Wind power converter open circuit and current sensor fault diagnosis method

Abstract

The invention provides a wind power converter open circuit and current sensor fault diagnosis method, and belongs to the technical field of wind turbine generator power electronic conversion and fault diagnosis. The technical problems that a current signal is not stable due to wind speed fluctuation, and an open-circuit fault of an insulated gate bipolar transistor (IGBT) of a converter and a fault of a current sensor are similar in feature and easy to confuse and difficult to joint and accurately identify are solved. According to the technical scheme, the method comprises the following steps that S1, a three-phase current signal of the wind turbine generator set side is collected and preprocessed; s2, obtaining a plurality of intrinsic mode functions (IMF) and residual components; s3, obtaining feature vector sets of different fault states; s4, obtaining an adjacent matrix for describing a feature association relationship; and S5, extracting to obtain a fault classification result. According to the method, the distinguishing capability of similar faults is enhanced, the misjudgment rate is reduced, and the diagnosis accuracy and robustness are improved.

Description

Technical Field

[0001] This invention relates to the field of power electronic conversion and fault diagnosis technology for wind turbine generators, and in particular to a method for diagnosing open circuit and current sensor faults in wind power converters. Background Technology

[0002] With the continuous increase in wind power installed capacity, permanent magnet synchronous motor wind power conversion systems have been widely used due to their advantages such as high efficiency, compact structure, and low maintenance costs. These systems typically employ a full-power conversion topology. In the turbine-side converter, the insulated gate bipolar transistor (IGBT) is a key power device responsible for power conversion and switching control. Affected by factors such as electrothermal stress and operating condition fluctuations, IGBT open-circuit faults are one of the more common device failure modes in wind power converters. IGBT open-circuit faults can lead to phase current distortion, output electromagnetic torque fluctuations, and increased harmonic content. If they cannot be detected and located in a timely and accurate manner, they will cause increased unit vibration, decreased efficiency, and even malfunctioning protection systems. In severe cases, they may trigger a chain reaction of faults and shutdown accidents.

[0003] Most existing diagnostic methods for IGBT faults in converters are based on the ideal assumption that the measurement circuit is normal and the sensors are fault-free. However, current sensors, as key measurement components for acquiring machine-side current information, are also susceptible to failure. Current sensor failures can lead to discrepancies between the feedback current and the actual current, causing the converter control circuit to generate incorrect adjustment commands, resulting in system operating point deviations, and even triggering incorrect fault judgments and protection actions. Furthermore, measurement errors can directly interfere with current signal-based fault diagnosis methods, causing fault characteristics to be masked or amplified, thereby reducing the accuracy and reliability of the diagnostic results.

[0004] Diagnostic studies on IGBT open-circuit faults and current sensor faults in generator-side converters are mostly conducted independently. One type of method only detects and locates faults in converter devices, typically assuming the current measurement link is intact. Another type of method monitors the sensors individually, often relying on redundant measurement channels or observer structures to estimate sensor status. These methods are effective in single fault scenarios, but when IGBT open-circuit faults and current sensor faults coexist or are coupled, they are prone to problems such as overlapping fault characteristics, misdiagnosis, or missed diagnosis, making it difficult to distinguish different fault sources and types in a timely and accurate manner. In addition, some existing fault diagnosis schemes rely on additional redundant sensors, increasing the hardware cost and structural complexity of the system. Other methods heavily rely on accurate system parameters and simplified mathematical models, and their diagnostic performance tends to deteriorate significantly and lack robustness under actual operating conditions with drastic wind speed fluctuations and high noise levels.

[0005] How to solve the above-mentioned technical problems is the challenge facing this invention. Summary of the Invention

[0006] The purpose of this invention is to provide a method for diagnosing open-circuit and current sensor faults in wind power converters. This method solves the technical problems of non-stationarity of current signals caused by wind speed fluctuations, and the difficulty in accurately identifying the similar characteristics of open-circuit faults in converter insulated gate bipolar transistors (IGBTs) and current sensor faults. By using multi-scale decomposition and feature screening, constructing adaptive feature maps, and performing spatiotemporal joint learning, this method can accurately distinguish and identify multiple types of faults, reduce false alarms, and improve stability and engineering applicability under wind speed disturbance and noise conditions.

[0007] The inventive concept of this invention is as follows: The method of this invention collects three-phase current signals from the wind turbine side to construct a diagnostic sample set. Then, it utilizes Adaptive Noise Complete Ensemble Empirical Mode Decomposition (CEEMDAN) to perform multi-scale decomposition on the samples. Based on the correlation with the original current, it filters components of discriminative Intrinsic Mode Functions (IMFs). Furthermore, it extracts composite features fused with time-frequency information from the components of the target IMFs and constructs an adaptive hybrid topology graph structure that simultaneously characterizes feature amplitude differences and directional similarities to obtain an adjacency matrix of feature associations. Finally, it inputs the composite features and their graph structure into a Spatiotemporal Feature Graph Convolutional Network (STFGCN) to achieve joint diagnosis of open-circuit faults in Insulated Gate Bipolar Transistors (IGBTs) and current sensor faults in wind power converters. This method can effectively distinguish different types of faults under single current measurement conditions, providing reliable technical support for wind turbine operation, maintenance, and fault early warning.

[0008] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: a method for diagnosing open circuit and current sensor faults in wind power converters, comprising the following steps: S1. Collect the three-phase current signal from the wind turbine generator side and preprocess it to build a diagnostic sample set covering normal operating conditions, open-circuit fault insulated gate bipolar transistors (IGBTs), and current sensor faults. S2. Using the adaptive noise complete set empirical mode decomposition CEEMDAN, the three-phase current signals in the diagnostic sample set are decomposed into multiple scales to obtain several intrinsic mode functions (IMFs) and residual components. S3. Based on the correlation between each Intrinsic Mode Function (IMF) and the corresponding original current signal, target Intrinsic Mode Functions (IMFs) with fault discrimination capability are selected, and composite features of fused time and frequency information are extracted to obtain feature vector sets for different fault states. S4. Construct an adaptive hybrid topology graph structure based on the magnitude difference and directional similarity among the feature vector sets to obtain an adjacency matrix describing the feature association relationship; S5. Input the composite features and their corresponding adjacency matrices into the spatiotemporal feature fusion module to form a spatiotemporal feature graph convolutional network, obtain the fault classification results, and realize the joint fault diagnosis of open circuit faults of insulated gate bipolar transistors (IGBTs) and current sensor faults in the generator-side converter of wind turbine units.

[0009] Further, in step S1, the three-phase current signal from the wind turbine generator side is collected and preprocessed, including: synchronously sampling the three-phase current at the stator side of the permanent magnet synchronous generator; dividing the original three-phase current sequence into several sample segments according to a fixed-length time window; and performing three-phase unified-scale normalization processing on each sample segment, that is, using the maximum absolute amplitude of the three-phase current within the sample segment as a unified scaling reference, and synchronously scaling the three-phase current signal by the same proportion so that the amplitude range of the normalized three-phase current is mapped to... A diagnostic sample set is constructed from normalized sample fragments.

[0010] Furthermore, in S1, a diagnostic sample set covering normal operating conditions, open-circuit faults in IGBTs, and current sensor faults is constructed. This includes classifying the research object into six fault states based on the different fault locations of the generator-side converter: 1) No Fault (NF); 2) Single open-circuit fault (SOCF); 3) Open phase fault (OPF); 4) Double open-circuit fault in different bridgearms (DOCF-D); 5) Double open-circuit fault in the same bridgearm (DOCF-S) 6) Current sensor fault (CSF).

[0011] For each of the six types of fault states mentioned above, several three-phase current signal samples are collected to form a joint diagnostic sample set covering multiple fault types.

[0012] Furthermore, in S1, the wind turbine generator-side converter has a three-phase two-level topology and contains six insulated-gate bipolar transistor (IGBT) switches. These switches can be combined in different open-circuit positions to form twenty-five open-circuit fault types, including: Single-tube open-circuit faults include individual open-circuit cases of T1, T2, T3, T4, T5, and T6; phase loss faults include the combination of two switching devices on any phase bridge being open in the same phase; double-tube open-circuit faults in different bridge arms are cross-open-circuit combinations between different phases; double-tube open-circuit faults in the same bridge arm are the simultaneous open circuits of the upper and lower tubes of the same bridge arm; current sensor faults correspond to the failure states of the three-phase sensors A, B, and C.

[0013] Furthermore, in S2, the three-phase current signals in the diagnostic sample set are decomposed into multiple scales using the adaptive noise-complete set empirical mode decomposition (CEEMDAN). The decomposition can be expressed as follows:

[0014] In the formula, Indicates phase as The three-phase current time-domain signal, These correspond to the phases of the three-phase current signals, respectively. Indicates the sampling point. For the first The components of an intrinsic mode function (IMF) The total number of components of the intrinsic mode function (IMF) obtained by decomposition. For residual components, The component index of the Intrinsic Mode Function (IMF) .

[0015] Further, in step S2, based on the correlation between each Intrinsic Mode Function (IMF) and the corresponding original current signal, target IMFs with fault detection capabilities are selected, including: Calculate the Pearson correlation coefficient between the components of each intrinsic mode function (IMF) and the corresponding initial phase current to quantitatively describe the degree of linear correlation: ; In the formula, This represents the Pearson correlation coefficient between the corresponding intrinsic mode function (IMF) and the original phase current signal. The number of sampling points in the signal sequence. For sequence The mean, For sequence The mean is defined as follows: .

[0016] Furthermore, in step S3, composite features of fused time-frequency information are extracted to comprehensively characterize the current change patterns under different fault states. These features mainly include energy entropy, instantaneous frequency entropy, waveform coefficients, standard deviation, and location parameters, resulting in a set of feature vectors for different fault states.

[0017] 1) Select the current for each phase. The components of the effective intrinsic mode functions (IMFs) are divided into: The energy entropy is obtained by calculating the energy percentage of each time segment. : ; In the formula, This represents the number of effective intrinsic mode functions (IMFs) obtained from the current signal of each phase. The total number of time segments obtained by dividing the components of each effective intrinsic mode function (IMF) is denoted as . For the first Number of sampling points within a time segment For time segment numbers, , Indicates phase as The The components of the effective intrinsic mode function (IMF) in the th... Values ​​within a time segment Indicates phase as The Within the first time segment The energy percentage of each component of an effective intrinsic mode function (IMF). For the sample In phase The energy entropy is calculated based on the energy percentage. It is a very small positive real number; 2) The phase current spectrum exhibits a center shift during a fault. The change in instantaneous frequency entropy characterizes the abnormal disturbance of the phase trajectory. A Hilbert transform is applied to the components of the Intrinsic Mode Function (IMF) to obtain its instantaneous frequency: ; In the formula, For the first The components of the intrinsic mode function (IMF) in phase instantaneous frequency below The instantaneous phase is obtained by Hilbert transforming the components of the Intrinsic Mode Function (IMF). Will Discretized There are several frequency intervals, and the probability of sample points within each interval is calculated. Thus, the instantaneous frequency entropy is obtained as: ; In the formula, The number of frequency intervals to divide. For frequency range indexing, , For the instantaneous frequency sample to fall into the first The probability of a frequency interval For the sample In phase The instantaneous frequency entropy below; 3) The distortion degree of the current waveform varies under different fault conditions, and the standard deviation values ​​differ accordingly. The normalized standard deviation of the current in each phase is expressed as follows: ; In the formula, For the sample In phase The standard deviation normalization feature is shown below.

[0018] 4) Describe the distortion characteristics of the current signal under different faults, using the quantization index of waveform coefficients: ; In the formula, For the sample In phase The characteristics of the normalized waveform coefficients.

[0019] 5) Add positioning factor It reflects the three-phase current asymmetry caused by an open-circuit fault, and its deviation from zero is used to describe the directional change of the output current: ; In the formula, For the sample In phase The positioning factor features below.

[0020] The above features are sequentially concatenated to form a complete feature row vector. And construct feature matrices for 25 types of fault states: .

[0021] Furthermore, in step S4, an adaptive hybrid topology graph structure is constructed to obtain an adjacency matrix describing the feature association relationships. The calculation method is as follows: For any node i and node j, their Euclidean distance matrix Similarity matrix with cosine Represented as: ; In the formula and They are nodes and nodes eigenvectors, For feature dimension, Represents the vector dot product. Represents the L2 norm, It is an exponential function. The median of the Euclidean distance matrix is ​​taken. To improve the adaptability of similarity, a weighting factor based on node variance is introduced: ; In the formula This represents the weight factor of node i. Representing the eigenvector variance The initial similarity matrix after fusion represents the maximum variance across all nodes. ; In the formula Represents a node and nodes The initial similarity matrix. Due to the presence of noisy and redundant edges weakening the propagation of effective information during graph convolution, [the following is done] respectively... and Performing the k-nearest neighbor (KNN) operation yields the Euclidean mask matrix. Sum and cosine mask matrix After performing intersection and union weighted summation, a fusion mask matrix is ​​constructed. The union retains the edges that are selected at least once under both metrics, and the intersection retains the edges that are selected simultaneously under both metrics. ; In the formula Represented by node Centered on, based on or For candidate nodes Sort and select the first A set of nearest neighbor nodes; Represents the union of two nearest neighbor masks. This represents the intersection of two types of nearest neighbor masks; This is an indicator function that takes the value 1 if the condition is true and 0 otherwise. where is the weighting coefficient for the intersection edges; Indicates element-wise multiplication; fusion mask Acting on Obtain a lightweight adjacency matrix .

[0022] For any two connected nodes and By calculating the number of public neighbors This reflects the structural similarity between the two nodes in their local neighborhood, and the edge weights are adjusted accordingly. ; In the formula, Represents a node and nodes The number of public neighbors; The maximum threshold for the number of public neighbors. Represents a node The degree is defined as , Define as a node The degree, Enhancement coefficient for public neighbors; Given the matrix after correction for common neighbor edge weights, for Matrix symmetric normalization yields the adjacency matrix of the direct input graph convolutional layer.

[0023] Furthermore, in S5, the spatiotemporal feature fusion module comprises two Transformer layers and one graph convolutional layer. The first Transformer operates on the temporal samples, modeling them in the temporal dimension through a self-attention mechanism. The graph convolution treats each feature dimension as a graph node, performing neighborhood aggregation and weighted fusion using the constructed adjacency matrix. The second Transformer performs global self-attention optimization again on the features fused by the graph convolution.

[0024] Furthermore, in S5, the spatiotemporal feature map convolutional network includes an input layer, two spatiotemporal feature fusion modules, a feature flattening layer, and a fully connected classification layer connected in sequence. The feature flattening layer maps the extracted high-dimensional spatiotemporal features into a one-dimensional vector representation; the fully connected classification layer performs nonlinear mapping and normalization processing on the flattened feature vector, outputting the probability distribution of each fault category, thereby achieving joint identification of open-circuit faults in the IGBTs of the machine-side converter and faults in the current sensor.

[0025] Furthermore, in S5, the spatiotemporal feature map convolutional network is trained in the following manner: The diagnostic sample set is divided into a training set, a validation set, and a test set. The composite features and their corresponding fault labels in the training set are used as supervision signals. The loss function is calculated through forward propagation, and the network parameters are updated via backpropagation using a gradient descent optimization algorithm until a preset convergence condition or a maximum number of training epochs is met. The training process is monitored using the validation set. When the accuracy of the validation set fluctuates within ±1% over five consecutive training epochs, the model is considered converged, and the model parameters corresponding to the last epoch are used as the final parameters. The performance of the trained spatiotemporal feature map convolutional network is evaluated using the test set to verify the accuracy and robustness of the joint fault diagnosis method.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention introduces adaptive noise complete set empirical mode decomposition (CEEMDAN) to decompose the three-phase current on the machine side at multiple scales. It also combines the correlation with the original current to screen the components of the target intrinsic mode function (IMF) that are representative and discriminative. This allows the feature aliasing caused by wind speed fluctuations and noise superposition to be separated and highlighted at different scales. This can effectively suppress the problem of single time-domain or frequency-domain features being sensitive to operating conditions, improve the stable expression of fault information at multiple time scales, and make it easier to extract and identify subtle differences in the distortion of similar waveforms. This enhances the effective information content and anti-interference ability of diagnostic samples from the source.

[0027] 2. On the screened Intrinsic Mode Function (IMF) channels, this invention extracts and fuses composite features reflecting multi-dimensional information such as energy, waveform morphology, and frequency band distribution. These features encompass both local transient changes and overall patterns, simultaneously characterizing current distortion mode changes caused by IGBT open circuits and systematic measurement biases introduced by sensor failures. Compared to methods relying on a few shallow features or fixed frequencies, the composite features of this invention provide more stable fault characterization under different wind speeds and noise levels, thereby reducing intra-class discrepancies, increasing inter-class differences, and minimizing confusion between similar faults.

[0028] 3. This invention simultaneously considers feature amplitude differences and directional similarity to construct an adaptive hybrid topology graph. This ensures that the resulting adjacency relationships reflect both differences in strength and consistency in changing trends, thus more accurately describing the inherent coupling and complementary relationships between composite features. This graph structure avoids the biases caused by fixed topologies or single similarity measures, allowing the spatial aggregation of graph convolutions to adaptively focus on key channels under data-driven conditions. This improves the effectiveness and interpretability of spatial feature learning and provides a more reliable structure for subsequent classification.

[0029] 4. This invention inputs composite features and their adaptive graph structure into a spatiotemporal feature graph convolutional network. It utilizes graph convolution to mine feature spatial correlations and employs a temporal modeling module to capture the time-dependent and evolutionary patterns of multi-scale features, achieving collaborative learning. Therefore, it maintains high diagnostic accuracy and generalization ability even under complex operating conditions such as wind speed fluctuations and strong noise interference. It can accurately distinguish between normal states, multi-position IGBT open-circuit faults, and current sensor failures under a single current measurement link, significantly reducing the risk of false alarms and missed alarms, meeting the needs of online monitoring and fault early warning, and thus improving the operational safety and economy of wind turbine units. Attached Figure Description

[0030] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0031] Figure 1 This is a topology diagram of the PMSG wind power system of the permanent magnet synchronous generator in this invention.

[0032] Figure 2 This is a schematic diagram of the wind speed curve and the three-phase current signal acquisition results in this invention; Among them, (a) is the power spectral density curve of wind speed; (b) is a partial segment of the generated time-varying wind speed sequence; and (c) is a partial segment of the three-phase current signal of the wind turbine generator side collected when there is no fault.

[0033] Figure 3 This is a schematic diagram showing the absolute values ​​of the Pearson correlation coefficients of the components of the Intrinsic Mode Function (IMF) under different fault conditions in this invention.

[0034] Figure 4 This is a radar diagram showing the characteristic distribution of three-phase current under different faults in this invention.

[0035] Figure 5 This is a schematic diagram of the t-SNE visualization results for twenty-five open-circuit fault types under four selected models in this invention; Among them, (a) is the visualization result of t-SNE of Long Short-Term Memory Network; (b) is the visualization result of t-SNE of Graph Convolutional Network; (c) is the visualization result of t-SNE of Spatiotemporal Difference Graph Convolutional Network; and (d) is the visualization result of t-SNE of Spatiotemporal Feature Graph Convolutional Network.

[0036] Figure 6 This is a schematic diagram of the confusion matrix of six fault states under four selected models in this invention; Among them, (a) is the confusion matrix diagram of the Long Short-Term Memory Network; (b) is the confusion matrix diagram of the Graph Convolutional Network; (c) is the confusion matrix diagram of the Spatiotemporal Difference Graph Convolutional Network; and (d) is the confusion matrix diagram of the Spatiotemporal Feature Graph Convolutional Network.

[0037] Figure 7 This is a flowchart of the combined fault diagnosis method in this invention. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0039] Example 1: See Figure 1 and Figure 7 This embodiment provides a method for diagnosing open circuit and current sensor faults in wind power converters, including the following steps: S1. Collect the three-phase current signal from the wind turbine generator side and preprocess it to build a diagnostic sample set covering normal operating conditions, open-circuit fault insulated gate bipolar transistors (IGBTs), and current sensor faults. S2. Using the adaptive noise complete set empirical mode decomposition CEEMDAN, the three-phase current signals in the diagnostic sample set are decomposed into multiple scales to obtain the components and residual components of several intrinsic mode functions (IMFs). S3. Based on the correlation between the components of each Intrinsic Mode Function (IMF) and the corresponding original current signal, the components of the target Intrinsic Mode Function (IMF) with fault discrimination capability are selected and the composite features of the fused time-frequency information are extracted to obtain the feature vector set of different fault states. S4. Construct an adaptive hybrid topology graph structure based on the magnitude difference and directional similarity among the feature vector sets to obtain an adjacency matrix describing the feature association relationship; S5. Input the composite features and their corresponding adjacency matrices into the spatiotemporal feature fusion module to form a spatiotemporal feature map convolutional network, extract the fault classification results, and realize the joint fault diagnosis of open circuit faults of insulated gate bipolar transistors (IGBTs) and current sensor faults in the machine-side converter.

[0040] Furthermore, in S1, the three-phase current is synchronously sampled on the stator side of the permanent magnet synchronous generator. The original three-phase current sequence is divided into several sample segments according to a fixed-length time window. Each sample segment is then normalized using a unified three-phase scale. That is, the maximum absolute amplitude of the three-phase current within that sample segment is used as a unified scaling reference, and the three-phase current signals are synchronously scaled by the same ratio, so that the amplitude range of the normalized three-phase current is mapped to... A diagnostic sample set is constructed from normalized sample fragments.

[0041] Based on the different fault locations of the wind turbine generator-side converter, the research objects are divided into six fault states: 1) No Fault (NF); 2) Single open-circuit fault (SOCF); 3) Open phase fault (OPF); 4) Double open-circuit fault indifferent bridge arms (DOCF-D); 5) Double open-circuit fault in the same bridge arms (DOCF-S); 6) Current sensor fault (CSF). For each fault state, several three-phase current signal samples are collected to form a joint diagnostic sample set covering multiple fault types.

[0042] Single-tube open-circuit faults include individual open-circuit cases for T1, T2, T3, T4, T5, and T6, such as... Figure 1 As shown; phase loss faults include the combination of two switching devices on any phase bridge being open in the same phase; double tube open circuit faults in different bridge arms are the combination of cross open circuits between different phases; double tube open circuit faults in the same bridge arm are the simultaneous open circuits of the upper and lower tubes of the same bridge arm; current sensor faults correspond to the failure states of the three-phase sensors A, B, and C.

[0043] Furthermore, in S2, the three-phase current signals in the diagnostic sample set are decomposed into multiple scales using the adaptive noise-complete set empirical mode decomposition (CEEMDAN). The decomposition can be expressed as:

[0044] In the formula, Indicates phase as The three-phase current time-domain signal, These correspond to the phases of the three-phase current signals, respectively. Indicates the sampling point. For the first The components of an intrinsic mode function (IMF) The total number of components of the intrinsic mode function (IMF) obtained by decomposition. For residual components, The component index of the Intrinsic Mode Function (IMF) .

[0045] Furthermore, in S3, the Pearson correlation coefficient between the components of each intrinsic mode function (IMF) and the corresponding initial phase current is calculated to quantitatively describe the degree of linear correlation: ; In the formula, This represents the Pearson correlation coefficient between the corresponding intrinsic mode function (IMF) and the original phase current signal. The number of sampling points in the signal sequence. For sequence The mean, For sequence The mean is defined as follows: .

[0046] Furthermore, in S3, composite features of fused time and frequency information are extracted to comprehensively characterize the current change patterns under different fault states. These features mainly include energy entropy, instantaneous frequency entropy, waveform coefficients, standard deviation, and location parameters, resulting in a set of feature vectors for different fault states.

[0047] 1) Select the current for each phase. The components of the effective intrinsic mode functions (IMFs) are divided into: The energy entropy is obtained by calculating the energy percentage of each time segment. : ; In the formula, This represents the number of effective intrinsic mode functions (IMFs) obtained from the current signal of each phase. The total number of time segments obtained by dividing the components of each effective intrinsic mode function (IMF) is denoted as . For the first Number of sampling points within a time segment For time segment numbers, , Indicates phase as The The components of the effective intrinsic mode function (IMF) in the th... Values ​​within a time segment Indicates phase as The Within the first time segment The energy percentage of each component of an effective intrinsic mode function (IMF). For the sample In phase The energy entropy is calculated based on the energy percentage. It is a very small positive real number; 2) The phase current spectrum exhibits a center shift during a fault. The change in instantaneous frequency entropy characterizes the abnormal disturbance of the phase trajectory. A Hilbert transform is applied to the components of the Intrinsic Mode Function (IMF) to obtain its instantaneous frequency: ; In the formula, For the first The components of the intrinsic mode function (IMF) in phase instantaneous frequency below The instantaneous phase is obtained by Hilbert transforming the components of the Intrinsic Mode Function (IMF). Will Discretized There are several frequency intervals, and the probability of sample points within each interval is calculated. Thus, the instantaneous frequency entropy is obtained as: ; In the formula, The number of frequency intervals to divide. For frequency range indexing, , For the instantaneous frequency sample to fall into the first The probability of a frequency interval For the sample In phase The instantaneous frequency entropy below; 3) The distortion degree of the current waveform varies under different fault conditions, and the standard deviation values ​​differ accordingly. The normalized standard deviation of the current in each phase is expressed as follows: ; In the formula, For the sample In phase The standard deviation normalization feature is shown below.

[0048] 4) Describe the distortion characteristics of the current signal under different faults, using the quantization index of waveform coefficients: ; In the formula, For the sample In phase The characteristics of the normalized waveform coefficients.

[0049] 5) Add positioning factor It reflects the three-phase current asymmetry caused by an open-circuit fault, and its deviation from zero is used to describe the directional change of the output current: ; In the formula, For the sample In phase The positioning factor features below.

[0050] The above features are sequentially concatenated to form a complete feature row vector. And construct feature matrices for 25 types of fault states: .

[0051] Furthermore, in S4, an adaptive hybrid topology graph structure is constructed to obtain the adjacency matrix describing the feature association relationships. The calculation method is as follows: For any node i and node j, their Euclidean distance matrix Similarity matrix with cosine Represented as: ; In the formula and They are nodes and nodes eigenvectors, For feature dimension, Represents the vector dot product. Represents the L2 norm, It is an exponential function. The median of the Euclidean distance matrix is ​​taken. To improve the adaptability of similarity, a weighting factor based on node variance is introduced: ; In the formula This represents the weight factor of node i. Representing the eigenvector variance The initial similarity matrix after fusion represents the maximum variance across all nodes. ; In the formula Represents a node and nodes The initial similarity matrix. Due to the presence of noisy and redundant edges weakening the propagation of effective information during graph convolution, [the following is done] respectively... and Performing the k-nearest neighbor (KNN) operation yields the Euclidean mask matrix. Sum and cosine mask matrix After performing intersection and union weighted summation, a fusion mask matrix is ​​constructed. The union retains the edges that are selected at least once under both metrics, and the intersection retains the edges that are selected simultaneously under both metrics. ; In the formula Represented by node Centered on, based on or For candidate nodes Sort and select the first A set of nearest neighbor nodes; Represents the union of two nearest neighbor masks. This represents the intersection of two types of nearest neighbor masks; This is an indicator function that takes the value 1 if the condition is true and 0 otherwise. where is the weighting coefficient for the intersection edges; Indicates element-wise multiplication; fusion mask Acting on Obtain a lightweight adjacency matrix .

[0052] For any two connected nodes and By calculating the number of public neighbors This reflects the structural similarity between the two nodes in their local neighborhood, and the edge weights are adjusted accordingly. ; In the formula, Represents a node and nodes The number of public neighbors; The maximum threshold for the number of public neighbors. Represents a node The degree is defined as (same principle) Define as a node (degree); Enhancement coefficient for public neighbors; Given the matrix after correction for common neighbor edge weights, for Matrix symmetric normalization yields the adjacency matrix of the direct input graph convolutional layer.

[0053] Furthermore, in S5, a spatiotemporal feature fusion module is constructed. Its structure includes two Transformer layers and one graph convolutional layer. The first Transformer operates on the temporal samples and models them in the time dimension through a self-attention mechanism. The graph convolution treats each feature dimension as a graph node, and performs neighborhood aggregation and weighted fusion using the adjacency matrix constructed based on S4. The second Transformer performs global self-attention optimization again on the features fused by the graph convolution.

[0054] Furthermore, in S5, the spatiotemporal feature map convolutional network comprises an input layer, two spatiotemporal feature fusion modules, a feature flattening layer, and a fully connected classification layer connected in sequence. The feature flattening layer maps the extracted high-dimensional spatiotemporal features into a one-dimensional vector representation; the fully connected classification layer performs nonlinear mapping and normalization on the flattened feature vectors, outputting the probability distribution of each fault category.

[0055] Furthermore, in S5, the spatiotemporal feature map convolutional network is trained in the following way: the diagnostic sample set is divided into a training set, a validation set, and a test set. The composite features and their corresponding fault labels in the training set are used as supervision signals. The loss function is calculated through forward propagation, and the network parameters are updated through backpropagation using a gradient descent optimization algorithm until the preset convergence condition or the upper limit of the number of training rounds is met. The training process is monitored using the validation set. When the accuracy of the validation set fluctuates by no more than ±1% within 5 consecutive training rounds, the model is considered to have converged, and the model parameters corresponding to the last round are used as the final parameters. The performance of the trained spatiotemporal feature map convolutional network is evaluated using the test set to verify the accuracy and robustness of the joint fault diagnosis method.

[0056] Example 2: To verify the effectiveness of the wind power converter open-circuit and current sensor fault diagnosis method based on adaptive noise complete set empirical mode decomposition CEEMDAN and spatiotemporal feature map convolutional network described in this invention, a 6MW permanent magnet synchronous generator (PMSG) wind power system simulation model was built in MATLAB / Simulink. The simulation was performed on a system equipped with an AMD Ryzen 7 5800H CPU and an RTX 3050Ti GPU. The spatiotemporal feature map convolutional network model was built based on the PyTorch 2.5.1 framework in Python 3.12. To avoid the influence of randomness in simulation results on the effectiveness of the method, all tests were repeated 5 times. The electrical topology of the PMSG wind power system is as follows: Figure 1 As shown in Table 1, the parameters of the PMSG wind power system are detailed in Table 1. The system sampling signal is the three-phase current output from the wind turbine generator-side converter, and the sampling frequency is set to 2kHz.

[0057] Table 1 shows the simulation parameters of the PMSG wind power system with permanent magnet synchronous generator:

[0058] Example 3: This example uses the Kaimal turbulence spectrum model in the IEC 61400-1 standard to generate a time-varying wind speed sequence, and its wind speed power spectral density is defined as:

[0059] In the formula For turbulence frequency, The average wind speed, For the integral scale, The standard deviation of wind speed is used to characterize the intensity of wind speed fluctuations. For turbulence intensity; set , , The sampling frequency is 2Hz. Figure 2 In (a) of the figure, the wind speed power spectral density curve is shown, which conforms to the power-law distribution of the Kaimal model. Figure 2 Image (b) shows a partial fragment of the generated time-varying wind speed sequence. Figure 2 (c) shows a partial segment of the three-phase current signal on the wind turbine side under fault-free conditions. For 25 fault conditions, current data for 300 seconds was collected for each condition and divided into 600 time series samples, totaling 15,000 samples.

[0060] Example 4: This example utilizes the adaptive noise-complete ensemble empirical mode decomposition (CEEMDAN) to perform multi-scale decomposition of the three-phase current signals in the diagnostic sample set, and calculates the Pearson correlation coefficient between the components of each intrinsic mode function (IMF) and the corresponding original phase current to quantitatively describe the degree of linear correlation. A larger absolute value of the correlation coefficient indicates that the components of the IMF have strong consistency with the original signal and can retain the main dynamic change characteristics. Figure 3 The diagram shows the absolute values ​​of the Pearson correlation coefficients of the components of each Intrinsic Mode Function (IMF) under different fault conditions. The components 3 to 6 of the IMF maintain a strong correlation with the original signal under each fault condition. Therefore, these components are selected as effective feature channels for feature extraction.

[0061] Example 5: This example extracts composite features from fused time-frequency information to comprehensively characterize the current change patterns under different fault states. These features mainly include energy entropy, instantaneous frequency entropy, waveform coefficients, standard deviation, and location parameters, resulting in a set of feature vectors for different fault states. Figure 4 The image shows a radar diagram of the three-phase current characteristic distribution under different faults. Different types of faults have significant differences in characteristic values.

[0062] Example 6: In this example, the fault dataset is divided into training, validation, and test sets in a 7:2:1 ratio to ensure that the samples in each set do not overlap. The model uses the cross-entropy loss function and the Adam optimizer, with a learning rate of 0.001, a batch size of 64, and 100 training iterations. Other key hyperparameters are listed in Table 2 below.

[0063] Table 2 shows the hierarchical structure of the spatiotemporal feature map convolutional network.

[0064] Example 7: This example compares the method with nine typical fault diagnosis methods, including feedforward networks (multilayer perceptron), convolutional networks (convolutional neural networks, residual networks), recurrent networks (long short-term memory neural networks, gated recurrent networks), graph convolutional networks, and their extensions. The multilayer perceptron uses three fully connected layers, with two hidden layers containing 256 and 128 units respectively. Both the long short-term memory neural network and the gated recurrent network use two recurrent layers, each containing 128 hidden units and employing a bidirectional structure. The convolutional neural network consists of two one-dimensional convolutional layers. The residual network uses an improved one-dimensional residual structure, outputting results through two residual blocks. The graph convolutional network contains two graph convolutional layers. Furthermore, it compares the method with several other existing spatiotemporal convolutional networks, including a spatiotemporal graph convolutional neural network based on SCADA data, a lightweight spatiotemporal graph convolutional network, and a spatiotemporal difference graph convolutional network.

[0065] Table 3 presents the comparative results of various methods on the test set, where test time represents the total time required for the model to complete predictions on 1150 test samples. The accuracy of the multilayer perceptron is 92.31%. Due to its reliance solely on nonlinear mapping of input features, it struggles to effectively capture temporal correlations, resulting in relatively low recognition performance. Long Short-Term Memory (LSTM) neural networks and gated recurrent networks have certain advantages in temporal modeling, with significantly improved accuracy compared to the multilayer perceptron, validating the effectiveness of temporal modeling. The accuracy of convolutional neural networks and residual networks are 91.62% and 91.60%, respectively, indicating that convolutional structures can extract local temporal patterns, but they have shortcomings in long-term dependency modeling. The test accuracy of graph convolutional networks is only 90.56%, lower than other methods, as modeling only spatial relationships is insufficient to reflect dynamic fault characteristics. In contrast, models combining temporal and graph structure modeling exhibit better diagnostic performance. The spatiotemporal graph convolutional neural network achieves an accuracy of 95.74%, which is further improved to 97.48% using a lightweight adjacency matrix. The accuracy of the spatiotemporal difference graph convolutional network is 94.68%, slightly lower than that of the spatiotemporal graph convolutional neural network. In contrast, the spatiotemporal feature graph convolutional network proposed in this invention achieves an average accuracy of 98.55% and a test time of 0.0491s. While meeting real-time requirements, it can more fully capture global spatiotemporal dependencies, demonstrating superior fault identification capabilities.

[0066] Table 3

[0067] Example 8: In this example, four typical models are selected, and the t-distributed random neighborhood embedding (t-SNE) is used to perform two-dimensional visualization of the feature space. Figure 5 This is a visualization of t-SNE results for 25 open-circuit fault types under four selected models. Different degrees of class confusion exist in the Long Short-Term Memory network, graph convolutional network, and spatiotemporal difference graph convolutional network. Figure 5 (As shown by the red dashed rectangles in (a) to (c) of section 5), these methods have shortcomings in identifying some fault types. In contrast, Figure 5 In (d) of this invention, the spatiotemporal feature map convolutional network forms highly compact clusters in the feature space, which can significantly distinguish different fault types and exhibit the best classification performance. Figure 6 This is a schematic diagram of the confusion matrix for six fault states under four selected models, where... Figure 6 The Long Short-Term Memory network, graph convolutional network, and spatiotemporal difference graph convolutional network in (a) to (c) of 6 all exhibit significant misclassifications across multiple categories. While the spatiotemporal difference graph convolutional network mitigates some misclassifications, it still shows bias in identifying similar faults such as open-circuit dual-pipe systems. Figure 6 In (d) of this invention, the inter-class misclassification of the spatiotemporal feature map convolutional network is greatly reduced, effectively alleviating the confusion problem between similar fault categories.

[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for diagnosing open circuit and current sensor faults in a wind power converter, characterized in that, Includes the following steps: S1. Collect the three-phase current signal from the wind turbine generator side and preprocess it to build a diagnostic sample set covering normal operating conditions, open-circuit fault insulated gate bipolar transistors (IGBTs), and current sensor faults. S2. Using the adaptive noise complete set empirical mode decomposition CEEMDAN, the three-phase current signals in the diagnostic sample set are decomposed into multiple scales to obtain several intrinsic mode functions (IMFs) and residual components. S3. Based on the correlation between each Intrinsic Mode Function (IMF) and the corresponding original current signal, target Intrinsic Mode Functions (IMFs) with fault discrimination capability are selected and composite features of fused time-frequency information are extracted to obtain feature vector sets for different fault states. S4. Construct an adaptive hybrid topology graph structure based on the magnitude difference and directional similarity among the feature vector sets to obtain an adjacency matrix describing the feature association relationship; S5. Input the composite features and their corresponding adjacency matrices into the spatiotemporal feature fusion module to form a spatiotemporal feature map convolutional network, extract the fault classification results, and realize the joint fault diagnosis of open circuit faults of insulated gate bipolar transistors (IGBTs) and current sensor faults in the generator-side converter of wind turbine units.

2. The method for diagnosing open circuit and current sensor faults in a wind power converter according to claim 1, characterized in that, The S1 acquires and preprocesses the three-phase current signals from the wind turbine generator side, including: Synchronous sampling of the three-phase currents is performed on the stator side of the permanent magnet synchronous generator. The original three-phase current sequence is divided into several sample segments according to a fixed-length time window. Each sample segment is then normalized using a unified three-phase scale. Specifically, the maximum absolute amplitude of the three-phase current within that sample segment is used as a unified scaling reference, and the three-phase current signals are synchronously scaled by the same ratio, so that the amplitude range of the normalized three-phase currents is mapped to... A diagnostic sample set is constructed from normalized sample fragments.

3. The method for diagnosing open circuit and current sensor faults in a wind power converter according to claim 1, characterized in that, In step S1, a diagnostic sample set is constructed covering normal operating conditions, open-circuit faults in insulated-gate bipolar transistors (IGBTs), and current sensor faults, including: Based on the different locations of the faults in the wind turbine generator-side converter, the fault states are divided into six categories: 1) Fault-free NF; 2) Single-tube open-circuit fault (SOCF); 3) Phase loss fault OPF; 4) Open circuit fault (DOCF-D) in different bridge arms; 5) Open circuit fault in two tubes of the same bridge arm (DOCF-S); 6) Current sensor failure (CSF); For each of the six types of fault states mentioned above, several three-phase current signal samples are collected to form a joint diagnostic sample set covering multiple fault types.

4. The method for diagnosing open circuit and current sensor faults in a wind power converter according to claim 3, characterized in that, In S1, the wind turbine generator-side converter has a three-phase two-level topology and contains six insulated-gate bipolar transistor (IGBT) switches. These switches are combined in different open-circuit positions to form twenty-five open-circuit fault types, among which: Single-tube open-circuit faults include individual open-circuit cases of T1, T2, T3, T4, T5, and T6; phase loss faults include the combination of two switching devices on any phase bridge being open in the same phase; double-tube open-circuit faults in different bridge arms are cross-open-circuit combinations between different phases; double-tube open-circuit faults in the same bridge arm are the simultaneous open circuits of the upper and lower tubes of the same bridge arm; current sensor faults correspond to the failure states of the three-phase sensors A, B, and C.

5. The method for diagnosing open circuit and current sensor faults in a wind power converter according to claim 1, characterized in that, In step S2, the three-phase current signals in the diagnostic sample set are decomposed into multiple scales using the adaptive noise complete set empirical mode decomposition (CEEMDAN). The decomposition form is expressed as follows: ; In the formula, Indicates phase as The three-phase current time-domain signal, These correspond to the phases of the three-phase current signals, respectively. Indicates the sampling point. For the first The components of an intrinsic mode function (IMF) The total number of components of the intrinsic mode function (IMF) obtained by decomposition. For residual components, The component index of the Intrinsic Mode Function (IMF) .

6. The method for diagnosing open circuit and current sensor faults in a wind power converter according to claim 1, characterized in that, In step S2, based on the correlation between each Intrinsic Mode Function (IMF) and the corresponding original current signal, target IMFs with fault detection capabilities are selected, including: Calculate the Pearson correlation coefficient between the components of each intrinsic mode function (IMF) and the corresponding initial phase current to quantitatively describe the degree of linear correlation: ； In the formula, This represents the Pearson correlation coefficient between the corresponding intrinsic mode function (IMF) and the original phase current signal. The number of sampling points in the signal sequence. For sequence The mean, For sequence The mean is defined as follows: .

7. The method for diagnosing open circuit and current sensor faults in a wind power converter according to claim 1, characterized in that, In step S3, the extraction of composite features from the fused time-frequency information is used to characterize the current variation patterns under different fault states, including energy entropy, instantaneous frequency entropy, waveform coefficients, standard deviation, and location parameters, to obtain a feature vector set for different fault states. This includes the following steps: S31, Select the current for each phase. The components of the effective intrinsic mode functions (IMFs) are divided into: The energy entropy is obtained by calculating the energy percentage of each time segment. : ； In the formula, This represents the number of effective intrinsic mode functions (IMFs) obtained from the current signal of each phase. The total number of time segments obtained by dividing the components of each effective intrinsic mode function (IMF) is denoted as . For the first Number of sampling points within a time segment For time segment numbers, , Indicates phase as The The components of the effective intrinsic mode function (IMF) in the th... Values ​​within a time segment Indicates phase as The Within the first time segment The energy percentage of each component of an effective intrinsic mode function (IMF). For the sample In phase The energy entropy is calculated based on the energy percentage. It is a very small positive real number; S32. The spectrum of the phase current exhibits a center shift during a fault. The change in instantaneous frequency entropy characterizes the abnormal disturbance of the phase trajectory. A Hilbert transform is applied to the components of the intrinsic mode function (IMF) to obtain its instantaneous frequency: ； In the formula, For the first The components of the intrinsic mode function (IMF) in phase instantaneous frequency below The instantaneous phase is obtained by Hilbert transforming the components of the Intrinsic Mode Function (IMF). Will Discretized There are several frequency intervals, and the probability of sample points within each interval is calculated. The instantaneous frequency entropy is obtained as: ； In the formula, The number of frequency intervals to divide. For frequency range indexing, , For the instantaneous frequency sample to fall into the first The probability of a frequency interval For the sample In phase The instantaneous frequency entropy below; S33. The distortion degree of the current waveform varies under different fault conditions, and the standard deviation values ​​differ accordingly. The normalized standard deviation of the current in each phase is expressed as follows: ； In the formula, For the sample In phase Standard deviation normalization feature; S34. Describe the distortion characteristics of the current signal under different faults, using the quantization index of waveform coefficients: ； In the formula, For the sample In phase Normalized waveform coefficient characteristics; S35, Add positioning factor It reflects the three-phase current asymmetry caused by an open-circuit fault, and its deviation from zero is used to describe the directional change of the output current: ； In the formula, For the sample In phase The following are the positioning factor characteristics; The above features are sequentially concatenated to form a complete feature row vector. And construct feature matrices for 25 types of fault states: 。 8. The method for diagnosing open circuit and current sensor faults in a wind power converter according to claim 1, characterized in that, In step S4, the method for constructing the adaptive hybrid topology graph structure and obtaining the adjacency matrix describing the feature association relationship is as follows: For any node i and node j, their Euclidean distance matrix Similarity matrix with cosine Represented as: ； In the formula and They are nodes and nodes eigenvectors, For feature dimension, Represents the vector dot product. Represents the L2 norm, It is an exponential function. Take the median of the Euclidean distance matrix; To improve the adaptability of similarity, a weighting factor based on node variance is introduced: ； In the formula This represents the weight factor of node i. Representing the eigenvector variance The initial similarity matrix after fusion represents the maximum variance across all nodes. ; In the formula Represents a node and nodes The initial similarity matrix, due to the presence of noisy and redundant edges, weakens the propagation of effective information during graph convolution. This is addressed by... and Performing the k-nearest neighbor (KNN) operation yields the Euclidean mask matrix. Sum and cosine mask matrix After performing intersection and union weighted summation, a fusion mask matrix is ​​constructed. The union retains the edges that are selected at least once under both metrics, and the intersection retains the edges that are selected simultaneously under both metrics. ； In the formula Represented by node Centered on, based on or For candidate nodes Sort and select the first A set of nearest neighbor nodes; Represents the union of two nearest neighbor masks. This represents the intersection of two types of nearest neighbor masks; For indicator functions; where is the weighting coefficient for the intersection edges; Indicates element-wise multiplication; fusion mask Acting on Obtain a lightweight adjacency matrix ; For any two connected nodes and By calculating the number of public neighbors This reflects the structural similarity between the two nodes in their local neighborhood, and the edge weights are adjusted accordingly. ； In the formula, Represents a node and nodes The number of public neighbors; The maximum threshold for the number of public neighbors. Represents a node The degree is defined as , Define as a node The degree; Enhancement coefficient for public neighbors; Given the matrix after correction for common neighbor edge weights, for Matrix symmetric normalization yields the adjacency matrix of the direct input graph convolutional layer.

9. A method for diagnosing open circuit and current sensor faults in a wind power converter according to claim 8, characterized in that, In S5, the spatiotemporal feature fusion module has a structure including two Transformer layers and one graph convolutional layer. The first Transformer acts on the temporal samples and models them in the time dimension through a self-attention mechanism. The graph convolution treats each feature dimension as a graph node and performs neighborhood aggregation and weighted fusion based on the constructed adjacency matrix. The second Transformer performs global self-attention optimization on the features fused by the graph convolution again. In S5, the spatiotemporal feature map convolutional network includes an input layer, two spatiotemporal feature fusion modules, a feature flattening layer, and a fully connected classification layer connected in sequence. The feature flattening layer maps the extracted high-dimensional spatiotemporal features into a one-dimensional vector representation; the fully connected classification layer performs nonlinear mapping and normalization on the flattened feature vectors, outputting the probability distribution of each fault category, thereby realizing the joint identification of open-circuit faults of IGBTs and current sensor faults in the machine-side converter.

10. A method for diagnosing open circuit and current sensor faults in a wind power converter according to claim 1, characterized in that, In S5, the spatiotemporal feature map convolutional network is trained in the following way: The diagnostic sample set is divided into a training set, a validation set, and a test set. The composite features and their corresponding fault labels in the training set are used as supervision signals. The loss function is calculated through forward propagation, and the network parameters are updated by backpropagation using the gradient descent optimization algorithm until the preset convergence condition or the upper limit of the number of training rounds is met. The training process is monitored using a validation set. When the validation set accuracy fluctuates by no more than ±1% within five consecutive training rounds, the model is considered to have converged, and the model parameters corresponding to the last round are used as the final parameters. The spatiotemporal feature map convolutional network trained is evaluated using a test set to verify the accuracy and robustness of the joint fault diagnosis method.

Images