Composite state decoupling identification method of high-voltage direct-current contactor

Abstract

The invention belongs to the technical field of composite state decoupling recognition, and provides a composite state decoupling recognition method of a high-voltage direct current contactor, which comprises the steps of working data acquisition and composite state decoupling recognition. The state recognition model training process comprises the steps of original collection data collection, additional frequency component calculation, CEEMDAN decomposition and filtering, comprehensive dynamic index calculation, sensitive component screening, SDP image calculation and OfficientNetV2-ViT model training. According to the method, through CEEMDAN decomposition and comprehensive dynamic index construction, an IMF containing an additional frequency is eliminated, signal purification is realized, and a high-value modal priority is dynamically enhanced; through the time interval and the angle amplification factor, it is ensured that the SDP image accurately presents the composite state features.

Description

Technical Field

[0001] This invention relates to the field of composite state decoupling identification technology, and in particular to a composite state decoupling identification method for high voltage DC contactors. Background Technology

[0002] With the rapid development of the new energy vehicle industry, high-voltage DC contactors, as core components of new energy vehicles and charging piles, are facing the dual challenges of surging demand and significantly increased reliability requirements. Therefore, intelligent identification of abnormal operating states of high-voltage DC contactors is of great significance for improving the reliability and safety of high-voltage systems in new energy vehicles and charging facilities. Vibration and sound signals can synchronously track real-time changes in the contactor's mechanical structure, comprehensively reflecting its operating status. Current research has used image features and deep learning methods to identify single operating states of AC contactors.

[0003] However, contactor degradation processes are diverse, and their abnormal modes may be coupled together. Yet, most current research on contactor state identification does not consider the case of composite states. Furthermore, existing methods often treat composite states as new modes independent of individual states, ignoring their potential relationship with individual states, making it difficult to accurately identify composite states from their essence. The challenges are twofold: first, how to extract important information related to composite states during signal preprocessing; and second, how to decouple and identify the coexistence of multiple abnormal states, accurately classifying composite states and their constituent individual states. Summary of the Invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of this invention is to provide a composite state decoupling identification method for high voltage DC contactors, which solves the problems of static nature, weak anti-interference, poor adaptability and black box nature, poor adaptability and low efficiency of existing contactor mode screening technology and existing SDP parameter optimization.

[0005] To achieve the above objectives, the present invention provides the following solution:

[0006] A method for decoupling and identifying the combined states of a high-voltage DC contactor includes:

[0007] Collect operating data of the target high-voltage DC contactor;

[0008] The working data is input into a pre-trained state recognition model for identification, resulting in a composite state decoupling recognition result; the training process of the state recognition model includes:

[0009] Vibration and acoustic signals of the contactor during operation under different operating conditions are collected to obtain raw data.

[0010] The characteristic frequencies of the signal during the operation of the severely worn contact are calculated based on the original acquired data to obtain additional frequency components; the severely worn contact is a contact with a roughness between 142μm and 183μm.

[0011] The original acquired data is decomposed using CEEMDAN to obtain the original IMF components. The original IMF components containing the additional frequency components are then removed to obtain the filtered IMF components.

[0012] Based on the filtered IMF components, the normalized correlation coefficient and normalized energy ratio are calculated and fused using variance weighting to obtain a comprehensive dynamic index.

[0013] The average value of all the comprehensive dynamic indicators is extracted to obtain the screening threshold. The screening threshold is then used to divide the filtered IMF components into sensitive components and screening components.

[0014] The SDP feature parameters of the sensitive components are calculated based on the time interval and angle magnification factor to obtain the SDP image. All the SDP images are then integrated to obtain the dataset.

[0015] The state recognition model is obtained by training the EfficientNetV2-ViT model using the dataset.

[0016] Preferably, the normalized correlation coefficient and normalized energy ratio are calculated and weighted by variance based on the filtered IMF components to obtain a comprehensive dynamic index, including:

[0017] The correlation degree of the filtered IMF components is calculated and normalized using the Pearson correlation coefficient method to obtain the normalized correlation coefficient;

[0018] The energy of each IMF in the filtered IMF component is calculated and normalized to obtain the normalized energy ratio.

[0019] The normalized correlation coefficient and the normalized energy ratio are weighted by variance to obtain the comprehensive dynamic index; the expression of the comprehensive dynamic index is: ;in, The aforementioned comprehensive dynamic index; and These are the variances of the normalized correlation coefficient and the normalized energy ratio, respectively. , These are the normalized correlation coefficient and the normalized energy ratio, respectively.

[0020] Preferably, the SDP feature parameters of the sensitive components are calculated according to the time interval and angle magnification factor to obtain SDP images. All SDP images are then integrated to obtain a dataset, including:

[0021] The time interval and the angle magnification factor of the sensitive components are calculated to obtain the determination parameters;

[0022] The sensitive components are transformed using the SDP transformation formula based on the determined parameters to obtain the SDP image;

[0023] The dataset is obtained by integrating the SDP images.

[0024] Preferably, the state recognition model is obtained by training the EfficientNetV2-ViT model using the dataset, including:

[0025] Construct the EfficientNetV2-ViT model; the EfficientNetV2-ViT model includes: an EfficientNetV2 network, a MobileViT network, and a network classifier connected in sequence; the network classifier adopts a multi-label classifier based on a classifier chain.

[0026] The dataset is input into the EfficientNetV2-ViT model for computation to obtain the network recognition results;

[0027] The network recognition result is calculated using binary cross-entropy loss to obtain the training loss, and the EfficientNetV2-ViT model is iteratively trained using the training loss to obtain the state recognition model.

[0028] The present invention discloses the following technical effects:

[0029] This invention provides a method for decoupling and identifying the composite state of a high-voltage DC contactor. By using CEEMDAN decomposition and comprehensive dynamic index construction, it solves the defects of existing contactor mode screening technology, such as static nature, weak anti-interference, and poor adaptability, and achieves signal purification. By using time interval and angle amplification factor, it solves the problems of black box, poor adaptability, and low efficiency in existing SDP parameter optimization, and achieves accurate extraction of composite state features. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram of the composite state decoupling identification process of a high-voltage DC contactor provided in an embodiment of the present invention;

[0032] Figure 2 The overall flowchart provided for embodiments of the present invention;

[0033] Figure 3 This is a flowchart of signal reconstruction provided in an embodiment of the present invention;

[0034] Figure 4 This is a schematic diagram of the classifier chain structure provided in an embodiment of the present invention;

[0035] Figure 5 The structure diagram of EfficientNetV2-ViT based on multi-label classification provided in the embodiments of the present invention;

[0036] Figure 6 The IMF screening results for high-voltage DC contactors provided in this embodiment of the invention. Figure 6 (a) shows the IMF screening results of the vibration signal. Figure 6 (b) shows the IMF filtering results for the audio signal;

[0037] Figure 7 These are SDP images of a high-voltage DC contactor under different states provided in an embodiment of the present invention. Figure 7 (a) is the SDP image under the condition of long adsorption time. Figure 7 (b) is the SDP image under the condition of short overtravel time. Figure 7 (c) is an SDP image showing a long release time. Figure 7 (d) is the SDP image of the normal closing state. Figure 7 (e) is the SDP image of the circuit breaker in normal tripping state. Figure 7 (f) is the SDP image of the composite state;

[0038] Figure 8 This is the composite state recognition result of the EfficientNetV2-ViT model provided in the embodiments of the present invention. Figure 8 (a) is the confusion matrix. Figure 8 (b) Visualization results of t-SNE classification;

[0039] Figure 9This is the composite state recognition result of the multi-label classification model provided in the embodiments of the present invention. Figure 9 (a) shows the multi-label classification results. Figure 9 (b) Activate the heatmap for the label. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] The purpose of this invention is to provide a decoupling identification method for composite states of high-voltage DC contactors, which solves the problems of static nature, weak anti-interference, poor adaptability, and black box nature, poor adaptability, and low efficiency of existing contactor mode screening technology, as well as the problems of existing SDP parameter optimization.

[0042] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0043] Figure 1 This is a schematic diagram of the composite state decoupling identification process of a high-voltage DC contactor provided in an embodiment of the present invention. Figure 2 The overall flowchart provided for the embodiments of the present invention is as follows: Figure 1 and Figure 2 As shown, the present invention provides a method for decoupling and identifying the composite state of a high-voltage DC contactor, comprising:

[0044] Step 100: Collect the operating data of the target high-voltage DC contactor;

[0045] Step 200: Input the working data into the pre-trained state recognition model for recognition, and obtain the composite state decoupling recognition result; the training process of the state recognition model includes:

[0046] Step 201: Collect vibration and acoustic signals of the contactor during operation under different operating conditions to obtain raw data;

[0047] Step 202: Calculate the characteristic frequencies of the signal during the operation of the severely worn contact based on the original acquired data to obtain additional frequency components; the severely worn contact is a contact with a roughness between 142μm and 183μm.

[0048] Step 203: Perform CEEMDAN decomposition on the original acquired data to obtain the original IMF components, and remove the original IMF components containing the additional frequency components to obtain the filtered IMF components;

[0049] Step 204: Calculate and perform variance-weighted fusion of the normalized correlation coefficient and normalized energy ratio based on the filtered IMF components to obtain a comprehensive dynamic index;

[0050] Step 205: Extract the average value of all the comprehensive dynamic indicators to obtain the screening threshold, and use the screening threshold to divide the filtered IMF components into sensitive components and screening components.

[0051] Step 206: Calculate the SDP feature parameters of the sensitive component according to the time interval and angle magnification factor to obtain the SDP image, and integrate all the SDP images to obtain the dataset;

[0052] Step 207: Train the EfficientNetV2-ViT model using the dataset to obtain the state recognition model.

[0053] Specifically, based on the filtered IMF components, the normalized correlation coefficient and normalized energy ratio are calculated and fused using variance weighting to obtain a comprehensive dynamic index, including:

[0054] The correlation degree of the filtered IMF components is calculated and normalized using the Pearson correlation coefficient method to obtain the normalized correlation coefficient;

[0055] The energy of each IMF in the filtered IMF component is calculated and normalized to obtain the normalized energy ratio.

[0056] The normalized correlation coefficient and the normalized energy ratio are weighted by variance to obtain the comprehensive dynamic index; the expression of the comprehensive dynamic index is: ;in, The aforementioned comprehensive dynamic index; and These are the variances of the normalized correlation coefficient and the normalized energy ratio, respectively. , These are the normalized correlation coefficient and the normalized energy ratio, respectively.

[0057] Preferably, the SDP feature parameters of the sensitive components are calculated according to the time interval and angle magnification factor to obtain SDP images. All SDP images are then integrated to obtain a dataset, including:

[0058] The time interval and the angle magnification factor of the sensitive components are calculated to obtain the determination parameters;

[0059] The sensitive components are transformed using the SDP transformation formula based on the determined parameters to obtain the SDP image;

[0060] The dataset is obtained by integrating the SDP images.

[0061] Furthermore, the EfficientNetV2-ViT model is trained using the dataset to obtain the state recognition model, including:

[0062] Construct the EfficientNetV2-ViT model; the EfficientNetV2-ViT model includes: an EfficientNetV2 network, a MobileViT network, and a network classifier connected in sequence; the network classifier adopts a multi-label classifier based on a classifier chain.

[0063] The dataset is input into the EfficientNetV2-ViT model for computation to obtain the network recognition results;

[0064] The network recognition result is calculated using binary cross-entropy loss to obtain the training loss, and the EfficientNetV2-ViT model is iteratively trained using the training loss to obtain the state recognition model.

[0065] Specifically, this embodiment provides a method for decoupling and identifying the composite state of a high-voltage DC contactor based on image features and deep learning (hereinafter referred to as the method), which includes the following steps:

[0066] Step 1: Collect vibration and acoustic signals of the contactor during operation under different operating conditions;

[0067] Step 2: Calculate the characteristic frequency of the signal during the collision of severely worn contacts and use this characteristic frequency as an additional frequency component; use a topography scanner to scan the topography of the contactor contacts, determine the roughness of the contactor contacts based on the root mean square deviation of the surface topography, and consider contacts with roughness in the range of 142μm to 183μm as severely worn contacts, and use the characteristic frequency of the signal of severely worn contacts as an additional frequency component.

[0068] Step 3: Adaptive Sensitive Mode Selection Based on Comprehensive Dynamic Indicators. The vibration and acoustic signals are decomposed using CEEMDAN, and the dominant frequencies of each IMF component are calculated. IMF components containing additional frequency components are removed. Based on this, a comprehensive dynamic index is constructed using the variance of the Pearson correlation coefficient and the energy ratio as weights, thereby adaptively selecting IMF components according to signal characteristics. The specific steps are as follows:

[0069] 1) Calculate the remaining frequency components after removing additional frequency components using the Pearson correlation coefficient method. The correlation coefficient is obtained by assessing the degree of correlation between each IMF and the original signal. , ,in Let M be the correlation coefficient between the Mth IMF and the original signal. For the index of the remaining IMF, The number of remaining IMFs, and for Normalization is performed:

[0070]

[0071] 2) Calculate the energy ratio of each IMF and normalize it. The calculation formula is as follows:

[0072]

[0073]

[0074]

[0075] In the formula, IMF m (t) is the th One eigenmode function For time, and The first Energy values ​​and energy ratios of individual IMFs For the energy set of all IMFs, For the first The normalized energy ratio of each IMF.

[0076] 3) Calculate the variance of the correlation coefficient and the energy ratio, and perform variance weighting to obtain the result. A composite dynamic indicator of the IMF The calculation formula is as follows:

[0077]

[0078]

[0079]

[0080] in, and Normalized correlation coefficients Ratio of normalized energy The mean, and These are the variances of the correlation coefficient and the energy ratio, respectively. The signal reconstruction process is as follows: Figure 3As shown.

[0081] After removing the IMF components containing additional frequency components, a comprehensive dynamic index is calculated for the remaining IMF components to obtain... ,in This serves as an index for the remaining IMFs. Subsequently, for these IMFs... Sort and The average value is set as a threshold, thus dividing the IMFs into sensitive components and filtered components. The signal is reconstructed using the sensitive components, which can effectively reflect the motion characteristics of the contactor while removing noise and other interference components.

[0082] Step 4: Calculate SDP feature parameters based on adaptive parameters of sensitive modal characteristics, generate SDP images using the effective modal components of vibration and acoustic signals, and then obtain multiple SDP images for different operating states to form a dataset. The specific steps are as follows:

[0083] 1) The SDP method converts time-series signals from Cartesian coordinates to polar coordinates, visually displaying signal changes through a two-dimensional mirror-symmetric image. The image's texture features can intuitively reflect the amplitude and frequency changes of the original time series. (Contactor vibration and effective modes of sound signals) Each point can be converted into a point in the polar coordinate system using SDP technology. ,in Let the instantaneous amplitude of the signal be the SDP transformation formula, which is as follows:

[0084]

[0085] In the formula, Number the current plane of symmetry. , The number of mirror-symmetric planes; The angle is the axis of symmetry of the mirror plane; The time domain signal corresponds to the time point amplitude, , Time domain signals The maximum and minimum amplitudes; for The radius in polar coordinates; and These are the angles by which the point deflects counterclockwise and clockwise along the mirror-symmetric plane in polar coordinates, respectively; For time interval parameters; It is the angle magnification factor, and .

[0086] 2) Calculate SDP parameters based on sensitive mode characteristics. When performing SDP transformation on the time-domain signal, the angle of the symmetry axis of the mirror-symmetric plane... Angle magnification factor and time interval parameters These factors collectively determine the morphological characteristics of SDP images. A quantitative correlation model between SDP parameters and IMF features is established to achieve physical interpretability of parameter calculations.

[0087] First, calculate the time interval l. The instantaneous frequency of the sensitive mode. The signal rate of change (high-frequency IMF changes rapidly, requiring a smaller value to capture details; low-frequency IMF changes slowly, requiring a larger value to avoid redundancy) and the energy percentage are determined by these factors. This reflects the correlation between the sensitive mode and the signal (high-energy IMFs are preferentially matched). Therefore, the formula for calculating the time interval l is:

[0088]

[0089] In the formula, The sampling frequency of the signal; Let m be the instantaneous frequency of the m-th sensitive mode; The energy percentage of the m-th sensitive mode; For composite state adaptation coefficients, the range for high-voltage DC contactors is 820 to 900. The number of sensitive modes.

[0090] Then, calculate the angle magnification factor. Angle magnification factor The angle magnification factor needs to match the IMF's ability to characterize the composite state and its amplitude variation characteristics. The calculation formula is as follows:

[0091]

[0092] In the formula, Based on the basic angle coefficient, the range of high-voltage DC contactors is from 10° to 40°; , is the comprehensive dynamic index of the m-th sensitive mode, reflecting the basic characterization ability of the sensitive mode for the composite state; This is the amplitude correction factor; The difference between the maximum and minimum amplitudes of the IMF reflects the amplitude coupling characteristics of the composite state.

[0093] Finally, multiple SDP images for each operating state are generated based on the optimized parameters to obtain the dataset.

[0094] Step 5: Construction of Multi-Label Classifier Chain and Decoupling of Composite States. Using EfficientNetV2-ViT as the feature extractor, leveraging its efficient feature extraction capabilities and the global modeling capabilities of the MobileViT module, local details and global dependencies in SDP images can be effectively captured. The classifier of the EfficientNetV2-ViT model is modified to construct a multi-label classifier based on a classifier chain, employing the Sigmoid activation function and calculating the classification loss using binary cross-entropy. The specific steps are as follows:

[0095] 1) Using the high-dimensional features output by EfficientNetV2-ViT as input, the construction of a multi-label classifier first involves multi-label transformation: That is, each sample correspond Tag ,in For the total number of tags, This is the tag index, with a value range of [value range missing]. Let s be a single sample and y be the set of multi-label samples. For the four single states of a DC contactor, the... Sample The label can be represented as a four-dimensional binary vector. Secondly, a classification chain is used to perform multi-label classification on the state features extracted by the feature extractor. The classifier chain consists of four binary classifiers, each predicting one label. In the chain structure, each classifier receives not only the original input features but also... It also uses the prediction result of the previous classifier as additional input (Sigmoid activation) to dynamically learn label dependencies, such as... Figure 4 As shown. The composite state recognition model based on multi-label classification is as follows. Figure 5 As shown.

[0096] 2) Train the state recognition model using the dataset, and then use the trained model to identify the operating state of the contactor. To optimize the training of the multi-label classification model, the binary cross-entropy loss (BCELoss) is typically used to calculate the classification loss. The expression for the binary cross-entropy loss is:

[0097]

[0098] In the formula, Indicates the total number of samples; The sample index and the range of values ​​is ; Indicates the first The first sample The actual value of each label (1 indicates existence, 0 indicates non-existence); The model predicts the first... The sample has the first The probability of each label is output by the Sigmoid function.

[0099] Specifically, in this embodiment, the method for decoupling and identifying the composite state of a high-voltage DC contactor includes the following steps:

[0100] Step 1: Four high-voltage DC contactors that had completed electrical life tests and two brand-new test samples were used as research objects. Each underwent 1000 no-load tests. Vibration and acoustic signals during the closing process of the contactors under different operating conditions were collected using an online monitoring platform. Based on the simulation results of contact collision sound generation, the vibration and sound pressure amplitude of the contact system were highest in the collision direction. Therefore, a KS78B10 accelerometer was installed at the bottom. To prevent the vibration sensor from being too close to the sound sensor and affecting the stability of the collected signals, an MPA201 sound sensor was placed on the side of the contactor. The signals were displayed and analyzed using a DH5922D dynamic test analyzer, with the sampling frequency set to 100kHz. Based on the parameter values ​​of this type of contactor provided by the manufacturer and combined with the calculation results, appropriate thresholds were selected to classify the contactor's operating state into the following five types: long overtravel time, short overtravel time, long release time, normal state (normal closing, normal opening), and composite state. Multiple labels were used to represent the sample state, and 0 and 1 in one-hot encoding indicated the presence of the state. The experimental dataset is shown in Table 1.

[0101] Table 1

[0102]

[0103] Step 2: Obtain the characteristic frequency of the signal during the collision of severely worn contacts based on spectrum analysis, and use this characteristic frequency as an additional frequency component.

[0104] Step 3: Adaptive screening of sensitive modes based on comprehensive dynamic indicators. Taking data of state A as an example, the mode screening results are as follows: Figure 6 As shown. Through spectral analysis, the IMF4 of the sound signal was removed due to the presence of additional components. Subsequently, the comprehensive dynamic index of the remaining IMFs was calculated, and the modal screening threshold for the vibration signal was set to 0.0734, and the threshold for the sound signal was set to 0.0750. Based on the threshold screening, the following was obtained: Figure 6 (a) Select IMF1, IMF2, and IMF3 of the vibration signal as sensitive components. Similarly, we obtain... Figure 6 (b) Select IMF1, IMF2, IMF3 and IMF5 of the audio signal as sensitive components. Then reconstruct the signal from the sensitive components.

[0105] Step 4: Calculate SDP feature parameters based on adaptive parameters of sensitive modal features. After multiple experimental measurements, set the adaptation coefficient. Given 820, β = 32, and k = 2, calculate the angle magnification factor of the SDP parameters. and time interval parameters To avoid excessive dispersion of feature information and reduced image interpretability, vibration and sound signals are segmented and mapped to two mirror-symmetric planes in the SDP transform, resulting in an SDP image containing four mirror-symmetric planes. Therefore, the symmetry axis angle is set. The angle is 90°. SDP images are generated using the effective modal components of vibration and acoustic signals, resulting in multiple SDP images for different operating states, forming a dataset. The SDP images corresponding to different states serve as references. Figure 7 (a) to Figure 7 (f).

[0106] Step 5: The EfficientNetV2-ViT model with a classifier chain was used for multi-label classification performance evaluation. The experimental dataset was divided into training, validation, and test sets in a 4:1:1 ratio. The model was trained using PyTorch 1.12 and Python 3.7.0, employing the AdamW optimizer with an initial learning rate of 0.0001 and a learning rate decay factor of 0.5. The batch size was set to 32, and the number of training epochs was 200. Binary cross-entropy loss was used as the loss function, and an early stopping mechanism was introduced to prevent overfitting; training stopped when the validation set loss did not decrease for five consecutive epochs. During model testing, subset accuracy, macro F1-Score, Hamming loss, and ranking loss were used to evaluate the model's overall accuracy, balance performance, label-level performance, and ranking ability.

[0107] To verify the effectiveness of the proposed SDP image generation technique, which uses modal adaptive screening based on comprehensive dynamic indices to select vibration and sound signals and calculates SDP parameters based on sensitive modal features, in the contactor composite state recognition task, SDP image samples generated by the following methods were used as input to the EfficientNetV2-ViT state recognition model based on the classifier chain, and the model's classification performance was evaluated. Method 1: SDP image generated by fusing original vibration and sound signals and optimizing parameters using an improved jumping spider algorithm. Method 2: SDP image generated by unweighted fused vibration and sound signals and optimizing parameters using an improved jumping spider algorithm. Method 3: SDP image generated by modal adaptive screening based on comprehensive dynamic indices and optimizing parameters using an improved jumping spider algorithm. Method 4: SDP image generated by the proposed method. Table 2 shows the state recognition results using SDP images generated by different methods as input. The table shows that the recognition accuracy of Methods 3 and 4 is higher than that of Methods 1 and 2, indicating that the proposed modal adaptive screening method based on comprehensive dynamic indices can effectively extract sensitive modes. The recognition accuracy of methods 3 and 4 is basically the same, but the time consumption of method 4 is much shorter than that of method 3, indicating that the calculation of SDP parameters based on sensitive modal features can improve the computational efficiency while maintaining recognition accuracy.

[0108] Table 2

[0109] Sample time / ms Accuracy / % F1-score / % Method 1 52.14 90.67 91.37 Method 2 52.07 92.50 94.25 Method 3 52.31 95.23 97.07 Method 4 3.56 95.21 97.16

[0110] To verify the proposed method's ability to decouple and identify complex anomalous states, the EfficientNetV2-ViT multi-class classification model based on a Softmax classifier was used as a comparison. The multi-class classification model treats complex states as new categories E independent of individual states during training to achieve complex state identification. To ensure fairness in the comparative experiments, the Softmax classifier-based model and the multi-label classification model constructed in this embodiment maintain the same structure, differing only in the design of the classification layer. The state identification results of the two methods are as follows: Figure 6 As shown.

[0111] Depend on Figure 8 (a) It can be seen that the EfficientNetV2-ViT model based on the Softmax classifier achieves an overall accuracy of 95.07% in the state recognition task, but it has limitations in recognizing the composite abnormal state E, mainly manifested in confusion with single states A and B. Figure 8 (b) As can be seen, further analysis through t-SNE feature visualization reveals that the clustering of single states is clear, while the feature clusters of the composite anomalous state E lie between states A and B, with relatively small spacing between the feature clusters. In contrast, the subset accuracy of the multi-label classification method reaches 95.21%, demonstrating strong composite state recognition capabilities. Figure 9 (a) shows that this method demonstrates good judgment for both single-label and multi-label applications. From Figure 9 (b) As can be seen, the label activation heatmap further verifies the model's ability to decouple composite states, clearly showing the activation status of sub-state labels in composite states, indicating that the model can accurately capture the inherent combination relationship of composite states. In addition, the multi-label classification method has strong flexibility and scalability in composite state recognition tasks, and can adapt to changes in various abnormal state combinations.

[0112] Furthermore, comparisons with other multi-label methods: To further verify the effectiveness of the proposed method in multi-label classification tasks, it was compared with other multi-label classification methods on the same experimental dataset. The comparison methods included binary association and label power sets in the problem transformation methods, and multi-label k-nearest neighbors (ML-KNN) and multi-label deep learning (MLDL) in the algorithm adaptation methods. ML-KNN finds the k nearest neighbors of each test sample in the training set and predicts the label of the test sample based on the label distribution of these nearest neighbors. MLDL utilizes a deep learning model to solve the multi-label classification problem; in this section's comparison experiments, MLDL specifically refers to the EfficientNetV2-ViT model based on the Sigmoid classifier. To avoid the randomness of the experimental results, 10 experiments were conducted on the test set, and the average value was taken as the final recognition result. The comparison results of the evaluation metrics of different methods are shown in Table 3.

[0113] Table 3

[0114] method Subset accuracy / % Macro F1-Score / % Hamming loss / % Sorting loss / % Binary association 88.80 93.03 3.56 2.47 Tag power set 90.60 93.83 2.93 1.91 ML-KNN 88.72 93.42 3.61 2.56 MLDL 93.07 96.18 2.07 1.29 Method of this embodiment 95.21 97.16 1.80 0.71

[0115] In problem transformation methods, compared to label power sets and classifier chains, binary association methods decompose multi-label problems into multiple independent binary classification problems, neglecting the dependencies between labels and resulting in lower subset accuracy. The classifier chain method used in this embodiment can dynamically capture label combination relationships and avoid the data sparsity problem of label power sets, significantly improving model performance. Furthermore, the EfficientNetV2-ViT model (MLDL method) based on the Sigmoid classifier can capture complex features of the data and output probability values ​​independently for each label, but neither it nor the ML-KNN method can explicitly model label dependencies, and its performance is still lower than the method in this embodiment. In summary, this embodiment achieves the decoupling of single abnormal state identification and complex state decoupling by leveraging the efficient feature extraction capabilities of the EfficientNetV2-ViT model and the dynamic capture advantage of label relationships in the classifier chain.

[0116] The beneficial effects of this invention are as follows:

[0117] This invention achieves signal purification by eliminating IMFs containing additional frequencies through CEEMDAN decomposition and comprehensive dynamic index construction, and dynamically enhances the priority of high-value modes; through time interval and angle magnification factors, it ensures that the SDP image accurately presents the composite state characteristics.

[0118] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0119] This embodiment uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for decoupling and identifying the composite state of a high-voltage DC contactor, characterized in that, The method comprises the following steps: Collecting working data of a target high-voltage direct-current contactor; Inputting the working data into a pre-trained state recognition model for recognition to obtain a composite state decoupling recognition result; The training process of the state recognition model comprises the following steps: Collecting vibration and sound signals of the contactor in the action process under different operating states to obtain original collected data; According to the original collected data, the characteristic frequency of the signal in the working process of a severely worn contact is calculated to obtain an additional frequency component; the severely worn contact is a contact with a roughness of 142 μm to 183 μm; The original IMF component is obtained by performing CEEMDAN decomposition on the original collected data, and the original IMF component containing the additional frequency component is removed to obtain a filtered IMF component; According to the filtered IMF component, the normalized correlation coefficient, the normalized energy ratio, and the variance weighted fusion are calculated to obtain a comprehensive dynamic index; The average value of all the comprehensive dynamic indexes is extracted to obtain a screening threshold, and the filtered IMF component is divided into a sensitive component and a screened component by using the screening threshold; According to the time interval and the angle amplification factor, the SDP feature parameters of the sensitive component are calculated to obtain an SDP image, and all the SDP images are integrated to obtain a data set; The EfficientNetV2-ViT model is trained by using the data set to obtain the state recognition model.

2. The method of claim 1, wherein the method is a composite state decoupling identification method of a high-voltage direct-current contactor. According to the filtered IMF component, the normalized correlation coefficient, the normalized energy ratio, and the variance weighted fusion are calculated to obtain a comprehensive dynamic index, which comprises the following steps: The correlation degree of the filtered IMF component is calculated and normalized by using the Pearson correlation coefficient method to obtain the normalized correlation coefficient; The energy of each IMF in the filtered IMF component is calculated and normalized to obtain the normalized energy ratio; The normalized correlation coefficient and the normalized energy ratio are subjected to variance-weighted calculation to obtain the comprehensive dynamic index; an expression of the comprehensive dynamic index is: ; wherein, is the comprehensive dynamic index; and are variances of the normalized correlation coefficient and the normalized energy ratio, respectively; , are the normalized correlation coefficient and the normalized energy ratio, respectively.

3. The method of claim 1, wherein the method is a composite state decoupling identification method of a high-voltage direct-current contactor. According to the time interval and the angle amplification factor, the SDP feature parameters of the sensitive component are calculated to obtain an SDP image, and all the SDP images are integrated to obtain a data set, which comprises the following steps: The time interval and the angle amplification factor of the sensitive component are calculated to obtain a decision parameter; According to the decision parameter, the sensitive component is converted by using an SDP transformation formula to obtain the SDP image; The SDP images are integrated to obtain the data set.

4. The method of claim 1, wherein the method is a composite state decoupling identification method of a high-voltage direct-current contactor. The EfficientNetV2-ViT model is trained by using the data set to obtain the state recognition model, which comprises the following steps: The EfficientNetV2-ViT model is constructed; the EfficientNetV2-ViT model comprises an EfficientNetV2 network, a MobileViT network, and a network classifier connected in sequence; the network classifier adopts a multi-label classifier based on a classifier chain; The data set is input into the EfficientNetV2-ViT model for calculation to obtain a network recognition result; The network recognition result is calculated by using a binary cross-entropy loss, a training loss is obtained, and the EfficientNetV2-ViT model is iteratively trained by using the training loss, so as to obtain the state recognition model.

Images