Converter valve saturable reactor fault identification method and system based on improved Gramb angle field

Abstract

The invention provides a converter valve saturable reactor fault identification method and system based on an improved Gramb angle field, and relates to the technical field of power equipment fault identification, and the method comprises the steps: collecting vibration voiceprint data during the operation of a target converter valve saturable reactor; performing data preprocessing on the vibration voiceprint data to obtain preprocessed time sequence data; performing feature extraction on the preprocessed time series data on the basis of an improved Gramb angle field and a principal component analysis method so as to convert the preprocessed time series data into two-dimensional image data; inputting the two-dimensional image data into a trained fault type identification model to obtain an identification result; and generating a fault diagnosis report according to the identification result. According to the method, the problem of accurate fault identification of the saturable reactor in a complex environment of a converter valve hall is effectively solved.

Description

Technical Field

[0001] The present invention relates to the technical field of power equipment fault identification, and in particular to a converter valve saturated reactor fault identification method and system based on an improved Gram angle field. Background Art

[0002] High-voltage direct current (HVDC) transmission technology has attracted attention in the field of power transmission because it can achieve long-distance, low-loss and high-efficiency power transmission. As a key device in the HVDC system, the operating status of the converter valve is directly related to the stability of the power grid and the effective distribution of energy. In a complex electromagnetic environment, vibrations and noises of various frequencies will be generated inside the converter valve, especially the saturated reactor. Under the action of high-frequency pulse voltage, the magnetic state of the iron core will change rapidly, which may cause the fasteners to loosen, thereby affecting the stable operation of the converter valve. Therefore, it is extremely important to monitor the vibration characteristics of the saturated reactor and the loose state of the fasteners. In the field of fault identification, especially in applications such as industrial equipment monitoring and health diagnosis, the efficiency and accuracy of processing high-dimensional data are crucial. The traditional Gramian Angle Field method has problems such as high computational complexity and noise interference in high-dimensional data.

[0003] Although the soundprint technology performs well in the test of a single saturated reactor, the following problems still exist in the actual application of the converter valve hall: the traditional short-term analysis method is difficult to capture the long-range dependency of the vibration soundprint signal of the saturated reactor. In addition, the internal structure of the converter valve is complex, and the collected sound signals are often aliased, and the existing feature extraction method cannot remove the high-dimensional data that does not contribute significantly to the angle calculation. Summary of the invention

[0004] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a method and system for identifying a saturated reactor fault in a converter valve based on an improved Gram angle field, which can effectively solve the problem of accurately identifying a saturated reactor fault in a complex environment of a converter valve hall.

[0005] To achieve the above object, the present invention provides the following solutions:

[0006] A converter valve saturated reactor fault identification method based on improved Gram angle field, comprising:

[0007] Collect the vibration soundprint data of the target converter valve saturated reactor during operation;

[0008] Performing data preprocessing on the vibration voiceprint data to obtain preprocessed time series data;

[0009] Based on the improved Gram angular field and principal component analysis method, feature extraction is performed on the pre-processed time series data to convert the pre-processed time series data into two-dimensional image data;

[0010] Inputting the two-dimensional image data into a trained fault type recognition model to obtain a recognition result;

[0011] A fault diagnosis report is generated according to the identification result.

[0012] Preferably, the data preprocessing includes signal denoising and normalization.

[0013] Preferably, based on the improved Gram's angle field and principal component analysis method, feature extraction is performed on the preprocessed time series data to convert the preprocessed time series data into two-dimensional image data, including:

[0014] Calculate the covariance matrix of the preprocessed time series data, extract the main components by eigendecomposition, select the first k main components with larger eigenvalues, and use the main components to reduce the dimension of the preprocessed time series data to obtain a feature space;

[0015] Calculate a Gram matrix based on the feature space;

[0016] Based on the Gram matrix, calculating the angle relationship using cosine similarity;

[0017] The angular relationship is mapped into a two-dimensional or three-dimensional angular field to form the two-dimensional image data.

[0018] Preferably, the training method of the fault type identification model includes:

[0019] Acquire a labeled sample data set, and establish a normal state and a fault state database based on the sample data set;

[0020] Dividing the data in the database into a training set, a validation set and a test set for training and testing the model;

[0021] The preset ResNet-50 deep network structure is trained, tested and verified using the training set, validation set and test set to obtain a trained fault type recognition model; each residual block in the fault type recognition model consists of two convolutional layers, the convolution kernel size is 3×3, and the activation function is ReLU; a batch normalization layer is also provided in the ResNet-50 deep network structure to improve the training efficiency and stability of the model.

[0022] Preferably, the loss function of the fault type identification model is a cross entropy loss function.

[0023] Preferably, after dividing the data in the database into a training set, a validation set and a test set, the method further comprises:

[0024] The training set is expanded by using data enhancement technology to obtain an expanded training set.

[0025] Preferably, the training method of the fault type identification model further includes:

[0026] Use Adam or SGD optimization algorithm to train the model and set the learning rate and decay strategy.

[0027] A converter valve saturated reactor fault identification system based on improved Gram angle field, comprising:

[0028] A data acquisition module is used to collect vibration soundprint data of the target converter valve saturated reactor during operation;

[0029] A data preprocessing module, used for performing data preprocessing on the vibration voiceprint data to obtain preprocessed time series data;

[0030] A feature extraction module, used for performing feature extraction on the pre-processed time series data based on an improved Gram's angle field and principal component analysis method, so as to convert the pre-processed time series data into two-dimensional image data;

[0031] A fault identification module, used for inputting the two-dimensional image data into a trained fault type identification model to obtain an identification result;

[0032] The result output module is used to generate a fault diagnosis report according to the recognition result.

[0033] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0034] The present invention provides a method and system for identifying faults of saturated reactors of converter valves based on improved Gram's angle field, the method comprising: collecting vibration soundprint data of the target converter valve saturated reactor during operation; performing data preprocessing on the vibration soundprint data to obtain preprocessed time series data; performing feature extraction on the preprocessed time series data based on improved Gram's angle field and principal component analysis method to convert the preprocessed time series data into two-dimensional image data; inputting the two-dimensional image data into a trained fault type recognition model to obtain a recognition result; and generating a fault diagnosis report according to the recognition result. The present invention effectively solves the problem of accurately identifying faults of saturated reactors in the complex environment of converter valve halls. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0036] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0037] Figure 2 A schematic diagram of a fault identification system framework provided by an embodiment of the present invention;

[0038] Figure 3 A schematic diagram of a GAF-PCA feature extraction framework provided in an embodiment of the present invention;

[0039] Figure 4 A schematic diagram of the Gram angular field transformation provided by an embodiment of the present invention;

[0040] Figure 5 Schematic diagram of the ResNet deep residual neural network model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] The purpose of the present invention is to provide a method and system for identifying faults of saturated reactors of converter valves based on an improved Gram angle field, which effectively solves the problem of accurately identifying faults of saturated reactors in a complex environment of a converter valve hall.

[0044] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] Figure 1 A flow chart of a method provided by an embodiment of the present invention, such as Figure 1 As shown, the present invention provides a method for identifying converter valve saturated reactor fault based on improved Gram angle field, comprising:

[0046] Step 100: Collect vibration soundprint data of the target converter valve saturated reactor during operation;

[0047] Step 200: preprocessing the vibration voiceprint data to obtain preprocessed time series data;

[0048] Step 300: Based on the improved Gram angular field and principal component analysis method, feature extraction is performed on the pre-processed time series data to convert the pre-processed time series data into two-dimensional image data;

[0049] Step 400: inputting the two-dimensional image data into the trained fault type recognition model to obtain a recognition result;

[0050] Step 500: Generate a fault diagnosis report based on the identification result.

[0051] Figure 2 A schematic diagram of a fault identification system framework provided by an embodiment of the present invention, such as Figure 2 As shown, the fault identification system of the present invention is mainly divided into the following modules: Data acquisition module: The vibration soundprint signals emitted by the saturated reactor of the converter valve are collected through the Sony ECM-VG1 gun-type electret condenser microphone, and these signals are transmitted to the central processing system for processing. Data preprocessing module: Due to the complex working environment of the equipment, the raw data collected by the sensor often contains a lot of noise. Therefore, it is necessary to perform preprocessing such as denoising and normalization on the collected vibration soundprint signals to eliminate the interference information in the data. Feature extraction module: Feature extraction is performed using the improved Gram angular field and principal component analysis (PCA) technology to convert the time series signal into two-dimensional image data. Fault identification module: The ResNet deep residual neural network is used to analyze the extracted features and identify the fault type. Result output module: According to the results of the model analysis, a fault diagnosis report is output to monitor the operating status of the saturated reactor in real time.

[0052] Specifically, the feature extraction part of the present invention adopts an improved Gram angular field fault recognition algorithm based on principal component analysis PCA technology. It mainly includes two parts. The first is to use Gram angular field combined with principal component analysis technology to separate the useful features and interference features in the one-dimensional time series vibration voiceprint signal to the greatest extent to prevent noise from interfering with the results, and then convert the parameters such as amplitude mutation, waveform complexity, and data change trend into a two-dimensional image for feature extraction; the second is to establish a database of normal state signals and fault state signals, and use the ResNet deep residual neural network algorithm to analyze the features extracted in the first step, and comprehensively diagnose the fault features.

[0053] Specifically, the implementation process of the method for extracting voiceprint signal features in this embodiment is as follows:

[0054] First, data collection and preprocessing are performed. The vibration soundprint data of the converter valve saturated reactor during operation is collected through sensors or equipment monitoring systems. Then the data is standardized so that the mean of the data is 0 and the variance is 1 to eliminate the influence of different dimensions. Secondly, the covariance matrix Cov(X) of the standardized data is calculated, and the eigenvalues ​​and corresponding eigenvectors of the covariance matrix are calculated through eigendecomposition or singular value decomposition, and the first k principal components are selected according to the eigenvalues. The k value is usually determined based on the cumulative variance explanation rate (such as 70%-90%). Use the selected principal components to reduce the dimension of the original data to obtain a new feature space: Z = X·W k , where W k is the selected eigenvector matrix. Finally, the Gram matrix is ​​calculated by using the reduced-dimensional data Z: G = Z T Z, use cosine similarity to calculate the angular relationship between samples: , the angle information is mapped into a two-dimensional or three-dimensional angle field to form a Grammar angle field map, and key features (such as angle distribution, clustering features, etc.) are extracted from the Grammar angle field.

[0055] Furthermore, the implementation process of the ResNet deep residual neural network algorithm of this embodiment is as follows:

[0056] First, the signal data processed in the previous step is marked, and a database of normal and fault states is established, which is divided into a training set, a validation set, and a test set. Then the deep residual network architecture is designed. According to the scale and feature complexity of the data set, the ResNet model in this paper uses the ResNet-50 network structure. The extracted feature matrix is ​​used as the input of ResNet, and the input shape needs to be adjusted according to the feature dimension. Secondly, in the design of the network layer, each residual block consists of two convolutional layers, the activation function is ReLU, and Batch Normalization is added. A global average pooling layer is added to reduce the parameters of the model and improve the computational efficiency; in the output layer, the Softmax activation function is used to output the probability distribution of the judgment result as normal or faulty. Finally, the cross entropy loss function is adopted, and the Adam or SGD optimization algorithm is used for model training, and the appropriate learning rate and its attenuation strategy are set to improve the convergence speed and stability. In addition, in order to prevent the model from overfitting, this embodiment uses Dropout and data enhancement technology to divide the training set and the validation set according to a certain ratio, and monitor the validation loss during the training process.

[0057] Figure 3It is a feature extraction section, which includes Gram angular field and principal component analysis PCA technology. First, this embodiment calculates the covariance matrix of the voiceprint signal data and extracts the main components through eigendecomposition. The first k principal components with larger eigenvalues ​​are selected to retain the key information in the data and reduce the interference of redundant noise. Through principal component analysis (PCA) dimensionality reduction, high-dimensional data can be mapped to a lower-dimensional feature space. Secondly, using the data processed by PCA, the angular relationship between the data is calculated to generate a Gram angular field diagram, such as Figure 4 As shown, the specific steps are as follows:

[0058] (1) Data preprocessing

[0059] Assume that the time series data is X=x1,x2,...,x N , where N is the number of all time points in the time series data. The original one-dimensional time series data is normalized to [0,1] to eliminate the influence of different dimensions on the results. The normalized value is but:

[0060]

[0061] Optionally, this embodiment also includes a process of denoising the original data, specifically using a wavelet denoising method, as follows:

[0062] Shift the original data to obtain a shifted signal;

[0063] Perform wavelet decomposition on the translated signal to obtain multiple wavelet coefficients;

[0064] Determine the filtering threshold according to the decomposition scale and length of the signal;

[0065] Constructing a denoising function according to the filtering threshold;

[0066] The denoising function is used to remove the noise of the translated signal to obtain the denoised original data signal.

[0067] Specifically, the filtering threshold of this embodiment is Among them, λ j represents the filtering threshold at the jth decomposition scale, σ j represents the standard deviation of the noise at the jth decomposition scale, N j represents the signal length at the jth decomposition scale, and j represents the decomposition scale.

[0068] Optionally, the filtering model of this embodiment is:

[0069]

[0070] Wherein, sign is the sign function, a is the first preset coefficient, b is the second preset coefficient, ω j,k represents the kth wavelet coefficient at the jth decomposition scale, Represents the wavelet coefficients after filtering.

[0071] Furthermore, in this embodiment, the wavelet coefficient ω is optimized by sparse representation. j,k , the specific formula is as follows:

[0072] min∥c∥1subject to∥Wx-c∥2≤∈

[0073] Where c is a sparse coefficient vector. The goal of sparse representation is to find a sparse coefficient vector c so that the signal can be approximately reconstructed by a linear combination of these coefficients. |c||1 is the L1 norm of c, that is, the sum of the absolute values ​​of all coefficients. By minimizing the L1 norm, this embodiment encourages sparse solutions, that is, most coefficients are zero or close to zero; W is a wavelet transform matrix. It converts the original signal x into the wavelet domain to obtain wavelet coefficients. Each row or column of the wavelet transform matrix W corresponds to a wavelet basis function, and x is the original signal vector. This embodiment hopes to approximate or reconstruct this signal through sparse representation. |Wx-c||2≤∈ is a constraint condition, indicating that the difference between the wavelet transformed signal Wx and the sparse coefficient c does not exceed a small error ∈ in the sense of the L2 norm. The L2 norm here is the Euclidean distance, which represents the error of signal reconstruction. ∈ is the allowed error threshold. It controls the accuracy of signal reconstruction and allows errors within a certain range in exchange for sparsity. Exemplarily, this embodiment sets ∈ as an estimated value of the noise level. This embodiment determines by analyzing the noise portion of the signal, for example, using the high frequency component of the signal or known noise characteristics.

[0074] (2) Polar coordinate conversion

[0075] Convert the normalized values ​​to polar coordinates and re-express the scaled time series in polar coordinates Right now:

[0076]

[0077] Where: t i For point x i Timestamp of i is the phase angle; r i is the polar coordinate radius; is the normalized and scaled X.

[0078] The above mapping equation has two important properties: Each time point data contains two pieces of information: the normalized value of the data point and its temporal position t i . Polar axis ri The relationship in time is retained, and the polar angle φ i The relationship in value is retained.

[0079] (3) Construct the Gram matrix

[0080] When the vector lengths are ignored, the inner product of two vectors can reflect the similarity between the two vectors. Therefore, this paper introduces GAF based on the Gram matrix and customizes a special inner product: <x1,x2<=cos(φ1+φ2), then:

[0081]

[0082] Compared with the original time series data, GAF adds information in one dimension, and its elements reflect the similarity of data between different moments. When time rolls, GAF converts each point of the time series data into the correlation relationship between this point and other points, and this relationship is measured by the custom inner product of GAF. When the values of the time series data are uniformly distributed in [-1,1], the values in the Gram matrix will also show a normal distribution, but the normal distribution is not conducive to applying convolutional neural networks to the result data. In addition, if the Gram matrix is used to define the inner product, it is difficult to distinguish Gaussian noise in the output data. Therefore, GAF is an inner product with a penalty term to eliminate the influence of Gaussian noise. In actual use, it is not necessary to calculate polar coordinates. This invention adopts the form of GASF (Gramian Angular Summation Field). Define I as the unit row vector [1,1,...,1], then:

[0083]

[0084] (4) Image generation

[0085] Take the angle matrix G represented by GAF as the pixel value of the image to generate a two-dimensional image. Each pixel value corresponds to the cosine value of the included angle between different moments in the time series data, thereby retaining the time information and dynamic characteristics of the data. These images reflect the angle distribution and clustering characteristics of the data and are used as the input of the subsequent deep learning model.

[0086] Figure 5It is a ResNet deep residual neural network model. After the feature extraction is completed, the fault recognition module analyzes the feature image through the ResNet neural network algorithm. The structure of the algorithm can effectively solve the gradient vanishing problem in the deep network through the residual block, thereby improving the training effect of the model. The specific steps are as follows: First, the extracted feature data is annotated to establish a database of normal state and fault state. The data set is divided into a training set, a validation set, and a test set for model training and testing. Secondly, the ResNet-50 deep network structure is used for training, and the input of the model is the extracted feature matrix. Each residual block consists of two convolutional layers with a convolution kernel size of 3×3 and an activation function of ReLU. A Batch Normalization layer is also added to the network to improve the training efficiency and stability of the model. Then, during the training process, the cross entropy loss function is used in combination with the Adam optimization algorithm to optimize the model. The learning rate adopts a dynamic adjustment strategy to improve the convergence speed of the model and prevent overfitting. At the same time, data enhancement technology is used to further expand the training set and enhance the generalization ability of the model. Finally, the trained model is used to identify the fault type of the saturated reactor and output the probability distribution of the fault. The system outputs a specific fault diagnosis report based on the probability distribution results.

[0087] The beneficial effects of the present invention are as follows:

[0088] (1) The improved feature extraction method proposed in the present invention has good diagnostic effect and excellent interpretability in the analysis of vibration soundprint signals of converter valve saturated reactors. The combined method of Gram angular field and PCA technology can capture long-range dependencies in the signal, enhance the correlation identification between signals, provide more reliable fault characteristics, and enable the model to understand the dynamic changes of the signal more comprehensively. At the same time, principal component analysis PCA can effectively remove redundant information, compress data dimensions by selecting the main components, and highlight the most valuable part of the signal for analysis in the feature extraction stage, reduce the interference of irrelevant features, and greatly reduce the computational burden.

[0089] (2) In the ResNet deep residual neural network model constructed by the present invention, the residual block connection allows the network to be deeper, overcoming the gradient vanishing problem in deep network training, so that the model can not only capture complex features, but also extract high-order representations of signals at a deeper level. In addition, ResNet can adaptively adjust the network weights through back propagation and optimization algorithms, gradually improve the accuracy of fault identification, and maintain high stability in noisy environments. In particular, when the complex internal structure of the converter valve causes signal aliasing, the robustness of fault feature identification is significantly enhanced.

[0090] (3) The efficient feature extraction and deep learning algorithm of the present invention can not only achieve rapid response and analysis of real-time data, which is suitable for practical applications in industrial sites, but also can expand the data set as the amount of data increases and improve the accuracy and adaptability of the model through retraining, thus ensuring the long-term effectiveness of the system.

[0091] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred 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 parts can be referred to the method part.

[0092] The principles and implementation methods of the present invention are described in this article using specific examples. The description of the above embodiments is only used to help understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for identifying converter valve saturated reactor fault based on improved Gram angle field, characterized in that: include: Collect the vibration soundprint data of the target converter valve saturated reactor during operation; Performing data preprocessing on the vibration voiceprint data to obtain preprocessed time series data; Based on the improved Gram angular field and principal component analysis method, feature extraction is performed on the pre-processed time series data to convert the pre-processed time series data into two-dimensional image data; Inputting the two-dimensional image data into a trained fault type recognition model to obtain a recognition result; A fault diagnosis report is generated according to the identification result.

2. The converter valve saturated reactor fault identification method based on improved Gram angle field according to claim 1, characterized in that: The data preprocessing includes signal denoising and normalization.

3. The converter valve saturated reactor fault identification method based on improved Gram angle field according to claim 1, characterized in that: Based on the improved Gram angle field and principal component analysis method, feature extraction is performed on the pre-processed time series data to convert the pre-processed time series data into two-dimensional image data, including: Calculate the covariance matrix of the preprocessed time series data, extract the main components by eigendecomposition, select the first k main components with larger eigenvalues, and use the main components to reduce the dimension of the preprocessed time series data to obtain a feature space; Calculate a Gram matrix based on the feature space; Based on the Gram matrix, calculating the angle relationship using cosine similarity; The angular relationship is mapped into a two-dimensional or three-dimensional angular field to form the two-dimensional image data.

4. The converter valve saturated reactor fault identification method based on improved Gram angle field according to claim 3, characterized in that: The training method of the fault type identification model includes: Acquire a labeled sample data set, and establish a normal state and a fault state database based on the sample data set; Dividing the data in the database into a training set, a validation set and a test set for training and testing the model; The preset ResNet-50 deep network structure is trained, tested and verified using the training set, validation set and test set to obtain a trained fault type recognition model; each residual block in the fault type recognition model consists of two convolutional layers, the convolution kernel size is 3×3, and the activation function is ReLU; a batch normalization layer is also provided in the ResNet-50 deep network structure to improve the training efficiency and stability of the model.

5. The converter valve saturated reactor fault identification method based on improved Gram angle field according to claim 4, characterized in that: The loss function of the fault type identification model is a cross entropy loss function.

6. The converter valve saturated reactor fault identification method based on improved Gram angle field according to claim 4, characterized in that: After dividing the data in the database into a training set, a validation set and a test set, the following steps are further included: The training set is expanded by using data enhancement technology to obtain an expanded training set.

7. The converter valve saturated reactor fault identification method based on improved Gram angle field according to claim 4, characterized in that: The training method of the fault type identification model also includes: Use Adam or SGD optimization algorithm to train the model and set the learning rate and decay strategy.

8. A converter valve saturated reactor fault identification system based on improved Gram angle field, characterized in that: include: A data acquisition module is used to collect vibration soundprint data of the target converter valve saturated reactor during operation; A data preprocessing module, used for performing data preprocessing on the vibration voiceprint data to obtain preprocessed time series data; A feature extraction module, used for performing feature extraction on the pre-processed time series data based on an improved Gram's angle field and principal component analysis method, so as to convert the pre-processed time series data into two-dimensional image data; A fault identification module, used for inputting the two-dimensional image data into a trained fault type identification model to obtain an identification result; The result output module is used to generate a fault diagnosis report according to the recognition result.

Images