Power converter open-circuit fault diagnosis method based on physical information transformer network

Abstract

The application discloses a power converter open-circuit fault diagnosis method based on a physical information transformer network, first, a training sample is acquired, and a power converter open-circuit fault diagnosis model based on a physical information transformer network is constructed, an amplitude and a curvature of an input three-phase current signal are extracted by a physical feature extraction module, and data point physical features are extracted therefrom, a dynamic weight generation module generates dynamic weights according to the data point physical features, a curvature exceeds feature generation module generates a curvature exceeds feature according to the curvature of the three-phase current signal, a transformer encoder is used for embedding the three-phase current signal to obtain signal features, then, attention features are extracted from the signal features, differential features of the signal features and the curvature exceeds feature according to the dynamic weights, and the attention features are fused, and finally, the fused attention features are sent to a detection module for detection after being processed by a feedforward neural network. The application enhances the fault information capturing capability and improves the open-circuit fault diagnosis accuracy by using a self-adaptive dimension selection mask strategy and a multi-branch self-attention.

Description

Technical Field

[0001] This invention belongs to the field of power converter fault diagnosis technology, and more specifically, relates to a power converter open-circuit fault diagnosis method based on physical information transformer network. Background Technology

[0002] With the global trend towards transitioning to environmentally friendly and renewable energy, renewable energy sources such as wind and solar power are gradually being integrated into public power grids, making power converters increasingly important. As key electronic devices, power converters are widely used in grid connection, wind power conversion, and photovoltaic power generation, improving energy conversion efficiency and reducing harmonic distortion. However, power converters such as IGBTs and MOSFETs are susceptible to failure due to high power losses, high thermal stress, and rising ambient temperatures. Fault types are mainly classified as short-circuit faults and open-circuit faults. Open-circuit faults, in particular, have insidious initial symptoms and can easily trigger more serious secondary faults, causing significant economic losses. Therefore, research on open-circuit fault diagnosis technology for power converters is of great importance. Existing power converter diagnostic technologies are mainly divided into two categories: model-based diagnostic methods and data-driven diagnostic methods.

[0003] The core principle of model-based diagnostic methods is to diagnose faults by evaluating the consistency between the predicted output of an industrial process mathematical model and measured reference values. Their effectiveness depends on a model that accurately simulates the dynamic characteristics of the power converter. For output series-interleaved boost converters, Xu et al. proposed a fault-tolerant control strategy using an immersion invariant observer (I&IO), which can quickly detect and identify open-circuit switch faults within two switching cycles. For voltage source inverters, An et al. proposed an open-circuit fault diagnosis method based on a switching function model, achieving rapid diagnosis by indirectly analyzing the collector-emitter voltage of the power switch, thus reducing system cost and complexity. However, such methods rely on accurate system models, which are difficult to construct accurately for complex industrial systems and may require additional sensors, increasing costs.

[0004] To address the limitations of model-based methods (especially their reliance on high-fidelity system models), data-driven approaches offer an effective solution. These methods do not require explicit physical models; instead, they leverage historical operational data to learn the complex relationships between system measurements and health status, thereby enabling fault diagnosis and classification. Related techniques include the Random Forest (RF) algorithm (suitable for systems with varying loads, but requires pre-computation and has insufficient accuracy), Support Vector Machines (SVM) (capable of detecting multiple faults in a short time, but requires additional physical validation), and Bayesian networks (simple to tune and implement, but cannot be tested online and is unsuitable for real-time fault diagnosis).

[0005] To address the problems of traditional data-driven methods, deep learning has become the mainstream approach for open-circuit fault diagnosis in power converters in recent years. Despite the significant achievements of deep learning-based methods, their inherent weakness in interpretability remains a major concern. In particular, these models fail to incorporate industrial physics context information into their internal mechanisms, making it difficult to guarantee that their outputs fully conform to fundamental physical laws. Furthermore, these methods heavily rely on expert experience for feature extraction, which significantly disrupts the inherent data structure of the original monitoring signals, leading to the loss of crucial fault information. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for diagnosing open-circuit faults in power converters based on a physical information transformer network. A physical information transformer network (PITN) based on a physical information self-attention mechanism is constructed, and an adaptive dimension selection masking strategy and multi-branch self-attention are combined to enhance the ability to capture fault information, thereby improving the accuracy of open-circuit fault diagnosis.

[0007] To achieve the above-mentioned objective, the power converter open-circuit fault diagnosis method based on physical information transformer networks of the present invention includes the following steps:

[0008] S1: Collect several fault samples under different open-circuit faults for the power converter. Each fault sample includes data points of length [length missing]. The three-phase current signals and their corresponding open-circuit fault tags. The value is determined based on actual needs;

[0009] S2: Construct a power converter open-circuit fault diagnosis model based on a physically-informed transformer network, including a physical feature extraction module, a dynamic weight generation module, a curvature exceedance feature generation module, a transformer encoder, and a detection module, wherein:

[0010] The physical feature extraction module is used to extract features from the input three-phase current signal. Extract the amplitude of each data point in each phase channel. and curvature ,in Indicates the channel index. , Indicates the index of the data point. Generate amplitude attention markers for each data point in each phase channel. and curvature attention mark ,in or This indicates that you are interested in the corresponding parameter. or This indicates that the corresponding parameters are not considered, and then the fused physical characteristics are obtained. For each data point, the maximum value is selected from the fused physical features of the three phase channels as the physical feature of the data point. And send it to the dynamic weight generation module, while simultaneously calculating the curvature of each data point in each phase channel. Send to the curvature exceedance feature generation module;

[0011] The dynamic weight generation module is used to generate weights based on the physical characteristics of data points. Generate each data point Dynamic weights Then it is sent to the transformer encoder;

[0012] The curvature exceedance feature generation module is used to generate features based on curvature. For the input three-phase current signal Generate curvature beyond features And sent to the transformer encoder, curvature exceeding the feature The generation method is as follows:

[0013] First, calculate the curvature feature matrix. Each element The calculation formula is as follows:

[0014] ,

[0015] in, These represent the preset upper and lower limits of curvature, respectively. Represents the ReLU activation function;

[0016] Then, the curvature excess feature is calculated using the following formula. :

[0017] ,

[0018] in, It represents the Hadamah accumulation. Represents a linear mapping function;

[0019] Transformer encoders are used for dynamic weight-based encoders. and curvature beyond features For the input three-phase current signal Feature extraction is performed to obtain signal features. And send it to the detection module; the Transformer encoder includes an input embedding layer, a position encoding layer, an embedding fusion module, a normalization module, a difference module, a multi-branch adaptive attention module, an attention feature fusion module, and a feedforward neural network, wherein:

[0020] The input embedding layer is used for the input three-phase current signal. Data embedding is performed, and the resulting embedded features are sent to the embedding fusion module;

[0021] The position coding layer is used for the input three-phase current signal. Perform position encoding and send the obtained position encoding to the embedding fusion module;

[0022] The embedding fusion module is used to fuse the position encoding with the embedding features, and sends the resulting fused features to the normalization module.

[0023] The normalization module is used to normalize the fused features, and the resulting normalized features are used as signal features. The data is sent to the differential module and the multi-branch adaptive attention module, respectively.

[0024] Differential modules are used to characterize signals. Perform first-order difference operations on the data points to obtain the difference features. And send it to the multi-branch adaptive attention module;

[0025] The multi-branch adaptive attention module is used to adjust attention based on dynamic weights. Signal characteristics Difference characteristics and curvature beyond features Feature extraction was performed using a self-attention mechanism to obtain attention features. , , And sent to the attention feature fusion module; attention features The generation method is as follows:

[0026] 1) Based on dynamic weights Calculate the dynamic mask range for each data point :

[0027] ,

[0028] in, , These represent the upper and lower limits of the preset dynamic mask, respectively. Indicates the preset coefficient;

[0029] 2) Based on the upper limit of dynamic mask Define a neighborhood for each data point :

[0030] ;

[0031] 3) For input features Calculate the query matrix Key matrix Sum matrix :

[0032] ,

[0033] in, , , , These represent the learnable weight matrices of the corresponding matrices;

[0034] 4) Based on neighborhood Calculated data points and its neighboring data points Attention scores between :

[0035] ,

[0036] in, The vector dimension of the key matrix is ​​represented by the superscript. Indicates transpose;

[0037] 5) Calculate the data points based on the dynamic mask. and its neighboring data points Adaptive mask between :

[0038] ,

[0039] in, Indicates rounding down;

[0040] 6) Calculate the data points using the following formula. and its neighboring data points Attention weights between :

[0041] ;

[0042] 7) Attention features are calculated using the following formula. :

[0043] ;

[0044] in, Representing attention features Medium data points eigenvalues;

[0045] The attention feature fusion module is used to perform dynamic weight fusion. Three attention features , , By performing fusion, fusion characteristics are obtained. The data is sent to the feedforward neural network, and the fusion formula is as follows:

[0046] ,

[0047] in, Represents a dynamic weight vector;

[0048] Feedforward neural networks are used for fused features Perform nonlinear transformation to obtain signal characteristics ;

[0049] The detection module is used to detect based on signal characteristics. Obtain the input three-phase current signal The probability of belonging to various open-circuit faults;

[0050] S3: Use the fault samples obtained in step S1 to train the power converter open circuit fault diagnosis model to obtain a trained power converter open circuit fault diagnosis model.

[0051] S4: When open-circuit fault diagnosis of the power converter is required, the number of data points to be collected is: The three-phase current signal is then input into the trained power converter open-circuit fault diagnosis model to obtain the open-circuit fault diagnosis result.

[0052] This invention relates to a power converter open-circuit fault diagnosis method based on a physical information transformer network. First, training samples are acquired. Then, a power converter open-circuit fault diagnosis model based on a physical information transformer network is constructed. A physical feature extraction module extracts the amplitude and curvature of the input three-phase current signal and extracts the physical features of the data points. A dynamic weight generation module generates dynamic weights based on the physical features of the data points. A curvature excess feature generation module generates curvature excess features on the three-phase current signal based on the curvature. A Transformer encoder is used to embed the three-phase current signal to obtain signal features. Then, attention features are extracted from the signal features, differential features of the signal features, and curvature excess features based on the dynamic weights, and these features are fused. Finally, a feedforward neural network processes the data and sends it to the detection module for detection.

[0053] The present invention has the following beneficial effects:

[0054] 1) This invention designs an innovative end-to-end diagnostic framework that embeds physical information. This framework enhances model interpretability while significantly reducing dependence on large-scale training data and improving adaptability in practical engineering applications.

[0055] 2) In this invention, dynamic weights are generated based on the physical characteristics of the data points of the three-phase current signal. Based on this, an adaptive dimension selection masking strategy is designed to reduce the training cost and memory pressure of long-term sequence data.

[0056] 3) This invention designs a multi-branch self-attention mechanism, which extracts attention features from signal features, differential features of signal features and curvature exceedance features according to dynamic weights and then fuses them, thereby enhancing the ability to capture fault information and improving the accuracy of fault diagnosis. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the specific implementation method of the power converter open-circuit fault diagnosis method based on physical information transformer networks;

[0058] Figure 2 This is a structural diagram of the power converter open-circuit fault diagnosis model in this invention;

[0059] Figure 3 This is an example diagram of the attention marker in this embodiment;

[0060] Figure 4 This is an example diagram of dynamic weights in this embodiment;

[0061] Figure 5 This is a flowchart of attention feature generation in this invention. Detailed Implementation

[0062] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0063] Example

[0064] Figure 1 This is a flowchart illustrating the specific implementation of a power converter open-circuit fault diagnosis method based on a physical information transformer network. (For example...) Figure 1 As shown, the power converter open-circuit fault diagnosis method based on physical information transformer network of the present invention includes the following steps:

[0065] S101: Obtain the training sample set:

[0066] Several fault samples were collected from the power converter under different open-circuit fault conditions. Each fault sample included data points of length [length missing]. The three-phase current signals and their corresponding open-circuit fault tags. The value is determined based on actual needs.

[0067] Generally, to achieve more comprehensive detection, open-circuit faults include single-tube and dual-tube faults. To improve detection accuracy, the data points of each three-phase current signal need to cover at least one fault cycle.

[0068] S102: Constructing a power converter open-circuit fault diagnosis model based on a physical information transformer network:

[0069] In this invention, a transformer network is selected as the backbone network to construct an open-circuit fault diagnosis model for power converters. Compared with other neural networks (such as CNN, LSTM, and GRU), the transformer network has significant advantages in processing long-term sequence data due to its self-attention mechanism. Furthermore, it can directly use three-phase current signals as input, effectively avoiding disruption of the inherent data structure. However, its computational complexity is... This leads to a heavy computational burden. Furthermore, the structural design of transformer networks lacks guidance from physical information derived from the monitoring signals of the power converter, resulting in poor interpretability. To address these issues, this invention constructs a Physical Information Transformer Network (PITN) based on a physical information self-attention mechanism, and uses this as the foundation to build an open-circuit fault diagnosis model for the power converter. Figure 2 This is a structural diagram of the power converter open-circuit fault diagnosis model in this invention. (See diagram below.) Figure 2 As shown, the power converter open-circuit fault diagnosis model in this invention includes a physical feature extraction module, a dynamic weight generation module, a curvature exceedance feature generation module, a transformer encoder, and a detection module. Each module will be described in detail below.

[0070] The physical feature extraction module is used to extract features from the input three-phase current signal. Extract the amplitude of each data point in each phase channel. and curvature ,in Indicates the channel index. , Indicates the index of the data point. Generate amplitude attention markers for each data point in each phase channel. and curvature attention mark ,in or This indicates that you are interested in the corresponding parameter. or This indicates that the corresponding parameters are not considered, and then the fused physical characteristics are obtained. For each data point, the maximum value is selected from the fused physical features of the three phase channels as the data point's physical feature. And send it to the dynamic weight generation module, while simultaneously calculating the curvature of each data point in each phase channel. The signal is sent to the curvature exceedance feature generation module. Amplitude and curvature, these two physical features, can effectively reflect the changes in the three-phase current signal under fault conditions, thereby extracting the fault characteristics.

[0071] Curvature in this embodiment The calculation method is as follows:

[0072] First, calculate the difference in amplitude. and :

[0073] ,

[0074] ,

[0075] Then, the curvature is calculated using the following formula. :

[0076] ,

[0077] in, This indicates mean filtering.

[0078] Since this invention implements fault detection, it is necessary to pay more attention to outliers. Therefore, in this embodiment, the focus is determined based on whether the corresponding parameter is an outlier. The specific method is as follows:

[0079] Calculate the mean and standard deviation of the corresponding parameter series:

[0080] , ,

[0081] , ,

[0082] Then, the attention marker is determined using the following formula:

[0083] ,

[0084] ,

[0085] in, , These represent the preset coefficients. This indicates an indicator function; its value is 1 when the content is true, and 0 otherwise.

[0086] Figure 3 This is an example diagram of the attention marker in this embodiment. For example... Figure 3 As shown, the curvature and amplitude "regions of interest" of each phase current signal differ significantly. Therefore, the combination problem of the entire three-phase current signal is solved by selecting the maximum value from the fused physical characteristics of each phase channel. The index of the selected channel is... Each data point is different, meaning that the physical characteristics will change dynamically as the data point index changes.

[0087] The dynamic weight generation module is used to generate weights based on the physical characteristics of data points. Generate each data point Dynamic weights Then it is sent to the transformer encoder.

[0088] Physical characteristics of data points While it can represent the main "region of interest" in a three-phase current signal, it cannot reflect the correlation between adjacent data points and has poor robustness to outlier data points. Therefore, dynamic weights are needed for calibration. In this embodiment, the dynamic weight generation module includes a convolutional layer, a normalization layer, and a nonlinear transform layer, wherein:

[0089] Convolutional layers are used to analyze the physical features of data points. Perform convolution to obtain convolutional features. And sent to the normalization layer. Convolutional features in this embodiment The calculation formula is as follows:

[0090] .

[0091] in, Represents the convolution kernel. It is a bias term.

[0092] Normalization layers are used to normalize convolutional features Normalize to the range [0,1], then normalize the features. The data is sent to the nonlinear transformation layer. The normalization calculation formula in this embodiment is as follows:

[0093] .

[0094] Nonlinear transformation layers are used for normalized features Dynamic weights are obtained by performing nonlinear transformations. The calculation formula for the nonlinear transformation in this embodiment is as follows:

[0095] ,

[0096] in, This indicates a preset threshold.

[0097] Figure 4 This is an example diagram of dynamic weights in this embodiment. For example... Figure 4 As shown, this value increases when the signal fluctuates sharply, and decreases when the signal tends to smooth out. This indicates that the weight... It can accurately capture physical characteristics, thereby ensuring the diagnostic performance of the PITN-based power converter open-circuit fault diagnosis model in this invention.

[0098] The curvature exceedance feature generation module is used to generate features based on curvature. For the input three-phase current signal Generate curvature beyond features And sent to the transformer encoder, curvature exceeding the feature The generation method is as follows:

[0099] First, calculate the curvature feature matrix. Each element The calculation formula is as follows:

[0100] ,

[0101] in, These represent the preset upper and lower limits of curvature, respectively. This represents the ReLU activation function.

[0102] Then, the curvature excess feature is calculated using the following formula. :

[0103] ,

[0104] in, It represents the Hadamah accumulation. This represents a linear mapping function.

[0105] It is evident that the curvature exceeds the characteristic It can effectively reflect the characteristic information of the curvature-focused region.

[0106] Transformer encoders are used for dynamic weight-based encoders. and curvature beyond features For the input three-phase current signal Feature extraction is performed to obtain signal features. And send it to the detection module. For example... Figure 2As shown, the Transformer encoder in this invention includes an input embedding layer, a position encoding layer, an embedding fusion module, a normalization module, a difference module, a multi-branch adaptive attention module, an attention feature fusion module, and a feedforward neural network, wherein:

[0107] The input embedding layer is used for the input three-phase current signal. Data embedding is performed, and the resulting embedded features are sent to the embedding fusion module.

[0108] The position coding layer is used for the input three-phase current signal. Perform position encoding and send the obtained position encoding to the embedding fusion module.

[0109] The embedding fusion module is used to fuse position encoding with embedding features, and the resulting fused features are sent to the normalization module.

[0110] The normalization module is used to normalize the fused features, and the resulting normalized features are used as signal features. The data is sent to the differential module and the multi-branch adaptive attention module, respectively.

[0111] Differential modules are used to characterize signals. Perform first-order difference operations on the data points to obtain the difference features. And send it to the multi-branch adaptive attention module. Differential features Each data point The calculation formula can be expressed as follows:

[0112] .

[0113] The multi-branch adaptive attention module is based on dynamic weights Signal characteristics Difference characteristics and curvature beyond features Feature extraction was performed using a self-attention mechanism to obtain attention features. , , It is then sent to the attention feature fusion module. Figure 5 This is a flowchart of attention feature generation in this invention. For example... Figure 5 As shown, the specific steps for generating attention features in this invention include:

[0114] S501: Generate dynamic mask range:

[0115] To reduce computational complexity, this invention designs an adaptive dimension selection masking strategy that embeds physical information. Specifically, for each data point, a masking strategy is developed based on dynamic weights. This determines the size of the dynamic mask within the attention module, thereby ensuring maximum utilization of computational resources. Dynamic mask range The calculation formula is as follows:

[0116] ,

[0117] in, , These represent the upper and lower limits of the preset dynamic mask, respectively. This indicates the preset coefficient.

[0118] S502: Determine the neighborhood of the data point:

[0119] To achieve GPU parallel computing, this invention is based on a dynamic mask upper limit. Define a neighborhood for each data point This allows us to limit the range of attention scores. Neighborhood The calculation formula is as follows:

[0120] .

[0121] S503: Calculate query-key-value matrix triples:

[0122] For input features Calculate the query matrix Key matrix Sum matrix :

[0123] ,

[0124] in, , , , These represent the learnable weight matrices of the corresponding matrices.

[0125] S504: Calculate attention score:

[0126] According to the neighborhood The data points can then be calculated. and its neighboring data points Attention scores between :

[0127] ,

[0128] in, The vector dimension of the key matrix is ​​represented by the superscript. This indicates transpose.

[0129] S505: Calculate the adaptive mask:

[0130] Data points are calculated based on dynamic masks. and its neighboring data points Adaptive mask between :

[0131] ,

[0132] in, This indicates rounding down to the nearest integer.

[0133] As can be seen, during the adaptive mask generation, the difference range of data point indices is... That is, the mask range is .

[0134] S506: Calculate attention weights:

[0135] The data points are calculated using the following formula. and its neighboring data points Attention weights between :

[0136] .

[0137] S507: Calculate attention features:

[0138] Attention features are calculated using the following formula. :

[0139] ,

[0140] in, Representing attention features Medium data points eigenvalues.

[0141] The attention feature fusion module is used to perform dynamic weight fusion. Three attention features , , By performing fusion, fusion characteristics are obtained. The data is sent to the feedforward neural network, and the fusion formula is as follows:

[0142] ,

[0143] in, This represents the dynamic weight vector.

[0144] Feedforward neural networks are used for fused features Perform nonlinear transformation to obtain signal characteristics .

[0145] The detection module is used to detect based on signal characteristics. Obtain the input three-phase current signal The probability of belonging to various open circuit faults.

[0146] S103: Training a power converter open-circuit fault diagnosis model:

[0147] The fault samples obtained in step S101 are used to train the power converter open-circuit fault diagnosis model to obtain a trained power converter open-circuit fault diagnosis model.

[0148] In this embodiment, the loss function adopted is the label-smoothed cross-entropy loss function. Compared with the conventional cross-entropy loss, this loss function can effectively reduce overfitting. Its calculation formula is as follows:

[0149] ,

[0150] in, This represents the total number of open-circuit fault categories. , Indicates belonging to the first The true probability and predicted probability of open-circuit faults This represents the label smoothing factor.

[0151] In this embodiment, the Adam optimizer is used to update the weight parameters, and a cosine annealing learning rate scheduler is used to ensure training stability. The formula for calculating the learning rate is as follows:

[0152] ,

[0153] in, Indicates the current training round. This represents the current learning rate. This represents the minimum learning rate. This represents the maximum learning rate. This represents the total number of training rounds.

[0154] S104: Power converter open circuit fault diagnosis:

[0155] When open-circuit fault diagnosis of a power converter is required, the number of data points to be collected is: The three-phase current signal is then input into the trained power converter open-circuit fault diagnosis model to obtain the open-circuit fault diagnosis result.

[0156] To better illustrate the technical effects of the present invention, specific examples are used to experimentally verify the present invention.

[0157] This embodiment presents a self-designed practical fault diagnosis hardware experimental platform, designed to simulate open-circuit faults and the dynamic characteristics of power converters in a real industrial environment, thereby collecting the dataset required for experimental verification. The experimental platform comprises several key subsystems: a power grid simulation power supply, a tekmdo34 microcontroller, an electronic load, a power supply circuit, an industrial personal computer, and a control sampling system. Table 1 shows the key parameters of the experimental platform in this embodiment.

[0158] parameter value parameter value Rated DC voltage 700V Power grid AC voltage measurement 220V / 50Hz Controller Model DSP283773 Rated power 10kW Filter capacitors 480uF Filter Inductor 4.3mL Sampling frequency 800kHz Control frequency 20kHz

[0159] Table 1

[0160] This embodiment simulates 21 types of open-circuit faults using a power supply circuit subsystem, including single-switch faults and mixed-switch faults, and simulates operating conditions of 2.5~10kW by controlling the electronic load subsystem. Table 2 is a list of open-circuit fault types and their corresponding labels in this embodiment.

[0161] Fault type Label Fault type Label Fault type Label [T1] 0 <![CDATA[T2-T3]]> 7 <![CDATA[T3-T6]]> 14 <![CDATA[T1-T2]]> 1 <![CDATA[T2-T4]]> 8 <![CDATA[T4]]> 15 <![CDATA[T1-T3]]> 2 <![CDATA[T2-T5]]> 9 <![CDATA[T4-T5]]> 16 <![CDATA[T1-T4]]> 3 <![CDATA[T2-T6]]> 10 <![CDATA[T4-T6]]> 17 <![CDATA[T1-T5]]> 4 <![CDATA[T3]]> 11 <![CDATA[T5]]> 18 <![CDATA[T1-T6]]> 5 <![CDATA[T3-T4]]> 12 <![CDATA[T5-T6]]> 19 <![CDATA[T2]]> 6 <![CDATA[T3-T5]]> 13 <![CDATA[T6]]> 20

[0162] Table 2

[0163] Using the experimental platform described above, this embodiment designs two fault diagnosis tasks: independent and identically distributed (i.i.d.) and non-independent and identically distributed (i.i.d.) faults. For the former, the training and test sets follow an independent and identically distributed (i.i.d.) distribution; for the latter, the training and test sets follow different distributions. It should be noted that the ratio of the training set to the test set is 4:1. Table 3 shows the dataset partitioning information for the independent and identically distributed fault diagnosis task in this embodiment. Table 4 shows the dataset partitioning information for the non-independent and identically distributed (i.i.d.) fault diagnosis task in this embodiment.

[0164] Dataset Load power Sample size a 2.5-3.5kW 22680 b 4-7.5kW 60480 c 7.5-10kW 45360

[0165] Table 3

[0166] Dataset Training set load power Load power of the test set a 2.5-6kW 8.5-9kW b 4-7.5kW 2.5-3kW c 6.5-10kW 3.5-4kW

[0167] Table 4

[0168] To demonstrate the superiority of the PITN-based open-circuit fault diagnosis method for power converters proposed in this invention, six existing methods were used as comparison methods in this embodiment. Three of these are traditional machine learning methods, including Support Vector Machine (SVM), Random Forest (RF), and k-Nearest Neighbors (KNN); the other three are deep learning methods, including Convolutional Neural Network (CNN), Long Short-Term Memory-Convolutional Neural Network (CNN-LSTM), and Transformer. Accuracy, precision, and F1 score in fault diagnosis were used as evaluation metrics. Table 5 compares the experimental results of the present invention and the comparison methods on the independent and identically distributed fault diagnosis task in this embodiment. Table 6 compares the experimental results of the present invention and the comparison methods on the independent and identically distributed fault diagnosis task in this embodiment.

[0169] method Accuracy (%) Accuracy (%) F1 score (%) SVM 55.59±18.28 58.45±17.46 55.24±18.74 KNN 71.83±5.66 72.03±5.86 71.57±5.85 RF 72.29±5.62 72.60±5.87 72.05±5.78 CNN 82.42±2.16 83.54±1.97 82.05±2.34 CNN-LSTM 93.51±1.16 93.58±1.14 93.49±1.18 Transformer 88.50±2.45 86.08±2.55 88.20±2.64 This invention 99.35±0.36 99.40±0.35 99.34±0.0.38

[0170] Table 5

[0171] method Accuracy (%) Accuracy (%) F1 score (%) SVM 46.13±6.65 52.31±6.62 45.37±5.77 KNN 51.99±6.86 53.07±5.33 51.66±6.33 RF 50.48±7.17 51.65±5.76 50.07±6.93 CNN 72.12±17.14 78.64±12.74 71.76±18.42 CNN-LSTM 80.71±1.55 82.02±1.86 80.60±1.87 Transformer 74.01±3.59 77.43±2.73 72.95±4.32 This invention 98.89±0.43 99.01±0.34 98.85±0.47

[0172] Table 6

[0173] As can be seen from Tables 5 and 6, the present invention has higher fault diagnosis accuracy and stronger generalization ability (bold indicates the highest accuracy under each fault diagnosis task). In particular, the average accuracy of the present invention on different fault diagnosis tasks is around 99%; compared with other fault diagnosis methods, it improves the diagnostic accuracy by at least 11% on fault diagnosis tasks with independent and identically distributed faults, and by at least 18% on fault diagnosis tasks with non-independent and identically distributed faults.

[0174] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A method for diagnosing open-circuit faults in power converters based on physical information transformer networks, characterized in that, Includes the following steps: S1: Collect several fault samples under different open-circuit faults for the power converter. Each fault sample includes data points of length [length missing]. The three-phase current signals and corresponding open-circuit fault tags. The value is determined based on actual needs; S2: Construct a power converter open-circuit fault diagnosis model based on a physically-informed transformer network, including a physical feature extraction module, a dynamic weight generation module, a curvature exceedance feature generation module, a transformer encoder, and a detection module, wherein: The physical feature extraction module is used to extract features from the input three-phase current signal. Extract the amplitude of each data point in each phase channel. and curvature ,in Indicates the channel index. , Indicates the index of the data point. Generate amplitude attention markers for each data point in each phase channel. and curvature attention mark ,in or This indicates that you are interested in the corresponding parameter. or This indicates that the corresponding parameters are not considered, and then the fused physical characteristics are obtained. For each data point, the maximum value is selected from the fused physical features of the three phase channels as the physical feature of the data point. And send it to the dynamic weight generation module, while simultaneously calculating the curvature of each data point in each phase channel. Send to the curvature exceedance feature generation module; The dynamic weight generation module is used to generate weights based on the physical characteristics of data points. Generate each data point Dynamic weights Then it is sent to the transformer encoder; The curvature exceedance feature generation module is used to generate features based on curvature. For the input three-phase current signal Generate curvature beyond features And sent to the transformer encoder, curvature exceeding the feature The generation method is as follows: First, calculate the curvature feature matrix. Each element The calculation formula is as follows: ， in, These represent the preset upper and lower limits of curvature, respectively. Represents the ReLU activation function; Then, the curvature excess feature is calculated using the following formula. : ， in, It represents the Hadamah accumulation. Represents a linear mapping function; Transformer encoders are used for dynamic weight-based encoders. and curvature beyond features For the input three-phase current signal Feature extraction is performed to obtain signal features. And send it to the detection module; the Transformer encoder includes an input embedding layer, a position encoding layer, an embedding fusion module, a normalization module, a difference module, a multi-branch adaptive attention module, an attention feature fusion module, and a feedforward neural network, wherein: The input embedding layer is used for the input three-phase current signal. Data embedding is performed, and the resulting embedded features are sent to the embedding fusion module; The position coding layer is used for the input three-phase current signal. Perform position encoding and send the obtained position encoding to the embedding fusion module; The embedding fusion module is used to fuse the position encoding with the embedding features, and sends the resulting fused features to the normalization module. The normalization module is used to normalize the fused features, and the resulting normalized features are used as signal features. The data is sent to the differential module and the multi-branch adaptive attention module, respectively. Differential modules are used to characterize signals. Perform first-order difference operations on the data points to obtain the difference features. And send it to the multi-branch adaptive attention module; The multi-branch adaptive attention module is used to adjust attention based on dynamic weights. Signal characteristics Difference characteristics and curvature beyond features Feature extraction was performed using a self-attention mechanism to obtain attention features. , , And sent to the attention feature fusion module; attention features The generation method is as follows: 1) Based on dynamic weights Calculate the dynamic mask range for each data point : ， in, , These represent the upper and lower limits of the preset dynamic mask, respectively. Indicates the preset coefficient; 2) Based on the upper limit of dynamic mask Define a neighborhood for each data point : ； 3) For input features Calculate the query matrix Key matrix Sum matrix : ， in, , , , These represent the learnable weight matrices of the corresponding matrices; 4) Based on neighborhood Calculated data points and its neighboring data points Attention scores between : ， in, The vector dimension of the key matrix is ​​represented by the superscript. Indicates transpose; 5) Calculate the data points based on the dynamic mask. and its neighboring data points Adaptive mask between : ， in, Indicates rounding down; 6) Calculate the data points using the following formula. and its neighboring data points Attention weights between : ； 7) Attention features are calculated using the following formula. : ； in, Representing attention features Mid-data points eigenvalues; The attention feature fusion module is used to perform dynamic weight fusion. Three attention features , , By performing fusion, fusion characteristics are obtained. The data is sent to the feedforward neural network, and the fusion formula is as follows: ， in, Represents a dynamic weight vector; Feedforward neural networks are used for fused features Perform nonlinear transformation to obtain signal characteristics ; The detection module is used to detect based on signal characteristics. Obtain the input three-phase current signal The probability of belonging to various open-circuit faults; S3: Use the fault samples obtained in step S1 to train the power converter open circuit fault diagnosis model to obtain a trained power converter open circuit fault diagnosis model. S4: When open-circuit fault diagnosis of the power converter is required, the number of data points to be collected is: The three-phase current signal is then input into the trained power converter open-circuit fault diagnosis model to obtain the open-circuit fault diagnosis result.

2. The power converter open-circuit fault diagnosis method according to claim 1, characterized in that, The open-circuit faults include single-tube faults and dual-tube faults.

3. The power converter open-circuit fault diagnosis method according to claim 1, characterized in that, The data points of each three-phase current signal need to cover at least one fault cycle.

4. The power converter open-circuit fault diagnosis method according to claim 1, characterized in that, The curvature The calculation method is as follows: First, calculate the difference in amplitude. and : ， ， Then, the curvature is calculated using the following formula. : ， in, This indicates mean filtering.

5. The power converter open-circuit fault diagnosis method according to claim 1, characterized in that, The method for calculating the attention marker is as follows: Calculate the mean and standard deviation of the corresponding parameter series: ， ， ， ， Then, the attention marker is determined using the following formula: ， ， in, , These represent the preset coefficients. This indicates an indicator function; its value is 1 when the content is true, and 0 otherwise.

6. The power converter open-circuit fault diagnosis method according to claim 1, characterized in that, The dynamic weight generation module includes convolutional layers, normalization layers, and nonlinear transformation layers, wherein: Convolutional layers are used to analyze the physical features of data points. Perform convolution to obtain convolutional features. And send it to the normalization layer; Normalization layers are used to normalize convolutional features Normalize to the range [0,1] and obtain the normalized features. Send to the nonlinear transformation layer; Nonlinear transformation layers are used for normalized features Dynamic weights are obtained by performing nonlinear transformations. The formula for calculating the nonlinear transformation is as follows: ， in, This indicates a preset threshold.

7. The power converter open-circuit fault diagnosis method according to claim 1, characterized in that, The loss function used during training of the power converter open-circuit fault diagnosis model is the label-smoothed cross-entropy loss function.

8. The power converter open-circuit fault diagnosis method according to claim 1, characterized in that, The power converter open-circuit fault diagnosis model is trained using the Adam optimizer to update the weight parameters, and the learning rate is calculated using the following formula: ， in, Indicates the current training round. This represents the current learning rate. This represents the minimum learning rate. This represents the maximum learning rate. This represents the total number of training rounds.

Images