An online diagnostic method and system for open-circuit faults in CETDC submodules' IGBTs

Abstract

This invention relates to the field of fault diagnosis technology for high-voltage, high-capacity DC-DC converters, and provides an online diagnostic method and system for open-circuit faults in IGBTs of CETDC submodules. The method includes: constructing a fusion representation of multi-channel capacitor voltages; establishing a VMD adaptive decomposition optimized by the INFO algorithm; constructing a composite fitness function and constraining the optimization process; establishing a sensitive mode screening based on envelope entropy statistical characteristics; constructing multi-domain feature vectors of sensitive modes and establishing a fault identification model; and establishing interpretability analysis and key feature compression. This invention maintains high-precision diagnosis under complex operating conditions and noise disturbances, and improves the real-time performance, stability, and engineering usability of online deployment through a closed-loop process of "parameter adaptive decomposition - sensitive mode screening - key feature compression - attention time-series identification - interpretability verification".

Description

Technical Field

[0001] This invention belongs to the field of fault diagnosis technology for high-voltage, high-capacity DC-DC converters, and particularly relates to an online diagnostic method and system for open-circuit faults of IGBTs in CETDC submodules. Background Technology

[0002] With the development of DC power transmission and DC power grids, high-voltage, high-capacity DC / DC converters serve as crucial interfaces between DC lines of different voltage levels, and their safe and stable operation is of great significance to system reliability. Capacitive energy transfer type DC / DC converters (CETDC) offer advantages in terms of low cost and high efficiency due to the elimination of AC component injection, reduced number of submodules, and the absence of filter inductors. However, as the number of output levels increases, the number of submodules and the scale of power devices also increase, making IGBTs a high-frequency fault point. Furthermore, short-circuit isolation often transforms into an open-circuit fault, leading to abnormal capacitor voltage, disruption of bridge arm voltage balance, and system distortion. Therefore, rapid and accurate open-circuit fault diagnosis and location are necessary.

[0003] Traditional IGBT open-circuit fault diagnosis methods mainly fall into two categories: hardware-based and model-based. Hardware-based diagnostic methods typically rely on adding sensors or modifying measurement points for rapid location. However, in engineered CETDC devices, voltage sensors are often fixedly positioned across the energy transfer capacitor for control and protection. It is difficult to adjust measurement points for diagnostic purposes in already operational or standardized equipment. Furthermore, introducing additional port-side measurement links may enhance the coupling between measurement and control and increase sensitivity to noise, thus introducing potential operational instability risks. Model-based methods, while not requiring additional sensors, are highly dependent on the accuracy of mathematical models and parameters. The computational and implementation complexity increases significantly with the size of submodules, and their robustness is limited under various operating conditions and noise disturbances.

[0004] In contrast, data-driven deep learning methods can automatically learn fault discrimination features from sampled data, making them more suitable for handling the nonlinear and non-stationary characteristics of CETDC capacitor voltage signals, and possessing better generalization and anti-interference capabilities. However, in engineering applications, they still face problems such as insufficient interpretability and redundant computation and excessive online load due to the high-dimensional input from multiple sub-modules. Therefore, there is an urgent need for an online diagnostic scheme for CETDC sub-module IGBT open-circuit faults that can be implemented using only the sub-module capacitor voltage without changing the existing sensor layout, while taking into account the separation quality of non-stationary impact features, robustness, real-time performance, and interpretability, to meet the requirements of rapid online diagnosis and engineering deployment of CETDC sub-module IGBT open-circuit faults. Summary of the Invention

[0005] The purpose of this invention is to provide an online diagnostic method and system for IGBT open-circuit faults in CETDC submodules, aiming to solve the problems mentioned in the background art.

[0006] This invention is implemented as follows: an online diagnostic method for IGBT open-circuit faults in CETDC submodules includes the following steps:

[0007] Step 1: Collect the capacitor voltage of each sub-module of CETDC, construct a fusion characterization of multi-channel capacitor voltage, and obtain the fusion single-channel voltage characterization.

[0008] Step 2: For the fused single-channel voltage characterization, establish a variational mode decomposition model, and use the INFO algorithm to calculate the number of modes in the variational mode decomposition. and penalty factor Perform joint iterative optimization to obtain the optimal parameter combination. One band-limited modal component;

[0009] Step 3: Construct a composite fitness function as the optimization objective function of the INFO algorithm. The composite fitness function integrates modal envelope entropy, intermodal orthogonality index, and modal number penalty term.

[0010] Step 4: Perform variational mode decomposition based on the optimal parameter combination to obtain... Each modal component has an envelope entropy. The envelope entropy of each modal component is calculated. An adaptive threshold is constructed based on the mean and standard deviation of the envelope entropy of all modal components. Modes with envelope entropy less than the adaptive threshold are selected as fault-sensitive modes.

[0011] Step 5: Construct multi-domain feature vectors for the selected fault-sensitive mode sequences, and input the multi-domain feature vectors into a bidirectional long short-term memory network model that incorporates channel attention mechanism and spatial attention mechanism for fault identification and classification;

[0012] Step 6: Attribution explanation of the decisions of the trained bidirectional long short-term memory network model based on SHAP value theory, obtain the contribution of each feature, and select the key feature set according to the contribution ranking for subsequent online diagnosis input feature compression.

[0013] Another objective of this invention is to provide an online diagnostic system for open-circuit faults in CETDC submodules' IGBTs, based on the aforementioned method, comprising:

[0014] The data acquisition module is used to acquire the capacitor voltage signals of each sub-module of CETDC in real time.

[0015] The feature fusion module is used to normalize and weight the multi-channel capacitor voltage signals using the entropy weight method, and output the fused voltage characterization signal.

[0016] The parameter optimization and decomposition module is used to optimize VMD parameters using the INFO algorithm and adaptively decompose the fused signal to output multiple intrinsic mode components.

[0017] The sensitive modality filtering module is used to construct an adaptive threshold based on the statistical characteristics of envelope entropy and to filter out a set of sensitive modalities.

[0018] The feature extraction module is used to perform joint time-domain and frequency-domain feature extraction on the sensitive modality set and construct multi-domain feature vectors;

[0019] The fault identification module incorporates a BiLSTM model with a CBAM attention mechanism to output fault diagnosis results based on the input multi-domain feature vectors.

[0020] The interpretability and feature compression module is used to interpret model decisions based on SHAP value theory and output Top-K key feature combinations for online diagnosis.

[0021] The diagnostic output module is used to display or transmit the final fault type, location of the fault submodule, and contribution information of key features.

[0022] The present invention provides an online diagnostic method and system for IGBT open-circuit faults in CETDC submodules, the advantages of which are as follows:

[0023] (1) Hardware-friendly and highly adaptable to engineering: This invention uses only the existing energy transfer capacitor voltage of each sub-module of CETDC as the only measurement quantity. Fault diagnosis is achieved through signal processing and deep learning technology. No new sensors or changes to the existing measurement point layout are required. It has excellent adaptability to already put into operation and finalized equipment.

[0024] (2) Adaptive and robust feature extraction: By optimizing the VMD decomposition parameters through the INFO algorithm, adaptive extraction of non-stationary fault impacts is achieved, overcoming the shortcomings of traditional fixed parameter decomposition methods that are easily affected by operating conditions and noise. Combined with sensitive mode screening based on envelope entropy statistical characteristics, noise and redundant components are effectively eliminated, improving the signal-to-noise ratio of features from the source.

[0025] (3) High diagnostic accuracy and strong localization capability: The constructed time-frequency domain joint multi-domain feature vector comprehensively describes the fault information. The BiLSTM model with the CBAM attention mechanism can adaptively strengthen key features and capture the bidirectional temporal dependence of fault evolution, which significantly improves the classification and localization accuracy of multi-submodule and multi-device faults and effectively alleviates the confusion problem between faults of adjacent submodules.

[0026] (4) The model is interpretable and easy to accept for engineering projects: The introduction of SHAP interpretability analysis quantifies the contribution of each feature to the diagnostic results, enhances the trust of maintenance personnel in the diagnostic conclusions, and is conducive to the engineering acceptance and promotion of the technology.

[0027] (5) High online real-time performance and low overhead: The Top-K key features are selected through SHAP analysis (e.g., the Top-4 are preferred), which significantly compresses the online input features. This not only greatly reduces the overhead of online feature calculation and data transmission, but also shifts the bottleneck of the diagnostic system from feature extraction to model inference, resulting in a significant reduction in end-to-end diagnostic latency and a significant increase in throughput, thus meeting the needs of online real-time diagnosis. Attached Figure Description

[0028] Figure 1 This is a flowchart of the signal preprocessing process based on INFO-VMD;

[0029] Figure 2 A general technical roadmap for an online diagnostic method for IGBT open-circuit faults in CETDC submodules provided by embodiments of the present invention;

[0030] Figure 3 A ranking graph of feature importance;

[0031] Figure 4 This is the result of the SHAP interpretability analysis;

[0032] Figure 5 This is a comparison chart of model accuracy under Top-K feature combinations;

[0033] Figure 6 A comparison chart of model loss values ​​under Top-K feature combinations;

[0034] Figure 7 This is the confusion matrix for the Top-4 feature combination model under System 1;

[0035] Figure 8 This is the confusion matrix of the Top-4 feature combination model under System 2. Detailed Implementation

[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0037] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0038] like Figure 1 and Figure 2As shown, an embodiment of the present invention provides an online diagnostic method for IGBT open-circuit faults in CETDC submodules, comprising the following steps:

[0039] Step 1: Construct a fusion characterization of multi-channel capacitor voltage;

[0040] The capacitor voltages of different submodules are normalized and their information content is evaluated. Weights are assigned based on the channel information content, and weighted fusion reconstruction is performed to enhance fault distortion characterization and suppress redundancy and noise. The circuit module capacitor voltage sequence forms a matrix. ,in Indicates the first The first channel in the The capacitance voltage values ​​at each sampling point are obtained; the range normalization of the voltage of each submodule is performed to obtain... The weights of each channel are calculated based on the entropy weight method. Finally, the fused single-channel voltage characterization was obtained. The calculation process is as follows:

[0041] (1)

[0042] In the formula, and They represent the first The maximum and minimum values ​​of each channel in the entire sample. The voltage of each submodule is after the range is normalized.

[0043] (2)

[0044] (3)

[0045] In the formula, For the first Information entropy of each channel; This represents the number of sampling points; For the first The first channel The proportion of each normalized sample.

[0046] (4)

[0047] In the formula, For the first Each channel weight, For the number of channels, For the first Information entropy of each channel; The difference coefficient is used to characterize the effective information content of the channel.

[0048] (5)

[0049] In the formula, For the first Each channel's original capacitor voltage signal; It is a time variable.

[0050] Step 2: Establish an adaptive VMD decomposition optimized by the INFO algorithm;

[0051] A VMD decomposition model is established to address the non-stationary characteristics of the fused voltage signal. The INFO algorithm is then used to iteratively optimize key VMD parameters to improve the separation quality of fault impact components and suppress mode mixing. The calculation process is as follows:

[0052] Characterization of the fused single-channel voltage Establish a VMD variational model to obtain band-limited modal components and its center frequency :

[0053] (6)

[0054] In the formula, Represents convolution; It is the Dirac function; For time differential operators; For the first One modal component; For the first The center frequency of each modal component.

[0055] Within the ADMM framework, the first The frequency domain iterative update of each mode is as follows:

[0056] (7)

[0057] In the formula, As a penalty factor; For Lagrange multipliers; This represents the number of iterations. For the first The modal component in the th ... Frequency domain representation at the next iteration; for Frequency domain representation; Frequency domain representation of Lagrange multipliers; To exclude the first Frequency domain characterization of modal components other than the first modal component; ω is the angular frequency.

[0058] The center frequency has been updated to:

[0059] (8)

[0060] In the formula, For the first The mode in the th ... The center frequency of the next iteration.

[0061] To avoid VMD key parameters and Relying on manual selection, this preferred solution uses the INFO algorithm to select the optimal solution. Joint iterative optimization is performed, and the objective function adopts the following composite fitness function (Equation (11)), and its minimization is used as the optimization criterion.

[0062] Step 3: Construct a composite fitness function and constrain the optimization process;

[0063] The impact significance is characterized by modal envelope entropy, modal independence is constrained by orthogonality index, and a modal number penalty term is introduced to avoid excessive decomposition, thus forming an objective function for INFO optimization. The calculation process is as follows:

[0064] For each mode obtained by VMD Hilbert demodulation was performed to obtain the envelope amplitude sequence, which was then normalized to obtain the probability distribution. The envelope entropy is defined as:

[0065] (9)

[0066] In the formula, For the first Envelope entropy of each mode; For the first The normalized probability of each sampling point; For the first The envelope signal of the i-th modal component after Hilbert transform is at the i-th modal component. The amplitude of each sampling point.

[0067] To constrain the independence between modes, an orthogonality index is introduced. :

[0068] (10)

[0069] In the formula, For covariance; For modality Standard deviation; The modal number; For modality Standard deviation; The smaller the value, the lower the correlation between modes and the stronger the independence.

[0070] By combining envelope entropy, orthogonality, and mode number penalty, a composite fitness function is constructed:

[0071] (11)

[0072] In the formula, To optimize the objective; and These are the weighting coefficients. The INFO algorithm updates candidate solutions iteratively, making... Gradually decrease the parameters until the optimal combination is output. .

[0073] Step 4: Establish a sensitive mode screening method based on the statistical properties of envelope entropy;

[0074] The optimal VMD decomposition yields Calculate the envelope entropy sequence for each modality. An adaptive threshold is constructed using its mean and standard deviation:

[0075] (12)

[0076] In the formula, The mean of the envelope entropy for all modes; Standard deviation; This is the threshold adjustment coefficient; This is an adaptive threshold.

[0077] Define a sensitive indicator function:

[0078] (13)

[0079] In the formula, For sensitive indicator functions, Indicates the first Each mode is a fault-sensitive mode; This mode is identified as invalid / noise mode and discarded; ultimately, it is retained. The sensitive modes consist of a set of sensitive modes. .

[0080] Step 5: Construct multi-domain feature vectors for sensitive modes and establish a fault identification model;

[0081] Multi-domain feature vectors are constructed for sensitive modal sequences and then input into the CBAM-BiLSTM model for classification and recognition. The channel attention module is as follows:

[0082] (14)

[0083] In the formula, Input features; and These are global average pooling and max pooling, respectively. It is a multilayer perceptron; For the Sigmoid function; This represents the channel attention weight.

[0084] The spatial attention module is:

[0085] (15)

[0086] In the formula, This is the intermediate feature vector obtained after weighting by the channel attention mechanism; its mathematical definition is: ; Indicates element-wise multiplication; This indicates splicing by channel; express convolution; Spatial attention weights; ultimately obtained , is the final enhanced feature vector obtained after weighting by the spatial attention mechanism, and its mathematical definition is: .

[0087] The gating update for the BiLSTM cell is as follows:

[0088] (16)

[0089] In the formula, For a moment Input features; and They are time points The hidden state and the memory unit state; Candidate memory cell state; for The state of the memory unit at any given moment; for The hidden state at any given moment; It is the hyperbolic tangent activation function; , and These are the forget gate, input gate, and output gate, respectively. These are the weight matrices corresponding to the forget gate, input gate, candidate memory unit, and output gate, respectively. These are the bias terms corresponding to the forget gate, input gate, candidate memory unit, and output gate, respectively. For Hadama accumulation.

[0090] BiLSTM bidirectional output splicing is as follows:

[0091] (17)

[0092] In the formula, For forward LSTM output, For output to the backward LSTM; This is a splicing technique for bidirectional output.

[0093] The classification output uses Softmax:

[0094] (18)

[0095] In the formula, For the fully connected layer to the first The class's logit output, For the fully connected layer to the first The class's logit output; Number of categories; For the first The sample belongs to the first The predicted probability of a class.

[0096] The training loss cross-entropy is:

[0097] (19)

[0098] In the formula, One-hot encoding of the actual label; Batch size; The training loss is cross-entropy.

[0099] Step 6: Establish interpretability analysis and key feature compression;

[0100] Based on SHAP, feature contribution is attributed and explained. A Top-K set of key features is determined according to global contribution ranking. The Top-4 key features are then selected for online input compression to reduce online computation and transmission overhead and improve diagnostic interpretability. The calculation process is as follows:

[0101] The SHAP interpretation framework is introduced into the trained CBAM-BiLSTM model to construct an additive interpretation model:

[0102] (20)

[0103] In the formula, Output the original model; To explain the model; Indicates the first Does the feature exist? For the first The contribution of each feature; Baseline output; This represents the total dimension of the input features.

[0104] The Shapley value is calculated as follows:

[0105] (twenty one)

[0106] In the formula, It is the complete feature set; For features not included A subset of features; Indicates the current sample Using only feature subsets The model output when the features are in the middle; Indicating in the feature subset Add the first The increment in model output caused by the first feature, i.e., the first feature... Features relative to subsets The marginal contribution.

[0107] An embodiment of the present invention provides an online diagnostic system for IGBT open-circuit faults in CETDC submodules, based on the above method, comprising:

[0108] The data acquisition module is used to acquire the capacitor voltage signals of each sub-module of CETDC in real time.

[0109] The feature fusion module is used to normalize and weight the multi-channel capacitor voltage signals using the entropy weight method, and output the fused voltage characterization signal.

[0110] The parameter optimization and decomposition module is used to optimize VMD parameters using the INFO algorithm and adaptively decompose the fused signal to output multiple intrinsic mode components.

[0111] The sensitive modality filtering module is used to construct an adaptive threshold based on the statistical characteristics of envelope entropy and to filter out a set of sensitive modalities.

[0112] The feature extraction module is used to perform joint time-domain and frequency-domain feature extraction on the sensitive modality set and construct multi-domain feature vectors;

[0113] The fault identification module incorporates a BiLSTM model with a CBAM attention mechanism to output fault diagnosis results based on the input multi-domain feature vectors.

[0114] The interpretability and feature compression module is used to interpret model decisions based on SHAP value theory and output Top-K key feature combinations for online diagnosis.

[0115] The diagnostic output module is used to display or transmit the final fault type, location of the fault submodule, and contribution information of key features.

[0116] As a preferred embodiment of the present invention, the global importance is calculated and sorted based on the contribution of all test samples to select the Top-K key feature set. The Top-4 key feature combination is preferred as the online input feature set to reduce the overhead of online feature calculation and transmission, and output the key feature contribution explanation information corresponding to the diagnostic conclusion.

[0117] like Figure 3 and Figure 4 As shown, interpretability analysis is performed on the trained model to obtain the feature importance ranking and contribution distribution, which are used to explain the model's discrimination criteria and identify key features that play a dominant role in the diagnostic results. Based on this, refer to... Figure 5 and Figure 6 The accuracy and loss curves of the model under different Top-K feature combinations were compared to determine the key feature set that minimizes the input dimension while ensuring diagnostic accuracy. Furthermore, combined with... Figure 7 and Figure 8 The confusion matrix results can verify the effectiveness of the key feature compression strategy in reducing the confusion of fault categories between adjacent sub-modules and improving the reliability of positioning.

[0118] The preferred Top-4 key features in this embodiment are: median, spectral centroid, power mean, and skewness, which are used as the online input feature set to significantly reduce the burden of online feature calculation and data transmission while maintaining high-precision diagnostic performance.

[0119] like Figure 3 As shown, the feature importance ranking results are presented, illustrating that different features contribute significantly differently to fault detection; (Refer to...) Figure 4 The results of SHAP interpretability analysis are presented, which can be used to explain the positive and negative contributions of key features to the output from both global and sample levels; (Refer to...) Figure 5 and Figure 6 The paper presents comparison charts of model accuracy and loss values ​​under Top-K feature combinations, showing that model performance gradually improves with the increase of the number of key features, but the gains tend to saturate after reaching a certain K value, thus providing a basis for feature compression; (Refer to...) Figure 7 and Figure 8 The confusion matrices of the Top-4 feature combination model for fault diagnosis in two systems are presented respectively, which can intuitively demonstrate the identification and localization effect of the present invention in different systems / different fault categories.

[0120] To verify the applicability, diagnostic effectiveness, and online real-time processing capability of this method under different system parameter conditions, two CETDC simulation systems were constructed and comparative experiments and real-time performance evaluations were conducted.

[0121] Two simulation systems, System 1 and System 2, were used to form a multi-condition dataset. The key parameter settings for the two systems are shown in Table 1. The voltage across the energy transfer capacitor of each submodule was collected under different operating conditions as the sole measurement. Following the procedures outlined in the previous embodiment, fusion characterization, INFO-VMD adaptive decomposition, and sensitive mode screening were completed. Furthermore, the time-domain and frequency-domain joint features of the sensitive modes were extracted to form feature vectors, which were then input into the fault identification model to obtain the fault category and location results. Simultaneously, based on the feature contribution analysis results, the Top-4 key features (median, spectral centroid, power mean, and skewness) were determined for online input compression, and combined with... Figures 3 to 8 The diagnostic effectiveness is validated by the given feature importance ranking, interpretability results, Top-K performance curves, and confusion matrix.

[0122] Table 1 Simulation parameter settings for System 1 and System 2

[0123]

[0124] To verify the effectiveness of this method compared with existing methods, this embodiment reproduces the comparison schemes such as WPD-PCA and SSAE, and sets LSTM and BiLSTM with full feature input as deep temporal benchmark models, and compares them with the preferred scheme CBAM-BiLSTM (Top-4 input). The comprehensive evaluation indexes obtained are shown in Table 2.

[0125] Table 2 Comparison of Diagnostic Performance of Different Models

[0126]

[0127] Table 2 shows that traditional schemes such as WPD-PCA and SSAE have low overall performance, mainly due to their insufficient adaptability to strong non-stationary fault impacts and multi-condition drift. Furthermore, linear dimensionality reduction or untargeted representation learning easily weakens the weak difference information between adjacent sub-modules, leading to increased localization confusion. While LSTM and BiLSTM can improve recognition capabilities using temporal information under full feature input conditions, they are still susceptible to non-critical features and noise disturbances due to the lack of an adaptive recalibration mechanism for the effectiveness and redundancy of input features. In contrast, the optimized scheme still achieves the best accuracy and global evaluation index even with only the Top-4 key features as input. This indicates that fusion representation, INFO-VMD decomposition, and sensitive modality screening can improve the feature signal-to-noise ratio from the source, while the attention mechanism can further enhance the contribution of key features, enabling the model to have clearer classification boundaries for faults in adjacent sub-modules under multiple conditions.

[0128] Furthermore, to verify the effectiveness of the key feature compression strategy in improving the real-time performance of online deployment, end-to-end real-time diagnostic evaluations were conducted on the Top-4 key feature scheme and the full feature input scheme under the same test platform and timing range. The test adopted a single-sample point-by-point input method (batch=1), evaluating a total of 1000 samples; the timing range covered both online feature computation and model inference, and the mean, standard deviation, and P99 quantile of the total latency were statistically analyzed. The end-to-end throughput (samples / s) was used to characterize the online processing capability, and the results are shown in Table 3.

[0129] Table 3. Comparison of end-to-end real-time diagnostic performance under different input feature dimensions

[0130]

[0131] In the table, "±" represents the standard deviation, which is used to characterize the discrete fluctuations of multiple measurement results; P99 is the 99th percentile, which is used to measure tail delay.

[0132] As shown in Table 3, compared to the full-feature input scheme, the Top-4 feature model scheme significantly reduces the online feature extraction time, decreasing the total end-to-end latency from 10.746ms to 2.191ms and increasing the throughput from 93.06 samples / s to 456.35 samples / s. This indicates that the key feature compression strategy of this method can effectively reduce the overhead of online feature computation and data transmission while ensuring diagnostic accuracy, thereby improving online real-time processing capabilities and response determinism. Furthermore, the time consumption percentage shows that the main bottleneck of the full-feature scheme is concentrated in the feature extraction stage, while the Top-4 scheme shifts the system bottleneck to the model inference stage, further illustrating the direct engineering value of compressed input dimensions for online deployment. Figure 7 and Figure 8 The confusion matrix results can further verify that, under different system parameter conditions, the identification model using Top-4 key feature input has fewer misclassifications of each fault category, which can reduce confusion between adjacent sub-modules, thereby improving the reliability of fault location and engineering availability.

[0133] The above description is merely 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 within the protection scope of the present invention.

Claims

1. An online diagnostic method for open-circuit faults in CETDC submodules' IGBTs, characterized in that, Includes the following steps: Step 1: Collect the capacitor voltage of each sub-module of CETDC, construct a fusion characterization of multi-channel capacitor voltage, and obtain the fusion single-channel voltage characterization. Step 2: For the fused single-channel voltage characterization, establish a variational mode decomposition model, and use the INFO algorithm to count the number of modes in the variational mode decomposition. and penalty factor Perform joint iterative optimization to obtain the optimal parameter combination. One band-limited modal component; Step 3: Construct a composite fitness function as the optimization objective function of the INFO algorithm. The composite fitness function integrates modal envelope entropy, intermodal orthogonality index, and modal number penalty term. Step 4: Perform variational mode decomposition based on the optimal parameter combination to obtain... Each modal component has an envelope entropy. The envelope entropy of each modal component is calculated. An adaptive threshold is constructed based on the mean and standard deviation of the envelope entropy of all modal components. Modes with envelope entropy less than the adaptive threshold are selected as fault-sensitive modes. Step 5: Construct multi-domain feature vectors for the selected fault-sensitive mode sequences, and input the multi-domain feature vectors into a bidirectional long short-term memory network model that incorporates channel attention mechanism and spatial attention mechanism for fault identification and classification; Step 6: Attribution explanation of the decisions of the trained bidirectional long short-term memory network model based on SHAP value theory, obtain the contribution of each feature, and select the key feature set according to the contribution ranking for subsequent online diagnosis input feature compression.

2. The online diagnostic method for IGBT open-circuit faults in CETDC submodules according to claim 1, characterized in that, Step 1 includes the following steps: Normalize and evaluate the information content of capacitor voltages in different sub-modules, assign weights based on channel information content, and perform weighted fusion reconstruction; collect data. The circuit module capacitor voltage sequence forms a matrix. ,in Indicates the first The first channel in the The capacitance voltage values ​​at each sampling point are obtained; the range normalization of the voltage of each submodule is performed to obtain... The weights of each channel are calculated based on the entropy weight method. Finally, the fused single-channel voltage characterization was obtained. The calculation process is as follows: （1） In the formula, and They represent the first The maximum and minimum values ​​of each channel in the entire sample. The voltage of each submodule after range normalization; （2） （3） In the formula, For the first Information entropy of each channel; This represents the number of sampling points; For the first The first channel The proportion of each normalized sample; （4） In the formula, For the first Each channel weight, For the number of channels, For the first Information entropy of each channel; The coefficient of variation; （5） In the formula, For the first Each channel's original capacitor voltage signal; It is a time variable.

3. The online diagnostic method for IGBT open-circuit faults in CETDC submodules according to claim 2, characterized in that, Step 2 includes the following specific steps: Characterization of the fused single-channel voltage Establish a VMD variational model to obtain band-limited modal components and its center frequency : （6） In the formula, Represents convolution; It is the Dirac function; For time differential operators; For the first One modal component; For the first The center frequency of each modal component; Within the ADMM framework, the first The frequency domain iterative update of each mode is as follows: （7） In the formula, As a penalty factor; For Lagrange multipliers; This represents the number of iterations. For the first The modal component in the ... Frequency domain representation at the next iteration; for Frequency domain representation; Frequency domain representation of Lagrange multipliers; To exclude the first Frequency domain characterization of modal components other than the first modal component; Angular frequency; The center frequency has been updated to: （8） In the formula, For the first The mode in the th ... The center frequency of the next iteration.

4. The online diagnostic method for IGBT open-circuit faults in CETDC submodules according to claim 3, characterized in that, Step 3 includes the following specific steps: For each mode obtained by VMD Hilbert demodulation was performed to obtain the envelope amplitude sequence, which was then normalized to obtain the probability distribution. The envelope entropy is defined as: （9） In the formula, For the first Envelope entropy of each mode; For the first The normalized probability of each sampling point; For the first The envelope signal of the i-th modal component after Hilbert transform is at the i-th modal component. The amplitude of each sampling point; To constrain the independence between modes, an orthogonality index is introduced. : （10） In the formula, For covariance; For modality Standard deviation; The modal number; For modality Standard deviation; The smaller the value, the lower the correlation between modes and the stronger the independence. By combining envelope entropy, orthogonality, and mode number penalty, a composite fitness function is constructed: （11） In the formula, To optimize the objective; and These are the weighting coefficients; the INFO algorithm updates candidate solutions iteratively, making... Gradually decrease the parameters until the optimal combination is output. .

5. The online diagnostic method for IGBT open-circuit faults in CETDC submodules according to claim 4, characterized in that, Step 4 includes the following specific steps: The optimal VMD decomposition yields Calculate the envelope entropy sequence for each modality. An adaptive threshold is constructed using its mean and standard deviation: （12） In the formula, The mean of the envelope entropy for all modes; Standard deviation; This is the threshold adjustment coefficient; An adaptive threshold; Define a sensitive indicator function: （13） In the formula, For sensitive indicator functions, Indicates the first Each mode is a fault-sensitive mode; This mode is identified as invalid / noise mode and discarded; ultimately, it is retained. The sensitive modes consist of a set of sensitive modes. .

6. The online diagnostic method for IGBT open-circuit faults in CETDC submodules according to claim 5, characterized in that, Step 5 includes the following specific steps: Multi-domain feature vectors are constructed for the sensitive modality sequences and then input into the CBAM-BiLSTM model for classification and recognition; the channel attention module is as follows: （14） In the formula, Input features; and These are global average pooling and max pooling, respectively. It is a multilayer perceptron; For the Sigmoid function; Channel attention weights; The spatial attention module is: （15） In the formula, The intermediate feature vector is obtained after weighting by the channel attention mechanism; its mathematical definition is: ; Indicates element-wise multiplication; This indicates splicing by channel; express convolution; Spatial attention weights; ultimately obtained , is the final enhanced feature vector obtained after weighting by the spatial attention mechanism, and its mathematical definition is: ; The gating update for the BiLSTM cell is as follows: （16） In the formula, For a moment Input features; and They are time points The hidden state and the memory unit state; Candidate memory cell state; for The state of the memory unit at any given moment; for The hidden state at any given moment; It is the hyperbolic tangent activation function; , and These are the forget gate, input gate, and output gate, respectively. These are the weight matrices corresponding to the forget gate, input gate, candidate memory unit, and output gate, respectively. These are the bias terms corresponding to the forget gate, input gate, candidate memory unit, and output gate, respectively. For Hadamah accumulation; BiLSTM bidirectional output splicing is as follows: （17） In the formula, For forward LSTM output, For output to the backward LSTM; This is a splicing technique for bidirectional output. The classification output uses Softmax: （18） In the formula, For the fully connected layer to the first The class's logit output; For the fully connected layer to the first The class's logit output; Number of categories; For the first The sample belongs to the first The predicted probability of a class; The training loss cross-entropy is: （19） In the formula, One-hot encoding of the actual label; Batch size; The training loss is cross-entropy.

7. The online diagnostic method for IGBT open-circuit faults in CETDC submodules according to claim 6, characterized in that, Step 6 includes the following specific steps: The SHAP interpretation framework is introduced into the trained CBAM-BiLSTM model to construct an additive interpretation model: （20） In the formula, Output the original model; To explain the model; Indicates the first Does the feature exist? For the first Shapley values ​​for each feature; Baseline output; The total dimension of the input features; The Shapley value is calculated as follows: （21） In the formula, It is the complete feature set; For features not included A subset of features; Indicates the current sample Using only feature subsets The model output when the features are in the middle; Indicating in the feature subset Add the first The increment in model output caused by the first feature, i.e., the first feature... Features relative to subsets The marginal contribution.

8. An online diagnostic system for open-circuit faults in CETDC submodules IGBTs, based on the online diagnostic method for open-circuit faults in CETDC submodules according to any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire the capacitor voltage signals of each sub-module of CETDC in real time. The feature fusion module is used to normalize and weight the multi-channel capacitor voltage signals using the entropy weight method, and output the fused voltage characterization signal. The parameter optimization and decomposition module is used to optimize VMD parameters using the INFO algorithm and adaptively decompose the fused signal to output multiple intrinsic mode components. The sensitive modality filtering module is used to construct an adaptive threshold based on the statistical characteristics of envelope entropy and to filter out a set of sensitive modalities. The feature extraction module is used to perform joint time-domain and frequency-domain feature extraction on the sensitive modality set and construct multi-domain feature vectors; The fault identification module incorporates a BiLSTM model with a CBAM attention mechanism to output fault diagnosis results based on the input multi-domain feature vectors. The interpretability and feature compression module is used to interpret model decisions based on SHAP value theory and output Top-K key feature combinations for online diagnosis. The diagnostic output module is used to display or transmit the final fault type, location of the fault submodule, and contribution information of key features.

Images