A method, system and program for diagnosing and evaluating mechanical defects in GIS / GIL equipment

Abstract

The application discloses a kind of related to GIS / GIL equipment mechanical defect diagnosis method, evaluation method and system, it is related to GIS / GIL equipment mechanical defect intelligent analysis technical field, by carrying out adjacent modal noise suppression GIS / GIL equipment mechanical vibration signal visual time-frequency analysis, then establish GIS / GIL vibration feature self-extraction and the improved light-weight convolutional neural network of multiple mark output, finally carry out load adaptive matching GIS / GIL equipment mechanical defect diagnosis and severity evaluation.The diagnosis, evaluation method is the defect diagnosis technology proposed to the actual field engineering, effectively overcome the defect feature information difficult to extract, low diagnostic accuracy and other problems brought by field environmental noise interference, dynamic running current, can be directly applied to the defect identification and state evaluation of GIS / GIL equipment mechanical defect in substation field, make reasonable operation and maintenance plan, discover and check internal mechanical defect early, with important engineering application value.

Description

Technical Field

[0001] This invention relates to the field of intelligent analysis technology for mechanical defects in GIS / GIL equipment, and more specifically, to a method, evaluation method, and system for diagnosing mechanical defects in GIS / GIL equipment. Background Technology

[0002] Gas-insulated switchgear (GIS / GIL) and gas-insulated transmission line (GIL) bear the heavy responsibility of load control, safety protection, and power transmission in power systems. The widespread adoption of GIS / GIL has effectively alleviated the power supply pressure caused by urban land scarcity, surging electricity loads, and high reliability requirements. However, due to the fully enclosed structure of GIS / GIL equipment, its internal status information is difficult to detect, and equipment failures can severely impact the normal operation of the system. Statistics show that GIS / GIL equipment failures are mainly related to insulation and mechanical defects generated during production, transportation, and operation. Mechanical defects are also a significant factor leading to insulation degradation; long-term abnormal noise and vibration can cause localized overheating, gas leakage, and even flashover breakdown. Therefore, effectively diagnosing mechanical defects in operating GIS / GIL equipment is crucial for ensuring its safe and stable operation.

[0003] Mechanical defects in GIS / GIL equipment are characterized by random loads and complex operating conditions. On the one hand, the operating current and internal defect status change in real time, and existing technologies do not consider the impact of load current on the diagnostic effect of mechanical defects in GIS / GIL equipment. On the other hand, environmental noise during the diagnostic process is not effectively filtered out, resulting in an inaccurate final diagnostic effect. Summary of the Invention

[0004] The first objective of this invention is to provide a method and system for diagnosing mechanical defects in GIS / GIL equipment. This method and system analyze and compare vibration signals and background signals in the operating environment of GIS / GIL equipment, and effectively filter out environmental noise interference by using adjacent mode noise suppression, thereby ensuring the accuracy of subsequent diagnostic results.

[0005] The second objective of this invention is to provide a method for assessing mechanical defects in GIS / GIL equipment. This method utilizes the diagnostic method described above and diagnoses the target based on defect information from samples in different intervals divided by load current, so as to simultaneously determine the type and severity of mechanical defects in GIS / GIL equipment under different operating currents.

[0006] The embodiments of the present invention are implemented as follows:

[0007] In a first aspect, a method for diagnosing mechanical defects in GIS / GIL equipment includes the following steps:

[0008] The mechanical vibration signal and background noise signal of the GIS / GIL equipment are acquired. The mechanical vibration signal and background noise signal of the GIS / GIL equipment are decomposed into multiple components to obtain the m modal components SVMF of the mechanical vibration signal and the n modal components NVMF of the background noise signal, where m and n are positive integers.

[0009] Solve for the barycenter frequency S of the modal components SVMF and NVMF respectively. f ={S f1 ,S f2 ,…,S fm} and N f ={N f1 N f2 ,…,N fn}, determine the NVMF that is closest to the centroid frequency of each modal component SVMF and perform matching;

[0010] For the pairwise matched modal components SVMF and NVMF, the neighborhood Γ is calculated with their respective barycentric frequencies as the center. Si ∈(S fi -δ,S fi +δ) and Γ Ni ∈(S fi -δ,S fi The amplitude and kurtosis of the frequency domain vibration signal within +δ) are used as threshold filtering criteria. Interference noise modal signals are filtered out by threshold comparison to obtain the target vibration signal. The modal component index i = 1, 2, 3, ..., K, where K is the total number of modal components; δ represents the width of the frequency range.

[0011] The target vibration signal is identified using a convolutional neural network model, and the identification results are obtained.

[0012] In an optional implementation, the GIS / GIL equipment mechanical vibration signal and background noise signal are subjected to multiple decomposition using a joint variational multimodal decomposition algorithm, and are calculated by introducing the Lagrange multiplication operator λ(t) and the quadratic penalty factor τ, as shown in expression (1) below:

[0013]

[0014] In equation (1), u k Represents a modal signal; ω k λ represents the center frequency; λ represents the Lagrange multiplier operator. The gradient operation is represented by δ(t), which is the unit impulse function; * represents convolution; t represents time; K represents the total number of modal components; k represents the modal component index; j is a complex number symbol; and f(t) represents the vibration time-domain sequence.

[0015] In an optional implementation, the frequency domain mode u is also included. k (ω) and the corresponding center frequency ω k The update steps, and their update calculation expressions are as shown in equations (2) and (3) respectively:

[0016]

[0017]

[0018] In an optional implementation, after filtering out interfering noise mode signals using threshold comparison, the following steps are further included:

[0019] The filtered modal signals are superimposed, and wavelet transform is performed on the superimposed signal to construct a three-dimensional time-frequency spectrum signal. The target vibration signal is obtained based on the time-frequency spectrum signal.

[0020] In an optional implementation, the process of superimposing the filtered modal signals and performing wavelet transform on the superimposed signals to construct a three-dimensional time-frequency spectrum signal includes the following steps:

[0021] Define a signal x(t), x(t)∈L 2 (R), based on x(t), determine the wavelet transform WT x (a,b), the expression is given by equation (4):

[0022]

[0023] In equation (4), t is time, a and b are the scale factor and translation parameter, respectively, and a is greater than 0; ψ a,b ψ(t) is a family of wavelet basis functions generated by shifting and scaling the mother wavelet function ψ(t).

[0024] In an alternative implementation, ψ a,b The relationship between (t) and the mother wavelet function ψ(t) is expressed as in equation (5):

[0025]

[0026] The constraint condition for ψ(t) is:

[0027]

[0028] In an optional implementation, the algorithmic architecture of the convolutional neural network model is expressed as equation (7):

[0029] Z = F(X|Θ) = f L (...f2(f1(X|θ1)|θ2|θ l (7)

[0030] In equation (7), F(X|Θ) represents a multi-layer nonlinear mapping model; f(.|θ) l ) is the mapping function; Θ = {θ} l ,θ2…θ l} represents the network parameter set.

[0031] In an optional implementation, the convolutional neural network model employs a Fire network stack module to extract sample features. Each Fire module includes a compression layer, an expansion layer, and a merging layer. The compression layer uses a 1×1 convolutional filter to apply the input feature matrix dataset C from the previous layer. l Convolution is performed; the extended layer uses 1×1 and 3×3 convolutional filters to perform grouped convolution to form connections in sparse channel domains; the merging layer fuses information between channels.

[0032] Secondly, a method for assessing mechanical defects in GIS / GIL equipment, applying the aforementioned method for diagnosing mechanical defects in GIS / GIL equipment, includes the following steps:

[0033] Based on load rate LO R The training sample set will be divided into three evaluation intervals: a first evaluation interval, a second evaluation interval, and a third evaluation interval. The first evaluation interval represents the LO (Local Order) interval. R ≤33%, the second evaluation interval represents 33%. <LO R ≤66%, the third evaluation interval represents LO R >66%; Load factor (LO) R For the equipment operating current I OM With the rated current I of the equipment RM The ratio;

[0034] The convolutional neural network neural model is trained using a training sample set.

[0035] Thirdly, a GIS / GIL equipment mechanical defect diagnosis system includes:

[0036] The decomposition module is used to acquire the mechanical vibration signal and background noise signal of the GIS / GIL equipment, and to perform multiple decompositions on the mechanical vibration signal and background noise signal of the GIS / GIL equipment to acquire the m modal components SVMF of the mechanical vibration signal and the n modal components NVMF of the background noise signal, where m and n are both positive integers.

[0037] The calculation module is used to solve for the barycenter frequency S of the modal components SVMF and NVMF, respectively. f ={S f1 ,S f2 ,…,S fm} and N f ={N f1 N f2 ,…,N fn}, determine the NVMF that is closest to the centroid frequency of each modal component SVMF and perform matching;

[0038] The judgment module is used to calculate the neighborhood Γ of each pairwise matched modal component SVMF and modal component NVMF, centered on their respective barycentric frequencies. Si ∈(S fi -δ,S fi +δ) and Γ Ni ∈(S fi -δ,S fi The amplitude and kurtosis of the frequency domain vibration signal within +δ) are used as threshold filtering criteria. Interference noise modal signals are filtered out by threshold comparison to obtain the target vibration signal. The modal component index i = 1, 2, 3, ..., K, where K is the total number of modal components; δ represents the width of the frequency range.

[0039] The identification module is used to identify the target vibration signal using a convolutional neural network model and obtain the identification result.

[0040] The beneficial effects of the embodiments of the present invention are:

[0041] The mechanical defect diagnosis method and system for GIS / GIL equipment provided in this invention extracts vibration signals and background noise signals from the working environment of the GIS / GIL equipment, and then performs multiple decompositions to obtain multiple modal components of the two signals. Based on the similarity of the center-of-gravity frequencies, pairwise matching is performed. The matched modal components are then centered on their center-of-gravity frequencies, and the vibration signal amplitude and kurtosis are calculated as judgment thresholds. After comparison, the modal signals of interference noise can be effectively filtered out, thereby obtaining a clearer target vibration signal, which can be used as input parameters for subsequent neural network models to ensure the accuracy of the output results.

[0042] The mechanical defect assessment method for GIS / GIL equipment provided in this embodiment of the invention utilizes the above-mentioned mechanical defect diagnosis method for GIS / GIL equipment and divides the sample set based on three intervals to train the selected algorithm model, so as to output a result that can evaluate the load current. That is, it considers the defect information of samples in different intervals of load current division to diagnose mechanical defects in GIS / GIL equipment and outputs the assessment results of defect type and severity. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 A flowchart illustrating the main steps of the diagnostic method provided in this embodiment of the invention;

[0045] Figure 2 A flowchart illustrating the main steps of a diagnostic method provided in another embodiment of the present invention;

[0046] Figure 3 A flowchart illustrating the main steps of the evaluation method provided in the embodiments of the present invention;

[0047] Figure 4 Schematic diagrams of normal and typical mechanical defects under different load currents provided in embodiments of the present invention;

[0048] Figure 5 A time-frequency comparison analysis diagram of two signals before and after filtering: normal GIS / GIL equipment signal and poor contact mechanical defect signal. This is provided for an embodiment of the present invention.

[0049] Figure 6 A schematic diagram illustrating the results of GIS / GIL mechanical defect diagnosis and severity assessment under different load currents provided in an embodiment of the present invention;

[0050] Figure 7 An exemplary module diagram of a diagnostic system provided in an embodiment of the present invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0052] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0053] It should be understood that the terms "system," "device," and / or "module" used in this invention are methods for distinguishing different components, elements, parts, sections, or assemblies at different levels. However, if other terms can achieve the same purpose, they may be replaced by other expressions.

[0054] As indicated in this invention and the claims, unless the context clearly indicates otherwise, the words "a," "an," "an," and / or "the" do not specifically refer to the singular and may also include the plural. Generally, the terms "comprising" and "including" only indicate the inclusion of explicitly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements.

[0055] Flowcharts are used in this invention to illustrate the operations performed by the system according to embodiments of this application. It should be understood that the preceding or following operations are not necessarily performed precisely in sequence. Instead, the steps can be processed in reverse order or simultaneously. Furthermore, other operations can be added to these processes, or one or more steps can be removed from them.

[0056] Example 1

[0057] In the field of mechanical defect identification technology for GIS / GIL equipment, previous methods have focused on feature analysis of vibration signals. For example, cameras are used to acquire images of the contact status of disconnector switches and bolt loosening. Furthermore, a support vector machine optimized by the "sparrow foraging" algorithm is used to diagnose mechanical defects. However, there are problems with the accuracy of image acquisition and recognition. Therefore, we use the easily obtainable vibration time-frequency signal as the starting point, while effectively filtering out environmental noise signals to achieve clear vibration signal acquisition.

[0058] Please see Figure 1 The mechanical defect diagnosis method for GIS / GIL equipment provided in this embodiment includes the following steps:

[0059] S100: Acquire the mechanical vibration signal and background noise signal of the GIS / GIL equipment, and perform multimodal decomposition on the mechanical vibration signal and the background noise signal to obtain the m modal components (SVMF) of the mechanical vibration signal and the n modal components (NVMF) of the background noise signal, where m and n are both positive integers; this step means performing multimodal decomposition on the vibration signal and the background noise signal, using the modal components of smaller dimensions as the basis for comparison calculation.

[0060] S200: Solve for the centroid frequencies S of the modal components SVMF and NVMF respectively. f ={S f1,S f2 ,…,S fm} and N f ={N f1 N f2 ,…,N fn The process involves determining and matching the modal component NVMF whose centroid frequency is closest to that of each modal component SVMF. This step involves determining the data centroid of each set of modal components SVMF and NVMF, such as the frequency of data occurrence or the lumped parameters of the data, and then pairing the modal components SVMF and NVMF whose centroid frequencies are closest. The signal that can be reconstructed from the paired modal components SVMF and NVMF is the mechanical vibration signal of the GIS / GIL equipment or a signal that is consistent with or highly similar to the mechanical vibration signal of the GIS / GIL equipment in the environment.

[0061] S300: For each pairwise matched modal component SVMF and modal component NVMF, the neighborhood Γ is calculated with their respective centroid frequencies as the center. Si ∈(S fi -δ,S fi +δ) and Γ Ni ∈(S fi -δ,S fi The amplitude and kurtosis of the frequency domain vibration signal within the +δ) range are used as threshold filtering criteria. Interference noise modal signals are filtered out using threshold comparison to obtain the target vibration signal. Here, the modal component indices i = 1, 2, 3, ..., K, where K is the total number of modal components; δ represents the width of the frequency range. This step involves using paired modal components SVMF and NVMF as a reference, and constructing a time-frequency spectrum of the vibration signal based on their centroid frequencies. The amplitude and kurtosis within the time-frequency spectrum are then used as the basis for threshold judgment and filtering. Modal signals outside the threshold (reconstructed from the modal components) are filtered out, and the final obtained vibration signal is the filtered high-precision target vibration signal. Then, step S500 is performed: the target vibration signal is identified using a convolutional neural network model to obtain the identification results, thereby providing data output for subsequent mechanical defect diagnosis.

[0062] Because GIS / GIL mechanical vibration signals exhibit both time-domain fluctuations and frequency-domain complexity, time-frequency spectra can better reflect the nonlinear vibration characteristics of internal mechanical defects in GIS / GIL equipment. Furthermore, GIS / GIL equipment operating in the field is subject to environmental vibration noise from equipment such as transformers and reactors, birdsong, and vehicle traffic. These noise components are random, and effective noise suppression is crucial for extracting feature information. Through the above technical solutions, a GIS / GIL equipment mechanical vibration time-frequency signal analysis technique with adjacent mode noise suppression was adopted, effectively filtering out environmental noise interference and achieving the goal of analyzing GIS / GIL equipment vibration signals from a time-frequency domain perspective.

[0063] In this embodiment, the mechanical vibration signal of the GIS / GIL equipment and the background noise signal are jointly decomposed using a variational mode decomposition (VMD) algorithm. The VMD method is a generalization of the classical Wiener filter across multiple adaptive frequency bands and is an effective means of obtaining different vibration information from the vibration signal. The VMD algorithm can non-recursively decompose the input signal into a series of orthogonal intrinsic mode sequences (IMFs) {u}. k The spectrum exhibits a certain degree of sparsity, with each modal signal u... k All closely revolve around its center frequency ω k The bandwidth is calculated using Gaussian smoothness. The VMD algorithm is essentially a constrained variational problem, which minimizes the sum of estimated bandwidths for each mode. By introducing the Lagrange multiplication operator λ(t) and the quadratic penalty factor τ, it is transformed into an unconstrained variational problem, as shown in Equation (1):

[0064]

[0065] In equation (1), u k Represents a modal signal; ω k λ represents the center frequency; λ represents the Lagrange multiplier operator. The gradient operation is represented by δ(t), which is the unit impulse function; * represents convolution; t represents time; K represents the total number of modal components; k represents the modal component index; j is a complex number symbol; and f(t) represents the vibration time-domain sequence.

[0066] Equation (1) can be applied to u using the alternating direction multiplier method. k ω k The signal is iteratively updated with λ to decompose it into K pre-defined optimal modal components using an optimal variational model. That is, this embodiment also includes updating the frequency domain mode u. k (ω) and the corresponding center frequency ω k The update steps, and their update calculation expressions are as shown in equations (2) and (3) respectively:

[0067]

[0068]

[0069] Based on the above solutions, please refer to Figure 2 After filtering out interfering noise modal signals using threshold comparison, the method further includes the following step S400: superimposing the filtered modal signals, performing wavelet transform on the superimposed signal to construct a three-dimensional time-frequency spectrum signal, and obtaining the target vibration signal based on the time-frequency spectrum signal. The wavelet transform algorithm, based on a time-frequency analysis window that changes with frequency, possesses high time resolution at high frequencies and high frequency resolution at low frequencies.

[0070] In this embodiment, the step of superimposing the filtered modal signals and performing wavelet transform on the superimposed signal to construct a three-dimensional time-frequency spectrum signal includes the following steps: determining a signal x(t), x(t)∈L 2 (R), based on the x(t), determine the wavelet transform WT x (a,b), the expression is given by equation (4):

[0071]

[0072] In equation (4), t is time, a and b are the scale factor and translation parameter, respectively, and a is greater than 0; ψ a,b (t) is a family of wavelet basis functions generated by shifting and scaling the mother wavelet function ψ(t). In this embodiment, for example, the relation adopts equation (5):

[0073]

[0074] The constraint condition for ψ(t) is:

[0075]

[0076] The above technical solution uses Morlet wavelets, which have high localization and analytical properties in both the time and frequency domains, to perform time-frequency analysis on the signal. The square of the amplitude of the wavelet transform coefficients is the wavelet time spectrum.

[0077] Deep convolutional neural network algorithms have powerful feature extraction and pattern recognition capabilities, and are widely used in image classification, image segmentation, object detection, and other fields. Lightweight convolutional neural network algorithms such as SqueezeNet and MobileNet have the advantages of fewer parameters and high distributed training efficiency, but can only output in a single mode. In this embodiment, the architecture of the traditional lightweight SqueezeNet algorithm is improved by using the random forest algorithm and a dual-channel parallel structure. The calculation formula of the SqueezeNet algorithm is as follows, that is, the algorithm architecture expression of the convolutional neural network neural model is as shown in equation (7):

[0078] Z = F(X|Θ) = f L (...f2(f1(X|θ1)|θ2)|θ l (7)

[0079] In equation (7), F(X|Θ) represents a multi-layer nonlinear mapping model; f(.|θ) l ) is the mapping function; Θ = {θ} l ,θ2…θ l} represents the network parameter set.

[0080] Building upon the above approach, the improved SqueezeNet network employs convolutional layers (Conv1) as input modules, followed by a Fire network stack module to extract sample features. Compared to traditional convolutional modules, the Fire stack module has fewer parameters and faster computational power. Each Fire module contains three core layers: a compression layer, an expansion layer, and a merging layer. The compression layer uses a 1×1 convolutional filter instead of a 3×3 filter on the input feature matrix C of the previous layer. l The convolution is expressed as in equation (8):

[0081]

[0082] Where l is the network sequence number, C in It is dataset C l Number of feature maps Let x be the output window matrix, f(.) be the convolution calculation function, and x be the output window matrix. l-1 j It is the input window matrix. It's weight. This serves as the baseline. Then, 1×1 and 3×3 convolutional filters are used in the extended layer for grouped convolution to form connections in the sparse channel domains. Finally, a merging layer is used to fuse the information between channels. ReLU pooling is used between layers for downsampling to improve the model's overfitting effect.

[0083]

[0084] Where d(.) represents the downsampling operation. The weights are used for step undersampling between specific layers (layers 1, 4, 8, and 10 in this patent), which preserves activation information to improve model accuracy.

[0085] In this embodiment, the network finally employs a dual-channel parallel convolutional layer to perform multi-target identification and output of the mechanical defect types and severity of GIS / GIL equipment. Specifically, the defect type identification channel consists of a convolutional layer, a global pooling layer, a softmax layer, and a classifier layer, and is connected in series with a defect matching layer to achieve model matching for the severity pattern identification channel. The severity identification channel converts the input feature matrix into a one-dimensional tensor by connecting a max pooling layer and a flattening layer to the output convolutional layer, which serves as the data input for a Random Forest (RF) classifier. This is then combined with the defect matching layer for model selection and training to achieve severity identification. The RF algorithm is a machine learning algorithm that combines Bagging ensemble learning technology with random subspace theory. By constructing different training sets, it increases the differences between classification models and improves the extrapolation prediction ability of the combined classification model. The specific steps are as follows:

[0086] Step 1: Use the Bootstrap resampling method to generate dataset S from the original training set:

[0087] S={(C i ,L i ),i=1,2,...,N} (10)

[0088] Among them (C) i ,L i )∈R d ×R, C i ,L i Let represent the source and target domain sample sets respectively for the i-th sampling;

[0089] Step 2: Randomly select m feature attribute values ​​from the d feature attribute values ​​of each node, and construct a decision tree h using the CART (Classification and Regression Tree) algorithm based on the Gini index. k ;

[0090] Step 3: Repeat Step 1 to train and obtain K classification model sequences {h1(C),h2(C),…,h K (C)} and construct a classification model, where the output of each tree is a specific label or class, and the majority voting rule is defined as in equation (11):

[0091]

[0092] Where arg is the combined classification model, I(C i) is the indicator function, h i It is a single decision tree classification model, and the arg function calculates the number of decision trees that classify the test sample into the k-th class.

[0093] Based on the above technical solutions, a lightweight deep learning algorithm SqueezeNet optimized by random forest is proposed. This algorithm establishes an improved lightweight convolutional neural network that can automatically extract GIS / GIL vibration features and output multiple labels, thereby improving the model's computation speed while also achieving automatic extraction of defect feature information.

[0094] Example 2

[0095] This embodiment provides a method for assessing mechanical defects in GIS / GIL equipment. Applying the mechanical defect diagnosis method for GIS / GIL equipment described in Embodiment 1, this assessment method further includes the following steps:

[0096] Based on load rate LO R The training sample set will be divided into a first evaluation interval, a second evaluation interval, and a third evaluation interval, where the first evaluation interval represents the LO (Local Evaluation) interval. R ≤33%, the second evaluation interval represents 33%. <LO R ≤66%, the third evaluation interval represents LO R >66%; Load factor (LO) R For the equipment operating current I OM With the rated current I of the equipment RM The ratio; the convolutional neural network neural model is trained based on the training sample set.

[0097] Based on the above technical solutions, and according to the load factor (LO)... R The vibration sample is divided into 3 intervals, and the load rate is calculated as shown in equation (12):

[0098] LO r =I OM I RM (12)

[0099] In equation (12), I OM I is the operating current of the equipment. RM This refers to the rated current of the equipment.

[0100] Among them, according to LO R The size divides the training sample set of the source domain into intervals (LO). R ≤33%), Interval 2 (33%) <LO R ≤66%), Interval 3 (LO) RIn three regions (>66%), a jointly improved lightweight deep SqueezeNet network algorithm was trained on the source domain sample set after load partitioning to construct a load-adaptive matching three-stage mechanical defect diagnosis model for GIS / GIL equipment. Finally, load matching was performed on the test sample set of the target domain and input into the corresponding defect identification network. The network then matched the defect identification results to a random forest model for severity assessment, obtaining multi-target diagnostic results of defect type and severity. This achieves the goal of using a load-adaptive convolutional neural network model to diagnose the mechanical defects of matched GIS / GIL equipment and identify the defect type and severity.

[0101] Please see Figure 3 Through the above technical solutions, the entire evaluation method first combines variational mode decomposition and wavelet transform algorithms, and performs step S10 based on sample vibration signals and background noise signals: a time-frequency analysis process for mechanical vibration of GIS / GIL equipment with neighboring mode noise suppression. Then, the architecture of the traditional lightweight SqueezeNet algorithm is improved by using random forest algorithm and dual-channel parallel structure, i.e., step S20: an improved lightweight convolutional neural network is implemented to achieve self-extraction of defect features of GIS / GIL vibration samples and multi-objective output of the network. Finally, the vibration samples are divided into 3 intervals according to the load rate, and step S30: a three-stage mechanical defect diagnosis and severity model of GIS / GIL equipment with load adaptive matching is trained and constructed based on the source domain sample set. After load matching of the test sample set, it is input into the corresponding defect diagnosis network to obtain multi-objective diagnosis results of defect type and severity, and finally achieve effective diagnosis of mechanical defects of GIS / GIL equipment with different loads.

[0102] To verify the model's effectiveness, this embodiment uses vibration signals from three typical mechanical defects—loose busbar base bolts, loose disconnector base bolts, and poor contact of the perforated contact finger—under different operating conditions (0–3000A) of a 110kV GIS / GIL device in normal operation and under different load ranges. Each defect has four severity levels: minor, moderate, and severe. Furthermore, vibration datasets of mechanical defects in GIS / GIL devices with different types and severity levels are constructed for different load ranges. These datasets are randomly divided into training and testing sets at a 2:1 ratio. The training set is further divided into sets S1–S3 based on the load range, and the testing set is S4. Samples of normal and typical mechanical defects under different load currents are shown below. Figure 4 As shown.

[0103] Please see Figure 5The time-frequency comparison analysis of the signals before and after filtering for normal GIS / GIL equipment and mechanical defects due to poor contact reveals that the original signal contains significant noise, exhibiting a clear amplitude distribution in the time-frequency graph. This noise affects the accurate feature extraction of defect information. After filtering, the normal signal primarily consists of vibration amplitude in the 100Hz frequency band, while the feature information of the defect signal in the 700Hz and 2000Hz frequency bands is further highlighted. Eliminating background noise interference improves the accuracy of defect identification.

[0104] Furthermore, test sample sets of mechanical defects in GIS / GIL equipment under different loads can be input into an adaptive three-stage mechanical defect identification and condition assessment model for diagnostic analysis. The results of defect type identification and severity assessment are as follows: Figure 6 As shown, it can be observed that the overall recognition accuracy of the training and testing sample sets is low under low load conditions, with training accuracy and testing accuracy being only 0.893 and 0.928 respectively at 300A. With increasing current, the overall recognition accuracy for different defect samples gradually increases, reaching over 0.992 under current conditions of 1200A and above.

[0105] Meanwhile, in terms of defect severity assessment, the accuracy of assessing the health status of different mechanical defects is relatively low when the load is low (range). As the current increases, the accuracy of identifying the health status of different defects is 0.93 or higher under load conditions of 1200A and above. The model effectively realizes the accurate diagnosis and assessment of mechanical defects in GIS / GIL equipment under different load currents.

[0106] In summary, this assessment method, based on an improved lightweight convolutional neural network and load adaptive matching, is a defect diagnosis technology proposed for actual field engineering. It effectively overcomes problems such as difficulty in extracting defect feature information caused by environmental noise interference and dynamic operating current, as well as low diagnostic accuracy. It can be directly applied to defect identification and condition assessment of GIS / GIL equipment in substations, enabling the formulation of reasonable operation and maintenance plans, and early detection and investigation of internal mechanical defects, thus possessing significant engineering application value.

[0107] Example 3

[0108] This embodiment also provides a GIS / GIL equipment mechanical defect diagnosis system 600. Please refer to [link / reference needed]. Figure 7The modular schematic diagram of the GIS / GIL equipment mechanical defect diagnosis system 600 is mainly used to divide the GIS / GIL equipment mechanical defect diagnosis system 600 into functional modules according to the embodiments of the above method. For example, it can be divided into individual functional modules, or two or more functions can be integrated into one processing module. The integrated module can be implemented in hardware or software functional modules. It should be noted that the module division in this invention is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods. For example, in the case of dividing each functional module according to its corresponding function, Figure 7 The diagram shown is only a schematic of a system / device. The GIS / GIL equipment mechanical defect diagnosis system 600 may include a decomposition module 610, a calculation module 620, a judgment module 630, and an identification module 650. The functions of each unit module are described below.

[0109] The decomposition module 610 is used to acquire the mechanical vibration signal and background noise signal of the GIS / GIL equipment, and to perform multiple decompositions on the mechanical vibration signal and the background noise signal to acquire the m modal components SVMF of the mechanical vibration signal and the n modal components NVMF of the background noise signal, respectively, where m and n are both positive integers;

[0110] Calculation module 620 is used to solve for the barycenter frequency S of the modal component SVMF and the modal component NVMF respectively. f ={S f1 ,S f2 ,…,S fm} and N f ={N f1 N f2 ,…,N fn}, determine the modal component NVMF that is closest to the centroid frequency of each of the modal components SVMF and perform matching;

[0111] The judgment module 630 is used to calculate the neighborhood Γ of each pair of matched modal components SVMF and NVMF, centered on their respective centroid frequencies. Si ∈(S fi -δ,S fi +δ) and Γ Ni ∈(S fi -δ,S fi The amplitude and kurtosis of the frequency domain vibration signal within +δ) are used as threshold filtering criteria. Interference noise modal signals are filtered out by threshold comparison to obtain the target vibration signal. The modal component index i = 1, 2, 3, ..., K, where K is the total number of modal components; δ represents the width of the frequency range.

[0112] The identification module 650 is used to identify the target vibration signal using a convolutional neural network model and obtain the identification result.

[0113] In some embodiments, the GIS / GIL equipment mechanical defect diagnosis system 600 further includes a superposition module 640, which is used to superimpose the filtered modal signals and perform wavelet transform on the superimposed signals to construct a three-dimensional time-frequency spectrum signal, and obtain the target vibration signal based on the time-frequency spectrum signal.

[0114] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., a solid-state drive (SSD)).

[0115] This application describes embodiments of methods, apparatus (systems), and computer program products according to embodiments of this application with reference to flowchart illustrations and / or block diagrams. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0116] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0117] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0118] Obviously, those skilled in the art can make various modifications and variations to the embodiments of this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations to the embodiments of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.

Claims

1. A method for diagnosing mechanical defects in GIS / GIL equipment, characterized in that, Includes the following steps: Acquire mechanical vibration signals and background noise signals of GIS / GIL equipment, and perform multiple decompositions on the mechanical vibration signals and background noise signals of GIS / GIL equipment to obtain m modal components SVMF of the mechanical vibration signals of GIS / GIL equipment and n modal components NVMF of the background noise signals, where m and n are both positive integers; Solve for the centroid frequencies S of the modal components SVMF and NVMF respectively. f ={S f1 , S f2 ,…, S fm } and N f ={N f1 , N f2 ,…, N fn }, determine the modal component NVMF that is closest to the centroid frequency of each of the modal components SVMF and perform matching; For each pairwise matched modal component SVMF and modal component NVMF, the neighborhood Γ is calculated with their respective centroid frequencies as the center. Si ∈(S fi -δ, S fi +δ) and Γ Ni ∈(S fi -δ, S fi The amplitude and kurtosis of the vibration signal in the frequency domain within +δ) are used as threshold filtering criteria. Interference noise modal signals are filtered out by threshold comparison to obtain the target vibration signal. The modal component index i = 1, 2, 3, …, K, where K is the total number of modal components; δ represents the width of the frequency range in the domain. The target vibration signal is identified using a convolutional neural network model to obtain the identification result.

2. The method for diagnosing mechanical defects in GIS / GIL equipment according to claim 1, characterized in that, The mechanical vibration signal of the GIS / GIL equipment and the background noise signal are decomposed multiple times using a joint variational multimodal decomposition algorithm, and the calculation is performed by introducing the Lagrange multiplication operator λ(t) and the quadratic penalty factor τ. The expression (1) is as follows: （1）； In equation (1), u k Represents a modal signal; ω k λ represents the center frequency; λ represents the Lagrange multiplication operator; ∂ represents the gradient operation; δ(t) is the unit impulse function; * represents convolution; t represents time; K represents the total number of modal components; k represents the modal component index; j is a complex number symbol; f(t) represents the vibration time-domain sequence.

3. The method for diagnosing mechanical defects in GIS / GIL equipment according to claim 2, characterized in that, It also includes frequency domain modes u k (ω) and the corresponding center frequency ω k The update steps, and their update calculation expressions are as shown in equations (2) and (3) respectively: （2）； （3）。 4. The method for diagnosing mechanical defects in GIS / GIL equipment according to claim 2, characterized in that, After filtering out interfering noise mode signals using threshold comparison, the following steps are also included: The filtered modal signals are superimposed, and wavelet transform is performed on the superimposed signal to construct a three-dimensional time-frequency spectrum signal. The target vibration signal is obtained based on the time-frequency spectrum signal.

5. The method for diagnosing mechanical defects in GIS / GIL equipment according to claim 4, characterized in that, The steps of superimposing the filtered modal signals and performing wavelet transform on the superimposed signals to construct a three-dimensional time-frequency spectrum signal include the following: Define a signal x(t), x(t)∈L 2 (R), based on the x(t), determine the wavelet transform WT x (a, b), the expression is given by equation (4): （4）； In equation (4), t is time, a and b are the scale factor and translation parameter, respectively, and a is greater than 0; ψ a,b ψ(t) is a family of wavelet basis functions generated by shifting and scaling the mother wavelet function ψ(t).

6. The method for diagnosing mechanical defects in GIS / GIL equipment according to claim 5, characterized in that, ψ a,b The relationship between (t) and the mother wavelet function ψ(t) is expressed as in equation (5): （5）； The constraint condition for ψ(t) is: （6）。 7. The method for diagnosing mechanical defects in GIS / GIL equipment according to claim 2, characterized in that, The algorithmic architecture of the convolutional neural network model is expressed as equation (7): （7）； In equation (7), F(X|Θ) represents a multi-layer nonlinear mapping model; f(.|θ) l ) is the mapping function; Θ={θ l ,θ2…θ l } represents the network parameter set.

8. The method for diagnosing mechanical defects in GIS / GIL equipment according to claim 7, characterized in that, In the convolutional neural network model, a Fire network stack module is used to extract sample features. This Fire network stack module includes a compression layer, an expansion layer, and a merging layer. The compression layer uses a 1×1 convolutional filter to apply the input feature matrix C from the previous layer. l Convolution is performed; the extended layer uses 1×1 and 3×3 convolutional filters to perform grouped convolution to form connections in sparse channel domains; the merging layer fuses information between channels.

9. A method for assessing mechanical defects in GIS / GIL equipment, characterized in that, The method for diagnosing mechanical defects in GIS / GIL equipment as described in any one of claims 1-8 further includes the following steps: Based on load rate LO R The training sample set will be divided into a first evaluation interval, a second evaluation interval, and a third evaluation interval, where the first evaluation interval represents the LO (Local Evaluation) interval. R ≤33%, the second evaluation range represents 33%. <LO R ≤66%, the third evaluation interval represents LO R >66%; Load factor (LO) R For the equipment operating current I OM With the rated current I of the equipment RM The ratio; The convolutional neural network neural model is trained using the training sample set.

10. A mechanical defect diagnosis system for GIS / GIL equipment, characterized in that, include: The decomposition module is used to acquire the mechanical vibration signal and background noise signal of the GIS / GIL equipment, and to perform multiple decompositions on the mechanical vibration signal and the background noise signal to acquire the m modal components SVMF of the mechanical vibration signal and the n modal components NVMF of the background noise signal, respectively, where m and n are both positive integers; The calculation module is used to solve for the barycenter frequency S of the modal component SVMF and the modal component NVMF, respectively. f ={S f1 , S f2 ,…, S fm } and N f ={N f1 , N f2 ,…, N fn }, determine the modal component NVMF that is closest to the centroid frequency of each of the modal components SVMF and perform matching; The judgment module is used to calculate the neighborhood Γ of each pairwise matched modal component SVMF and modal component NVMF, centered on their respective centroid frequencies. Si ∈(S fi -δ, S fi +δ) and Γ Ni ∈(S fi -δ, S fi The amplitude and kurtosis of the vibration signal in the frequency domain within +δ) are used as threshold filtering criteria. Interference noise modal signals are filtered out by threshold comparison to obtain the target vibration signal. The modal component index i = 1, 2, 3, …, K, where K is the total number of modal components; δ represents the width of the frequency range in the domain. The identification module is used to identify the target vibration signal using a convolutional neural network model and obtain the identification result.

Images