Contactor running state online identification method based on image features and deep learning

Abstract

The application discloses a kind of contactor operating state online identification method based on image feature and deep learning, first, the vibration and acoustic signal of contactor in the action process under different operating states are collected;Then, the characteristic frequency of signal in the collision process of seriously worn contact is calculated, and the characteristic frequency is used as additional frequency component;The vibration and acoustic signal are decomposed by CEEMDAN, to obtain multiple IMF components;The main frequency of each IMF component is calculated, and the IMF component containing additional frequency component is eliminated;The correlation coefficient and energy value between each remaining IMF component and original signal are calculated;The IMF component with correlation coefficient and energy value greater than average is retained, to obtain the effective modal component of vibration and acoustic signal;Then, the time interval parameter and angle amplification factor of SDP conversion technology are optimized using intelligent optimization algorithm, according to the optimized parameters, the effective modal component of vibration signal and acoustic signal is used to generate SDP image, to obtain multiple SDP images of different operating states;Finally, a state recognition model is constructed and trained, and the trained model is used for state recognition. This method avoids the influence of interference component on effective mode, improves the feature expression effect of SDP image, and improves the recognition accuracy.

Description

Technical Field

[0001] This invention belongs to the field of AC contactor operating status recognition technology, specifically a contactor operating status online recognition method based on image features and deep learning. Background Technology

[0002] AC contactors, as common low-voltage switching devices, are widely used in power distribution systems. With the increasing demand for contactors across various sectors, higher requirements are being placed on their reliability. Contactor failures will affect the stability of the power distribution system; therefore, online monitoring of their status is of great significance for improving the reliability and safety of power distribution and control systems.

[0003] The selection of measurement signals is a prerequisite for achieving state recognition. The vibration signals generated by the contactor during the opening and closing process can reflect the dynamic changes of the mechanical structure, while the acoustic signals contain comprehensive electrical and mechanical information. Both of these signals can characterize the operating state of the contactor. Moreover, the analysis method based on vibration and acoustic signals has the advantage of non-invasive detection, requiring no modification to the electrical circuit or structure, thus avoiding the impact on the stability of the electrical system. For example, Liu Shuxin et al. (Liu Shuxin, Song Jian, Liu Yang, et al. Research on motion analysis and fault diagnosis of AC contactor contact system [J]. Journal of Electrical Engineering, 2021, 36(S2): 477-486.) proposed a motion analysis method for AC contactor contact system based on vibration signals, which effectively improved the operating state recognition rate. You Yingmin et al. (You Yingmin, Wang Jingqin, Shu Liang, et al. A method for predicting the electrical life of AC contactors based on audio features [J]. Journal of Electrical Engineering, 2021, 36(09): 1986-1998.) used Mel-frequency cepstral coefficients (MFCC) to extract acoustic signal features and analyzed the differences in acoustic signals under different contactor states. Li Haiying et al. (Li Haiying, Sun Yue, Zhang Xiao, et al. A method for fault diagnosis of vacuum contactors based on the fusion of modal time-frequency diagrams and ResNet50 [J]. High Voltage Engineering, 2023, 49(05): 1831-1840.) proposed a method for fault diagnosis of vacuum contactors based on wavelet time-frequency diagrams and residual networks. However, the influence of contact electrical wear on vibration and acoustic signal characteristics is ignored in the analysis of the operating status of AC contactors, and all of them are based on a single signal perspective, ignoring the fusion of multi-sensor information. Compared to joint analysis of multiple signals, single-signal analysis is more susceptible to interference from environmental factors. Secondly, while image features reflect operational status, time-frequency analysis heavily relies on expert knowledge, and time-frequency graphs are not readily representative of status information. Furthermore, due to the high dimensionality and complex semantic information of image data, most models enhance feature representation by increasing network width and depth; however, this is prone to gradient vanishing, increases training time and computational resources, and places high demands on computer hardware. Therefore, considering the weak intensity and short duration of vibration and acoustic signals generated during contactor movement, and to further reduce interference and obtain more accurate and comprehensive operational status information, further in-depth research is needed on methods for identifying the operational status of AC contactors using appropriate feature fusion methods and deep learning models. Summary of the Invention

[0004] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is to provide an online identification method for contactor operating status based on image features and deep learning.

[0005] The technical solution adopted by the present invention to solve the aforementioned technical problem is as follows:

[0006] An online method for identifying the operating status of a contactor based on image features and deep learning includes the following steps:

[0007] Step 1: Collect vibration and acoustic signals of the contactor during operation under different operating conditions;

[0008] 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;

[0009] The vibration and sound signals are decomposed using CEEMDAN to obtain multiple IMF components. The main frequencies of each IMF component are calculated, and IMF components containing additional frequency components are removed. The correlation coefficient and energy value between each remaining IMF component and the original signal are calculated. IMF components with correlation coefficients and energy values ​​greater than the average value are retained to obtain the effective modal components of the vibration and sound signals.

[0010] Step 3: Optimize the time interval parameters and angle magnification factor of SDP conversion technology using intelligent optimization algorithms. Based on the optimized time interval parameters and angle magnification factor, generate SDP images using the effective modal components of vibration and acoustic signals, thereby obtaining multiple SDP images under different operating conditions and forming a dataset.

[0011] The expression for converting effective modal components into an SDP image is:

[0012] (1)

[0013] (2)

[0014] (3)

[0015] (4)

[0016] In the formula, The rotation angle is the angle of the plane of mirror symmetry. The number of mirror-symmetric planes. The index of the mirror-symmetric plane. For point The radius in polar coordinates, For point time domain coordinates, , These represent the maximum and minimum amplitudes of the effective modal components; and Points The angles of counterclockwise and clockwise deflection along the mirror-symmetric plane in polar coordinates. For time interval parameters, For point coordinates It is the angle magnification factor, and ;

[0017] Step 4: Construct a state recognition model based on the EfficientNetV2 network. This model consists of 5 layers. The SDP image is processed through convolution, batch normalization, and activation operations before being passed through each layer. The first layer includes two Fused-MBConv modules, the second layer includes four Fused-MBConv modules, the third layer includes one MBConv module and two MobileViT modules, the fourth layer includes one MBConv module and four MobileViT modules, and the fifth layer includes one MBConv module and three MobileViT modules. The output features of the fifth layer are passed through a fully connected layer to obtain the recognition result. The Mish activation function and ECA attention block replace the Swish activation function and SE block of the original MBConv module to obtain the MBConv modules of the fourth and fifth layers.

[0018] Step 5: Train the state recognition model using the dataset, and use the trained state recognition model to identify the operating state of the contactor.

[0019] Furthermore, when the intelligent optimization algorithm employs the spider jumping algorithm, the spider position is initialized according to the following formula:

[0020] (5)

[0021] (6)

[0022] In the formula, Indicates the first The initial position of the jumping spider. Indicates the index of the parameter being optimized. Indicates the number of parameters being optimized. , Indicates the first The upper and lower bounds of the parameter to be optimized. Indicates the first The initial values ​​of the optimized parameters, This indicates a random number generation operation;

[0023] The difference value between SDP images of different operating states corresponding to each jumping spider is quantitatively calculated according to formula (7). The difference values ​​between different SDP images are accumulated to obtain the fitness of the jumping spider.

[0024] (7)

[0025] In the formula, For the first SDP image corresponding to jumping spider and The difference between them , For SDP images and pixel values, and These are SDP images. and The average of all pixel values , This represents the number of pixels in the length and width directions of the SDP image;

[0026] The jumping spider's position is updated based on fitness, using the following formula:

[0027] (8)

[0028] (9)

[0029] (10)

[0030] In the formula, , It is the first Only the position of the jumping spider before and after the update. It is the first The stride length of a jumping spider. It is the location of the neighbor with the highest fitness within the current jumping spider's perception range. It is the first The fitness of jumping spiders, It is a coefficient that controls the movement step size. , These are the maximum and minimum values ​​of the coefficient. , These are the maximum and minimum fitness values.

[0031] Furthermore, the expansion ratios of the Fused-MBConv modules in the first and second layers of the state recognition model are 1 and 4, respectively, while the expansion ratios of the MBConv modules in the third to fifth layers are all 4.

[0032] Furthermore, the roughness of the severely worn contact is 142-183 μm.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] (1) In the contactor operation status identification based on motion characteristics, the contact electrical wear state affects the vibration and sound signal characteristics, resulting in the problem that the vibration and sound signal characteristics are not obvious. Through frequency characteristic analysis, it was found that the signal generated by the worn contact during collision contact contains additional frequency components. Therefore, in the process of modal component screening, the modal components containing additional frequency components are removed to avoid the influence of interference components, and then the effective modal components are extracted to achieve information decoupling between electrical wear state and mechanical state at the data layer.

[0035] (2) The IMF components of vibration and sound signals are converted into images, and state information from different frequency bands is fused to intuitively display the subtle changes in vibration and sound signals under different operating conditions. This highlights the differences in information from different operating conditions and overcomes the limitations of traditional time-frequency analysis methods that rely on expert knowledge and are not easy to intuitively represent state information. Based on the improved jumping spider algorithm, the parameters of the SDP image are optimized. The jumping spider search strategy is dynamically adjusted through an adaptive step size mechanism, which enhances the global search capability and local search accuracy of the algorithm and further improves the feature expression effect of the SDP image. In addition, compared with the time-frequency image, the SDP image has stronger robustness and provides more reliable feature input for state recognition.

[0036] (3) Construct a state recognition model combining EfficientNetV2 and MobileViT. To improve the local feature extraction efficiency and multi-level feature extraction capability of MobileViT, EfficientNetV2 is introduced. By fully utilizing the local feature extraction capability of EfficientNetV2 and the global modeling capability of Transformer, the model can not only perform better in resource-constrained environments, but also improve the generalization capability of the model. First, in the shallow structure of the EfficientNetV2-ViT model, the expansion ratio is gradually increased to achieve fast extraction of low-level features with low computational cost, and the number of feature map channels is increased to enrich feature expression, thereby improving the model's running efficiency. Second, the MBConv module is introduced in Layer 3, Layer 4 and Layer 5 of MobileViT to improve the feature extraction accuracy and effectiveness of the original model. In addition, the Mish activation function and ECA attention block are introduced in Layer 4 and Layer 5 to replace the original activation function and SE block, which can enhance the nonlinear expression capability of the model, avoid gradient vanishing, and improve the accuracy and computational efficiency of state recognition. Attached Figure Description

[0037] Figure 1 This is an overall flowchart of the present invention;

[0038] Figure 2 A flowchart for generating an SDP image;

[0039] Figure 3 This is a schematic diagram of the state recognition model.

[0040] Figure 4 A schematic diagram of the Fused-MBConv module when the scaling factor is equal to 1;

[0041] Figure 5 A schematic diagram of the Fused-MBConv module when the scaling factor is not equal to 1;

[0042] Figure 6 This is a schematic diagram of the MBConv module in the third layer;

[0043] Figure 7 This is a structural diagram of the MBConv modules in the fourth and fifth layers;

[0044] Figure 8 This is a structural diagram of the MobileViT module;

[0045] Figure 9 The spectrum diagrams show the sound pressure simulation results for smooth and rough contacts;

[0046] Figure 10 The spectrum of the acoustic signal during the collision process of smooth contacts;

[0047] Figure 11 The spectrum of the acoustic signal during the collision process of the rough contact;

[0048] Figure 12 The diagram shows the CEEMDAN decomposition results of the acoustic signal.

[0049] Figure 13 The image shows the CEEMDAN decomposition results of the vibration signal.

[0050] Figure 14 This is an analysis diagram of acoustic signal IMFs;

[0051] Figure 15 This is an analysis diagram of vibration signal IMFs;

[0052] Figure 16 A graph showing the change in fitness values ​​during the SDP image parameter optimization process;

[0053] Figure 17 SDP image of the contactor under long overtravel time condition;

[0054] Figure 18 SDP image of the contactor in a long-engagement state;

[0055] Figure 19 This is an SDP image of the contactor under conditions of high contact bounce frequency;

[0056] Figure 20 This is an SDP image of the contactor when the iron core is closing at a high speed.

[0057] Figure 21 This is an SDP image of the contactor when the contact bounce time is long.

[0058] Figure 22 This is an SDP image of the contactor in its normal operating state.

[0059] Figure 23 The accuracy and loss curves for the state recognition model training process;

[0060] Figure 24 This is a comparison chart of recognition accuracy for different signal-to-noise ratio values. Detailed Implementation

[0061] The technical solution of the present invention will be further described below with reference to the accompanying drawings and this embodiment, but it is not intended to limit the scope of protection of the present invention.

[0062] This invention provides an online identification method for contactor operating status based on image features and deep learning (hereinafter referred to as the method, see [link]). Figures 1-24 ), including the following steps:

[0063] Step 1: Collect vibration and sound signals of the contactor during operation under different operating conditions using an online monitoring platform;

[0064] Step 2: After performing spectral analysis on the vibration and acoustic signals during the contactor contact collision process, it was found that the roughness of the contact affects the frequency characteristics of the signal compared to a brand new contactor. Contactors with different roughness all generate new characteristic frequencies. Therefore, a topography scanner was used to scan the topography of the contactor contacts. The roughness of the contactor contacts was determined based on the root mean square deviation of the surface topography. The larger the root mean square deviation of the surface topography, the greater the roughness of the contact. Contacts with roughness in the range of 142-183μm were considered as severely worn contacts, and the characteristic frequency of the signal from severely worn contacts was used as an additional frequency component.

[0065] Vibration and acoustic signals are decomposed using Ensemble Empirical Mode Decomposition (CEEMDAN) to obtain multiple IMF components and a residual component. The IMF components of the vibration and acoustic signals are then filtered based on three indicators: characteristic frequency, correlation, and energy value. Taking the vibration signal as an example, spectral analysis is performed on each IMF component to calculate its principal frequency. If the principal frequency contains additional frequency components, the corresponding IMF component is discarded. Then, the correlation coefficient between the remaining IMF components and the original vibration signal is calculated using the Pearson correlation coefficient method. The energy value of the remaining IMF components is also calculated. IMF components with both correlation coefficients and energy values ​​greater than the average are retained, thus obtaining the effective modal components of the vibration signal. Similarly, the effective modal components of the acoustic signal are obtained. This decoupling of contact wear and operating status information at the data layer is achieved, which is beneficial for improving the accuracy of identifying abnormal operating states of the contactor.

[0066] Step 3: Optimize the parameters of SDP conversion technology using intelligent optimization algorithms. Based on the optimized parameters, generate SDP images using the effective modal components of vibration and acoustic signals, thereby obtaining multiple SDP images under different operating conditions and forming a dataset.

[0067] 3-1. Using Symmetric Point Mode Conversion (SDP) technology, an SDP image is generated based on the effective modal components of the vibration and acoustic signals. The one-dimensional effective modal components are converted into polar coordinates, achieving data fusion from different data sources. The SDP image contains multiple mirror-symmetric planes, with the number of mirror-symmetric planes equal to the number of effective modal components. The expression for converting the effective modal components into an SDP image is:

[0068] (1)

[0069] (2)

[0070] (3)

[0071] (4)

[0072] In the formula, The rotation angle is the angle of the plane of mirror symmetry. The number of mirror-symmetric planes. The index of the mirror-symmetric plane. For point The radius in polar coordinates, For point time domain coordinates, , These represent the maximum and minimum amplitudes of the effective modal components; and Points The angles of counterclockwise and clockwise deflection along the mirror-symmetric plane in polar coordinates. For time interval parameters, For point coordinates It is the angle magnification factor, and .

[0073] 3-2. Intelligent optimization algorithms include genetic algorithms, spider jumping algorithms, particle swarm optimization, ant colony optimization, and gray wolf optimization algorithms. Taking the spider jumping algorithm as an example, this paper improves the spider jumping algorithm and uses the improved algorithm to optimize the parameters of SDP images, specifically the time interval parameter. and angle magnification factor ;

[0074] First, determine the number of jumping spiders, the maximum number of iterations, and the perception range factor. Initialize the range of the optimized parameters, the position of the jumping spiders, the fitness, and the global optimum position, and define the adaptive step size range. Each jumping spider represents a set of optimized parameters (including the time interval parameters and angle magnification factors for all mirror-symmetric planes of the SDP image). The position initialization formula is as follows:

[0075] (5)

[0076] (6)

[0077] In the formula, Indicates the first The initial position of the jumping spider. Indicates the index of the parameter being optimized. Indicates the number of parameters being optimized. , Indicates the first The upper and lower bounds of the parameter to be optimized. Indicates the first The initial values ​​of the optimized parameters, This indicates a random number generation operation;

[0078] Next, based on the image correlation analysis method, the difference value between SDP images of different operating states corresponding to each jumping spider is quantitatively calculated according to Equation (7). The difference values ​​between different SDP images are accumulated to obtain the fitness of the jumping spider.

[0079] (7)

[0080] In the formula, For the first SDP image corresponding to jumping spider and The difference between them , For SDP images and pixel values, and These are SDP images. and The average of all pixel values , This represents the number of pixels in the length and width directions of the SDP image;

[0081] The greater the fitness, the greater the difference between the two SDP images. If the fitness of the jumping spider is greater than the current global best fitness, then the global best fitness and global best position are updated using the jumping spider.

[0082] Then, the jumping spider's position is updated based on its fitness. The distance between each jumping spider and other jumping spiders is calculated. Based on the perception range, the neighbor with the highest fitness is found. This local information is used to guide the jumping spider's search direction, and the position is updated based on the movement step size, so that the jumping spider gradually moves closer to the location of the best neighbor, thereby searching for the optimal solution in the global space. The jumping spider position update formula is as follows:

[0083] (8)

[0084] (9)

[0085] (10)

[0086] In the formula, , It is the first Only the position of the jumping spider before and after the update. It is the first The stride length of a jumping spider. It is the location of the neighbor with the highest fitness within the current jumping spider's perception range; It is the first The fitness of jumping spiders, It is a coefficient that controls the movement step size and is dynamically adjusted according to the jumping spider's fitness. , These are the maximum and minimum values ​​of the coefficient. , These are the maximum and minimum fitness values;

[0087] Finally, after the iteration is completed, the global optimal position is obtained, which is the set of optimized parameters with the greatest difference in SDP images of different running states; based on the optimized parameters, multiple SDP images of each running state are generated to obtain the dataset.

[0088] Step 4: Construct the state recognition model EfficientNet-ViT based on the EfficientNetV2 network; for example... Figure 3 As shown, the EfficientNet-ViT model mainly consists of 5 layers. The SDP image undergoes convolution (Conv2D), batch normalization (BN), and activation (Siwh) operations before passing through each layer sequentially. Layer 1 includes two Fused-MBConv modules with a scaling factor of 1; Layer 2 includes four Fused-MBConv modules with a scaling factor of 4; Layer 3 includes one MBConv module and two MobileViT modules; Layer 4 includes one MBConv module and four MobileViT modules; and Layer 5 includes one MBConv module and three MobileViT modules. The output features of Layer 5 are passed through a fully connected layer to obtain the output of the EfficientNetV2-ViT model. The fully connected layer flattens the input feature vector into a one-dimensional vector and uses a multi-classifier based on the Softmax function to perform image classification. The structure of the Fused-MBConv modules in Layer 1 and Layer 2 is shown in [reference needed]. Figure 4 and Figure 5 The scaling factor for each MBConv module is 4. The structure of the MBConv module in Layer 3 is as follows: Figure 6 As shown; the structure of the MBConv modules in Layer 4 and Layer 5 is as follows Figure 7 As shown, the Swish activation function and SE block of the original MBConv module are replaced with the Mish activation function and ECA attention block; the structure of the MobileViT module is as follows. Figure 8 As shown.

[0089] In the shallow structure (layers 1 and 2) of the EfficientNetV2-ViT model, gradually increasing the expansion ratio enables rapid extraction of low-level features with low computational cost, improving model efficiency. Secondly, the local fine features extracted by the MBConv module in Layers 3 to 5 serve as input to the MobileViT module. The Transformer module within MobileViT utilizes the local feature location information extracted by MBConv, combined with its own self-attention mechanism, to more accurately establish the connections between different regions in the SDP image, thereby improving the overall quality of feature representation. The combination of MBConv and MobileViT modules enables collaborative extraction of local and global features, with the extracted features complementing each other at different levels, enriching the semantic information of the features. Introducing the Mish activation function and ECA attention blocks in Layer 4 and Layer 5 can improve the model's ability to capture complex patterns and its computational efficiency. Specifically, different attention mechanisms are used at different layers, enabling the model to process and learn data from multiple perspectives. The SE block in Layer 3 delves into the dependencies between channels, while the ECA attention blocks in Layers 4 and 5 can quickly filter features, allowing the model to learn more comprehensive and representative features. In addition, this synergistic effect can improve the model's computational efficiency. ECA captures the dependencies between channels through one-dimensional convolution, avoiding the large number of parameters brought by the fully connected layers in the SE mechanism. Replacing the Swish activation function with the Mish activation function can improve the model's nonlinear expressive ability and reduce the risk of overfitting. Compared with the Swish activation function, the Mish activation function has a certain gradient response in the negative region, enabling it to learn more complex nonlinear feature relationships and accelerate the model's convergence speed.

[0090] Step 5: Train the state recognition model using the dataset, and use the trained state recognition model to identify the operating state of the contactor;

[0091] During training, the RAdam optimizer was used for parameter optimization, with a learning rate of 0.001 and 32 samples per iteration, for a maximum of 200 iterations. The model parameters were adjusted by minimizing the cross-entropy loss function using the backpropagation algorithm to make the model's predicted probability distribution as close as possible to the true label distribution, thereby improving classification performance. The calculation formula is as follows:

[0092] (11)

[0093] In the formula, For the number of categories, For the true value of the sample, This represents the probability value predicted by the model.

[0094] Example

[0095] This embodiment uses the H8C-12 AC contactor as an example to illustrate an online contactor operating status identification method based on image features and deep learning. The specific steps are as follows:

[0096] Step 1: Ten H8C-12 AC contactors that had undergone different number of load life tests and two brand-new test pieces were used as the research objects. Each underwent 1000 no-load tests. Vibration and acoustic signals of the contactors during the closing process 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 directly above the external slot of the moving contact. To prevent the vibration sensor and sound sensor from being too close and affecting the stability of the collected signals, an MPA201 sound sensor was placed directly above the contactor, 0.1m away from the vibration sensor. 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, the appropriate threshold is selected to classify the contactor's operating status into the following 6 types: long overtravel time, long engagement time, many bounces, fast core closing speed, long bounce time, and normal. The sample information for each operating status is shown in Table 1.

[0097]

[0098] Step 2: Based on the spectral analysis of the simulated and measured signals, it was found that the frequency characteristics of the signals generated by the collision between smooth and rough contacts are different. Figure 9 The image shows the frequency spectrum of the sound pressure simulation results for smooth and rough contacts. Figure 10 , 11 The image shows the spectrum of the acoustic signal during the collision process of smooth and rough contacts. It can be seen from the image that the signal during the collision process of rough contacts contains additional frequency components.

[0099] CEEMDAN decomposition was performed on the vibration and acoustic signals under various operating conditions. The CEEMDAN decomposition results for the acoustic and vibration signals are as follows: Figure 12 and 13 As shown; perform spectral analysis on each IMF component and calculate the main frequencies, such as... Figure 14 and 15 As shown, during modal screening, the IMF components containing additional frequency components are removed. Then, the correlation coefficient and energy value of each remaining IMF component with the original signal are calculated. The first three IMF components whose correlation coefficient and energy value are both greater than the average value are retained as effective modal components. Thus, the vibration signal and the sound signal each obtain three effective modal components.

[0100] Step 3: Generate SDP images using the effective modal components of vibration and acoustic signals, thus obtaining SDP images for each operating state; optimize the parameters of the SDP images based on the improved jumping spider algorithm, selecting 200 SDP images for each operating state for parameter optimization. The number of jumping spiders is set to 20, the maximum number of iterations to 100, and the angle magnification factor of the mirror symmetry plane is set. The range of values ​​is Time interval parameter The range of values ​​is The initial position of the jumping spider The perception range factor is 0.2. For example... Figure 16 As shown, after 80 iterations, the fitness value reaches its maximum, at which point the jumping spider's global optimum position is... The SDP images of the contactor show the greatest difference across its six operating states. This difference is most pronounced when the rotation angle of the mirror symmetry plane... At this time, the SDP images of the contactor in the six operating states are as follows: Figures 17-22 As shown, the shape of the mirror-symmetric plane in the SDP image can reflect the operating state of the contactor.

[0101] Step 4: Construct the state recognition model EfficientNet-ViT based on the EfficientNetV2 network;

[0102] Step 5: Train the state recognition model and use the trained state recognition model to identify the operating state of the contactor.

[0103] To illustrate the advantages of the effective modal components of vibration and acoustic signals in contactor condition identification, the original vibration and acoustic signals were converted into SDP images, resulting in vibration SDP images and acoustic SDP images. The SDP images obtained from the conversion of the effective modal components of vibration and acoustic signals were denoted as vibration-acoustic SDP images. These three types of images were input into the EfficientNet-ViT model for contactor operating condition identification. Accuracy and F1-Score were introduced as evaluation metrics to assess the accuracy of the condition identification model. The identification results are shown in Table 2. Compared to the vibration-acoustic SDP images, the identification accuracy and F1 score of the vibration and acoustic SDP images decreased significantly, indicating that the interference from contact electrical wear cannot be ignored, and it is necessary to remove the additional frequency components caused by contact surface roughness. Simultaneously, the joint analysis of the two signals enriched the condition information.

[0104] To illustrate the advantages of SDP images in state recognition tasks, vibration-acoustic signals were converted into Short Time Fourier Transform (STFT) images and Hilbert-Huang Transform (HHT) images, respectively. The recognition results are shown in Table 2. Using SDP images as model input yields the highest recognition accuracy. When HHT images are used as input, the recognition accuracy is lower than SDP images but higher than STFT images. SDP images are obtained by directly converting the time-series signal to polar coordinates, thus preserving the signal's feature information. In contrast, STFT is a time-frequency image, and processing the signal with a window function is crucial; the choice of window function affects the time-frequency image, resulting in the loss of some feature information and thus reducing recognition accuracy. HHT images are obtained by using Hilbert transform after obtaining a series of IMFs from the signal through Empirical Mode Decomposition (EMD). EMD is an adaptive time-frequency localization analysis method that, based on the data itself, can effectively extract information reflecting the local features of the signal, thus achieving a higher recognition accuracy than STFT images. However, EMD decomposition suffers from mode aliasing and endpoint effects, affecting the feature information in HHT images and resulting in lower recognition accuracy than SDP images.

[0105]

[0106] To demonstrate the effectiveness and applicability of the EfficientNet-ViT model, ResNet50, MobileViT, and EfficientNetV2 models were used for recognition based on the same dataset. To eliminate the influence of random errors, each model was tested 10 times, and the average of the 10 test results was taken. The recognition performance of different models is shown in Table 3. The EfficientNet-ViT model achieved the highest accuracy, which was 15% higher than RestNet50. It also exhibited good stability and required fewer computational parameters, reducing the computational burden. Figure 23 The training loss and validation accuracy curves of the EfficientNet-ViT model are shown. When the number of training iterations approaches 50, the model's validation accuracy reaches a high level, and the loss gradually decreases. The results indicate that the model learns effective features from SDP images, improving its fitting ability.

[0107]

[0108] In real-world working environments, noise signals inevitably mix into the signals acquired by sensors. To test the anti-interference capability of the SDP image in this method, Gaussian white noise of 5dB, 10dB, 15dB, 20dB, and 60dB was added to the original signal, respectively. Using the signal preprocessing methods in Table 2, namely the SDP image of the vibration-acoustic signal (Method A), the SDP image of the vibration signal (Method B), the SDP image of the sound signal (Method C), the STFT image of the vibration-acoustic signal (Method D), and the HHT image (Method E), the EfficientNet-ViT model was trained. The recognition accuracy at different signal-to-noise ratios was compared. Figure 24 As shown, the recognition accuracy based on SDP images decreases as the noise environment worsens, but the decrease is less severe than that of the time-frequency map-based method. Experiments show that noise has some impact on feature extraction from SDP images; however, SDP images are more robust to interference than time-frequency maps.

[0109] Compared with existing technologies, considering the impact of contact electrical wear state on the recognition effect of contactor motion characteristics, this invention first studies the differences in frequency characteristics of instantaneous collision signals of contacts under worn and smooth states through simulation and actual experiments. Second, it uses CEEMDAN to decompose the signal and filters modes based on the frequency components of each mode, their correlation with the original signal, and energy analysis: retaining components sensitive to changes in motion characteristics and eliminating additional frequency components caused by contact electrical wear, thereby reducing the interference of contact electrical wear state. Then, it optimizes the parameters of the SDP image based on an improved jumping spider algorithm, fusing state-related information from different frequency bands in vibration and acoustic signals to form an SDP image that visually displays the subtle changes in signals under various operating states. Finally, it constructs an EfficientNet-ViT model for image recognition, which quickly and efficiently captures different scale features of the SDP image with limited computational resources. The synergistic effect of the SE and ECA attention modules improves the model's computational efficiency and enhances feature extraction capabilities, while the introduced Mish activation function improves the model's recognition accuracy and generalization ability. Compared to other state recognition methods, this method does not perform excessive signal processing, and the feature information is well preserved. At the same time, the simple SDP technology fuses vibration and sound signals, enriching the state information. Introducing EfficientNetV2 into MobileViT not only reduces parameter calculation and ensures training efficiency, but also enhances the model's feature expression ability, enabling the contactor state recognition accuracy to reach 95.86%.

[0110] Any aspects not covered in this invention are applicable to existing technologies.

Claims

1. A method for online identification of contactor operating status based on image features and deep learning, characterized in that, Includes the following steps: Step 1: Collect vibration and acoustic signals of the contactor during operation under different operating conditions; 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; The vibration and sound signals are decomposed using CEEMDAN to obtain multiple IMF components. The main frequencies of each IMF component are calculated, and IMF components containing additional frequency components are removed. The correlation coefficient and energy value between each remaining IMF component and the original signal are calculated. IMF components with correlation coefficients and energy values ​​greater than the average value are retained to obtain the effective modal components of the vibration and sound signals. Step 3: Optimize the time interval parameters and angle magnification factor of SDP conversion technology using intelligent optimization algorithms. Based on the optimized time interval parameters and angle magnification factor, generate SDP images using the effective modal components of vibration and acoustic signals, thereby obtaining multiple SDP images under different operating conditions and forming a dataset. The expression for converting effective modal components into an SDP image is: (1) (2) (3) (4) In the formula, The rotation angle is the plane of mirror symmetry. The number of mirror-symmetric planes. The index of the mirror-symmetric plane. For point The radius in polar coordinates, For point time domain coordinates, , These represent the maximum and minimum amplitudes of the effective modal components; and Points The angles of counterclockwise and clockwise deflection along the mirror-symmetric plane in polar coordinates. For time interval parameters, For point coordinates It is the angle magnification factor, and ; Step 4: Construct a state recognition model based on the EfficientNetV2 network. This model consists of 5 layers. The SDP image is processed through convolution, batch normalization, and activation operations before being passed through each layer. The first layer includes two Fused-MBConv modules, the second layer includes four Fused-MBConv modules, the third layer includes one MBConv module and two MobileViT modules, the fourth layer includes one MBConv module and four MobileViT modules, and the fifth layer includes one MBConv module and three MobileViT modules. The output features of the fifth layer are passed through a fully connected layer to obtain the recognition result. The Mish activation function and ECA attention block replace the Swish activation function and SE block of the original MBConv module to obtain the MBConv modules of the fourth and fifth layers. Step 5: Train the state recognition model using the dataset, and use the trained state recognition model to identify the operating state of the contactor.

2. The online contactor operating status identification method based on image features and deep learning according to claim 1, characterized in that, When the intelligent optimization algorithm uses the jumping spider algorithm, the jumping spider position is initialized according to the following formula: (5) (6) In the formula, Indicates the first The initial position of the jumping spider. Indicates the index of the parameter being optimized. Indicates the number of parameters being optimized. , Indicates the first The upper and lower bounds of the parameter to be optimized. Indicates the first The initial values ​​of the optimized parameters, This indicates a random number generation operation; The difference value between SDP images of different operating states corresponding to each jumping spider is quantitatively calculated according to formula (7). The difference values ​​between different SDP images are accumulated to obtain the fitness of the jumping spider. (7) In the formula, For the first SDP image corresponding to jumping spider and The difference between them , For SDP images and pixel values, and These are SDP images. and The average of all pixel values , This represents the number of pixels in the length and width directions of the SDP image; The jumping spider's position is updated based on fitness, using the following formula: (8) (9) (10) In the formula, , It is the first Only the position of the jumping spider before and after the update. It is the first The stride length of a jumping spider. It is the location of the neighbor with the highest fitness within the current jumping spider's perception range. It is the first The fitness of jumping spiders, It is a coefficient that controls the movement step size. , These are the maximum and minimum values ​​of the coefficient. , These are the maximum and minimum fitness values.

3. The online contactor operating status identification method based on image features and deep learning according to claim 1 or 2, characterized in that, The expansion ratios of the Fused-MBConv modules in the first and second layers of the state recognition model are 1 and 4, respectively, while the expansion ratios of the MBConv modules in the third to fifth layers are all 4.

4. The online contactor operating status identification method based on image features and deep learning according to claim 1, characterized in that, The roughness of the severely worn contact is 142-183 μm.

Images