GIS equipment abnormal sound vibration identification method based on improved residual network

Abstract

The invention discloses a GIS equipment abnormal sound vibration identification method based on an improved residual network, and belongs to the technical field of GIS equipment fault detection and identification. According to the method, the traditional residual error network (SENet) model is optimally designed, the compression excitation model based on the improved residual error network (SENet) is constructed, and the model is relatively high in recognition accuracy and relatively small in error. The model can accurately detect and identify the mechanical vibration caused by poor contact of the isolating switch contact of the gas insulated switchgear (GIS), and the identification effect is remarkable; and a new feasible scheme is provided for researching the detection and fault identification of the mechanical abnormal sound vibration defect of the gas insulated switchgear (GIS). According to the recognition method provided by the invention, the early-stage latent problem of the gas insulated switchgear (GIS) can be found and recognized early, early warning is realized, greater loss caused by major faults is reduced as much as possible, reliable supply of electric power is further ensured, and the recognition method has very important significance and practical value.

Description

Technical Field

[0001] The invention relates to the technical field of GIS equipment fault detection and identification, and in particular to a GIS equipment abnormal sound and vibration identification method based on an improved residual network. Background Art

[0002] With the widespread application of gas insulated switchgear (GIS) in power grids, it optimizes the design and orderly combines the primary equipment such as circuit breakers, disconnectors, grounding switches, voltage and current transformers, lightning arresters, busbars, bushings, etc., encapsulates them in a metal shell, and fills them with SF6 gas as arc extinguishing and insulating medium to form a closed combination of electrical appliances. Compared with traditional open equipment, GIS equipment occupies a small area, has high reliability, strong safety, and less operation and maintenance workload. However, due to its fully enclosed structure, once a fault occurs, it is difficult to accurately locate and quickly repair it.

[0003] According to statistics, there are more than 20,000 GIS equipment failures every year, most of which are internal failures of GIS equipment. Due to the complex internal structure of GIS equipment and the large differences in manufacturing process levels, there are many types of internal failures, and most of them are non-contact failures such as microcracks and internal discharges, which are difficult to detect. When GIS equipment is abnormal, it often manifests as obvious vibration or abnormal sound. If on-site personnel fail to detect equipment abnormalities in time, serious consequences will often occur.

[0004] In summary, the traditional detection methods have many inherent deficiencies, and detecting GIS equipment faults through vibration signals is relatively novel and worthy of further study. In view of this, the present invention proposes a GIS equipment abnormal vibration identification method based on an improved residual network to achieve GIS abnormal vibration fault detection and identification. Summary of the invention

[0005] 1. Technical Problems to be Solved by the Present Invention

[0006] The purpose of the present invention is to propose a method for identifying abnormal noise and vibration of GIS equipment based on an improved residual network to solve the problems raised in the background technology.

[0007] 2. Technical solution

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] A method for identifying abnormal noise and vibration of gas insulated switchgear (GIS) equipment based on improved residual network (SENet) comprises the following steps:

[0010] S1. Using a vibration sensor to collect vibration signals as a target sample set, the vibration signals of various components of a gas insulated switchgear (GIS) device are integrated into a vibration training sample set and a vibration test sample set;

[0011] S2, input the vibration signal collected in S1 into a vibration computer for processing, remove irrelevant vibration signals caused by external influences, denoise and enhance the signal, and extract the vibration signal feature quantity containing the abnormal sound characteristics of the gas insulated switchgear (GIS);

[0012] S3, randomly extracting the continuous feature data before the abnormal sound and vibration and the data during the abnormal sound and vibration from the data set as a group of data for training, so as to make the model more integrated;

[0013] S4, improve the traditional residual network (SENet) model, and use the improved residual network (SENet) model to process the data;

[0014] S5. Randomly select abnormal sound vibration signals and normal vibration signals before the abnormal sound vibration signals from the training set and the test set, divide the selected abnormal sound vibration signals and the normal vibration signals before the abnormal sound vibration signals into 1:49, and then use the AdamW optimization algorithm to further train the residual network (SENet) model;

[0015] S6. Process and identify the vibration signals in the vibration test sample set according to the trained model to predict whether abnormal noise and vibration will occur.

[0016] Preferably, the S1 specifically includes the following contents:

[0017] Vibration sensors are installed on the circuit breakers, disconnectors, earthing switches, voltage and current transformers, lightning arresters, busbars, and bushing casings of the gas insulated switchgear (GIS). Sufficient vibration signals are collected by the vibration sensors as target sample sets, and then vibration training sample sets and vibration test sample sets of different contact states of various components of the gas insulated switchgear (GIS) are constructed; the vibration training sample sets and vibration test sample sets specifically include the following data: ① vibration data of the circuit breaker; ② vibration data of the disconnector; ③ vibration data of the earthing switch; ④ vibration data of the voltage transformer; ⑤ vibration data of the current transformer; ⑥ vibration data of the voltage transformer; ⑦ vibration data of the lightning arrester; ⑧ vibration data of the busbar; ⑨ casing vibration data.

[0018] Preferably, the data input into the vibration computer in S2 are the shell vibration data and other relevant parameter data of the gas insulated switchgear (GIS) during operation before the abnormal sound and vibration. After removing irrelevant signals, signal denoising and signal enhancement processing, the final output is the shell vibration data of the gas insulated switchgear (GIS) during the operation of the abnormal sound and vibration.

[0019] Preferably, the improvement of the traditional residual network (SENet) model described in S4 specifically includes the following contents: improving the convolution kernel size, stride and padding of each layer of the traditional residual network (SENet) model to adapt to the structured vibration data of gas insulated switchgear (GIS); in addition, changing the output dimension of the FC layer in the final output part of the residual network (SENet) model to make it equal to the time step number of the vibration abnormal sound to be detected.

[0020] Preferably, the processing of data using the improved residual network (SENet) model in S4 specifically includes the following contents:

[0021] S4.1. Construct a residual network: denote the expected function as H(x), denote the learned result as x, and the next layer of the network learns the residual F(x) of H(x) and x. Therefore, the output of the next layer is the shortcut connection sum of F(x) and x.

[0022] S4.2, Compression operation: The residual network (SENet) model performs feature compression along the spatial dimension, converting each two-dimensional feature channel into a real number with a global receptive field. The dimension of its output matches the number of input feature channels. The specific function is expressed as:

[0023]

[0024] Among them, F sq To use global average pooling of channels, Z c is the cth element of Z; H and W are spatial dimensions, C is the number of feature channels, and u c is the current feature channel;

[0025] S4.3, Excitation operation: The residual network (SENet) model builds the correlation between channels through two fully connected (FC) layers and outputs the same number of weights as the input features. The first fully connected (FC) layer in the bottleneck structure containing two fully connected (FC) layers is used to reduce the dimension, and the dimension reduction coefficient r is a hyperparameter; then the linear rectification function ReLU is used for activation, and the last second fully connected (FC) layer restores the original dimension; the specific function is expressed as:

[0026] s=F ex (z,W)=δ[W 2 ReLU(W 1 z)]

[0027] Among them, s is the weight, F ex is the global average pooling, z is the zth element, ReLU is the linear rectification function, δ is the function of ReLU,

[0028] S4.4 Application of attention mechanism: The weight coefficients between each feature channel are learned through the SE module, and the residual network (SENet) model that weights each feature channel item by item through Scale is more recognizable for the features of each channel.

[0029] Preferably, the further training of the residual network (SENet) model using the AdamW optimization algorithm in S5 specifically refers to performing bias correction on the AdamW algorithm so that each iterative learning rate has a certain range, thereby stabilizing the parameters;

[0030] The AdamW optimization algorithm is as follows:

[0031] 1) Calculation of momentum and second-order moment:

[0032] m t =β 1 m t-1 +(1-β 1 ) t

[0033]

[0034] Among them, m t represents the first-order momentum of the gradient, β 1 represents the momentum decay rate, usually taken as 0.9; v t represents the second-order momentum of the square of the gradient; β 2 represents the decay rate of the square of momentum, usually taken as 0.999; g t is the gradient value;

[0035] 2) Parameter update:

[0036]

[0037] Among them, θ t is a parameter term, η is the learning rate, Represents the gradient update part of Adam; Represents an independent weight decay term, independent of gradient updates.

[0038] 3. Beneficial effects

[0039] The present invention optimizes the design of the traditional residual network (SENet) model and constructs a compression excitation model based on the improved residual network (SENet). The model has high recognition accuracy and small error. The test results show that the model can accurately detect and identify the mechanical vibration caused by poor contact of the disconnector contacts of the gas insulated switchgear (GIS), and the recognition effect is significant; it provides a new feasible solution for the detection and fault identification of mechanical abnormal vibration defects of gas insulated switchgear (GIS).

[0040] The identification method proposed in the present invention can detect and identify early potential problems of gas insulated switchgear (GIS) at an early stage, provide early warning, and minimize greater losses caused by major failures, thereby ensuring a reliable supply of electricity, which has very important significance and practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of the abnormal sound and vibration identification method of GIS equipment based on the improved residual network proposed in Example 1 of the present invention;

[0042] Figure 2 It is the attention mechanism proposed in Embodiment 1 of the present invention;

[0043] Figure 3 This is the SE-ResNet diagram proposed in Example 1 of the present invention. DETAILED DESCRIPTION

[0044] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0045] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0046] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0047] Embodiment 1:

[0048] The present invention proposes a method for identifying abnormal noise and vibration of GIS equipment based on an improved residual network, the purpose of which is to:

[0049] Objective 1: To propose identifiable vibration characteristics for abnormal noise and vibration of GIS equipment;

[0050] Objective 2: Based on the improved residual network vibration and noise detection model, a method for identifying abnormal noise and vibration of GIS equipment is established.

[0051] See also Figure 1 The overall idea of ​​the abnormal sound and vibration identification method of GIS equipment based on improved residual network proposed in the present invention is as follows:

[0052] Step 1: construct a vibration training sample set and a vibration test sample set of GIS equipment, wherein the vibration training sample set and the vibration test sample set are collected and constructed through vibration sensor detection;

[0053] Step 2: Obtain the vibration abnormal sound front-end signal and the vibration abnormal sound signal through the vibration computer, wherein the vibration abnormal sound front-end signal and the vibration abnormal sound signal are from the vibration training sample set and the vibration test sample set of the GIS equipment;

[0054] Step 3: Use the continuous feature data before the abnormal sound and vibration and the data during the abnormal sound and vibration in the randomly selected data set as a set of data for training to make the model more integrated;

[0055] Step 4: Combine SENet to build a GIS equipment vibration recognition model through compression and excitation operations;

[0056] Step 5: Use the AdamW optimization algorithm to train the SENet model faster and make the parameters more stable;

[0057] Step 6: Based on the trained model, the vibration signal before the abnormal noise and vibration occurs can be predicted and identified, and it is determined that the abnormal noise and vibration will occur.

[0058] Purpose 1 of the present invention is specifically achieved through the following technical solutions.

[0059] In step 1:

[0060] In order to better extract the characteristics of vibration signals, the present invention uses a vibration computer to process the vibration signals, removes irrelevant vibration signals generated by external influences, and performs signal denoising and enhancement processing to extract vibration signal characteristic quantities containing abnormal sound characteristics of GIS equipment.

[0061] In step 2:

[0062] The extracted vibration signals containing abnormal noise and vibration of GIS equipment are integrated into sample sets and training sets. In order to establish a more accurate prediction model and a vibration model of multi-parameter interaction that is closer to the reality, a data-driven approach is adopted, that is, machine learning or deep learning algorithms are used to model according to the input parameters, and features similar to the real conditions are abstracted or learned internally and output.

[0063] In step 3:

[0064] The vibration data before and during abnormal vibration in the training set are randomly extracted as a set of data for training. The purpose is to allow the model to analyze enough data to achieve better abnormal vibration recognition results. By extracting data for training the model through random sampling, the trained model can pay more attention to the mutual influence relationship between multiple parameters, and learn the influence on vibration parameters, making the model more integrated and paying more attention to the changes in vibration caused by parameter changes in practice.

[0065] Purpose 2 of the present invention is specifically achieved through the following technical solutions.

[0066] In step S4:

[0067] Construct a residual network. Let the expected function be represented as H(x) and the learned result be represented as x. Then the network of the next layer does not learn a new x as before, but learns the residual F(x) of H(x) and x. Therefore, the output of the next layer is the shortcut connection sum of F(x) and x.

[0068] The shortcut connection in the residual network does not add new parameters and does not increase the computational complexity. Learning F(x) to fit the residual also solves the problem that F(x) can be learned to be 0 when x is already close to H(x). In this way, research on large-scale data based on a deeper Resnet will obtain better results than conventional deep networks.

[0069] SENet starts with feature channels, builds the relationship between feature channels through compression and excitation operations, automatically obtains the importance of each feature channel, that is, weight, through learning, and then enhances useful features and suppresses features that are not very useful for the current task according to the learned weights.

[0070] Compression operation, feature compression along the spatial dimension, converting each two-bit feature channel into a real number so that this real number has a global receptive field to some extent, and the output dimension matches the number of input feature channels, where H and W are spatial dimensions, C is the number of feature channels, and u c is the current feature channel.

[0071]

[0072] Among them, F sq To use global average pooling of channels, Z c is the cth element of Z; H and W are spatial dimensions, C is the number of feature channels, and u c is the current feature channel.

[0073] The excitation operation builds the correlation between channels through two fully connected (FC) layers and outputs the same number of weights as the input features:

[0074] s=F ex (z,W)=δ[W 2 ReLU(W 1 z)]

[0075] Among them, s is the weight, F ex is the global average pooling, z is the zth element, ReLU is the linear rectification function, δ is the function of ReLU,

[0076] Here, a bottleneck structure consisting of two FC layers is adopted, in which the first FC layer plays the role of dimensionality reduction, and the dimensionality reduction coefficient r is a hyperparameter. Then the linear rectification function ReLU is used for activation, and the last FC layer restores the original dimension.

[0077] The SE module learns the weight coefficients between each feature channel, and weights each feature channel item by item through Scale, so that the model has more recognition ability for the features of each channel. This is the application of the attention mechanism on the feature channel (such as Figure 2 shown).

[0078] The improved SENet model is based on the original 18-layer SENet, and the convolution kernel size, stride and padding of each layer are improved to be suitable for the structured vibration data of GIS. In addition, the output dimension of the FC layer is changed in the final output part of the network to be equal to the abnormal sound and vibration data to be identified.

[0079] See also Figure 3 , the 18-layer network provided by the original model, namely 4×2 SE modules, is used. Each block contains two convolutional layers, plus the input convolutional layer and the last FC layer, as a verification to explore the feasibility of vibration recognition. And the input and output dimensions, as well as the internal partial layer convolution kernel size, stride size and padding size are improved to better adapt to the dimensions of multi-parameter data input and the purpose of research.

[0080] In step S5:

[0081] Using the AdamW optimization algorithm, after bias correction, the learning rate of each iteration has a certain range, making the parameters relatively stable.

[0082] The AdamW optimization algorithm is as follows:

[0083] 1) Calculation of momentum and second-order moment:

[0084] m t =β 1 m t-1 +(1-β 1 ) t

[0085]

[0086] Among them, m t represents the first-order momentum of the gradient, β 1 represents the momentum decay rate, usually taken as 0.9; v t represents the second-order momentum of the square of the gradient; β 2 represents the decay rate of the square of momentum, usually taken as 0.999; g t is the gradient value.

[0087] 2) Parameter update:

[0088]

[0089] Among them, θ t is a parameter term, η is the learning rate, Represents the gradient update part of Adam; Represents an independent weight decay term, independent of gradient updates.

[0090] In step S6:

[0091] Based on the trained model, the vibration signal before the abnormal noise and vibration occurs can be predicted and identified, and it can be determined that the abnormal noise and vibration will occur.

[0092] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and improved concepts of the present invention within the technical scope disclosed by the present invention, and they should be covered by the protection scope of the present invention.

Claims

1. A method for identifying abnormal noise and vibration of GIS equipment based on an improved residual network, characterized in that: The following steps are involved: S1, using vibration sensors to collect vibration signals as target sample sets, and integrating vibration signals of various components of GIS equipment into vibration training sample sets and vibration test sample sets; S2, input the vibration signal collected in S1 into the vibration computer for processing, remove irrelevant vibration signals caused by external influences, denoise and enhance the signal, and extract the vibration signal feature quantity containing the abnormal sound characteristics of the GIS equipment; S3, randomly extracting the continuous feature data before the abnormal sound and vibration and the data during the abnormal sound and vibration from the data set as a group of data for training, so as to make the model more integrated; S4, improve the traditional SENet model and use the improved SENet model to process the data; S5. Randomly select abnormal sound vibration signals and normal vibration signals before the abnormal sound vibration signals from the training set and the test set, divide the selected abnormal sound vibration signals and the normal vibration signals before the abnormal sound vibration signals into 1:49, and then use the AdamW optimization algorithm to further train the SENet model; S6. Process and identify the vibration signals in the vibration test sample set according to the trained model to predict whether abnormal noise and vibration will occur.

2. According to the method for identifying abnormal sound and vibration of GIS equipment based on improved residual network in claim 1, it is characterized in that: The S1 specifically includes the following contents: Vibration sensors are installed on the circuit breakers, disconnectors, grounding switches, voltage and current transformers, lightning arresters, busbars, and bushing casings of the GIS. Sufficient vibration signals are collected by the vibration sensors as target sample sets, and then vibration training sample sets and vibration test sample sets of different contact states of various components of the GIS equipment are constructed; the vibration training sample sets and vibration test sample sets specifically include the following data: ① vibration data of the circuit breaker; ② vibration data of the disconnector; ③ vibration data of the grounding switch; ④ vibration data of the voltage transformer; ⑤ vibration data of the current transformer; ⑥ vibration data of the voltage transformer; ⑦ vibration data of the lightning arrester; ⑧ vibration data of the busbar; ⑨ casing vibration data.

3. The method for identifying abnormal sound and vibration of GIS equipment based on improved residual network according to claim 1 is characterized in that: The data input into the vibration computer in S2 is the shell vibration data and other relevant parameter data when the GIS is working before the abnormal sound and vibration. After removing irrelevant signals, signal denoising and signal enhancement processing, the final output is the shell vibration data when the abnormal sound and vibration occur during the GIS working process.

4. The method for identifying abnormal sound and vibration of GIS equipment based on improved residual network according to claim 1 is characterized in that: The improvements to the traditional SENet model described in S4 specifically include the following: improving the convolution kernel size, stride, and padding of each layer of the traditional SENet model to adapt to GIS structured vibration data; in addition, changing the output dimension of the FC layer in the final output part of the SENet model to make it equal to the time step number of the vibration abnormal sound to be detected.

5. The method for identifying abnormal sound and vibration of GIS equipment based on improved residual network according to claim 4 is characterized in that: The data processing using the improved SENet model described in S4 specifically includes the following contents: S4.

1. Construct a residual network: denote the expected function as H(x), denote the learned result as x, and the next layer of the network learns the residual F(x) of H(x) and x. Therefore, the output of the next layer is the shortcut connection sum of F(x) and x. S4.2, Compression operation: SENet performs feature compression along the spatial dimension, converting each two-dimensional feature channel into a real number with a global receptive field. The dimension of its output matches the number of input feature channels. The specific function is expressed as: Among them, F sq To use global average pooling of channels, Z c is the cth element of Z; H and W are spatial dimensions, C is the number of feature channels, and u c is the current feature channel; S4.3, Excitation operation: SENet builds the correlation between channels through two fully connected layers and outputs the same number of weights as the input features. The first fully connected layer in the bottleneck structure containing two fully connected layers is used to reduce the dimension, and the dimension reduction coefficient r is a hyperparameter; then the linear rectification function ReLU is used for activation, and the last second fully connected layer restores the original dimension; the specific function is expressed as: s=F ex (z,W)=δ[W2ReLU(W1z)] Among them, s is the weight, F ex is the global average pooling, z is the zth element, ReLU is the linear rectification function, δ is the function of ReLU, S4.4 Application of attention mechanism: The SE module learns the weight coefficients between each feature channel, and the SENet model that weights each feature channel item by item through Scale is more recognizable for the features of each channel.

6. The method for identifying abnormal sound and vibration of GIS equipment based on improved residual network according to claim 1 is characterized in that: The use of the AdamW optimization algorithm to further train the SENet model described in S5 specifically refers to performing bias correction on the AdamW algorithm so that each iterative learning rate has a certain range, thereby making the parameters tend to be stable; The AdamW optimization algorithm is as follows: 1) Calculation of momentum and second-order moment: m t =β1m t-1 +(1-β1)g t Among them, m t represents the first-order momentum of the gradient, β1 represents the momentum decay rate; v t represents the second-order momentum of the square of the gradient; β2 represents the decay rate of the square of momentum, g t is the gradient value; 2) Parameter update: Among them, θ t is a parameter term, η is the learning rate, Represents the gradient update part of Adam; Represents an independent weight decay term, independent of gradient updates.

Images