Online parallel fault diagnosis method and system for power grid power equipment

Abstract

The invention relates to the technical field of multivariate data monitoring and diagnosis in the power grid power industry, and provides an online parallel fault diagnosis method and system for power grid power equipment, and the method comprises the steps: receiving power grid feature vector data in real time through an IRichSpout interface, and transmitting the data to a Spout source assembly, thereby forming a to-be-processed data stream; the method comprises the following steps: encapsulating a data stream to be processed into a plurality of Tuple tuples, and generating a unique ID for each Tuple tuple; receiving the Tuple tuple by using a PreBolt component, and preprocessing a data set in the Tuple tuple through a standard fractional method to obtain a standardized sample to be classified; and processing the standardized to-be-classified sample by using a deep classification diagnosis model constructed through unsupervised pre-training, supervised learning, supervised fine tuning and classifier online learning to obtain a fault diagnosis result of the power equipment. The method is high in classification accuracy, and can accurately classify the streaming data of the power grid equipment.

Description

Technical Field

[0001] The present invention relates to the technical field of multivariate data monitoring and diagnosis in the power grid industry, and in particular to an online parallel fault diagnosis method and system for power grid equipment. Background Art

[0002] With the development of power systems, power equipment failures have a significant impact on people's lives, necessitating continuous monitoring of equipment status. Advances in sensor and communication technologies have led to an exponential growth in power grid data. This data is real-time, volatile, and infinite, creating streaming data that requires continuous monitoring. Hadoop, the existing platform, can process batch data but suffers from poor real-time performance. Storm, an open-source distributed real-time computing architecture, can rapidly process massive data streams, addressing Hadoop's real-time processing limitations.

[0003] With the rise of Storm, some applications have emerged in the power industry. Storm implements a time-based sliding window processing method and uses thresholds to detect anomalies in power grid data streams. It also rapidly processes alarm data from power grid equipment and uses clustering algorithms to process related data streams.

[0004] Deep learning algorithms belong to machine learning algorithms, which can be divided into supervised learning and unsupervised learning. In the real world, most samples are unlabeled, so unsupervised learning is more widely used than supervised learning.

[0005] Deep learning models such as restricted Boltzmann machines, autoencoders, and extreme learning machine autoencoders can effectively utilize large numbers of unlabeled samples to solve complex machine learning challenges. However, autoencoders have a simple structure and lack probabilistic generative models. RBMs are overly complex and have difficulty processing continuous-valued data. Extreme learning machine autoencoders have limited generalization capabilities. Furthermore, diagnostic models can experience concept drift during continuous diagnostics in data streams. Consequently, conventional deep learning algorithms suffer from low classification accuracy and efficiency, as well as poor noise immunity. Summary of the Invention

[0006] The purpose of the present invention is to solve at least one technical problem in the background technology and to provide a method and system for online parallel fault diagnosis of power grid equipment.

[0007] To achieve the above objectives, the present invention provides an online parallel fault diagnosis method for power grid power equipment, comprising:

[0008] Receive power grid feature vector data in real time through the IRichSpout interface and send it to the Spout source component to form a data stream to be processed;

[0009] Encapsulate the data stream to be processed into multiple tuples and generate a unique ID for each tuple;

[0010] Use the PreBolt component to receive the Tuple tuple and preprocess the data set in the Tuple tuple using the standard score method to obtain standardized samples to be classified;

[0011] The deep classification diagnosis model constructed through unsupervised pre-training, supervised learning, supervised fine-tuning and classifier online learning is used to process standardized samples to obtain fault diagnosis results of power equipment.

[0012] According to one aspect of the present invention, the method for constructing the deep classification diagnosis model includes: respectively deploying a multi-layer denoising autoencoder, a multi-layer Gaussian restricted Boltzmann machine, a multi-layer extreme learning machine autoencoder, and an online sequence extreme learning machine in different Bolt components of the Storm platform to obtain a DAEBolt component, a GRBMBolt component, a DELEFBolt component, and an OS-ELMBolt component;

[0013] Connect the DAEBolt component, GRBMBolt component, DELEFBolt component and OS-ELMBolt component in sequence to obtain the deep classification diagnosis model deployed on the Storm platform.

[0014] According to one aspect of the present invention, the unsupervised pre-training of the deep classification diagnosis model includes:

[0015] Connect historical power grid data to the Spout source component of the Storm platform through the IRichSpout interface, and encapsulate the historical power grid data into multiple tuples. Each tuple is configured with a unique ID.

[0016] Use the PreBolt component to receive the Tuple tuple and preprocess the data set in the Tuple tuple using the standard score method to obtain model training samples;

[0017] The model training samples are input into the deep classification diagnosis model, and the unlabeled samples are used to perform unsupervised pre-training on the DAEBolt component, GRBMBolt component, and DELEFBolt component in the deep classification diagnosis model;

[0018] The DAEBolt component encapsulates a multi-layer denoising autoencoder, which receives model training samples as input for multi-layer feature extraction. The output of each layer of the multi-layer denoising autoencoder is used as the input of the next layer for unsupervised pre-training.

[0019] The GRBMBolt component encapsulates a multi-layer Gaussian restricted Boltzmann machine. The GRBMBolt component receives the output of the trained DAEBolt component as its input for multi-layer feature extraction. The output of each layer between the multi-layer Gaussian restricted Boltzmann machines is used as the input of the lower layer for unsupervised pre-training. The GRBMBolt component uses the contrastive divergence algorithm for training;

[0020] The DELEFBolt component encapsulates a multi-layer extreme learning machine autoencoder. The DELEFBolt component receives the output of the trained GRBMBolt component as the input of this component to perform multi-layer feature extraction. The output of each hidden layer between the multi-layer extreme learning machine autoencoders is used as the input of the next hidden layer for unsupervised pre-training.

[0021] According to one aspect of the present invention, the supervised learning of the deep classification diagnostic model includes:

[0022] After unsupervised pre-training, supervised learning is performed on the OS-ELMBolt component. The hidden layer parameters in the DAEBolt and GRBMBolt components are kept fixed. The input samples are continuously subjected to multi-layer feature extraction through a multi-layer denoising autoencoder, a multi-layer Gaussian restricted Boltzmann machine, and a multi-layer extreme learning machine autoencoder. The final hidden layer output extracted by the DELEFBolt component is used as the feature input to the OS-ELMBolt component.

[0023] The OS-ELMBolt component performs supervised learning using labeled samples. An online sequence extreme learning machine is deployed within the OS-ELMBolt component, which is responsible for the final classification diagnosis of the entire deep classification diagnosis model. The OS-ELMBolt component receives the output of the DELEFBolt component as its input, and adjusts the output layer weights of the online sequence extreme learning machine through least squares calculation based on the input data, thereby completing the supervised classification training of the online sequence extreme learning machine.

[0024] According to one aspect of the present invention, the supervised fine-tuning of the deep classification diagnosis model is:

[0025] The back-propagation algorithm is used to perform supervised fine-tuning on the parameters of the DAEBolt and GRBMBolt components.

[0026] According to one aspect of the present invention, the online learning of the classifier of the deep classification diagnosis model is:

[0027] The dataset formed by the final hidden layer output matrix obtained by the training samples in the sequential learning phase of the online sequential extreme learning machine and the category label matrix of the training samples is used as input to train the online sequential extreme learning machine in the OS-ELMBolt component.

[0028] According to one aspect of the present invention, the fault diagnosis result is stored in a database via a DatabaseBolt component, and the fault diagnosis result is stored in a data file via a FileBolt component.

[0029] To achieve the above object, the present invention further provides an online parallel fault diagnosis system for power grid power equipment, comprising:

[0030] The data stream acquisition module receives the power grid feature vector data in real time through the IRichSpout interface and sends it to the Spout source component to form a data stream to be processed;

[0031] The data stream encapsulation module encapsulates the data stream to be processed into multiple tuples and generates a unique ID for each tuple;

[0032] The standardized sample acquisition module uses the PreBolt component to receive the tuple tuple and preprocesses the data set in the tuple tuple using the standard score method to obtain the standardized sample to be classified;

[0033] The fault diagnosis result acquisition module uses a deep classification diagnosis model constructed through unsupervised pre-training, supervised learning, supervised fine-tuning and classifier online learning to process standardized samples to be classified and obtain the fault diagnosis results of the power equipment.

[0034] To achieve the above-mentioned purpose, the present invention also provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and runnable on the processor, wherein the computer program, when executed by the processor, implements the above-mentioned online parallel fault diagnosis method for power grid equipment.

[0035] To achieve the above objectives, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned online parallel fault diagnosis method for power grid equipment.

[0036] According to the solution of the present invention, the classification accuracy of the present invention is relatively high, and the streaming data of power grid equipment can be accurately classified.

[0037] The present invention is suitable for streaming data of power equipment because these data are basically unlabeled data, and the initial model of the present invention does not require a large amount of labeled sample data, but only a small amount of labeled sample data for accurate classification.

[0038] The present invention can realize online learning of the classifier during the data flow diagnosis process, and the classifier output weight matrix can be continuously optimized according to actual needs, and the diagnostic capability of new data is stable and reliable.

[0039] The present invention improves the efficiency of streaming data processing of power grid equipment, that is, deploying the classification model on the Storm platform, and improving the diagnostic efficiency of fault processing by adaptively configuring the number of processes, meta-components and the parallelism of processing components.

[0040] The present invention can diagnose online faults of power data of power grid equipment in real time. By deploying a deep hybrid fault diagnosis model on the Storm platform, it can achieve efficient classification and processing of streaming data of power grid equipment. In addition, this algorithm introduces a multi-layer denoising autoencoder, a multi-layer Gaussian restricted Boltzmann machine, and a deep extreme learning feature extractor. Its parameters are pre-trained by unlabeled input feature vector data, and then the classifier is supervised through labeled samples. Some model parameters are fine-tuned in a supervised manner, and the classifier is used to achieve online learning. These processes enable the network's discrimination performance to be continuously optimized, thereby achieving accurate classification. In addition, this method's monitoring of power equipment fault types can ensure the safe operation of power equipment, reduce losses to residents' production and life, and detect various equipment faults early to avoid catastrophic accidents. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A flowchart schematically illustrates a method for online parallel fault diagnosis of power equipment in a power grid according to an embodiment of the present invention;

[0042] Figure 2 A diagram schematically showing a process of processing power grid characteristic vector data according to an embodiment of the present invention;

[0043] Figure 3 The data access flow chart according to one embodiment of the present invention is schematically shown. DETAILED DESCRIPTION

[0044] The present invention will now be discussed with reference to exemplary embodiments. It should be understood that the embodiments discussed are only for enabling those skilled in the art to better understand and implement the present invention, rather than implying any limitation on the scope of the present invention.

[0045] As used herein, the term "including" and variations thereof are to be interpreted as open-ended terms meaning "including, but not limited to." The term "based on" is to be interpreted as "based, at least in part, on." The terms "one embodiment" and "an embodiment" are to be interpreted as "at least one embodiment."

[0046] Figure 1 The flowchart schematically shows a method for online parallel fault diagnosis of power equipment in a power grid according to an embodiment of the present invention. Figure 1As shown, in this embodiment, the online parallel fault diagnosis method for power grid power equipment includes: receiving power grid feature vector data in real time through the IRichSpout interface and sending it to the Spout source component to form a data stream to be processed;

[0047] Encapsulate the data stream to be processed into multiple tuples and generate a unique ID for each tuple;

[0048] Use the PreBolt component to receive the Tuple tuple and preprocess the data set in the Tuple tuple using the standard score method to obtain standardized samples to be classified;

[0049] The deep classification diagnosis model constructed through unsupervised pre-training, supervised learning, supervised fine-tuning and classifier online learning is used to process standardized samples to obtain fault diagnosis results of power equipment.

[0050] Further, Figure 2 The following diagram schematically shows a process diagram of processing power grid characteristic vector data according to an embodiment of the present invention. Figure 2 As shown, in this implementation, the Storm distributed real-time computing platform is introduced to process streaming data. Storm is a distributed real-time data stream processing platform that can handle massive data streams. The Storm cluster can be dynamically expanded as needed, has high fault tolerance, and can successfully process every piece of data through the ACK mechanism.

[0051] The topology is composed of a Spout source component and several logical processing components Bolt. The data tuple is sent by the Spout and flows to the subsequent Bolt, forming a directed acyclic graph. The present invention is a topology designed for online parallel fault diagnosis of power equipment. Figure 2 As shown in the dotted box. Figure 2 The SendClient is a client designed and implemented by the present invention, which can achieve a certain degree of control by sending instructions to the running Topology.

[0052] Storm uses a master-slave architecture, consisting of two types of nodes: Nimbus and Supervisor. There is only one master node, Nimbus, in the entire cluster, while there can be multiple Supervisor nodes. Nimbus is responsible for distributing code within the cluster. Worker nodes are nodes that run the Supervisor daemon.

[0053] Further, Figure 3 Schematically shows a data access flow chart according to an embodiment of the present invention. Figure 3As shown, in this embodiment, the collected power grid feature vector data is connected to the Spout source component through the IRichSpout interface. The power grid feature vector data is a data stream with no intervals. These feature vector data are continuously sent to the Spout source component, forming a data stream to be processed, that is, the samples to be diagnosed. These samples are encapsulated into multiple tuples, and the Spout source component sends these tuples to the downstream component.

[0054] In this implementation, a tuple represents the tuple of data flowing between components. Each tuple encapsulates an appropriate amount of data (metadata). Here, each tuple encapsulates 1,000 pieces of data, called a dataset. Each tuple encapsulates a dataset, which is then sent to a queue for processing. During subsequent processing, to facilitate the processing of diagnostic results and ensure the sequential processing of tuples, each sent tuple, or the corresponding dataset within each tuple, is tagged with a unique ID. This ID indicates the position of the tuple, or dataset within the tuple, in the data stream.

[0055] Furthermore, according to one embodiment of the present invention, the downstream preprocessing component PreBolt receives a tuple sent by the upstream Spout source component, and the tuple encapsulates the sample to be diagnosed, and each tuple contains 1,000 feature vector data. The preprocessing component PreBolt preprocesses the feature vector data in the received tuple. In this embodiment, taking the oil chromatography data for transformer fault diagnosis as an example, the five gas contents of H2, CH4, C2H6, C2H2, and C2H4 are selected as input and preprocessed. The preprocessing here is to process these data using the standard score method. This method can reduce the skewness of sample data and improve the accuracy of diagnosis. The formula is as follows:

[0056]

[0057] Among them, x new is the standardized gas content value, x is the original gas content value, μ is the average gas content in the training set or test set X, and σ is the standard deviation of the gas content in X;

[0058] After each tuple of data in the preprocessing component PreBolt is preprocessed, the downstream component receives a tuple passed from the preprocessing component PreBolt. The data in this tuple is a standardized deformation of the data set corresponding to the corresponding ID number in the tuple passed to the preprocessing component PreBolt. It is called a standardized sample to be classified. These samples participate in subsequent fault diagnosis processing and are finally summarized by number.

[0059] Furthermore, according to one embodiment of the present invention, a method for constructing a deep classification diagnosis model includes: respectively deploying a multi-layer denoising autoencoder, a multi-layer Gaussian restricted Boltzmann machine, a multi-layer extreme learning machine autoencoder, and an online sequence extreme learning machine in different Bolt components of the Storm platform to obtain a DAEBolt component, a GRBMBolt component, a DELEFBolt component, and an OS-ELMBolt component;

[0060] Connect the DAEBolt component, GRBMBolt component, DELEFBolt component and OS-ELMBolt component in sequence to obtain the deep classification diagnosis model deployed on the Storm platform.

[0061] Furthermore, according to one embodiment of the present invention, unsupervised pre-training of a deep classification diagnosis model includes:

[0062] Connect historical power grid data to the Spout source component of the Storm platform through the IRichSpout interface, and encapsulate the historical power grid data into multiple tuples. Each tuple is configured with a unique ID.

[0063] The PreBolt component is used to receive tuples and preprocess the data set in the tuples using the standard score method to obtain model training samples. In this embodiment, in order to prevent data set skew, when collecting data, for example, oil chromatography detection data of the same model transformers before and after failure at multiple engineering sites are selected. These data contain normal data and fault data and are unlabeled samples, which are used as model training samples.

[0064] The model training samples are input into the deep classification diagnosis model, and the unlabeled samples are used to perform unsupervised pre-training on the DAEBolt component, GRBMBolt component, and DELEFBolt component in the deep classification diagnosis model;

[0065] The DAEBolt component encapsulates a multi-layer denoising autoencoder. The DAEBolt component receives model training samples as input for multi-layer feature extraction. The output of each layer of the multi-layer denoising autoencoder is used as the input of the next layer for unsupervised pre-training.

[0066] The GRBMBolt component encapsulates a multi-layer Gaussian restricted Boltzmann machine. The GRBMBolt component receives the output of the trained DAEBolt component as its input for multi-layer feature extraction. The output of each layer of the multi-layer Gaussian restricted Boltzmann machine is used as the input of the lower layer for unsupervised pre-training. The GRBMBolt component uses the contrastive divergence algorithm for training.

[0067] The DELEFBolt component encapsulates a multi-layer extreme learning machine autoencoder. The DELEFBolt component receives the output of the trained GRBMBolt component as its input for multi-layer feature extraction. The output of each hidden layer between the multi-layer extreme learning machine autoencoders is used as the input of the next hidden layer for unsupervised pre-training.

[0068] In this embodiment, the denoising autoencoder pre-training (unsupervised pre-training) process includes:

[0069] The denoising autoencoder (DAE) is trained using an unsupervised method, with the goal of making the output reproduce the input signal as much as possible, reducing the impact of measurement errors, and then obtaining a relatively optimal initialization parameter θ1 = {W 1 ,b 1}.

[0070] The structure of the denoising autoencoder mainly consists of two parts: the encoder and the decoder.

[0071] The encoder and decoder functions are as follows:

[0072] y=f θ (x)=s(Wx+b) (2)

[0073] a=f θ′ (y) = s(W′y + b) (3)

[0074] Where s is a nonlinear function, such as the sigmoid function. θ = {W, b}, θ′ = {W′, b′} is a parameter set, where W and W′ satisfy W′ = WT.

[0075] The encoder transforms the original input data x into y, and the decoder uses the transformation formula (3) to reversely transform the activation value y into a reconstructed representation z of the original input x. Finally, the BP algorithm is used to update the parameters W,b so that the average reconstruction error C is minimized. The cross entropy loss function is defined as:

[0076]

[0077] When the probability distributions of the test and training samples differ significantly, the autoencoder performs poorly. To improve its robustness, this implementation uses a denoising autoencoder. This denoising autoencoder overcomes some of the shortcomings of traditional autoencoders, enabling the hidden layer to learn a more robust representation, preventing it from simply learning an equivalent representation, and improving the robustness and generalization of the system.

[0078] The DAEBolt component encapsulates a multi-layer denoising autoencoder (DAE). This component receives the output of the upstream preprocessing component as its input for multi-layer feature extraction. The output of each layer between the multi-layer DAEs is used as the input of the lower layer for training.

[0079] In this embodiment, the Gaussian restricted Boltzmann machine pre-training process includes:

[0080] The Gaussian restricted Boltzmann machine is also trained using an unsupervised method. Using probabilistic real values as input can more effectively reconstruct and reproduce the useful information in the input signal and obtain better parameters θ. n , thereby improving the network classification accuracy.

[0081] A Gaussian restricted Boltzmann machine is a special type of restricted Boltzmann machine (RBM) in which the visible units have a Gaussian distribution, while the hidden units typically have a binary or Bernoulli distribution. This configuration makes the Gaussian restricted Boltzmann machine (GRBM) more suitable for processing continuous-valued data.

[0082] GRBMs consist of two layers: a layer of visible units (v) and a layer of hidden units (h). Like standard RBMs, the layers in GRBMs are fully connected, meaning that every two visible and hidden units are connected, but there are no connections between units within the same layer. The units in the visible layer receive input data, while the units in the hidden layer are responsible for capturing high-level features of the visible units.

[0083] As an energy-based probability model, the joint probability distribution of the visual layer v and the hidden layer h is determined by the following energy function:

[0084]

[0085] Where: σ i is the visible layer node v i The standard deviation of the corresponding Gaussian noise; network parameters θ={W ij ,a i ,b j}, where W ij is the visible layer node v i With hidden layer node h j The connection weight between i Represents the visible layer node v i Bias of b j Represents the hidden layer node h j The bias of , all parameters are real numbers.

[0086] The joint probability distribution of the state (v, h) can be obtained through the energy function, as shown in formula (6); combined with the structural characteristics of GRBM, the conditional probability distribution of the corresponding hidden layer nodes and visible layer nodes can be obtained, as shown in formulas (7) and (8).

[0087]

[0088] Where: partition function is the normalization factor; is the activation function; N represents Gaussian distribution.

[0089] The goal of GRBM learning is to obtain network parameters θ so that the hidden layer can better represent the visible layer. The contrastive divergence (CD) algorithm proposed by G. Hinton is usually used to quickly update the GRBM model parameters. Using Gibbs sampling, starting from any state in the training sample set, the state of the hidden layer nodes is calculated according to formula (7), and the nodes of this layer are fixed. Then, the visible layer nodes are reconstructed according to formula (8). The update rule of the network parameter θ is obtained as follows:

[0090] θ=θ+ε(<v i h j > data -<v i h j > recon ) (9)

[0091] Where: ε is the learning rate; <v i h j > data is the mathematical expectation of the input data; <v i h j > recon is the mathematical expectation of the reconstructed data.

[0092] The GRBMBolt component encapsulates a multi-layer Gaussian restricted Boltzmann machine (GRBM). After the DAEBolt component is trained, it receives the output of the upstream DAEBolt component as its input for multi-layer feature extraction. The output of each layer of the multi-layer GRBM is used as the input of the next layer for pre-training. This component uses the contrastive divergence algorithm for training.

[0093] In this embodiment, the pre-training process of the deep extreme learning feature extractor (i.e., the multi-layer extreme learning machine autoencoder) includes:

[0094] a) Extreme Learning Machine Training

[0095] The extreme learning machine (ELM) is a single hidden layer feedforward neural network (SLFN). The input layer weights and hidden layer biases are randomly generated. During the execution process, only the number of hidden layer nodes needs to be set, and the weight matrix between the hidden layer and the output layer is obtained through learning or training.

[0096] Specifically, N sample data (P, Q), represents the input matrix, represents the expected output matrix, where X i =[x i1 ,x i2 ,…,xin ] T ∈R n , Y i =[y i1 ,y i2 ,…,y im ] T ∈R m .

[0097] The input weights and biases between the input layer and the hidden layer are:

[0098] ω=[ω1ω2...ω L ] n×L (10)

[0099] B=[b1 b2...b L ] 1×L (11)

[0100] Among them, ω i =[w i1 ,w i2 ,…,w iL ].

[0101] Select a suitable activation function g(x) to map the output layer data from the original space to the feature space of ELM:

[0102]

[0103] The relationship between the hidden layer and the output layer is:

[0104] H N×L β L×m =Q N×m (13) Where β represents the output weight matrix, and the hidden layer H is as follows:

[0105]

[0106] The output weight matrix is obtained using formula (13). In order to ensure the accuracy of the network, the minimum error function is defined. Because the hidden layer output matrix is not a square matrix, it is easy to cause ill-posed problems when solving it using the traditional least squares method. It is usually necessary to add regularization to ensure numerical stability:

[0107]

[0108] Where λ is an adjustable positive parameter. Solve equation (13) to obtain the output weight matrix:

[0109] β=(I / λ+H T H)-1H T Q (16)

[0110] Substituting the calculated output weight matrix into formula (16), we can obtain the output matrix predicted by the model.

[0111] b) Deep Extreme Learning Feature Extractor Training

[0112] An extreme learning machine-based autoencoder (AE) can be trained to replicate its input to its output. Therefore, the AE concept is applied to the ELM, so that the ELM's input data is also used for its output, i.e., the output Q = P, which serves as the extreme learning machine (ELM-AE) for the autoencoder.

[0113] The optimization goal is:

[0114]

[0115] Output weight matrix:

[0116] β=(I / λ+H T H)-1H T P (18)

[0117] In this embodiment, ELM-AE is stacked to construct a deep extreme learning feature extractor, which also uses a layer-by-layer greedy training method to train the network. The input weights of each hidden layer of the deep extreme learning feature extractor are initialized using ELM-AE to perform hierarchical unsupervised pre-training.

[0118] The feature extraction process is:

[0119] Remember H i The output of the i-th hidden layer of the deep extreme learning feature extractor, H0 = P, H0 is used as the input to train the first ELM-AE, and the weight matrix β1 between the hidden layer and the output layer is obtained. Then, the connection parameters between the input layer and the hidden layer are used Instead, calculate the first hidden layer output H1, and then use H1 as the input and target output of the next ELM-AE, and so on for layer-by-layer training.

[0120] Calculated by the formula:

[0121]

[0122] As the depth of training increases, each hidden layer of the deep learning model can extract feature representations from the data that are more abstract and easier to classify than the previous layer.

[0123] The DELEFBolt component encapsulates a deep extreme learning feature extractor. After the GRBMBolt component is trained, it receives the output of the upstream GRBMBolt component as its input for multi-layer feature extraction. The output of each hidden layer between the multi-layer extreme learning machine autoencoders is used as the input of the next hidden layer for pre-training, and the final hidden layer output of the DELEFBolt component is used as the input of the final classifier component OS_ELMBolt for classification training and fault diagnosis. The deep extreme learning feature extractor is different from other deep learning structures. Once the parameters of the previous layer are determined, the parameters of the current layer are also determined, and no iterative supervised fine-tuning is required.

[0124] The number of hidden layers for the stacked denoising autoencoder in the DAEBolt component, the stacked Gaussian restricted Boltzmann machine in the GRBMBolt component, and the deep extreme learning feature extractor in the DELEFBolt component is set according to the actual classification accuracy.

[0125] The purpose of pre-training the DAEBolt, GRBMBolt, and DELEFBolt components is to optimize model parameters through unsupervised learning, deeply learn the hierarchical feature representation of data, alleviate overfitting, and accelerate model convergence, thereby systematically optimizing parameters for the entire network.

[0126] Furthermore, according to one embodiment of the present invention, supervised learning of a deep classification diagnostic model includes:

[0127] After unsupervised pre-training, supervised learning is performed on the OS-ELMBolt component (classifier component). The hidden layer parameters in the DAEBolt component and the GRBMBolt component are kept fixed. The input samples are continuously subjected to multi-layer feature extraction through a multi-layer denoising autoencoder, a multi-layer Gaussian restricted Boltzmann machine, and a multi-layer extreme learning machine autoencoder. The final hidden layer output extracted by the DELEFBolt component is used as the feature input to the OS-ELMBolt component.

[0128] The OS-ELMBolt component performs supervised learning using labeled samples. An online sequence extreme learning machine (OS-ELM) is deployed within the OS-ELMBolt component, which is responsible for the final classification diagnosis of the entire deep classification diagnosis model. The OS-ELMBolt component receives the output of the DELEFBolt component as its input and adjusts the output layer weights of the online sequence extreme learning machine through least squares calculation based on the input data, thereby completing the supervised classification training of the online sequence extreme learning machine.

[0129] In this embodiment, the algorithm principle of OS-ELM is as follows:

[0130] 1) Online Sequential Extreme Learning Machine

[0131] The initial training data set is N0, and each sample is (X i ,Y i ), where X i =[x i1 ,x i2 ,…,x in ] T ∈R n , Y i =[y i1 ,y i2 ,…,y im ] T ∈R m The hidden layer output function is g(x), and the number of hidden layer nodes is L. The algorithm is divided into the initialization phase and the online sequence learning phase.

[0132] Step 1: The initialization phase of the algorithm is as follows:

[0133] Randomly initialize the input weights ω of the hidden layer nodes i and bias b i , i=1,...L;

[0134] Obtain the hidden layer output matrix H0 through formula (14);

[0135] Calculate the initial output weights in:

[0136]

[0137] Set the batch size of sample data k=0;

[0138] Step 2: The sequential learning phase of the algorithm is as follows:

[0139] When a new batch of training data arrives, it is assumed to be the k+1th batch of sample data. N K+1 is the number of training samples in the current batch.

[0140] Calculate the hidden layer output matrix H of the current batch of training samples k+1 ,

[0141]

[0142] The category label matrix of the current batch of training samples is:

[0143]

[0144] Calculate the output weight matrix β k+1 ,

[0145]

[0146] Set k=k+1 and return to step 2.

[0147] For a limited number of datasets, if the total number of training samples is N, the number of training samples in the initialization phase of OS-ELM is N0. When N = N0, the OS-ELM algorithm is equivalent to the original ELM algorithm. During the online training phase, the output weight β of the single hidden layer neural network is adjusted using data blocks consisting of one or more samples.

[0148] Furthermore, according to one embodiment of the present invention, the supervised fine-tuning of the deep classification diagnosis model is:

[0149] The back-propagation algorithm is used to perform supervised fine-tuning on the parameters of the DAEBolt and GRBMBolt components.

[0150] In this embodiment, the DAEBolt and GRBMBolt components are pre-trained to obtain a relatively optimal initialization parameter set θ, which has a certain improvement in the accuracy of power equipment fault diagnosis. However, the parameter set θ is not optimal, and the diagnostic accuracy can still be improved through fine-tuning.

[0151] After the output weight matrix of the OS-ELM algorithm within the OS-ELMBolt component is determined through supervised learning, the backpropagation (BP) algorithm is used to further optimize the parameters θ within the DAEBolt and GRBMBolt components. Only the stacked DAE and stacked GRBM algorithm parameters are adjusted, while the parameters of each layer of the Deep Extreme Learning feature extractor and the OS-ELM output weights remain unchanged. The goal of fine-tuning is to further optimize the feature representation by incorporating task label information and improve classification performance.

[0152] After supervised fine-tuning, the deep initial diagnosis model is trained and can perform online fault diagnosis of power grid data streams. When new labeled sample data is accumulated, the diagnosis model can also be learned online.

[0153] Furthermore, according to one embodiment of the present invention, the online learning of the classifier of the deep classification diagnosis model is:

[0154] The dataset formed by the final hidden layer output matrix obtained in the sequence learning phase of the online sequence extreme learning machine and the category label matrix of the training samples is used as input to train the online sequence extreme learning machine in the OS-ELMBolt component.

[0155] In this embodiment, OS-ELM is the final classifier of the present invention, and OS-ELM can realize online learning function. When new training samples are accumulated, such as the k+1 batch of data, the input feature vector data of the newly added k+1 batch of samples is sent to Spout through SendClient. The relevant data is processed by the PreBolt preprocessing component to obtain standardized samples to be classified. The samples are continuously extracted by the multi-layer denoising autoencoder, multi-layer Gaussian restricted Boltzmann machine and deep extreme learning feature extractor encapsulated in the DAEBolt, GRBMBolt and DELEFBolt components to obtain the final feature output. Will The OS-ELM algorithm in the OS-ELMBolt component is trained using this data as input, and the original model is updated according to the algorithm used in the sequential learning phase. Tuple processing can be paused while the model is being updated and resumed after completion. This online model update increases the parallelism of each processing component, resulting in a relatively short processing time. This prevents the accumulation of unprocessed tuples in OS-ELMBolt, and the entire process does not affect the normal operation of the topology.

[0156] Furthermore, according to an embodiment of the present invention, the fault diagnosis result is stored in a database via the DatabaseBolt component, and the fault diagnosis result is stored in a data file via the FileBolt component.

[0157] According to the above solution of the present invention, the classification accuracy of the present invention is relatively high, and the streaming data of power grid equipment can be accurately classified.

[0158] The present invention is suitable for streaming data of power equipment because these data are basically unlabeled data, and the initial model of the present invention does not require a large amount of labeled sample data, but only a small amount of labeled sample data for accurate classification.

[0159] The present invention can realize online learning of the classifier during the data flow diagnosis process, and the classifier output weight matrix can be continuously optimized according to actual needs, and the diagnostic capability of new data is stable and reliable.

[0160] The present invention improves the efficiency of streaming data processing of power grid equipment, that is, deploying the classification model on the Storm platform, and improving the diagnostic efficiency of fault processing by adaptively configuring the number of processes, meta-components and the parallelism of processing components.

[0161] The present invention can diagnose online faults of power data of power grid equipment in real time. By deploying a deep hybrid fault diagnosis model on the Storm platform, it can achieve efficient classification and processing of streaming data of power grid equipment. In addition, this algorithm introduces a multi-layer denoising autoencoder, a multi-layer Gaussian restricted Boltzmann machine, and a deep extreme learning feature extractor. Its parameters are pre-trained by unlabeled input feature vector data, and then the classifier is supervised through labeled samples. Some model parameters are fine-tuned in a supervised manner, and the classifier is used to achieve online learning. These processes enable the network's discrimination performance to be continuously optimized, thereby achieving accurate classification. In addition, this method's monitoring of power equipment fault types can ensure the safe operation of power equipment, reduce losses to residents' production and life, and detect various equipment faults early to avoid catastrophic accidents.

[0162] Furthermore, to achieve the above-mentioned object, the present invention also provides an online parallel fault diagnosis system for power grid power equipment, comprising:

[0163] The data stream acquisition module receives the power grid feature vector data in real time through the IRichSpout interface and sends it to the Spout source component to form a data stream to be processed;

[0164] The data stream encapsulation module encapsulates the data stream to be processed into multiple tuples and generates a unique ID for each tuple;

[0165] The standardized sample acquisition module uses the PreBolt component to receive the tuple tuple and preprocesses the data set in the tuple tuple using the standard score method to obtain the standardized sample to be classified;

[0166] The fault diagnosis result acquisition module uses a deep classification diagnosis model constructed through unsupervised pre-training, supervised learning, supervised fine-tuning and classifier online learning to process standardized samples to be classified and obtain the fault diagnosis results of the power equipment.

[0167] The above-mentioned online parallel fault diagnosis system for power grid electric equipment according to the present invention can implement the above-mentioned online parallel fault diagnosis method for power grid electric equipment. The specific process steps are as described above and will not be repeated here.

[0168] Furthermore, to achieve the above-mentioned purpose, the present invention also provides an electronic device, comprising a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the online parallel fault diagnosis method for power grid equipment as described above is implemented.

[0169] Furthermore, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the above-mentioned online parallel fault diagnosis method for power grid equipment is implemented.

[0170] Those skilled in the art will appreciate that the modules and algorithm steps described in conjunction with the embodiments disclosed herein can be implemented using electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0171] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described devices and equipment can refer to the corresponding processes in the aforementioned method implementation methods and will not be repeated here.

[0172] In the embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the modules is merely a logical function division. In actual implementation, there may be other division methods, such as multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or modules, which can be electrical, mechanical or other forms.

[0173] The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules, that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the objectives of the embodiments of the present invention.

[0174] In addition, each functional module in the embodiment of the present invention may be integrated into one processing module, or each module may exist physically separately, or two or more modules may be integrated into one module.

[0175] If the functions are implemented as software modules and sold or used as standalone products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the energy-saving signal transmission / reception method according to various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, ROM, RAM, a magnetic disk, or an optical disk.

[0176] The above description is merely a preferred embodiment of the present application and an illustration of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to the technical solutions formed by the specific combination of the above-mentioned technical features, but also encompasses other technical solutions formed by any combination of the above-mentioned technical features or their equivalents without departing from the inventive concept. For example, a technical solution formed by replacing the above-mentioned features with (but not limited to) technical features with similar functions disclosed in this application.

[0177] It should be understood that the size of the serial numbers of each step in the content of the invention and the implementation methods of the present invention does not absolutely mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the implementation methods of the present invention.

Claims

1. A method for online parallel fault diagnosis of power grid equipment, characterized in that: include: Receive power grid feature vector data in real time through the IRichSpout interface and send it to the Spout source component to form a data stream to be processed; Encapsulate the data stream to be processed into multiple tuples and generate a unique ID for each tuple; Use the PreBolt component to receive the Tuple tuple and preprocess the data set in the Tuple tuple using the standard score method to obtain standardized samples to be classified; The deep classification diagnosis model constructed through unsupervised pre-training, supervised learning, supervised fine-tuning and classifier online learning is used to process standardized samples to obtain fault diagnosis results of power equipment.

2. The online parallel fault diagnosis method for power grid equipment according to claim 1, characterized in that: The method for constructing the deep classification diagnosis model includes: respectively deploying a multi-layer denoising autoencoder, a multi-layer Gaussian restricted Boltzmann machine, a multi-layer extreme learning machine autoencoder, and an online sequence extreme learning machine in different Bolt components of the Storm platform to obtain a DAEBolt component, a GRBMBolt component, a DELEFBolt component, and an OS-ELMBolt component; Connect the DAEBolt component, GRBMBolt component, DELEFBolt component and OS-ELMBolt component in sequence to obtain the deep classification diagnosis model deployed on the Storm platform.

3. The online parallel fault diagnosis method for power grid equipment according to claim 2, characterized in that: The unsupervised pre-training of the deep classification diagnostic model includes: Connect historical power grid data to the Spout source component of the Storm platform through the IRichSpout interface, and encapsulate the historical power grid data into multiple tuples. Each tuple is configured with a unique ID. Use the PreBolt component to receive the Tuple tuple and preprocess the data set in the Tuple tuple using the standard score method to obtain model training samples; The model training samples are input into the deep classification diagnosis model, and the unlabeled samples are used to perform unsupervised pre-training on the DAEBolt component, GRBMBolt component, and DELEFBolt component in the deep classification diagnosis model; The DAEBolt component encapsulates a multi-layer denoising autoencoder, which receives model training samples as input for multi-layer feature extraction. The output of each layer of the multi-layer denoising autoencoder is used as the input of the next layer for unsupervised pre-training. The GRBMBolt component encapsulates a multi-layer Gaussian restricted Boltzmann machine. The GRBMBolt component receives the output of the trained DAEBolt component as its input for multi-layer feature extraction. The output of each layer between the multi-layer Gaussian restricted Boltzmann machines is used as the input of the lower layer for unsupervised pre-training. The GRBMBolt component uses the contrastive divergence algorithm for training; The DELEFBolt component encapsulates a multi-layer extreme learning machine autoencoder. The DELEFBolt component receives the output of the trained GRBMBolt component as the input of this component to perform multi-layer feature extraction. The output of each hidden layer between the multi-layer extreme learning machine autoencoders is used as the input of the next hidden layer for unsupervised pre-training.

4. The online parallel fault diagnosis method for power grid equipment according to claim 2, characterized in that: The supervised learning of the deep classification diagnostic model includes: After unsupervised pre-training, supervised learning is performed on the OS-ELMBolt component. The hidden layer parameters in the DAEBolt and GRBMBolt components are kept fixed. The input samples are continuously subjected to multi-layer feature extraction through a multi-layer denoising autoencoder, a multi-layer Gaussian restricted Boltzmann machine, and a multi-layer extreme learning machine autoencoder. The final hidden layer output extracted by the DELEFBolt component is used as the feature input to the OS-ELMBolt component. The OS-ELMBolt component performs supervised learning using labeled samples. An online sequence extreme learning machine is deployed within the OS-ELMBolt component, which is responsible for the final classification diagnosis of the entire deep classification diagnosis model. The OS-ELMBolt component receives the output of the DELEFBolt component as its input, and adjusts the output layer weights of the online sequence extreme learning machine through least squares calculation based on the input data, thereby completing the supervised classification training of the online sequence extreme learning machine.

5. The online parallel fault diagnosis method for power grid equipment according to claim 2, characterized in that: The supervised fine-tuning of the deep classification diagnosis model is: The back-propagation algorithm is used to perform supervised fine-tuning on the parameters of the DAEBolt and GRBMBolt components.

6. The online parallel fault diagnosis method for power grid equipment according to claim 2, characterized in that: The online learning of the classifier of the deep classification diagnosis model is: The dataset formed by the final hidden layer output matrix obtained by the training samples in the sequential learning phase of the online sequential extreme learning machine and the category label matrix of the training samples is used as input to train the online sequential extreme learning machine in the OS-ELMBolt component.

7. The method for online parallel fault diagnosis of power equipment in a power grid according to any one of claims 1 to 6, characterized in that: The fault diagnosis result is stored in a database via the DatabaseBolt component, and the fault diagnosis result is stored in a data file via the FileBolt component.

8. Online parallel fault diagnosis system for power grid equipment, characterized by: include: The data stream acquisition module receives the power grid feature vector data in real time through the IRichSpout interface and sends it to the Spout source component to form a data stream to be processed; The data stream encapsulation module encapsulates the data stream to be processed into multiple tuples and generates a unique ID for each tuple; The standardized sample acquisition module uses the PreBolt component to receive the tuple tuple and preprocesses the data set in the tuple tuple using the standard score method to obtain the standardized sample to be classified; The fault diagnosis result acquisition module uses a deep classification diagnosis model constructed through unsupervised pre-training, supervised learning, supervised fine-tuning and classifier online learning to process standardized samples to be classified and obtain the fault diagnosis results of the power equipment.

9. An electronic device, characterized in that The invention comprises a processor, a memory and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed by the processor, the method for online parallel fault diagnosis of power equipment in a power grid as claimed in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method for online parallel fault diagnosis of power equipment in a power grid according to any one of claims 1 to 7 is implemented.

Images