Static var generator fault diagnosis method and system based on neural network

Abstract

The invention provides a static var generator fault diagnosis method and system based on a neural network. The method comprises the following steps: collecting operation parameter data of a static var generator SVG in a fault state and a normal state during operation; and preprocessing operation parameter data in the fault diagnosis database when the SVG operates in a fault state and a normal state. And constructing and training a neural network model of dominant faults and hidden faults based on the preprocessed data. And inputting operation parameter data in a test set into the trained LSTM model to predict dominant faults. And assigning the hidden fault characteristics of the SVG in the test set under the normal operation condition, inputting the hidden fault characteristics into the trained BP neural network model, and predicting the hidden fault of the SVG. According to the SVG fault diagnosis method, fault diagnosis is carried out on SVG operation parameter data through the constructed and trained neural network model, the SVG fault diagnosis efficiency is improved, dominant and hidden fault treatment strategies can be given, and the method is beneficial to assisting in rapidly removing faults.

Description

Technical Field

[0001] The invention relates to the technical field of power electronics, and in particular to a static VAR generator fault diagnosis method and system based on a neural network. Background Art

[0002] As an important reactive power compensation device in modern power systems, static VAR generators (SVGs) can quickly and accurately adjust reactive power, and play a vital role in maintaining grid voltage stability and improving power quality. However, in actual operation, SVGs may experience various faults, such as damage to power electronic devices, control circuit failure, connection reactor or transformer failure, etc. If these faults are not discovered and diagnosed in time, they will not only affect the normal operation of the SVG itself, causing the reactive power compensation function to fail, but may also pose a serious threat to the safe and stable operation of the entire power system.

[0003] At present, the existing SVG fault diagnosis methods have certain limitations. For example, some methods can only diagnose a single type of fault and cannot fully cover various fault conditions that may occur in SVG; some methods have low diagnostic accuracy and are prone to misjudgment or missed judgment. Secondly, the real-time performance of some diagnostic methods is poor, and they cannot accurately detect the fault and issue an alarm at the first time when the fault occurs. Thirdly, static VAR generators generally need to work in a high-voltage and high-power environment for a long time. The power switch tube works under high temperature and high frequency conditions for a long time, which will increase the probability of failure. When an IGBT open-circuit fault occurs, SVG can still run with the fault for a period of time, and the fault may not be noticed for a long time. However, this "running with a fault" situation cannot effectively provide reactive current to the power grid, and at the same time causes fluctuations in the voltage of the DC bus capacitor, affecting the performance of the compensation and bringing more serious losses. Therefore, research on fault diagnosis of SVG is of great significance.

[0004] Therefore, it is necessary to develop a method that can diagnose SVG faults comprehensively, accurately and in real time. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a static VAR generator fault diagnosis method and system based on a neural network to solve the deficiencies of the prior art described in the background technology.

[0006] To solve the above technical problems, the embodiments of the present invention provide the following technical solutions:

[0007] A static VAR generator fault diagnosis method based on a neural network comprises the following steps:

[0008] Collect the operating parameter data of the static VAR generator SVG in the fault state and normal state, record the corresponding processing data and fault result data in the fault state, collect the compensation current data of SVG in normal operation, and establish a fault diagnosis database;

[0009] Preprocess the data in the fault diagnosis database, including:

[0010] Perform data cleaning, data normalization, data standardization, and feature engineering on the operating parameter data; perform feature extraction on the pre-processed operating parameter data, and extract key features that can reflect the operating status of SVG by establishing a feature extraction model;

[0011] Extract keywords from processing data and fault result data, encode the extracted keywords, and map the encoded processing data with the fault result data to build a fault handling database;

[0012] For the compensation current data of SVG under normal operation, the amplitude and phase characteristics of the compensation current fundamental wave and each harmonic when a single-tube open-circuit fault occurs in SVG are extracted through Fourier transform, and the hidden fault characteristics are identified;

[0013] Based on the preprocessed operating parameter data and operating result data in the fault diagnosis database, an LSTM model data set is constructed, including a training set, a validation set, and a test set, which are divided according to a ratio of 70%:15%:15%, and then a fault prediction model of the LSTM neural network is constructed, and optimized based on an improved genetic algorithm, and the training set data is input into the LSTM model for model training; at the same time, based on the hidden fault characteristics identified by the SVG under normal operating conditions, the hidden fault characteristics are assigned and encoded, and the hidden fault characteristic values ​​are normalized, a BP neural network model is established, and the BP neural network model is trained based on the normalized values ​​of the hidden fault characteristics and the fault coding;

[0014] Input the running parameter data in the test set into the trained LSTM model, use the Softmax classifier to identify whether a fault occurs and perform multi-classification of the fault category, and obtain the confidence vector data of the fault diagnosis respectively, wherein the category corresponding to the highest confidence in the vector data is the diagnosis result, and input the diagnosis result into the fault handling database, and obtain the fault handling strategy through data mapping and matching;

[0015] The hidden fault features of the SVG in the test set under normal operation are assigned values ​​and input into the trained BP neural network model, the fault code value is calculated, the fault code value is decoded to obtain the hidden fault of the SVG under normal operation, and corresponding processing is performed based on the hidden fault.

[0016] Optionally, the processing of data cleaning, data normalization, data standardization, and feature engineering on the operating parameter data is specifically:

[0017] Data cleaning, including missing value processing, duplicate value processing, outlier value processing, and error value processing;

[0018] Data normalization, the formula is:

[0019]

[0020] Among them, x is the original sample data, x min is the minimum value in the sample data, x max is the maximum value in the sample data, and y is the normalized data;

[0021] Data standardization, including time series data standardization and data format standardization;

[0022] Feature engineering, including feature selection and feature combination construction.

[0023] Optionally, the feature extraction of the preprocessed operating parameter data specifically includes:

[0024] Time domain feature extraction, including mean value calculation, effective value calculation, value and peak-to-peak value determination, variance and standard deviation calculation;

[0025] Frequency domain feature extraction, including spectrum analysis, power spectrum density calculation, and harmonic analysis;

[0026] Time-frequency domain feature extraction, specifically using wavelet transform analysis;

[0027] Event-based feature extraction, including fault event features and operation event features;

[0028] Correlation feature extraction, including device correlation features and time correlation features.

[0029] Optionally, the keyword extraction of the processing data is specifically:

[0030] Perform data cleaning on the text content of the processing data to remove extra spaces, punctuation marks and abnormal characters in the text;

[0031] Perform grammatical analysis on the text content after data cleaning, including classification of Chinese nouns and technical English nouns;

[0032] Tag the text with parts of speech, match the tagged words and sentences with the pre-established vocabulary, and obtain initial keywords;

[0033] For the initial keywords, stop words are removed, synonyms are merged, and keyword weights are adjusted, where the weight order is maintenance object noun > maintenance action word > parameter quantifier.

[0034] Optionally, the keyword extraction is performed on the extracted processing data keywords and fault result data, the extracted keywords are encoded, and the encoded processing data and the fault result data are mapped, specifically:

[0035] Query the Unicode code corresponding to each Chinese character for the key words in the processing data, and convert the Unicode code into binary;

[0036] Convert Unicode code to binary and use dictionary encoding technology to compress data;

[0037] Create an array mapping between the encoded processing data keywords and the fault result data.

[0038] Optionally, the fault prediction model of the LSTM neural network is constructed, optimized based on an improved genetic algorithm, and the training set data is input into the LSTM model for model training, specifically:

[0039] The constructed LSTM model includes input layer, hidden layer and output layer. The input layer has at least 10 neurons, the hidden layer has three layers, the output layer has at least 5 neurons, and Softmax is used as the activation function. A fully connected layer is added, and a Dropout layer is added before the fully connected layer to prevent overfitting.

[0040] Construct an evaluation function for the improved genetic algorithm, use binary coding to initialize the population, and set various parameters; set the fitness function, calculate the fitness value of each individual, and retain the individual with the largest fitness; determine whether the algorithm termination conditions are met. If the convergence conditions are met, output the result, otherwise perform crossover, mutation, and selection operations; determine whether the result of the selection operation meets the algorithm termination conditions. If so, end the algorithm;

[0041] The constructed LSTM neural network model is used as the main architecture of the fault prediction model, the number of input layers is set to 3, and the output layer outputs the state matrix; RMSE is determined as the error evaluation standard, and RMSE is also used as the evaluation function of the improved genetic algorithm;

[0042] The improved genetic algorithm is used to dynamically calculate the fitness value of the step length of each LSTM training, and the step length with the optimal fitness value is determined as the LSTM neural network training step length;

[0043] Taking the operating parameter data in the training set as input, taking the fault result data corresponding to the operating parameter data in the training set as output, taking the cross entropy as the loss function of the model, training is performed according to the loss function and the improved genetic algorithm, and the weights between the neurons are continuously adjusted according to the value of the loss function, so that the value of the loss function is gradually reduced;

[0044] After each complete training data set traversal, the loss function value of the validation set is recorded and verified. When the loss function value meets the set threshold, the training ends and a trained LSTM neural network model is obtained.

[0045] Optionally, the cross entropy loss function is specifically:

[0046]

[0047] Among them, Y j is the true label of the jth sample, is the predicted probability of the jth sample.

[0048] Optionally, the fault handling strategy is obtained through data mapping and matching, and the fault handling strategy is obtained through mapping and matching based on a dictionary or a lookup table.

[0049] Optionally, the operating parameter data includes, the fault result data includes but is not limited to overcurrent, overvoltage, overtemperature, IGBT module damage, communication failure, control failure, the operating parameter data includes but is not limited to SVG output voltage value, SVG output current value, SVG reactive power output, SVG power factor, SVG operating temperature, SVG operating environment parameters.

[0050] The present invention also proposes a diagnostic system for implementing the above-mentioned static VAR generator fault diagnosis method based on neural network, comprising:

[0051] A data acquisition unit is used to collect operating parameter data of the static VAR generator SVG in a fault state and a normal state when the static VAR generator SVG is in operation, record the corresponding processing data and fault result data in the fault state, collect the compensation current data of the SVG in a normal operation, and establish a fault diagnosis database;

[0052] The data preprocessing unit is used to preprocess the data in the fault diagnosis database, including:

[0053] Perform data cleaning, data normalization, data standardization, and feature engineering on the operating parameter data; perform feature extraction on the pre-processed operating parameter data, and extract key features that can reflect the operating status of SVG by establishing a feature extraction model;

[0054] Extract keywords from processing data and fault result data, encode the extracted keywords, and map the encoded processing data with the fault result data to build a fault handling database;

[0055] For the compensation current data of SVG under normal operation, the amplitude and phase characteristics of the compensation current fundamental wave and each harmonic when a single-tube open-circuit fault occurs in SVG are extracted through Fourier transform, and the hidden fault characteristics are identified;

[0056] A prediction model building unit is used to build an LSTM model data set based on the preprocessed operation parameter data and operation result data in the fault diagnosis database, including a training set, a validation set, and a test set, which are divided according to the ratio of 70%:15%:15%, and then build a fault prediction model of the LSTM neural network, and optimize it based on the improved genetic algorithm, and input the training set data into the LSTM model for model training; at the same time, based on the hidden fault characteristics identified by the SVG under normal operation, assign and encode the hidden fault characteristics, and normalize the hidden fault characteristic values, establish a BP neural network model, and train the BP neural network model based on the normalized values ​​of the hidden fault characteristics and the fault coding;

[0057] The explicit fault diagnosis unit inputs the running parameter data in the test set into the trained LSTM model, uses the Softmax classifier to identify whether a fault occurs and performs multi-classification of the fault category, and obtains the confidence vector data of the fault diagnosis respectively, wherein the category corresponding to the highest confidence in the vector data is the diagnosis result, and the diagnosis result is input into the fault handling database, and the fault handling strategy is obtained through data mapping and matching;

[0058] The hidden fault diagnosis unit assigns hidden fault features of the SVG in the test set under normal operation and inputs them into the trained BP neural network model, calculates the fault code value, decodes the fault code value to obtain the hidden fault of the SVG under normal operation, and performs corresponding processing based on the hidden fault.

[0059] The beneficial effects of the above technical solution of the present invention are as follows:

[0060] 1. The present invention constructs a diagnostic database based on the operating parameter data of SVG in a fault state and a normal state, as well as the corresponding processing data and fault result data in a fault state. The preprocessed operating parameter data and operating result data in the fault diagnosis database construct an LSTM model data set, then construct an LSTM model and input the training set data into the LSTM model for model training, and then input the operating parameter data in the test set into the trained LSTM model, use the Softmax classifier to identify whether a fault occurs and perform multi-classification of the fault category, and obtain the confidence vector data of the fault diagnosis respectively, wherein the category corresponding to the highest confidence in the vector data is the diagnosis result, and finally the diagnosis result is input into the fault handling database, and the fault handling strategy is obtained through data mapping and matching. By constructing and training a neural network model to perform fault diagnosis on the SVG operating parameter data, the efficiency of SVG fault diagnosis is improved, and a fault handling strategy can be automatically given, which is conducive to assisting in rapid troubleshooting.

[0061] 2. Before the neural network model is constructed and trained, the present invention performs different preprocessing measures on the SVG operating parameter data and the corresponding processing data and fault result data under the fault state; among them, the operating parameter data is processed by data cleaning, data normalization, data standardization, and feature engineering; the preprocessed operating parameter data is subjected to feature extraction, and by establishing a feature extraction model, key features that can reflect the operating state of the SVG are extracted; finally, keywords of the processing data and fault result data are extracted, the extracted keywords are encoded, and the encoded processing data is mapped to the fault result data to construct a fault handling database. After these preliminary treatments, the accuracy of the neural network model construction and training and the data processing efficiency are improved.

[0062] 3. In addition to predicting obvious faults of SVG, the present invention also extracts the amplitude and phase characteristics of the compensation current fundamental wave and each harmonic when a single-tube open-circuit fault occurs in SVG based on the compensation current data of SVG under normal operation through Fourier transform, and identifies the characteristics of hidden faults. Based on the hidden fault characteristic data, a BP neural network model is constructed and trained to predict hidden faults, providing technical support for the prediction of hidden faults of SVG under normal operation, making the operation reliability of SVG higher.

[0063] 4. When constructing and training the LSTM model, the present invention introduces an improved genetic algorithm to dynamically adjust the training step size of the neural network, giving the neural network at different times different training step sizes, thereby reducing the training error and improving the accuracy of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1It is a flow chart of the static VAR generator fault diagnosis method based on neural network of the present invention;

[0065] Figure 2 It is a principle block diagram of the static VAR generator fault diagnosis system based on neural network of the present invention;

[0066] Figure 3 It is a circuit schematic diagram of a three-level SVG reactive power compensation system in the prior art;

[0067] Figure 4 for Figure 3 Specific circuit diagram of the three-level converter. DETAILED DESCRIPTION

[0068] In order to make the technical problems, technical solutions and advantages to be solved by the present invention more clear, a detailed description will be given below with reference to the accompanying drawings and specific embodiments.

[0069] like Figure 1 As shown, the present invention proposes a static VAR generator fault diagnosis method based on a neural network, comprising the following steps:

[0070] S1. Collecting operating parameter data of the static VAR generator SVG in a fault state and a normal state, recording corresponding processing data and fault result data in a fault state, collecting compensation current data of the SVG in a normal operation, and establishing a fault diagnosis database;

[0071] S2. Preprocessing the data in the fault diagnosis database, including:

[0072] Perform data cleaning, data normalization, data standardization, and feature engineering on the operating parameter data; perform feature extraction on the pre-processed operating parameter data, and extract key features that can reflect the operating status of SVG by establishing a feature extraction model;

[0073] Extract keywords from processing data and fault result data, encode the extracted keywords, and map the encoded processing data with the fault result data to build a fault handling database;

[0074] For the compensation current data of SVG under normal operation, the amplitude and phase characteristics of the compensation current fundamental wave and each harmonic when a single-tube open-circuit fault occurs in SVG are extracted through Fourier transform, and the hidden fault characteristics are identified;

[0075] S3. Based on the preprocessed operation parameter data and operation result data in the fault diagnosis database, an LSTM model data set is constructed, including a training set, a validation set, and a test set, which are divided according to a ratio of 70%:15%:15%, and then a fault prediction model of the LSTM neural network is constructed, and optimized based on an improved genetic algorithm, and the training set data is input into the LSTM model for model training; at the same time, based on the hidden fault features identified by the SVG under normal operating conditions, the hidden fault features are assigned and encoded, and the hidden fault feature values ​​are normalized, a BP neural network model is established, and the BP neural network model is trained based on the normalized values ​​of the hidden fault features and the fault coding;

[0076] S4. Input the running parameter data in the test set into the trained LSTM model, use the Softmax classifier to identify whether a fault occurs and perform multi-classification of the fault category, and obtain the confidence vector data of the fault diagnosis respectively, wherein the category corresponding to the highest confidence in the vector data is the diagnosis result, and input the diagnosis result into the fault handling database, and obtain the fault handling strategy through data mapping and matching;

[0077] S5. Assign values ​​to the hidden fault features of the SVG in the test set under normal operation and input them into the trained BP neural network model to calculate the fault code value, decode the fault code value to obtain the hidden fault of the SVG under normal operation, and perform corresponding processing based on the hidden fault.

[0078] The following is a detailed description of the specific implementation process of the above steps:

[0079] The SVG operation parameter data in step S1 include fault result data including but not limited to overcurrent, overvoltage, overtemperature, IGBT module damage, communication failure, and control failure, and the operation parameter data include but not limited to SVG output voltage value, SVG output current value, SVG reactive power output, SVG power factor, SVG operation temperature, and SVG operation environment parameters.

[0080] The processing data is the handling method adopted when encountering the above-mentioned SVG fault. The present invention records the fault handling process before establishing the database, such as "If the IGBT module is not blown up, disconnect the external connection of the IGBT module, use the diode and resistance ranges of the multimeter to measure the GE and CE terminals of the IGBT module, and the faulty IGBT can generally be found through comparative measurement", "Check whether the collector plug-in terminals on the unit drive board side and the IGBT side are detached"

[0081] Fault result data, including the definition of fault type after fault handling, such as XX overcurrent, XX overvoltage, XX overtemperature, IGBT module damage, XX communication failure, XX control failure, where "XX" represents the name of the object.

[0082] Step S2 performs data cleaning, data normalization, data standardization, and feature engineering on the operating parameter data, specifically:

[0083] (I) Data cleaning, including missing value processing, duplicate value processing, outlier processing, and error value processing, is the process of reviewing and verifying the collected data. The purpose is to identify and correct errors, missing values, duplicate data, outliers, and other problems in the data to improve data quality, including:

[0084] (1) Missing value processing includes:

[0085] Identify missing values: Determine which data is missing by checking whether there are empty values ​​(such as empty cells in a table) or specific placeholders (such as "NA" and "NULL") in the data set.

[0086] Fill missing values: including filling with mean, median or mode. For numerical data, the mean or median of the column data can be used to fill missing values; for categorical data, the mode (the value with the most occurrences) can be used to fill missing values.

[0087] (2) Handling Duplicate Data

[0088] Identify duplicate data: Compare each record in the data set to find records that are identical (same values ​​in all columns) or have the same values ​​in key columns.

[0089] Deleting duplicate data: Generally, one of the records is retained and the remaining duplicate records are deleted to reduce data redundancy and avoid bias in the results in subsequent analysis.

[0090] (3) Handling Error Data

[0091] Format error: For example, the date format should be "YYYY-MM-DD", but some data records are in "DD-MM-YYYY" or other incorrect formats, so they need to be converted to the correct format. For example, a column should be numeric data, but some text data is mixed in it. The text data needs to be converted to numeric data (if possible) or processed accordingly (such as deleting the wrong data records).

[0092] (4) Handling outliers

[0093] If the outlier is caused by data entry errors and has a significant impact on the analysis results, you can consider deleting it. If you can determine that the outlier is due to special circumstances but not errors (such as a power outage caused by an unexpected external reason), you can try to correct it through reasonable methods, such as taking the average of the adjacent normal data to replace the outlier.

[0094] Through effective data cleaning, data can be made more accurate, complete and standardized, thus providing a reliable foundation for subsequent in-depth data analysis, data mining and data-based decision-making.

[0095] (II) Data normalization, the formula is:

[0096]

[0097] Among them, x is the original sample data, x min is the minimum value in the sample data, x max is the maximum value in the sample data, and y is the normalized data. By performing a linear transformation on the original data, the value range of the data can meet specific requirements, such as unifying the original data features of different magnitudes and value ranges into a standard new value range that is easy to compare and subsequently process. For example, after normalizing the voltage amplitude related to SVG fault diagnosis, the neural network can process these data consistently, avoiding excessive influence of features with large values ​​on the training results.

[0098] (3) Data standardization, including time series data standardization and data format standardization.

[0099] Standardization of time series data. For time series data in the power system (such as changes in power load over time, voltage fluctuation series, etc.), it can be standardized by moving average, difference and other methods. Moving average can smooth short-term fluctuations in the data, and difference can remove the trend of the data. For example, for the power load time series, a 3-hour moving average can be used to obtain a smoother load curve, and then difference processing (current moment load minus previous moment load) can be performed to convert the non-stationary time series into a stationary series, making it easier for the neural network to learn its dynamic change law.

[0100] Standardize the data format to ensure that the power data formats from different sources are consistent. For example, the timestamp format should be unified. Some data timestamps may be in the format of "year-month-day-hour: minute: second", while others may be digital representations of timestamps, which need to be unified. At the same time, the units of data should also be standardized, such as the power unit is unified as kilowatt (kW) or megawatt (MW), and the voltage unit is unified as kilovolt (kV). This can avoid neural network training errors caused by confusing data formats.

[0101] (iv) Feature engineering, including feature selection and feature combination construction, including:

[0102] Feature selection, select the most valuable features for neural network training from a large number of power data features. In SVG fault prediction, in addition to power parameter data, external environmental data (temperature, humidity, dust, etc.) may have an important impact on the role of the fault. Correlation analysis (such as calculating the Pearson correlation coefficient) can be used to select features with high correlation with the target variable (such as the possibility of fault). For features with low correlation, they can be discarded to reduce the data dimension and improve the training efficiency of the neural network.

[0103] Feature combination construction combines the features of the original power data or constructs new features. For example, when analyzing the stability of the power system when SVG performs power compensation, the voltage amplitude and phase angle are combined into a complex voltage vector as a new feature. Or new features are constructed based on the operating principle of power equipment. For example, for transformers, the product of the load rate (the ratio of actual load power to rated power) and the oil temperature change rate is constructed as a new feature to reflect the comprehensive operating status of the transformer.

[0104] Step S2 extracts features from the preprocessed operating parameter data, specifically including:

[0105] Time domain feature extraction, including average value calculation, effective value calculation, value and peak-to-peak value determination, variance and standard deviation calculation, the specific content is not described in detail;

[0106] Frequency domain feature extraction, including spectrum analysis, power spectrum density calculation, and harmonic analysis. The specific contents will not be described in detail.

[0107] Time-frequency domain feature extraction specifically uses wavelet transform analysis; wavelet transform has the characteristics of multi-resolution analysis. It can analyze the time-frequency characteristics of signals at different scales (corresponding to different scaling factors), and can capture the high-frequency detail information of the signal as well as analyze the low-frequency trend information of the signal.

[0108] Event-based feature extraction, including fault event features and operation event features, such as external power grid failures and irregular operation factors, which affect the normal operation of SVG;

[0109] Correlation feature extraction includes device correlation features and time correlation features, such as device status information related to SVG power compensation and characteristics of the device operating time period. These features can indirectly affect the working performance of SVG.

[0110] In step S2, keywords are extracted from the processing data, specifically:

[0111] Perform data cleaning on the text content of the processing data to remove redundant spaces, punctuation, and abnormal characters in the text; for example, replace multiple consecutive spaces in the text with a single space, and delete some meaningless special characters (such as tabs, line breaks, etc., unless they have specific meanings in the text).

[0112] Perform grammatical analysis on the text content after data cleaning, including classification of Chinese nouns and technical English nouns; grammatical analysis is mainly used to distinguish professional and technical nouns in the processing data to avoid these nouns interfering with subsequent data training.

[0113] Perform part-of-speech tagging on the text, match the tagged words and sentences with the pre-established vocabulary, and obtain the initial keywords. Perform part-of-speech tagging on the text, that is, determine the part of speech of each word (such as noun, verb, adjective, etc.), and then determine the keywords according to specific rules.

[0114] For the initial keywords, we remove stop words, merge synonyms, and adjust keyword weights, where the weight order is maintenance object nouns > maintenance action words > parameter quantity words. By adjusting the keyword weights, we can locate the key steps in the daily troubleshooting process more quickly, which is conducive to improving the accuracy of model training. The most critical thing in the maintenance process is to locate the object first, and then how to repair it. The number of parameters can be obtained through on-site debugging or testing.

[0115] In step S2, keywords are extracted from the extracted processing data keywords and fault result data, the extracted keywords are encoded, and a mapping is established between the encoded processing data and the fault result data, specifically:

[0116] Query the Unicode code corresponding to each Chinese character for the key words in the processing data, and convert the Unicode code into binary. Unicode (Unicode, Unicode, unicode) is an industry standard in the field of computer science. It organizes and encodes most of the world's writing systems, allowing computers to process and display characters in multiple languages.

[0117] Dictionary encoding technology is used to compress the data when converting Unicode code to binary. Since the binary data after the text keyword conversion is relatively large, data compression can effectively store and transmit data, thereby reducing the computing load and improving the efficiency of the algorithm.

[0118] The encoded processing data keywords and fault result data are mapped to an array to provide support for finding the corresponding solution strategy through the fault result after the fault result is diagnosed by the neural network. Array mapping is used in real time to define an array containing the SVG fault processing process, which contains different keyword attributes (such as recovery, replacement, welding, overcurrent, overvoltage, etc.) of the text sequence after each SVG fault processing process is processed. Array mapping is equivalent to defining a function and associating it with the fault result. The attributes of each SVG fault processing process are recombined or adjusted according to the new fault handling requirements, so as to convert it into a mapping array of another situation, such as replacing the keyword attribute "replacement" with "welding" to meet the needs of this situation.

[0119] In step S3, a fault prediction model of the LSTM neural network is constructed and optimized based on the improved genetic algorithm, and the training set data is input into the LSTM model for model training, specifically:

[0120] The constructed LSTM model includes input layer, hidden layer and output layer. The input layer has at least 10 neurons, the hidden layer has three layers, the output layer has at least 5 neurons, and Softmax is used as the activation function. A fully connected layer is added, and a Dropout layer is added before the fully connected layer to prevent overfitting.

[0121] Construct an evaluation function for the improved genetic algorithm, use binary coding to initialize the population, and set various parameters; set the fitness function, calculate the fitness value of each individual, and retain the individual with the largest fitness; determine whether the algorithm termination conditions are met. If the convergence conditions are met, output the result, otherwise perform crossover, mutation, and selection operations; determine whether the result of the selection operation meets the algorithm termination conditions. If so, end the algorithm;

[0122] The constructed LSTM neural network model is used as the main architecture of the fault prediction model, the number of input layers is set to 3, and the output layer outputs the state matrix; RMSE is determined as the error evaluation standard, and RMSE is also used as the evaluation function of the improved genetic algorithm;

[0123] The improved genetic algorithm is used to dynamically calculate the fitness value of the step length of each LSTM training, and the step length with the optimal fitness value is determined as the LSTM neural network training step length;

[0124] Taking the operating parameter data in the training set as input, taking the fault result data corresponding to the operating parameter data in the training set as output, taking the cross entropy as the loss function of the model, training is performed according to the loss function and the improved genetic algorithm, and the weights between the neurons are continuously adjusted according to the value of the loss function, so that the value of the loss function is gradually reduced;

[0125] After each complete training data set traversal, the loss function value of the validation set is recorded and verified. When the loss function value meets the set threshold, the training ends and a trained LSTM neural network model is obtained.

[0126] Among them, the above cross entropy loss function is specifically:

[0127]

[0128] Among them, Y j is the true label of the jth sample, is the predicted probability of the jth sample.

[0129] From the perspective of information theory, cross entropy can be understood as a measure of the "distance" between two probability distributions. In classification tasks, the probability distribution of the true label is determined (under one-hot encoding, only one category has a probability of 1, and the rest are 0), and the model predicts an estimated probability distribution. The cross entropy loss function is to make the probability distribution predicted by the model as close to the true probability distribution as possible. When the probability distribution predicted by the model is exactly the same as the true probability distribution, the cross entropy loss value is 0. As the difference between the two increases, the cross entropy loss value increases, which prompts the model to continuously adjust parameters to reduce the loss value, thereby improving the accuracy of the prediction.

[0130] Step S3 assigns and encodes the hidden fault features based on the hidden fault features identified by the SVG under normal operation, normalizes the hidden fault feature values, establishes a BP neural network model, and trains the BP neural network model based on the hidden fault feature normalization values ​​and fault codes, specifically:

[0131] When an IGBT open circuit fault occurs, the SVG can still operate with the fault for a period of time, and the fault may not be noticed for a long time. However, this "operation with fault" situation cannot effectively provide reactive current to the grid, and at the same time causes fluctuations in the DC bus capacitor voltage, affecting the compensation performance. Therefore, the present invention defines the diagnosis of SVG hidden faults.

[0132] IGBT fault types are generally divided into open circuit faults and short circuit faults, among which the causes of IGBT open circuit faults are mainly divided into overcurrent burnout and drive circuit breakage. In the power electronic circuit, each IGBT is connected in series with a fuse. When an IGBT short-circuit fault occurs, the faulty IGBT branch current is too large, which will burn out the series fuse in a very short time, so the IGBT short-circuit fault eventually manifests as an open circuit fault. Therefore, the IGBT faults considered in the present invention are all open circuit faults.

[0133] When an open-circuit fault occurs in the IGBT, the AC voltage waveform output by the diode-clamped three-level topology will be significantly distorted, and when different power tubes fail, the voltage waveform distortion shapes have their own characteristics. Therefore, the output voltage of the three-level inverter can reflect the operating status of the system. Since the multi-level voltage output by the SVG is converted into the output compensation current when it is working, the operating status of the IGBT can also be reflected by detecting the compensation current. Similarly, the compensation current output by the three-level topology changes the phase and harmonics of the grid current, so when an IGBT open-circuit fault occurs in the SVG, the grid current will also be distorted. Because three current sensors are required for current tracking in the SVG system, and if the voltage signal is selected as the fault sampling point, the system needs to add an additional voltage Hall to sample the phase voltage or line voltage. This method not only increases the cost, but also occupies hardware space. Therefore, the present invention selects the three-phase compensation current as the fault sampling point of the three-level SVG. Figure 3 As shown, it is a circuit schematic diagram of a three-level SVG reactive power compensation system in the prior art. Figure 4 for Figure 3 The specific circuit diagram of the three-level converter is shown in FIG. 1 , and the three-level converter is the core executive device of SVG for reactive power compensation. In the figure, S1a, S2a, S3a, S4a and S1b, S2b, S3b, S4b and S1c, S2c, S3c, S4 are IGBT devices.

[0134] According to the test, when a single-tube open-circuit fault occurs in phase A, not only the waveform of the fault phase is distorted, but also the current of the non-fault phase is affected. When S1a fails, the current waveform of the negative half cycle is basically unchanged, and the positive half cycle only works normally for half of the time; when S2a fails, there is almost no waveform output in the positive half cycle, and the waveform of the negative half cycle is basically unchanged; when S3a fails, there is almost no current waveform in the negative half cycle, and the positive half cycle is basically unchanged; when S4a fails, the basic current amplitude of the positive half cycle is basically unchanged, and the negative half cycle only works normally for half of the time. According to the analysis, each fault waveform has its own characteristics. The SVG compensation current can be used as a fault sampling point. Next, the fault condition is subjected to FFT analysis for the extraction of fault parameters.

[0135] Therefore, the present invention first extracts the amplitude and phase features of the compensation current fundamental wave and each harmonic of the SVG when a single-tube open-circuit fault occurs through Fourier transform FFT from the compensation current data of the SVG under normal operation, and identifies the hidden fault features. At the same time, based on the hidden fault features identified by the SVG under normal operation, the hidden fault features are assigned and encoded, and the hidden fault feature values ​​are normalized to establish a BP neural network model, and the BP neural network model is trained based on the normalized values ​​of the hidden fault features and the fault coding.

[0136] Under normal circumstances, after the amplitude and phase angle are obtained through FFT analysis, the amplitude and phase (including DC component) of the first second harmonic can be used to identify the fault characteristics. In order to further improve the accuracy of diagnosis, the present invention selects the three-phase compensation currents A, B, and C as the fault characteristic sampling points, and extracts the DC component D, fundamental amplitude Al, fundamental phase angle Bl, and second harmonic phase angle B2 of each phase current as fault characteristic signals. For the convenience of analysis, 12 fault states can be divided according to the fault tube sequence number. If the SVG system is regarded as a special fault state when it is operating normally without fault, there are 13 fault states in total. Each fault state consists of 12 characteristic values. Before the fault waveform characteristic values ​​are trained with the neural network, each fault condition needs to be fault-coded. Next, the 12 characteristic values ​​of each fault state are normalized, and then a BP neural network model is established with the corresponding fault code.

[0137] Step S4 inputs the running parameter data in the test set into the trained LSTM model, uses the Softmax classifier to identify whether a fault occurs and performs multi-classification of the fault category, and obtains the confidence vector data of the fault diagnosis, where the category corresponding to the highest confidence in the vector data is the diagnosis result.

[0138] In multi-class classification problems, our goal is to classify input data into multiple different categories. For example, in image classification tasks, we need to classify input images into different animal categories such as cats, dogs, rabbits, etc.; in the text classification task of the present invention, we need to classify input text into different categories such as voltage fault, current fault, temperature fault, communication fault, etc.

[0139] The Softmax classifier makes classification decisions based on probability distribution. It converts the output value of the last layer (output layer) of the neural network into probability values ​​corresponding to each category, and then determines the category to which the input data belongs based on these probability values. First, the input data passes through the previous layers of the neural network (such as the input layer, hidden layer, etc.) for feature extraction and nonlinear transformation. This part of the calculation is similar to the general neural network calculation process. Through the weighted summation between neurons in each layer and the action of the activation function, the original input data is gradually converted into a feature representation that is more suitable for classification decisions. After being processed by the previous layers, the data reaches the output layer. Assume that the output layer has N neurons, corresponding to N possible categories. For a certain input data that is processed by the neural network and reaches the output layer, the output value of each neuron is recorded as z j (where j = 1, 2, ..., N). Next, these output values ​​z of the output layer j Apply the Softmax function to convert it into the probability value corresponding to each category. The calculation formula of the Softmax function is as follows:

[0140]

[0141] Among them, p j is the probability that the input data is judged as category j, and e is the base of the natural logarithm (approximately equal to 2.71828). In other words, the original output value z of the output layer is converted to j Normalization is performed so that the probability values ​​corresponding to all categories are p j The sum is 1, so we get a probability distribution about the possibility of each category.

[0142] According to the probability value p corresponding to each category j The present invention adopts the maximum probability decision for classification decision: the category with the largest probability value is selected as the category to which the input data belongs. That is, if for a certain input data, among the probability values ​​obtained after the Softmax function processing, p k is the largest (where k is a category index), then the input data is determined to belong to category k.

[0143] Finally, the diagnosis results are input into the fault handling database, and the fault handling strategy is obtained through data mapping and matching. The fault handling strategy is obtained by mapping and matching based on dictionaries or lookup tables. The dictionary contains a series of keys and corresponding values. In the dictionary-based mapping and matching, the input data elements are compared with the keys in the dictionary according to the known key-value pair relationship. After finding the matching key, the corresponding value is obtained to achieve data mapping and matching. The lookup table (LUT) is essentially a predefined data structure used to store the correspondence between specific data elements, so that mapping and matching operations can be quickly performed in the subsequent processing process to find the output result corresponding to the input data.

[0144] Step S5 assigns the hidden fault features of the SVG in the test set under normal operation and inputs them into the trained BP neural network model, calculates the fault code value, decodes the fault code value to obtain the hidden fault of the SVG under normal operation, and performs corresponding processing based on the hidden fault. This is the hidden fault diagnosis step of SVG. Based on the BP neural network model trained in step S3, by inputting the hidden fault feature assignment data of the SVG in the test set under normal operation, fault prediction can be achieved. Due to the small number of fault handling types, the present invention does not independently establish a corresponding handling strategy database.

[0145] like Figure 2 As shown, the present invention also proposes a diagnostic system for implementing the above-mentioned static VAR generator fault diagnosis method based on neural network, comprising:

[0146] The data acquisition unit 101 is used to collect operating parameter data of the static VAR generator SVG in a fault state and a normal state, record the corresponding processing data and fault result data in the fault state, collect the compensation current data of the SVG in the normal operation, and establish a fault diagnosis database;

[0147] The data preprocessing unit 102 is used to preprocess the data in the fault diagnosis database, including:

[0148] Perform data cleaning, data normalization, data standardization, and feature engineering on the operating parameter data; perform feature extraction on the pre-processed operating parameter data, and extract key features that can reflect the operating status of SVG by establishing a feature extraction model;

[0149] Extract keywords from processing data and fault result data, encode the extracted keywords, and map the encoded processing data with the fault result data to build a fault handling database;

[0150] For the compensation current data of SVG under normal operation, the amplitude and phase characteristics of the compensation current fundamental wave and each harmonic when a single-tube open-circuit fault occurs in SVG are extracted through Fourier transform, and the hidden fault characteristics are identified;

[0151] The prediction model construction unit 103 is used to construct an LSTM model data set based on the preprocessed operation parameter data and operation result data in the fault diagnosis database, including a training set, a validation set, and a test set, which are divided according to the ratio of 70%:15%:15%, and then construct a fault prediction model of the LSTM neural network, and optimize it based on the improved genetic algorithm, and input the training set data into the LSTM model for model training; at the same time, based on the hidden fault characteristics identified by the SVG under normal operation, the hidden fault characteristics are assigned and encoded, and the hidden fault characteristic values ​​are normalized to establish a BP neural network model, and the BP neural network model is trained based on the normalized values ​​of the hidden fault characteristics and the fault coding;

[0152] The explicit fault diagnosis unit 104 inputs the running parameter data in the test set into the trained LSTM model, uses the Softmax classifier to identify whether a fault occurs and performs multi-classification of the fault category, and obtains the confidence vector data of the fault diagnosis respectively, wherein the category corresponding to the highest confidence in the vector data is the diagnosis result, and the diagnosis result is input into the fault handling database, and the fault handling strategy is obtained through data mapping and matching;

[0153] The hidden fault diagnosis unit 105 assigns hidden fault features of the SVG in the test set under normal operation and inputs them into the trained BP neural network model, calculates the fault code value, decodes the fault code value to obtain the hidden fault of the SVG under normal operation, and performs corresponding processing based on the hidden fault.

[0154] When the above module is in operation, the content of the static VAR generator fault diagnosis method based on neural network is adopted, which will not be repeated here.

[0155] The above is 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 static VAR generator fault diagnosis method based on neural network, characterized in that: The following steps are involved: Collect the operating parameter data of the static VAR generator SVG in the fault state and normal state, record the corresponding processing data and fault result data in the fault state, collect the compensation current data of the SVG in the normal operation, and establish a fault diagnosis database; Preprocess the data in the fault diagnosis database, including: Perform data cleaning, data normalization, data standardization, and feature engineering on the operating parameter data; perform feature extraction on the pre-processed operating parameter data, and extract key features that can reflect the operating status of SVG by establishing a feature extraction model; Extract keywords from processing data and fault result data, encode the extracted keywords, and map the encoded processing data with the fault result data to build a fault handling database; For the compensation current data of SVG under normal operation, the amplitude and phase characteristics of the compensation current fundamental wave and each harmonic when a single-tube open-circuit fault occurs in SVG are extracted through Fourier transform, and the hidden fault characteristics are identified; Based on the preprocessed operating parameter data and operating result data in the fault diagnosis database, an LSTM model data set is constructed, including a training set, a validation set, and a test set, which are divided according to a ratio of 70%:15%:15%, and then a fault prediction model of the LSTM neural network is constructed, and optimized based on an improved genetic algorithm, and the training set data is input into the LSTM model for model training; at the same time, based on the hidden fault characteristics identified by the SVG under normal operating conditions, the hidden fault characteristics are assigned and encoded, and the hidden fault characteristic values ​​are normalized, a BP neural network model is established, and the BP neural network model is trained based on the normalized values ​​of the hidden fault characteristics and the fault coding; Input the running parameter data in the test set into the trained LSTM model, use the Softmax classifier to identify whether a fault occurs and perform multi-classification of the fault category, and obtain the confidence vector data of the fault diagnosis respectively, wherein the category corresponding to the highest confidence in the vector data is the diagnosis result, and input the diagnosis result into the fault handling database, and obtain the fault handling strategy through data mapping and matching; The hidden fault features of the SVG in the test set under normal operation are assigned values ​​and input into the trained BP neural network model to calculate the fault code value, decode the fault code value to obtain the hidden fault of the SVG under normal operation, and process it accordingly.

2. The neural network-based static VAR generator fault diagnosis method according to claim 1, characterized in that: The processing of data cleaning, data normalization, data standardization, and feature engineering on the operating parameter data is specifically as follows: Data cleaning, including missing value processing, duplicate value processing, outlier value processing, and error value processing; Data normalization, the formula is: Among them, x is the original sample data, x min is the minimum value in the sample data, x max is the maximum value in the sample data, and y is the normalized data; Data standardization, including time series data standardization and data format standardization; Feature engineering, including feature selection and feature combination construction.

3. The neural network-based static VAR generator fault diagnosis method according to claim 1, characterized in that: The feature extraction of the preprocessed operating parameter data specifically includes: Time domain feature extraction, including mean value calculation, effective value calculation, value and peak-to-peak value determination, variance and standard deviation calculation; Frequency domain feature extraction, including spectrum analysis, power spectrum density calculation, and harmonic analysis; Time-frequency domain feature extraction, specifically using wavelet transform analysis; Event-based feature extraction, including fault event features and operation event features; Correlation feature extraction, including device correlation features and time correlation features.

4. The neural network-based static VAR generator fault diagnosis method according to claim 1, characterized in that: The keyword extraction of the processing data is specifically as follows: Perform data cleaning on the text content of the processing data to remove extra spaces, punctuation marks and abnormal characters in the text; Perform grammatical analysis on the text content after data cleaning, including classification of Chinese nouns and technical English nouns; Tag the text with parts of speech, match the tagged words and sentences with the pre-established vocabulary, and obtain initial keywords; For the initial keywords, stop words are removed, synonyms are merged, and keyword weights are adjusted, where the weight order is maintenance object noun > maintenance action word > parameter quantifier.

5. The method for diagnosing static VAR generator faults based on neural network according to claim 1, characterized in that: The keyword extraction is performed on the extracted processing data keywords and fault result data, the extracted keywords are encoded, and the encoded processing data and the fault result data are mapped, specifically: Query the Unicode code corresponding to each Chinese character for the key words in the processing data, and convert the Unicode code into binary; Convert Unicode code to binary and use dictionary encoding technology to compress data; Create an array mapping between the encoded processing data keywords and the fault result data.

6. The neural network-based static VAR generator fault diagnosis method according to claim 1, characterized in that: The fault prediction model of the LSTM neural network is constructed, optimized based on the improved genetic algorithm, and the training set data is input into the LSTM model for model training, specifically: The constructed LSTM model includes input layer, hidden layer and output layer. The input layer has at least 10 neurons, the hidden layer has three layers, the output layer has at least 5 neurons, and Softmax is used as the activation function. A fully connected layer is added, and a Dropout layer is added before the fully connected layer to prevent overfitting. Construct an evaluation function for the improved genetic algorithm, use binary coding to initialize the population, and set various parameters; set the fitness function, calculate the fitness value of each individual, and retain the individual with the largest fitness; determine whether the algorithm termination conditions are met. If the convergence conditions are met, output the result, otherwise perform crossover, mutation, and selection operations; determine whether the result of the selection operation meets the algorithm termination conditions. If so, end the algorithm; The constructed LSTM neural network model is used as the main architecture of the fault prediction model, the number of input layers is set to 3, and the output layer outputs the state matrix; RMSE is determined as the error evaluation standard, and RMSE is also used as the evaluation function of the improved genetic algorithm; The improved genetic algorithm is used to dynamically calculate the fitness value of the step length of each LSTM training, and the step length with the optimal fitness value is determined as the LSTM neural network training step length; Taking the operating parameter data in the training set as input, taking the fault result data corresponding to the operating parameter data in the training set as output, taking the cross entropy as the loss function of the model, training is performed according to the loss function and the improved genetic algorithm, and the weights between the neurons are continuously adjusted according to the value of the loss function, so that the value of the loss function is gradually reduced; After each complete training data set traversal, the loss function value of the validation set is recorded and verified. When the loss function value meets the set threshold, the training ends and a trained LSTM neural network model is obtained.

7. The neural network-based static VAR generator fault diagnosis method according to claim 1, characterized in that: The cross entropy loss function is specifically: Among them, Y j is the true label of the jth sample, is the predicted probability of the jth sample.

8. The neural network-based static VAR generator fault diagnosis method according to claim 1, characterized in that: The fault handling strategy is obtained through data mapping and matching, and the fault handling strategy is obtained through mapping and matching based on a dictionary or a lookup table.

9. The method for diagnosing static VAR generator faults based on neural network according to claim 1, characterized in that: The operating parameter data include, the fault result data include but are not limited to overcurrent, overvoltage, overtemperature, IGBT module damage, communication failure, control failure, the operating parameter data include but are not limited to SVG output voltage value, SVG output current value, SVG reactive power output, SVG power factor, SVG operating temperature, SVG operating environment parameters.

10. A diagnostic system for implementing the neural network-based static VAR generator fault diagnosis method as described in any one of claims 1 to 9, characterized in that: include: A data acquisition unit is used to collect operating parameter data of the static VAR generator SVG in a fault state and a normal state when the static VAR generator SVG is in operation, record the corresponding processing data and fault result data in the fault state, collect the compensation current data of the SVG in a normal operation, and establish a fault diagnosis database; The data preprocessing unit is used to preprocess the data in the fault diagnosis database, including: Perform data cleaning, data normalization, data standardization, and feature engineering on the operating parameter data; perform feature extraction on the pre-processed operating parameter data, and extract key features that can reflect the operating status of SVG by establishing a feature extraction model; Extract keywords from processing data and fault result data, encode the extracted keywords, and map the encoded processing data with the fault result data to build a fault handling database; For the compensation current data of SVG under normal operation, the amplitude and phase characteristics of the compensation current fundamental wave and each harmonic when a single-tube open-circuit fault occurs in SVG are extracted through Fourier transform, and the hidden fault characteristics are identified; A prediction model building unit is used to build an LSTM model data set based on the preprocessed operation parameter data and operation result data in the fault diagnosis database, including a training set, a validation set, and a test set, which are divided according to the ratio of 70%:15%:15%, and then build a fault prediction model of the LSTM neural network, and optimize it based on the improved genetic algorithm, and input the training set data into the LSTM model for model training; at the same time, based on the hidden fault characteristics identified by the SVG under normal operation, assign and encode the hidden fault characteristics, and normalize the hidden fault characteristic values, establish a BP neural network model, and train the BP neural network model based on the normalized values ​​of the hidden fault characteristics and the fault coding; The explicit fault diagnosis unit inputs the running parameter data in the test set into the trained LSTM model, uses the Softmax classifier to identify whether a fault occurs and performs multi-classification of the fault category, and obtains the confidence vector data of the fault diagnosis respectively, wherein the category corresponding to the highest confidence in the vector data is the diagnosis result, and the diagnosis result is input into the fault handling database, and the fault handling strategy is obtained through data mapping and matching; The hidden fault diagnosis unit assigns hidden fault features of the SVG in the test set under normal operation and inputs them into the trained BP neural network model, calculates the fault code value, decodes the fault code value to obtain the hidden fault of the SVG under normal operation, and performs corresponding processing based on the hidden fault.

Images