Optical fiber current transformer fault diagnosis method and related equipment based on improved dense network

Abstract

The present invention discloses a fiber optic current transformer fault diagnosis method and related equipment based on an improved dense network. The method collects fiber optic current transformer operating status monitoring data and the transformer's final output signal data; uses a time sliding window to extract time domain features, and combines the operating status monitoring data to form high-dimensional multi-source information; uses principal component analysis to reduce the dimensionality of the multi-source information, and uses Markov transition fields to convert the reduced multi-source information into Markov transition field image data; and inputs the reduced multi-source information time series data and Markov transition field image data into a trained dual-input improved dense connection network model to implement current transformer fault diagnosis. The present invention can accurately identify various fiber optic current transformer fault types and ensure the transformer's accurate operation.

Description

Technical Field

[0001] The invention belongs to the technical field of optical fiber current transformer fault detection, and in particular relates to an optical fiber current transformer fault diagnosis method based on an improved dense network. Background Art

[0002] Fiber-optic current transformers (FOCTs) offer advantages such as simple insulation, high precision, compact size, no ferromagnetic saturation or resonance, strong anti-interference capabilities, no risk of open circuits on the low-voltage side, wide dynamic range, wide frequency bandwidth, and convenient digital communication. They have been gradually adopted in converter stations and substations, replacing traditional electromagnetic current transformers. However, due to long-term operation in complex environments such as strong electromagnetic interference, wide temperature fluctuations, and vibration, the complex optoelectronic components within FOCTs can degrade, affecting their accuracy, stability, and reliability, and even causing failures.

[0003] For fiber-optic current transformers, research focuses on improving fiber optic materials and transformer structures, enhancing transformer temperature performance, sensitivity, and noise reduction. However, given the difficulties of fiber-optic current transformers, such as multiple fault types, complex fault causes, and difficult-to-locate fault locations, it is imperative to seek efficient and accurate methods for diagnosing fiber-optic current transformer faults. Current fiber-optic current transformer fault diagnosis methods mainly include: methods based on mathematical analytical models; methods based on signal processing; and methods based on data processing and artificial intelligence. The method based on mathematical analytical models uses the principle and structure analysis of the fiber-optic current transformer to establish a mathematical analytical model of the fault output as prior information, then takes a difference from the actual observed signal, analyzes and judges the difference signal, and thus achieves transformer fault diagnosis. The method based on signal processing uses signal processing algorithms such as time-frequency analysis and signal decomposition to extract the fault characteristics of the fiber-optic current transformer output signal, judge the equipment working status based on the characteristics, and achieve fault diagnosis. Methods based on data processing and artificial intelligence fall into two categories: one is to preprocess the fiber optic current transformer fault data through time-frequency analysis and feature extraction to construct a fault feature dataset, which is then used through classification and diagnosis algorithms such as machine learning. The other is to preprocess the fault data through noise reduction filtering, enhancement and expansion, and then directly input it into a deep neural network to obtain diagnostic results. The complex and multi-dimensional structure of fiber optic current transformers makes it difficult to establish an accurate model and to achieve multiple types of diagnosis through single signal decomposition and comparison. Methods based on data processing and artificial intelligence can fully utilize the advantages of data to avoid these problems. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the deficiencies in the prior art, the present invention provides a method for diagnosing optical fiber current transformer faults based on an improved dense network. A dual-input feature data set is constructed by extracting features from multiple state monitoring data and output signals. The transformer fault diagnosis is realized through the dual-input improved dense connection network, thereby realizing the diagnosis and identification of complex fault types of optical fiber current transformers.

[0005] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is:

[0006] A method for diagnosing optical fiber current transformer faults based on an improved dense network comprises the following steps:

[0007] Step 1: Collect the optical fiber current transformer operating status monitoring data and the transformer final output signal data.

[0008] Step 2: Use a time sliding window to extract time domain features from the collected data, and combine the operating status monitoring data to form high-dimensional multi-source information.

[0009] Step 3: Use principal component analysis to reduce the dimensionality of high-dimensional multi-source information, preliminarily extract features, and achieve data fusion.

[0010] Step 4: Use the Markov transition field to convert the multi-source information after dimensionality reduction into Markov transition field image data.

[0011] In step 5, the multi-source information time series data and Markov transfer field image data after dimensionality reduction are input into the trained dual-input improved dense connection network model to realize current transformer fault diagnosis.

[0012] Preferably: the dual-input improved densely connected network model includes a time series input path, an image input path, a splicing layer, a dense structure, a batch normalization layer, a global average pooling layer, a fully connected layer, and an output layer. The time series input path and the image input path are respectively connected to the splicing layer, and the splicing layer, the dense structure, the third batch normalization layer, the global average pooling layer, the fully connected layer, and the output layer are connected in sequence. The time series input path includes a first convolutional layer, a first batch normalization layer, and a first pooling layer connected in sequence. The image input path includes a second convolutional layer, a second batch normalization layer, a second pooling layer, and a CBAM spatial attention module connected in sequence. The dense structure includes a first SEDB block, a first transition layer, a second SEDB block, a second transition layer, a third SEDB block, a third transition layer, and a fourth SEDB block connected in sequence, and the output of the first SEDB block enters the input of the first transition layer, the second transition layer, and the third transition layer respectively, the output of the first transition layer enters the input of the second transition layer and the third transition layer respectively, and the output of the second transition layer enters the input of the third transition layer. The first SEDB block, the second SEDB block, the third SEDB block, and the fourth SEDB block have 3, 6, 12, and 8 C-SE units respectively. The C-SE unit includes a first BN layer, a first Relu activation function, a 1×1 convolution layer, a second BN layer, a second Relu activation function, a 1×3 convolution layer, and an SE channel attention module connected in sequence.

[0013] Preferably, the method of step 2 includes:

[0014] The time sliding window length used is Ns, and the sliding step size is Ns.

[0015] The extracted time domain features include mean, standard deviation, variance, peak-to-peak value, absolute mean, root mean square amplitude, root mean square, kurtosis, crest factor, margin index, shape factor, and pulse factor.

[0016] The extracted multi-dimensional time domain features of the output signal and the collected multi-dimensional status monitoring data of the optical fiber current transformer are spliced ​​in time sequence to form a 27-dimensional multi-source information sequence.

[0017] Preferred: Method of using principal component analysis in step 3 to reduce the dimension of high-dimensional multi-source information:

[0018] Step (3-1) uses the multi-source information features of the optical fiber current transformer to construct the sample matrix X:

[0019]

[0020] In the formula, m is the number of samples, and n is the feature dimension of each sample.

[0021] Step (3-2) calculates the mean of each column vector of the sample matrix X by column and standard deviation s i , for each element After standardization, we get the standardized sample matrix X':

[0022]

[0023] Step (3-3) calculates the covariance matrix R of the standardized sample matrix:

[0024]

[0025] Where, are the vector means of the i-th row or j-th column of the matrix X', respectively.

[0026] Steps (3-4) calculate the eigenvalues ​​and eigenvectors of the covariance matrix R, and arrange the eigenvalues ​​{λ i ,i=1,2,...,n}, calculate the cumulative contribution rate of the first i principal components:

[0027]

[0028] Steps (3-5) retain the principal components with cumulative contribution rates exceeding 90%, and construct a reduced-dimensional feature matrix from the corresponding eigenvectors and the sample matrix. Ultimately, principal component analysis is used to reduce the 27-dimensional multi-source information features of the fiber optic current transformer to 9 dimensions.

[0029] Preferably: the method of using the Markov transition field to convert the multi-source information after dimensionality reduction into Markov transition field image data in step 4 includes:

[0030] For the 9-dimensional mutual inductor multi-source information time series, each feature information time series X=[x1,x2,…,x n ], n represents the length of the time series. First, the subsequence X is evenly divided into Q intervals and the corresponding quantiles are determined. Each element x of the time series is assigned i The corresponding quantile q j Sequence number, calculate the quantile interval transfer along the time axis with a first-order Markov chain, construct the Markov transfer matrix, expand the transfer matrix by arranging each probability in time sequence, and obtain the Markov transfer field matrix M:

[0031]

[0032] Where m ij Indicates the quantile q i The state is transferred to quantile q j The conditional probability of the state, that is, m ij=P(x t ∈q i |x t-1 ∈q j ).

[0033] The hyperparameter of the Markov transition field, the number of quantile intervals Q, is taken as the sequence length. The obtained Markov transition field matrix is ​​generated as a color image.

[0034] Preferably, collecting the optical fiber current transformer operating status monitoring data and the transformer final output signal data includes:

[0035] The output signal is collected through the output part of the optical fiber current transformer.

[0036] The status monitoring data collected for the optical fiber current transformer includes SLD temperature, optical module temperature, sensor ring temperature, circuit board temperature, TEC voltage, positive power supply voltage, system voltage, negative power supply voltage, optical module compensation, sensor ring compensation, SLD optical power, absolute optical power, system current, SLD current, and optical module humidity.

[0037] The working status of the mutual inductor during information collection includes: normal state, insulation structure contamination bias fault, optical path performance degradation drift fault, linear polarizer extinction ratio ratio fault, optical path interruption failure fault, light source drive attenuation ratio fault, light source temperature control drift fault, light source coupling misalignment failure fault, and light source chip breakdown failure fault.

[0038] Preferably: the method for training a dual-input improved densely connected network includes:

[0039] A dual-input dataset was constructed by mapping the 9-dimensional multi-source information sequence to the generated Markov transition field image. Each dual-input sample was then labeled according to the transformer operating state sequence (0-8). The dataset was divided into training, validation, and test sets with an 8:1:1 ratio. The hyperparameters for model training were set as follows: a maximum number of training iterations of 50, a learning rate of 0.0001, a batch size of 4, an Adam optimizer, and a cross-entropy loss function. Network parameters were optimized based on the validation set accuracy to determine the final diagnostic network.

[0040] Preferably, the multi-source information time series data and Markov transition field image data after dimensionality reduction are input into the trained dual-input improved dense connection network model for recognition, and the judgment probabilities of 9 working states are obtained, among which the working state with the highest probability is the fault diagnosis result of the current optical fiber current transformer.

[0041] Another object of the present invention is to provide a fiber optic current transformer fault diagnosis system based on an improved dense network, comprising an information acquisition unit, a time sliding unit, a principal component analysis unit, a Markov transfer unit, a dual-input improved dense connection network model unit, and an output unit, wherein:

[0042] The information acquisition unit is used to collect the optical fiber current transformer operating status monitoring data and the transformer final output signal data.

[0043] The time sliding unit is used to extract time domain features from the collected data using a time sliding window, and combine the operation status monitoring data to form high-dimensional multi-source information.

[0044] The principal component analysis unit uses the principal component analysis method to reduce the dimension of high-dimensional multi-source information, preliminarily extract features, and realize data fusion.

[0045] The Markov transfer unit converts the multi-source information after dimensionality reduction into Markov transfer field image data using the Markov transfer field.

[0046] The dual-input improved dense connection network model unit is used to input the reduced-dimensional multi-source information time series data and Markov transfer field image data into the trained dual-input improved dense connection network model to obtain the fault diagnosis result of the current optical fiber current transformer.

[0047] The output unit is used to output the current fault diagnosis result of the optical fiber current transformer.

[0048] Another object of the present invention is to provide an electronic device comprising: at least one processor, at least one memory, and a communication interface. The processor, memory, and communication interface communicate with each other. The memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the improved dense network-based optical fiber current transformer fault diagnosis method.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] This invention combines time-domain feature extraction from output and status detection signals with PCA multi-source information feature dimensionality reduction. It further proposes using the Markov transition field image method to enhance feature information within time series. This improves fault diagnosis types, enabling more refined component fault diagnosis and facilitating more accurate transformer repair and replacement. The proposed fiber-optic current transformer fault diagnosis network, based on a dual-channel improved densely connected network, achieves a recognition accuracy rate of 98.04%, ensuring diagnostic accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1It is a flow chart of the method disclosed in the present invention.

[0052] Figure 2 This is a structural diagram of the dual-input improved densely connected network model in the method disclosed in the present invention.

[0053] Figure 3 is the confusion matrix of the model test results.

[0054] Figure 4 It is a performance index diagram of the model test results. DETAILED DESCRIPTION

[0055] The present invention is further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0056] A fiber optic current transformer fault diagnosis method based on an improved dense network requires obtaining the fault output signal and multiple state monitoring signals of the fiber optic current transformer; extracting multiple time domain features of the output signal through a sliding window and combining them with multiple state monitoring signals; reducing the dimensionality of the combined high-dimensional information by principal component analysis; converting the reduced dimensionality multi-source information time series into image data using a Markov transition field; inputting the reduced dimensionality multi-source information time series data and the Markov transition field image data into a trained dual-input improved dense connection network model for diagnosis and identification, such as Figure 1 As shown, Figure 1 As shown, the specific steps include:

[0057] Step 1: Collect the optical fiber current transformer operating status monitoring data and the transformer final output signal data.

[0058] In another embodiment, multiple status monitoring signals such as temperature, humidity, power, voltage, current, compensation, etc. of multiple modules of the transformer and the final output signal data of the transformer are collected.

[0059] In another embodiment, the final output signal is collected through the output portion of the optical fiber current transformer at a sampling frequency of 10 kHz. Other sensors are used to collect status monitoring data of the optical fiber current transformer, including SLD temperature, optical module temperature, sensor ring temperature, circuit board temperature, TEC voltage, positive power supply voltage, system voltage, negative power supply voltage, optical module compensation, sensor ring compensation, SLD optical power, absolute optical power, system current, SLD current, and optical module humidity, at a sampling frequency of 1 Hz. The transformer operating states during information collection include: normal state, insulation structure contamination bias fault, optical path performance degradation drift fault, linear polarizer extinction ratio ratio fault, optical path interruption failure fault, light source drive attenuation ratio fault, light source temperature control drift fault, light source coupling misalignment failure fault, and light source chip breakdown failure fault.

[0060] Step 2: Use a time sliding window to extract time domain features from the collected data, and combine the operating status monitoring data to form high-dimensional multi-source information.

[0061] In another embodiment, the method for constructing high-dimensional multi-source information is as follows:

[0062] The time sliding window used was 1 second long, with a sliding step of 1 second. The extracted time-domain features included mean, standard deviation, variance, peak-to-peak value, absolute mean, root mean square amplitude, root mean square (RMS), kurtosis, crest factor, margin index, form factor, and pulse factor. The extracted multidimensional time-domain features of the output signal and the collected multidimensional condition monitoring data of the fiber optic current transformer were spliced ​​in time sequence to form a 27-dimensional multi-source information sequence.

[0063] Step 3: Use principal component analysis to reduce the dimensionality of high-dimensional multi-source information, preliminarily extract features, and achieve data fusion.

[0064] In another embodiment, a method for reducing the dimensionality of high-dimensional multi-source information using principal component analysis is as follows:

[0065] Step (3-1) uses the multi-source information features of the optical fiber current transformer to construct the sample matrix X:

[0066]

[0067] In the formula, m is the number of samples, and n is the feature dimension of each sample.

[0068] Step (3-2) calculates the mean of each column vector of the sample matrix X by column and standard deviation s i , for each element After standardization, we get the standardized sample matrix X':

[0069]

[0070] Step (3-3) calculates the covariance matrix R of the standardized sample matrix:

[0071]

[0072] Where, are the vector means of the i-th row or j-th column of the matrix X', respectively.

[0073] Steps (3-4) calculate the eigenvalues ​​and eigenvectors of the covariance matrix R, and arrange the eigenvalues ​​{λ i ,i=1,2,...,n}, calculate the cumulative contribution rate of the first i principal components:

[0074]

[0075] Steps (3-5) retain the principal components with cumulative contribution rates exceeding 90%, and construct a reduced-dimensional feature matrix from the corresponding eigenvectors and the sample matrix. Ultimately, principal component analysis is used to reduce the 27-dimensional multi-source information features of the fiber optic current transformer to 9 dimensions.

[0076] Step 4: Use the Markov transition field to convert the multi-source information after dimensionality reduction into Markov transition field image data.

[0077] In another embodiment, a method for converting dimensionality-reduced multi-source information into Markov transition field image data using a Markov transition field includes:

[0078] For the 9-dimensional mutual inductor multi-source information time series, each feature information time series X=[x1,x2,…,x n ], n represents the length of the time series. First, the subsequence X is evenly divided into Q intervals and the corresponding quantiles are determined. Each element x of the time series is assigned i The corresponding quantile q j Sequence number, calculate the quantile interval transfer along the time axis with a first-order Markov chain, construct the Markov transfer matrix, expand the transfer matrix by arranging each probability in time sequence, and obtain the Markov transfer field matrix M:

[0079]

[0080] Where m ij Indicates the quantile q i The state is transferred to quantile q j The conditional probability of the state, that is, m ij =P(x t ∈q i |x t-1 ∈q j ).

[0081] In another embodiment, the hyperparameter of the Markov transition field, the number of quantile intervals Q, is the sequence length. The obtained Markov transition field matrix is ​​generated in the form of a color image.

[0082] In step 5, the multi-source information time series data and Markov transfer field image data after dimensionality reduction are input into the trained dual-input improved dense connection network model to realize current transformer fault diagnosis.

[0083] In another embodiment, if Figure 2 As shown, the dual-input improved dense connection network model includes a time series input path, an image input path, a splicing layer, a dense structure, a batch normalization layer, a global average pooling layer, a fully connected layer, and an output layer. The time series input path and the image input path are respectively connected to the splicing layer, and the splicing layer, the dense structure, the third batch normalization layer, the global average pooling layer, the fully connected layer, and the output layer are connected in sequence. The time series input path includes the first convolutional layer, the first batch normalization layer, and the first pooling layer connected in sequence. The image input path includes the second convolutional layer, the second batch normalization layer, the second pooling layer, and the CBAM spatial attention module connected in sequence. The dense structure includes a first SEDB block, a first transition layer, a second SEDB block, a second transition layer, a third SEDB block, a third transition layer, and a fourth SEDB block connected in sequence, and the output of the first SEDB block enters the input of the first transition layer, the second transition layer, and the third transition layer respectively, the output of the first transition layer enters the input of the second transition layer and the third transition layer respectively, and the output of the second transition layer enters the input of the third transition layer. The first SEDB block, the second SEDB block, the third SEDB block, and the fourth SEDB block have 3, 6, 12, and 8 C-SE units respectively. The C-SE unit includes a first BN layer, a first Relu activation function, a 1×1 convolution layer, a second BN layer, a second Relu activation function, a 1×3 convolution layer, and an SE channel attention module connected in sequence.

[0084] like Figure 2 As shown in Figure 2, the formula for the input part of the dual-input improved dense connection network model is as follows:

[0085] x1=MaxPool(BN(Conv1D(x seq )));

[0086] x2=CBAM(MaxPool(BN(Conv2d(x matrix ))));

[0087] x in =x1⊕HAvePool(x2)⊕WAvePool(x2);

[0088] Where Conv1D() and Conv2D() are 1D and 2D convolutions respectively, BN() is batch normalization, MaxPool() is maximum pooling, CBAM() is spatial attention module, HAVEPool() and WAVEPool() are average pooling of height and width respectively, ⊕ is vector concatenation, x seq is the sequence input, x matrix is the Markov transition field input.

[0089] The formula for the model feature extraction backbone network is as follows:

[0090] x fea =SEDB(tran(SEDB(tran(SEDB(tran(SEDB(x in ));

[0091] Where SEDB() is the dense block and tran() is the transition layer.

[0092] The formula for the model output is as follows:

[0093] y out =softmax(FC(GAP(BN(x seq ))))

[0094] Where GAP() is global average pooling, FC() is full connection, and softmax() is the normalized exponential function.

[0095] The specific structure is as follows: the model input consists of a 9×64×1 time series input and a 9×64×64 Markov transition field image input. The time series input path includes a convolutional layer, a batch normalization layer, and a max pooling layer, which extract features and downsample the time series, converting the input to 64×32×1. The image input path includes a convolutional layer, a batch normalization layer, a max pooling layer, and a CBAM spatial attention module, which extracts features and downsamples the image data, using an attention mechanism to focus on local spatial information, converting the input to 64×16×16. The image feature map is then average pooled in its height and width dimensions to obtain 64×16×1 and 64×1×16 feature maps, respectively. These are then transposed and concatenated to 64×32×1 and merged with the sequence feature map to form a new sequence feature map of size 64×64×1. The merged sequence is then fed into a dense structure consisting of four SEDB blocks and three transition layers for deep feature extraction. Finally, a batch normalization layer, a global average pooling layer, a fully connected layer, and an output layer output the diagnostic type of the fiber current transformer. The four SEDB blocks have 3, 6, 12, and 8 C-SE units, respectively, with a channel growth rate of 16. Each C-SE unit consists of a batch normalization layer, a 1×1 convolutional layer, a Relu activation function, a batch normalization layer, a 1×3 convolutional layer, a Relu activation function, and an SE channel attention module.

[0096] Methods for training dual-input improved densely connected networks include:

[0097] A dual-input dataset was constructed by mapping the 9-dimensional multi-source information sequence to the generated Markov transition field image. Each dual-input sample was then labeled according to the transformer operating state sequence (0-8). The dataset was divided into training, validation, and test sets with an 8:1:1 ratio. The hyperparameters for model training were set as follows: a maximum number of training iterations of 50, a learning rate of 0.0001, a batch size of 4, an Adam optimizer, and a cross-entropy loss function. Network parameters were optimized based on the validation set accuracy to determine the final diagnostic network.

[0098] The dimensionality-reduced multi-source information time series data and Markov transition field image data are input into the trained dual-input improved densely connected network model for recognition, and the judgment probabilities of nine working states are obtained. Among them, the working state with the highest probability is the fault diagnosis result of the current fiber optic current transformer.

[0099] In another embodiment, a fiber optic current transformer fault diagnosis system based on an improved dense network is provided, comprising an information acquisition unit, a time sliding unit, a principal component analysis unit, a Markov transfer unit, a dual-input improved dense connection network model unit, and an output unit, wherein:

[0100] The information acquisition unit is used to collect the optical fiber current transformer operating status monitoring data and the transformer final output signal data.

[0101] The time sliding unit is used to extract time domain features from the collected data using a time sliding window, and combine the operation status monitoring data to form high-dimensional multi-source information.

[0102] The principal component analysis unit uses the principal component analysis method to reduce the dimension of high-dimensional multi-source information, preliminarily extract features, and realize data fusion.

[0103] The Markov transfer unit converts the multi-source information after dimensionality reduction into Markov transfer field image data using the Markov transfer field.

[0104] The dual-input improved dense connection network model unit is used to input the reduced-dimensional multi-source information time series data and Markov transfer field image data into the trained dual-input improved dense connection network model to obtain the fault diagnosis result of the current optical fiber current transformer.

[0105] The output unit is used to output the current fault diagnosis result of the optical fiber current transformer.

[0106] In another embodiment, an electronic device is provided, comprising: at least one processor, at least one memory, and a communication interface. The processor, memory, and communication interface communicate with each other. The memory stores program instructions executable by the processor, and the processor invokes the program instructions to execute the improved dense network-based optical fiber current transformer fault diagnosis method.

[0107] This embodiment can accurately identify various fault types of optical fiber current transformers and ensure the accurate operation of the transformer.

[0108] The effectiveness of the proposed multi-source information-based fiber optic current transformer fault diagnosis method was tested using samples from the test set. Specific test indicators include: the confusion matrix directly counts and displays the model's classification results for each test sample, and the classification results of each sample belong to one of four categories: true positive (TP), false positive (FP), false negative (FN), and true negative (TN). Accuracy refers to the probability of a sample being correctly predicted as the true value among all test samples. Precision refers to the probability that a sample actually belongs to that category among all samples predicted to be of that category. Recall refers to the probability that a sample actually belongs to that category among all samples actually belonging to that category. The F1 indicator can be used to comprehensively evaluate the model's precision and recall capabilities. The calculation formulas for the four indicators are as follows:

[0109]

[0110] like Figure 3 is the confusion matrix of the model on the test set, where 0-8 correspond to normal state, insulation structure contamination bias fault, optical path performance degradation drift fault, linear polarizer extinction ratio ratio change fault, optical path interruption failure fault, light source drive attenuation ratio change fault, light source temperature control drift fault, light source coupling misalignment failure fault, and light source chip breakdown failure fault. Four normal state faults were identified as insulation structure contamination bias faults, and one was identified as a linear polarizer extinction ratio ratio change fault; one insulation structure contamination bias fault was identified as a linear polarizer extinction ratio ratio change fault; two optical path performance degradation drift faults were identified as light source temperature control drift faults; both the linear polarizer extinction ratio ratio change fault and the optical path interruption failure fault were correctly identified; two light source drive attenuation ratio faults were identified as normal; one light source temperature control drift fault was identified as normal; the light source coupling misalignment failure fault was correctly identified; one light source chip breakdown failure fault was identified as normal, and two were identified as optical path interruption failure faults.

[0111] The network's diagnostic accuracy was 98.04%, Figure 4 Statistical charts show the model's precision, recall, and F1 index for each category. Overall, the model performs best for light source coupling misalignment failure, with all indicators achieving 100%. The model performs worst for light source temperature control drift, with all three indicators ranking at the bottom of the classification. In terms of precision, the model performs best for linear polarizer extinction ratio failure, optical path interruption failure, and light source coupling misalignment failure, all achieving 100%. The model performs worst for light source chip breakdown failure, achieving only 92.68%. In terms of recall, the model performs best for optical path performance degradation drift, light source driver attenuation ratio failure, light source coupling misalignment failure, and light source chip breakdown failure, all achieving 100%. The model performs worst for light source temperature control drift, achieving only 93.10%. In terms of F1 index, the model performs best for light source coupling misalignment failure, achieving 100%, while the model performs worst for light source temperature control drift, achieving 94.74%. The average precision of the model for all categories is 97.40%, the average recall is 97.63%, and the average F1 value is 97.47%, which proves that the designed model has achieved good results in fiber optic current transformer fault diagnosis.

[0112] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for diagnosing optical fiber current transformer faults based on an improved dense network, characterized in that: The following steps are involved: Step 1: Collecting the optical fiber current transformer operating status monitoring data and the transformer final output signal data; Step 2: Use a time sliding window to extract time domain features from the collected data, and combine the operating status monitoring data to form high-dimensional multi-source information; Step 3: Use principal component analysis to reduce the dimensionality of high-dimensional multi-source information, extract features preliminarily, and achieve data fusion; Step 4: Use the Markov transfer field to convert the multi-source information after dimensionality reduction into Markov transfer field image data; The method of converting the multi-source information after dimensionality reduction into Markov transition field image data using the Markov transition field includes: For the 9-dimensional mutual inductor multi-source information time series, each feature information time series is processed separately according to the feature dimension , Indicates the length of the time series; first, the subsequence Evenly divided into intervals and determine the corresponding quantiles, assigning each element of the time series The corresponding quantile Sequence number, calculate the quantile interval transfer along the time axis with a first-order Markov chain, construct the Markov transfer matrix, expand the transfer matrix by arranging each probability in time sequence, and obtain the Markov transfer field matrix : Where, Indicates the quantile State transfer to quantile The conditional probability of the state, that is ; Hyperparameters of the Markov transition field, the number of quantile intervals Take the sequence length; generate the obtained Markov transition field matrix into a color image form; In step 5, the multi-source information time series data and Markov transfer field image data after dimensionality reduction are input into the trained dual-input improved dense connection network model to realize current transformer fault diagnosis.

2. The optical fiber current transformer fault diagnosis method based on the improved dense network according to claim 1 is characterized in that: The dual-input improved dense connection network model includes a time series input path, an image input path, a splicing layer, a dense structure, a batch normalization layer, a global average pooling layer, a fully connected layer and an output layer. The time series input path and the image input path are respectively connected to the splicing layer, and the splicing layer, the dense structure, the third batch normalization layer, the global average pooling layer, the fully connected layer and the output layer are connected in sequence; the time series input path includes the first convolutional layer, the first batch normalization layer and the first pooling layer connected in sequence; the image input path includes the second convolutional layer, the second batch normalization layer, the second pooling layer and the CBAM spatial attention module connected in sequence; the dense structure includes the first SEDB block, the first transition layer, the second S EDB block, second transition layer, third SEDB block, third transition layer, fourth SEDB block, and the output of the first SEDB block enters the input of the first transition layer, the second transition layer, and the third transition layer respectively, the output of the first transition layer enters the input of the second transition layer and the third transition layer respectively, and the output of the second transition layer enters the input of the third transition layer. The first SEDB block, the second SEDB block, the third SEDB block, and the fourth SEDB block have 3, 6, 12, and 8 C-SE units respectively; the C-SE unit includes a first BN layer, a first Relu activation function, a 1×1 convolutional layer, a second BN layer, a second Relu activation function, a 1×3 convolutional layer, and an SE channel attention module connected in sequence.

3. The method for diagnosing faults of optical fiber current transformers based on an improved dense network according to claim 2, characterized in that: The method for step 2 includes: The time sliding window length used is Ns, and the sliding step size is Ns; The extracted time domain features include mean, standard deviation, variance, peak-to-peak value, absolute mean, root mean square amplitude, root mean square, kurtosis, crest factor, margin index, shape factor, and pulse factor; The extracted multi-dimensional time domain features of the output signal and the collected multi-dimensional status monitoring data of the optical fiber current transformer are spliced ​​in time sequence to form a 27-dimensional multi-source information sequence.

4. The method for diagnosing faults of optical fiber current transformers based on an improved dense network according to claim 3, characterized in that: The method of using principal component analysis in step 3 to reduce the dimensionality of high-dimensional multi-source information: Step (3-1) uses the multi-source information features of the optical fiber current transformer to construct a sample matrix : Where, is the number of samples, is the feature dimension of each sample; Step (3-2) calculates the sample matrix by column The mean of each column vector and standard deviation , for each element Standardization processing to obtain the standardized sample matrix : Step (3-3) calculates the covariance matrix of the standardized sample matrix : Where, , The matrices No. Line or vector mean of columns; Steps (3-4) calculate the covariance matrix The eigenvalues ​​and eigenvectors of are arranged in descending order. , before calculation The cumulative contribution rate of the principal components: Steps (3-5) retain the principal components when the cumulative contribution rate exceeds 90%, and construct a reduced-dimensional feature matrix from the corresponding eigenvectors and sample matrix; finally, the 27-dimensional multi-source information features of the optical fiber current transformer are reduced to 9 dimensions through principal component analysis.

5. The method for diagnosing faults of optical fiber current transformers based on an improved dense network according to claim 4, characterized in that: The collection of optical fiber current transformer operating status monitoring data and transformer final output signal data includes: The output signal is collected through the output part of the optical fiber current transformer; Collects fiber optic current transformer status monitoring data including SLD temperature, optical module temperature, sensor ring temperature, circuit board temperature, TEC voltage, positive power supply voltage, system voltage, negative power supply voltage, optical module compensation, sensor ring compensation, SLD optical power, absolute optical power, system current, SLD current, and optical module humidity; The working status of the mutual inductor during information collection includes: normal state, insulation structure contamination bias fault, optical path performance degradation drift fault, linear polarizer extinction ratio ratio fault, optical path interruption failure fault, light source drive attenuation ratio fault, light source temperature control drift fault, light source coupling misalignment failure fault, and light source chip breakdown failure fault.

6. The method for diagnosing optical fiber current transformer faults based on an improved dense network according to claim 5, characterized in that: Methods for training dual-input improved densely connected networks include: A dual-input dataset was constructed by mapping the 9-dimensional multi-source information sequence to the generated Markov transition field image. Each dual-input sample was then labeled according to the working state sequence of the mutual inductor (0-8). The dataset was divided into training, validation, and test sets at an 8:1:1 ratio. The hyperparameters in model training were set as follows: the maximum number of model training times was 50, the learning rate was 0.0001, the number of batch samples was 4, the optimizer was Adam, and the cross-entropy loss function was selected as the loss function. The network parameters were optimized based on the accuracy of the validation set to determine the final diagnostic network.

7. The method for diagnosing faults of optical fiber current transformers based on an improved dense network according to claim 6, characterized in that: The dimensionality-reduced multi-source information time series data and Markov transition field image data are input into the trained dual-input improved densely connected network model for recognition, and the judgment probabilities of nine working states are obtained. Among them, the working state with the highest probability is the fault diagnosis result of the current fiber optic current transformer.

8. A fiber optic current transformer fault diagnosis system based on an improved dense network, characterized in that: The optical fiber current transformer fault diagnosis method based on the improved dense network according to any one of claims 1 to 7 comprises an information acquisition unit, a time sliding unit, a principal component analysis unit, a Markov transfer unit, a dual-input improved dense connection network model unit, and an output unit, wherein: The information acquisition unit is used to collect the optical fiber current transformer operating status monitoring data and the transformer final output signal data; The time sliding unit is used to extract time domain features from the collected data using a time sliding window, and combine the operation status monitoring data to form high-dimensional multi-source information; The principal component analysis unit uses principal component analysis to reduce the dimensionality of high-dimensional multi-source information, preliminarily extract features, and achieve data fusion; The Markov transfer unit converts the multi-source information after dimensionality reduction into Markov transfer field image data using the Markov transfer field; The dual-input improved dense connection network model unit is used to input the multi-source information time series data and Markov transfer field image data after dimensionality reduction into the trained dual-input improved dense connection network model to obtain the fault diagnosis result of the current optical fiber current transformer; The output unit is used to output the current fault diagnosis result of the optical fiber current transformer.

9. An electronic device, characterized in that: include: At least one processor, at least one memory and a communication interface; the processor, memory and communication interface communicate with each other; The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the optical fiber current transformer fault diagnosis method based on the improved dense network according to any one of claims 1 to 7.

Images