Man-machine interaction type UHVDC fault diagnosis method

Abstract

The invention discloses a man-machine interaction type UHVDC fault diagnosis method, which comprises the following steps of S1, inputting original data into a Qwen2.5 VL large model, respectively performing unified stage extraction and directional stage extraction on features of the original data corresponding to keywords, performing extraction in the unified stage according to unified keywords, performing extraction in the directional stage according to directional keywords, and performing extraction in the directional stage according to directional keywords; discretizing the features extracted in the unification stage and the orientation stage to form a matrix X; and S2, inputting the matrix X into any one of a TabPFN model, an EAPCR model, an SVM model, a NaiveBayes model and an XGboost model, and then outputting a diagnosis result. According to the diagnosis method, the Qwen2.5 VL large model with high applicability is adopted, the accuracy of fault diagnosis can be improved by combining the directional extraction features of the unification stage and the directional stage, and particularly, the fault diagnosis capability of the fault diagnosis method combined with the EAPCE model is the strongest.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and fault diagnosis of power systems, and in particular to a human-machine interactive method for diagnosing UHVDC faults. Background Technology

[0002] High-voltage direct current (UHVDC) transmission, as a crucial component of modern power systems, is characterized by high voltage levels (above ±800kV), large transmission capacity (multi-gigawatts), and long-distance transmission (thousands of kilometers). It enables efficient long-distance, large-scale power transmission and is widely used in inter-regional energy allocation and renewable energy grid integration. However, UHVDC faults often lead to widespread power outages, equipment damage, and economic losses, severely impacting energy supply security and grid stability, and increasing the economic burden on power companies and society. Traditional power equipment fault detection methods typically rely on manual analysis of waveform data and fault messages. Engineers determine the fault type and cause based on information in the fault waveform and message. These methods are inefficient and lack real-time processing capabilities, failing to meet the demands for rapid and accurate fault diagnosis in complex power environments.

[0003] Existing methods are limited by the following aspects: 1. Learning signal features based solely on waveform data is susceptible to sample imbalance and noise, leading to a high misjudgment rate; 2. Rule-based judgment based solely on fault messages (events / protection actions) is insufficient for location resolution and cannot accurately locate specific sites and fault types; 3. In UHVDC scenarios, a single site may generate thousands of waveform files, and traditional methods cannot efficiently process all the data.

[0004] As a major engineering project, UHVDC's data is highly confidential and cannot be publicly released, resulting in the absence of a publicly available multimodal fault dataset. This data scarcity makes it difficult for deep learning models to be adequately trained and generalized, especially under small sample conditions, where models are prone to overfitting or insufficient generalization ability. Furthermore, due to the system's high stability, actual faults occur very rarely, further exacerbating the sample scarcity problem. Simultaneously, the fault data is multi-source and heterogeneous, including waveform recordings (time-series waveforms), fault messages (text events), and other data types. These data have different formats and dimensions, making unified processing and effective fusion difficult, further intensifying the bottleneck in data utilization.

[0005] Existing research often involves "early splicing" or "late voting" of information such as waveform recordings and messages, lacking modeling of explicit high-order interactions between multimodal features. Insufficient utilization of cross-modal feature information leads to limitations in diagnostic performance and interpretability. In industrial scenarios such as UHVDC fault diagnosis, which require "high recall, low false alarms, and interpretability," it is difficult to obtain a large number of precise samples due to the rarity of fault events, strong data confidentiality, and engineering complexity. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a method for diagnosing human-computer interactive UHVDC faults.

[0007] The objective of this invention is achieved through the following technical solution:

[0008] A method for diagnosing UHVDC faults using a human-computer interactive interface includes the following steps:

[0009] S1. Input the original data into the Qwen2.5_VL large model, and extract the features of the original data corresponding to each keyword in the unified stage and the targeted stage respectively. The unified stage extracts based on the unified keyword, and the targeted stage extracts based on the targeted keyword. Discretize the features extracted in the unified stage and the targeted stage to form matrix X.

[0010] S2. Input matrix X into any one of the TabPFN model, EAPCR model, SVM model, NaiveBayes model, and XGboost model, and output the diagnostic results.

[0011] Preferably, the EAPCR model maps matrix X into a dense representation E by feeding it into a learnable embedding layer; then, it uses a bilinear attention mechanism to calculate the relationship between any two features to obtain matrix A; next, it introduces the permutation matrix P of matrix A, processes matrix A and matrix P in parallel on a two-branch lightweight convolutional network, extracts local and non-local interaction features respectively, and then flattens and concatenates the outputs of the two branches; finally, it maps the concatenated vector to a five-dimensional soft-max probability through a multilayer perceptron with residual connections, and outputs fault diagnosis results for classes 0-4.

[0012] Preferably, each unified keyword includes unified general keywords, unified special keywords, and unified fixed keywords. The unified general keywords consist of a unified general waveform recording and a unified general prompt word. The unified general waveform recording includes DC line voltage, neutral bus voltage, DC current on the DC pole bus valve side, DC current on the DC neutral bus valve side, DC current on the DC pole bus line side, grounding electrode 1 current, grounding electrode 2 current, the 50Hz component of the DC line voltage, the 50Hz component of the DC pole bus line side DC current, the difference between the pole bus current and the neutral bus current, DC differential protection stage 1 setting, DC differential protection stage 2 setting, common-mode transient, common-mode integral, differential-mode integral, differential 1 of the DC line voltage, differential 2 of the DC line voltage, the maximum value of the three-phase AC current on the Y-bridge valve side, the maximum value of the three-phase AC current on the D-bridge valve side, and the connection voltage of the high and low voltage valve groups. The unified general keywords include maximum positive peak value, maximum negative peak value, maximum slope peak value, abnormal duration percentage, transient synchronicity, oscillation count, baseline drift, and noise spikes. The unified special keywords... It consists of a unified special waveform recording and a unified special keyword. The unified special waveform recording includes the AC voltage of phase A, phase B, and phase C; the AC current of phase A on the Y-bridge valve side; the AC current of phase B on the Y-bridge valve side; the AC current of phase C on the Y-bridge valve side; the AC current of phase A on the D-bridge valve side; the AC current of phase B on the D-bridge valve side; and the AC current of phase C on the D-bridge valve side. The unified special keyword includes the maximum positive peak value, the maximum negative peak value, the maximum slope peak value, the percentage of abnormal duration, the synchronization of abrupt changes, the number of oscillations, the baseline drift, noise glitches, the symmetry of positive and negative peak values, and... Half-cycle symmetry; unified fixed keywords include three-phase AC voltage imbalance, three-phase AC voltage imbalance on the Y-bridge valve side, three-phase AC voltage imbalance on the D-bridge valve side, AC voltage phase hysteresis, AC voltage phase hysteresis on the Y-bridge valve side, AC voltage phase hysteresis on the D-bridge valve side, maximum positive peak value difference of AC voltage, maximum positive peak value difference of AC voltage on the Y-bridge valve side, maximum positive peak value difference of AC voltage on the D-bridge valve side, maximum negative peak value difference of AC voltage, maximum negative peak value difference of AC voltage on the Y-bridge valve side, and maximum negative peak value difference of AC voltage on the D-bridge valve side.

[0013] Preferably, the directional keywords include the maximum positive / maximum negative peak value of the DC current on the valve side of the 100Hz protection DC neutral bus, the maximum positive / maximum negative peak value of the 100Hz component of the DC current on the valve side of the 100Hz protection DC neutral bus, the maximum positive / maximum negative peak value of all waveforms of the DC current on the line side of the grounding electrode differential protection station (grounding wire current - grounding electrode 1 current - grounding electrode 2 current - the other pole line side), the waveform difference 1 between phase A AC current on the Y-bridge valve side and phase A AC current on the D-bridge valve side of the bridge differential protection, the waveform difference 1 between phase B AC current on the Y-bridge valve side and phase B AC current on the D-bridge valve side of the bridge differential protection, and the waveform difference between phase C AC current on the Y-bridge valve side and phase C AC current on the D-bridge valve side of the bridge differential protection. Value 1: Waveform difference between phase A AC on the Y-bridge valve side and phase A AC on the D-bridge valve side under bridge differential protection. Value 2: Waveform difference between phase C AC on the Y-bridge valve side and phase C AC on the D-bridge valve side under bridge differential protection. Value 3: Waveform difference between DC current on the DC pole valve side, DC current on the DC neutral bus valve side, and phase A AC on the Y-bridge valve side under commutation failure protection. Value 4: Waveform difference between DC current on the DC pole valve side, DC current on the DC neutral bus valve side, and phase B AC on the Y-bridge valve side under commutation failure protection. Value 5: Waveform difference between DC current on the DC pole valve side, DC current on the DC neutral bus valve side, and phase C AC on the Y-bridge valve side under commutation failure protection. Value 6: Waveform difference between DC current on the DC pole valve side, DC current on the DC neutral bus valve side, and phase A AC on the D-bridge valve side under commutation failure protection. Inter-phase difference, DC current on the DC pole bus valve side - DC current on the DC neutral bus valve side - DC current on the D bridge valve side B phase AC waveform difference, DC current on the DC pole bus valve side - DC current on the DC neutral bus valve side - DC current on the D bridge valve side C phase AC waveform difference, DC current on the DC pole bus valve side - DC current on the DC neutral bus valve side - DC current on the Y bridge valve side A phase AC - DC current on the D bridge valve side A phase AC waveform difference, DC current on the DC pole bus valve side - DC current on the DC neutral bus valve side - DC current on the Y bridge valve side B phase AC - DC current on the D bridge valve side B phase AC waveform difference, DC current on the DC pole bus valve side - DC current on the DC neutral bus valve side - DC current on the Y bridge valve side C phase AC - D bridge The waveform difference between phase C AC current on the valve side, the waveform difference between the DC current on the valve side of the converter short circuit protection DC pole bus valve side - DC current on the valve side of the DC neutral bus - A phase AC current on the Y bridge valve side - A phase AC current on the D bridge valve side (1 / 2), the waveform difference between the DC current on the valve side of the converter short circuit protection DC pole bus valve side - DC current on the valve side of the DC neutral bus - B phase AC current on the Y bridge valve side - B phase AC current on the D bridge valve side (1 / 2), the waveform difference between the DC current on the valve side of the converter short circuit protection DC pole bus valve side - DC current on the valve side of the DC neutral bus - C phase AC current on the Y bridge valve side - C phase AC current on the D bridge valve side (1 / 2), and the absolute value of the waveform difference between the DC current on the valve side of the DC pole bus valve side and the DC current on the valve side of the DC differential protection DC current.DC backup differential protection DC line voltage maximum peak absolute value, AC / DC overcurrent protection DC pole bus valve side DC current - DC neutral bus valve side DC current - Y bridge valve side A phase AC current - D bridge valve side A phase AC current maximum peak absolute value, AC / DC overcurrent protection DC pole bus valve side DC current - DC neutral bus valve side DC current - Y bridge valve side B phase AC current - D bridge valve side B phase AC current maximum peak absolute value, AC / DC overcurrent protection DC pole bus valve side DC current - DC The absolute value of the maximum peak of the DC current on the neutral bus valve side - the AC current on the C-phase of the Y-bridge valve side - the AC current on the C-phase of the D-bridge valve side; the absolute value of the difference between the DC line voltage and the voltage waveform of the high and low voltage valve group connection; the absolute value of the difference between the voltage waveform of the high and low voltage valve group connection and the neutral bus voltage; the absolute value of the maximum peak of the common-mode transient change in the DC line traveling wave protection; the absolute value of the maximum peak of the common-mode integral in the DC line traveling wave protection; the absolute value of the maximum peak of the differential-mode integral in the DC line traveling wave protection; DC... The following parameters are considered: deviation of DC line voltage from steady state under low voltage protection; absolute value of the difference between DC line voltage and high / low voltage valve group connection voltage waveforms under low voltage protection; absolute value of the maximum peak value of DC line voltage under low voltage protection; absolute value of the difference between the waveforms of the other pole line current and the DC current on the other pole line side under metallic loop differential protection; absolute value of the maximum peak value of DC line voltage under DC line sudden change protection; maximum value of the DC voltage waveform slope under DC line sudden change protection; absolute value of the difference between the waveforms of the other pole line current and the DC current on the other pole line side under DC line differential protection; absolute value of the maximum peak value of the 50Hz component of the DC current on the neutral bus valve side under 50Hz protection; absolute value of the maximum peak value of DC line voltage under low voltage protection; absolute value of the maximum peak value of the neutral bus voltage under grounding electrode open circuit protection; absolute value of the maximum peak value of the DC current on the other pole line side under grounding electrode open circuit protection; and absolute value of the maximum peak value between the current of grounding electrode 1 and the current of grounding electrode 2 under grounding electrode open circuit protection.

[0014] Preferably, the keywords are defined by ultra-high voltage direct current experts in combination with electrical principles and related knowledge.

[0015] Preferably, when extracting features during the targeted phase, the setting for features that are automated but do not trigger the protection execution process is 0.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0017] By employing the Qwen2.5_VL large model, combining keyword feature extraction in the unified and targeted phases, and integrating a deep interactive learning model, cross-modal feature interactive learning is achieved, enabling more accurate targeted feature extraction and improving the overall fault diagnosis performance. The targeted phase accurately extracts fault features, avoiding full analysis and improving diagnostic efficiency. The Qwen2.5_VL large model is applicable to some existing deep interactive learning models, demonstrating a wide range of applicability. This invention's diagnostic method can diagnose faults accurately in real time, providing a reliable basis for intelligent operation and maintenance of power systems. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the dataset fault information collection points for a specific implementation method. Detailed Implementation

[0019] To make the technical problems solved by the present invention, the technical solutions adopted, and the technical effects achieved clearer, the technical solutions of the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] This specific embodiment provides a method for diagnosing human-computer interactive UHVDC faults, comprising the following steps:

[0021] S1. Input the raw data into the Qwen2.5_VL large model. Extract the features of the raw data corresponding to each keyword in the unified stage and the directional stage respectively. The keywords are defined by UHVDC experts in combination with electrical principles and related knowledge to define the initial feature dimensions related to the fault. The definition rules are input into the Qwen2.5_VL large model. In the unified stage, extract based on the unified keywords to capture common fault modes. In the directional stage, extract based on the directional keywords. When extracting features in the directional stage, set the features that do not trigger protection to perform similar processes but are automated to 0, so as to maintain the consistency of feature dimensions and reduce noise interference. Discretize the features extracted in the unified stage and the directional stage to form matrix X.

[0022] S2. By feeding matrix X into a learnable embedding layer, it is mapped to a dense representation E, preserving feature semantics and relationships, alleviating dimensionality explosion, automatically learning implicit relationships, and improving generalization performance; then, a bilinear attention mechanism is used to calculate the relationship between any two features, resulting in matrix A; A = Tanh(EE) T ), where Tanh(X) is the hyperbolic tangent function, matrix A shows the relationships between features, providing structural information about cross-modal interactions; then, the permutation matrix P of matrix A is introduced, P = MAMT This method breaks down originally adjacent features and brings distant features closer together, revealing feature combination patterns and providing interactive structural information to capture nonlinear interactions and enhance interpretability. Matrices A and P are processed in parallel on a dual-branch lightweight convolutional network to extract local and non-local interactive features, adjust feature relationships, obtain comprehensive cross-modal information, break sequential dependencies, and avoid missing key information. The outputs of the two branches are then flattened and concatenated. Finally, the concatenated vector is mapped to a five-dimensional soft-max probability through a residual-connected multilayer perceptron, outputting 0-4 class fault diagnosis results to diagnose ultra-high voltage fault types / regions / causes.

[0023] S3. Training and evaluation of diagnostic results: (1) Training process: 5-fold cross-validation

[0024] The purpose of this step is to objectively evaluate generalization performance by dividing the data into five parts, training and validating them sequentially, and eliminating bias from a single split.

[0025] Compared to a single validation set, it is more robust and has smaller errors, truly reflecting the model's generalization ability under small sample conditions and avoiding overfitting.

[0026] (2) Loss function: Cross-entropy loss;

[0027] The purpose of this step is to measure the difference between the predicted probability and the true class, and to optimize the model's classification accuracy.

[0028] Reasons for selection: Suitable for multi-classification tasks (such as fault type diagnosis), efficiently handles class imbalance, and promotes the learning of discrimination boundaries.

[0029] (3) Optimizer: AdamW (learning rate = 1e-3);

[0030] Function of the steps: Automatically adjust parameters, combine momentum and adaptive learning rate to accelerate convergence.

[0031] Reason for selection: Compared with traditional optimizers, it is more stable on multimodal data and avoids overfitting through weight decay, thus achieving excellent generalization performance.

[0032] like Figure 1 As shown, the dataset of this invention consists of relevant data collected at 20 potential failure points F1-F20.

[0033] Each unified keyword includes unified general keywords, unified special keywords, and unified fixed keywords. A unified general keyword consists of a unified general waveform recorder and a unified general prompt word. The unified general waveform recorder includes DC line voltage, neutral bus voltage, DC current on the DC pole bus valve side, DC current on the DC neutral bus valve side, DC current on the DC pole bus line side, grounding electrode 1 current, grounding electrode 2 current, the 50Hz component of the DC line voltage, the 50Hz component of the DC pole bus line side DC current, the difference between the pole bus current and the neutral bus current, DC differential protection stage 1 setting, DC differential protection stage 2 setting, common-mode transient, common-mode integral, differential-mode integral, differential 1 of the DC line voltage, differential 2 of the DC line voltage, maximum three-phase AC current on the Y-bridge valve side, maximum three-phase AC current on the D-bridge valve side, and high / low voltage valve group connection voltage. For example, unified general keywords include maximum positive peak value, maximum negative peak value, maximum slope peak value, abnormal duration percentage, transient synchronicity, number of oscillations, baseline drift, and noise spikes. For example, common keywords include: maximum positive peak value of DC line voltage, maximum slope peak value of DC current on the DC pole bus valve side, and percentage of abnormal continuous 50Hz component of DC line voltage.

[0034] The unified special keywords consist of a unified special waveform recording and a unified special keyword. The unified special waveform recording includes the AC voltage of phase A, phase B, and phase C, the AC current of phase A on the Y-bridge valve side, the AC current of phase B on the Y-bridge valve side, the AC current of phase C on the Y-bridge valve side, the AC current of phase A on the D-bridge valve side, the AC current of phase B on the D-bridge valve side, and the AC current of phase C on the D-bridge valve side. The unified special keywords include the maximum positive peak value, the maximum negative peak value, the maximum slope peak value, the percentage of abnormal duration, the synchronization of abrupt changes, the number of oscillations, the baseline drift, noise glitches, the symmetry of positive and negative peak values, and the symmetry of half-cycles; for example, the synchronization of abrupt changes in phase A AC voltage, the baseline drift of phase A AC current on the D-bridge valve side, and the noise glitches of phase C AC current on the D-bridge valve side.

[0035] The unified fixed keywords include AC voltage three-phase unbalance, Y-bridge valve side AC three-phase unbalance, D-bridge valve side AC three-phase unbalance, AC voltage phase hysteresis, Y-bridge valve side AC phase hysteresis, D-bridge valve side AC phase hysteresis, AC voltage maximum positive peak value difference, Y-bridge valve side AC maximum positive peak value difference, D-bridge valve side AC maximum positive peak value difference, AC voltage maximum negative peak value difference, Y-bridge valve side AC maximum negative peak value difference, and D-bridge valve side AC maximum negative peak value difference.

[0036] The targeted keywords include: maximum positive / negative peak value of the DC current on the valve side of the 100Hz protection DC neutral busbar; maximum positive / negative peak value of the 100Hz component of the DC current on the valve side of the 100Hz protection DC neutral busbar; maximum positive / negative peak value of all waveforms of the DC current on the line side of the grounding electrode differential protection station (grounding wire current - grounding electrode 1 current - grounding electrode 2 current - other electrode line side); absolute value of the waveform difference between phase A AC current on the Y-bridge valve side and phase A AC current on the D-bridge valve side; waveform difference between phase B AC current on the Y-bridge valve side and phase B AC current on the D-bridge valve side; waveform difference between phase C AC current on the Y-bridge valve side and phase C AC current on the D-bridge valve side; and the waveform difference between phase C AC current on the Y-bridge valve side and phase C AC current on the D-bridge valve side. Waveform difference between phase A AC on bridge valve side and phase A AC on bridge valve side; Waveform difference between phase C AC on bridge valve side and phase C AC on bridge valve side; Waveform difference between DC current on DC pole bus valve side and DC current on DC neutral bus valve side; Waveform difference between phase A AC on bridge valve side; Waveform difference between phase B AC on bridge valve side; Waveform difference between DC current on DC pole bus valve side and DC current on DC neutral bus valve side; Waveform difference between phase C AC on bridge valve side; Waveform difference between phase A AC on bridge valve side; Waveform difference between phase B AC on bridge valve side; Waveform difference between phase C AC on bridge valve side; Waveform difference between phase A AC on bridge valve side; Waveform difference between phase B AC on bridge valve side; Waveform difference between phase C AC on bridge valve side; Waveform difference between phase A AC on bridge valve side; Waveform difference between phase B AC on bridge valve side; Waveform difference between phase C AC on bridge valve side; Waveform difference between phase C AC on bridge valve side; Waveform difference between phase A AC on bridge valve side; Waveform difference between phase B AC on bridge valve side; Waveform difference between phase C AC on bridge valve side; Waveform difference between phase C AC on bridge valve side; Waveform difference between phase C AC on bridge valve side; Waveform difference between phase A AC on bridge valve side; Waveform difference between phase C ... The waveform difference between DC current on the bus valve side, DC current on the neutral bus valve side, and AC current on the B phase of the D-bridge valve side; the waveform difference between DC current on the DC pole bus valve side, DC current on the neutral bus valve side, and AC current on the C phase of the D-bridge valve side under commutation failure protection; the waveform difference between DC current on the DC pole bus valve side, DC current on the neutral bus valve side, DC current on the Y-bridge valve side, AC current on the A phase of the D-bridge valve side; the waveform difference between DC current on the DC pole bus valve side, DC current on the neutral bus valve side, DC current on the Y-bridge valve side, AC current on the B phase of the D-bridge valve side; the waveform difference between DC current on the DC pole bus valve side, DC current on the neutral bus valve side, DC current on the Y-bridge valve side, AC current on the B phase of the D-bridge valve side; the waveform difference between DC current on the DC pole bus valve side, DC current on the neutral bus valve side, AC current on the Y-bridge valve side, AC current on the C phase of the D-bridge valve side; and converter short circuit. The waveform difference between the DC current on the DC pole bus valve side of the circuit protection, the DC current on the DC neutral bus valve side, the AC current on the Y-bridge valve side (phase A) and the AC current on the D-bridge valve side (phase A) is 1 / 2. The waveform difference between the DC current on the DC pole bus valve side of the converter short-circuit protection, the DC current on the DC neutral bus valve side, the DC current on the Y-bridge valve side (phase B) and the AC current on the D-bridge valve side (phase B) is 1 / 2. The waveform difference between the DC current on the DC pole bus valve side of the converter short-circuit protection, the DC current on the DC neutral bus valve side, the DC current on the Y-bridge valve side (phase C) and the AC current on the D-bridge valve side (phase C) is 1 / 2. The absolute value of the waveform difference between the DC current on the DC pole bus valve side and the DC current on the DC neutral bus valve side of the DC differential protection, and the absolute value of the maximum peak value of the DC line voltage for DC backup differential protection.The following are examples of overcurrent protection technologies: DC current on the DC pole bus valve side - DC current on the DC neutral bus valve side - A-phase AC current on the Y-bridge valve side - A-phase AC current on the D-bridge valve side; DC current on the DC pole bus valve side - DC current on the DC neutral bus valve side - B-phase AC current on the Y-bridge valve side - B-phase AC current on the D-bridge valve side; DC current on the DC pole bus valve side - DC current on the DC neutral bus valve side - C-phase AC current on the Y-bridge valve side - C-phase AC current on the D-bridge valve side; DC overvoltage protection: DC line voltage - absolute value of the waveform difference between high and low voltage valve group connection voltages; DC overvoltage protection: absolute value of the waveform difference between high and low voltage valve group connection voltages and neutral bus voltages; DC line traveling wave protection: maximum peak absolute value of common-mode transient; DC line traveling wave protection: maximum peak absolute value of common-mode integral; DC line traveling wave protection: maximum peak absolute value of differential-mode integral; DC undervoltage protection: DC line voltage. The following parameters are considered: Degree of deviation from steady state; Absolute value of the difference between the DC line voltage and the high / low voltage valve group connection voltage waveforms for DC undervoltage protection; Absolute value of the maximum peak value of the DC line voltage for DC undervoltage protection; Absolute value of the difference between the waveforms of the other pole line current and the DC current on the other pole line side for metallic loop differential protection; Absolute value of the maximum peak value of the DC line voltage for DC line sudden change protection; Maximum value of the slope of the DC voltage waveform for DC line sudden change protection; Absolute value of the difference between the waveforms of the other pole line current and the DC current on the other pole line side for DC line differential protection; Absolute value of the maximum peak value of the 50Hz component of the DC current on the neutral bus valve side for 50Hz protection; Absolute value of the maximum peak value of the DC voltage for DC line undervoltage protection; Absolute value of the maximum peak value of the neutral bus voltage for grounding electrode open circuit protection; Absolute value of the maximum peak value of the DC current on the other pole line side for grounding electrode open circuit protection; and Absolute value of the maximum peak value of the current between grounding electrode 1 and grounding electrode 2 for grounding electrode open circuit protection.

[0037] In this specific implementation, the TabPFN model, CatBoost model, XGBoost model, NaiveBayes model, LogReg model, GBDT model, SVM model, and RandomForest model are all existing models.

[0038] The following is a brief description of the feature mining and anomaly methods of various existing models.

[0039] 1. Methods based on TabPFN deep models

[0040] 1) Input features: a unified feature set;

[0041] 2) Feature preprocessing: Standardization;

[0042] 3) Model structure: A table-based classification model based on Transformer;

[0043] 4) Model training: Five-fold cross-validation;

[0044] 5) Output prediction: Transformer directly outputs the category.

[0045] 2. CatBoost-based method

[0046] 1) Input features: a unified feature set;

[0047] 2) Feature preprocessing: Automatically process category features;

[0048] 3) Model structure: Gradient boosting tree, supporting categorical features;

[0049] 4) Hyperparameter settings: Learning rate = 0.03, Number of iterations = 1000;

[0050] 5) Model training: Five-fold cross-validation;

[0051] 6) Output prediction: weighted tree voting.

[0052] 3. XGBoost-based methods

[0053] 1) Input features: a unified feature set;

[0054] 2) Feature preprocessing: numerical standardization;

[0055] 3) Model structure: Gradient boosting tree;

[0056] 4) Hyperparameter settings: Learning rate = 0.1, Number of trees = 100, Maximum depth = 5;

[0057] 5) Model training: Five-fold cross-validation;

[0058] 6) Output prediction: Tree ensemble voting.

[0059] 4. Naive Bayes-based methods

[0060] 1) Input features: a unified feature set;

[0061] 2) Feature preprocessing: Standardizing continuous features;

[0062] 3) Model assumptions: Features are independently distributed;

[0063] 4) Model type: Gaussian Naive Bayes;

[0064] Hyperparameter setting: Variance smoothing = 1e-1;

[0065] 5) Model training: Calculate conditional probabilities;

[0066] 6) Output prediction: the class with the highest posterior probability.

[0067] 5. Logistic Regression-Based Methods

[0068] 1) Input features: a unified feature set;

[0069] 2) Feature preprocessing: Standardization;

[0070] 3) Model structure: Multi-class logistic regression;

[0071] 4) Hyperparameter settings: Maximum number of iterations = 1000, regularization strength = 0.01;

[0072] 5) Model training: Five-fold cross-validation;

[0073] 6) Output prediction: softmax outputs the probability, selecting the category with the maximum value.

[0074] 6. Gradient Boosting Decision Tree (GBDT) based method

[0075] 1) Input features: a unified feature set;

[0076] 2) Feature preprocessing: numerical feature standardization;

[0077] 3) Model structure: Iterative improvement using multiple weak learners;

[0078] 4) Hyperparameter settings: learning rate = 0.05, number of estimators = 200, maximum depth = 2, subsampling ratio = 0.7, minimum leaf sample = 2, random seed = 42;

[0079] 5) Model training: Five-fold cross-validation;

[0080] 6) Output prediction: Weighted voting determines the category.

[0081] 7. Support Vector Machine (SVM) based methods

[0082] 1) Input features: a unified feature set;

[0083] 2) Feature preprocessing: Input after standardization (StandardScaler);

[0084] 3) Kernel function: Radial basis function (RBF);

[0085] 4) Hyperparameter settings: Kernel function = radial basis function, penalty coefficient = 10, kernel parameter = scale, enable probability estimation;

[0086] 5) Model training: Five-fold cross-validation;

[0087] 6) Output prediction: Use the decision function threshold of 0.0 to classify the categories.

[0088] 8. Random Forest-based methods

[0089] 1) Input features: a unified feature set;

[0090] 2) Feature preprocessing: Numerical features are standardized, and categorical features are directly input;

[0091] 3) Model structure: Composed of multiple decision trees;

[0092] 4) Hyperparameter settings: Number of decision trees = 200, maximum tree depth = 3, minimum leaf sample = 2, enable out-of-bag scores, random seed = 42;

[0093] 5) Model training: Five-fold cross-validation;

[0094] 6) Output prediction: the majority class of all decision trees vote.

[0095] In this specific implementation, step S1 is the same for all models. The number of parameters and the model size of Qwen2.5_VL are selected as 32 bytes. Figure 1 The generated UHVDC dataset was used for fault diagnosis, and the diagnostic performance of different models was compared. The results are shown in Table 1.

[0096] Table 1. Performance comparison of different UHVDC fault diagnosis models (average of five-fold cross-validation)

[0097] Model accuracy Accuracy Recall rate F1 score RandomForest 0.6545 0.6422 0.6545 0.6074 SVM 0.7455 0.7862 0.7455 0.7388 GBDT 0.6786 0.7071 0.6786 0.6657 LogReg 0.7273 0.7470 0.7273 0.7252 NaiveBayes 0.7636 0.7930 0.7636 0.7615 XGboost 0.7455 0.7581 0.7455 0.7407 catboost 0.7091 0.7455 0.7091 0.6927 TabPFN 0.7818 0.7906 0.7818 0.7773 EAPCR 0.7818 0.7952 0.7818 0.7823

[0098] As shown in Table 1, Qwen2.5_VL has good practicality with SVM, NaiveBayes, catboost, TabPFN and EAPCR models, with accuracy rates all above 0.74. In particular, when combined with the EAPCR model, the accuracy, precision, recall and F1 score are the highest, indicating that this diagnostic method has strong diagnostic capabilities for UHVDC faults.

[0099] To verify the effectiveness of keyword extraction in the unified stage + targeted stage, the Qwen2.5_VL parameter count and model size were set to 32 bytes. The accuracy of each model is shown in Table 2.

[0100] Table 2. Ablation Comparison of Multimodal Feature Data Extracted Unified by Qwen-2.5-VL Cue Words

[0101]

[0102] As shown in Table 2, when using the large model Qwen2.5-VL, the accuracy of fault diagnosis is improved by adding the directional feature stage.

[0103] The above embodiments are only some preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-substantial changes and substitutions made by those skilled in the art based on the present invention shall fall within the scope of protection claimed by the present invention.

Claims

1. A method of diagnosing human-interactive UHVDC faults, characterized in that, Includes the following steps: S1. Input the original data into the Qwen2.5_VL large model, and extract the features of the original data corresponding to each keyword in a unified stage and a targeted stage respectively. The unified stage extracts based on the unified keyword, and the targeted stage extracts based on the targeted keyword. Discretize the features extracted in the unified stage and the targeted stage to form a matrix X. S2. Input the matrix X into any one of the TabPFN model, EAPCR model, SVM model, NaiveBayes model, and XGboost model, and then output the diagnostic results.

2. The method of claim 1, wherein the method further comprises: The EAPCR model maps the matrix X into a dense representation E by feeding it into a learnable embedding layer; then it uses a bilinear attention mechanism to calculate the relationship between any two features to obtain matrix A. Then, the permutation matrix P of matrix A is introduced, and matrices A and P are processed in parallel on a dual-branch lightweight convolutional network to extract local and non-local interaction features respectively. The outputs of the two branches are then flattened and concatenated. Finally, the concatenated vector is mapped to a five-dimensional soft-max probability through a residual-connected multilayer perceptron to output fault diagnosis results of 0-4 classes.

3. The method of claim 2, wherein the method further comprises: The unified keywords include unified general keywords, unified special keywords, and unified fixed keywords; the unified general keywords consist of a unified general waveform and a unified general prompt word. The unified general waveform includes DC line voltage, neutral bus voltage, DC current on the DC pole bus valve side, DC current on the DC neutral bus valve side, DC current on the DC pole bus line side, grounding electrode 1 current, grounding electrode 2 current, the 50Hz component of the DC line voltage, the 50Hz component of the DC pole bus line side DC current, and the pole... The difference between bus current and neutral bus current, DC differential protection stage 1 setting, DC differential protection stage 2 setting, common-mode transient, common-mode integral, differential-mode integral, derivative 1 of DC line voltage, derivative 2 of DC line voltage, maximum value of three-phase AC current on the Y-bridge valve side, maximum value of three-phase AC current on the D-bridge valve side, and connection voltage of high and low voltage valve groups; the unified general keywords include maximum positive peak value, maximum negative peak value, maximum slope peak value, abnormal duration ratio, transient synchronicity, number of oscillations, baseline drift, and noise glitches; the unified special keywords The term consists of a unified special waveform and a unified special keyword. The unified special waveform includes the AC voltage of phase A, phase B, and phase C; the AC voltage of phase A on the Y-bridge valve side; the AC voltage of phase B on the Y-bridge valve side; the AC voltage of phase C on the Y-bridge valve side; the AC voltage of phase A on the D-bridge valve side; the AC voltage of phase B on the D-bridge valve side; and the AC voltage of phase C on the D-bridge valve side. The unified special keyword includes the maximum positive peak value, the maximum negative peak value, the maximum slope peak value, the percentage of abnormal duration, the synchronization of abrupt changes, the number of oscillations, the baseline drift, noise glitches, and the symmetry of positive and negative peak values. Sex and half-cycle symmetry; unified fixed keywords include three-phase unbalance of AC voltage, three-phase unbalance of AC power on the Y-bridge valve side, three-phase unbalance of AC power on the D-bridge valve side, phase hysteresis of AC voltage, phase hysteresis of AC power on the Y-bridge valve side, phase hysteresis of AC power on the D-bridge valve side, maximum positive peak value difference of AC voltage, maximum positive peak value difference of AC power on the Y-bridge valve side, maximum positive peak value difference of AC power on the D-bridge valve side, maximum negative peak value difference of AC voltage, maximum negative peak value difference of AC power on the Y-bridge valve side, and maximum negative peak value difference of AC power on the D-bridge valve side.

4. The diagnostic method for human-computer interactive UHVDC faults according to claim 3, characterized in that, The directional keywords include maximum positive peak value / maximum negative peak value of 100Hz protection DC neutral bus valve side DC current, maximum positive peak value / maximum negative peak value of 100Hz component of 100Hz protection DC neutral bus valve side DC current, all waveform maximum positive peak value / all waveform maximum negative peak value / all waveform difference absolute value of ground pole bus differential protection station ground wire current-ground pole line 1 current-ground pole line 2 current-another pole line route side DC current, waveform difference 1 of Y bridge valve side A phase alternating current-D bridge valve side A phase alternating current of bridge difference protection, waveform difference 1 of Y bridge valve side B phase alternating current-D bridge valve side B phase alternating current of bridge difference protection, waveform difference 1 of Y bridge valve side C phase alternating current-D bridge valve side C phase alternating current of bridge difference protection, waveform difference 2 of Y bridge valve side A phase alternating current-D bridge valve side A phase alternating current of bridge difference protection, waveform difference 2 of Y bridge valve side C phase alternating current-D bridge valve side C phase alternating current of bridge difference protection, waveform difference between Y bridge valve side A phase alternating current and D bridge valve side A phase alternating current of commutation failure protection DC pole bus valve side DC current-DC neutral bus valve side DC current, waveform difference between Y bridge valve side B phase alternating current and D bridge valve side B phase alternating current of commutation failure protection DC pole bus valve side DC current-DC neutral bus valve side DC current, waveform difference between Y bridge valve side C phase alternating current and D bridge valve side C phase alternating current of commutation failure protection DC pole bus valve side DC current-DC neutral bus valve side DC current, waveform difference between D bridge valve side A phase alternating current and Y bridge valve side A phase alternating current of commutation failure protection DC pole bus valve side DC current-DC neutral bus valve side DC current, waveform difference between D bridge valve side B phase alternating current and Y bridge valve side B phase alternating current of commutation failure protection DC pole bus valve side DC current-DC neutral bus valve side DC current, waveform difference between D bridge valve side C phase alternating current and Y bridge valve side C phase alternating current of commutation failure protection DC pole bus valve side DC current-DC neutral bus valve side DC current, waveform difference between Y bridge valve side A phase alternating current and D bridge valve side A phase alternating current of group difference protection DC pole bus valve side DC current-DC neutral bus valve side DC current, waveform difference between Y bridge valve side B phase alternating current and D bridge valve side B phase alternating current of group difference protection DC pole bus valve side DC current-DC neutral bus valve side DC current, waveform difference between Y bridge valve side C phase alternating current and D bridge valve side C phase alternating current of group difference protection DC pole bus valve side DC current-DC neutral bus valve side DC current, waveform difference 1 / waveform difference 2 between Y bridge valve side A phase alternating current and D bridge valve side A phase alternating current of converter short circuit protection DC pole bus valve side DC current-DC neutral bus valve side DC current, waveform difference 1 / waveform difference 2 between Y bridge valve side B phase alternating current and D bridge valve side B phase alternating current of converter short circuit protection DC pole bus valve side DC current-DC neutral bus valve side DC current, waveform difference 1 / waveform difference 2 between Y bridge valve side C phase alternating current and D bridge valve side C phase alternating current of converter short circuit protection DC pole bus valve side DC current-DC neutral bus valve side DC current, waveform difference absolute value of DC pole bus valve side DC current-DC neutral bus valve side DC current of DC differential protection, maximum peak value absolute value of DC line voltage of DC backup differential protection,AC / DC overcurrent protection DC pole busbar valve side DC current - DC neutral busbar valve side DC current - Y bridge valve side A phase AC current - D bridge valve side A phase AC current maximum peak absolute value, AC / DC overcurrent protection DC pole busbar valve side DC current - DC neutral busbar valve side DC current - Y bridge valve side B phase AC current - D bridge valve side B phase AC current maximum peak absolute value, AC / DC overcurrent protection DC pole busbar valve side DC current - DC neutral busbar valve side DC current - Y bridge valve side C phase AC current - D bridge valve side C phase AC current maximum peak absolute value, DC overvoltage protection DC line voltage - high-low pressure valve group connection voltage waveform difference absolute value, DC overvoltage protection high-low pressure valve group connection voltage - neutral bus voltage waveform difference absolute value, DC line traveling wave protection common mode abruptness maximum peak absolute value, DC line traveling wave protection common mode integral quantity maximum peak absolute value, DC line traveling wave protection differential mode integral quantity maximum peak absolute value, DC low voltage protection DC line voltage relative to steady state deviation degree, DC low voltage protection DC line voltage - high-low pressure valve group connection voltage waveform difference absolute value, DC low voltage protection DC line voltage maximum peak absolute value, metal return longitudinal difference protection opposite station another pole line current - another pole line side DC current waveform difference absolute value, DC line abruptness protection DC line voltage maximum peak absolute value, DC line abruptness protection DC line voltage waveform slope maximum value, DC line longitudinal difference protection opposite station another pole line current - another pole line side DC current waveform difference absolute value, 50Hz protection DC neutral busbar valve side DC current 50Hz component maximum peak absolute value, DC line low voltage protection DC line voltage maximum peak absolute value, ground pole open circuit protection neutral bus voltage maximum peak absolute value, ground pole open circuit protection another pole line side DC current maximum peak absolute value, and ground pole open circuit protection ground pole line 1 current - ground pole line 2 current maximum peak absolute value.

5. The diagnostic method for human-computer interactive UHVDC faults according to claim 1, characterized in that, The keywords were defined by ultra-high voltage direct current experts based on electrical principles and related knowledge.

6. The diagnostic method for human-computer interactive UHVDC faults according to claim 1, characterized in that, When extracting features during the targeted phase, the setting for features that are automatically executed but have not triggered the protection execution process is 0.

Images