Switch cabinet contact fault prediction method and system
Through the improved multi-layer perceptron neural network model, the dual-channel hidden layer and the improved cross-entropy function are used to solve the problem of early prediction of switch cabinet contact fault detection in the prior art and the problem of insufficient prediction accuracy, achieving higher fault prediction accuracy and model generalization capabilities.
Patent Information
- Application Number
- CN202411522308.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-29
- Publication Date
- 2025-02-25
AI Technical Summary
The existing switch cabinet contact fault detection methods mainly focus on situations where there have been or are obvious faults, and the failure cannot be predicted in advance. The neural network-based methods have insufficient data transmission and generalization capabilities, resulting in low prediction accuracy.
The improved multi-layer perceptron neural network model is adopted, and the hidden layer is improved as a dual-channel hidden layer, including a fully connected neuron layer and an attention-based neuron layer, and the cross entropy function is improved as a loss function to improve the model's feature extraction ability and generalization ability.
Through the improved neural network model, switching cabinet contact failure can be predicted more accurately in advance, improving prediction accuracy and model generalization capabilities, and reducing information loss in data transmission.
Smart Images
Figure CN119514615A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of power equipment fault prediction, and in particular to a switch cabinet contact fault prediction method and system. Background Art
[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.
[0003] As an indispensable and important equipment in the power system, the switchgear plays a key role in the distribution, control and protection of power. The contact part inside is the core component to realize the circuit connection and disconnection function. In the long-term operation process, the switchgear contacts are affected by various factors and are prone to failure. Once the switchgear contacts fail, it will cause the interruption of power transmission, affect the normal operation of the power grid, make the power supply unstable, and cause unplanned power outages. Therefore, it is very important to predict the switchgear contact failure in advance.
[0004] Many traditional fault detection methods for switch cabinet contacts focus on detecting faults that have already occurred or obvious signs of faults have appeared, such as threshold-based detection methods, which only sound an alarm when the contact temperature exceeds a pre-set high temperature threshold or the current shows obvious abnormal fluctuations. This detection method is often lagging and cannot predict faults in the early stages of their occurrence. With the continuous development of neural network technology, technicians in this field have tried to apply neural networks to switch cabinet contact fault prediction. However, based on the current neural network structure, when making predictions, some information is easily lost during the transmission of data in the hidden layer. In addition, the existing neural network structure has limited generalization ability and is difficult to capture complex feature relationships when faced with diverse data distributions, which leads to low accuracy of prediction results. All of these restrict the development of neural networks in the field of switch cabinet contact fault prediction. Summary of the invention
[0005] In order to solve the above problems, the present invention proposes a switch cabinet contact fault prediction method and system. The present invention selects a multi-layer perceptron neural network model, and improves the traditional multi-layer perceptron neural network model, changes the traditional hidden layer into a dual-channel hidden layer, and improves the error function of the traditional multi-layer perceptron neural network model, effectively solving the technical problems existing in the prior art.
[0006] In order to achieve the above object, the present invention adopts the following technical solution:
[0007] In a first aspect, the present invention provides a switch cabinet contact fault prediction method, comprising:
[0008] Acquire insulator data, wherein the insulator data includes insulator property data, historical operation data, and historical environment data, and pre-process the acquired insulator data;
[0009] Collect physical data of the switch cabinet contacts to be tested and pre-process the collected physical data;
[0010] An improved multi-layer perceptron neural network model is constructed, wherein the hidden layer in the improved multi-layer perceptron neural network model is a dual-channel hidden layer, wherein one channel is a fully connected neuron layer, and the other channel is a neuron layer based on an attention mechanism; the improved multi-layer perceptron neural network model is trained, an improved cross entropy function is used as a loss function of the improved multi-layer perceptron neural network model, and the trained improved multi-layer perceptron neural network model is used as a contact fault prediction model; the current switch cabinet contact physical data is collected, and the physical data is preprocessed, and the preprocessed physical data is input into the contact fault prediction model, and the fault probability and fault type probability distribution are output.
[0011] According to a further technical solution, the physical data of the switch cabinet contacts include: temperature data, vibration data, displacement data of the switch cabinet contacts and current data flowing through the switch cabinet contacts; and the temperature data, vibration data, displacement data and current data are preprocessed.
[0012] In a further technical solution, the preprocessing includes data cleaning, and the specific method of data cleaning is: dividing the collected physical data of the switch cabinet contacts to be tested into several sub-intervals according to a fixed time interval, setting multiple weights for each sub-interval, and obtaining the final weight w i , according to the final weight w i Perform weighted calculation and determine the dynamic threshold; physical data exceeding the dynamic threshold is considered abnormal data and the abnormal data is eliminated;
[0013] The preprocessing also includes normalizing the physical data after data cleaning.
[0014] A further technical solution is to obtain the final weight w i The method is as follows: First, determine the time weight according to the distance from the collected physical data to the current time The formula is: Where β is the attenuation coefficient, t i The time difference between the collected physical data and the current time; secondly, the density weight is determined according to the data density of the sub-interval where the collected physical data is located The formula is: Among them, n i is the number of physical data in the subinterval where the physical data is located, is the average number of physical data in all subintervals; the final weight
[0015]
[0016] The method for weighted calculation is: using the final weight w i Calculate the weighted mean μ of each subinterval w and weighted standard deviation mean σ w , the formula is: Among them, x i is the physical data in the subinterval, N is the total number of physical data in the subinterval; the dynamic threshold is μ w ±kσ w , where k is the proportionality coefficient.
[0017] In a further technical solution, each neuron in the fully connected neuron layer is connected to a neuron in the input layer, that is: ij =x i × ij , where z ij is the input of the fully connected neuron layer, x i is the neuron in the input layer, w ij is a connection weight; summing the inputs of the fully connected neuron layer to obtain a first total input, and performing a nonlinear transformation on the first total input to obtain a first output;
[0018] In the attention mechanism-based neuron layer, attention weights, weighted sums, and nonlinear transformations are calculated in sequence. Specifically, the correlation degree a between each feature output by the input layer and the neurons in the attention mechanism-based neuron layer is calculated. ij , and as the attention weight, the specific formula is: a ij =σ(y i × a +b a ), where σ is the Sigmoid function, y i is the feature output by the input layer, w a is the weight of a fully connected network with only one layer, b a is the bias term; the features of the input layer output are multiplied by the corresponding attention weights and summed to obtain the second total input, and the second total input is nonlinearly transformed to obtain the second output.
[0019] A further technical solution is to merge the first output in the fully connected neuron layer with the second output in the attention mechanism-based neuron layer as input to the next layer of neurons.
[0020] A further technical solution is that the improved cross entropy function is specifically:
[0021] Among them, m i is the true label, is the predicted value, C i is the cost weight, μ is the regularization parameter, k is the number of weights, C j is the cost weight associated with weight j.
[0022] In a second aspect, the present invention provides a switch cabinet contact fault prediction system, comprising:
[0023] The data acquisition module is configured to: collect physical data of the switch cabinet contacts to be tested, and pre-process the collected physical data;
[0024] The network model building module is configured to: build an improved multi-layer perceptron neural network model, wherein the hidden layer in the improved multi-layer perceptron neural network model is a dual-channel hidden layer, wherein one channel is a fully connected neuron layer, and the other channel is a neuron layer based on an attention mechanism; train the improved multi-layer perceptron neural network model, use an improved cross entropy function as a loss function of the improved multi-layer perceptron neural network model, and use the trained improved multi-layer perceptron neural network model as a contact fault prediction model;
[0025] The contact fault prediction module is configured to: collect the current switch cabinet contact physical data, pre-process the physical data, input the pre-processed physical data into the contact fault prediction model, and output the fault probability and fault type probability distribution.
[0026] Compared with the prior art, the present invention has the following advantages and beneficial effects:
[0027] The present invention selects a multi-layer perceptron neural network model and improves the traditional multi-layer perceptron neural network model, changing the traditional hidden layer into a dual-channel hidden layer, wherein one channel is a fully connected neuron layer and the other channel is an attention mechanism-based neuron layer; the fully connected neuron layer can perform comprehensive feature extraction on input physical data and can learn the complex relationship between different features of the physical data; the attention mechanism-based neuron layer focuses on the extraction of local important features, and can highlight features that have an important impact on the task by calculating the attention weight, and enhance the generalization ability of the model, and the two are combined to further improve the feature extraction ability; and the dual-channel hidden layer can reduce data loss during transmission and can better process different types of data.
[0028] The present invention also improves the error function of the traditional multi-layer perceptron neural network model, and uses the improved cross entropy function as the loss function of the improved multi-layer perceptron neural network model to solve the problem that the error function in the prior art lacks consideration of features. The cost adjustment of the improved cross entropy function based on feature importance in the present invention can make the model pay more attention to features that have an important impact on classification, reduce misclassification caused by ignoring key features, and thus improve the prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their description are used to explain the present invention but do not constitute a limitation of the present invention.
[0030] Figure 1 It is a flow chart of the method of embodiment 1 of the present invention. DETAILED DESCRIPTION
[0031] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.
[0032] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.
[0033] Embodiment 1
[0034] like Figure 1 As shown, this embodiment proposes a switch cabinet contact fault prediction method, which specifically includes the following steps:
[0035] S1: Collecting physical data of the switch cabinet contacts to be tested and preprocessing the collected physical data;
[0036] S2: constructing an improved multi-layer perceptron neural network model, wherein the hidden layer in the improved multi-layer perceptron neural network model is a dual-channel hidden layer, wherein one channel is a fully connected neuron layer, and the other channel is a neuron layer based on an attention mechanism; training the improved multi-layer perceptron neural network model, using an improved cross entropy function as a loss function of the improved multi-layer perceptron neural network model, and using the trained improved multi-layer perceptron neural network model as a contact fault prediction model;
[0037] S3: Collect the current physical data of the switch cabinet contacts, pre-process the physical data, input the pre-processed physical data into the contact fault prediction model, and output the fault probability and fault type probability distribution.
[0038] Among them, in step S1, the physical data of the switch cabinet contacts to be tested collected include but are not limited to temperature data, vibration data, displacement data of the switch cabinet contacts and current data flowing through the switch cabinet contacts. Specifically: a variety of sensors are installed around the switch cabinet contacts, including high-precision temperature sensors for monitoring and collecting contact temperature data, because problems such as poor contact and overload of the contacts often cause temperature changes; vibration sensors for collecting vibration data of the contacts, and the vibration sensors can capture the vibration of the contacts during opening and closing operations and operation; high-precision current sensors for measuring the current flowing through the contacts; displacement sensors for collecting the displacement of the contacts.
[0039] In step S1, the collected temperature data, vibration data, displacement data and current data are also preprocessed. In this embodiment, the preprocessing includes data cleaning and normalizing the physical data after data cleaning. The specific method of data cleaning is: the collected physical data of the switch cabinet contact to be tested is divided into several sub-intervals according to fixed time intervals, such as dividing it into a sub-interval every 20 minutes, and multiple weights are set for each sub-interval, including time weight and density weight And get the final weight w i , according to the final weight w i Perform weighted calculation and determine the dynamic threshold; physical data exceeding the dynamic threshold is considered abnormal data and is removed; specifically, the final weight w is obtained. i The method is as follows: for each sub-interval, different weights are assigned according to the distance from the physical data to the current time. The closer the physical data is to the current time, the greater the weight. First, the time weight is determined according to the distance from the collected physical data to the current time. The formula is: Where β is the attenuation coefficient, t i The time difference between the collected physical data and the current time; secondly, the density weight is determined according to the data density of the sub-interval where the collected physical data is located The formula is: Among them, n i is the number of physical data in the subinterval where the physical data is located, is the average number of physical data in all subintervals; the final weight Reusing the final weight w i Perform weighted calculation, specifically: calculate the weighted mean μ of each subinterval w and weighted standard deviation mean σ w , the formulas are: and Among them, x iis the physical data in the subinterval, N is the total number of physical data in the subinterval; the dynamic threshold is μ w ±kσ w , where k is the proportionality coefficient. In this embodiment, k is 3. The above data cleaning method takes into account the data characteristics more comprehensively, that is, it takes into account the time weight and density weight at the same time, and can more comprehensively reflect the characteristics of the data. For example, in some cases, the data may appear in a certain time period. The density weight can make the data in this time period play a more reasonable role in the threshold calculation, avoiding the deviation caused by relying solely on the time weight. At the same time, the calculation of the final weight makes the calculation of the weighted mean and weighted standard deviation more accurate, so that the determined dynamic threshold can more accurately identify abnormal data, especially in the case of uneven data distribution. In this embodiment, this method can better adapt to the characteristics of the data and reduce misjudgment.
[0040] In step S2, each neuron in the fully connected neuron layer is connected to the neurons in the input layer, that is: ij =x i × ij , where z ij is the input of the fully connected neuron layer, x i (i=1,2,3,…,e) are neurons in the input layer, w ij is the connection weight; the input z of the fully connected neuron layer ij The sum is obtained to obtain the first total input, and the first total input is subjected to nonlinear transformation to obtain the first output. Specifically: the first total input Among them, b1 is the bias term, and the first output H1 is obtained by performing nonlinear transformation using the activation function (ReLU function can be selected), H1=ReLU(A1), which can enhance the nonlinear expression ability of the network and enable the network to learn more complex functional relationships.
[0041] In the attention mechanism-based neuron layer, attention weights, weighted sums, and nonlinear transformations are calculated in turn. Specifically, the degree of association a between each feature output by the input layer and the neurons in the attention mechanism-based neuron layer is calculated. ij , and as the attention weight, the specific formula is: a ij =σ(y i × a +b a ), where σ is the Sigmoid function, y i is the feature output by the input layer, w a is the weight of a fully connected network with only one layer, b ais a bias term; it can dynamically calculate the importance weight of each feature for the attention mechanism-based neuron layer according to the input features; multiply the features output by the input layer by the corresponding attention weights and sum them up to obtain the second total input A2, Where b2 is the bias term, and the second output H2 is obtained by performing nonlinear transformation using the improved activation function, where the improved activation function is a variant of the Sigmoid function Sigmoid variant (x), specifically: Where δ is a coefficient greater than 1, which makes the function have a steeper gradient to enhance the attention effect.
[0042] H2=Sigmoid variant (A2), merge the first output H1 in the fully connected neuron layer with the second output H2 in the attention mechanism-based neuron layer, that is, add the two together as the input of the next layer of neurons. This dual-channel hidden layer structure can simultaneously utilize the learning ability of the fully connected neuron layer and the ability of the attention mechanism-based neuron layer to focus on important features. For switch cabinet contact fault prediction, different fault modes may be associated with different feature combinations. The fully connected neuron layer can learn the general patterns in the data, while the attention mechanism-based neuron layer can focus on key features according to feature importance, thereby better capturing these complex feature relationships and improving the network's fault prediction accuracy.
[0043] In step S2, the improved cross entropy function is specifically: Among them, m i is the true label, is the predicted value, C i is the cost weight, μ is the regularization parameter, k is the number of weights, C j is the cost weight associated with weight j. The improved cross entropy function is used as the loss function of the improved multi-layer perceptron neural network model to solve the problem that the error function in the prior art lacks consideration of features. The cost adjustment of the improved cross entropy function based on feature importance in the present invention can make the model pay more attention to features that have an important impact on classification, reduce misclassification caused by ignoring key features, and thus improve the prediction accuracy of the model.
[0044] Embodiment 2
[0045] This embodiment provides a switch cabinet contact fault prediction system, which specifically includes the following modules:
[0046] The data acquisition module is configured to: collect physical data of the switch cabinet contacts to be tested, and pre-process the collected physical data;
[0047] The network model building module is configured to: build an improved multi-layer perceptron neural network model, wherein the hidden layer in the improved multi-layer perceptron neural network model is a dual-channel hidden layer, wherein one channel is a fully connected neuron layer, and the other channel is a neuron layer based on an attention mechanism; train the improved multi-layer perceptron neural network model, use an improved cross entropy function as a loss function of the improved multi-layer perceptron neural network model, and use the trained improved multi-layer perceptron neural network model as a contact fault prediction model;
[0048] The contact fault prediction module is configured to: collect the current switch cabinet contact physical data, pre-process the physical data, input the pre-processed physical data into the contact fault prediction model, and output the fault probability and fault type probability distribution.
[0049] The implementation of the specific modules in this embodiment refers to the steps of an insulator flashover voltage prediction method described in Embodiment 1, and will not be described in detail here.
[0050] The above-described embodiments only express several implementation methods of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.
Claims
1. A switch cabinet contact fault prediction method, characterized in that: include: Collect physical data of the switch cabinet contacts to be tested and pre-process the collected physical data; Constructing an improved multi-layer perceptron neural network model, wherein the hidden layer in the improved multi-layer perceptron neural network model is a dual-channel hidden layer, wherein one channel is a fully connected neuron layer, and the other channel is a neuron layer based on an attention mechanism; training the improved multi-layer perceptron neural network model, using an improved cross entropy function as a loss function of the improved multi-layer perceptron neural network model, and using the trained improved multi-layer perceptron neural network model as a contact fault prediction model; The current physical data of the switch cabinet contacts are collected and preprocessed, and the preprocessed physical data are input into the contact fault prediction model to output the fault probability and the fault type probability distribution.
2. A switch cabinet contact fault prediction method according to claim 1, characterized in that: The physical data of the switch cabinet contact includes: temperature data, vibration data, displacement data of the switch cabinet contact and current data flowing through the switch cabinet contact; and the temperature data, vibration data, displacement data and current data are preprocessed.
3. A switch cabinet contact fault prediction method according to claim 2, characterized in that: The preprocessing includes data cleaning. The specific method of data cleaning is: dividing the collected physical data of the switch cabinet contacts to be tested into several sub-intervals according to fixed time intervals, setting multiple weights for each sub-interval, and obtaining the final weight w i , according to the final weight w i Perform weighted calculation and determine the dynamic threshold; physical data exceeding the dynamic threshold is considered abnormal data and the abnormal data is eliminated; The preprocessing also includes normalizing the physical data after data cleaning.
4. A switch cabinet contact fault prediction method according to claim 3, characterized in that: The final weight w is obtained i The method is as follows: First, determine the time weight according to the distance from the collected physical data to the current time The formula is: Where β is the attenuation coefficient, t i The time difference between the collected physical data and the current time; secondly, the density weight is determined according to the data density of the sub-interval where the collected physical data is located The formula is: Among them, n i is the number of physical data in the subinterval where the physical data is located, is the average number of physical data in all subintervals; the final weight The method for weighted calculation is: using the final weight w i Calculate the weighted mean μ of each subinterval w and weighted standard deviation mean σ w , the formula is: Among them, x i is the physical data in the subinterval, N is the total number of physical data in the subinterval; the dynamic threshold is μ w ±kσ w , where k is the proportionality coefficient.
5. A switch cabinet contact fault prediction method according to claim 1, characterized in that: Each neuron in the fully connected neuron layer is connected to a neuron in the input layer, that is: ij =x i × ij , where z ij is the input of the fully connected neuron layer, x i is the neuron in the input layer, w ij is the connection weight; Summing the inputs of the fully connected neuron layer to obtain a first total input, and performing a nonlinear transformation on the first total input to obtain a first output; In the attention mechanism-based neuron layer, attention weights, weighted sums, and nonlinear transformations are calculated in sequence. Specifically, the correlation degree a between each feature output by the input layer and the neurons in the attention mechanism-based neuron layer is calculated. ij , and as the attention weight, the specific formula is: a ij =σ(y i × a +b a ), where σ is the Sigmoid function, y i is the feature output by the input layer, w a is the weight of a fully connected network with only one layer, b a is the bias term; the features of the input layer output are multiplied by the corresponding attention weights and summed to obtain the second total input, and the second total input is nonlinearly transformed to obtain the second output.
6. A switch cabinet contact fault prediction method according to claim 5, characterized in that: The first output in the fully connected neuron layer is combined with the second output in the attention mechanism-based neuron layer as input of the next layer of neurons.
7. A switch cabinet contact fault prediction method according to claim 1, characterized in that: The improved cross entropy function is specifically: Among them, m i is the true label, is the predicted value, C i is the cost weight, μ is the regularization parameter, k is the number of weights, C j is the cost weight associated with weight j.
8. A switch cabinet contact fault prediction system, characterized in that: include: The data acquisition module is configured to: collect physical data of the switch cabinet contacts to be tested, and pre-process the collected physical data; The network model building module is configured to: build an improved multi-layer perceptron neural network model, wherein the hidden layer in the improved multi-layer perceptron neural network model is a dual-channel hidden layer, wherein one channel is a fully connected neuron layer, and the other channel is a neuron layer based on an attention mechanism; train the improved multi-layer perceptron neural network model, use an improved cross entropy function as a loss function of the improved multi-layer perceptron neural network model, and use the trained improved multi-layer perceptron neural network model as a contact fault prediction model; The contact fault prediction module is configured to: collect the current switch cabinet contact physical data, pre-process the physical data, input the pre-processed physical data into the contact fault prediction model, and output the fault probability and fault type probability distribution.
9. A computer-readable storage medium having a program stored thereon, characterized in that: When the program is executed by a processor, the steps in the switch cabinet contact fault prediction method as described in any one of claims 1 to 7 are implemented.
10. An electronic device comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the switch cabinet contact fault prediction method according to any one of claims 1 to 7 are implemented.