Converter valve recording data extraction method based on adaptive filtering and data driving

Abstract

A converter valve recording data extraction method based on adaptive filtering and data driving belongs to the technical field of converter valve fault identification, and comprises the following steps: recording data of a converter valve under different working conditions are extracted and processed, and the recording data comprise core variable parameters under each converter valve fault type; based on the converter valve fault types, constructing a core variable parameter sensitivity analysis and calculation model, extracting the core variable parameter sensitivity degree under each converter valve fault type, and then performing sorting according to the core variable parameter sensitivity degree; a 3D-CNN three-dimensional convolutional neural network is constructed, corresponding three-dimensional time-frequency representation is constructed through a core variable parameter sensitive degree sorting result to serve as network input, then a self-attention module is introduced to highlight key features, and finally a corresponding high-dimensional feature vector capable of accurately representing the fault type is extracted from a network middle layer. Through improvement of an algorithm, accurate and rapid extraction of converter valve fault recording data characteristics is realized.

Description

Technical Field

[0001] This invention belongs to the field of converter valve fault identification technology, specifically relating to a converter valve waveform data extraction method based on adaptive filtering and data-driven approach. Background Technology

[0002] Effective processing and accurate fault identification of converter valve waveform data are crucial for ensuring their reliable operation. Currently, waveform data from converter valves manufactured by different companies exists in various formats, lacking a unified framework for reading and analysis, leading to incompatibility issues between data from different devices. In this context, the inefficiency of data preprocessing is significant. Furthermore, waveform data is susceptible to power grid harmonics and electromagnetic interference, and traditional fixed-parameter filtering algorithms cannot dynamically adapt to noise changes. In addition, fault identification often relies on manually extracted empirical features to fully correlate fault types with the inherent patterns of waveform data fluctuations. These problems directly result in delayed converter valve fault warnings and high false alarm rates, which can, in severe cases, lead to valve group shutdowns, power grid deficits, and compromised power supply reliability. Therefore, research on methods for extracting characteristics from converter valve waveform data is urgently needed. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a method for extracting waveform data of converter valves based on adaptive filtering and data driving. Through algorithm improvement, the characteristics of fault waveform data of converter valves can be extracted accurately and quickly.

[0004] The technical solution adopted in this invention is: a method for extracting waveform data of converter valves based on adaptive filtering and data-driven methods. The method includes the following steps: Step S10, extracting and processing waveform data of converter valves under different operating conditions, the waveform data including core variable parameters under each converter valve fault type, the core variable parameters including converter valve equipment parameters and power grid interference intensity; Step S20, based on the converter valve fault type, constructing a sensitivity analysis calculation model for core variable parameters, extracting the sensitivity of core variable parameters under each converter valve fault type, and then sorting them according to the magnitude of the sensitivity of core variable parameters; Step S30, constructing a 3D-CNN three-dimensional convolutional neural network, constructing a corresponding three-dimensional time-frequency representation based on the ranking result of the sensitivity of core variable parameters as the network input, then introducing a self-attention module to highlight key features, and finally extracting a high-dimensional feature vector that can accurately represent the fault type from the middle layer of the network.

[0005] The beneficial effects of this invention are as follows: This invention sequentially extracts and processes waveform data of converter valves under different operating conditions, analyzes the sensitivity of core variables to fault types and prioritizes influencing factors, and uses a three-dimensional time-frequency convolutional network for automated feature learning and extraction of high-dimensional feature vectors. This provides a unified data processing framework and multi-dimensional feature expression for accurate identification and rapid diagnosis of converter valve faults, making fault warnings more timely and significantly improving identification accuracy. It effectively overcomes the limitations of traditional methods in dealing with complex interference and dynamic noise, and has the characteristics of solid theoretical foundation and flexible technical application. It helps to improve the operational reliability of converter valve equipment, reduce maintenance costs, and ensure the safety, stability, and economical dispatch of the power grid. Attached Figure Description

[0006] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0007] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0008] See appendix Figure 1 This invention provides a method for extracting waveform data from converter valves based on adaptive filtering and data-driven approaches. The method includes the following steps: Step S10, extracting and processing waveform data from converter valves under different operating conditions. The waveform data includes core variable parameters for each converter valve fault type, including converter valve equipment parameters and power grid interference intensity; Step S20, constructing a sensitivity analysis calculation model for core variable parameters based on converter valve fault types, extracting the sensitivity of core variable parameters for each fault type, and then sorting them according to the magnitude of the sensitivity; Step S30, constructing a 3D-CNN three-dimensional convolutional neural network, constructing a corresponding three-dimensional time-frequency representation based on the sensitivity ranking results of core variable parameters as network input, then introducing a self-attention module to highlight key features, and finally extracting a high-dimensional feature vector that can accurately represent the fault type from the intermediate layer of the network.

[0009] In step S10, waveform data of the converter valve under different operating conditions is extracted and processed. The waveform data includes two core variables: converter valve equipment parameters and power grid interference intensity for each converter valve fault type. The converter valve equipment parameters include: U value Measured voltage across the converter valve (unit: kV), I arm The current flowing through the converter valve bridge arm (unit: kA), C eq Equivalent submodule capacitance (unit: mF), S modThe submodule switching status is a discrete sequence consisting of 0 (disconnected) and 1 (connected). Grid interference intensity includes: THD. u Total harmonic distortion (THD) of AC side voltage (%), F dip A composite index consisting of voltage sag depth and duration. Converter valve fault type F type This includes: Fault 1 is a submodule fault (IGBT open circuit, capacitor failure); Fault 2 is an inter-module synchronization fault (trigger pulse loss or delay); Fault 3 is other types of faults (such as valve tower water circuit blockage leading to overheating, reflected as thermal model parameter drift).

[0010] In step S20, based on the converter valve fault type, a sensitivity analysis calculation model for core variable parameters is constructed. The sensitivity of core variable parameters under each converter valve fault type is extracted, and then ranked according to the magnitude of their sensitivity. The basic idea of ​​this step is to analyze the sensitivity of each core variable parameter under a certain fault type, in order to extract several types of waveform data that are strongly correlated with a certain fault type of the converter valve. First, for a certain fault type F... type =c (c = 1, 2, 3), to obtain a large amount of waveform data when this type of fault occurs. Then, a sensitivity analysis calculation model for core variable parameters is constructed to quantify the impact of different core variable parameters on the accuracy of fault type judgment, so as to quantify the correlation strength between the core variable parameter and the fault type. Based on the quantification sensitivity of the contribution of core variable parameters to fault identification effect, the core logic is: if a core variable parameter is crucial to fault identification, removing it will lead to a significant decrease in the model's identification accuracy; conversely, if the core variable parameter is insignificant, removing it will result in minimal change in accuracy.

[0011] The process of constructing the sensitivity analysis calculation model for the core variable parameters in step S20 includes: Step 201: Construct the feature set V of the core variable parameters for each converter valve fault type. c Each parameter is a vector of length N. V c ={V1,V2,...,V K}={U value ,I arm C eq ,S mod THD u ,F dip}, Where K represents the number of core variable parameters, and c represents the fault type identifier; Step 202: Construct a fault identification model based on random forest. And fault identification accuracy model, In the formula, I is an indicator function, which takes the value 1 when the condition is met and 0 otherwise; Step 203: Remove the fault identification accuracy model after removing the kth core variable parameter. in V c,k This represents the feature set after removing the kth core variable parameter; Step 204, the formula for the sensitivity of the k-th variable to the fault type identification accuracy model. Importance k =Acc-Acc k , Importance k The larger the value, the greater the sensitivity of the k-th core variable parameter to the identification of this fault type; Step 205: Output the sensitivity ranking V′ of the core variable parameters for each fault type. V c ′={V (1) V (2) ,...,V (K)}, In the formula, V c ′ represents the sensitivity ranking of all core variable parameters in fault type c, V (1) V represents the core variable parameter with the highest sensitivity ranking. (K) The core variable parameter indicating the lowest sensitivity ranking; Step 206: Select the top three core variable parameters based on the sensitivity of each fault type, and use them as the set of key variables. V′ selected ={V (1) V (2) V (3)}

[0012] Step S30: Construct a 3D-CNN three-dimensional convolutional neural network. The corresponding three-dimensional time-frequency representation is constructed by ranking the sensitivity of core variable parameters as the network input. Then, a self-attention module is introduced to highlight key features. Finally, a high-dimensional feature vector that can accurately represent the fault type is extracted from the middle layer of the network.

[0013] First, each core variable parameter is divided into D consecutive segments of a fixed time window length T. Each segment contains L sampling points. A short-time Fourier transform is performed on the signal segment within each time window, and its mathematical expression is as follows: F = (L / 2) + 1, f = {0.1…F-1}, d = {0.1…D-1}, a = {0.1…L-1}, Where, x k (a) represents the core variable parameter V k The amplitude at the a-th sampling point, ω(·) is the window function, d is the time window index, f is the frequency index, H is the window step size, and F is the number of frequency points; This transformation yields the core variable parameter V. k Corresponding time-frequency matrix Matrix element X k (d,f) represents the spectral amplitude at time window d and frequency point f; Finally, the time-frequency matrices of all selected core variable parameters are stacked along the channel dimension to construct a three-dimensional time-frequency input. C=|V′ selected |; Using three-dimensional time and frequency as input to a 3D-CNN three-dimensional convolutional neural network, the convolution kernel W slides along the time and frequency dimensions to generate a feature map Z through convolution operations; The value of the g-th feature map in layer l at position (d, f) Calculated by the following formula: in, b represents the kernel weights. (l,g) As a bias term, C in σ represents the number of channels in the input feature map, and σ is a non-linear activation function.

[0014] To further enhance feature representation capabilities, a self-attention module is introduced after convolutional layer processing. This module calculates the attention weight for each feature vector in the feature map Z to reflect its importance for fault classification. First, the feature saliency s is calculated through autocorrelation, with its i-th element calculated as follows: in, Let E be the i-th feature vector after the feature map is expanded, and E be the feature dimension.

[0015] Subsequently, the saliency is transformed into normalized attention weights a using the Softmax function: Where N is the total number of feature vectors. The weighted feature maps Z′ are fused through a fully connected layer and mapped to a high-dimensional space to obtain the feature representation f used for classification: z′ i =a i·z i , f = φ FC (Z′), Where, φ FC (·) indicates multi-layer fully connected network operation.

[0016] The model uses the cross-entropy loss function as the optimization objective: Where C is the number of fault categories, y c One-hot encoding of the real label. This represents the class probabilities predicted by the model.

[0017] The model is trained using the backpropagation algorithm and gradient descent optimizer, continuously adjusting the network weights to gradually strengthen the correlation between intermediate layer features and fault types. The backpropagation algorithm calculates the gradient of the loss function with respect to the weight parameters of each layer of the network based on the chain rule.

[0018] For any layer weight parameter θ, its gradient is calculated as follows: The gradient indicates the direction in which the loss function changes with the weights. By employing a gradient descent optimizer and iteratively updating the network weights, the feature map Z′ output by the self-attention module can more accurately focus on time-frequency patterns highly correlated with the fault type, thereby enhancing the discriminative power of the features.

[0019] Finally, from the trained classification model, we obtain feature vectors that can directly represent the fault type and have strong discriminative power. Specifically, we select the feature map Z′ output by the self-attention module after the last convolutional layer and before the fully connected layer in the network as the intermediate feature source.

[0020] The feature map contains both the time-frequency structure information learned by convolution and the key feature distribution enhanced by the self-attention module.

[0021] To obtain a fixed-dimensional feature vector, the feature map is... Global average pooling is performed across time, frequency, and spatial dimensions to aggregate spatial information and obtain channel-level feature description vectors. Where, z′ d,f,c Let v be the value of the c-th fault type at location (d, f) in the feature map Z′. The resulting vector v c This refers to the fault feature vector extracted from the model.

[0022] The present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, characterized in that the processor executes the computer program to implement the steps of a converter valve waveform data extraction method based on adaptive filtering and data driving.

[0023] The present invention also provides a non-transitory computer-readable storage medium storing a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of a converter valve waveform data extraction method based on adaptive filtering and data driving.

[0024] The present invention also provides a computer program product, including a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of a converter valve waveform data extraction method based on adaptive filtering and data driving.

[0025] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for extracting waveform recording data from a converter valve based on adaptive filtering and data-driven methods, characterized in that, The method includes the following steps: Step S10: Extract and process the waveform data of the converter valve under different operating conditions. The waveform data includes the core variable parameters under each converter valve fault type. The core variable parameters include the converter valve equipment parameters and the power grid interference intensity. Step S20: Based on the converter valve fault type, construct a core variable parameter sensitivity analysis calculation model, extract the sensitivity of the core variable parameters under each converter valve fault type, and then sort them according to the magnitude of the core variable parameter sensitivity. Step S30: Construct a 3D-CNN three-dimensional convolutional neural network. The corresponding three-dimensional time-frequency representation is constructed by ranking the sensitivity of core variable parameters as the network input. Then, a self-attention module is introduced to highlight key features. Finally, a high-dimensional feature vector that can accurately represent the fault type is extracted from the middle layer of the network.

2. The method for extracting waveform data from a converter valve according to claim 1, characterized in that, The waveform data recorded by the flow valve in step S10 still needs to be cleaned. The data cleaning and analysis method is as follows: First, the 3σ-Grubbs fusion criterion is used to remove extreme outliers from the waveform data; Secondly, a time-series window linear-spline hybrid interpolation completion algorithm is used to complete the missing values ​​of the waveform data.

3. The method for extracting waveform data from a converter valve according to claim 1, characterized in that, The fault types of the converter valve include submodule faults, with the fault type identifier recorded as 1; inter-module synchronization faults, with the fault type identifier recorded as 2; and other types of faults, with the fault type identifier recorded as 3. The set of converter valve fault types is F. type =c, c={1, 2, 3}, where c represents the fault type identifier; The parameters of the converter valve equipment include the measured voltage U at both ends of the converter valve. value The current I flowing through the converter valve bridge arm arm Equivalent submodule capacitance value C eq Submodule switching status S mod ; The power grid interference intensity includes the total harmonic distortion (THD) of the AC side voltage. u The composite index F, consisting of voltage sag depth and duration. dip .

4. The method for extracting waveform data from a converter valve according to claim 1, characterized in that, The process of constructing the sensitivity analysis calculation model for the core variable parameters in step S20 includes: Step 201: Construct the feature set V of the core variable parameters for each converter valve fault type. c Each parameter is a vector of length N. V c ={V1,V2,...,V K }={U value ,I arm ,C eq ,S mod ,THD u ,F dip }, Where K represents the number of core variable parameters; Step 202: Construct a fault identification model based on random forest. And fault identification accuracy model, In the formula, I is an indicator function, which takes the value 1 when the condition is met and 0 otherwise; Step 203: Remove the fault identification accuracy model after removing the kth core variable parameter. in V c,k This represents the feature set after removing the kth core variable parameter; Step 204, the formula for the sensitivity of the k-th variable to the fault type identification accuracy model. Importance k =Acc-Acc k , Importance k The larger the value, the greater the sensitivity of the k-th core variable parameter to the identification of this fault type; Step 205: Output the sensitivity ranking V′ of the core variable parameters for each fault type. V′ c ={V (1) ,V (2) ,...,V (K) }, In the formula, V c ′ represents the sensitivity ranking of all core variable parameters in fault type c, V (1) V represents the core variable parameter with the highest sensitivity ranking. (K) The core variable parameter indicating the lowest sensitivity ranking; Step 206: Select the top three core variable parameters based on the sensitivity of each fault type, and use them as the set of key variables. V′ selected ={V (1) ,V (2) ,V (3) }。 5. The method for extracting waveform data from a converter valve according to claim 1, characterized in that, The process of constructing the corresponding three-dimensional time-frequency data from the sensitivity ranking results of the core variable parameters in step S30 includes: First, each core variable parameter is divided into D consecutive segments of a fixed time window length T. Each segment contains L sampling points. A short-time Fourier transform is performed on the signal segment within each time window, and its mathematical expression is as follows: F = (L / 2) + 1, f = {0.1…F-1}, d = {0.1…D-1}, a = {0.1…L-1}, Where, x k (a) represents the core variable parameter V k The amplitude at the a-th sampling point, ω(·) is the window function, d is the time window index, f is the frequency index, H is the window step size, and F is the number of frequency points; Secondly, through this transformation, the core variable parameter V is obtained. k Corresponding time-frequency matrix Matrix element X k (d,f) represents the spectral amplitude at time window d and frequency point f; Finally, the time-frequency matrices of all selected core variable parameters are stacked along the channel dimension to construct a three-dimensional time-frequency input. C=|V′ selected |; Using three-dimensional time and frequency as input to the 3D-CNN three-dimensional convolutional neural network, the convolution kernel W slides in the time and frequency dimensions to generate feature map Z through convolution operations.

6. The method for extracting waveform data from a converter valve according to claim 1, characterized in that, In step 30, a self-attention module is introduced after the convolutional layer processing. This module calculates the attention weight for each feature vector in the feature map Z to reflect its importance for fault classification.

7. The method for extracting waveform data from a converter valve according to claim 1, characterized in that, In step 30, a high-dimensional feature vector v that can accurately characterize the fault type is extracted from the intermediate layer of the network. c The expression: Where, z′ d,f,c Let Z′ be the value of the c-th fault type at position (d,f) in the feature map Z′, where Z′ represents the feature map output by the self-attention module.

8. An electronic device, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-7.

9. A non-transitory computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1-7.

Images