SCD (Substation Configuration Description) file standardization detection report generation method based on correctness check

Abstract

The invention discloses an SCD (Substation Configuration Description) file standardized detection report generation method based on correctness check, which is characterized in that a CNN (Convolutional Neural Network) part and an RNN (Recurrent Neural Network) part in a CRNN model are respectively improved into a Transform model and an LSTM (Long Short Term Memory) network model based on the thought of the CRNN model, and finally an SCD file standardized detection report is generated. According to the method, a large amount of manual tedious work can be saved, repeated work is templated to the maximum extent, variable work is automated, the working efficiency is improved, full-automatic inspection of secondary equipment configuration and virtual loop connection and automatic generation of an SCD file standardized detection report are achieved, and the method is suitable for popularization and application. The intelligent substation SCD file review and detection efficiency is effectively improved, the review comprehensiveness and accuracy are improved, and the secondary equipment virtual loop detection time is shortened.

Description

Technical Field

[0001] The invention belongs to the technical field of intelligent substation fault detection, and in particular relates to a design of a SCD file standardized detection report generation method based on correctness checking. Background Art

[0002] Existing fault detection technology depends largely on the decision-making of technicians and experts, but the analysis conducted by experts is easily affected by fatigue, stress, knowledge level, psychological quality, skill level, etc. The analysis results are far from the actual situation, resulting in misjudgment. At present, fault detection technology is mainly divided into three categories: fault detection technology based on signal processing, fault detection technology based on analytical model, and fault detection technology based on artificial intelligence. The most systematic and earliest fault detection method is the fault detection method based on analytical model, which can deeply understand the dynamic characteristics of the system and perform real-time diagnosis. It does not rely on historical data or known fault data. However, it is difficult to obtain the system model. At the same time, the robustness problem caused by modeling interference, error and noise gradually emerges. It is becoming more and more difficult to establish an accurate mathematical model. Since the fault detection method based on signal processing does not require an accurate mathematical model, it avoids the difficulties of the fault diagnosis method based on analytical model. Based on the measured signal data, this method of judging faults based on signal characteristic values ​​is basically independent of the model and is applicable to nonlinear and linear systems. However, it only analyzes the data of the device under test, does not analyze the coupling relationship and correlation between high-dimensional signals in the system, and does not further use the deep measurement information. Continuous technological progress has effectively improved the acquisition, processing, analysis, transmission, storage and application of system and equipment data. Deep learning can quickly process large amounts of data and analyze the advantages of useful information. It can detect difficult-to-detect faults and effectively improve the accuracy of fault detection. Therefore, deep learning is increasingly used for fault detection in various fields.

[0003] Machine learning is divided into three categories: supervised learning, semi-supervised learning, and unsupervised learning. As a technology for implementing machine learning, deep learning has enabled machine learning to achieve more applications. Fault detection methods based on deep learning have the following advantages over traditional fault detection methods: (1) Deep learning has strong self-learning capabilities and can avoid relying on expert diagnostic experience; (2) Deep learning's own classifier can extract feature parameters and perform identification, reducing the impact of human selection. Machine learning applications are very extensive, and its applications in the field of fault detection mainly include neural networks, decision trees, support vector machines, etc. Among them, neural networks have nonlinear mapping capabilities, self-learning capabilities, parallel computing capabilities, the ability to approximate arbitrary functions, and fault tolerance capabilities. They can perform fault detection based on massive data collected during the operation of equipment or systems.

[0004] At present, the main types of neural networks are: (1) Convolutional neural networks (CNN) and deep convolutional neural networks (DCNN), which can directly input raw data without tedious pre-processing of complex signals such as language and images, and can learn to mine features at different levels; (2) Recurrent neural networks, which can process list and sequence data, mine time series information in time series data, and use previously input information to enhance the judgment of current information, and have strong semantic mining and expression capabilities; (3) Deep belief networks (DBN), which are constructed by superimposing multiple restricted Boltzmann machines (RBMs). Compared with traditional fault detection methods, they do not require much diagnostic experience and signal technology, and have relatively strong versatility, adaptability, and the ability to process nonlinear and high-dimensional data. In order to further improve the accuracy of the model, it is possible to consider combining the neural network algorithm with other algorithms, such as combining artificial neural networks with gray models to form gray neural networks.

[0005] At present, deep learning has limitations in fault detection. In equipment fault detection, it is mostly necessary to know the equipment fault type and then classify it through neural networks. With the development of technology, the objects to be diagnosed are becoming more and more complex, and it will become more and more difficult to obtain effective and accurate fault types. New fault types cannot be predicted. Unsupervised clustering algorithms have good scalability, and the combination of clustering algorithms and deep learning has received widespread attention. Combining fuzzy C-means clustering and convolutional neural networks to predict the photovoltaic output at the minute level before the day, the prediction accuracy and robustness have been successfully improved. After using K-means clustering, BPNN models are constructed according to each category to predict the air conditioning load of subway stations hour by hour. The prediction accuracy is better than that of continuously optimizing and updating the BPNN model structure using optimization algorithms, and the effect is better. The neighbor propagation (AP) clustering algorithm is used to classify sample data, and the long short-term memory network model is established to realize the grid-connected power prediction of solar thermal power stations. The K-means clustering center is used as the pattern layer neuron to train the probabilistic neural network. The complexity of the model is greatly reduced, and the time for fault detection is greatly reduced.

[0006] In the face of fault detection in smart grids, a feature extractor is constructed using a convolutional neural network to extract high-level features from time series data. Then, the extracted features are semi-supervised clustered to obtain sample labels. After sampling the fault samples through a sampling algorithm, a classifier constructed using a recurrent neural network is used for classification and identification. Using time series data clustering, in preparation for further research on time series data, it is proposed to use a long short-term memory network (LSTM) to learn the intrinsic connections of time series data, and then use different distance measurement k-means clustering algorithms to cluster the LSTM learning results.

[0007] In order to enhance the scalability of the fault detection model, ensure high diagnostic accuracy for existing fault types, and realize the identification and diagnosis of new faults, a clustering algorithm is introduced to cluster the fault data diagnosed by the neural network model with similar fault data, and finally use the centroid distance to determine whether it is a new fault.

[0008] In the review process of SCD (Substation Configuration Description) files of smart substations, there is currently no comprehensive and reliable technical means to realize the intelligent inspection of the correctness of the secondary virtual circuit of SCD files. The most important information such as secondary circuit wiring and terminal connection information that the on-site electrical personnel focus on is not checked and can only be implemented manually. The disadvantages are that it is time-consuming and difficult to ensure the accuracy. Manual review and inspection of each item is time-consuming and labor-intensive, and the accuracy and comprehensiveness cannot be guaranteed, which brings many inconveniences to the operation and maintenance, expansion and other work of smart substations, and brings hidden dangers to safe operation. Summary of the invention

[0009] The purpose of the present invention is to solve the problem that the existing SCD file review method for smart substations is time-consuming and labor-intensive and the accuracy and comprehensiveness cannot be guaranteed. A method for generating a standardized test report for SCD files based on correctness checking is proposed.

[0010] The technical solution of the present invention is: a method for generating a SCD file standardized test report based on correctness check, comprising the following steps:

[0011] S1. Input the SCD file into the Transformer model to extract text features, and output the text features of the SCD file.

[0012] S2. Input the text features of the SCD file into the LSTM network model and output the prediction vector of the text elements of the SCD file.

[0013] S3. Construct a CTC loss function based on the prediction vector of the text element of the SCD file, perform end-to-end joint training on the Transformer model and the LSTM network model based on the CTC loss function, and obtain the probability distribution vector of the text element of the SCD file with the goal of minimizing the CTC loss function.

[0014] S4. Select the token with the highest predicted probability in the probability distribution vector and convert it into a string to obtain a standardized detection report of the SCD file.

[0015] Furthermore, step S1 includes the following sub-steps:

[0016] S11. Get the representation vector of each word in the SCD file.

[0017] S12. Input the representation vector of the word into the encoder of the Transformer model to obtain the encoding information matrix of the word.

[0018] S13. Input the word encoding information matrix into the decoder of the Transformer model and decode to obtain the text features of the SCD file.

[0019] Furthermore, in step S11, the representation vector of the word is obtained by adding the embedding of the word and the embedding of the word position.

[0020] Furthermore, the embedding of word positions is represented by positional encoding, and the calculation formula is:

[0021] PE (pos,2i) =sin(pos / 10000 2i / d )

[0022] PE (pos,2i+1) =cos(pos / 10000 2i / d )

[0023] Where PE represents position encoding, pos represents the position of the word in the sentence, d represents the dimension of the position encoding, 2i represents even dimension, 2i+1 represents odd dimension, and 2i+1≤d.

[0024] Furthermore, in step S12, the encoder includes 6 encoder blocks connected in sequence, each encoder block has the same structure, and includes a self-attention layer and a feedforward neural network.

[0025] Furthermore, in step S13, the decoder includes a masked self-attention layer, an encoder-decoder attention layer, and a feedforward neural network connected in sequence.

[0026] Furthermore, in step S13, the decoder decodes the i+1th word in sequence according to the currently decoded 1st to ith words, and when decoding to the i+1th word, the decoder covers the words after the i+1th word through a masking operation.

[0027] Further, step S2 includes the following sub-steps:

[0028] S21. Input the text features of the SCD file into the LSTM network model, determine the information that needs to be discarded through the forget gate, and determine the information that needs to be retained through the input gate:

[0029] f t =σ(W fi x t +W fh h t-1 +b f )

[0030] i t =σ(W xi x t +W xh h t-1 +b i )

[0031] where f t represents the output of the forget gate at the current time step, σ represents the sigmoid activation function, W fi and W fh are the weight matrices of the forget gate, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, b f is the bias term of the forget gate, i t represents the output of the input gate at the current time step, W xi and W xh are the weight matrices of the input gate, b i is the bias term of the input gate.

[0032] S22. Input the information retained by the input gate to the output gate, determine the hidden state of the current time step and obtain the prediction vector of the text element of the SCD file:

[0033] C t =tanh(W xc x t +W hc h t-1 +b c )

[0034] o t =σ(W xo x t +W xh h t-1 +b o )

[0035] h t =o t *tanh(C t )

[0036] Among them C t Represents the memory cell state at the current time step, W xc and W hc are all weight matrices of memory units, b c is the bias term of the memory unit, o t Represents the prediction vector of the text element of the SCD file output by the output gate at the current time step, W xo and W xh are the weight matrices of the output gate, b o is the bias term of the output gate, ht Represents the hidden state at the current time step.

[0037] Further, step S3 includes the following sub-steps:

[0038] S31. Transcribe the prediction vector of the text element of the SCD file into a label sequence.

[0039] S32. Construct the CTC loss function based on the prediction vector of the text element of the SCD file and the corresponding label sequence:

[0040] L(y)=-log(p(l|x))

[0041] Where L(y) represents the CTC loss function, x represents the prediction vector of the text element of the SCD file, l represents the label sequence, and p(l|x) represents the probability of observing the label sequence l given the prediction vector x.

[0042] S33. The prediction vector corresponding to the minimum CTC loss function value is used as the probability distribution vector of the text element of the SCD file.

[0043] The beneficial effects of the present invention are as follows: the present invention is based on the idea of ​​the CRNN model and uses artificial intelligence to carry out SCD detection, which can save a lot of redundant manual work, template repetitive work to the maximum extent, automate variable work, improve work efficiency, realize full-automatic inspection of secondary equipment configuration and virtual circuit connection and automatic generation of SCD file standardized detection report, effectively improve the SCD file review and detection efficiency of smart substations, increase the comprehensiveness and accuracy of the review, and reduce the virtual circuit detection time of secondary equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 The figure is a flow chart of a method for generating a standardized detection report of an SCD file based on correctness checking provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] Now, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. It should be understood that the embodiments shown and described in the accompanying drawings are only exemplary and are intended to explain the principles and spirit of the present invention, rather than to limit the scope of the present invention.

[0046] The embodiment of the present invention provides a method for generating a SCD file standardized test report based on correctness checking, such as Figure 1 As shown, the following steps S1 to S4 are included:

[0047] S1. Input the SCD file into the Transformer model to extract text features, and output the text features of the SCD file.

[0048] Step S1 includes the following sub-steps S11 to S13:

[0049] S11. Get the representation vector of each word in the SCD file.

[0050] In the embodiment of the present invention, the representation vector of a word is obtained by adding the embedding of the word and the embedding of the word position, where the embedding refers to features extracted from the original data.

[0051] In the embodiment of the present invention, the embedding of the word position is represented by position coding, and the calculation formula is:

[0052] PE (pos,2i) =sin(pos / 10000 2i / d )

[0053] PE (pos,2i+1) =cos(pos / 10000 2i / d )

[0054] Wherein PE represents position encoding, pos represents the position of a word in a sentence, d represents the dimension of the position encoding (which is the same as the embedding dimension of a word in the embodiment of the present invention), 2i represents an even dimension, 2i+1 represents an odd dimension, and 2i+1≤d.

[0055] In addition to word embedding, Transformer also needs to use position embedding to indicate the position of a word in a sentence. Because Transformer does not use the RNN structure, but uses global information, it cannot use the order information of words, which is very important for natural language processing. Therefore, Transformer uses position embedding to save the relative or absolute position of a word in a sequence.

[0056] S12. Input the representation vector of the word into the encoder of the Transformer model to obtain the encoding information matrix of the word.

[0057] In an embodiment of the present invention, the encoder includes 6 encoder blocks connected in sequence, each encoder block has the same structure, including a self-attention layer and a feedforward neural network, and the matrix dimension of each encoder block output is exactly the same as the input.

[0058] S13. Input the word encoding information matrix into the decoder of the Transformer model and decode to obtain the text features of the SCD file.

[0059] In an embodiment of the present invention, the decoder includes a masked self-attention layer, an encoder-decoder attention layer, and a feedforward neural network connected in sequence.

[0060] The decoder decodes the i+1th word according to the currently decoded 1st to ith words in sequence, and when decoding to the i+1th word, it covers the words after the i+1th word through a masking operation.

[0061] The core of Transformer is the self-attention mechanism, which dynamically assigns attention weights to other positions at each position in the sequence. In this way, the model can effectively capture the dependencies between words in a sentence, no matter how far apart they are. At the same time, Transformer also uses a multi-head attention mechanism, which means that multiple self-attention mechanisms are calculated in parallel, and each head processes different parts independently. Finally, the outputs of these heads are spliced ​​together to improve the model's ability to capture different semantic relationships.

[0062] S2. Input the text features of the SCD file into the LSTM network model and output the prediction vector of the text elements of the SCD file.

[0063] Since LSTM is unidirectional, it only uses past information. However, in image-based sequences, the contexts in both directions are mutually useful and complementary, so the embodiment of the present invention combines two LSTMs, one forward and one backward, into a bidirectional LSTM. In addition, multiple layers of bidirectional LSTM can be stacked, and the deep structure allows for a higher level of abstraction than shallow abstractions. The embodiment of the present invention uses a two-layer bidirectional LSTM network model with 256 units each.

[0064] Step S2 includes the following sub-steps S21-S22:

[0065] S21. Input the text features of the SCD file into the LSTM network model, determine the information that needs to be discarded through the forget gate, and determine the information that needs to be retained through the input gate:

[0066] f t =σ(W fi x t +W fh h t-1 +b f )

[0067] i t =σ(W xi x t +W xh h t-1 +b i )

[0068] where f t represents the output of the forget gate at the current time step, σ represents the sigmoid activation function, W fi and W fh are the weight matrices of the forget gate, xt represents the input of the current time step, h t-1 represents the hidden state of the previous time step, b f is the bias term of the forget gate, i t represents the output of the input gate at the current time step, W xi and W xh are the weight matrices of the input gate, b i is the bias term of the input gate.

[0069] S22. Input the information retained by the input gate to the output gate, determine the hidden state of the current time step and obtain the prediction vector of the text element of the SCD file:

[0070] C t =tanh(W xc x t +W hc h t-1 +b c )

[0071] o t =σ(W xo x t +W xh h t-1 +b o )

[0072] h t =o t *tanh(C t )

[0073] Among them C t Represents the memory cell state at the current time step, W xc and W hc are all weight matrices of memory units, b c is the bias term of the memory unit, o t Represents the prediction vector of the text element of the SCD file output by the output gate at the current time step, W xo and W xh are the weight matrices of the output gate, b o is the bias term of the output gate, h t Represents the hidden state at the current time step.

[0074] S3. Construct a CTC (Connectionist Temporal Classification) loss function based on the prediction vector of the text element of the SCD file. Perform end-to-end joint training on the Transformer model and the LSTM network model based on the CTC loss function. Obtain the probability distribution vector of the text element of the SCD file with the goal of minimizing the CTC loss function.

[0075] Step S3 includes the following sub-steps S31 to S33:

[0076] S31. Transcribe the prediction vector of the text element of the SCD file into a label sequence.

[0077] Mathematically, transcription refers to finding the label sequence with the highest probability combination based on the predictions for each frame.

[0078] S32. Construct the CTC loss function based on the prediction vector of the text element of the SCD file and the corresponding label sequence:

[0079] L(y)=-log(p(l|x))

[0080] Where L(y) represents the CTC loss function, x represents the prediction vector of the text element of the SCD file, l represents the label sequence, and p(l|x) represents the probability of observing the label sequence l given the prediction vector x.

[0081] S33. The prediction vector corresponding to the minimum CTC loss function value is used as the probability distribution vector of the text element of the SCD file.

[0082] S4. Select the token with the highest prediction probability in the probability distribution vector and convert it into a string to obtain the true prediction result, so as to check the SCD file and obtain the SCD file standardization detection report.

[0083] Those skilled in the art will appreciate that the embodiments described herein are intended to help readers understand the principles of the present invention, and should be understood that the protection scope of the present invention is not limited to such specific statements and embodiments. Those skilled in the art can make various other specific variations and combinations that do not deviate from the essence of the present invention based on the technical revelations disclosed by the present invention, and these variations and combinations are still within the protection scope of the present invention.

Claims

1. A method for generating a standardized test report of an SCD file based on correctness checking, characterized in that: The following steps are involved: S1. Input the SCD file into the Transformer model to extract text features, and output the text features of the SCD file; S2, input the text features of the SCD file into the LSTM network model, and output the prediction vector of the text element of the SCD file; S3. Construct a CTC loss function based on the prediction vector of the text element of the SCD file, perform end-to-end joint training on the Transformer model and the LSTM network model based on the CTC loss function, and obtain the probability distribution vector of the text element of the SCD file with the goal of minimizing the CTC loss function; S4. Select the token with the highest predicted probability in the probability distribution vector and convert it into a string to obtain a standardized detection report of the SCD file.

2. The method for generating a SCD file standardization test report based on correctness check according to claim 1, characterized in that: The step S1 comprises the following sub-steps: S11, obtain the representation vector of each word in the SCD file; S12, input the representation vector of the word into the encoder of the Transformer model to obtain the encoding information matrix of the word; S13. Input the word encoding information matrix into the decoder of the Transformer model and decode to obtain the text features of the SCD file.

3. The method for generating a SCD file standardization test report based on correctness check according to claim 2, characterized in that: The representation vector of the word in step S11 is obtained by adding the embedding of the word and the embedding of the word position.

4. The method for generating a SCD file standardization test report based on correctness check according to claim 3 is characterized in that: The embedding of the word position is represented by position encoding, and the calculation formula is: ON (pos,2i) =sin(pos / 10000 2i / d ) ON (pos,2i+1) =cos(pos / 10000 2i / d ) Where PE represents position encoding, pos represents the position of the word in the sentence, d represents the dimension of the position encoding, 2i represents even dimension, 2i+1 represents odd dimension, and 2i+1≤d.

5. The method for generating a SCD file standardization test report based on correctness check according to claim 2, characterized in that: The encoder in step S12 includes 6 encoder blocks connected in sequence, each of which has the same structure and includes a self-attention layer and a feedforward neural network.

6. The method for generating a SCD file standardization test report based on correctness check according to claim 2, characterized in that: The decoder in step S13 includes a masked self-attention layer, an encoder-decoder attention layer and a feedforward neural network connected in sequence.

7. The method for generating a SCD file standardization test report based on correctness check according to claim 2, characterized in that: In step S13, the decoder decodes the i+1th word in sequence according to the currently decoded 1st to ith words, and when decoding the i+1th word, it covers the words after the i+1th word through a masking operation.

8. The method for generating a SCD file standardization test report based on correctness check according to claim 1, characterized in that: The step S2 comprises the following sub-steps: S21. Input the text features of the SCD file into the LSTM network model, determine the information that needs to be discarded through the forget gate, and determine the information that needs to be retained through the input gate: f t =σ(W fi x t +W fh h t-1 +b f ) i t =σ(W xi x t +W xh h t-1 +b i ) where f t represents the output of the forget gate at the current time step, σ represents the sigmoid activation function, W fi and W fh are the weight matrices of the forget gate, x t represents the input of the current time step, h t-1 represents the hidden state of the previous time step, b f is the bias term of the forget gate, i t represents the output of the input gate at the current time step, W xi and W xh are the weight matrices of the input gate, b i is the bias term of the input gate; S22. Input the information retained by the input gate to the output gate, determine the hidden state of the current time step and obtain the prediction vector of the text element of the SCD file: C t =tanh(W xc x t +W hc h t-1 +b c ) o t =σ(W xo x t +W xh h t-1 +b o ) h t =o t *tanh(C t ) Among them C t Represents the memory cell state at the current time step, W xc and W hc are all weight matrices of memory units, b c is the bias term of the memory unit, o t Represents the prediction vector of the text element of the SCD file output by the output gate at the current time step, W xo and W xh are the weight matrices of the output gate, b o is the bias term of the output gate, h t Represents the hidden state at the current time step.

9. The method for generating a SCD file standardization test report based on correctness check according to claim 1, characterized in that: The step S3 comprises the following sub-steps: S31, transcribing the prediction vector of the text element of the SCD file into a label sequence; S32. Construct the CTC loss function based on the prediction vector of the text element of the SCD file and the corresponding label sequence: L(y)=-log(p(l|x)) Where L(y) represents the CTC loss function, x represents the prediction vector of the text element of the SCD file, l represents the label sequence, and p(l|x) represents the probability of observing the label sequence l given the prediction vector x; S33. The prediction vector corresponding to the minimum CTC loss function value is used as the probability distribution vector of the text element of the SCD file.

Images