CN117347803A - Partial discharge detection method, system, equipment and medium - Google Patents
Partial discharge detection method, system, equipment and medium Download PDFInfo
- Publication number
- CN117347803A CN117347803A CN202311387592.2A CN202311387592A CN117347803A CN 117347803 A CN117347803 A CN 117347803A CN 202311387592 A CN202311387592 A CN 202311387592A CN 117347803 A CN117347803 A CN 117347803A
- Authority
- CN
- China
- Prior art keywords
- data
- partial discharge
- time sequence
- training
- target
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000001514 detection method Methods 0.000 title claims abstract description 62
- 230000004927 fusion Effects 0.000 claims abstract description 49
- 230000007246 mechanism Effects 0.000 claims abstract description 48
- 238000004364 calculation method Methods 0.000 claims abstract description 47
- 230000015654 memory Effects 0.000 claims abstract description 42
- 230000009466 transformation Effects 0.000 claims abstract description 35
- 230000002457 bidirectional effect Effects 0.000 claims abstract description 33
- 238000000034 method Methods 0.000 claims abstract description 30
- 238000000605 extraction Methods 0.000 claims abstract description 19
- 238000013507 mapping Methods 0.000 claims abstract description 14
- 238000012549 training Methods 0.000 claims description 93
- 230000006870 function Effects 0.000 claims description 32
- 230000002441 reversible effect Effects 0.000 claims description 10
- 238000004590 computer program Methods 0.000 claims description 9
- 230000004913 activation Effects 0.000 claims description 7
- 230000006403 short-term memory Effects 0.000 claims description 6
- 230000007787 long-term memory Effects 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 2
- 239000010410 layer Substances 0.000 description 66
- 238000013135 deep learning Methods 0.000 description 13
- 210000002569 neuron Anatomy 0.000 description 7
- 238000013136 deep learning model Methods 0.000 description 5
- 239000003550 marker Substances 0.000 description 4
- 238000013145 classification model Methods 0.000 description 3
- 238000011161 development Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- 239000002356 single layer Substances 0.000 description 3
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 2
- 238000004422 calculation algorithm Methods 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 108010076504 Protein Sorting Signals Proteins 0.000 description 1
- 230000003213 activating effect Effects 0.000 description 1
- 230000015556 catabolic process Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000007796 conventional method Methods 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000006731 degradation reaction Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000000750 progressive effect Effects 0.000 description 1
- 238000007637 random forest analysis Methods 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 239000002699 waste material Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/12—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing
- G01R31/1227—Testing dielectric strength or breakdown voltage ; Testing or monitoring effectiveness or level of insulation, e.g. of a cable or of an apparatus, for example using partial discharge measurements; Electrostatic testing of components, parts or materials
Landscapes
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a partial discharge detection method, a partial discharge detection system, partial discharge detection equipment and a partial discharge detection medium, and relates to the field of partial discharge detection; the method comprises the following steps: acquiring operation data of target electrical equipment; inputting the operation data into a partial discharge discrimination model to obtain a discrimination result of the target electrical equipment; the partial discharge discrimination model comprises a bidirectional long-short-time memory network, an attention mechanism layer, a full connection layer and a trained optimizer; the bidirectional long-short-time memory network performs time sequence calculation and feature extraction on the operation data to obtain target time sequence feature data; the attention mechanism layer performs attention mechanism calculation and weighted fusion calculation on the target time sequence feature data to obtain target fusion feature data; the full connection layer performs linear transformation and nonlinear mapping on the target fusion characteristic data to obtain target transformation data; the trained optimizer determines a discrimination result of the target electrical equipment according to the target transformation data. The invention can realize partial discharge detection and improve detection accuracy.
Description
Technical Field
The present invention relates to the field of partial discharge detection, and in particular, to a method, a system, an apparatus, and a medium for detecting partial discharge.
Background
Partial discharge (Partial Discharge, PD) is one of the common failure phenomena in electrical equipment, and the high frequency signals generated by it have a significant impact on the reliability and safety of the equipment. Therefore, developing an efficient and accurate partial discharge detection system is critical to the power industry. Conventional partial discharge detection methods rely primarily on expert experience and manual feature extraction, and have many limitations and challenges. On the one hand, the artificial feature extraction needs to rely on the knowledge and experience of domain experts, so that the method is time-consuming and labor-consuming, and complex nonlinear relations are difficult to capture. On the other hand, the conventional method cannot be effectively utilized for unlabeled data, resulting in degradation of model performance.
In recent years, rapid development of deep learning technology brings new opportunities for partial discharge detection. The deep learning model can automatically learn the characteristic representation of the input data without manually designing the characteristics, thereby improving the accuracy and the robustness of the detection system. However, deep learning typically requires a large amount of marker data for training, and in practical applications, the marker data is often difficult to obtain, which limits the application of deep learning in partial discharge detection.
Disclosure of Invention
The invention aims to provide a partial discharge detection method, a partial discharge detection system, partial discharge detection equipment and a partial discharge detection medium, so as to realize partial discharge detection and improve detection accuracy.
In order to achieve the above object, the present invention provides the following solutions:
a partial discharge detection method, the method comprising:
acquiring operation data of target electrical equipment;
inputting the operation data into a partial discharge discrimination model to obtain a discrimination result of the target electrical equipment; the discrimination result comprises: a high frequency signal in which a partial discharge phenomenon occurs;
wherein the partial discharge discrimination model includes: a bidirectional long-short-time memory network, an attention mechanism layer, a full connection layer and a trained optimizer;
the bidirectional long and short-term memory network is used for carrying out time sequence calculation on the operation data to obtain target time sequence data, and carrying out feature extraction on the target time sequence data to obtain target time sequence feature data;
the attention mechanism layer is used for carrying out attention mechanism calculation and weighted fusion calculation on the target time sequence feature data to obtain target fusion feature data;
the full connection layer is used for carrying out linear transformation and nonlinear mapping on the target fusion characteristic data to obtain target transformation data;
the trained optimizer is used for determining a judging result of the target electrical equipment according to the target transformation data.
Optionally, the method for determining the partial discharge discrimination model specifically includes:
acquiring training data of the target electrical equipment; the training data includes: tagged operational data and untagged operational data; the label is a discrimination result;
constructing a partial discharge discrimination network; the partial discharge discrimination network includes: the system comprises a bidirectional long-short-time memory network, an attention mechanism layer, a full-connection layer and an optimizer which are connected in sequence;
inputting the training data into a two-way long and short time memory network in the partial discharge discrimination network to perform time sequence calculation to obtain training time sequence data, and performing feature extraction on the training time sequence data to obtain training time sequence feature data;
taking the training time sequence characteristic data as the input of the attention mechanism layer, and performing attention mechanism calculation and weighted fusion calculation on the training time sequence characteristic data to obtain training fusion characteristic data;
taking the training fusion characteristic data as the input of the full connection layer, and performing linear transformation and nonlinear mapping on the training fusion characteristic data to obtain training transformation data;
taking the training transformation data as input of the optimizer, taking the minimum cross entropy loss function as a target, and training parameters of the optimizer by adopting a back propagation method to obtain a trained optimizer; the cross entropy loss function is determined by adopting a semi-supervised learning method based on unbalanced consistency regularization; the cross entropy loss function comprises a labeled loss function and a label-free loss function;
the partial discharge discrimination model includes: the two-way long short-term memory network, the attention mechanism layer, the full connection layer and the trained optimizer.
Optionally, the training data is input to a bidirectional long and short time memory network in the partial discharge discrimination network to perform time sequence calculation, so as to obtain training time sequence data, and feature extraction is performed on the training time sequence data, so as to obtain training time sequence feature data, which specifically includes:
inputting the training data to a time sequence reset layer in the bidirectional long-short-time memory network, and performing nonlinear activation operation to obtain a nonlinear result; the time sequence reset layer comprises a plurality of forward long-short time memory network units and a plurality of reverse long-short time memory network units; the nonlinear result comprises a forward nonlinear result and a reverse nonlinear result;
inputting the nonlinear result to an updating layer in a bidirectional long-short-time memory network, splicing the forward nonlinear result and the reverse nonlinear result, and capturing the characteristics to obtain an output result;
and inputting the output result to an output layer in the bidirectional long-short-time memory network to perform multiplication operation, so as to obtain the training time sequence characteristic data.
Optionally, the training time sequence feature data is used as input of the attention mechanism layer, attention mechanism calculation and weighted fusion calculation are performed on the training time sequence feature data, and training fusion feature data is obtained, which specifically includes:
according to the training time sequence characteristic data, carrying out normalization correlation calculation by adopting an attention mechanism, and determining attention distribution corresponding to the training time sequence characteristic data;
and carrying out weighted summation according to the attention distribution and the training time sequence characteristic data to obtain training fusion characteristic data.
Optionally, the cross entropy loss function has the expression:
wherein l u Is a cross entropy loss function; i is a sequence number; n is the number of tagged operational data; μn is the number of unlabeled running data; q i For predicting class distribution; τ is a threshold; h is a cross entropy function;is a pseudo tag; p is p m (y|z i ) For predicting class distribution; z is training data; x is the operation data with labels; u is unlabeled running data.
A partial discharge detection system, the system comprising:
the data acquisition module is used for acquiring the operation data of the target electrical equipment;
the judging module is used for inputting the operation data into a partial discharge judging model to obtain a judging result of the target electrical equipment; the discrimination result comprises: a high frequency signal in which a partial discharge phenomenon occurs;
wherein the partial discharge discrimination model includes: a bidirectional long-short-time memory network, an attention mechanism layer, a full connection layer and a trained optimizer;
the bidirectional long and short-term memory network is used for carrying out time sequence calculation on the operation data to obtain target time sequence data, and carrying out feature extraction on the target time sequence data to obtain target time sequence feature data;
the attention mechanism layer is used for carrying out attention mechanism calculation and weighted fusion calculation on the target time sequence feature data to obtain target fusion feature data;
the full connection layer is used for carrying out linear transformation and nonlinear mapping on the target fusion characteristic data to obtain target transformation data;
the trained optimizer is used for determining a judging result of the target electrical equipment according to the target transformation data.
An electronic device comprising a memory for storing a computer program and a processor for running the computer program to cause the electronic device to perform the partial discharge detection method described above.
A computer readable storage medium storing a computer program which when executed by a processor implements the partial discharge detection method described above.
According to the specific embodiment provided by the invention, the invention discloses the following technical effects:
the invention provides a partial discharge detection method, a partial discharge detection system, partial discharge detection equipment and a partial discharge detection medium, wherein operation data of target electrical equipment are obtained; inputting the operation data into a partial discharge discrimination model to obtain a discrimination result of the target electrical equipment; the partial discharge discrimination model comprises a bidirectional long-short-time memory network, an attention mechanism layer, a full connection layer and a trained optimizer; the bidirectional long-short-time memory network performs time sequence calculation and feature extraction on the operation data to obtain target time sequence feature data; the attention mechanism layer performs attention mechanism calculation and weighted fusion calculation on the target time sequence feature data to obtain target fusion feature data; the full connection layer performs linear transformation and nonlinear mapping on the target fusion characteristic data to obtain target transformation data; the trained optimizer determines a discrimination result of the target electrical equipment according to the target transformation data; therefore, the invention can realize partial discharge detection and improve detection accuracy.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the drawings that are needed in the embodiments will be briefly described below, it being obvious that the drawings in the following description are only some embodiments of the present invention, and that other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
FIG. 1 is a flow chart of a partial discharge detection method according to an embodiment of the present invention;
fig. 2 is a block diagram of a partial discharge detection system according to an embodiment of the present invention.
Symbol description:
a data acquisition module-1 and a discrimination module-2.
Detailed Description
The following description of the embodiments of the present invention will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present invention, but not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
In order to solve the problems in the prior art, researchers begin to combine deep learning with semi-supervised learning, and a partial discharge detection system integrating the deep learning and the semi-supervised learning is provided. Semi-supervised learning exploits large amounts of unlabeled data to improve model performance by making efficient use of these unlabeled data. Specifically, the partial discharge detection system integrating deep learning and semi-supervised learning comprises the following key steps:
first, an initial deep learning model is trained using a small amount of marker data. The marking data is manually marked by an expert and contains samples of normal and partial discharges. By using these marker data, the deep learning model can learn the characteristic representation of the partial discharge. Next, semi-supervised learning is performed using the unlabeled data. Unlabeled data is a large amount of data collected from the power equipment that contains both normal samples and unknown partial discharge samples. Semi-supervised learning further optimizes the deep learning model with a large amount of unlabeled data by integrating the unlabeled data with the labeled data. Thus, the model can better capture the difference between normal and partial discharge, and the performance of the detection system is improved. Finally, the new samples are predicted and classified. The trained fusion model can predict a new sample and judge whether partial discharge exists. Meanwhile, the model can classify partial discharge and identify different types of discharge faults.
The development of a partial discharge detection system integrating deep learning and semi-supervised learning provides an efficient and accurate fault detection method for the power industry. By utilizing the capability of deep learning model for automatically learning feature representation and semi-supervised learning, the system can effectively utilize unlabeled data on the basis of a small amount of labeled data, and improve the performance and reliability of the detection system. In the future, with the further development of deep learning and semi-supervised learning, the partial discharge detection system integrating the deep learning and the semi-supervised learning will play a greater role in the field of power equipment fault detection.
The invention aims to provide a partial discharge detection method, a partial discharge detection system, partial discharge detection equipment and a partial discharge detection medium, so as to realize partial discharge detection and improve detection accuracy.
In order that the above-recited objects, features and advantages of the present invention will become more readily apparent, a more particular description of the invention will be rendered by reference to the appended drawings and appended detailed description.
Example 1
As shown in fig. 1, an embodiment of the present invention provides a partial discharge detection method, which includes:
step 100: and acquiring the operation data of the target electrical equipment.
Step 200: inputting the operation data into a partial discharge discrimination model to obtain a discrimination result of the target electrical equipment; the discrimination result includes: a high frequency signal in which a partial discharge phenomenon occurs.
Wherein, partial discharge discriminant model includes: a bidirectional long-short-time memory network, an attention mechanism layer, a full connection layer and a trained optimizer.
The bidirectional long-short time memory network is used for carrying out time sequence calculation on the operation data to obtain target time sequence data, and carrying out feature extraction on the target time sequence data to obtain target time sequence feature data.
The attention mechanism layer is used for carrying out attention mechanism calculation and weighted fusion calculation on the target time sequence feature data to obtain target fusion feature data.
The full connection layer is used for carrying out linear transformation and nonlinear mapping on the target fusion characteristic data to obtain target transformation data.
The trained optimizer is used for determining the judging result of the target electrical equipment according to the target transformation data.
The method for determining the partial discharge discrimination model specifically comprises the following steps:
acquiring training data of target electrical equipment; the training data includes: tagged operational data and untagged operational data; the label is the discrimination result.
Constructing a partial discharge discrimination network; a partial discharge discrimination network comprising: the system comprises a bidirectional long-short-time memory network, an attention mechanism layer, a full-connection layer and an optimizer which are connected in sequence.
And inputting the training data into a two-way long and short time memory network in the partial discharge judging network to perform time sequence calculation to obtain training time sequence data, and performing feature extraction on the training time sequence data to obtain training time sequence feature data.
And taking the training time sequence characteristic data as input of an attention mechanism layer, and performing attention mechanism calculation and weighted fusion calculation on the training time sequence characteristic data to obtain training fusion characteristic data.
And taking the training fusion characteristic data as the input of the full connection layer, and performing linear transformation and nonlinear mapping on the training fusion characteristic data to obtain training transformation data.
Training parameters of the optimizer by using training transformation data as input of the optimizer and taking the minimum cross entropy loss function as a target and adopting a back propagation method to obtain the trained optimizer; the cross entropy loss function is determined by adopting a semi-supervised learning method based on imbalance consistency regularization; the cross entropy loss function includes a labeled loss function and an unlabeled loss function.
Specifically, the cross entropy loss function is expressed as:
wherein l u Is a cross entropy loss function; i is a sequence number; n is the number of tagged operational data; μn is the number of unlabeled running data; q i For predicting class distribution; τ is a threshold; h is a cross entropy function;is a pseudo tag; p is p m (y|z i ) For predicting class distribution; z is training data; x is the operation data with labels; u is unlabeled running data.
A partial discharge discriminant model comprising: a bidirectional long-short-time memory network, an attention mechanism layer, a full connection layer and a trained optimizer.
The method comprises the steps of inputting training data into a two-way long and short time memory network in a partial discharge judging network to perform time sequence calculation to obtain training time sequence data, and performing feature extraction on the training time sequence data to obtain training time sequence feature data, wherein the method specifically comprises the following steps of:
inputting training data to a time sequence reset layer in a bidirectional long-short-time memory network, and performing nonlinear activation operation to obtain a nonlinear result; the time sequence resetting layer comprises a plurality of forward long-short time memory network units and a plurality of reverse long-short time memory network units; the nonlinear results include forward nonlinear results and reverse nonlinear results.
And inputting the nonlinear result to an updating layer in the bidirectional long-short-time memory network, splicing the forward nonlinear result and the reverse nonlinear result, and capturing the characteristics to obtain an output result.
And inputting the output result to an output layer in the bidirectional long-short-time memory network to perform multiplication operation, so as to obtain training time sequence characteristic data.
In addition, training time sequence feature data is used as input of an attention mechanism layer, attention mechanism calculation and weighted fusion calculation are carried out on the training time sequence feature data, and training fusion feature data is obtained, wherein the training fusion feature data specifically comprises:
and carrying out normalized correlation calculation by adopting an attention mechanism according to the training time sequence characteristic data, and determining the attention distribution corresponding to the training time sequence characteristic data.
And carrying out weighted summation according to the attention distribution and the training time sequence characteristic data to obtain training fusion characteristic data.
Example 2
As shown in fig. 2, an embodiment of the present invention provides a partial discharge detection system, including: a data acquisition module 1 and a discrimination module 2.
And the data acquisition module 1 is used for acquiring the operation data of the target electrical equipment.
The judging module 2 is used for inputting the operation data into the partial discharge judging model to obtain a judging result of the target electrical equipment; the discrimination result includes: a high frequency signal in which a partial discharge phenomenon occurs.
Wherein, partial discharge discriminant model includes: a bidirectional long-short-time memory network, an attention mechanism layer, a full connection layer and a trained optimizer.
The bidirectional long-short time memory network is used for carrying out time sequence calculation on the operation data to obtain target time sequence data, and carrying out feature extraction on the target time sequence data to obtain target time sequence feature data.
The attention mechanism layer is used for carrying out attention mechanism calculation and weighted fusion calculation on the target time sequence feature data to obtain target fusion feature data.
The full connection layer is used for carrying out linear transformation and nonlinear mapping on the target fusion characteristic data to obtain target transformation data.
The trained optimizer is used for determining the judging result of the target electrical equipment according to the target transformation data.
Example 3
An embodiment of the present invention provides an electronic device including a memory for storing a computer program and a processor that runs the computer program to cause the electronic device to execute the partial discharge detection method in embodiment 1.
In one embodiment, the present invention also provides a computer-readable storage medium storing a computer program which, when executed by a processor, implements the partial discharge detection method in embodiment 1.
Example 4
Aiming at the problems that the traditional method needs to rely on the knowledge and experience of field experts to perform feature extraction, lacks a large amount of marking data, improves the performance and reliability of a detection system and the like, the partial discharge detection method provided by the invention is a partial discharge detection method integrating deep learning and semi-supervised learning.
In practical application, the specific operation steps of the method can be as follows:
constructing a classification model by using Bi-LSTM and an attention mechanism in deep learning, wherein the model comprises the following components:
a) An input layer for receiving input data. I.e. receive operational data of the target electrical device.
b) The Bi-LSTM layer, namely a bidirectional long and short time memory network, comprises a forward and a backward bidirectional long and short time memory network, and is used for inputting data to perform time sequence modeling and feature extraction.
Specifically, bi-LSTM captures long-term dependencies in time series data better by considering both past and future information over a time step. In timing modeling, biLSTM divides an input sequence into two parts in time order: forward and backward sequences. The forward sequence is entered into Bi-LSTM step by step in order starting at time step t=1. The backward sequence is gradually input into Bi-LSTM in reverse order starting from time step t=t. Thus, the BiLSTM can more fully model the time series data by taking both past and future context information into account. Bi-LSTM plays an important role in feature extraction of time series data. By calculating the hidden states in the forward and backward directions at each time step, the Bi-LSTM is able to capture local and global features in the input sequence. The forward hidden state contains the context information from the beginning of the sequence to the current time step, while the backward hidden state contains the context information from the end of the sequence to the current time step. These hidden states can be used to extract features about signal sequence variations.
c) And the attention mechanism layer is used for carrying out weighted fusion on the output of the Bi-LSTM layer so as to extract key characteristics and realize the function of data dimension reduction.
Assume that the set of inputs is h= [ H 1 ,h 2 ,h 3 ,...,h n ]The use of the attention mechanism can extract more important content according to the correlation of the output q and the input information, so as to improve the accuracy and reduce the calculation amount. s is the calculation of q and h i Correlation between them. These scores are then normalized with softmax to yield q at each input h i Attention distribution a= [ a ] 1 ,a 2 ,a 3 ,...,a n ]。
Information can be selectively extracted from the input information based on these attention profiles, i.e. the input information is weighted summed based on the attention profiles.
h i Refers to a group of input information, a i Refers to the importance of each set of input information.
d) And the full connection layer is used for completing linear transformation and nonlinear mapping of the features. The full connection layer is to connect all neurons of the previous layer with all neurons of the current layer, and establish a linear relation between input and output by learning weight parameters to complete mapping.
Specifically: 1. input feature conversion: the output features of the previous layer are linearly transformed, typically including matrix multiplication and offset addition. This process can be expressed as z=xw+b, where X is the output feature of the previous layer, W is the weight matrix, b is the bias vector, and Z is the input feature of the current layer. 2. Activating function application: and carrying out nonlinear mapping on the input characteristics after linear transformation through an activation function so as to increase the expression capacity of the network. Common activation functions include ReLU, sigmoid, tanh, etc. 3. Calculating output characteristics: the input feature mapped by the activation function is taken as the output feature of the current layer, namely Y=f (Z), wherein Y is the output feature of the current layer, and f is the activation function.
e) And the output layer is used for outputting the classification result.
After the model is built, the network model is trained by using the labeled data.
The dimension of input data is N multiplied by T multiplied by L (N is the number of signals, T is the number of signal segments for dividing signals, L is the number of characteristics), the model is 7 layers altogether, the first layer is an input layer, the second layer is Bi-LSTM of 128 neurons, the third layer is Bi-LSTM of 64 neurons, the fourth layer is an Attention layer, key information is extracted, the effect of reducing the dimension is realized, the fifth layer is a single-layer network of 32 neurons, the sixth layer is a single-layer network of 16 neurons, and the seventh layer is a single-layer network of 2 neurons. The original data comprises tag data D l ＝{(x i ,y i )|i∈(1,...,n)}，x i Is input, y i Is one-hot label, no label data D u ＝{u i I e (1,..n) }. The model is trained by using local discharge signal data with labels, cross entropy is selected as a loss function in a supervised learning mode, and model parameters are optimized by using an Adam optimizer, so that fault discrimination of signals is realized. Training process: training data (input features and corresponding labels) are input into the model, gradients are calculated using a back propagation method, and model parameters are updated using Adam optimization algorithms. In the training process, the model is gradually optimized through iterating a plurality of epochs until a preset stopping condition is reached.
Traditional supervised learning builds a model using only labeled samples, resulting in D u Waste of information in the system. And the generalization ability is low due to the small number and class imbalance. Although label information is not directly contained in unlabeled samples, they are sampled independently from labeled samples and can provide efficient information for modeling. Thus propose a use of D u Semi-supervised learning method deep FuseNet for realizing multi-source data fusion by utilizing deep learning modelDifferent types of data and information are combined. Consistency regularization is an important component of the most advanced Semi-supervised learning (Semi-supervised learning, SSL) algorithm, which assumes that predictions of perturbed versions of the same set of signals should remain unchanged. A semi-supervised learning method based on imbalance consistency regularization is presented. The method includes two cross entropy losses: with label loss l s And no label loss. Signal data enhancement is achieved by randomly varying a number of peak point waveforms. Then respectively identifying the original data and the enhanced data by using the model, q i ＝p m (y|u i ) Generating a pseudo tag, retaining the pseudo tag for the unlabeled exemplar if the predicted categories are consistent and the probability exceeds a predetermined threshold:
μn is the number of unlabeled operating data, H is the cross entropy function,is a pseudo tag without a tag, q i Is the probability distribution without labels.
Wherein,
cross entropy loss based on tagged sample data:
n is the number of tagged operational data, p i Is a pseudo tag under a label, p m (y|x i ) Is the probability distribution under the label.
Combining unlabeled samples with labeled samples that retain pseudo labels, calculating cross entropy loss for samples containing unlabeled samples:
data enhancement of partial discharge signals in unlabeled data, including but not limited to signal shifting, noise adding, etc., the enhanced data being D h ＝{h i :i∈(1,…,n)}。
Evaluation index: for the problem of unbalanced classification of data samples, the accuracy is used as an evaluation index, so that the model cannot be accurately evaluated. Herein, ma Xiusi correlation coefficient (MCC), precision, recall, F1-score, G-mean, and Area Under Curve (AUC) were used as evaluation indexes to evaluate performance, and the formula was defined as follows:
where TP represents a true case, TN represents a true case, FP represents a false case, and FN represents a false case.
After the model is initially built by using the labeled data sample, the signal data after the enhancement of the unlabeled sample and the original signal data are respectively distinguished by using the classification model, so that corresponding distinguishing results are obtained, and because the problems are classified into two types, 0 represents normal and 1 represents fault. Then, probability values of 0 and 1 corresponding to each group of unlabeled data are obtained, and the sum of the two probabilities is 1. Then the label value corresponding to the large probability is used as the prediction result of the label-free sample, and if the label value is consistent with the prediction label of the same sample corresponding to the original data after enhancement, a pseudo label is given to the label-free sample.
Because the probability of failure during the operation of the device is low, the collected tagged data or untagged number is unbalanced data. In order to alleviate the unbalance problem, the unlabeled sample is judged to be faulty, and the data with the pseudo label is combined with the labeled data, so that the model is retrained to finish parameter tuning, and the judging capability of the model is improved.
By combining deep learning and semi-supervised learning, the information in the labeled and unlabeled data sets is fully utilized, an accurate partial discharge judgment model is constructed, whether the partial discharge phenomenon occurs in the signal is finally determined, and a judgment result is provided.
In order to ensure the stability and the fidelity of the model prediction result, a cross-validation method is adopted, the labeled data set is divided into 5 mutually exclusive subsets with consistent sizes, and the mutual exclusive subsets are obtained through hierarchical sampling, so that the consistency of data distribution is maintained. Then training and verifying for 5 times, taking 4 subsets as training sets, taking the rest subsets as test sets, and taking the average value of the 5 test results as the final result of the model.
The amount of data in the different categories in the tagged dataset is quite different, which can have an impact on classifier construction. The information of the unlabeled dataset is then used to augment a minority of classes of samples, and the size of the threshold in the semi-supervised learning strategy will determine the number of fault data identified, obviously the size of the threshold is relative to the number of pseudo tags. When the threshold is too high, the model sets pseudo labels for unlabeled samples with high prediction probability, and the number of the pseudo labels is small; as the threshold value is smaller, a sample with a small probability is also given a pseudo tag, and the amount of data determined as a failure sample among the unlabeled samples increases. Although this is advantageous for balancing the ratio of the amount of data between the normal and faulty signals in the training set, the probability of erroneous samples in the training set is increased because the pseudo tag is not a real tag, affecting the determination of the classification limits in the classifier, reducing the accuracy of the model. In order to achieve balance, an optimal threshold is found, a broad search mode is adopted, and the results are compared.
When the threshold is set to different values, the corresponding prediction results are shown in table 1. Different thresholds can be seen, the predicted results are inconsistent, and the fact that the different thresholds affect the classification model accuracy is also explained. The MCC is highest at a threshold of 0.7. Then the threshold is set to be between 0.5 and 0.95 by adopting extensive searching, and the optimal threshold is found by analyzing different prediction results. The experimental results are shown in the graph, and when the threshold is set to 0.7, the effect of the discrimination model is optimal.
Table 1 comparison of the predicted results for different thresholds
| θ | MCC | Precision | Recall | F1-score | AUC |
| 0.7 | 0.776 | 0.845 | 0.736 | 0.786 | 0.936 |
| 0.75 | 0.732 | 0.772 | 0.725 | 0.748 | 0.914 |
| 0.8 | 0.749 | 0.782 | 0.747 | 0.747 | 0.947 |
| 0.85 | 0.771 | 0.848 | 0.725 | 0.781 | 0.928 |
| 0.9 | 0.753 | 0.795 | 0.742 | 0.767 | 0.921 |
The performance of the classifier is compared with that of other classifier, such as XGBoost, random forest and lightGBM, and then compared with a deep network model without label data, and the accuracy of the classifier is still lower than that of deep FuseNet. The comparison of the prediction results under different classifiers is shown in Table 2.
Table 2 comparison table of prediction results under different classifiers
In the present specification, each embodiment is described in a progressive manner, and each embodiment is mainly described in a different point from other embodiments, and identical and similar parts between the embodiments are all enough to refer to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant points refer to the description of the method section.
The principles and embodiments of the present invention have been described herein with reference to specific examples, the description of which is intended only to assist in understanding the methods of the present invention and the core ideas thereof; also, it is within the scope of the present invention to be modified by those of ordinary skill in the art in light of the present teachings. In view of the foregoing, this description should not be construed as limiting the invention.
Claims (8)
1. A partial discharge detection method, the method comprising:
acquiring operation data of target electrical equipment;
inputting the operation data into a partial discharge discrimination model to obtain a discrimination result of the target electrical equipment; the discrimination result comprises: a high frequency signal in which a partial discharge phenomenon occurs;
wherein the partial discharge discrimination model includes: a bidirectional long-short-time memory network, an attention mechanism layer, a full connection layer and a trained optimizer;
the bidirectional long and short-term memory network is used for carrying out time sequence calculation on the operation data to obtain target time sequence data, and carrying out feature extraction on the target time sequence data to obtain target time sequence feature data;
the attention mechanism layer is used for carrying out attention mechanism calculation and weighted fusion calculation on the target time sequence feature data to obtain target fusion feature data;
the full connection layer is used for carrying out linear transformation and nonlinear mapping on the target fusion characteristic data to obtain target transformation data;
the trained optimizer is used for determining a judging result of the target electrical equipment according to the target transformation data.
2. The partial discharge detection method according to claim 1, wherein the method for determining the partial discharge discrimination model specifically comprises:
acquiring training data of the target electrical equipment; the training data includes: tagged operational data and untagged operational data; the label is a discrimination result;
constructing a partial discharge discrimination network; the partial discharge discrimination network includes: the system comprises a bidirectional long-short-time memory network, an attention mechanism layer, a full-connection layer and an optimizer which are connected in sequence;
inputting the training data into a two-way long and short time memory network in the partial discharge discrimination network to perform time sequence calculation to obtain training time sequence data, and performing feature extraction on the training time sequence data to obtain training time sequence feature data;
taking the training time sequence characteristic data as the input of the attention mechanism layer, and performing attention mechanism calculation and weighted fusion calculation on the training time sequence characteristic data to obtain training fusion characteristic data;
taking the training fusion characteristic data as the input of the full connection layer, and performing linear transformation and nonlinear mapping on the training fusion characteristic data to obtain training transformation data;
taking the training transformation data as input of the optimizer, taking the minimum cross entropy loss function as a target, and training parameters of the optimizer by adopting a back propagation method to obtain a trained optimizer; the cross entropy loss function is determined by adopting a semi-supervised learning method based on unbalanced consistency regularization; the cross entropy loss function comprises a labeled loss function and a label-free loss function;
the partial discharge discrimination model includes: the two-way long short-term memory network, the attention mechanism layer, the full connection layer and the trained optimizer.
3. The partial discharge detection method according to claim 2, wherein the training data is input to a bidirectional long and short time memory network in the partial discharge discrimination network to perform time sequence calculation to obtain training time sequence data, and the training time sequence data is subjected to feature extraction to obtain training time sequence feature data, and the method specifically comprises:
inputting the training data to a time sequence reset layer in the bidirectional long-short-time memory network, and performing nonlinear activation operation to obtain a nonlinear result; the time sequence reset layer comprises a plurality of forward long-short time memory network units and a plurality of reverse long-short time memory network units; the nonlinear result comprises a forward nonlinear result and a reverse nonlinear result;
inputting the nonlinear result to an updating layer in a bidirectional long-short-time memory network, splicing the forward nonlinear result and the reverse nonlinear result, and capturing the characteristics to obtain an output result;
and inputting the output result to an output layer in the bidirectional long-short-time memory network to perform multiplication operation, so as to obtain the training time sequence characteristic data.
4. The partial discharge detection method according to claim 2, wherein the training time sequence feature data is used as an input of the attention mechanism layer, and attention mechanism calculation and weighted fusion calculation are performed on the training time sequence feature data to obtain training fusion feature data, and specifically includes:
according to the training time sequence characteristic data, carrying out normalization correlation calculation by adopting an attention mechanism, and determining attention distribution corresponding to the training time sequence characteristic data;
and carrying out weighted summation according to the attention distribution and the training time sequence characteristic data to obtain training fusion characteristic data.
5. The partial discharge detection method according to claim 2, wherein the cross entropy loss function has an expression of:
wherein l u Is a cross entropy loss function; i is a sequence number; n is the number of tagged operational data; μn is the number of unlabeled running data; q i For predicting class distribution; τ is a threshold; h is a cross entropy function;is a pseudo tag; p is p m (y|z i ) For predicting class distribution; z is training data; x is the operation data with labels; u is unlabeled running data.
6. A partial discharge detection system, the system comprising:
the data acquisition module is used for acquiring the operation data of the target electrical equipment;
the judging module is used for inputting the operation data into a partial discharge judging model to obtain a judging result of the target electrical equipment; the discrimination result comprises: a high frequency signal in which a partial discharge phenomenon occurs;
wherein the partial discharge discrimination model includes: a bidirectional long-short-time memory network, an attention mechanism layer, a full connection layer and a trained optimizer;
the bidirectional long and short-term memory network is used for carrying out time sequence calculation on the operation data to obtain target time sequence data, and carrying out feature extraction on the target time sequence data to obtain target time sequence feature data;
the attention mechanism layer is used for carrying out attention mechanism calculation and weighted fusion calculation on the target time sequence feature data to obtain target fusion feature data;
the full connection layer is used for carrying out linear transformation and nonlinear mapping on the target fusion characteristic data to obtain target transformation data;
the trained optimizer is used for determining a judging result of the target electrical equipment according to the target transformation data.
7. An electronic device comprising a memory for storing a computer program and a processor that runs the computer program to cause the electronic device to perform the partial discharge detection method according to any one of claims 1 to 5.
8. A computer-readable storage medium, characterized in that it stores a computer program which, when executed by a processor, implements the partial discharge detection method according to any one of claims 1 to 5.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311387592.2A CN117347803A (en) | 2023-10-25 | 2023-10-25 | Partial discharge detection method, system, equipment and medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311387592.2A CN117347803A (en) | 2023-10-25 | 2023-10-25 | Partial discharge detection method, system, equipment and medium |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN117347803A true CN117347803A (en) | 2024-01-05 |
Family
ID=89366393
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311387592.2A Pending CN117347803A (en) | 2023-10-25 | 2023-10-25 | Partial discharge detection method, system, equipment and medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117347803A (en) |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111062495A (en) * | 2019-11-28 | 2020-04-24 | 深圳市华尊科技股份有限公司 | Machine learning method and related device |
| CN113920472A (en) * | 2021-10-15 | 2022-01-11 | 中国海洋大学 | Unsupervised target re-identification method and system based on attention mechanism |
| CN113988215A (en) * | 2021-11-15 | 2022-01-28 | 国网四川省电力公司营销服务中心 | Power distribution network metering cabinet state detection method and system |
| CN114239725A (en) * | 2021-12-17 | 2022-03-25 | 四川大学 | Electricity stealing detection method oriented to data virus throwing attack |
| CN114638256A (en) * | 2022-02-22 | 2022-06-17 | 合肥华威自动化有限公司 | Transformer fault detection method and system based on sound wave signals and attention network |
| CN115272777A (en) * | 2022-09-26 | 2022-11-01 | 山东大学 | Semi-supervised image analysis method for power transmission scene |
| WO2023045278A1 (en) * | 2021-09-27 | 2023-03-30 | 西安交通大学 | Data dual-drive method, apparatus, and device for predicting power grid failure during typhoon |
| CN116304979A (en) * | 2023-03-02 | 2023-06-23 | 兰州交通大学 | Attention mechanism-based multi-feature fusion partial discharge type identification method |
| CN116821783A (en) * | 2023-04-27 | 2023-09-29 | 天津大学 | Fault detection method, system, equipment and medium |
-
2023
- 2023-10-25 CN CN202311387592.2A patent/CN117347803A/en active Pending
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111062495A (en) * | 2019-11-28 | 2020-04-24 | 深圳市华尊科技股份有限公司 | Machine learning method and related device |
| WO2023045278A1 (en) * | 2021-09-27 | 2023-03-30 | 西安交通大学 | Data dual-drive method, apparatus, and device for predicting power grid failure during typhoon |
| CN113920472A (en) * | 2021-10-15 | 2022-01-11 | 中国海洋大学 | Unsupervised target re-identification method and system based on attention mechanism |
| CN113988215A (en) * | 2021-11-15 | 2022-01-28 | 国网四川省电力公司营销服务中心 | Power distribution network metering cabinet state detection method and system |
| CN114239725A (en) * | 2021-12-17 | 2022-03-25 | 四川大学 | Electricity stealing detection method oriented to data virus throwing attack |
| CN114638256A (en) * | 2022-02-22 | 2022-06-17 | 合肥华威自动化有限公司 | Transformer fault detection method and system based on sound wave signals and attention network |
| CN115272777A (en) * | 2022-09-26 | 2022-11-01 | 山东大学 | Semi-supervised image analysis method for power transmission scene |
| CN116304979A (en) * | 2023-03-02 | 2023-06-23 | 兰州交通大学 | Attention mechanism-based multi-feature fusion partial discharge type identification method |
| CN116821783A (en) * | 2023-04-27 | 2023-09-29 | 天津大学 | Fault detection method, system, equipment and medium |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109408389B (en) | Code defect detection method and device based on deep learning | |
| CN112784881B (en) | Network abnormal flow detection method, model and system | |
| CN113312447B (en) | Semi-supervised log anomaly detection method based on probability label estimation | |
| CN114760098A (en) | CNN-GRU-based power grid false data injection detection method and device | |
| CN112541532A (en) | Target detection method based on dense connection structure | |
| CN115272776B (en) | Hyperspectral image classification method based on double-path convolution and double attention and storage medium | |
| CN114048468A (en) | Intrusion detection method, intrusion detection model training method, device and medium | |
| CN111967535A (en) | Fault diagnosis method and device for temperature sensor in grain storage management scene | |
| Wang et al. | Multi-task Joint Sparse Representation Classification Based on Fisher Discrimination Dictionary Learning. | |
| CN113343581A (en) | Transformer fault diagnosis method based on graph Markov neural network | |
| CN115587335A (en) | Training method of abnormal value detection model, abnormal value detection method and system | |
| CN117081831A (en) | Network intrusion detection method and system based on data generation and attention mechanism | |
| CN116304853A (en) | Data classification method and system combining extreme learning machine and feature extraction | |
| CN117034143A (en) | Distributed system fault diagnosis method and device based on machine learning | |
| CN114897085A (en) | Clustering method based on closed subgraph link prediction and computer equipment | |
| CN106815209B (en) | Uygur agricultural technical term identification method | |
| CN117516937A (en) | Rolling bearing unknown fault detection method based on multi-mode feature fusion enhancement | |
| CN111782805A (en) | Text label classification method and system | |
| CN116467720A (en) | Intelligent contract vulnerability detection method based on graph neural network and electronic equipment | |
| CN117347803A (en) | Partial discharge detection method, system, equipment and medium | |
| CN116226748A (en) | Multi-label co-occurrence network discrimination-based electricity stealing type detection method and system | |
| CN106816871B (en) | State similarity analysis method for power system | |
| Nag et al. | CNN based approach for post disaster damage assessment | |
| CN113535928A (en) | Service discovery method and system of long-term and short-term memory network based on attention mechanism | |
| CN115288994B (en) | Improved DCGAN-based compressor abnormal state detection method |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |