are clustered to generate cluster-level hyperedges; and the hypergraph correlation matrix is ​​periodically updated through weight fusion, achieving the coordinated evolution of "node features—hyperedge structure—propagation operator." The hypergraph structure can be reconstructed in real time according to insulation state and feature distribution, while also taking into account short-term local nonlinear changes and cross-frequency band and cross-stage global degradation modes. Compared to graph / hypergraph models with one-time construction and fixed structures, this significantly improves the expressive power for multi-scale, high-order correlation features and adaptability to changes in operating conditions.

[0148] Addressing the challenges of scarce defect samples and complex on-site conditions in vehicle-mounted cable terminals, this invention constructs a self-supervised contrastive learning framework based on dynamic feature hypergraphs. It builds multi-view hypergraphs from the same HFCT signal, extracts embeddings using a shared encoder, and employs InfoNCE loss based on cosine similarity to approximate the multi-view representations of the same sample, extrapolating representations from different samples. This allows for structural pre-training on a large amount of unlabeled data, requiring only a few labeled samples for fine-tuning to achieve high-precision insulation condition classification. This mechanism fully utilizes the structural information in unlabeled HFCT data, significantly reducing reliance on extensive manual annotation, while enhancing the robustness of features to noise and operational disturbances. This enables the invention to maintain good generalization performance even under small sample sizes and complex on-site conditions, making it suitable for long-term online monitoring and condition assessment.

[0149] The above description is not intended to limit the present invention in any way. Although the present invention has been disclosed through the above embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A hypergraph neural network evaluation method for vehicle-mounted cable terminals based on differential dynamic features, characterized in that, Includes the following steps: Step S10: Obtain the partial discharge pulse signal of the EMU cable terminal; The time-threshold dual-pulse method was used to extract partial discharge pulse signals from the raw partial discharge data. Step S20: Normalize the partial discharge pulse signal and perform fast Fourier transform to obtain the pulse spectrum. Calculate the multi-order odd derivatives of its amplitude spectrum and combine the multi-order odd derivatives into a multi-dimensional dynamic feature vector. Step S21: Normalize the amplitude of the partial discharge pulse signal to obtain a preprocessed signal; In the formula: The normalized pulse signal; The mean of the sampled sequence, Standard deviation; The original pulse signal; Step S22: Multiply the preprocessed signal by a window function and perform a fast Fourier transform to obtain the spectrum; Frequency resolution is The corresponding frequency is ; Its amplitude spectrum is as follows: or Step S23: Calculate the first, third, and fifth derivatives of its amplitude spectrum; Taking the central difference of the spectrum along the frequency direction, the first derivative is approximately: The third derivative uses the central third-order difference acceleration operator, and its correct form is: Its corresponding difference coefficient vector is: The fifth derivative is implemented using the central higher-order difference operator commonly used in this field, and its discrete form can be expressed as: in For a pre-selected fifth-order central difference operator; Step S24: Combine the first-order, third-order, and fifth-order derivatives into a three-dimensional dynamic feature vector; For each order The derivatives are normalized according to the maximum absolute value within the window: In the formula: Frequency index before normalization place First derivative, Normalized frequency index place First derivative; To prevent small positive numbers from being divided by zero; Based on this, index the same frequency. Synthesizing three-dimensional dynamic eigenvectors from multiple derivative arrays at a given point: Step S30: Divide its amplitude spectrum into multiple frequency bands, initialize the hypergraph structure based on the Knn algorithm, construct hyperedges based on the nearest neighbor nodes, and calculate the hyperedge correlation matrix; Step S31: Divide the entire frequency band of its amplitude spectrum into M frequency bands in an equal-width or non-equal-width manner; Step S32: Each frequency band center frequency corresponds to a graph node, and the average amplitude of the frequency band is taken as the node feature; Step S33: Perform K-nearest neighbor search with node features as input, and construct a hyperedge for each node that includes itself and its K nearest neighbors; Step S34: Construct hyperedges based on nearest neighbor nodes and calculate the hyperedge correlation matrix; Step S35: Simultaneously assign weights to each hyperedge to obtain the hypergraph; Step S40: Construct a dynamic hypergraph neural network model, perform feature propagation through hypergraph convolution, and introduce a cluster-based hyperedge adaptive update mechanism to enable node features-hyperedge structure-propagation operator to form a closed-loop evolution. Step S50: Construct a self-supervised dynamic hypergraph contrastive learning framework to learn the structural representation of cable terminal insulation degradation of EMU trains. Pre-train the dynamic hypergraph neural network model on unlabeled partial discharge pulse signal samples to obtain the trained dynamic hypergraph neural network model. Step S60: Fine-tune the trained dynamic hypergraph neural network model using a small number of labeled partial discharge pulse signal samples; Step S70: The new partial discharge pulse signal constructs a multidimensional dynamic feature vector and a hypergraph, and inputs it into the fine-tuned dynamic hypergraph neural network model. The corresponding insulation status category or degradation degree is output by the classifier to realize the automated insulation status assessment and early warning of the vehicle-mounted cable terminal.

2. The method for evaluating vehicle-mounted cable terminals using a hypergraph neural network based on differential dynamic features according to claim 1, characterized in that, In step S10, a partial discharge test is conducted on the cable terminal of the EMU, and the partial discharge pulse signal is extracted from the original partial discharge data.

3. The method for evaluating vehicle-mounted cable terminals using a hypergraph neural network based on differential dynamic features according to claim 1, characterized in that, The workflow of the cluster-based hyperedge adaptive update mechanism in step S40 is as follows: Step S41: After training for several rounds, perform K-means clustering on the features of the current layer nodes to divide the nodes into several clusters, with each cluster corresponding to a new candidate hyperedge; Step S42: Construct a new correlation matrix based on the new candidate hyperedges; Step S43: Linearly merge the initial correlation matrix and the new correlation matrix according to the weight coefficients; Step S44: Recalculate the node degree, hyperedge degree, and hypergraph propagation operator using the fused correlation matrix, and continue subsequent training.

4. The method for evaluating vehicle-mounted cable terminals using a hypergraph neural network based on differential dynamic features according to claim 1, characterized in that, The training process of the dynamic hypergraph neural network model in step S50 is as follows: Step S51: Generate multiple views of the same partial discharge pulse signal and obtain two sets of different but semantically consistent hypergraphs; Step 52: Input the two sets of hypergraphs into the parameter-shared dynamic hypergraph neural network model to obtain the corresponding sample-level embedding vectors; Step S53: Take the sample-level embedding vector of the given sample as the positive sample pair, and the embeddings of other samples as the negative samples, and calculate the contrast loss between the two. Step S54: On a large amount of unlabeled partial discharge pulse signal data, the dynamic hypergraph neural network model is pre-trained using the above-mentioned contrast loss to obtain a trained dynamic hypergraph neural network model.

5. The method for evaluating the differential dynamic characteristics of a vehicle-mounted cable terminal using a hypergraph neural network according to claim 4, characterized in that, In step S51, multiple views are generated by adding Gaussian noise, randomly perturbing some edge weights, or deleting a small number of nodes / superedges.

6. The method for evaluating vehicle-mounted cable terminals using a hypergraph neural network based on differential dynamic features according to claim 1, characterized in that, In step S60, one or more fully connected classifiers are added to the top of the pre-trained dynamic hypergraph neural network model, and cross-entropy loss is used for supervised fine-tuning.

7. The method for evaluating the differential dynamic characteristics of a vehicle-mounted cable terminal using a hypergraph neural network according to claim 6, characterized in that, The supervision loss function in step S60 is: In the formula: For loss; Indexed by category; For the first One-hot vectors of the true labels of each sample; Output class probabilities for the model.

Images