Diagnosis method for multi-tube open-circuit fault of doubly-fed wind power converter

Abstract

The invention discloses a double-fed wind power converter multi-tube open-circuit fault-oriented diagnosis method, which is characterized in that an ICAM module is connected in series on the basis of a ConvNeXt-T model, and an ICAM-ConvNeXt model is constructed and trained; and carrying out real-time monitoring and online diagnosis on the running state of the DFIG by adopting the trained ICAM-ConvNeXt model. The method has the advantages that the training data volume can be reduced, the model diagnosis speed is high, the fault position can be accurately positioned, the number of iterations is small, the positioning precision is high, and high diagnosis accuracy can still be ensured under multiple working conditions and extreme conditions.

Description

Technical Field

[0001] The invention relates to the field of electric power technology, and in particular to a diagnosis method for multi-tube open-circuit faults of a double-fed wind power converter. Background Art

[0002] With the vigorous development of renewable energy, the application of wind power technology in various fields has also received more and more attention. In terms of wind power technology selection, the doubly-fed induction generator (DFIG) has been regarded as one of the important technologies for improving energy structure and security capabilities due to its advantages such as low cost, controllable power and flexible grid connection. This technology helps to improve energy independence, reliability and environmental adaptability. However, the converter that drives the DFIG to work restricts the development of motors due to its high failure rate and maintenance cost. The converter failure is mainly concentrated in the insulated-gate bipolar transistor (IGBT). Common IGBT failures in the converter include short circuit and open circuit failures. Although the short circuit failure is highly destructive, the method of implanting a fuse in the drive circuit is usually adopted to convert the short circuit failure into an open circuit failure; the change in the open circuit fault characteristic is relatively subtle and difficult to detect manually. The lack of accurate and efficient diagnostic technology will cause problems such as distortion and asymmetry of the three-phase output waveform, affecting the normal operation of the DFIG. Moreover, in DFIG, compared with the grid-side converter, the rotor-side converter has a higher probability of IGBT open-circuit failure under long-term high heat load operation, and most of the current research samples and processes the fault signals of the rotor-side converter in DFIG under the condition of sufficient samples, but in actual engineering, it often takes a huge cost and time to build a data set containing rich feature samples. When the data set is seriously unbalanced and insufficient, it is difficult for the model to fully learn the fault characteristics through a small amount of data, and it is easy to produce overfitting. Therefore, studying the IGBT open-circuit fault diagnosis method in the DFIG rotor-side converter under small samples is of great significance to the normal operation of the wind power system.

[0003] At present, the research on the open-circuit fault diagnosis of IGBT in the rotor-side converter of DFIG is divided into two categories, namely, based on traditional algorithms and artificial intelligence. The diagnosis method based on artificial intelligence gradually replaces the traditional algorithm because it overcomes the dependence on mathematical models and realizes accurate diagnosis of faults by extracting features and recognizing patterns of state variables during motor faults. Common image transformation methods used in fault diagnosis include voiceprints, Markov transfer fields, Gram angle fields, recursive graphs, Wigner-Vile distributions, and wavelet transforms. Although the images transformed by these methods can retain the spatiotemporal dependencies in the original fault signals, the single two-dimensional image mapping transformation has certain limitations, and the intelligent model cannot fully utilize its ability to extract high-dimensional features in the case of small samples. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the technical problem to be solved by the present invention is: how to provide a diagnostic method that can reduce the amount of training data, has a fast model diagnosis speed, can accurately locate the fault location, has a small number of iterations, has high positioning accuracy, and can still ensure a high diagnostic accuracy rate under multiple working conditions and extreme conditions.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A diagnostic method for multi-tube open-circuit faults in a doubly-fed wind turbine converter is proposed. An ICAM module is connected in series on the basis of a ConvNeXt-T model to construct and train an ICAM-ConvNeXt model. The trained ICAM-ConvNeXt model is used to perform real-time monitoring and online diagnosis of the operating status of DFIG.

[0007] Furthermore, the training steps of the ICAM-ConvNeXt model are as follows:

[0008] S1. According to the working mechanism of DFIG converter under normal and fault conditions, a DFIG simulation model is established and three-phase current is obtained as an effective fault parameter;

[0009] S2, using Gram angle difference field coding to map the three-phase current signals into color GADF images respectively, extracting the images of the red channel, green channel and blue channel from the three GADF images respectively, and fusing them according to the channel dimension to obtain a fused image;

[0010] S3, randomly dividing the fused image dataset into a training set for model training, a validation set for model adjustment, and a test set for performance evaluation;

[0011] S4. Input the training set into the initialized ICAM-ConvNeXt model and perform multiple rounds of iterative training. Use the validation set to evaluate the optimal performance of the model and record the best training parameters. Use the test set to evaluate the classification and diagnostic accuracy of the model. If the preset performance standard is not met, the model will re-enter the training stage and be optimized again until the classification and diagnostic accuracy of the model reaches the preset performance standard.

[0012] Furthermore, during diagnosis, the three-phase current is measured and sampled in real time by the current sensor of the DFIG, and the three-phase current signals are mapped into color GADF images respectively. The images of the red channel, green channel and blue channel are extracted from the three GADF images respectively, and fused according to the channel dimension to obtain a real-time fused image; the trained ICAM-ConvNeXt model is used to extract features of the real-time fused image, and fault classification and identification are realized.

[0013] Furthermore, the ICAM module is based on the CAM module and adopts depthwise separable convolution DSC to replace the ordinary full convolution in the CAM module. The depthwise separable convolution DSC is provided with a BN layer and a Relu layer.

[0014] In summary, the present invention has the advantages of being able to reduce the amount of training data, fast model diagnosis speed, accurate positioning of the fault location, small number of iterations, high positioning accuracy, and still being able to ensure a high diagnostic accuracy rate under multiple working conditions and extreme conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is the DFIG system structure.

[0016] Figure 2 It is the control structure of DFIG system.

[0017] Figure 3 Schematic diagram of the conversion from the original three-phase current signal to the multi-channel fusion image.

[0018] Figure 4 Schematic diagram of the conversion from the original three-phase current signal to the multi-channel fusion image.

[0019] Figure 5 It is the basic structure of DSC network.

[0020] Figure 6 This is the ICAM-ConvNeXt model structure.

[0021] Figure 7 This is a picture of a real-time digital semi-physical platform.

[0022] Figure 8 Output three-phase current for DFIG under partial fault category.

[0023] Fig. 9 This is a multi-channel fusion image under partial multi-tube failure.

[0024] Fig.10 Validation accuracy of the diagnostic model for different batch sizes, optimizers, and initial learning rates.

[0025] Fig.11 This is a curve chart of accuracy and loss rate during training.

[0026] Fig.12 Diagnostic confusion matrix results during testing.

[0027] Fig.13 Schematic diagram of t-SNE visualization at different stages of the multi-channel fusion image test set.

[0028] Fig.14 are the feature images under different image processing methods.

[0029] Fig.15 Radar chart for fault diagnosis performance evaluation.

[0030] Fig.16 Confusion matrix of the test process under different working conditions.

[0031] Fig.17 This is the accuracy and loss rate curve of the training process under imbalanced data.

[0032] Fig.18 Confusion matrix for the testing process under imbalanced data. DETAILED DESCRIPTION

[0033] The present invention is further described in detail below in conjunction with embodiments.

[0034] DFIG basic structure and control: In DFIG, back-to-back converter technology is used to excite the rotor. The rotor-side converter is connected to the DFIG rotor, and the grid-side converter is connected to the grid. The two converters are connected through a DC bus to form a closed loop to achieve rotor excitation. The DFIG system structure is as follows: Figure 1 shown.

[0035] As an efficient and flexible power generation equipment, DFIG is widely used in wind power generation and power system peak regulation. Its core advantage is that it can realize dynamic support for the power grid through the power flow on the rotor side. Therefore, the rotor side converter is a key component in the DFIG system. This embodiment adopts a model predictive current control algorithm, regards the rotor current as a state variable, and uses the forward Euler method to dynamically predict the rotor current value at the next moment. The core of the control algorithm is to find the optimal voltage vector combination. This vector combination is selected when the error between the rotor current and the ideal current is minimized, and is output to the rotor side converter. The DFIG system control structure is as follows: Figure 2 shown.

[0036] Analysis of DFIG fault characteristics: The power switch devices in the DFIG rotor-side converter are subject to electromagnetic interference, causing open circuit faults, and the faults may occur at T 1 ~T 6 Among them, the probability of single-tube failure is the highest, double-tube failure has a certain probability of occurrence, and three-tube and above failures almost never occur. Most of the research on motor drive circuits based on artificial intelligence considers single-tube failures, and there is little research on double-tube failures, and double-tube failures often cause greater harm. This embodiment divides the IGBT failures in the DFIG rotor-side converter into two categories, namely, double-tube failures on opposite sides and out-of-phase double-tube failures on the same side for research and analysis.

[0037] The fault of two tubes on different sides and in different phases is T 1 and T 6 Fault as an example, when T 1 and T 6 After an open circuit fault occurs at the same time, the positive half-cycle waveform of the A phase current is close to 0, the negative half-cycle waveform of the B phase current is close to 0, and the output current of the C phase is distorted and increases. Therefore, the characteristics of the double-tube open circuit fault on the opposite side and in opposite phases can be summarized as follows: when the faulty IGBT is in the upper bridge arm, the phase current of the faulty bridge arm is less than or equal to zero; when the faulty IGBT is in the lower bridge arm, the phase current of the faulty bridge arm is greater than or equal to zero; the phase current of the non-faulty bridge arm will be distorted to a certain extent. 1 and T 3 Fault as an example, when T 1 and T 3 After the open circuit fault occurs at the same time, the current of phase A and phase B is less than or equal to zero, and the output current of phase C is slightly distorted. Therefore, the characteristics of the double-tube open circuit fault on the same side with different phases can be summarized as follows: when the two faulty IGBTs are in the upper bridge arm, the phase current of the faulty bridge arm is less than or equal to zero; when the two faulty IGBTs are in the lower bridge arm, the phase current of the faulty bridge arm is greater than or equal to zero; the phase current of the non-faulty bridge arm is slightly distorted.

[0038] Based on the above analysis, and considering that the DFIG is equipped with current sensors that can measure and sample the three-phase current in real time, and the characteristics of the three-phase current waveform under different open-circuit fault states will have different distortion phenomena, the three-phase current can be selected as the characteristic variable for diagnosing multi-tube faults in the DFIG rotor-side converter.

[0039] Feature enhancement algorithm based on multi-channel fusion: In DFIG, three-phase current sensors are essential electrical components used to monitor and control current to ensure the effective operation and optimized performance of the system. An IGBT open-circuit fault in the rotor-side converter of DFIG will cause a certain distortion of the three-phase current. The uniqueness of the degree of current distortion caused by various faults is the basis for the study of this embodiment. Therefore, the degree of current distortion can be used to locate the fault. The rotor-side current signal is a one-dimensional time series signal. After an open-circuit fault occurs in the IGBT, the complexity of the three-phase current is extremely high from a time domain perspective. The difference between them is often difficult to clearly distinguish from a one-dimensional perspective, but from a two-dimensional image perspective, the fault implications of the three-phase current can be revealed. The GAF coding technology can effectively record the fluctuation information in the original signal and highlight the hidden features of the signal in the time domain. It includes two coding methods: Gramian Angular Summation Field (GASF) and GADF. The specific implementation steps are as follows:

[0040] Assume that there is a time series signal X with N time points. 1 ,x2,…,x i ,…,x N The Gram matrix is ​​defined as shown in formula (1). The inner product of vectors can represent the correlation between vectors.

[0041]

[0042] Where: G represents the Gram matrix; <·> represents the inner product.

[0043] But one-dimensional time series is not a vector, so The data is mapped to the polar coordinate system and converted into a vector, and the conversion is shown in formula (2).

[0044]

[0045] Where: θ i is the polar angle; r i is the polar diameter; n is the spatial regularization constant.

[0046] From formula (2), in the polar coordinate system, when time increases, The encoding obtains points with different angles and radii. Since the cosine function is monotonic on [0,π], that is, the mapping result of a time series data has a unique corresponding value in the polar coordinate system, and the inverse mapping is also unique, that is, the GAF algorithm is a double mapping, which can retain complete feature information, and the dependence on time is guaranteed by the polar radius r. Therefore, the feature map generated by processing the one-dimensional time data with the GAF algorithm will not lose any modal information and contains the connection between signals at different time intervals.

[0047] However, if the Gram matrix shown in formula (2) is used to define the inner product, the output data will be difficult to distinguish the interference of Gaussian noise.

[0048] In order to solve this problem, the inner product form with penalty term is defined as shown in formula (3).

[0049]

[0050] For the two new inner product forms, GASF and GADF can be obtained by calculating the sum and difference of trigonometric chord functions in equation (4) and equation (5) respectively.

[0051]

[0052]

[0053] Where: φ i (i=1,2,…,n) is the i-th sequence angle.

[0054] After GAF encoding, the upper triangular elements in GADF and the lower triangular elements in GASF can more completely explain the potential states and information in the original signal. When the amplitude of the signal slows down, the corresponding GADF or GASF image will map a large area of ​​connected areas with similar color textures; when the amplitude of the signal increases greatly, the corresponding GADF or GASF image will display horizontal textures or vertical textures. With respect to the GASF and GADF encoding methods, the GADF encoding method is unique in time series mapping, and the encoded mapping image contains element sparsity, in which most of the redundant information has been removed. The GADF image not only contains the time correlation in the original signal, but also retains the potential state correlation in the non-stationary time-varying signal. Compared with the GASF encoding method, it also retains the differences between multiple modalities, which can cause the maximum difference between different fault features. That is, this embodiment uses the GADF encoding method to perform a feature extraction method for fault signal dimensionality transformation and multi-channel fusion, and its conversion process is as follows: Figure 3 shown.

[0055] exist Figure 3In the above, the three-phase current signals collected by the sensor are converted into corresponding GADF images in sequence using the GADF encoding technology. Since each image is composed of three primary colors of red, green and blue, one of the three primary colors of RGB is extracted from each phase current under comprehensive consideration and used as the fused image, that is, the red primary color is extracted from phase A to obtain the grayscale image I. A -R, B phase extracts the green primary color to obtain the grayscale image I B - G, C phase extracts the blue primary color to obtain grayscale image I C -B, I A -R, I B -G and I C -B is fused into a new image by channel. The multi-channel fusion image will contain the internal characteristic information of the three-phase current. Its contour and color features jointly depict the fault information of the DFIG rotor-side converter, which can effectively realize the concentration of fault characteristics, and the training data is reduced to 1 / 3 of the original three-phase current data, which can effectively reduce the complexity of data processing. The intelligent model identifies various types of faults by analyzing the relevant fault features in the multi-channel fusion image. In this process, complex coordinate transformation steps can be avoided, and it does not rely on the in-depth knowledge and manual operation experience of experts. The original signal is simply converted directly into an image format. Using an "end-to-end" strategy, different fault types can be quickly and accurately diagnosed.

[0056] ICAM-ConvNeXt model: ConvNeXt is an advanced deep learning architecture developed by Facebook AI Research Institute and other organizations in 2022. It uses residual network as the basic skeleton design principle, successfully solves the common gradient vanishing or exploding problems in deep networks, and can further improve the diagnostic accuracy. ConvNeXt also draws on the training strategy of Transformer and integrates the training process of other neural networks. In terms of training efficiency and accuracy, it surpasses the popular Swin Transformer architecture. In order to make the model adaptively focus on the intrinsic fault characteristics of multi-channel fusion images, this embodiment integrates the ICAM module with the ConvNeXt network to enhance the performance of the overall model, and obtains the new ICAM-ConvNeXt model for application in the diagnosis of multi-tube open-circuit faults of DFIG rotor-side converters.

[0057] ICAM structure: The CAM structure adaptively adjusts the importance weights of different channels. By weighting them independently, the model can enhance its ability to capture key channel features, achieve decoupled expression of fault features, improve the quality of feature representation and overall model performance, and help improve the generalization and interpretability of the model. Figure 4 Shown is the basic structure of the CAM module.

[0058] Figure 4 In the example, the shape and size of the input feature map is F∈R C×H×W , where C is the number of feature channels, H and W are the height and width of the feature map, respectively. The input features are globally averaged and max-pooled to obtain 1×1×W feature vectors, which are then input into a shared multi-layer perceptron (MLP). The feature vectors are extracted through two layers of full convolutional layers, and then the outputs are summed and input into the sigmoid activation function to compress the feature values ​​to between 0 and 1 to indicate the importance of each channel. The output result is multiplied by the input feature to obtain a feature map distributed by channel weights. The calculation process is as follows:

[0059]

[0060] Where: σ is the sigmoid activation function; F avg and F max is the feature after global average pooling and maximum pooling; W 0 and W 1 is the weight of the full convolutional layer.

[0061] Since the multi-channel fusion image is input into the model as a three-channel color image during the feature extraction process, the traditional CAM module will generate repeated and redundant channel features when calculating the channel weight score. To solve this problem, this embodiment replaces the original ordinary full convolution in the CAM module with Depth Separable Convolution (DSC), and then adds BN layer and Relu layer after DSC to reduce the complexity of the model and enhance the overall nonlinear fitting ability of the model. DSC can be divided into layer-by-layer convolution and point-by-point convolution. The basic structures of the two are as follows: Figure 5 shown.

[0062] Figure 5 In , the amount of computation of the input feature map after ordinary convolution is as follows:

[0063] Q 1 =K 2 ·S·C·H·W (7)

[0064] Where: H and W represent the length and width of the input feature; K represents the size of the convolution kernel; C is the number of input feature channels; S is the number of output feature channels.

[0065] The amount of computation of the input feature map after DSC is as follows:

[0066] Q 2 =K 2 ·C·H·W+S·C·H·W (8)

[0067] The ratio of the two is obtained from equations (7) and (8):

[0068]

[0069] In formula (9), for the same size of input data, DSC has much less computational complexity than ordinary convolution. Now, ordinary convolution in the traditional CAM structure is replaced by depthwise separable convolution to obtain a new ICAM structure, which can effectively focus on useful fault features and capture rich contextual channel information without significantly increasing parameters and computational costs, thereby improving the recognition ability of fault image details and upper and lower distance dependencies, enhancing channel feature expression, reducing overfitting risks, and improving its generalization and flexibility.

[0070] ConvNeXt structure: ConvNeXt is improved on the basis of residual network, including increasing the stacking depth of independent modules to enhance learning ability; replacing traditional convolution to optimize feature extraction; introducing reverse bottleneck structure to improve computational efficiency; expanding convolution kernel to capture rich spatial information; and using GELU activation function and LN layer to improve the performance of downsampling module. There are 5 versions of ConvNeXt network, namely T, S, B, L and XL. The essential difference between different ConvNeXt networks is the number of channels of the input feature layer in each individual stage and the number of times the stacked modules are repeated in each stage. After continuous model debugging, this embodiment adopts ConvNeXt-T model, whose number of channels and the number of times the stacked modules are repeated in each stage are (96, 192, 384, 768) and (3, 3, 9, 3) respectively. In order to improve the classification and diagnostic ability and efficiency of the model, the ConvNeXt structure is serially connected with the ICAM module, and its model structure is as follows: Figure 6 shown.

[0071] In the fault diagnosis process of DFIG rotor-side converter, the improved fusion of ICAM and ConvNeXt can significantly improve the accuracy and efficiency of fault identification. The ICAM module focuses on key fault feature areas, enhances the model's sensitivity to fault modes, reduces the number of parameters and the amount of calculation, and improves the robustness of the model in complex environments. This serial fusion can help the model accurately identify the feature areas related to the fault in complex fault images and improve the accuracy of fault diagnosis.

[0072] DFIG rotor-side converter fault diagnosis process: In order to solve the problem of multi-tube open-circuit fault diagnosis of the rotor-side converter in DFIG, a method combining multi-channel fusion graph and ICAM-ConvNeXt model is proposed. The specific steps are as follows:

[0073] (1) Taking DFIG as the diagnosis object, the experimental platform is used to simulate the operation of multiple open-circuit switches of different switching power devices, and the three-phase currents under different fault conditions are sampled and normalized.

[0074] (2) The sliding window feature extraction technology is applied to the three-phase current signal under different fault conditions to obtain fault sample data, thereby constructing the original data set. The GADF encoding technology is used to map the three-phase current signals into color GADF images respectively. The images of the red channel, green channel and blue channel are extracted from the three GADF images respectively, and fused according to the channel dimension to create a fused image containing multiple channels. After preprocessing, these images form a multi-channel fused image data set. The data set is randomly divided into three non-overlapping subsets: a training set for model training, a validation set for model adjustment, and a test set for performance evaluation, which provides fault sample data for subsequent experimental research.

[0075] (3) The training set in (2) is input into the initialized ICAM-ConvNeXt model, and multiple rounds of iterative training are performed. The validation set is used to evaluate the optimal performance of the model, and the optimal training parameters are recorded. The classification and diagnosis accuracy of the model is evaluated using the test set. If the test result does not meet the optimal performance standard, the model will re-enter the training stage and be optimized again until the classification and diagnosis accuracy of the model reaches a certain standard. At this point, the optimal performance diagnostic model has been trained offline and can be deployed on the edge device for real-time monitoring and online diagnosis of the DFIG operation status in real-world situations.

[0076] Through the above steps, different states of multi-tube faults in the DFIG rotor-side converter can be effectively diagnosed and identified, providing technical theoretical support for the subsequent formulation of DFIG protection strategies.

[0077] DFIG rotor-side converter diagnostic data set preparation: In order to verify the effectiveness and correctness of the proposed method, this embodiment built a three-phase two-level DFIG system based on the MATLAB / Simulink platform. Some of its parameters are shown in Table 1. Figure 7 The real-time digital semi-physical laboratory (RT-LAB) shown in the figure is used to verify the feasibility of the DFIG system built.

[0078] Table 1 Partial parameters of DFIG system

[0079]

[0080] Figure 7In the experiment, the waveform in the oscilloscope is compared with the simulation waveform, and the two are consistent, which proves that the constructed DFIG system is reasonable and can be used in the subsequent diagnosis of multi-tube open-circuit faults of the DFIG rotor-side converter. In the experiment, the DFIG system is set to run for a total of 6 seconds. At the 3rd second, a multi-tube open-circuit fault occurs in the DFIG rotor-side converter. From Table 2, 13 groups of bridge arm voltages under different multi-tube fault categories can be collected, of which the first group of signals is generated in normal operation, and the remaining 12 groups of signals are generated under different fault categories. The three-phase current signals under some fault categories are as follows: Figure 8 shown.

[0081] Table 2 Fault category sample labels and quantities

[0082]

[0083] Through Figure 8 The time domain waveform observation shows that the DFIG has great differences in normal operation and multi-tube fault conditions, and effective fault information can be extracted. However, when multi-tube faults occur on the same side but in different phases, there is little difference in the time domain observation. It is difficult to extract effective fault information based on expert experience, and more sophisticated feature extraction methods are needed for processing. The article "Fault Diagnosis Method of Divergence Multi-Module Sliding Window Slow Feature Analysis" published in the Journal of National University of Defense Technology in 2023 proposed a sliding window feature extraction algorithm to process offline fault data. It can dynamically update data and expand the sample size. After comprehensive consideration, the sliding window size is designed to be 2500 sampling points and the sliding step is 800 sampling points. The three-phase current signals under different fault categories can be divided into 372 samples, and the total number of samples is 372×13=4836. By performing image transformation and image fusion processing on the divided sample data sets, a total of 4836 multi-channel fused images are obtained, which are constructed as a DFIG rotor-side converter multi-tube fault diagnosis data set, such as Fig. 9 The figure shows the mapping result of the 200th sample under some fault categories. In order to avoid wasting computing resources and be more suitable for the small sample task environment, the resolution of each multi-channel fusion image is set to 120×120 after comprehensive consideration. In fault diagnosis problems, small samples usually refer to the case where the ratio of the number of training sets to the total number of (training sets + test sets) is α≤0.5. The multi-channel fusion image dataset is randomly divided into training set, validation set and test set according to 3:1:6. The three sets do not interfere with each other for subsequent model training, storage and testing.

[0084] Fig. 9In the figure, when the DFIG rotor-side converter has multiple open-circuit faults of different types, the diversity of faults is fully reflected in the multi-channel fusion image, and the differences in details, colors and lines provide clear clues for fault classification. Compared with the one-dimensional time series waveform, the multi-channel fusion image shows high discrimination characteristics in identifying and distinguishing different fault types. This characteristic can greatly promote the training process of the ICAM-ConvNeXt model, enabling the model to learn and identify fault features more accurately, thereby improving the diagnostic performance and reliability.

[0085] Diagnostic results and comparative analysis: The platform parameters used for the experimental model are: Intel(R)Core(TM)i7-13650HX processor, NVIDIA GeForce RTX 4050Laptop GPU graphics card, CPU and GPU hybrid training to accelerate the model training process, deep learning framework is Pytorch, and programming language is Python3.11.

[0086] The rationality of hyperparameter selection has a key impact on whether the model can reach the optimal state. The accuracy of the validation set is used as the evaluation indicator for diagnosing the model. The hyperparameter selection of the model is studied. Different hyperparameter indicators such as batch size, optimizer and initial learning rate have a significant impact on the deep learning model. The batch size affects the stability of model training and the corresponding memory requirements. A larger batch size can improve the computing efficiency to a certain extent but lead to unstable training trends. The optimizer determines the way the model parameters are updated. The choice of different optimizers may affect the convergence speed. The initial learning rate determines the initial step size of the model parameter update. Too low a rate will lead to slow convergence, while too high a rate will cause the training process to miss the optimal solution.

[0087] To obtain the best hyperparameters, the model was validated by selecting different batch sizes, optimizers, and initial learning rates. Fig.10 Shown are the validation accuracy of the diagnostic models for different batch sizes, optimizers, and initial learning rates.

[0088] Fig.10 In the experiment, the experimental results show that when the batch size is 32, the Adam gradient descent optimizer is used and the initial learning rate is set to 0.001, the model can achieve a high accuracy rate. This parameter design can not only balance the training efficiency and model performance, but also help the model to converge quickly during the training process and maintain good generalization ability. In order to further improve the performance of the model, the learning rate is adjusted in time during the training process, such as every 10 iterations, the learning rate is reduced by 10%, and the relevant regularization methods are combined to ensure the stability of the model.

[0089] In order to verify the effectiveness of the proposed feature extraction method and diagnostic model, this embodiment follows the above hyperparameter design and inputs the multi-feature fusion image training set into the ICAM-ConvNeXt diagnostic model. Fig.11 As shown in the figure, during the diagnostic model training process, as the number of iterations continued to increase, the model training accuracy was significantly improved, and the loss rate decreased rapidly. At the 15th iteration, the model had shown signs of convergence, and both the accuracy and loss rate reached a relatively stable level. When it was further iterated to the 100th time, the model achieved 100% accuracy in the classification ability of the multi-channel fusion image dataset, and the loss rate dropped to 2.3644×10 -8 , which indicates that the model has not only reached convergence, but also has an excellent fitting effect, that is, the 100th iteration model is saved as the target model for multi-tube open-circuit fault diagnosis of DFIG rotor-side converter.

[0090] In order to verify the generalization performance of the proposed diagnostic model, the test set in the multi-channel fusion image dataset was input into the saved target model, and the average test accuracy reached 98.69%. This embodiment uses the confusion matrix to quantitatively analyze the diagnostic results. The horizontal axis represents the diagnostic prediction results, the vertical axis represents the actual diagnostic results, and the main diagonal elements represent the number of samples with the same predicted results and actual results. The darker the color, the higher the diagnostic accuracy of the fault category. The test diagnosis result confusion matrix is ​​shown in Figure 1. Fig.12 shown.

[0091] Fig.12 The samples under different multi-tube fault categories can be correctly classified, the number of misjudged samples is very small, the classification accuracy of each sample is high, and the model has strong generalization ability. It can classify and diagnose different faults based on the abstraction and combination of features such as color and texture in the multi-channel fusion image. Although the time for offline training of the model is part of the diagnosis process, the response speed during online diagnosis has a greater impact on the diagnosis process. Therefore, the performance of the model in real-time diagnosis is the key indicator for measuring the diagnosis project. In this embodiment, the test set is input into the trained target model, and the time taken to diagnose the occurrence of multi-tube faults is about 4.21s. The model has a response time of seconds, which meets the needs of real-time diagnosis. In the future, it can be considered to deploy the method of this embodiment in edge devices such as the Internet of Things, and perform online diagnosis and real-time monitoring during the actual DFIG operation. This rapid diagnosis capability is very important for real-time detection, ensuring that downtime and potential hazards are minimized when a fault occurs.

[0092] In order to further highlight the ability of the method in this embodiment to learn the characteristics of DFIG rotor-side converters in various fault categories, an unsupervised dimensionality reduction method t-SNE is introduced, which is widely used in high-dimensional visualization and verification of fault sample aggregation capabilities. Fig.13Shown are the t-SNE visualization results of the unprocessed original data, model input layer, and model output layer in the multi-channel fusion image test set.

[0093] Fig.13 In the example, the raw data without any processing showed a large area of ​​overlapping distribution, resulting in poor clustering effect of each fault type and inability to achieve effective distinction; however, after applying the feature extraction method proposed in this embodiment, the clustering distribution of the samples improved, although there was still a certain degree of confusion in the boundary area of ​​the fault type; after the fault data passed through the output layer of the ICAM-ConvNeXt model, the diagnostic model extracted and fully learned the fault features again, and various fault features successfully completed separation and convergence, significantly improving the separation between samples. In the output layer, it can be seen that the same features are clustered and different features are separated, and the overlapping area is small, that is, the effect of step-by-step feature extraction of the raw data after the method proposed in this embodiment is very significant, and the effect of multi-tube fault classification in the DFIG rotor-side converter is excellent. The t-SNE visualization results prove that the method of this embodiment has high sensitivity and strong recognition ability for the open circuit fault diagnosis of the DFIG rotor-side converter under small samples.

[0094] (1) Ablation experiment: To verify the advantages of the feature extraction method of this embodiment, the experiment is carried out while keeping the ICAM-ConvNeXt model unchanged. The GADF method in the image fusion processing is replaced by the RP method, the MTF method, the GASF method and the wavelet transform method. To save space, only the following are shown. Fig.14 The DFIG rotor-side converter shown in Figure 1 is 1 and T 6 The 150th feature image generated by different schemes after the fault occurs. The feature images generated by different schemes are used as model inputs respectively. To avoid the occurrence of low-probability phenomena, multiple experiments are conducted. In order to evaluate the performance of different feature maps for fault diagnosis, accuracy, precision, recall rate and F1 score are selected as evaluation indicators. The fault diagnosis performance evaluation indicators under different feature maps are as follows: Fig.15 (a) shown.

[0095] (2) Comparative experiment: To verify the advantages of the ICAM-ConvNeXt model in this embodiment, the experiment was conducted while ensuring that the feature extraction method remained unchanged. Multiple comparative experiments were conducted with models ①-④: the unimproved CAM-ConvNeXt model, the ConvNeXt-T model, the ResNet model, and the Swin-Transformer model. The samples in the multi-channel fusion image dataset were used as the input of each model, and the above-mentioned evaluation indicators were also used to measure the quality of the diagnosis model. The fault diagnosis performance evaluation indicators under different models are as follows: Fig.15 (b) as shown.

[0096] Fig.15 In (a), under the condition of ensuring the consistency of the ICAM-ConvNeXt model, the feature extraction method of this embodiment is superior to other solutions in different evaluation indicators, and the accuracy is improved by 29.22%, 4.48%, 1.61% and 34.63% compared with other solutions. The GADF method, which has the best application effect of the feature enhancement algorithm based on multi-channel fusion proposed in this embodiment, has the advantage of converting one-dimensional signals into two-dimensional images in that it can intuitively reveal the time series dynamic changes of data by calculating the angle difference between time series data points, which not only retains the original information of the time series, but also provides an intuitive representation form through the generated angle difference image, which is particularly suitable for capturing and analyzing local trends in time series in small task learning, thereby showing high efficiency and accuracy in fault diagnosis; Fig.15 In (b), while ensuring the consistency of feature extraction methods, the ICAM-ConvNeXt model of this embodiment outperforms other solutions in different evaluation indicators, and improves the accuracy by 1.68%, 2.57%, 8.7% and 4.47% compared with other solutions. By introducing an improved channel attention mechanism, the ICAM-ConvNeXt model not only reduces the computational complexity of the model, but also adaptively adjusts the important weights of channel features, enhances the model's ability to capture key fault features, and thus improves the accuracy and efficiency of fault identification.

[0097] In summary, the combination of the feature enhancement algorithm based on multi-channel fusion and the ICAM-ConvNeXt model can accurately and efficiently classify and diagnose the multi-tube open-circuit faults of the DFIG rotor-side converter.

[0098] Performance analysis of the diagnostic method under multiple operating conditions: In actual DFIG operation, the motor often exists in multiple operating modes, so it is also very important for the diagnostic model to perform fault diagnosis under multiple operating conditions. Based on the DFIG rotor-side converter multi-tube open-circuit fault diagnosis target model, the fault data of operating condition 1 with a motor speed of 1600r / min and operating condition 2 with a motor speed of 1700r / min are used for testing to verify the adaptability of the model under multiple operating conditions, such as Fig.16 Shown is the confusion matrix classification result.

[0099] Fig.16 The method of this embodiment can accurately identify multi-tube open circuit faults at different speeds in working conditions 1 and 2, with fewer misjudgments. The fault diagnosis accuracy rate reaches 98.00% and 98.17%. The addition of ICAM effectively improves the global feature extraction capability of the ConvNeXt model, and further enhances the generalization ability of the model.

[0100] Performance analysis of the diagnostic method under extreme conditions: During the actual operation of DFIG, when a small sample occurs, there may be interference from extreme conditions such as sample imbalance and external noise. In order to verify the applicability of the method of this embodiment under extreme conditions, the following experiments are conducted:

[0101] (1) In the DFIG fault diagnosis process, although single-tube faults are more harmful to the system, their frequency of occurrence is relatively low, and the probability of double-tube failures is even rarer. This makes it difficult to collect sufficient fault samples in practical applications. This sample scarcity phenomenon will cause a serious imbalance in the number ratio between normal data and fault data, which will pose a challenge to the training and test evaluation of the diagnosis model and make it difficult to explore the internal fault rules. In order to better meet the multi-tube open-circuit faults of the DFIG rotor-side converter under real working conditions, the training data simulates the sample unbalanced data under real working conditions for experiments. According to the frequency of occurrence of different types of faults in the DFIG rotor-side converter, the multi-channel fusion image data set is redistributed into 300 groups of normal categories and 200 groups of other fault categories, totaling 300+200×12=2700 groups of images, which are randomly divided at a ratio of 2:3 and sent to the ICAM-ConvNeXt model for training and testing. The training and test results under the unbalanced data set are shown in Figure 2. Fig.17 and Fig.18 shown.

[0102] Fig.17 In the process of training, as the iterations continue, the accuracy of the model increases rapidly and the loss rate decreases rapidly. When the number of iterations of the multi-channel fusion image imbalanced data set is 18, the accuracy and loss rate reach a convergence state. When the number of iterations is 100, the best effect is achieved, with a training accuracy of 100% and a loss rate of 2.5856×10 -6 At this moment, the training model has reached convergence and the fitting effect is perfect. Fig.18 The average test accuracy of the unbalanced data set of multi-channel fusion images reached 96.54%. The test results show that the method of this embodiment has high recognition accuracy of various faults under unbalanced sample conditions and low misjudgment rate, and can diagnose the location of multi-tube faults more accurately, proving that the proposed method also has good generalization and accuracy for imbalance problems under small samples.

[0103] (2) In actual DFIG projects, the friction and vibration between the parts of the motor during mechanical rotation will generate noise, which will cause the three-phase signal collected by the current sensor to be mixed with noise, covering up the real fault information and causing the model to deviate during the diagnosis process. In order to verify the noise resistance of this model in reality, the above multi-channel fusion image dataset is used to add white noise with Gaussian distribution to the dataset at different signal-to-noise ratios (SNR) to simulate actual work.

[0104] Noise signals of -10dB, -5dB, 0dB, 5dB and 10dB were added to the multi-channel fusion image dataset to train and test the model of this embodiment and different comparison models in the comparative experiment (unimproved CAM-ConvNeXt model, ConvNeXt-T model, ResNet model, Swin-Transformer model). The number of iterations, number of samples and division method were consistent with the previous text. The experiment was carried out multiple times and the average accuracy was calculated. The anti-noise performance results under balanced datasets and unbalanced datasets are shown in Table 3.

[0105] Table 3 Noise resistance performance of different models

[0106]

[0107]

[0108] In Table 3, regardless of the type of data set, the model of this embodiment shows a diagnostic accuracy that exceeds other models under different SNR conditions, especially in the extreme noise environment of SNR = -10dB. Combined with the imbalance of the data set and the scarcity of samples, the average diagnostic accuracy of the model of this embodiment can still reach 86.38%, which is significantly higher than other existing models. With the improvement of SNR, although the diagnostic accuracy of other models has increased, it still does not exceed the performance of the model of this embodiment. Because this embodiment combines the multi-channel fusion image with the ICAM-ConvNeXt model in an innovative way, the method can fully capture the fault characteristics and effectively eliminate the interference of redundant information, thereby weakening the noise interference to a certain extent, giving the model a better anti-noise characteristic. The experimental results show that the method proposed in this study can achieve efficient and accurate fault diagnosis for the DFIG rotor-side converter after a multi-tube open-circuit fault occurs, and has a strong ability to resist noise interference.

[0109] In summary, this embodiment combines information fusion enhancement with deep learning networks and introduces them into the problem of power equipment fault diagnosis. It proposes a DFIG rotor-side converter multi-tube open-circuit fault diagnosis method based on multi-channel fusion and ICAM-ConvNeXt model under small samples, which solves the problems of insufficient generalization performance of the diagnosis model, decreased accuracy, and difficulty in diagnosing the location of the faulty IGBT caused by the scarcity of target domain samples. It can provide new ideas for the open-circuit fault diagnosis problems related to the motor field. The following conclusions are drawn:

[0110] (1) A feature enhancement algorithm based on multi-channel fusion is proposed. The GADF technology is used to transform and fuse the DFIG three-phase current signals according to different channels into a single image. The image contains the three-phase fault feature information and can distinguish the hidden features of the DFIG rotor-side converter under different fault types, realize feature complementary enhancement display, and provide a rich feature basis for subsequent fault diagnosis in small sample tasks.

[0111] (2) An ICAM-ConvNeXt model based on an improved channel attention mechanism is proposed, which significantly improves the model's feature extraction capability and diagnostic performance. The ICAM module effectively reduces the model complexity by replacing ordinary convolution with deep separable convolution, while maintaining sensitivity to the relationship between channels, which helps to more accurately identify minor faults in power equipment. ConvNeXt has a powerful feature representation capability. When combined with ICAM, it can more efficiently capture fault features, thereby achieving accurate fault diagnosis. This combination provides a more advanced technical means for power equipment fault diagnosis.

[0112] (3) The method of this embodiment shows better performance in small sample tasks. Compared with other methods, it has better diagnostic accuracy for multiple tube fault types and strong anti-disturbance ability under multiple working conditions and extreme conditions. It has high anti-interference and robustness.

[0113] 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 and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A diagnostic method for multi-tube open circuit faults in a doubly-fed wind power converter, characterized in that: Based on the ConvNeXt-T model, the ICAM modules are connected in series to build and train the ICAM-ConvNeXt model. The trained ICAM-ConvNeXt model is used to perform real-time monitoring and online diagnosis of the operating status of DFIG.

2. The method for diagnosing multi-tube open circuit faults in a doubly-fed wind power converter according to claim 1, characterized in that: The training steps of the ICAM-ConvNeXt model are as follows: S1. According to the working mechanism of DFIG converter under normal and fault conditions, a DFIG simulation model is established and three-phase current is obtained as an effective fault parameter; S2, using Gram angle difference field coding to map the three-phase current signals into color GADF images respectively, extracting the images of the red channel, green channel and blue channel from the three GADF images respectively, and fusing them according to the channel dimension to obtain a fused image; S3, randomly dividing the fused image dataset into a training set for model training, a validation set for model adjustment, and a test set for performance evaluation; S4. Input the training set into the initialized ICAM-ConvNeXt model and perform multiple rounds of iterative training. Use the validation set to evaluate the optimal performance of the model and record the best training parameters. Use the test set to evaluate the classification and diagnostic accuracy of the model. If the preset performance standard is not met, the model will re-enter the training stage and be optimized again until the classification and diagnostic accuracy of the model reaches the preset performance standard.

3. The diagnostic method for multi-tube open circuit fault of a doubly-fed wind power converter according to claim 2, characterized in that: During diagnosis, the three-phase current is measured and sampled in real time through the current sensor of DFIG, and the three-phase current signals are mapped into color GADF images respectively. The images of the red channel, green channel and blue channel are extracted from the three GADF images respectively, and fused according to the channel dimension to obtain a real-time fused image; the trained ICAM-ConvNeXt model is used to extract features of the real-time fused image, and fault classification and identification are realized.

4. The method for diagnosing multi-tube open circuit faults in a doubly-fed wind power converter according to claim 1, characterized in that: The ICAM module is based on the CAM module and adopts depthwise separable convolution DSC to replace the ordinary full convolution in the CAM module. The depthwise separable convolution DSC is followed by a BN layer and a Relu layer.

Images