Online fault classification method, equipment and storage medium for high-voltage electronic switchgear

Abstract

The present invention discloses a method, device, and storage medium for online fault classification of high-voltage electronic switchgear. The method randomly selects and segments monitoring parameters of the high-voltage electronic switchgear to reconstruct a one-dimensional data set. The method then processes the one-dimensional data set using a sliding window to obtain multiple one-dimensional subsequences. If the distance between two adjacent one-dimensional subsequences is less than a set distance threshold and the two adjacent one-dimensional subsequences correspond to the same category, one of the two adjacent one-dimensional subsequences is deleted, and the K one-dimensional subsequences with the largest weights are selected. A training set is constructed using the one-dimensional subsequences, and a convolutional neural network is trained to obtain a classification model. The method utilizes a sliding window to identify subsequences and performs pruning, thereby improving data quality and resolving the problem in prior art of long model training times and inability to improve classification accuracy due to insufficient data volume.

Description

Technical Field

[0001] The present invention relates to the technical field of state monitoring and fault classification of high-voltage electronic switches, and in particular to an online fault classification method, device and storage medium for high-voltage electronic switch equipment. Background Art

[0002] As a primary component for power grid control and protection, high-voltage electronic switchgear is crucial to the safe and reliable operation of the grid. A high-voltage circuit breaker essentially performs an instantaneous closing and opening action. This involves transferring the energy stored in the operating mechanism's energy storage mechanism to a series of mechanical components, such as the motor, over a very short period of time. During this energy transfer, the mechanical components move, rub, and collide with each other.

[0003] Throughout this process, mechanical components operate at high speeds and experience intense impacts. Accelerations can reach 100g, and contacts can accelerate to several meters per second within milliseconds. The impacts are even more intense during braking and buffering. This high-intensity impact response can lead to increased friction in transmission components, stalling, and phase loss in the circuit breaker after numerous opening and closing operations. These issues can cause malfunctions, malfunctions, or refusal to close. After numerous opening and closing operations, the circuit breaker's surface inevitably develops uneven surfaces due to long-term friction. These increases the friction between mechanical components during the opening and closing process, increasing the torque required for opening and closing. Stalling often occurs in the moving arcing contacts, shift fork, and sickle arm, often due to assembly errors and lack of lubrication in moving parts. This type of fault can have more serious consequences than those caused by increased friction. When the power supply system is operating normally, a phase loss fault can occur if the opening and closing of a switch are out of sync. Circuit breakers can also fail due to refusal to close. Phase loss in circuit breakers is primarily related to mechanical failures, including operating mechanism failure, circuit breaker inoperability, and transmission failure. The consequences of a phase loss in a circuit breaker are severe, potentially damaging power equipment.

[0004] With the rapid improvement and development of data mining technology and artificial intelligence algorithms, mining the vast amount of historical data accumulated during the operation of high-voltage electronic switchgear to perform condition monitoring and fault warning has gradually become a new research direction. However, this research direction relies on a large amount of fault data for preliminary training. The more training samples and the more complete the fault data, the higher the accuracy of the resulting diagnostic model.

[0005] For online equipment, especially small-batch equipment, it is impossible to effectively identify abnormal equipment conditions based solely on the equipment's historical normal operating conditions data, and the classification effect is poor.

[0006] CN115293208A provides a disconnector fault diagnosis method and system that utilizes a deep weighted fusion model of an SVM classifier to diagnose disconnector faults. Due to the characteristics of the SVM classifier, it is not ideal for multi-classification problems and is only suitable for binary classification. Although weighted fusion is used to compensate for overfitting caused by incorrect parameter selection, the fixed penalty coefficient can also lead to a decrease in classification accuracy. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a method, device and storage medium for online fault classification of high-voltage electronic switchgear in view of the shortcomings of the existing technology, so as to improve the accuracy of fault classification.

[0008] To solve the above technical problems, the present invention adopts a technical solution: a method for online fault classification of high-voltage electronic switchgear, comprising the following steps:

[0009] Randomly select and segment the monitoring parameters of high-voltage electronic switchgear and reconstruct a one-dimensional data set;

[0010] Processing the one-dimensional data set using a sliding window to obtain multiple one-dimensional subsequences; if the distance between two adjacent one-dimensional subsequences is less than a set distance threshold, and the two adjacent one-dimensional subsequences correspond to the same category, deleting one of the two adjacent one-dimensional subsequences and selecting K one-dimensional subsequences with the largest weights;

[0011] The K one-dimensional subsequences with the largest weights are selected to construct a training set, and the convolutional neural network is trained to obtain a classification model.

[0012] This invention uses a sliding window to identify subsequences and performs pruning (i.e., removing redundant one-dimensional subsequences), improving data quality and resolving the existing issues of long model training times and limited classification accuracy due to insufficient data. This invention eliminates the need for complex methods such as weighted fusion and fixed penalty coefficients, resulting in high classification accuracy.

[0013] In the present invention, the monitoring parameters of the high-voltage electronic switchgear include electromagnetic torque Tr, actual current Iq curve, and actual current Id curve.

[0014] In the present invention, the specific implementation process of reconstructing a one-dimensional data set includes:

[0015] Obtain multivariate historical data of high-voltage electronic switchgear under normal and abnormal operating conditions as a training matrix T;

[0016] Normalize the training matrix T;

[0017] Randomly extract some vectors from the normalized training matrix T to obtain a one-dimensional data set.

[0018] The present invention utilizes historical data of normal and abnormal high-voltage electronic switches to construct a historical memory matrix, which can process and classify multi-dimensional complex data. When selecting matrix data to construct the memory matrix and predict the output, since different monitoring data of high-voltage electronic switch equipment have different dimensions and the absolute values ​​of different monitoring data vary greatly, in order to ensure that the distance between different observation vectors is correctly measured using a nonlinear operator, the present invention normalizes each monitoring parameter separately.

[0019] In the present invention, the one-dimensional subsequence X est The expression is: Xest = argmax (ω); where, D is the time series of monitoring parameters, Xsub is the subsequence selected by the sliding window, ω is the subsequence weight, Represents the operator for finding the Euclidean distance between vectors, is the union of time series whose distance is less than d, It is the union of time series whose distance is greater than d, where d is a given value.

[0020] The present invention uses weights rather than simple distance calculations to preliminarily select multiple subsequences with obvious features and combines them with residual spectrum analysis, making the time series feature data more obvious and easier to classify, making the modeling process simpler, meeting the real-time working conditions on site, and further improving the classification accuracy.

[0021] The specific implementation process of constructing a training set using the K one-dimensional subsequences with the largest weights and training the convolutional neural network to obtain the classification model includes:

[0022] Reconstructing the K one-dimensional subsequences with the largest weights into multiple two-dimensional grayscale images, obtaining saliency images using the two-dimensional grayscale images, and using the multiple saliency images to form an image dataset;

[0023] The image dataset is split into a training set and a validation set, and a convolutional neural network is trained to obtain a classification model.

[0024] The present invention converts the one-dimensional subsequence (time series) into a picture form, making the time series feature data more obvious and convenient for classification. It fully utilizes the advantages of convolutional neural network data mining, further improves the accuracy of the classification model, and thus further improves the fault classification accuracy.

[0025] In the present invention, the expression of the saliency image S(x) of the two-dimensional grayscale image I(x) is: S(x) = g(I(x))*F -1 [exp(R(f)+P(f))] 2; Where, P(f)=Sh(F[I(x)]), R(f)=L(f)-h(I(x))*L(f), L(f)=log(A(f)), A(f)=R(F[I(x)]), F and F -1 They represent the forward and inverse Fourier transforms, respectively. g(I(x)) is Gaussian filtering, h(I(x)) is mean filtering, x is a two-dimensional grayscale image, and f is the value obtained after Fourier transform of x.

[0026] The present invention uses a residual spectrum analysis method to make the features more distinct, thereby further improving the fault classification accuracy.

[0027] The specific implementation process of splitting the image dataset into a training set and a validation set, training a convolutional neural network, and obtaining a classification model includes:

[0028] Randomly extract two parts of data from the training set to obtain two sub-training sets S and T.

[0029] The four data sets S, T, S∩T, and S∪T are used to train a convolutional neural network model using the MAML method to obtain the classification model.

[0030] This invention introduces a convex game strategy to further combine the salient features obtained after data augmentation with the generalization capabilities of meta-learning, significantly improving fault classification accuracy. To better extract features, this invention embeds a convolutional neural network into the meta-learning framework, enabling rapid and high fault classification accuracy with minimal data during training.

[0031] Furthermore, the method of the present invention further comprises:

[0032] The classification model is used to classify monitoring parameters of high-voltage electronic switchgear acquired in real time.

[0033] As an inventive concept, the present invention further provides a terminal device, comprising:

[0034] one or more processors;

[0035] A memory having one or more programs stored thereon, which, when executed by the one or more processors, enables the one or more processors to implement the steps of the above method of the present invention.

[0036] As an inventive concept, the present invention also provides a computer-readable storage medium storing a computer program, which implements the steps of the above method when executed by a processor.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. The present invention uses the historical data of normal and abnormal high-voltage electronic switches to form a historical memory matrix for modeling, which can process and classify multi-dimensional complex data;

[0039] 2. This invention introduces a sliding window and performs pruning processing. Compared with the traditional dynamic programming and sliding window methods used in classification, it improves the quality of data. By using weights rather than simple distance calculations, K subsequences with obvious characteristics are initially selected and combined with residual spectrum analysis to make the time series feature data more obvious and easier to classify. It has the advantages of simple modeling and meeting the real-time working conditions on site, thereby improving the classification accuracy.

[0040] 3. Unlike previous methods that directly process time series data, this method converts time series data into images, reshaping one-dimensional data into two-dimensional images. Furthermore, to better utilize the advantages of convolutional neural network (CNN) data mining, residual spectrum analysis is used to make features more distinct, which is conducive to improving classification accuracy.

[0041] 4. This invention proposes a convex game strategy for time series data. Convex games are a type of cooperative game. Meta-learning itself is proposed to improve generalization ability, while strategies such as residual spectrum analysis are aimed at feature enhancement and data augmentation. However, feature enhancement and model generalization ability are not positively correlated. Therefore, this invention introduces a convex game strategy to further combine the significant features obtained after data augmentation with the generalization ability of meta-learning, greatly improving fault classification accuracy.

[0042] 5. The present invention adopts a meta-learning-based method. Meta-learning itself is a framework. In order to better perform feature extraction, the present invention embeds a convolutional neural network into the meta-learning framework. During training, it can quickly achieve high fault classification accuracy based on a small amount of data. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 Schematic diagram of the fault classification principle based on meta-learning in an embodiment of the present invention;

[0044] Figure 2 It is the three-closed-loop control principle of the motor;

[0045] Figure 3 It is a structural diagram of a double-acting circuit breaker;

[0046] Figure 4 Schematic diagram of the meta-learning method. DETAILED DESCRIPTION

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0048] In this document, the terms "include," "comprising," and similar expressions are intended to indicate a logical relationship and should not be construed as indicating a spatial relationship. For example, "A includes B" is intended to indicate that B logically belongs to A, not that B is spatially located within A. Furthermore, the terms "include," "comprising," and similar expressions should be considered open-ended, not closed-ended. For example, "A includes B" is intended to indicate that B belongs to A, but B does not necessarily constitute the entirety of A; A may also include other elements such as C, D, and E.

[0049] Example 1

[0050] like Figure 1 As shown, this embodiment provides a method for online fault classification of high-voltage electronic switchgear, comprising the following steps:

[0051] S1. Randomly select and split multidimensional data to reconstruct the data set.

[0052] In the embodiment of the present invention, the multidimensional dataset in step S1 above can be constructed in the following manner:

[0053] Step 1, select key monitoring parameters: by analyzing the working principle of high-voltage electronic switch equipment, select the key monitoring parameters of high-voltage electronic switch equipment;

[0054] Step 2: Select the training matrix T: Obtain multivariate historical data of the high-voltage electronic switchgear under normal and abnormal operating conditions as the training matrix T;

[0055] Step 3, data preprocessing: normalize the training matrix T;

[0056] Step 4, synthesize the dataset D: randomly extract some vectors from the normalized training matrix T and perform dimensionality reduction.

[0057] In the above steps, Step 1 to Step 4 are offline processing stages.

[0058] In Step 1, the host computer interface of the high-voltage electronic switch two-in-one device is mainly used to observe the action waveform of the servo motor drive mechanism. Its observable parameters include position, motor d-axis / q-axis voltage, motor d-axis / q-axis current, speed, and torque:

[0059] (1) Pos is the actual position of the motor. In the open position, the value is 6.28, and in the closed position, the value is 3.14. PosRef is the motor position value set by the controller.

[0060] (2) Iq represents the actual current value of the q-axis, and IqRef represents the q-axis current value set by the controller;

[0061] (3) Id represents the actual current value of the d-axis, and IdRef represents the q-axis current value set by the controller;

[0062] (4) Ud and Uq are the d-axis and q-axis voltage values, respectively. Their values ​​are synthesized by the three-phase voltages of A, B, and C. The voltage value is not the main basis for judging the status of the circuit breaker;

[0063] (5) Speed ​​is the measured speed waveform, and SpeedRef represents the speed waveform set by the controller;

[0064] (7) Tr is the electromagnetic torque of the motor, which is calculated from the current.

[0065] The servo motor itself cannot operate independently. The operation of the servo motor comes from the "drive" of the servo driver, and the servo driver drives the servo motor based on three-loop control (such as Figure 2 Three-loop control refers to a control loop consisting of three control loops: position loop, speed loop, and current loop. Each control loop forms a separate closed-loop control. For example, in the position loop, the position controller (servo driver) sends a command (PosRef) to the servo motor. The servo motor receives the command and moves to the specified position (Pos). Pos then returns to the input of the position controller in a specific manner, exerting a control influence on the input, forming a system control method with feedback information.

[0066] The movement of the servo motor will then cause the circuit breaker to move. The circuit breaker is divided into many forms, such as double-acting circuit breakers (such as Figure 3 Take the example of a double-acting circuit breaker (shown in Figure 1). A double-acting circuit breaker consists of eight parts: the static main contact 1, the sickle crank arm 2, the shift fork 3, the static arc contact 4, the moving arc contact 5, the moving main contact 6, the insulating pull rod 7, and the main shaft crank arm 8. Specifically, the arc extinguishing chamber is composed of components such as the moving and static main contacts and the moving and static arc contacts. During closing, the main shaft crank arm rotates under gravity, driving the insulating pull rod and sickle crank arm, bringing the moving and static main contacts and the moving and static arc contacts into contact, and the shift fork engages, completing a closing operation. During opening, the shift fork opens, and the spring moves the pull rod and crank arm, separating the moving and static main contacts and the moving and static arc contacts, completing a closing operation.

[0067] To determine whether a single opening and closing action is normal, it is mainly necessary to compare whether the parameters involved in the three-loop control are normal. Observe the electromagnetic torque Tr, actual current Iq curve, actual current Id curve, actual speed waveform Speed, and the degree of fit with the IqRef and IdRef curves set by the controller, and whether there are spikes and curve fluctuations.

[0068] In summary, through the analysis of the working principle of high-voltage electronic switchgear, the key monitoring parameters of high-voltage electronic switchgear are selected: electromagnetic torque Tr, actual current Iq curve, and actual current Id curve.

[0069] Assume that there are n interrelated variables in the “related variable set” of a certain device, and the vector composed of the observed values ​​of these n monitoring parameters is recorded as the observation vector X(t i ), t i For the i-th moment, the expression is as follows:

[0070]

[0071] For example, in the embodiment of the present invention, the relevant variable set in the high-voltage electronic switch has three interrelated observation variables, and the vector composed of the observation values ​​of these three monitoring parameters is recorded as the observation vector X(t i ), the expression is as follows:

[0072]

[0073] x Tr (t i ),x Iq (t i ),x Id (t i ) represent t i The observed values ​​of electromagnetic torque, actual current Iq and actual current Id at the moment. Similarly, the monitoring parameter x j The sequence of all observations at each time state can be represented as a vector X j :

[0074] X j =[x j (t1),x j (t2),…,x j (t m )].

[0075] In Step 3, the measured values ​​of various monitoring parameters of the high-voltage electronic switchgear under normal and abnormal operating conditions are selected to form a training matrix, which is represented by the letter T, and m represents the mth moment:

[0076]

[0077] Normal operation refers to the operating state of high-voltage electronic switchgear during periods of low failure rates. From the perspective of the data for each monitored parameter, normal operation refers to the stable variation of each monitored parameter, which also includes the normal degradation of each monitored parameter. Abnormal conditions can be categorized based on the changes in current and torque: a slight increase in current and torque compared to normal; a sharp increase in current and torque; and a decrease in current and torque compared to normal opening and closing values.

[0078] Regarding the selection of the training data matrix T, the following prerequisites are required:

[0079] 1) Cover a sufficiently long operating period, including the operation of high-voltage electronic switchgear in different seasons and under different loads, and all dynamic changes of each monitoring parameter during normal and abnormal operation of the equipment;

[0080] 2) The time states contained in the training matrix T can all express a state of the device object;

[0081] 3) The simultaneity of each monitoring parameter in each set of sampled values ​​must be satisfied, and they must be sampled at the same time. In summary, assuming that the training data contains a total of k time states, the training matrix T can be expressed as:

[0082] T=[X(t1),X(t2),…,X(t k )].

[0083] In Step 3, when selecting matrix data to construct the memory matrix and predict the output, since the dimensions of different monitoring data of high-voltage electronic switchgear are different and the absolute values ​​of different monitoring data vary greatly, in order to ensure that the nonlinear operator is used to correctly measure the distance between different observation vectors, each monitoring parameter needs to be normalized according to its own extreme value (a row in the training matrix and the observation matrix) so that the actual measurement value is mapped to the range [0, 1]. The mapping relationship is as follows:

[0084]

[0085] Among them, x is the data before normalization, x * is the normalized data, mean(x) and std(x) are the mean and standard deviation of the data before normalization of the monitoring parameters.

[0086] In Step 4, random seed slicing (i.e., dimensionality reduction) is used to select some observation vectors X(t i ), forming new data, namely the data set Data.

[0087] S2. Input the data set into the sliding window and analyze and select subsequences.

[0088] Candidate subsequence X est The calculation formula for (i.e. one-dimensional subsequence) is:

[0089]

[0090]

[0091] Xest = argmax(ω);

[0092] Where D is the time series, Xsub is the subsequence selected by the sliding window, and ω is the subsequence weight. Represents the operator for calculating the Euclidean distance between vectors. To make the nonlinear state estimation method more intuitive and avoid irreversibility after dot product due to correlation between variables, the Euclidean distance is used to reflect the relationship between two vectors, that is, the similarity between two vectors. The relationship between the subsequence selected by the sliding window and the original sequence is the Euclidean distance operation:

[0093]

[0094] in The operator for finding the Euclidean distance between vectors is used to represent the size of similarity:

[0095]

[0096] The initially selected subsequences exhibit significant sequence repetition, necessitating further pruning to remove subsequences with overly similar waveforms. Assume the subsequences are X1 and X2, their corresponding categories are Y1 and Y2, and the distance between the two subsequences is d12, where dx is a given distance threshold. If the conditions d12 < dx and Y1 = Y2 are met, the two subsequences are considered similar, and X2 is removed from the candidate set. Pruning improves the quality of the dataset, and finally, the K candidate subsequences with the highest weights are extracted to construct a new dataset.

[0097] The physical meaning of the above formula is: the Euclidean distance between each subsequence and the original sequence vector is calculated to reflect the similarity between them. The most similar ones have the smallest Euclidean distance between them, the data features are more obvious, and their contribution to the classification result is the greatest.

[0098] S3. Reconstruct the subsequence into a grayscale image and perform significant feature analysis.

[0099] In an embodiment of the present invention, in order to utilize the feature extraction and noise reduction capabilities of a two-dimensional CNN, the original continuous one-dimensional time series data is stacked into two-dimensional data to construct a grayscale image.

[0100] In order to make the characteristics of the data more obvious, the spectral residual method is used to perform saliency detection on the converted image. The process is as follows:

[0101] Assume that the grayscale image corresponding to a piece of data in the dataset is I(x).

[0102] By calculation, we can get the amplitude spectrum A(f)=R(F[I(x)]), the phase spectrum P(f)=Sh(F[I(x)]) and the logarithmic spectrum L(f)=log(A(f)).

[0103] R(f)=L(f)-h(I(x))*L(f)

[0104] Obtain its remaining graph, and finally according to the formula

[0105] S(x)=g(I(x))*F -1 [exp(R(f)+P(f))] 2 ;

[0106] The saliency image is obtained to replace the grayscale image and store it in the data set. -1 represent the forward and inverse Fourier transform respectively.

[0107] S4. Split the final image obtained in step S3 into a training set and a validation set for meta-learning classification.

[0108] Since simply splitting the dataset into training and validation sets cannot guarantee a better combination of feature enhancement and meta-learning generalization ability, a convex game strategy is introduced in feature data classification.

[0109] v(S∪T)+v(s∩T)≥v(S)+v(T);

[0110] In this embodiment, v(S) in the formula represents the loss function, namely CE(S) below. Specifically, the dataset is first divided into a training set and a validation set, and then the training set is randomly divided again, and two parts S and T are randomly extracted. Meta-learning is performed on each of these four components (S, T, S∪T, and S∩T). The resulting loss functions are fine-tuned during meta-learning and then retrained to obtain the final loss functions for each component. These four loss functions are calculated using the above formula to obtain the final loss Lmeta, which is then backpropagated to update the model. CE(S) represents the cross-entropy loss of S, which is used in backpropagation and model updates.

[0111] Lmeta=CE(S∪T)+CE(S∩T)-CE(S)-CE(T).

[0112] As a framework, meta-learning uses a convolutional neural network (CNN) model for internal embedding. The network model structure consists of a four-layer convolutional network and a batch normalization layer, followed by a linear layer.

[0113] In this embodiment of the present invention, since the training set is divided into four categories, the following MAML algorithm will be applied to each category. Each of the four categories obtained by convex game will go through two stages: pre-training and fine-tuning.

[0114] The meta-learning method of the embodiment of the present invention is the MAML algorithm, which is divided into two stages: pre-training and fine-tuning. In pre-training, after initializing the CNN model, several tasks are randomly sampled and the first gradient calculation is performed. Where θ is a random initial parameter, α is a learning rate parameter, is gradient descent, is the loss function, where the loss is cross entropy loss, and the gradient update is performed to get θ':

[0115]

[0116] After a batch is completed, a secondary gradient update is performed to improve the generalization ability of the model so that it can better learn out-of-domain knowledge. The updated gradient θ N for:

[0117]

[0118] Where β is the learning rate parameter, is the loss function of a sampling period and . Repeat the above process until the end of the period. After this part of the training is completed, a preliminary model can be obtained.

[0119] Based on the preliminary model, further fine-tuning is required. The fine-tuning process is different from the pre-training process in several ways. The initial parameter θ of the fine-tuning process directly uses the θ obtained by pre-training. N , rather than random initialization, the data loading part no longer needs to be looped because only one task is extracted instead of a batch. Finally, there is no second gradient update, but the parameters are updated directly using the results of the first gradient calculation. Finally, multiple gradient calculation results are averaged to avoid extreme cases.

[0120] In an embodiment of the present invention, MAML training is performed on S, T, S∪T, and S∩T respectively, and each training cycle includes the above-mentioned pre-training process and fine-tuning process. For each data set (S, T, S∪T, S∩T), a cross entropy loss will be obtained after each training cycle is completed. After all four data sets have completed a training cycle, a total loss Lmeta will be obtained. The total loss is used to adjust the parameters of the CNN model, and the adjusted parameters are used as the initial parameters of the next training cycle, and so on, until the training cycle reaches the set number of times. After each parameter adjustment, the model obtained by each parameter adjustment will be retained. For example, if the total number of training cycles is N, N CNN models will be obtained, and the model with the best performance after verification with the verification set will be the final classification model.

[0121] Example 2

[0122] This embodiment provides an online fault classification system for high-voltage electronic switchgear, including:

[0123] A data set reconstruction unit, used to randomly select and segment monitoring parameters of high-voltage electronic switchgear and reconstruct a one-dimensional data set;

[0124] a subsequence construction unit, configured to process the one-dimensional data set using a sliding window to obtain a plurality of one-dimensional subsequences; if the distance between two adjacent one-dimensional subsequences is less than a set distance threshold, and the two adjacent one-dimensional subsequences correspond to the same category, then deleting one of the two adjacent one-dimensional subsequences and selecting K one-dimensional subsequences with the largest weights;

[0125] The output unit is used to construct a training set using the one-dimensional subsequence obtained by the subsequence construction unit, train the convolutional neural network, and obtain a classification model.

[0126] Furthermore, the output unit of this embodiment includes:

[0127] An image dataset construction module is used to reconstruct the one-dimensional subsequence into multiple two-dimensional grayscale images, obtain saliency images using the two-dimensional grayscale images, and the multiple saliency images constitute the image dataset;

[0128] The training module is used to split the image dataset into a training set and a validation set, train a convolutional neural network, and obtain a classification model.

[0129] The system of this embodiment also includes:

[0130] The classification unit is used to classify the monitoring parameters of the high-voltage electronic switchgear acquired in real time by using a classification model.

[0131] The process of each unit / module implementing the corresponding function in this embodiment is the same as that in embodiment 1 and will not be repeated here.

[0132] Example 3

[0133] Embodiment 3 of the present invention provides a terminal device corresponding to the above-mentioned embodiment 1. The terminal device may be a processing device for a client, such as a mobile phone, a laptop computer, a tablet computer, a desktop computer, etc., to execute the method of the above-mentioned embodiment.

[0134] The terminal device of this embodiment includes a memory, a processor, and a computer program stored in the memory; the processor executes the computer program in the memory to implement the steps of the method in the above-mentioned embodiment 1.

[0135] In some implementations, the memory may be a high-speed random access memory (RAM), and may also include a non-volatile memory, such as at least one disk storage.

[0136] In other implementations, the processor may be a central processing unit (CPU), a digital signal processor (DSP), or other general-purpose processors, which are not limited herein.

[0137] Example 4

[0138] Embodiment 4 of the present invention provides a computer-readable storage medium corresponding to the above-mentioned embodiment 1, on which a computer program / instruction is stored. When the computer program / instruction is executed by a processor, the steps of the method of the above-mentioned embodiment 1 are implemented.

[0139] Computer readable storage media can be tangible devices that hold and store instructions used by instruction execution devices. Computer readable storage media can be, for example, but not limited to, electronic storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any combination thereof.

[0140] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiment of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal translation scripting language JavaScript, etc.

[0141] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0142] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0143] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present application.

[0144] Obviously, those skilled in the art may make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application is intended to include these modifications and variations.

Claims

1. A method for online fault classification of high-voltage electronic switchgear, characterized in that: The following steps are involved: Randomly select and segment the monitoring parameters of high-voltage electronic switchgear and reconstruct a one-dimensional data set; Processing the one-dimensional data set using a sliding window to obtain multiple one-dimensional subsequences; If the distance between two adjacent one-dimensional subsequences is less than the set distance threshold, and the categories corresponding to the two adjacent one-dimensional subsequences are the same, then one of the two adjacent one-dimensional subsequences is deleted and the K one-dimensional subsequences with the largest weights are selected; The K one-dimensional subsequences with the largest weights are used to construct a training set, train the convolutional neural network, and obtain a classification model; The monitoring parameters of the high-voltage electronic switchgear include electromagnetic torque Tr, actual current Iq curve, and actual current Id curve; The specific implementation process of constructing a training set using the K one-dimensional subsequences with the largest weights and training the convolutional neural network to obtain the classification model includes: Reconstructing the K one-dimensional subsequences with the largest weights into multiple two-dimensional grayscale images, obtaining saliency images using the two-dimensional grayscale images, and using the multiple saliency images to form an image dataset; Splitting the image dataset into a training set and a validation set, training a convolutional neural network, and obtaining a classification model; The expression of the saliency image S(x) of the two-dimensional grayscale image I(x) is: S(x)=g(I(x))*F -1 [exp(R(f)+P(f))] 2 ; Where, P(f)=Sh(F[I(x)]), R(f)=L(f)-h(I(x))*L(f), L(f)=log(A(f)), A(f)=R(F[I(x)]), F and F -1 They represent the forward and inverse Fourier transforms, respectively. g(I(x)) is Gaussian filtering, h(I(x)) is mean filtering, P(f) is the phase spectrum, A(f) is the amplitude spectrum, x is a two-dimensional grayscale image, and f is the value obtained by Fourier transforming x.

2. The method for online fault classification of high-voltage electronic switchgear according to claim 1, characterized in that: The specific implementation process of reconstructing a one-dimensional data set includes: Obtain multivariate historical data of high-voltage electronic switchgear under normal and abnormal operating conditions as a training matrix T; Normalize the training matrix T; Randomly extract some vectors from the normalized training matrix T to obtain a one-dimensional data set.

3. The method for online fault classification of high-voltage electronic switchgear according to claim 1, characterized in that: One-dimensional subsequence X est The expression is: Xest = argmax (ω); where, D is the time series of monitoring parameters, Xsub is the subsequence selected by the sliding window, ω is the subsequence weight, Represents the operator for finding the Euclidean distance between vectors, which is the union of time series whose distance is less than d. It is the union of time series whose distance is greater than d, where d is a given value.

4. The method for online fault classification of high-voltage electronic switchgear according to claim 1, characterized in that: The specific implementation process of splitting the image dataset into a training set and a validation set, training a convolutional neural network, and obtaining a classification model includes: Randomly extract two parts of data from the training set to obtain two sub-training sets S and T. The four data sets S, T, S∩T, and S∪T are used to train a convolutional neural network model using the MAML method to obtain the classification model.

5. The method for online fault classification of high-voltage electronic switchgear according to claim 1, characterized in that: Also includes: The classification model is used to classify monitoring parameters of high-voltage electronic switchgear acquired in real time.

6. A terminal device, characterized in that: include: one or more processors; A memory having one or more programs stored thereon, which, when executed by the one or more processors, enables the one or more processors to implement the steps of the method according to any one of claims 1 to 5.

7. A computer-readable storage medium, characterized in that The device stores a computer program, which implements the steps of the method according to any one of claims 1 to 5 when executed by a processor.

Images