Converter valve fault diagnosis method and system based on state analysis

Abstract

The invention discloses a converter valve fault diagnosis method and system based on state analysis. Acquiring operation data of a converter valve thyristor assembly, a valve cooling assembly and a valve lightning arrester, and processing the non-numeric state quantity into numeric data according to the influence degree of the non-numeric state quantity on the operation of the converter valve; constructing a multi-criterion feature selection method for the preprocessed operation data to screen feature subsets in the operation data; based on the data samples of the screened feature subsets, constructing a converter valve fault diagnosis model by adopting a dual-channel hybrid neural network in combination with multi-task branch output; operating data of the converter valve are collected in real time, and after preprocessing and feature screening, the fault level distribution probability of each assembly is output through the converter valve fault diagnosis model. According to the converter valve fault diagnosis method provided by the invention, intelligent and refined diagnosis and division of the overall operation state of the converter valve can be realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of power equipment monitoring, and particularly relates to a converter valve fault diagnosis method and system based on state analysis. BACKGROUND

[0002] The converter valve is a core device of the extra-high voltage direct current transmission and plays an important role in rectification and inversion. If problems such as commutation failure occur, it may lead to a decrease in the voltage on the direct current side and an increase in the current, and even cause power transmission interruption in severe cases, endangering the safe operation of the power grid. In the past few years, many researchers have actively explored various methods for fault diagnosis and prediction of converter valve components.

[0003] The patent application with publication number CN117746197A captures converter valve images through three different light sources (such as visible light, infrared, and ultraviolet), extracts and fuses multiple features (such as convolution layers and feature splicing dimension reduction), generates detailed fusion features, and finally outputs fault diagnosis results by a prediction module. The patent application with publication number CN119513744A proposes a converter valve fault diagnosis method based on non-electrical data, which can improve the reliability and anti-interference ability of fault diagnosis through modal decomposition and feature extraction of non-electrical signals. The patent application with publication number CN113762486A discloses a converter valve fault diagnosis model construction method based on a convolution gate recurrent neural network, which utilizes the end-to-end feature extraction capability of deep neural networks to automatically learn the time sequence features in the converter valve operation signals, and enhances the recognition ability of the model to long-term dependence based on the gating mechanism, improving the generalization performance of the diagnosis model under complex working conditions.

[0004] However, the existing technology ignores the comprehensive features of complex time sequence dependence and time-frequency structure in the converter valve operation data, making it difficult to adapt to the needs of long-time, multi-state, and multi-source heterogeneous data fusion, limiting the generalization ability and sensitivity of the diagnosis model, and further affecting the accurate evaluation of the running state of the converter valve, and further affecting the evaluation of the running state. SUMMARY

[0005] To solve the problems in the prior art, the application provides a converter valve fault diagnosis method and system based on state analysis. Running data of a converter valve thyristor assembly, a valve cooling assembly and a valve lightning arrester are collected, and non-numerical state quantities are processed into numerical data according to their influence on the running of the converter valve. After the preprocessed running data, a multi-criteria feature selection method is used to screen a feature subset in the running data. Based on the data samples screened by the feature subset, a double-channel hybrid neural network combined with a multi-task branch output is used to construct a converter valve fault diagnosis model. Real-time running data of the converter valve are collected, and after preprocessing and feature screening, the converter valve fault diagnosis model is used to output the fault level distribution probability of each component.

[0006] The first aspect of the application provides a converter valve fault diagnosis method based on state analysis, which adopts the following technical scheme:

[0007] The running data of the converter valve thyristor assembly, the valve cooling assembly and the valve lightning arrester are collected, and the non-numerical state quantities are processed into numerical data according to their influence on the running of the converter valve.

[0008] After the preprocessed running data, a multi-criteria feature selection method is used to screen a feature subset in the running data.

[0009] Based on the data samples screened by the feature subset, a double-channel hybrid neural network combined with a multi-task branch output is used to construct a converter valve fault diagnosis model.

[0010] Real-time running data of the converter valve are collected, and after preprocessing and feature screening, the converter valve fault diagnosis model is used to output the fault level distribution probability of each component.

[0011] Further, the running data includes numerical state quantities and non-numerical state quantities.

[0012] For each non-numerical state quantity, an initial score is given according to its influence on the safe running of the converter valve, and a level is drawn. Each level corresponds to a different weight.

[0013] The level weight of the non-numerical state quantity is multiplied by 10 to calculate the maximum score of the non-numerical state quantity. The ratio of the initial score of the non-numerical state quantity to the maximum score is taken as the score of the non-numerical state quantity, and the non-numerical state quantity is processed into numerical data.

[0014] Further, for the damping capacitor, the component capacitor and the voltage equalizing capacitor in the thyristor assembly running data:

[0015] The thyristor level is equivalent to a combined circuit model containing a damping circuit, a voltage equalizing circuit and a power taking circuit.

[0016] By applying DC voltage signals and two frequency AC signals at both ends of the thyristor level respectively, the DC impedance and the comprehensive impedance at two frequencies are calculated and obtained;

[0017] A fault tree model of the converter valve is constructed, and the DC impedance and the comprehensive impedance of the equivalent circuit of the thyristor level are mapped to the fault tree model, so that the numerical detection of the damping capacitor, the component capacitor and the grading capacitor of the thyristor assembly is realized without disconnection.

[0018] Further, the feature subset obtaining step comprises:

[0019] The isolated forest is used to perform anomaly detection and cleaning on the operation data, an abnormal score threshold is set to retain samples to form a sample set, and the sample set after cleaning is classified according to states, and a sample category balancing operation is performed;

[0020] For the feature state quantity in the data sample, the distance measure between comprehensive features, the feature distribution characteristics and the feature local preservation ability are comprehensively considered, the importance of each feature state quantity is calculated, and the features with importance higher than a threshold are retained to form a feature subset.

[0021] Further, the construction process of the converter valve fault diagnosis model comprises:

[0022] The data sample screened by the feature subset is taken as input data, and the dimension is

[0023] The is input into the Transformer branch to capture long-term and short-term time sequence dependencies and dynamic change characteristics; simultaneously, the is converted into a two-dimensional time-frequency image sequence by continuous wavelet transform, and is input into the CNN branch to extract multi-scale time-frequency local features;

[0024] The outputs of the two channels are spliced into a fusion feature vector, and are divided into branch fusion feature vectors corresponding to the thyristor assembly, the valve cooling assembly and the valve arrester;

[0025] Each branch fusion feature vector of each component corresponds to a task branch, and each task branch passes through a fully connected layer and a softmax to output a probability distribution of a fault level corresponding to each component.

[0026] Further, the Transformer branch uses a bidirectional GRU to extract bidirectional dynamic time features in X, and accesses a multi-head self-attention encoder to capture global time sequence dependency patterns and generate time sequence context embedding vectors;

[0027] The CNN branch is provided with a plurality of parallelly arranged multi-scale convolution kernels to extract local features of the time-frequency image, and the parallel structure of the multi-scale convolution kernels comprises at least two convolution kernel branches with different sizes;

[0028] A deep separable convolution layer is connected after the multi-scale convolution to reduce the complexity of the model; and a middle-high layer feature extraction network is constructed by stacking the reverse residual structure;

[0029] A channel attention mechanism is embedded in the reverse residual structure to generate a mode embedding feature vector of a local time-frequency structure in the time-frequency image.

[0030] Further, the output of the two channels is spliced into a fusion feature vector, comprising:

[0031] The time sequence context embedding vector output by the Transformer branch and the mode embedding feature vector output by the CNN branch are respectively subjected to global average pooling to obtain respective corresponding global channel features;

[0032] The global channel features of the two branches are input into a shared two-layer fully connected neural network to generate channel attention weight coefficients of the global channel features of the two branches;

[0033] The time sequence context embedding vector and the mode embedding feature vector are respectively multiplied by the respective channel attention weight coefficients to generate a fusion feature vector through weighted combination.

[0034] Further, a component index identifier is introduced to divide the fusion feature vector into branch fusion feature vectors corresponding to the thyristor component, the valve cooling component and the valve arrester;

[0035] The branch fusion feature vector of each component corresponds to a task branch, and in each task branch, the branch fusion feature vector is subjected to global average pooling to further compress the dimension, and then is output through a fully connected layer and a softmax to output a probability distribution of the fault grade corresponding to the component.

[0036] Further, a converter valve fault diagnosis model is used to diagnose the real-time converter valve operation data, comprising:

[0037] The operation data of each component of the converter valve are collected in real time, and after non-numerical state quantity processing, pretreatment and feature subset screening, the converter valve fault diagnosis model is used to output the probability distribution of each fault grade of the thyristor component, the valve cooling component and the valve arrester;

[0038] According to the probability distribution output by the converter valve fault diagnosis model, the maximum probability principle is used to determine the fault grade of each component, and each component corresponds to an independent fault grade diagnosis result.

[0039] The second aspect of the application provides a state analysis-based converter valve fault diagnosis system, which uses the converter valve fault diagnosis method provided by the first aspect of the application.

[0040] A data acquisition and processing module is configured to acquire operation data of the converter valve thyristor assembly, the valve cooling assembly and the valve arrester, and process non-numerical state quantities into numerical data according to their influence on the operation of the converter valve.

[0041] A state quantity feature screening module is configured to construct a multi-criteria feature selection method to screen a feature subset from the preprocessed operation data.

[0042] A model construction and reasoning module is configured to construct a converter valve fault diagnosis model based on the screened data samples, and adopt a dual-channel hybrid neural network combined with a multi-task branch output mode.

[0043] A model running and classification module is configured to acquire real-time operation data of the converter valve, and output a fault level distribution probability of each component by using the converter valve fault diagnosis model after preprocessing and feature screening.

[0044] The application has the following advantages compared with the prior art,

[0045] 1. The application comprehensively acquires numerical and non-numerical state quantities of three components (thyristor assembly, valve cooling assembly and valve arrester) of the converter valve, covering key operation parameters such as damping capacitor, component capacitor, voltage equalization capacitor, cooling flow rate, inlet and outlet water temperature difference, voltage offset, arrester leakage current, etc. By introducing a quantitative processing mechanism for non-numerical state quantities, the non-numerical state quantities are converted into unified numerical indicators, thereby enhancing the fusion ability and time-sensitive perception ability of multi-source heterogeneous data. A hanging multi-task fault diagnosis model constructed based on these operation data can independently output probability distribution of fault levels of the three components based on input data, thereby improving the fault recognition granularity, response efficiency and diagnosis accuracy of the converter valve system, and providing strong data support for fine operation and maintenance and preventive maintenance.

[0046] 2、The application proposes a dual-channel hybrid neural network architecture, which uses two branches to model the temporal context and local time-frequency patterns respectively. The Transformer branch uses bidirectional GRU and multi-head self-attention mechanism to accurately extract long-term and short-term dependencies and mutation features in the component operating state; the CNN branch extracts abnormal patterns in the local time-frequency image through continuous wavelet transform and multi-scale convolution structure, and combines channel attention mechanism and reverse residual structure to improve the expression ability of local features. The outputs of the two channels are weighted and fused and divided by component index, corresponding to each task branch, realizing end-to-end structural decoupling, multi-scale expression and feature alignment. This method not only improves the recognition ability of different types of faults, but also significantly enhances the generalization, robustness and adaptability of the model. BRIEF DESCRIPTION OF DRAWINGS

[0047] Figure 1 Flowchart of the fault diagnosis method provided for the embodiment;

[0048] Figure 2 Equivalent circuit diagram of thyristor;

[0049] Figure 3 Fault tree diagram provided for the embodiment;

[0050] Figure 4 Training accuracy diagram under different iteration numbers;

[0051] Figure 5 Model recall rate comparison diagram;

[0052] Figure 6 Confusion matrix of original data;

[0053] Figure 7 Confusion matrix of data processed by SMOTE;

[0054] Figure 8 Confusion matrix of data processed by Isolation Forest-Smote. DETAILED DESCRIPTION

[0055] To make the purpose, technical scheme and advantages of the present application clearer, the technical scheme of the present application will be described clearly and completely below in combination with the drawings in the embodiments of the present application. The embodiments described in the present application are only a part of the embodiments of the present application, not all the embodiments. Based on the spirit of the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0056] The present embodiment provides a converter valve fault diagnosis method based on state analysis and hybrid neural network algorithm, referring to Figure 1 , the steps include:

[0057] S1: Obtain and preprocess the operating data of each component of the converter valve. Specifically:

[0058] 1.1: Set the time interval for data sampling, collect the operating data of the three major components of the converter valve according to the operating state quantities recorded in the current industry guidelines, including the thyristor component, the valve cooling component, and the valve surge arrester, covering structural, thermal, electrical, acoustic vibration characteristics, alarm history, maintenance records, etc., which are divided into numerical and non-numerical types. In this embodiment, 39 operating state quantities are collected as the characteristics of each data collection.

[0059] In this embodiment, the sampling interval is set to 15 minutes, and Table 1 shows some examples of operating state quantities.

[0060] Table 1: Partial examples of operating state quantities of the converter valve

[0061]

[0062] Among the 39 state quantities, the differences between the damping capacitance and the initial value, the component capacitance and the initial value, and the grading capacitance and the initial value are quickly determined. Specifically:

[0063] 1.1.1: The thyristor stage is equivalent to a combined circuit model containing damping circuit, grading circuit, and energy taking circuit; without disassembling the internal connections, a certain amplitude of DC voltage signal is applied across the thyristor stage; since the capacitor has high resistance to DC, only the grading circuit is turned on at this time, and the damping and energy taking circuits are in open circuit state due to the blocking of the capacitor.

[0064] Therefore, the applied voltage U DC and the measured current I DC are collected, and the DC impedance Z DC is calculated according to Ohm's law ; the DC impedance reflects the state of the grading circuit resistance.

[0065] 1.1.2: On the same circuit port as 1.1.1, a 100Hz and a 100kHz sinusoidal AC signal are applied respectively, and the voltage and current peak values U 100Hz , I 100Hz , U 100kHz and I 100kHz at the two frequencies are collected.

[0066] The comprehensive impedances Z 100Hz and Z 100kHz at the two frequencies are calculated respectively:

[0067]

[0068] 1.1.3: Based on the impedances calculated above, according to and The relationship between reactive power and active power and impedance (resistance and reactance) can be established, and thus a plurality of sets of active and reactive power response equations can be established, where X represents impedance; the total equivalent resistance and capacitance of the loop are solved by simultaneously solving the equations to determine whether they deviate from the design values.

[0069] Referring to Figure 2 , the equivalent circuit of the thyristor stage in Figure 2 is taken as an example, the total equivalent resistance and capacitance of the loop are solved by the following formulas;

[0070]

[0071] Among them, A2, A3, B1, B2 are intermediate variables set for convenient calculation:

[0072]

[0073] In the above algebraic equation set, P and Q are active power and reactive power, ω is the angular frequency, U is the total voltage; R1, R2, C1 and C2 are the resistance and capacitance of the damping loop in Figure 2 , R3 and C3 are the resistance and capacitance of the energy extraction loop in Figure 2 , R dc1 and R dc2 are the resistances of the voltage equalization loop in Figure 2 .

[0074] 1.1.3: Introduce the converter valve fault tree model (FTA), as shown in Figure 3 ; use the model to associate impedance abnormalities with typical fault elements;

[0075] Map the DC impedance, 100Hz, 100kHz impedance as input indicators to the FTA model, and deduce the possible fault paths and fault elements. For example, the mapping rules are as follows:

[0076] (1) Voltage equalization loop mapping rule: if in the DC impedance test, the measured voltage equalization loop impedance Z DC deviates from the set reference beyond the limit, then the "voltage equalization resistance failure" branch in the fault tree is triggered.

[0077] (2) Damping loop mapping rule: if in the 100Hz test, the capacitive impedance is abnormal (such as current weakening, capacitive reactance mutation), then map the "damping loop capacitance failure" node in the fault tree; if in the 100kHz test, the resistive impedance is abnormal (such as abnormal high resistance), then map to the "damping loop resistance failure" branch.

[0078] (3) The mapping rule of the energy extraction loop: if there is a capacitance anomaly in the 100Hz test or an impedance curve anomaly in the 100kHz test, the failure type is judged in combination with the design value of the branch, and is mapped to the corresponding capacitor and resistor nodes of the energy extraction loop.

[0079] Based on the method of multi-frequency impedance test + fault tree analysis (FTA), the failure type of the internal circuit below the thyristor level of the converter valve can be accurately located:

[0080] Damping circuit: through 100Hz (capacitive dominant) and 100kHz (resistive dominant) impedance tests, it can be judged whether the resistance (such as R2, R3) or capacitance (such as C2, C3) in the damping circuit has a short circuit, an open circuit, or a significant increase in resistance / capacitance.

[0081] Voltage equalization circuit: through direct current impedance (DC) test, it is judged whether the resistance element in the voltage equalization circuit has a failure (short circuit / open circuit / resistance value deviating from the nominal range), so as to realize the rapid discrimination of the voltage equalization circuit failure.

[0082] Energy extraction circuit: also relying on low-frequency / high-frequency impedance test, the abnormal condition of the resistance or capacitance in the energy extraction circuit is qualitatively analyzed under the premise of not disassembling the wire, and in combination with the FTA model, the energy extraction circuit and the damping circuit failure can be distinguished and located to the specific element.

[0083] In this embodiment, the thyristor equivalent circuit model is combined with fault tree analysis (FTA), and is supplemented by multi-frequency impedance test, which can realize efficient diagnosis of multiple functional circuits (such as damping, voltage equalization, and energy extraction) inside the converter valve. This method not only supports rapid measurement of the comprehensive impedance characteristics of key elements without disassembling the wire, but also can map the measured impedance value to the specific failure path in the FTA model, so as to accurately locate the position of the problem element. At the same time, for the parameter changes of the damping capacitance, the component capacitance, and the voltage equalization capacitance in the 39 state quantities, this method can realize rapid investigation and qualitative identification through equation set solving and fault tree linkage analysis, avoid tedious item-by-item test, and concentrate numerical detection in the suspect area, thereby significantly reducing the test cost and calculation complexity.

[0084] 1.2: The numerical state quantity and the non-numerical state quantity are classified; for the non-numerical state quantity such as family defect and rectification condition, valve tower main water leakage condition, and valve arrester rust condition, first, an expert rule base corresponding to the condition of each component of the converter valve is constructed, an initial score is given by comparison with the expert rule base, and is divided into four levels (normal, attention, abnormal, and serious). Each level corresponds to different weights.

[0085] Then, the s i = λ iX10get the maximum score of the ith state quantity; wherein, s i X10is the maximum score of the ith state quantity, and λ i is the grade weight corresponding to the ith state quantity.

[0086] Finally, the ratio of the initial score to the maximum score is taken as the score of the ith state quantity, and the abstract non-numeric state quantity is processed into numeric data.

[0087] S2: The numeric state quantities in the collected operation data are preprocessed through the Pandas library in Python, and then combined with the state quantities converted from non-numeric, to obtain a training data set. This includes:

[0088] 2.1: An abnormality detection method based on tree structure is adopted, a sub-sample is randomly extracted from the numeric collected data, and binary division is recursively performed, the abnormality degree is quantified based on the path length of the sample in the tree structure, so as to eliminate outliers; wherein, the abnormality score threshold is dynamically set according to the state grade (normal, attention, abnormal, serious) to which the sample belongs; Table 3 shows an example of threshold setting;

[0089] Table 3: Example of abnormal threshold setting

[0090] State level Abnormal score threshold Engineering basis Normal s>0.65 Strict cleaning (low tolerance for false positives) Caution s>0.55 Moderate reservation of edge samples Abnormal s>0.45 Loose handling (avoiding loss of real faults) Severe s>0.35 Maximum reservation of critical samples

[0091] 2.2: The class imbalance problem is further processed for the cleaned sample set; specifically:

[0092] The state grade is taken as the classification, the proportion of each class in the sample set is counted, and the target distribution is set to be balanced among the four classes; according to this, the expansion multiples required for the “attention”, “abnormal” and “serious” classes are calculated, and oversampling is used for sample expansion. The normal class samples are not subjected to oversampling.

[0093] S3: The state quantities during the operation of the converter valve are selected and divided; specifically:

[0094] The state quantity of the ultra-high voltage converter valve refers to various detectable or derivable variables that can reflect the working state, health degree and operation performance of the converter valve. However, the state quantities collected in this embodiment involve comprehensive equipment, but at the same time, there are problems such as high correlation between some state quantities and low contribution to the evaluation result. Therefore, a multi-criteria feature selection method including three types of feature selection algorithms is adopted to study the feature importance in this embodiment.

[0095] The multi-criteria feature selection method is to sort the importance of the above-mentioned 39 feature state quantities based on distance measurement, sample distribution characteristics, and feature local preservation ability, respectively, to obtain sequences I RF , I FS and I LS. Specifically,

[0096] In the training data set, each data represents the running state of the converter valve at a certain time, containing 39 characteristic state quantities.

[0097] For each characteristic state quantity, a score is obtained by calculating the distance difference (such as Euclidean distance or Manhattan distance) between the same class samples and different class samples, and the distance scores of all characteristics constitute I RF sequence;

[0098] For each characteristic state quantity, the mean difference between samples in different categories (inter-class difference) and the variance within the same category (intra-class difference) are calculated, and then the Fisher score is calculated by the ratio of inter-class difference to intra-class difference, and the Fisher scores of all characteristics constitute I FS sequence;

[0099] For each data sample, the local field is determined based on the similarity with other samples, and then the Laplacian score of each characteristic state quantity in the local field of the corresponding sample is calculated, and the Laplacian scores of all characteristics constitute I LS .

[0100] Then the importance I of each characteristic state quantity is calculated by the following calculation method:

[0101] I i =I RF,i +I FS,i +I LS,i , i = 1, 2, …, 39; wherein I i represents the importance of the i-th characteristic state quantity, I RF,i , I FS,i and I LS,i are the I RF sequence, I FS sequence and I FS sequence of the i-th characteristic state quantity.

[0102] The greater the value of I i , the more important the feature, which should be considered first in feature selection; otherwise, it means that the importance of the feature is lower, which can be discarded directly if necessary.

[0103] For each sample, the 39 characteristic state quantities are screened according to the importance threshold value, and the characteristics with importance higher than the threshold value are retained to form a feature subset, which is used as the input of the subsequent diagnostic model. The importance threshold value is set by the field professional according to the actual diagnostic requirements.

[0104] S4: Based on the sample screened by the feature subset, a converter valve fault diagnosis model is constructed; specifically:

[0105] 4.1: The data samples screened by the feature subset are composed into time series data samples as input data, with dimensions of where T is the length of the time window, and d is the number of features retained in the feature subset.

[0106] 4.2: The is input into the Transformer branch to capture long-term and short-term time series dependencies and dynamic change features; at the same time, the is converted into a two-dimensional time-frequency image sequence by continuous wavelet transform and input into the CNN branch to extract multi-scale time-frequency local features. Specifically:

[0107] For the Transformer branch, bidirectional GRU is used to extract bidirectional dynamic time features in X, and a multi-head self-attention encoder is connected to capture the global time series dependency mode of the running state of the three major components of the converter valve, and generate a time series context embedding vector;

[0108] For the CNN branch, multi-scale convolution kernels (such as 3x3, 5x5) are used to extract local features under different receptive fields in parallel, and depth separable convolution layers are introduced to reduce model complexity; in addition, reverse residual structures are stacked in the middle and high layers to enhance feature flow, and channel attention mechanisms are configured to enhance key response areas related to faults, and generate mode embedding feature vectors of local time-frequency structures in time-frequency images.

[0109] 4.3: The feature vectors output by each branch are spliced according to the feature dimensions to form a unified fused representation vector; wherein the channel weights of each branch feature are allocated through an attention weighted fusion mechanism to obtain a weighted comprehensive feature vector. Specifically:

[0110] The time series context embedding vector F T output by the Transformer branch and the mode embedding feature vector F C output by the CNN branch are respectively obtained by global average pooling (GAP);

[0111] The above GAP output vectors are input into a shared two-layer fully connected neural network (MLP) to generate channel attention weight coefficients ω T and ω C of the global channel features of the two branches, satisfying ω T + ω C = 1.

[0112] F T and F C are multiplied by their respective weights ω T and ω C , respectively, and then combined to generate a fused feature vector F.

[0113] 4.4: Introduce component index identification, divide the fusion feature vector F into the corresponding fusion feature vectors F of the thyristor component, valve cooling component and valve surge arrester thy , F cool and F arre ;

[0114] The fusion feature vector of the three components corresponds to one task branch respectively. In each task branch, the fusion feature vector is globally averaged and pooled to further compress the dimension. After passing through the fully connected layer and softmax, the probability distribution of the fault level corresponding to the component is output.

[0115] S5: Real-time acquisition of operation data of each element of the converter valve, after processing by steps S2 and S3, the converter valve fault diagnosis model is used to output the probability distribution of each fault level of the thyristor component, valve cooling component and valve surge arrester.

[0116] According to the probability distribution output by the model, the maximum probability principle is used to determine the fault level (normal, attention, abnormal, serious) of each component. Each component corresponds to an independent diagnosis result, and the corresponding strategy is set to cope with it.

[0117] For example, if the fault level of the thyristor component is "abnormal", the fault level of the valve cooling component is "attention", and the fault level of the valve surge arrester is "normal", the converter valve system can automatically push the thyristor inspection suggestion, monitor the running state of the valve cooling component, and maintain the current running state of the valve surge arrester unchanged.

[0118] As an optional step of this embodiment, when the fault level of a component reaches the attention level and above, an alarm mechanism is automatically triggered to send a warning notice to the maintenance system and the operator. The fault data and the diagnosis result are stored in the database synchronously to support subsequent analysis and tracking.

[0119] Combined with the maintenance record and the subsequent manual confirmation result, a feedback loop is constructed to update the training data set regularly, realize online fine-tuning of the model and performance improvement.

[0120] In order to verify the effectiveness of the method proposed in this patent, the historical data of the converter valve of a certain hydropower station in Jiangbei, Nanjing, Jiangsu Province is used in the experiment part.

[0121] Figure 4 The model accuracy under different data preprocessing conditions is shown, and the accuracy distribution of the original data exposes the fatal defect of class imbalance. The accuracy of some classes is as low as 0.2-0.4, reflecting the serious shortage of the model's recognition ability for minority classes. This phenomenon confirms the limitations of traditional classification methods when the data distribution is skewed-the model tends to "majority class hegemony", ignoring the feature expression of minority classes.

[0122] The introduction of SMOTE (Synthetic Minority Over-sampling Technique) provides a preliminary solution to this problem. By generating synthetic samples to balance the class distribution, the precision of the majority class is significantly improved to the range of 0.6-0.8. However, the limitations of SMOTE cannot be ignored: its mechanical sample generation mechanism may introduce noise, leading to "false prosperity" in some classes, i.e., the precision is improved while the risk of overfitting increases. For example, some classes may achieve a precision of 0.8, but in actual application, they may not perform stably due to noise interference.

[0123] To this end, the hybrid strategy of iForest SMOTE shows a breakthrough value. iForest SMOTE not only significantly increases the number of classes with high precision (>0.6), but also almost eliminates the extremely low value (<0.2). Especially in the wide class range of 0-341, the precision of some minority classes is close to 0.9. This proves that the synergy of anomaly detection and over-sampling can effectively balance the dual requirements of data distribution and data quality.

[0124] Figure 5 The recall rate under different data and processing conditions is shown. From the recall rate distribution of the original data (Original Data), the recall values of class 2 and class 3 are generally lower than 0.4, especially in some subclasses, close to 0.2 or even close to zero. This phenomenon directly reveals the severity of the class imbalance problem.

[0125] After introducing SMOTE (Synthetic Minority Over-sampling Technique), the recall rate shows local improvement. The chart data shows that the recall values of class 2 and class 3 are generally improved to the range of 0.4-0.6, and some subclasses even reach 0.8. This confirms the effectiveness of SMOTE in balancing data distribution through synthetic samples.

[0126] After iForest SMOTE processing, the recall rates of class 2 and class 3 further climb to the range of 0.6-0.8, and the extremely low value (<0.2) almost disappears. More importantly, the stability of high recall rate (>0.6) is significantly enhanced, indicating that the model reduces missed detection while avoiding performance fluctuations caused by noise. This proves that the iForest SMOTE strategy that combines anomaly detection and over-sampling significantly improves the recall ability of the model for minority classes, and is more robust than the traditional SMOTE method.

[0127] The confusion matrix of the original data (such as Figure 6) shows that there is a significant deviation between the distribution of true labels and predicted labels. The diagonal values of most classes (such as labels 0-3) are high, indicating that the model can better identify these classes; while the diagonal values of a few classes (such as labels 4-23) are very low, some even zero, indicating that the model can hardly capture their true characteristics, and a large number of samples are misclassified into the majority classes. This phenomenon directly exposes the severity of the class imbalance problem.

[0128] The confusion matrix (such as Figure 7 ) after SMOTE processing shows that the number of classes has expanded to 562, reflecting that the oversampling technique has expanded the data volume of the minority classes by synthesizing samples. The diagonal values of the minority classes have generally increased to a medium level (such as the number of correctly predicted samples for some labels increasing from 0 to 50-100), indicating that the model's recognition ability for the originally scarce classes has improved. However, the non-diagonal elements have also increased significantly, which implies that SMOTE may introduce noise or duplicate samples, causing the model to fall into overfitting in local areas and reducing the generalization performance.

[0129] The confusion matrix (such as Figure 8 ) after iForest-SMOTE processing shows better balance. After removing abnormal samples by the isolation forest algorithm, the noise interference is greatly reduced, the diagonal values of the minority classes are further improved (such as the number of correctly classified samples in the label interval 50-100 increasing by about 20%-30%), and the high misclassification area (such as the cross confusion between labels 200-300) is significantly shrunk. It is worth noting that the extreme misclassification (such as a class being completely misclassified as another class) almost disappears, indicating that iForest-SMOTE, while purifying data quality, retains the balancing advantage of SMOTE, enabling the model to capture the characteristics of minority classes while avoiding performance fluctuations caused by noise. The comparison between the original data and SMOTE confirms the limitations of simple oversampling - it solves the problem of data volume, but ignores data quality. iForest-SMOTE achieves the coordinated optimization of data distribution and data quality through the dual mechanisms of "denoising" and "oversampling". This strategy not only improves the classification accuracy of the model for high-dimensional unbalanced data, but also enhances the reliability of the results by reducing misclassification.

[0130] In the comparative experiment, the LSTM is 91.6%, which has a certain improvement compared with the traditional BPNN of 89.3%. BPNN performs well in solving non-linear problems and processing a small amount of data, but its learning ability for large-scale data is limited, and it cannot deeply mine complex data. SVM can extract fault features, but it cannot directly classify multiple faults, and needs to manually define thresholds, reducing the utilization of original data. The method of the present application can effectively solve the above problems.

[0131] Table 4 Evaluation index of prediction result

[0132]

[0133] In summary, according to the given evaluation index data, it can be concluded that the converter valve fault diagnosis method based on state analysis and hybrid neural network algorithm proposed in the application performs best and has the optimal prediction effect, which embodies the effectiveness of the method.

[0134] Finally, it should be noted that the above examples are only used to illustrate the technical solutions of the present application and not to limit it. Although the present application has been described in detail with reference to the above examples, those skilled in the art should understand that the specific embodiments of the present application can be modified or replaced equivalently without departing from the spirit and scope of the present application. Any modification or equivalent replacement without departing from the spirit and scope of the present application should be covered within the protection scope of the claims of the present application.

Claims

1. A converter valve fault diagnosis method based on state analysis, characterized in that: include: Collecting the operating data of the converter valve thyristor assembly, valve cooling assembly, and valve arrester, and processing the non-numeric state variables into numerical data based on their impact on the converter valve operation; For the pre-processed running data, a multi-criteria feature selection method is constructed to screen the feature subsets in the running data; Based on the data samples of the filtered feature subset, a converter valve fault diagnosis model is constructed by using a dual-channel hybrid neural network combined with multi-task branch output. The converter valve operation data is collected in real time, and after preprocessing and feature screening, the converter valve fault diagnosis model is used to output the fault level distribution probability of each component.

2. A converter valve fault diagnosis method based on state analysis according to claim 1, characterized in that: The operation data includes numerical state quantities and non-numerical state quantities; For each non-numerical state variable, an initial score is given according to its impact on the safe operation of the converter valve and the grades are divided into different levels with different weights. The maximum score of the non-numeric state quantity is calculated by multiplying the level weight of the non-numeric state quantity by 10; the ratio of the initial score of the non-numeric state quantity to the maximum score is used as the score of the non-numeric state quantity, and the non-numeric state quantity is processed into numerical data.

3. The converter valve fault diagnosis method based on state analysis according to claim 1, characterized in that: For the damping capacitor, component capacitance, and grading capacitor in the thyristor component operating data: The thyristor stage is equivalent to a combined circuit model containing a damping circuit, a voltage-sharing circuit, and an energy-taking circuit; By applying a DC voltage signal and two-frequency AC signals at both ends of the thyristor stage, the DC impedance and the comprehensive impedance at the two frequencies are calculated. A fault tree model of the converter valve is constructed, and the DC impedance and comprehensive impedance of the thyristor-level equivalent circuit are mapped to the fault tree model. Without removing the wires, the damping capacitance, component capacitance, and grading capacitance of the thyristor assembly can be numerically detected.

4. The converter valve fault diagnosis method based on state analysis according to claim 1, characterized in that: The step of obtaining the feature subset includes: Performing anomaly detection and cleaning on the operating data, setting an anomaly score threshold to retain samples to form a sample set; and classifying the cleaned sample set by state level, performing a sample category balancing operation; For the feature state quantities in the data sample, the distance measurement between features, feature distribution characteristics and feature local retention ability are comprehensively considered to calculate the importance of each feature state quantity, and the features with importance higher than the threshold are retained and included in the feature subset.

5. The converter valve fault diagnosis method based on state analysis according to claim 1, characterized in that: The construction process of the converter valve fault diagnosis model includes: The data samples after filtering the feature subset are used as input data, with the dimension of Will Input to the Transformer branch captures long-term and short-term temporal dependencies and dynamic change characteristics; synchronously, Perform continuous wavelet transform to convert it into a two-dimensional time-frequency image sequence, and input it into the CNN branch to extract multi-scale time-frequency local features; The outputs of the two channels are spliced ​​into a fused feature vector, and divided into branch fused feature vectors corresponding to the thyristor component, valve cooling component and valve arrester; The branch fusion feature vector of each component corresponds to a task branch. Each task branch passes through a fully connected layer and softmax to output the probability distribution of the fault level corresponding to each component.

6. A converter valve fault diagnosis method based on state analysis according to claim 5, characterized in that: The Transformer branch uses a bidirectional GRU to extract bidirectional dynamic temporal features in X and connects to a multi-head self-attention encoder to capture the global temporal dependency pattern and generate a temporal context embedding vector; The multi-scale convolution kernels arranged in parallel in the CNN branches extract local features of the time-frequency image, and the parallel structure of the multi-scale convolution kernels includes at least two convolution kernel branches of different sizes; Connecting depth-separable convolution layers after multi-scale convolution reduces model complexity; Construct a mid- and high-level feature extraction network by stacking the reverse residual structure; A channel attention mechanism is embedded in the inverse residual structure to generate a pattern embedding feature vector of the local time-frequency structure in the time-frequency image.

7. The converter valve fault diagnosis method based on state analysis according to claim 5, characterized in that: The output of the two channels is spliced ​​into a fused feature vector, including: The temporal context embedding vector output by the Transformer branch and the pattern embedding feature vector output by the CNN branch are respectively subjected to global average pooling to obtain their corresponding global channel features; The two-branch global channel features are input into a shared two-layer fully connected neural network to generate the channel attention weight coefficients of the two-branch global channel features; The temporal context embedding vector and the pattern embedding feature vector are multiplied by their respective channel attention weight coefficients, and the weighted combination is used to generate a fusion feature vector.

8. The converter valve fault diagnosis method based on state analysis according to claim 5, characterized in that: Component index identification is introduced to divide the fused feature vector into branch fused feature vectors corresponding to thyristor component, valve cooling component and valve arrester; The branch fusion feature vector of each component corresponds to a task branch. In each task branch, the branch fusion feature vector is globally averaged and pooled to further compress the dimension. After passing through a fully connected layer and softmax, the probability distribution of the fault level corresponding to the component is output.

9. The converter valve fault diagnosis method based on state analysis according to claim 1, characterized in that: The converter valve fault diagnosis model is used to perform fault diagnosis on real-time converter valve operation data, including: The operating data of each converter valve component is collected in real time. After non-numerical state quantity processing, pre-processing, and feature subset screening, the converter valve fault diagnosis model is used to output the probability distribution of each fault level of the thyristor component, valve cooling component, and valve arrester. According to the probability distribution output by the converter valve fault diagnosis model, the maximum probability principle is adopted to determine the fault level of each component. Each component corresponds to an independent fault level diagnosis result.

10. A converter valve fault diagnosis system based on state analysis, using the converter valve fault diagnosis method according to any one of claims 1 to 9, characterized in that: The system comprises: Data acquisition and processing module: collects the operating data of the converter valve thyristor assembly, valve cooling assembly and valve arrester, and processes the non-numeric state quantity into numerical data according to the degree of its influence on the converter valve operation; State quantity feature screening module: for pre-processed operation data, a multi-criteria feature selection method is constructed to screen the feature subset in the operation data; Model building and reasoning module: Based on data samples that have been filtered through feature subsets, a dual-channel hybrid neural network combined with multi-task branch output is used to build a converter valve fault diagnosis model; The model operation and classification module collects the operation data of the converter valve in real time, and after preprocessing and feature screening, uses the converter valve fault diagnosis model to output the fault level distribution probability of each component.

Images