GIS disconnector fault diagnosis method based on RGB time-frequency representation and attention fusion

Abstract

The present application relates to a GIS disconnector fault diagnosis method based on RGB time-frequency representation and attention fusion, belonging to the technical field of high-voltage electrical equipment state monitoring and fault diagnosis. The method collects the vibration acceleration signal of the GIS disconnector shell, applies the Morlet wave continuous wavelet transform to each segment after the sliding window overlap segmentation, and extracts the time-frequency amplitude matrix of the low-frequency band 0-1000Hz, the medium-frequency band 1000-4000Hz and the high-frequency band 4000-5000Hz respectively; the three-frequency band matrix is mapped to the R, G and B channels of the RGB image after normalization, and a 64x64x3 RGB time-frequency feature image is generated; the image is input into the convolutional neural network embedded with the SE attention module, the SE module adaptively weights the feature channels through compression, excitation and re-labeling, enhances the fault sensitive frequency band and suppresses the secondary frequency band, and outputs the diagnosis result. The present application integrates multi-band time-frequency features through RGB three-channel fusion, makes up for the defects that a single transformation cannot consider both global trend and local mutation, and realizes adaptive focusing of the fault sensitive frequency band.

Description

Technical Field

[0001] This invention belongs to the field of high-voltage electrical equipment condition monitoring and fault diagnosis technology, and relates to a fault diagnosis method for GIS disconnect switches based on RGB time-frequency representation and attention fusion. Background Technology

[0002] During long-term operation, GIS disconnect switches, due to their completely enclosed structure, make it difficult to directly observe their internal mechanical condition. Maintenance personnel cannot diagnose typical mechanical faults such as poor contact, loose shielding, or foreign object retention using conventional methods. A serious fault could lead to widespread power outages and significant economic losses. Therefore, researching accurate and reliable methods for diagnosing mechanical faults in GIS disconnect switches has significant engineering value.

[0003] Existing methods for diagnosing mechanical faults in GIS disconnect switches are mainly divided into two categories: traditional signal processing and shallow machine learning methods, and deep learning methods based on one-dimensional convolutional neural networks.

[0004] The first category comprises traditional signal processing and shallow machine learning methods. This method uses accelerometers or laser Doppler vibration meters to collect vibration acceleration signals of the GIS shell. After preprocessing such as denoising and filtering, it employs time-frequency analysis using methods such as short-time Fourier transform, wavelet packet decomposition, or empirical mode decomposition to generate time-frequency images. Then, it extracts statistical features such as energy entropy, kurtosis, skewness, and peak-to-peak value. Finally, the feature vectors are input into shallow classifiers such as support vector machines or random forests to determine whether the equipment is in a normal state or has a mechanical fault. The process is as follows: Figure 1 As shown. The principle of this type of method is that GIS disconnect switches have specific vibration response modes under different mechanical states, and their vibration acceleration signals exhibit identifiable differences in time-frequency representation. However, this type of method has the following shortcomings: First, the feature extraction process relies on human experience and lacks adaptability; second, the time-frequency transformation method used is relatively simple, and the generated time-frequency representation has limited dimensions, making it difficult to simultaneously take into account both the global trend of signal change and local abrupt change information. For fault types with high spectral similarity—such as... Figure 2 The spectrum diagrams showing the internal foreign object state and the normal state—the characteristics of artificial structures are difficult to distinguish accurately, which can easily lead to misjudgment.

[0005] The second category is the one-dimensional convolutional neural network (1D-CNN) deep learning method. This method directly uses the original one-dimensional vibration acceleration signal as network input, automatically extracting features through multiple stacked one-dimensional convolutional and pooling layers. The diagnostic results are output by a fully connected layer and a softmax classifier. Its architecture is as follows: Figure 3As shown. However, the receptive field of the convolution kernel in a one-dimensional convolutional neural network is limited to a single time dimension, and it cannot fully utilize the joint distribution characteristics of multi-dimensional information in vibration acceleration signals. Therefore, when faced with faults with high spectral similarity, 1D-CNN also struggles to effectively distinguish them and is prone to misjudgment.

[0006] In summary, the common shortcomings of existing methods are: insufficient utilization of multi-band information of vibration acceleration signals and lack of adaptive focusing capability for fault-sensitive frequency bands. Summary of the Invention

[0007] In view of this, this invention addresses the problems of GIS disconnect switches during long-term operation, where the internal mechanical state is difficult to observe directly due to the completely enclosed structure of the equipment, making it impossible for maintenance personnel to diagnose typical mechanical faults such as poor contact of contacts, loose shielding, and retention of foreign objects through conventional means. It also addresses the shortcomings of existing fault diagnosis methods, such as difficulty in accurately distinguishing faults with high spectral similarity, reliance on manual experience for feature extraction, limited receptive field of one-dimensional convolutional networks, and lack of frequency band adaptive weighting mechanisms. Therefore, this invention proposes a fault diagnosis method for GIS disconnect switches based on RGB time-frequency representation and attention fusion. This method fully utilizes multi-band information from vibration acceleration signals and achieves adaptive frequency band focusing, ensuring the accuracy of mechanical fault diagnosis for GIS disconnect switches while improving the adaptability of the diagnostic method under complex field conditions.

[0008] To achieve the above objectives, the present invention provides the following technical solution: A fault diagnosis method for GIS disconnect switches based on RGB time-frequency representation and attention fusion includes the following steps: S1: Collect vibration acceleration signals of the GIS disconnect switch housing; S2: Perform time-frequency transformation on the vibration acceleration signal to extract time-frequency features of multiple different frequency bands; S3: Map the time-frequency features of the multiple different frequency bands to different color channels of an RGB color image to generate an RGB time-frequency feature image that fuses multi-frequency band information; S4: Input the RGB time-frequency feature image into a convolutional neural network diagnostic model with an attention mechanism, and adaptively weight the feature channels through the attention mechanism to output the fault diagnosis result of the GIS disconnect switch.

[0009] Further, in step S1, the vibration acceleration signal of the GIS disconnect switch housing is acquired by a laser Doppler vibration meter or an accelerometer, with a sampling frequency of 20kHz; the acquired vibration acceleration signal is then preprocessed by denoising and filtering.

[0010] Further, in step S2, the vibration acceleration signal is segmented using a sliding window to obtain multiple overlapping time segments, and then the time-frequency features are extracted from each time segment; the sliding window has a window length of 1024 sampling points, a step size of 512 sampling points, and an overlap rate of 50% between adjacent windows.

[0011] Furthermore, in step S2, the time-frequency transformation is a continuous wavelet transform, and the wavelet mother function used in the continuous wavelet transform is the Morlet wavelet.

[0012] Furthermore, the multiple different frequency bands include low frequency band, mid frequency band and high frequency band, the frequency range of the low frequency band is 0-1000Hz, the frequency range of the mid frequency band is 1000-4000Hz, and the frequency range of the high frequency band is 4000-5000Hz; after continuous wavelet transform, the time-frequency amplitude matrix corresponding to each frequency band is obtained respectively.

[0013] Further, in step S3, the time-frequency amplitude matrices corresponding to the low-frequency, mid-frequency, and high-frequency bands are normalized respectively, and their values ​​are mapped to the [0,1] interval. Then, the normalized low-frequency band time-frequency amplitude matrix is ​​mapped to the R channel of the RGB color image, the mid-frequency band time-frequency amplitude matrix is ​​mapped to the G channel, and the high-frequency band time-frequency amplitude matrix is ​​mapped to the B channel, and the two are merged to generate the RGB time-frequency feature image. The RGB time-frequency feature image is uniformly scaled to 64×64 pixels by bilinear interpolation.

[0014] Further, in step S4, the convolutional neural network diagnostic model embedded with an attention mechanism includes a first convolutional layer, an SE attention module, a first pooling layer, a second convolutional layer, a second pooling layer, a global average pooling layer, a Dropout layer, and a fully connected output layer connected in sequence; the dropout rate of the Dropout layer is 0.4.

[0015] Furthermore, the first convolutional layer contains 32 3×3 convolutional kernels, uses the same padding method, has a stride of 1, and outputs a feature map size of 64×64×32; the second convolutional layer contains 64 3×3 convolutional kernels, uses the same padding method, has a stride of 1, and outputs a feature map size of 32×32×64; both the first pooling layer and the second pooling layer use max pooling with a 2×2 window and a stride of 2.

[0016] Furthermore, the SE attention module first compresses the spatial information of each feature channel into a global descriptor through global average pooling, then learns the importance weights of each channel through two fully connected layers, and finally multiplies the importance weights with the original feature map channel by channel to achieve enhancement of fault-sensitive frequency bands and suppression of secondary frequency bands.

[0017] Furthermore, during the model training phase, the generated RGB time-frequency feature images are randomly divided into a training set and a test set in an 8:2 ratio. The training set is used for model parameter learning, and the test set is used to evaluate model performance.

[0018] The beneficial effects of this invention are as follows: (1) This invention decomposes the original one-dimensional vibration acceleration signal into time-frequency amplitude matrices of three frequency bands—low frequency (0-1000Hz), mid frequency (1000-4000Hz), and high frequency (4000-5000Hz)—through continuous wavelet transform, and maps them to the R, G, and B channels of an RGB color image, respectively, to generate an RGB time-frequency feature image that integrates multi-frequency band information. The three-band division scheme is determined based on the physical vibration characteristics of different fault types of GIS disconnect switches: the energy of normal operation and internal foreign matter states is concentrated in the low frequency band below 1000Hz; the loose shield generates higher harmonics under the action of an alternating electric field, and the energy is concentrated in the mid frequency band of 1000-4000Hz; poor contact of the contacts excites partial discharge and micro-arcs, generating high-frequency impact components that extend to 5000Hz, corresponding to the high frequency band of 4000-5000Hz. By separating, encoding, and visually fusing the features of different frequency bands, the RGB three-channel image can simultaneously present the time-frequency distribution characteristics of the low, medium, and high frequency bands. This fully utilizes the joint distribution characteristics of multi-dimensional information in the vibration acceleration signal, enabling the subsequent diagnostic model to synchronously obtain discrimination information of different frequency bands from the image. This effectively compensates for the shortcomings of a single transformation method that cannot take into account both the global trend and local abrupt change information of the signal.

[0019] (2) This invention uses Morlet wavelet as the wavelet mother function of continuous wavelet transform. Morlet wavelet is a complex exponential function under Gaussian envelope, which has good localization characteristics in both time and frequency domains. It can obtain high frequency resolution in low frequency band and high time resolution in high frequency band, overcoming the inherent defect of short time Fourier transform time and frequency resolution being constrained by fixed window length. At the same time, the waveform characteristics of Morlet wavelet are highly matched with the transient vibration acceleration signal generated by GIS disconnect switch connection operation and mechanical impact. Its oscillation attenuation characteristics can effectively capture the time and frequency aggregation characteristics of fault impact. The generated time and frequency amplitude matrix has clear energy aggregation patterns in both time and frequency dimensions, which is beneficial for feature extraction of subsequent deep learning models.

[0020] (3) This invention embeds an SE (Squeeze-and-Excitation) attention module into the convolutional neural network. This module is located after the first convolutional layer and before the first pooling layer, and recalibrates each feature channel in the early stage of feature extraction. The shallow feature map has a high spatial resolution (64×64×32) and retains rich local time-frequency details. The SE module compresses the spatial information of each channel into a global descriptor through global average pooling, and then automatically learns the importance weights of each channel through two fully connected layers. Finally, the weights are multiplied with the original feature map channel by channel, so that the model can automatically identify and enhance the sensitive frequency band features related to the fault during the training process, while suppressing the secondary frequency band information unrelated to the fault. Compared with the shortcomings of existing 1D-CNN methods that treat each frequency band information indiscriminately and lack the ability to focus on frequency bands, the SE attention module enables the model to focus more specifically on the key frequency band features that can distinguish between the two fault types, such as internal foreign objects and normal states, when facing fault types with high spectral similarity.

[0021] (4) This invention adopts an end-to-end diagnostic architecture, which automates the entire process from raw vibration acceleration signal acquisition, sliding window segmentation, continuous wavelet transform, RGB image generation to CNN-SE model diagnostic output, without the need for manual design of feature parameters. Compared with the traditional signal processing method that relies on domain experts to manually construct statistical features, this invention avoids the empirical bias introduced in the manual feature extraction process and improves the adaptability and versatility of the diagnostic method.

[0022] (5) Experimental results show that, on a 220kV GIS disconnector experimental platform, the method of this invention achieves a final diagnostic accuracy of over 94% for four operating states: internal foreign objects, poor contact of contacts, normal operation, and loose shielding. Compared with five existing methods, including SVM, DBN, 1D-CNN, ViT, and ResNet-18, the loss function curve of this invention decreases rapidly in the early stages of training, with the smallest fluctuation amplitude, and eventually converges to the lowest loss value; the accuracy curve has the highest initial accuracy and the fastest convergence speed, and tends to stabilize after fewer iterations. Confusion matrix analysis shows that the accuracy, precision, recall, and F1-Score of each state all reach over 93%, verifying the comprehensive effectiveness of this invention in multi-class fault identification tasks.

[0023] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0024] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating the traditional signal processing and SVM diagnostic methods in existing technologies.

[0025] Figure 2 This is a schematic diagram comparing the spectrum of the GIS disconnect switch under foreign object conditions and the normal state.

[0026] Figure 3 This is a schematic diagram of the architecture of the 1D-CNN deep learning method in the existing technology.

[0027] Figure 4 This is a schematic diagram of the sample segmentation and three-band vibration acceleration signal generation process provided in the embodiment of the present invention.

[0028] Figure 5 This is a schematic diagram of the CNN-SE attention fusion model structure provided in an embodiment of the present invention.

[0029] Figure 6 This is a schematic diagram comparing the training loss function curve and accuracy curve of the method provided in this embodiment of the invention with existing methods.

[0030] Figure 7 This is a schematic diagram of the confusion matrix of the method provided in the embodiments of the present invention under four operating states.

[0031] Figure 8 This is a schematic diagram of the overall process of the GIS disconnector fault diagnosis method based on RGB time-frequency representation and attention fusion provided in the embodiments of the present invention.

[0032] Figure 9 This is a schematic diagram of the main steps of the present invention. Detailed Implementation

[0033] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0034] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0035] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0036] To clearly illustrate the technical contributions of this invention, we will first combine... Figures 1 to 3 The shortcomings of existing technologies are analyzed.

[0037] As described in the background section, existing methods for diagnosing mechanical faults in GIS disconnect switches are mainly divided into two categories. The first category consists of traditional signal processing and shallow machine learning methods, the typical process of which is as follows: Figure 1 As shown: After acquiring vibration acceleration signals of the GIS shell using an accelerometer or laser Doppler vibrometer, the signals undergo preprocessing such as denoising and filtering. Time-frequency analysis is then performed using methods such as short-time Fourier transform, wavelet packet decomposition, or empirical mode decomposition to generate time-frequency images. Statistical features such as energy entropy, kurtosis, skewness, and peak-to-peak value are then extracted. Finally, the feature vectors are input into shallow classifiers such as support vector machines or random forests for diagnosis. However, the time-frequency transformation techniques used in this type of method are relatively simple, and the generated time-frequency representation has limited dimensions, making it difficult to simultaneously consider both the global trend and local abrupt changes of the signal. Figure 2 As shown, the spectrum diagrams of the internal foreign object state and the normal state are highly similar in the low-frequency band. The artificially constructed features make it difficult to accurately distinguish between the two, which can easily lead to misjudgment.

[0038] The second category is one-dimensional convolutional neural network (1D-CNN) deep learning methods, whose architecture is as follows: Figure 3As shown, the original one-dimensional vibration acceleration signal is directly used as the network input. Features are automatically extracted through multiple stacked one-dimensional convolutional and pooling layers, and the diagnostic results are output by a fully connected layer and a Softmax classifier. However, the receptive field of the convolutional kernel of the one-dimensional convolutional neural network is limited to a single time dimension, which cannot fully utilize the joint distribution characteristics of multi-dimensional information in the vibration acceleration signal. It is also difficult to effectively distinguish faults with high spectral similarity.

[0039] To address the common shortcomings of the two aforementioned methods—insufficient utilization of multi-frequency information from vibration acceleration signals and a lack of adaptive focusing capability for fault-sensitive frequency bands—this invention proposes a GIS disconnector fault diagnosis method based on RGB time-frequency representation and attention fusion. The following is a combination of... Figures 4 to 8 The specific implementation steps of the method of the present invention will be described in detail. For example... Figure 9 As shown, the present invention includes the following steps: Step S1: Collect vibration acceleration signals of the GIS disconnect switch housing. The vibration acceleration signal x(t) of the GIS disconnector casing is acquired using a laser Doppler vibration meter or an accelerometer, with a sampling frequency of 20kHz. The selection of the sampling frequency of 20kHz satisfies the Nyquist sampling theorem—the main modal frequencies of the GIS disconnector casing and the effective vibration components induced by mechanical faults are concentrated within 5kHz, and the sampling frequency of 20kHz provides sufficient margin for the transition band of the anti-aliasing filter.

[0040] The acquired vibration acceleration signals are preprocessed by denoising and filtering to eliminate the influence of environmental noise and power frequency interference on subsequent time-frequency analysis.

[0041] Step S2: Perform time-frequency transformation on the vibration acceleration signal to extract time-frequency features from multiple different frequency bands. S21: Sliding window splitting.

[0042] The original vibration acceleration signal x(t) is segmented using a sliding window. The window length is 1024 sampling points, and the step size is 512 sampling points. Multiple time segments are obtained through overlapping sliding. Figure 4 As shown in the sliding window segmentation section, the overlapping sliding window strategy can ensure that fault features can be effectively captured no matter where they appear in the signal.

[0043] A window length of 1024 points corresponds to a time domain span of 51.2 ms, covering approximately 5.12 fundamental frequency cycles of the 100 Hz vibration frequency of the GIS shell, which is sufficient to completely capture the entire process of mechanical impact energy attenuation induced by contact closure and structural failure. Simultaneously, this window length does not introduce excessive steady-state background noise and power frequency harmonic components, effectively preventing the transient characteristics of the fault from being diluted by background noise. A step size of 512 points achieves 50% overlap between adjacent windows, which on the one hand suppresses the edge effect of continuous wavelet transform at the window boundaries, avoiding distortion of time-frequency characteristics at the boundaries; on the other hand, as a data augmentation strategy, it enriches the diversity of training samples without changing the underlying signal characteristics.

[0044] To verify the rationality of the above sliding window parameter selection, this invention compared the diagnostic performance under various combinations of window length and step size. When the window length is set too short, its time domain span is insufficient to fully cover the decay process of the mechanical impact, resulting in truncated transient features and incomplete energy accumulation patterns in the time-frequency image. When the window length is set too long, excessive steady-state background noise and power frequency harmonic components are introduced, diluting the proportion of transient fault features and reducing the inter-class discrimination of the time-frequency image. When the step size is set too small, resulting in excessively high overlap, the training samples are highly redundant, significantly increasing computational overhead while providing limited improvement in classification accuracy. When the step size is set too large, resulting in excessively low or even no overlap, the edge effect of continuous wavelet transform at the window boundary causes distortion of time-frequency features, and features are lost when some fault impacts are located at the boundary, leading to decreased classification stability. Considering both classification accuracy and computational efficiency, this invention selects a window length of 1024 points and a step size of 512 points.

[0045] S22: Continuous wavelet transform and extraction of three-band time-frequency features.

[0046] For each signal segment extracted by the sliding window, a continuous wavelet transform is applied to extract the time-frequency features of the low-frequency, mid-frequency, and high-frequency bands, respectively. The mathematical definition of the continuous wavelet transform is as follows:

[0047] In the formula: Ψ*a,b(t) are the conjugate wavelet functions of Ψa,b(t), a is the scaling factor, which is used to control the scaling transformation of the wavelet function and corresponds to the frequency resolution of the frequency axis; b is the time shift factor, which is used to control the translation transformation of the wavelet function and corresponds to the time resolution of the time axis; Ψa,b(t) is generated by the wavelet mother function ψ(t) through translation and scaling operations.

[0048] This invention employs the Morlet wavelet as the mother wavelet function. The Morlet wavelet is a complex exponential function under a Gaussian envelope, exhibiting excellent localization characteristics in both the time and frequency domains. It lacks a scaling function and can simultaneously achieve high frequency resolution in the low-frequency band and high time resolution in the high-frequency band, overcoming the inherent limitation of the short-time Fourier transform's time-frequency resolution being constrained by a fixed window length. Furthermore, the waveform characteristics of the Morlet wavelet highly match the transient vibration acceleration signals generated by GIS disconnector operation and mechanical impact. Its oscillation decay characteristics effectively capture the time-frequency clustering characteristics of fault impacts, resulting in a generated time-frequency amplitude matrix with clear energy clustering patterns in both the time and frequency dimensions, which is beneficial for feature extraction in subsequent deep learning models.

[0049] To verify the superiority of the Morlet wavelet, this invention also compared the feature extraction performance using different wavelet basis functions. When using the Daubechies wavelet, its time-domain support range is relatively short, resulting in insufficient frequency resolution for low-frequency steady-state vibration components. While the Symlets or Coiflets wavelets offer good waveform symmetry, their sensitivity to transient shocks is lower than that of the Morlet wavelet. However, the Morlet wavelet exhibits optimal time-frequency clustering, producing the most significant differences in energy distribution among various fault states in the generated time-frequency image, resulting in the highest classification accuracy. Therefore, this invention selects the Morlet wavelet as the mother wavelet function for continuous wavelet transform.

[0050] After continuous wavelet transform, three time-frequency amplitude matrices are obtained for the low-frequency, mid-frequency, and high-frequency bands, as shown below: W L = [w L(i,j) ], W M = [w M(i,j) ], W H = [w H(i,j) ] Based on the physical vibration characteristics of different fault types of GIS disconnect switches, this invention divides the entire frequency band into three physical frequency bands: low frequency, mid frequency, and high frequency. The specific division scheme is as follows: The low-frequency band is 0-1000Hz. During normal operation, the vibration energy of the GIS disconnector is mainly concentrated around 1000Hz, with harmonic components showing a decreasing trend, a clear spectral structure, and no significant high-frequency components. Under the condition of internal foreign objects, although metal debris or wear particles generate random jumps under the influence of an electric field, their vibration energy is limited by the size and trajectory of the foreign object, making it difficult to excite sustained high-frequency components; the energy is also mainly distributed within 1000Hz. Therefore, the 0-1000Hz low-frequency band can completely cover the main energy concentration area under normal operation and internal foreign object conditions. It should be particularly noted here that... Figure 2As shown, the spectral characteristics of the normal state and the internal foreign object state are highly similar in the low-frequency band, which is the fundamental reason why existing traditional methods are difficult to distinguish between the two. This invention overcomes this problem by using subsequent RGB multi-band fusion and SE attention mechanism.

[0051] The mid-frequency band is 1000-4000Hz. When the shielding becomes loose, the three-phase conductors experience uneven stress, generating higher harmonics under the action of an alternating electric field. The energy is concentrated around 2000Hz, and the overall spectrum shifts towards the mid-to-high frequency range. Meanwhile, the vibration spectrum of mechanical defects in GIS is usually within 4kHz. Therefore, the mid-frequency band of 1000-4000Hz is used to capture the harmonic and resonance characteristics induced by faults such as loose shielding.

[0052] The high-frequency band is 4000-5000Hz. When there is poor contact between the contacts, the contact resistance increases and the contact pressure decreases, resulting in unstable electrical contact between the moving and stationary contacts. This triggers partial discharge and micro-arcs, generating high-frequency impact components with energy extending up to 5000Hz. Therefore, the 4000-5000Hz high-frequency band is used to capture the high-frequency impact and partial discharge characteristics caused by poor contact. The boundary values ​​of the three frequency bands are determined based on the physical vibration characteristics of different fault types of GIS disconnect switches, and a reasonable frequency margin (e.g., a 1000Hz transition margin between the three frequency bands) is reserved to accommodate individual frequency variations under different operating conditions.

[0053] To verify the effectiveness of the above frequency band division scheme, this invention further compared the diagnostic accuracy under various different frequency band boundary divisions. When the upper limit of the low-frequency band is set too low, the main energy of the normal operating state is truncated, resulting in the loss of effective features; when the upper limit of the mid-frequency band is set too high or the distinction between high and low frequency bands is too low, the spectral aliasing of loose shielding and poor contact increases, making it difficult to distinguish between the two types of faults; when the starting frequency of the high-frequency band is set too low, background noise is introduced into the high-frequency channel, interfering with the extraction of the impact features of poor contact. The comparative results show that when using the frequency band division scheme of this invention, the energy concentration areas of each fault state are accurately separated, and the classification accuracy is the highest. The frequency band division scheme of this invention can effectively separate the feature frequency bands of various faults, providing a physically clear input basis for the subsequent adaptive weighting of the attention mechanism.

[0054] Step S3: Map the time-frequency features of multiple frequency bands to different color channels of an RGB color image to generate an RGB time-frequency feature image that fuses information from multiple frequency bands. S31: Normalization process.

[0055] The three time-frequency amplitude matrices extracted in step S22—low-frequency, mid-frequency, and high-frequency—are normalized, mapping the values ​​of each element to the [0,1] interval. The normalization formula is as follows:

[0056] In the formula, W(i,j) represents the element value of the time-frequency amplitude matrix at coordinate position (i,j), min(W) and max(W) represent the minimum and maximum values ​​among all elements of the matrix, respectively, and N(i,j) represents the normalized value. The normalization process eliminates the data scale inconsistency problem caused by the difference in amplitude magnitude between different frequency bands.

[0057] S32: RGB three-channel mapping.

[0058] The normalized low-frequency time-frequency amplitude matrix is ​​mapped to the R (red) channel of an RGB color image, the mid-frequency time-frequency amplitude matrix is ​​mapped to the G (green) channel, and the high-frequency time-frequency amplitude matrix is ​​mapped to the B (blue) channel.

[0059] This mapping method encodes the time-frequency features of the low, medium, and high frequency bands into three independent channels of the color image, allowing information from different frequency bands to remain independently identifiable while forming a collaborative and complementary joint representation within the same image frame. Figure 1 Compared with the existing technology that uses only a single time-frequency transformation method to generate a single-dimensional time-frequency representation, the RGB three-channel image of the present invention can simultaneously present the time-frequency distribution characteristics of the low, medium and high frequency bands, making full use of the joint distribution characteristics of multi-dimensional information in the vibration acceleration signal, and effectively making up for the defect that a single transformation method cannot take into account both the global trend and local abrupt change information of the signal.

[0060] S33: Image scaling and generation.

[0061] The three-channel images are uniformly scaled to 64×64 pixels using bilinear interpolation, generating a 64×64×3 RGB three-channel time-frequency feature image IRGB = [IR, IG, IB]. The complete processing flow from S21 to S33 is as follows: Figure 4 As shown, Figure 4 It fully demonstrates the entire process from the original vibration acceleration signal through sliding window segmentation, continuous wavelet transform three-band extraction, normalization processing to RGB three-channel mapping and image generation.

[0062] S34: Dataset partitioning.

[0063] All generated RGB time-frequency feature image samples were randomly divided into training and test sets in an 8:2 ratio. The training set was used for learning and optimizing model parameters, while the test set was used to evaluate the model's generalization performance and diagnostic accuracy.

[0064] Step S4: Input the RGB time-frequency feature image into a convolutional neural network diagnostic model embedded with an attention mechanism. Adaptively weight the feature channels through the attention mechanism, and output the fault diagnosis result of the GIS disconnect switch. The overall structure of the CNN-SE attention fusion diagnostic model used is as follows: Figure 5 As shown. With Figure 3 Unlike existing 1D-CNN methods that extract features solely through stacked one-dimensional convolutional and pooling layers and lack frequency band adaptive weighting capabilities, the diagnostic model of this invention embeds an SE attention module between convolutional layers, achieving adaptive focusing on fault-sensitive frequency bands. This model employs a two-layer convolutional structure, consisting of a first convolutional layer, an SE attention module, a first pooling layer, a second convolutional layer, a second pooling layer, a global average pooling layer, a Dropout layer, and a fully connected output layer connected sequentially. The following is a combination of... Figure 5 Each component is described in detail.

[0065] S41: First convolutional layer Conv1.

[0066] The 64×64×3 RGB time-frequency feature image generated in step S3 is input into the first convolutional layer Conv1. The first convolutional layer contains 32 3×3 convolutional kernels, using Same padding with a stride of 1, and outputs a feature map of size 64×64×32. Two-dimensional convolution is defined as the convolutional kernel performing a sliding inner product operation on the two-dimensional feature map to extract local features in both spatial and channel dimensions. The calculation formula is as follows:

[0067] Where f represents the activation function; W represents the convolution kernel weight matrix; represents the convolution operation; I represents the input feature map; B represents the bias term.

[0068] S42: SE Attention Module.

[0069] The feature map output from the first convolutional layer is input into the SE attention module. For example... Figure 5 As shown, the SE attention module is embedded after the first convolutional layer and before the first max pooling layer, marked with a dashed box in the model structure diagram. The reason for choosing this embedding position is that the shallow feature map has high spatial resolution (64×64×32), preserving rich local time-frequency details, including the start and end times of the impact signal and the frequency band location of energy accumulation, among other key discriminative features. Embedding the SE module at this position allows for recalibration of each feature channel in the early stages of feature extraction, ensuring that subsequent pooling operations and deep convolutions are based on the enhanced fault-sensitive features, preventing the dilution of fault information during deep propagation. The SE attention module does not change the spatial size of the feature map; it only adaptively weights the feature channels.

[0070] To verify the rationality of the above embedding positions, this invention compares the diagnostic performance of the SE module at different embedding positions. When the SE module is embedded after the first pooling layer and before the second convolutional layer, the spatial resolution of the feature map is halved, some local details are lost, and the channel response of the fault-sensitive frequency band is not prominent enough. When the SE module is embedded after the second convolutional layer and before global average pooling, although the deep feature semantic information is rich, the spatial resolution is low, the differences between channels tend to stabilize, and the discrimination effect of the attention mechanism is limited. When the SE module is embedded after both convolutional layers, the number of parameters and computational cost increases significantly, while the improvement in classification accuracy is limited. The comparative results show that embedding the SE module after the first convolutional layer and before the first max pooling layer in this invention can achieve the best classification performance.

[0071] like Figure 5 As shown in the SE attention module section, the processing flow of the SE attention module is divided into three stages: squeezing, excitation, and scaling. (1) Compression: The spatial information of each feature channel is compressed into a scalar through global average pooling, which serves as the global descriptor for that channel, as shown in the following formula:

[0072] (2) Incentive: The nonlinear interaction relationship between channels is learned through the bottleneck structure composed of two fully connected layers, and the importance weight of each channel is generated as shown in the following formula:

[0073] Where W1 is the weight matrix of the first fully connected layer, used to compress and reduce the channel dimension; W2 is the weight matrix of the second fully connected layer, used to restore the original channel dimension; the output sc represents the importance weight of the c-th channel; σ represents the ReLU activation function, which introduces non-linear transformation capability into the model; δ represents the Sigmoid activation function, which constrains the output weights to the interval [0,1].

[0074] (3) Recalibration: The learned channel importance weights are multiplied with the original feature map channel by channel along the channel dimension to enhance the features of fault-sensitive frequency bands and suppress the features of secondary frequency bands, as shown in the following formula:

[0075] Through the three stages described above, the SE attention module enables the model to automatically identify and enhance fault-related sensitive frequency band features during training, while suppressing secondary frequency band information unrelated to faults. This makes the method of this invention applicable to situations such as... Figure 2When encountering fault types such as internal foreign objects and normal states with high spectral similarity, it can more effectively focus on the key frequency band characteristics that distinguish the two, overcoming the limitations of methods like... Figure 1 The existing traditional methods shown are as follows: Figure 3 The 1D-CNN methods shown all lack the ability to focus on specific frequency bands.

[0076] S43: First pooling layer.

[0077] The weighted feature map output by the SE attention module is input into the first pooling layer. Max pooling uses a 2×2 window with a stride of 2, reducing the spatial size of the feature map to half its original size. This effectively controls the computational complexity of the model while preserving key spatial information, as shown in the following equation:

[0078] S44: Second convolutional layer Conv2.

[0079] The feature map output from the first pooling layer is input into the second convolutional layer, Conv2. The second convolutional layer contains 64 3×3 convolutional kernels, uses Same padding, has a stride of 1, and outputs a feature map with a size of 32×32×64.

[0080] S45: Second pooling layer.

[0081] The feature map output from the second convolutional layer is input into the second pooling layer, and max pooling operation with a 2×2 window and a stride of 2 is also used.

[0082] S46: Global average pooling layer.

[0083] The pooled feature maps are input into a global average pooling layer, which performs a global average on the spatial dimensions of each feature map, compressing each feature map into a scalar value and outputting a 64-dimensional feature vector, as shown in the following formula:

[0084] S47: Dropout layer.

[0085] A Dropout layer is applied after the global average pooling layer, with a dropout rate of 0.4, meaning that some neurons are randomly deactivated with a probability of 0.4 during training. The Dropout regularization strategy effectively suppresses overfitting by forcing the network to learn more robust feature representations even with missing neurons.

[0086] S48: Fully connected output layer.

[0087] Finally, a fully connected layer is connected, and the classification probability distribution is output through the Softmax normalized exponential function. The category corresponding to the highest probability is taken as the mechanical state recognition result of the GIS disconnect switch, as shown in the following formula:

[0088] The working principle of this invention can be combined with Figure 8 The overall process can be summarized as follows. For example... Figure 8 As shown, the energy distribution patterns of the GIS disconnect switch's casing vibration acceleration signal differ in the low-frequency, mid-frequency, and high-frequency bands under different mechanical conditions (normal, poor contact, loose shielding, foreign object retention, etc.). Vibration acceleration signals from the GIS disconnect switch casing are acquired using a laser Doppler vibrometer or accelerometer. After preprocessing such as denoising and filtering, the original vibration acceleration signal is segmented using a sliding window. For each signal segment, a Morlet wavelet continuous wavelet transform is applied to decompose the one-dimensional vibration acceleration signal into three independent time-frequency amplitude matrices: low-frequency (0-1000Hz), mid-frequency (1000-4000Hz), and high-frequency (4000-5000Hz). These matrices are then encoded into the R, G, and B channels of an RGB color image, thus converting the multi-frequency information into a unified 64×64×3 two-dimensional image representation. Based on this, as... Figure 5 As shown, the CNN-SE model extracts local spatial-channel features of an image through two convolutional layers. An SE attention module embedded after the first convolutional layer adaptively weights the feature channels, automatically enhancing sensitive frequency bands relevant to fault detection and suppressing irrelevant or secondary frequency bands. Finally, a fully connected layer and a Softmax classifier output the diagnostic results. Figure 8 As can be seen from the overall process, this invention adopts an end-to-end diagnostic architecture, which automates the entire process from the acquisition of raw vibration acceleration signals, sliding window segmentation, continuous wavelet transform, RGB image generation to CNN-SE model diagnostic output, without the need for manual design of feature parameters.

[0089] To verify the effectiveness of the method of this invention, vibration acceleration signal acquisition experiments were conducted on a 220kV GIS disconnector experimental platform under four operating conditions: internal foreign objects, poor contact of contacts, normal operation, and loose shielding. More than 600 sets of vibration acceleration signals from closing operations were collected for each condition. A sliding window with a window length of 1024 points and a step size of 512 points was used to segment the vibration acceleration signal of each closing operation, dividing each operation into 8 overlapping local sample segments. All samples were randomly divided into training and test sets at an 8:2 ratio, with all segments from the same operation belonging to the same set, ultimately forming a standardized dataset of 15840 samples in the training set and 3960 samples in the test set.

[0090] To comprehensively evaluate the classification performance of the method of this invention, it is compared with five existing methods: Support Vector Machine (SVM), Deep Belief Network (DBN), Traditional One-Dimensional Convolutional Neural Network (1D-CNN), Visual Transformer (ViT), and ResNet-18. All comparison methods use the same training and test sets and have undergone hyperparameter optimization to ensure the fairness of the comparison.

[0091] like Figure 6 As shown, Figure 6 The loss function curves and accuracy curves of the method of this invention and five comparative methods during the training process are presented. From Figure 6 As can be seen from the loss function curve in (a), the loss value of the method of the present invention decreases rapidly in the early stage of training, exhibits the smallest fluctuation amplitude and the best stability throughout the entire training process, and finally converges to the lowest loss value; while DBN and traditional CNN show a certain degree of oscillation in the middle of training, and ViT also shows slight fluctuations, indicating that their training stability is weaker than that of the method of the present invention. Figure 6 As shown in the accuracy curve (b), the method of this invention exhibits the highest initial accuracy in the early stages of training. The accuracy increases rapidly with the number of iterations, stabilizes after a few iterations, and finally converges to an accuracy of over 94%. In contrast, the accuracy curves of traditional CNN, DBN, and SVM methods show a more gradual increase, with significantly slower convergence speeds, and their final convergence accuracy is lower than that of the method of this invention. Although ViT and ResNet-18 have faster convergence speeds than traditional shallow models, their final accuracy still lags behind that of the method of this invention. The above comparison results demonstrate that the introduction of the SE attention mechanism not only effectively improves the model's classification accuracy through adaptive frequency band weighting but also accelerates model convergence and enhances training stability by suppressing interference from irrelevant feature channels.

[0092] like Figure 7 As shown, Figure 7 This is the confusion matrix of the method of the present invention under four operating states. The method of the present invention can achieve an accuracy, precision, recall, and F1-score of over 93% for all four operating states, with balanced and stable performance across all metrics. Taking the internal foreign object state as an example, out of 985 internal foreign object test samples, 925 were correctly predicted as internal foreign objects, 25 were misclassified as poor contact, 18 were misclassified as normal operation, and 17 were misclassified as loose shielding. Based on this, the true number of internal foreign object states is TP=925, the false negative number is FN=60, the false positive number is FP=61, the classification accuracy is 93.91%, the precision is 93.81%, the recall is 93.91%, and the F1-score is 93.86%. From the misclassification distribution of the confusion matrix, the number of misclassifications is highest between the internal foreign object state and the normal operation state. This is because their spectral characteristics in the low-frequency band are highly similar (e.g., ...). Figure 2 As shown in the figure, the energy is mainly concentrated around 1000Hz, making it the most challenging task in classification. However, the proportion of major misclassifications for each state is at a low level, indicating that the method of this invention has excellent robustness and classification stability in distinguishing spectrally similar faults.

[0093] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A fault diagnosis method for GIS disconnect switches based on RGB time-frequency representation and attention fusion, characterized in that, Includes the following steps: S1: Collect vibration acceleration signals of the GIS disconnect switch housing; S2: Perform time-frequency transformation on the vibration acceleration signal to extract time-frequency features of multiple different frequency bands; S3: Map the time-frequency features of the multiple different frequency bands to different color channels of an RGB color image to generate an RGB time-frequency feature image that fuses multi-frequency band information; S4: Input the RGB time-frequency feature image into a convolutional neural network diagnostic model with an attention mechanism, and adaptively weight the feature channels through the attention mechanism to output the fault diagnosis result of the GIS disconnect switch.

2. The method according to claim 1, characterized in that, In step S1, the vibration acceleration signal of the GIS disconnect switch housing is acquired by a laser Doppler vibration meter or an accelerometer, with a sampling frequency of 20kHz; the acquired vibration acceleration signal is then preprocessed by denoising and filtering.

3. The method according to claim 1, characterized in that, In step S2, the vibration acceleration signal is segmented using a sliding window to obtain multiple overlapping time segments, and then the time-frequency features are extracted from each time segment. The sliding window has a window length of 1024 sampling points, a step size of 512 sampling points, and an overlap rate of 50% between adjacent windows.

4. The method according to claim 1, characterized in that, In step S2, the time-frequency transformation is a continuous wavelet transform, and the wavelet mother function used in the continuous wavelet transform is the Morlet wavelet.

5. The method according to claim 4, characterized in that, The multiple different frequency bands include low frequency band, mid frequency band and high frequency band. The frequency range of the low frequency band is 0-1000Hz, the frequency range of the mid frequency band is 1000-4000Hz, and the frequency range of the high frequency band is 4000-5000Hz. After continuous wavelet transform, the time-frequency amplitude matrix corresponding to each frequency band is obtained.

6. The method according to claim 5, characterized in that, In step S3, the time-frequency amplitude matrices corresponding to the low-frequency band, mid-frequency band, and high-frequency band are normalized respectively, and their values ​​are mapped to the [0,1] interval. Then, the normalized low-frequency band time-frequency amplitude matrix is ​​mapped to the R channel of the RGB color image, the mid-frequency band time-frequency amplitude matrix is ​​mapped to the G channel, and the high-frequency band time-frequency amplitude matrix is ​​mapped to the B channel. The matrices are then merged to generate the RGB time-frequency feature image. The RGB time-frequency feature image is uniformly scaled to 64×64 pixels using bilinear interpolation.

7. The method according to claim 1, characterized in that, In step S4, the convolutional neural network diagnostic model with embedded attention mechanism includes a first convolutional layer, an SE attention module, a first pooling layer, a second convolutional layer, a second pooling layer, a global average pooling layer, a Dropout layer, and a fully connected output layer connected in sequence; the dropout rate of the Dropout layer is 0.

4.

8. The method according to claim 7, characterized in that, The first convolutional layer contains 32 3×3 convolutional kernels, uses Same padding, has a stride of 1, and outputs a feature map size of 64×64×32. The second convolutional layer contains 64 3×3 convolutional kernels, uses Same padding, has a stride of 1, and outputs a feature map size of 32×32×64. Both the first and second pooling layers use max pooling with a 2×2 window and a stride of 2.

9. The method according to claim 7, characterized in that, The processing flow of the SE attention module is as follows: First, the spatial information of each feature channel is compressed into a global descriptor through global average pooling; then, the importance weights of each channel are learned through two fully connected layers, where the first fully connected layer compresses and reduces the dimensionality of the channel, and the second fully connected layer restores the original channel dimension; finally, the importance weights are multiplied with the original feature map channel by channel to enhance the fault-sensitive frequency band and suppress the secondary frequency band.

10. The method according to claim 1, characterized in that, During the model training phase, the generated RGB time-frequency feature images are randomly divided into a training set and a test set in an 8:2 ratio. The training set is used for model parameter learning, and the test set is used to evaluate model performance.

Images