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

Abstract

The invention discloses a contactor operation state on-line identification method based on image features and deep learning. The method comprises the following steps: firstly, collecting vibration and sound signals of a contactor in different operation states in an action process; then, calculating the characteristic frequency of the signal in the collision process of the seriously worn contact, and taking the characteristic frequency as an additional frequency component; performing CEEMDAN decomposition on the vibration and sound signals to obtain a plurality of IMF components; the main frequency of each IMF component is calculated, and IMF components containing additional frequency components are removed; calculating a correlation coefficient and an energy value between each remaining IMF component and the original signal; keeping the IMF components of which the correlation coefficients and the energy values are greater than the average value to obtain effective modal components of the vibration and sound signals; thirdly, optimizing a time interval parameter and an angle amplification factor of the SDP conversion technology by utilizing an intelligent optimization algorithm, and generating SDP images by utilizing effective modal components of the vibration signals and the sound signals according to the optimized parameters, so as to obtain a plurality of SDP images in different operation states; and finally, a state recognition model is constructed and trained, and the trained model is used for state recognition. According to the method, the influence of interference components on the effective mode is avoided, the feature expression effect of the SDP image is improved, and the recognition precision is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of AC contactor operating status recognition, and specifically is an online contactor operating status recognition method based on image features and deep learning. Background Art

[0002] AC contactors are common low-voltage switching devices that are widely used in power distribution systems. As the demand for contactors in various fields continues to increase, higher requirements are also placed on their reliability. Contactor failure will affect the stability of the power distribution system, so online monitoring of its status is of great significance to improving the reliability and safety of the power distribution system and control system.

[0003] The selection of measurement signals is a prerequisite for state recognition. The vibration signal generated by the contactor during the opening and closing process can reflect the dynamic changes of the mechanical structure, and the acoustic signal contains comprehensive electrical and mechanical information. Both 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. There is no need to change 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]. Transactions of China Electrotechnical Society, 2021, 36(S2): 477-486.) proposed a motion analysis method for AC contactor contact system based on vibration signals, which effectively improved the recognition rate of operating status. You Yingmin et al. (You Yingmin, Wang Jingqin, Shu Liang, et al. Prediction method of AC contactor electrical life based on audio features [J]. Transactions of China Electrotechnical Society, 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 states of the contactor. Li Haiying et al. (Li Haiying, Sun Yue, Zhang Xiao, et al. Vacuum contactor fault diagnosis method based on modal time-frequency diagram and ResNet50 fusion [J]. High Voltage Engineering, 2023, 49(05): 1831-1840.) proposed a vacuum contactor fault diagnosis method based on wavelet time-frequency diagram and residual network. However, in the analysis of the operating state of the AC contactor, the influence of contact electrical wear on the vibration and acoustic signal characteristics is ignored, and it is all based on a single signal perspective, ignoring the fusion of multi-sensor information. Compared with the joint analysis of multiple signals, a single signal is easily disturbed by environmental factors; secondly, the time-frequency analysis method is highly dependent on expert knowledge to reflect the operating status with image features, and the time-frequency diagram is not easy to intuitively represent the status information; in addition, due to the high latitude and complex semantic information of image data, most models enhance the feature expression ability by increasing the network width and depth, but gradient disappearance is prone to occur, which will also increase the training time and computing resources of the model, and have high requirements on computer hardware. Therefore, considering that the vibration and sound signals generated by the contactor movement process have the characteristics of weak intensity and short duration, and in order to further reduce interference and obtain more accurate and comprehensive operating status information, the method of using appropriate feature fusion methods and deep learning models to identify the operating status of AC contactors needs further in-depth research. Summary of the invention

[0004] In view of the deficiencies in the prior art, the technical problem that the present invention intends to solve is to provide an online identification method for the operating status of a contactor based on image features and deep learning.

[0005] The present invention solves the technical problem by adopting the following technical solutions: A method for online identification of contactor operating status based on image features and deep learning, comprising the following steps: Step 1: Collect the vibration and sound 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 the characteristic frequency as an additional frequency component; Perform CEEMDAN decomposition on vibration and acoustic signals to obtain multiple IMF components respectively; calculate the main frequency of each IMF component and remove the IMF components containing additional frequency components; calculate the correlation coefficient and energy value between the remaining IMF components and the original signal; retain the IMF components whose correlation coefficient and energy value are greater than the average value to obtain the effective modal components of vibration and acoustic signals; Step 3: Use the intelligent optimization algorithm to optimize the time interval parameters and angle magnification factor of the SDP conversion technology. According to the optimized time interval parameters and angle magnification factor, the effective modal components of the vibration signal and the acoustic signal are used to generate the SDP image, and then multiple SDP images of different operating states are obtained to form a data set. The expression for converting the effective modal component into the SDP image is: (1) (2) (3) (4) In the formula, is the rotation angle of the mirror symmetry plane, is the number of mirror symmetry planes, is the serial number of the mirror symmetry plane, For point The radius in polar coordinates, For point The time domain coordinates of , are the maximum and minimum amplitudes of the effective modal components; and Points The angle of deflection counterclockwise and clockwise along the mirror symmetry plane in polar coordinates, is the time interval parameter, For point The coordinates of is the angle magnification factor, and ; Step 4: Build a state recognition model based on the EfficientNetV2 network. The model consists of 5 layers. The SDP image passes through each layer in turn after convolution, batch normalization and activation operations. 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 pass through the fully connected layer to obtain the recognition results. The Mish activation function and the ECA attention block replace the Swish activation function and the SE block of the original MBConv module to obtain the MBConv modules of the fourth and fifth layers. Step 5: Use the data set to train the state recognition model, and use the trained state recognition model to identify the operating state of the contactor.

[0006] Furthermore, when the intelligent optimization algorithm adopts the jumping spider algorithm, the jumping spider position is initialized according to the following formula: (5) (6) In the formula, Indicates The initial position of the jumping spider, Indicates the serial number of the optimized parameter, represents the number of optimized parameters, , Indicates The upper and lower bounds of the optimized parameters, Indicates The initial values ​​of the optimized parameters, Represents a random number generation operation; According to formula (7), the difference between the SDP images of different running states corresponding to each jumping spider is quantitatively calculated, and the difference between different SDP images is accumulated to obtain the fitness of the jumping spider; (7) In the formula, For the SDP image corresponding to jumping spider and The difference between , For SDP images and The pixel value of and They are SDP images and The average value of all pixel values, , is the number of pixels in the length and width directions of the SDP image; Update the jumping spider position according to the fitness. The jumping spider position update formula is as follows: (8) (9) (10) In the formula, , It is Only the jumping spider updates its front and back positions. It is The jumping spider's step length, is the location of the neighbor with the highest fitness within the current jumping spider’s sensing range, It is The fitness of jumping spiders, is the coefficient that controls the moving step size, , are the maximum and minimum values ​​of the coefficients, , are the maximum and minimum values ​​of fitness.

[0007] Furthermore, the expansion ratio coefficients of the Fused-MBConv modules of the first and second layers of the state recognition model are 1 and 4 respectively, and the expansion ratio coefficients of the MBConv modules of the third to fifth layers are all 4.

[0008] Furthermore, the roughness of the severely worn contact is 142-183 μm.

[0009] Compared with the prior art, the present invention has the following beneficial effects: (1) In the identification of contactor operating status based on motion characteristics, the electrical wear state of the contact has an impact on the characteristics of vibration and acoustic signals, resulting in unclear characteristics of vibration and acoustic signals. Through frequency characteristic analysis, it is 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 realize the information decoupling of electrical wear state and mechanical state at the data layer.

[0010] (2) The IMF components of vibration and acoustic signals are converted into images, and the state information of different frequency bands is integrated to intuitively display the subtle changes of vibration and acoustic signals under different operating states, highlighting the differences in different operating state information. It also overcomes the limitations of traditional time-frequency analysis methods that rely on expert knowledge and are not easy to intuitively represent state information. The parameters of the SDP image are optimized based on the improved jumping spider algorithm, and 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 diagram, the SDP image has stronger robustness and provides more reliable feature input for state recognition.

[0011] (3) Construct a state recognition model combining EfficientNetV2 and MobileViT. In order to improve the local feature extraction efficiency and multi-level feature extraction capability of MobileViT, EfficientNetV2 is introduced to make full use of the local feature extraction capability of EfficientNetV2 and the global modeling capability of Transformer, which not only enables the model to have better performance in resource-constrained environments, but also improves the generalization ability of the model. First, in the shallow structure of the EfficientNetV2-ViT model, by gradually increasing the expansion scale coefficient, the rapid extraction of low-level features is achieved at a low computational cost, and the number of feature map channels is increased to enrich the feature expression, thereby improving the model operation efficiency; secondly, the MBConv module is introduced in Layer3, Layer4 and Layer5 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 Layer4 and Layer5 to replace the original activation function and SE block, which can enhance the nonlinear expression ability of the model, avoid gradient disappearance, and improve the accuracy and computational efficiency of state recognition. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is the overall flow chart of the present invention; Figure 2 Flowchart for generating SDP image; Figure 3 It is a structural diagram of the state recognition model; Figure 4 It is a structural diagram of the Fused-MBConv module when the expansion ratio coefficient is equal to 1; Figure 5 This is a schematic diagram of the structure of the Fused-MBConv module when the expansion scale factor is not equal to 1; Figure 6 It is a schematic diagram of the structure of the MBConv module of the third layer; Figure 7Schematic diagram of the structure of the MBConv module of the fourth and fifth layers; Figure 8 It is a schematic diagram of the structure of the MobileViT module; Fig. 9 The spectrum diagram of the acoustic pressure simulation results of smooth contact and rough contact; Fig.10 It is the spectrum diagram of the acoustic signal during the collision process of smooth contacts; Fig.11 It is the spectrum diagram of the acoustic signal during the collision process of the rough contact; Fig.12 This is the CEEMDAN decomposition result diagram of the acoustic signal; Fig.13 This is the CEEMDAN decomposition result diagram of the vibration signal; Fig.14 It is the analysis diagram of IMFs of acoustic signal; Fig.15 It is the analysis diagram of vibration signal IMFs; Fig.16 It is the fitness value change curve of the SDP image parameter optimization process; Fig.17 This is the SDP image of the contactor in the state of long overtravel time; Fig.18 This is the SDP image of the contactor in the state of long pull-in time; Fig.19 This is the SDP image of the contactor when the number of contact bounces is high; Fig. 20 This is the SDP image of the contactor when the core closing speed is fast; Fig.21 This is the SDP image of the contactor when the contact bounce time is long; Fig. 22 This is the SDP image of the contactor in normal state; Fig.23 The accuracy and loss curves of the state recognition model training process; Fig.24 Comparison chart of recognition accuracy for different signal-to-noise ratio values. DETAILED DESCRIPTION

[0013] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and the present embodiment, but this is not intended to limit the protection scope of the present invention.

[0014] The present invention provides a contactor operating state online recognition method based on image features and deep learning (hereinafter referred to as method, see Figures 1 to 24 ), including the following steps: Step 1: Use the online operating status monitoring platform to collect vibration and sound signals of the contactor during operation under different operating conditions; Step 2: After performing spectrum analysis on the vibration and acoustic signals during the contactor contact collision process, compared with a new contactor, the roughness of the contact will affect the frequency characteristics of the signal, and contactors with different roughness will generate new characteristic frequencies. Therefore, a topography scanner is used to perform a topography scan on the contactor contact, and the roughness of the contactor contact is determined based on the root mean square deviation of the surface topography. The greater the root mean square deviation of the surface topography, the greater the roughness of the contact; the contact with a roughness in the range of 142-183μm is regarded as a severely worn contact, and the characteristic frequency of the severely worn contact signal is used as an additional frequency component; Based on the ensemble empirical mode decomposition (CEEMDAN), the vibration and acoustic signals are decomposed to obtain multiple IMF components and a residual component respectively; the IMF components of the vibration and acoustic signals are screened based on the three indicators of characteristic frequency, correlation and energy value; taking the vibration signal as an example, the spectrum analysis of each IMF component of the vibration signal is performed to calculate the main frequency of each IMF component. If the main frequency contains additional frequency components, the corresponding IMF component is eliminated; then, the correlation coefficient between the remaining IMF components and the original vibration signal is calculated according to the Pearson correlation coefficient method; the energy value of the remaining IMF components is calculated; the IMF components whose correlation coefficients and energy values ​​are both greater than the average value are retained, thereby obtaining the effective modal components of the vibration signal; similarly, the effective modal components of the acoustic signal are obtained; in this way, the decoupling of contact wear and operating status information at the data layer can be achieved, which is beneficial to improving the recognition accuracy of abnormal operating status of the contactor.

[0015] Step 3: Use the intelligent optimization algorithm to optimize the parameters of the SDP conversion technology. According to the optimized parameters, the effective modal components of the vibration signal and the acoustic signal are used to generate SDP images, and then multiple SDP images of different operating states are obtained to form a data set. 3-1. Using the symmetric point pattern (SDP) conversion technology, the SDP image is generated according to the effective modal components of the vibration signal and the acoustic signal, and the one-dimensional effective modal components are converted into polar coordinates to achieve data fusion from different data sources; the SDP image contains multiple mirror symmetry planes, and the number of mirror symmetry planes is equal to that of effective modal components. The expression for converting the effective modal components into the SDP image is: (1) (2) (3) (4) In the formula, is the rotation angle of the mirror symmetry plane, is the number of mirror symmetry planes, is the serial number of the mirror symmetry plane, For point The radius in polar coordinates, For point The time domain coordinates of , are the maximum and minimum amplitudes of the effective modal components; and Points The angle of deflection counterclockwise and clockwise along the mirror symmetry plane in polar coordinates, is the time interval parameter, For point The coordinates of is the angle magnification factor, and .

[0016] 3-2. Intelligent optimization algorithms include genetic algorithm, jumping spider algorithm, particle swarm algorithm, ant colony algorithm, gray wolf optimization algorithm, etc. Taking the jumping spider algorithm as an example, the jumping spider algorithm is improved and the parameters of the SDP image are optimized using the improved jumping spider algorithm, that is, the time interval parameter is optimized. and the angular magnification factor ; First, the number of jumping spiders, the maximum number of iterations, and the perception range factor are determined, the range of optimized parameters, the position of jumping spiders, the fitness, and the global optimal position are initialized, and the adaptive step range is defined; each jumping spider represents a set of optimized parameters (including the time interval parameters and angle magnification factors of all mirror symmetry planes of the SDP image), and the position initialization formula is as follows: (5) (6) In the formula, Indicates The initial position of the jumping spider, Indicates the serial number of the optimized parameter, represents the number of optimized parameters, , Indicates The upper and lower bounds of the optimized parameters, Indicates The initial values ​​of the optimized parameters, Represents a random number generation operation; Next, based on the image correlation analysis method, the difference values ​​between the SDP images of different running states corresponding to each jumping spider are quantitatively calculated according to formula (7), and the difference values ​​between different SDP images are accumulated to obtain the fitness of the jumping spider; (7) In the formula, For the SDP image corresponding to jumping spider and The difference between , For SDP images and The pixel value of and They are SDP images and The average value of all pixel values, , is the number of pixels in the length and width directions of the SDP image; The larger the fitness, the greater the difference between the two SDP images. If the fitness of the jumping spider is greater than the current global optimal fitness, the jumping spider is used to update the global optimal fitness and the global optimal position. Then, the jumping spider position is updated according to the fitness; the distance between each jumping spider and other jumping spiders is calculated, and the neighbor with the largest fitness is found according to the perception range, thereby using local information to guide the jumping spider's search direction, and the jumping spider position is updated according to the moving step length, so that the jumping spider gradually approaches the position of the optimal neighbor, thereby searching for the optimal solution in the global space; the jumping spider position update formula is as follows: (8) (9) (10) In the formula, , It is Only the jumping spider updates its front and back positions. It is The jumping spider's step length, is the location of the neighbor with the highest fitness within the current sensing range of the jumping spider; It is The fitness of jumping spiders, is the coefficient that controls the moving step length, which is dynamically adjusted according to the fitness of the jumping spider; , are the maximum and minimum values ​​of the coefficients, , is the maximum and minimum value of fitness; Finally, after the iteration is completed, the global optimal position is obtained, that is, a set of optimized parameters with the largest difference in SDP images of different operating states; multiple SDP images of each operating state are generated according to the optimized parameters to obtain a data set.

[0017] Step 4: Build the state recognition model EfficientNet-ViT based on the EfficientNetV2 network; Figure 3 As shown in the figure, the EfficientNet-ViT model mainly includes 5 layers. The SDP image passes through each layer in turn after convolution (Conv2D), batch normalization (BN), and activation (Siwh) operations. The first layer Layer1 includes two Fused-MBConv modules with an expansion ratio of 1, and the second layer Layer2 includes four Fused-MBConv modules with an expansion ratio of 4; the third layer Layer3 includes one MBConv module and two MobileViT modules, the fourth layer Layer4 includes one MBConv module and four MobileViT modules, and the fifth layer Layer5 includes one MBConv module and three MobileViT modules. The output features of the fifth layer Layer5 pass through the fully connected layer to obtain the output of the EffficientNetV2-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 implement image classification. The structure of the Fused-MBConv module of the first layer Layer1 and the second layer Layer2 can be found in Figure 4 and Figure 5 The expansion ratio of each MBConv module is 4. The structure of the MBConv module of the third layer Layer3 is as follows: Figure 6 As shown; the structure of the MBConv module of the fourth layer Layer4 and the fifth layer Layer5 is as follows Figure 7 As shown in Figure 2, the Mish activation function and ECA attention block are used to replace the Swish activation function and SE block of the original MBConv module; the structure of the MobileViT module is as follows Figure 8 shown.

[0018] In the shallow structure (first and second layers) of the EfficientNetV2-ViT model, by gradually increasing the expansion ratio coefficient, low-level features can be quickly extracted at a low computational cost, improving the model's operating efficiency; secondly, the local fine features extracted by the MBConv module in Layer3 to Layer5 are used as the input of the MobileViT module. The Transformer module in the MobileViT module uses the local feature position information extracted by the MBConv module and combines its own self-attention mechanism to more accurately establish the association between different regions in the SDP image, thereby improving the quality of the overall feature representation. The combination of the MBConv and MobileViT modules can achieve the coordinated extraction of local and global features. The extracted features complement each other at different levels, enriching the semantic information of the features. Introducing the Mish activation function and ECA attention block 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 levels, so that the model can process and learn data from multiple angles. The SE block in Layer 3 deeply explores the dependencies between channels, and the ECA attention blocks in Layer 4 and 5 can quickly screen features, so that the model can learn more comprehensive and representative features. In addition, this synergy can improve the computational efficiency of the model. ECA captures the dependencies between channels through one-dimensional convolution, avoiding the large number of parameters brought by the fully connected layer in the SE mechanism; using the Mish activation function instead of the Swish activation function can improve the nonlinear expression ability of the model 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, which enables it to learn more complex nonlinear feature relationships and accelerate the convergence of the model.

[0019] Step 5: Use the data set to train the state recognition model, and use the trained state recognition model to identify the operating state of the contactor; During the training process, the RAdam optimizer is used for parameter optimization, the learning rate is set to 0.001, and 32 samples are operated in each iteration, with a maximum number of iterations of 200. The model parameters are adjusted by minimizing the cross entropy loss function through the back propagation algorithm to make the model's predicted probability distribution as close as possible to the true label distribution to improve the classification performance. The calculation formula is as follows: (11) In the formula, is the number of categories, is the true value of the sample, is the probability value predicted by the model.

[0020] Example This embodiment takes the H8C-12 AC contactor as an example, and the contactor operating status online recognition method based on image features and deep learning is specifically described as follows: Step 1: Take 10 H8C-12 AC contactors after different operation times of the load life test and 2 new test products as the research objects, conduct 1000 no-load tests respectively, and use the online monitoring platform of the operating status to collect the vibration and sound signals of the contactor during the closing process under different operating conditions. Based on the simulation results of contact collision sound, the vibration and sound pressure amplitude of the contact system in the collision direction are the largest, and the KS78B10 acceleration sensor is installed directly above the external card slot of the moving contact; in order to prevent the vibration sensor and the sound sensor from being close to each other and affecting the stability of the collected signal, the MPA201 sound sensor is placed directly above the contactor and 0.1m away from the vibration sensor. The signal is displayed and analyzed based on the DH5922D dynamic test analyzer, and the sampling frequency of the test is set to 100kHz. According to the parameter values ​​of this type of contactor provided by the manufacturer and combined with the calculation results, appropriate thresholds are selected to divide the operating status of the contactor into the following 6 types: long overtravel time, long pull-in time, many bounce times, fast core closing speed, long bounce time and normal. The sample information under various operating conditions is shown in Table 1.

[0021]

[0022] Step 2: Based on the spectrum analysis of the simulated and measured signals, it is found that the frequency characteristics of the signals generated by the collision between the smooth contact and the rough contact are different. Fig. 9 The spectrum diagram of the sound pressure simulation results of smooth contact and rough contact. Fig.10 , 11 This is the spectrum diagram of the acoustic signal during the collision process between a smooth contact and a rough contact. It can be seen from the figure that the signal during the collision process of the rough contact contains additional frequency components.

[0023] The vibration and acoustic signals of each operating state are decomposed by CEEMDAN. The CEEMDAN decomposition results of acoustic and vibration signals are as follows: Fig.12 and 13 As shown; perform spectrum analysis on each IMF component and calculate the main frequency, such as Fig.14 and 15 As shown; during modal screening, the IMF components containing additional frequency components are removed, and then the correlation coefficients and energy values ​​of the remaining IMF components and the original signal are calculated, and the first three IMF components whose correlation coefficients and energy values ​​are greater than the average value are retained as effective modal components, then the vibration signal and the sound signal respectively obtain three effective modal components.

[0024] Step 3: Generate SDP images using effective modal components of vibration and acoustic signals, and then obtain SDP images under various operating conditions; optimize the parameters of SDP images based on the improved jumping spider algorithm, and select 200 SDP images for each operating state for parameter optimization. Set the number of jumping spiders to 20, the maximum number of iterations to 100, and the angle magnification factor of the mirror symmetry plane to The value range is , time interval parameter The value range is , the jumping spider's initial position , the perception range factor is 0.2. Fig.16 As shown in Figure 2, after 80 iterations, the fitness value reaches the maximum, and the global optimal position of the jumping spider is , the SDP images of the contactor in the six operating states have the largest difference. At this time, the SDP images of the contactor in the six operating states are as follows: Figures 17 to 22 As shown in Figure 3, the shape of the mirror symmetry plane in the SDP image can reflect the operating status of the contactor.

[0025] Step 4: Build the state recognition model EfficientNet-ViT based on the EfficientNetV2 network; Step 5: Train the state recognition model, and use the trained state recognition model to identify the operating state of the contactor.

[0026] In order to illustrate the advantages of the effective modal components of vibration and acoustic signals in the contactor state recognition task in this method, the single original vibration signal and sound signal are converted into SDP images respectively to obtain vibration signal SDP images and acoustic signal SDP images. The SDP images obtained by converting the effective modal components of vibration and acoustic signals are recorded as vibration-acoustic signal SDP images. The three types of images are input into the EfficientNet-ViT model for contactor operation state recognition. The two evaluation indicators of accuracy and F1-Score are introduced to evaluate the accuracy of the state recognition model. The recognition results are shown in Table 2. Compared with the vibration-acoustic signal SDP image, the recognition accuracy and F1 score of the vibration signal SDP image and the acoustic signal SDP image are significantly reduced, indicating that the interference of contact electrical wear cannot be ignored, and it is necessary to eliminate the additional frequency components caused by the roughness of the contact surface; at the same time, the joint analysis of the two signals enriches the state information.

[0027] To illustrate the advantages of SDP images in state recognition tasks, the vibration-acoustic signals are converted into short-time Fourier transform (STFT) images and Hilbert-Huang transform (HHT) images, respectively. The recognition results are shown in Table 2. The SDP image has the highest recognition accuracy when used as the model input. The recognition accuracy of the HHT image is lower than that of the SDP image, but higher than that of the STFT image. The SDP image is obtained by directly converting the time series signal into the polar coordinate system, and the characteristic information of the signal will not be lost. In contrast, the STFT is a time-frequency image. By adding a window function to process the signal, the choice of the window function will affect the time-frequency graph, and some characteristic information will be lost, thus reducing the recognition accuracy. The HHT image is obtained by using the Hilbert transform after obtaining a series of IMFs of the signal through empirical mode decomposition (EMD). EMD is an adaptive time-frequency localization analysis method. Based on the decomposition of the data itself, it can effectively extract information that can reflect the local characteristics of the signal, so its recognition accuracy is higher than that of the STFT image. However, EMD decomposition has problems of modal aliasing and endpoint effect, which affects the characteristic information in the HHT image, making the recognition accuracy lower than that of the SDP image.

[0028]

[0029] In order to illustrate the effectiveness and applicability of the EfficientNet-ViT model, the ResNet50, MobileViT and EfficientNetV2 models were used for recognition based on the same data set. In order to eliminate the influence of accidental errors, each model was tested 10 times, and the average of the results of the 10 tests was taken. The recognition performance of different models is shown in Table 3. The EfficientNet-ViT model has the highest accuracy, which is 15% higher than that of RestNet50. It has good stability, fewer calculation parameters, and reduces the amount of calculation. Fig.23 The training loss and verification accuracy curve of the EfficientNet-ViT model. When the number of training times is close to 50, the verification accuracy of the model reaches a high level and the loss gradually decreases. The results show that the model learns effective features from the SDP image and improves the model's fitting ability.

[0030]

[0031] In the actual working environment, the signals collected by the sensor will inevitably be mixed with noise signals. In order to test the anti-interference ability of the SDP image in this method, 5db, 10dB, 15dB, 20dB and 60dB of Gaussian white noise are 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), and then training the EfficientNet-ViT model, the recognition accuracy comparison at different signal-to-noise ratio values ​​is as follows: Fig.24 As shown in the figure, as the noise environment becomes worse, the recognition accuracy based on SDP images decreases, but the decrease is weaker than that based on the time-frequency graph method. Experiments show that noise has a certain impact on the feature extraction of SDP images, but SDP images have stronger anti-interference ability than time-frequency graphs.

[0032] Compared with the prior art, considering that the electrical wear state of the contact has an impact on the recognition effect of the contactor's motion characteristics, the present invention first combines simulation and actual experiments to study the frequency characteristic differences of the instantaneous collision signal of the contact in the worn and smooth states; secondly, CEEMDAN is used to decompose the signal, and the mode is screened according to the frequency components of each mode, the degree of correlation with the original signal, and energy analysis: the components sensitive to the change of motion characteristics are retained, and the additional frequency components caused by the electrical wear of the contact are eliminated, thereby reducing the interference of the electrical wear state of the contact; then, based on the improved jumping spider algorithm, the parameters of the SDP image are optimized, and the state-related information of different frequency bands in the vibration and sound signals is integrated to form an SDP image, which intuitively displays the slight changes of the signal under each operating state. Finally, the EfficientNet-ViT model is constructed for image recognition. With less computing resources, the different scale features of the SDP image are quickly and efficiently captured. The synergy of the SE and ECA attention modules improves the model calculation efficiency and enhances the feature extraction capability. At the same time, the introduced Mish activation function improves the model recognition accuracy and generalization capability. Compared with other state recognition methods, this method does not process the signal too much, and the feature information is well preserved. At the same time, the simple SDP technology fuses the vibration and acoustic signals to enrich the state information. The introduction of EfficientNetV2 in MobileViT not only reduces parameter calculation and ensures training efficiency, but also enhances the feature expression ability of the model, making the accuracy of contactor state recognition reach 95.86%.

[0033] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A contactor operating status online recognition method based on image features and deep learning, characterized in that: The following steps are involved: Step 1: Collect the vibration and sound 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 the characteristic frequency as an additional frequency component; Perform CEEMDAN decomposition on vibration and acoustic signals to obtain multiple IMF components respectively; calculate the main frequency of each IMF component and remove the IMF components containing additional frequency components; calculate the correlation coefficient and energy value between the remaining IMF components and the original signal; retain the IMF components whose correlation coefficient and energy value are greater than the average value to obtain the effective modal components of vibration and acoustic signals; Step 3: Use the intelligent optimization algorithm to optimize the time interval parameters and angle magnification factor of the SDP conversion technology. According to the optimized time interval parameters and angle magnification factor, the effective modal components of the vibration signal and the acoustic signal are used to generate the SDP image, and then multiple SDP images of different operating states are obtained to form a data set. The expression for converting the effective modal component into the SDP image is: (1) (2) (3) (4) In the formula, is the rotation angle of the mirror symmetry plane, is the number of mirror symmetry planes, is the serial number of the mirror symmetry plane, For point The radius in polar coordinates, For point The time domain coordinates of , are the maximum and minimum amplitudes of the effective modal components; and Points The angle of deflection counterclockwise and clockwise along the mirror symmetry plane in polar coordinates, is the time interval parameter, For point The coordinates of is the angle magnification factor, and ; Step 4: Build a state recognition model based on the EfficientNetV2 network. The model consists of 5 layers. The SDP image passes through each layer in turn after convolution, batch normalization and activation operations. 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 pass through the fully connected layer to obtain the recognition results. The Mish activation function and the ECA attention block replace the Swish activation function and the SE block of the original MBConv module to obtain the MBConv modules of the fourth and fifth layers. Step 5: Use the data set to train the state recognition model, and use the trained state recognition model to identify the operating state of the contactor.

2. The method for online identification of contactor operating status based on image features and deep learning according to claim 1 is characterized in that: When the intelligent optimization algorithm adopts the jumping spider algorithm, the jumping spider position is initialized according to the following formula: (5) (6) In the formula, Indicates The initial position of the jumping spider, Indicates the serial number of the optimized parameter, represents the number of optimized parameters, , Indicates The upper and lower bounds of the optimized parameters, Indicates The initial values ​​of the optimized parameters, Represents a random number generation operation; According to formula (7), the difference between the SDP images of different running states corresponding to each jumping spider is quantitatively calculated, and the difference between different SDP images is accumulated to obtain the fitness of the jumping spider; (7) In the formula, For the SDP image corresponding to jumping spider and The difference between , For SDP images and The pixel value of and They are SDP images and The average value of all pixel values, , is the number of pixels in the length and width directions of the SDP image; Update the jumping spider position according to the fitness. The jumping spider position update formula is as follows: (8) (9) (10) In the formula, , It is Only the jumping spider updates its front and back positions. It is The jumping spider's step length, is the location of the neighbor with the highest fitness within the current jumping spider’s sensing range, It is The fitness of jumping spiders, is the coefficient that controls the moving step size, , are the maximum and minimum values ​​of the coefficients, , are the maximum and minimum values ​​of fitness.

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

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

Images