Converter valve submodule monitoring method based on status word and computer system

Abstract

The invention relates to a converter valve submodule monitoring method based on status words and a computer system, and belongs to the technical field of power system fault prediction. According to the invention, the fault state of the sub-module after the current time period is predicted according to the electrical quantity time sequence data and the status word sequence of the sub-module in the current time period, the prediction result is output in the form of the status word, and the sub-module is monitored according to the predicted status word. The electrical quantity time sequence data and the status word sequence of the sub-module are selected for prediction, fault related information of the sub-module is effectively reflected, the potential relation between each fault state and the abnormal state of the sub-module is captured through prediction of the fault state of the sub-module, accurate prediction and monitoring of the sub-module are achieved, and the fault diagnosis accuracy of the sub-module is improved. The fault prediction and monitoring capability of the flexible direct current converter valve sub-module is effectively improved, and the problem that potential abnormity cannot be found and processed in time due to the fact that real-time online analysis cannot be carried out on the sub-module in the prior art is solved.

Description

Technical Field

[0001] This invention belongs to the field of power system fault prediction technology, specifically relating to a monitoring method and computer system for converter valve submodules based on status words. Background Technology

[0002] In flexible DC transmission systems, converter valves, as core equipment of the DC grid, undertake the crucial task of AC-DC conversion. The bridge arm of a flexible DC converter valve consists of hundreds of sub-modules connected in series, and the reliability of these sub-modules directly affects the overall stability of the system. To improve system reliability, it is necessary to monitor the health status of the flexible DC converter valve sub-modules in real time and provide early warnings before sub-module failures occur, reducing the risk of system outages. Existing flexible DC converter valve status monitoring functions mainly focus on the detection and handling of sub-module failures and leakage monitoring of valve towers, but lack effective monitoring methods for other critical components. Especially during sub-module operation, the lack of real-time online analysis tools makes it impossible to promptly detect and handle potential anomalies, posing a significant risk to the safe operation of the DC system. Summary of the Invention

[0003] The purpose of this invention is to provide a monitoring method and computer system for converter valve submodules based on status words, in order to solve the problem that the existing technology cannot perform real-time online analysis of submodules, resulting in the inability to detect and handle potential anomalies in a timely manner.

[0004] To address the aforementioned technical problems, this invention provides a monitoring method for a converter valve submodule based on status words, the method comprising:

[0005] Obtain the feature set of the converter valve submodule for the current time period. The feature set includes the electrical quantity time series data and status word sequence of the submodule. The status word contains status indication and fault indication. The status indication indicates whether the normal operation status of the submodule is activated, and the fault indication indicates whether the fault or abnormality of the submodule has occurred.

[0006] The acquired feature set is input into the trained submodule fault prediction model to obtain the submodule status word for the predicted time period after the current time period; the dataset used for training the submodule fault prediction model includes the historical feature set and the corresponding submodule status word.

[0007] The status of the converter valve submodule is monitored based on the submodule status word of the predicted time period.

[0008] Furthermore, the feature set also includes statistical features for each data segment, which is generated from electrical quantity time series data; the statistical features include at least one of the mean, standard deviation, and rate of change.

[0009] Furthermore, the feature set also includes trend features for each data segment, which is generated based on electrical quantity time series data; the calculation method for the trend features includes: comparing the average value of the previous data segment with the average value of the current data segment; if the difference between the average value of the current data segment and the average value of the previous data segment is greater than the upper limit of the trend feature set threshold, the trend feature of the current data segment is marked as "rising"; if the difference between the average value of the current data segment and the average value of the previous data segment is less than the lower limit of the trend feature set threshold, the trend feature of the current data segment is marked as "falling"; if the difference between the average value of the current data segment and the average value of the previous data segment is less than or equal to the upper limit of the trend feature set threshold and greater than or equal to the lower limit of the trend feature set threshold, the trend feature of the current data segment is marked as "stable".

[0010] Furthermore, the feature set also includes status word change features for each data segment, the data segments being generated based on the status word sequence; the calculation method for the status word change features includes: comparing the status word trigger percentage of the previous data segment with the status word trigger percentage of the current data segment, the status word trigger percentage being the percentage of status words with a value of 1 in the status word sequence; if the difference between the status word trigger percentage of the current data segment and the status word trigger percentage of the previous data segment is greater than the upper limit of the status word change feature threshold, the status word change feature of the current data segment is marked as "increasing"; if the difference between the status word trigger percentage of the current data segment and the status word trigger percentage of the previous data segment is less than the lower limit of the status word change feature threshold, the status word change feature of the current data segment is marked as "decreasing"; if the difference between the status word trigger percentage of the current data segment and the status word trigger percentage of the previous data segment is less than or equal to the upper limit of the status word change feature threshold and greater than or equal to the lower limit of the status word change feature threshold, the status word change feature of the current data segment is marked as "stable".

[0011] Furthermore, the methods for generating data segments from electrical quantity time series data include: the electrical quantity time series data is represented as {d1, d2, ..., d...} n}, where d i This is the electrical data at the i-th time point. The segment length is set to K and the step size to L. Data segment P is generated using the sliding window method. j ={d j ,d j+1 ,...,d j+K-1}; where j is the starting position of the sliding window, and the generated fragment P j It contains data at K consecutive time points.

[0012] Furthermore, the electrical quantity time series data of the submodule is obtained by preprocessing the original collected electrical quantity time series data of the submodule, and the preprocessing method includes removing outliers and / or processing missing values ​​using interpolation techniques based on time series models; the method for identifying outliers is the interquartile range method.

[0013] Furthermore, the status word of the submodule is obtained by preprocessing the original acquired submodule status word sequence. The preprocessing method includes automatically identifying and marking abnormal bits in the status word through a trained self-supervised learning model, removing the marked abnormal values, encoding the status word after removing the abnormal values, and encoding each bit of the status word as a separate column.

[0014] Furthermore, the sub-module fault prediction model includes a convolutional neural network and a gated recurrent unit. The convolutional neural network is used to extract spatial feature information, and the gated recurrent unit is used to capture the dependencies of time-series data based on the spatial feature information.

[0015] Furthermore, the feature set also includes environmental factors and / or historical operation records, wherein the environmental factors include temperature and / or humidity, and the historical operation records include at least one of the following: submodule component replacement time, number of historical submodule failures, and number of submodule maintenance.

[0016] Furthermore, the method for monitoring the status of the converter valve submodule based on the submodule status word of the predicted time period includes: determining the fault point based on the predicted submodule status word, calculating the fault duration of the fault point, and triggering an alarm if the fault duration is greater than a set time threshold.

[0017] Furthermore, after an alarm is triggered, the mean square error between the predicted submodule status word and the corresponding true status word is calculated. If the mean square error is greater than the set mean square error threshold, the true status value is added to the offline sample library. The submodule fault prediction model is retrained using the offline sample library, and the retrained submodule fault prediction model is used to predict the submodule status.

[0018] Furthermore, the electrical quantity time series data of the submodule includes: physical quantity data, operating status data, and control command data of the converter valve; the physical quantity data of the converter valve includes at least one of the following: submodule capacitor voltage, bridge arm current, capacitance value, and bridge arm current direction.

[0019] To address the aforementioned technical problems, the present invention also provides a computer system, including a processor, which executes a computer program to implement the steps of a status word-based monitoring method for a converter valve submodule.

[0020] The beneficial effects of the above technical solution are as follows: This invention is a pioneering invention. When a fault occurs in the converter valve submodule, it will be reflected in the electrical quantity time series data and status word sequence of the submodule. This invention predicts the fault state of the submodule after the current time period based on the electrical quantity time series data and status word sequence of the submodule in the current time period. The prediction result is output in the form of status words. The submodule is monitored based on the predicted status words. By selecting the electrical quantity time series data and status word sequence of the submodule for prediction, the fault-related information of the submodule is effectively reflected. By predicting the fault state of the submodule, the potential relationship between each fault state and abnormal state of the submodule is captured, realizing accurate prediction and monitoring of the submodule. This effectively improves the fault prediction and monitoring capability of the flexible DC converter valve submodule and solves the problem in the prior art that it is impossible to perform real-time online analysis of the submodule, resulting in the inability to detect and handle potential anomalies in a timely manner. Attached Figure Description

[0021] Figure 1 This is a flowchart of an embodiment of the converter valve submodule monitoring method based on status words of the present invention. Detailed Implementation

[0022] The focus of this invention is to provide a monitoring method for a converter valve submodule based on status words, the method comprising:

[0023] Obtain the feature set of the converter valve submodule for the current time period. The feature set includes the electrical quantity time series data and status word sequence of the submodule. The status word contains status indication and fault indication. The status indication indicates whether the normal operation status of the submodule is activated, and the fault indication indicates whether the fault or abnormality of the submodule has occurred.

[0024] The acquired feature set is input into the trained submodule fault prediction model to obtain the submodule status word for the predicted time period after the current time period; the dataset used for training the submodule fault prediction model includes the historical feature set and the corresponding submodule status word.

[0025] The status of the converter valve submodule is monitored based on the submodule status word of the predicted time period.

[0026] When a fault occurs in the converter valve submodule, it will be reflected in the electrical quantity time series data and status word sequence of the submodule. This invention predicts the fault state of the submodule after the current time period based on the electrical quantity time series data and status word sequence of the submodule in the current time period. The prediction result is output in the form of status words. The submodule is monitored based on the predicted status words. By selecting the electrical quantity time series data and status word sequence of the submodule for prediction, the fault-related information of the submodule is effectively reflected. By predicting the fault state of the submodule, the potential relationship between each fault state and abnormal state of the submodule is captured, realizing accurate prediction and monitoring of the submodule. This effectively improves the fault prediction and monitoring capability of the flexible DC converter valve submodule and solves the problem in the prior art that it is impossible to perform real-time online analysis of the submodule, resulting in the inability to detect and handle potential anomalies in a timely manner.

[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0028] Example of a monitoring method for converter valve submodules based on status words:

[0029] This embodiment presents a status word-based monitoring method for converter valve submodules. By constructing an effective prediction model, this method enables reliable online monitoring of flexible DC converter valve submodules by predicting changes in status words. Certain abnormal states of the submodule are correlated with specific fault states. By predicting status words, these correlations and trends can be captured more accurately. By predicting changes in status words, it is possible to predict in advance whether the submodule will experience abnormalities or faults in the future. For example, if there is a high correlation between power supply faults and overvoltage faults, the prediction model can improve the prediction probability of overvoltage faults when a power supply fault is detected. Through this correlation analysis, faults can be prevented and responded to more effectively.

[0030] The specific implementation steps of the status word-based converter valve submodule monitoring method are as follows:

[0031] Step 1: Data collection.

[0032] The system collects electrical quantity data (i.e., electrical quantity time series data) from submodules in real time. This data includes physical quantity data, operating status data, and control command data for the converter valves. The physical quantity data for the converter valves includes at least one of the following: submodule capacitor voltage, bridge arm current, capacitance value, and bridge arm current direction. Simultaneously, the system collects status word data (i.e., status word sequences) from the submodules and expands the data acquisition dimensions by combining environmental data (temperature, humidity, etc.) and operation records to construct a more comprehensive dataset.

[0033] In this embodiment, the status word data is designed to consist of three 16-bit binary numbers, defined as status word 1, status word 2, and status word 3, respectively. Each bit of the status word represents a specific status or fault indication of the submodule. The status (0 or 1) of each status bit or fault indication bit can provide real-time information about the current operating status of the submodule. The meanings of the status words are as follows:

[0034] 1) Status indication: Some bits in the status word may indicate the normal operating status of the submodule, such as unlock mode, natural charging mode, etc. These bits being 0 or 1 can indicate whether these states are active.

[0035] 2) Fault indication: Other bits in the status word are used to indicate potential faults or abnormalities, such as communication failures, drive failures, voltage problems, etc. When these bits are 1, it indicates that the corresponding fault or abnormality has occurred.

[0036] Step 2: Data preprocessing.

[0037] The collected data undergoes cleaning, including outlier removal, missing value handling, and data normalization. For electrical quantity data, the interquartile range (IQR) method is first used to identify and remove outliers. Then, advanced interpolation methods, such as time series model-based interpolation, are used to handle missing data, further improving data reliability. Subsequently, adaptive normalization is applied to normalize the data, adapting it to different operating environments. For status word data, multi-level validity checks are first used to ensure data integrity and format correctness. Then, a self-supervised learning model is used to automatically mark and identify outlier bits in the status word, and the marked outliers are removed, ensuring intelligent processing. Finally, dynamic encoding technology is used to encode the status word, treating each bit as a separate column for one-hot encoding, converting it into a series of binary columns. Each status bit or fault indication bit corresponds to one column, with bits indicating a status of 1 marked as 1 in the corresponding column and bits indicating a status of 0 marked as 0 in the corresponding column. The above data preprocessing steps can effectively improve data quality and provide a reliable data foundation for subsequent submodule fault diagnosis, performance evaluation and optimization control.

[0038] When electrical quantity data is relatively good, one of the two processing methods can be chosen: removing outliers and handling missing values.

[0039] Step 3: Feature construction.

[0040] The purpose of feature construction is to design the most meaningful attributes for submodule fault diagnosis and performance evaluation from the raw data. First, the electrical quantity time series data of each submodule is divided into several continuous data segments of length K. The specific steps are as follows:

[0041] Let the time series data D be represented as {d1, d2, ..., d...} n}, where d i This is the observation value at the i-th time point. Set the segment length to K and the step size to L, and generate data segment P using the sliding window method. j ={d j d j+1 , ..., d j+K-1}. Where j is the starting position of the sliding window, and the generated fragment P j It contains data at K consecutive time points, representing the running status of the submodule within K time steps.

[0042] Next, for each data segment P j Extract statistical features, which include at least one of the following: mean μ, standard deviation σ, and rate of change Δ. The calculation method is as follows:

[0043]

[0044] In the formula, μ j Let d be the average value of the j-th data segment, K be the time step, and d be the average value of the j-th data segment. j+i-1 It is the observation value at the (j+i-1)th time point.

[0045]

[0046] In the formula, σ j Let be the standard deviation of the j-th data segment.

[0047]

[0048] In the formula, Δ j Let d be the rate of change of the j-th data segment. j It is the observation value at the j-th time point.

[0049] The feature set also includes trend features for each data segment, which is generated based on electrical quantity time series data. The calculation method for trend features includes: comparing the average value of the previous data segment with the average value of the current data segment; if the difference between the average value of the current data segment and the average value of the previous data segment is greater than the upper limit of the trend feature threshold, the trend feature of the current data segment is marked as "rising"; if the difference between the average value of the current data segment and the average value of the previous data segment is less than the lower limit of the trend feature threshold, the trend feature of the current data segment is marked as "falling"; if the difference between the average value of the current data segment and the average value of the previous data segment is less than or equal to the upper limit of the trend feature threshold and greater than or equal to the lower limit of the trend feature threshold, the trend feature of the current data segment is marked as "stable".

[0050] The feature set also includes status word change features for each data segment, which is generated based on the status word sequence. The calculation method for status word change features includes: comparing the status word trigger percentage of the previous data segment with the status word trigger percentage of the current data segment. The status word trigger percentage is the percentage of status words with a value of 1 in the status word sequence. If the difference between the status word trigger percentage of the current data segment and the status word trigger percentage of the previous data segment is greater than the upper limit of the status word change feature threshold, the status word change feature of the current data segment is marked as "rising". If the difference between the status word trigger percentage of the current data segment and the status word trigger percentage of the previous data segment is less than the lower limit of the status word change feature threshold, the status word change feature of the current data segment is marked as "falling". If the difference between the status word trigger percentage of the current data segment and the status word trigger percentage of the previous data segment is less than or equal to the upper limit of the status word change feature threshold and greater than or equal to the lower limit of the status word change feature threshold, the status word change feature of the current data segment is marked as "stable".

[0051] T j =trend(P j )

[0052] Among them, trend(P j ) is for data fragment P j The results after applying the trend detection algorithm are labeled as "rising," "falling," or stable. Environmental factors and / or historical operation records are integrated as additional features. Environmental factors include temperature and / or humidity, and historical operation records include at least one of the following: submodule component replacement time, number of historical submodule failures, and number of submodule maintenance visits. The resulting feature vector...

[0053] F combined =[Statistical characteristics, trend characteristics, state word change characteristics, environmental factors, historical records]. Finally, the feature vector F... combined Compared with the original data F original The features are fused together to generate the final feature set F. final :

[0054] F final =[F original F combined ]

[0055] The feature set may include at least the original collected electrical quantity time series data and submodule status word sequence of the submodule.

[0056] The data is processed by predicting M steps every K steps. The state word data from the M consecutive steps following the time window are used as labels. Each bit of the state word is treated as an independent label with a value of 0 or 1, indicating whether the state is active. This allows the model to learn how to predict the state of each state bit in the next M steps. The final feature set is divided into training, validation, and test sets in a 7:2:1 ratio to ensure sufficient data for training and independent datasets for model validation and testing.

[0057] Step 4: Model training.

[0058] The submodule fault prediction model comprises a convolutional neural network and gated recurrent units. It is trained using the training set partitioned during feature construction. The specific construction method of the submodule fault prediction model is as follows:

[0059] First, the model uses a Convolutional Neural Network (CNN) to process the input feature set. The CNN extracts features through multiple convolutional layers, captures local spatial patterns of the features, and effectively reduces feature dimensionality while retaining key information through pooling layers. Next, the extracted spatial features are input into a Gated Recurrent Unit (GRU) to capture the dependencies in the time-series data. The GRU is a recurrent neural network (RNN) model for processing time-series data, which introduces a gating mechanism to effectively solve the gradient vanishing problem in traditional RNNs with long-series data. A self-attention mechanism can also be introduced into the GRU. This mechanism can flexibly allocate weights in long sequences, measure the importance of each element in the input sequence, and identify the features that have the greatest impact on the submodule state prediction, allowing the model to more accurately capture complex temporal dependencies. Furthermore, adaptive optimization techniques, such as the Adam optimizer, are used during the training of the submodule fault prediction model. By dynamically adjusting the learning rate and gradient, the convergence speed and stability of the model during training are ensured.

[0060] Step 5: Data Prediction.

[0061] Obtain the feature set of the current time period of the converter valve submodule. The feature set includes the electrical quantity time series data and status word sequence of the submodule. The status word contains status indication and fault indication. The status indication indicates whether the normal operation status of the submodule is activated, and the fault indication indicates whether the fault or abnormality of the submodule has occurred. Input the obtained feature set into the trained submodule fault prediction model to obtain the submodule status word of the predicted time period after the current time period.

[0062] The status of the converter valve submodule is monitored based on the submodule status word for the predicted time period; the dataset used for training the submodule fault prediction model includes historical feature sets and corresponding submodule status words. Specific steps are as follows: Figure 1 As shown.

[0063] First, the method receives electrical quantity data and status word data from the submodule communication module. The electrical quantity data includes submodule capacitor voltage, bridge arm current, capacitance value, system operating status, submodule control commands, bridge arm current direction, and submodule status. Location information is used to mark the data source and the submodule's position within the converter valve.

[0064] Next, the data undergoes preprocessing, including outlier removal, data normalization, and data fragment generation. During data fragment generation, a sliding window method is used to process historical electrical data over a period of time. For example, the sliding window has a length of Lwin and slides gradually from the beginning of the original data sequence with a step size Lstep, obtaining a new data fragment with each slide until the end of the historical data. If the number of remaining samples is less than the sliding window step size, the window stops sliding, waiting for new data to arrive.

[0065] Trend and statistical features are extracted from each generated data segment, and these features, together with the original collected electrical quantity time series data and status word sequence data, constitute a feature set. The feature set also includes environmental factors and historical operation records.

[0066] The feature set is input into the previously trained submodule fault prediction model. The model extracts and analyzes patterns and trends related to the submodule's fault state from the feature set, and predicts the submodule's operational status over a future period. The prediction result is output as a submodule status word. Based on the alarm and fault bits in the output submodule status word, it is determined whether the submodule has experienced an anomaly or fault within the future period. If an anomaly or fault is detected, the fault point of the submodule is located based on the alarm or fault bit, and the fault duration is calculated. It is then determined whether the fault duration exceeds a set time threshold (2ms in this embodiment). If it exceeds the set time threshold, an alarm is triggered, notifying maintenance personnel to take preventative measures to avoid potential faults. Furthermore, the predicted submodule status word is temporarily stored, awaiting the arrival of the next segment of real status data. The mean squared error (MSE) between the predicted submodule status word and the actual status word is calculated. When the MSE exceeds a set MSE threshold, this segment of data is added to the offline sample library, awaiting training of the prediction model. The trained prediction model updates and replaces the original prediction model to provide more accurate predictions.

[0067] The formula for calculating the mean square error between the predicted submodule status word and the corresponding actual status word value is as follows:

[0068]

[0069] In the formula, MSE is the mean square error, n is the total number of bits in the status word, and a i To predict the i-th bit of the state word, b i It is the i-th bit of the real state word.

[0070] Computer system example:

[0071] This invention also provides a computer system, including a processor, which executes a computer program to implement the steps of a status word-based converter valve submodule monitoring method. The specific processes, principles, and effects have been described in detail in the embodiments of the status word-based converter valve submodule monitoring method, and will not be repeated here.

[0072] This invention predicts changes in status words to anticipate whether a submodule will experience anomalies or faults in the future. In the feature construction phase, this method innovatively combines statistical features of electrical quantity data with trend features of status word data. By generating data fragments, extracting statistical features, and analyzing trend features, the raw data is transformed into a more informative feature set, providing a richer data foundation for subsequent model training. This not only expands the feature set but also enhances the sensitivity of fault prediction, enabling the model to more fully utilize the information in the data and improve prediction accuracy. This invention also employs a deep learning hybrid network model combining convolutional neural networks (CNNs) and gated recurrent units (GRUs) as the submodule fault prediction model. A multi-label output layer further improves the predictive ability for complex fault states, exhibiting extremely high accuracy and robustness, especially when handling multiple fault states of submodules. Through this innovative fault prediction model and method design, submodule faults can be represented by status words, and the accuracy and stability of predictions are further improved by combining the prediction results of multiple models through ensemble learning. Utilizing historical operating data of the submodule avoids interference from human experience, making prediction and monitoring more reliable and efficient. In summary, this invention has advantages such as high real-time performance, accuracy and stability, and wide applicability. It can effectively improve the fault prediction and monitoring capabilities of the flexible DC converter valve submodule, reduce the risk of system failure and shutdown, and has high innovation and practical value.

[0073] Specific implementation methods have been given above, but the present invention is not limited to the described implementation methods. The basic idea of ​​the present invention lies in the above basic scheme. For those skilled in the art, designing various modified models, formulas, and parameters based on the teachings of the present invention does not require creative effort. Changes, modifications, substitutions, and variations made to the implementation methods without departing from the principles and spirit of the present invention still fall within the protection scope of the present invention.

Claims

1. A monitoring method for a converter valve submodule based on status words, characterized in that, The method includes: Obtain the feature set of the current time period of the converter valve submodule. The feature set includes the electrical quantity time series data and status word sequence of the submodule. The status word contains status indication and fault indication. The status indication indicates whether the normal operation status of the submodule is activated, and the fault indication indicates whether the fault or abnormality of the submodule has occurred. The acquired feature set is input into the trained submodule fault prediction model to obtain the submodule status word for the predicted time period after the current time period; the dataset used for training the submodule fault prediction model includes the historical feature set and the corresponding submodule status word. The status of the converter valve submodule is monitored based on the submodule status word of the predicted time period.

2. The monitoring method for the converter valve submodule based on status words according to claim 1, characterized in that, The feature set also includes statistical features for each data segment, which is generated from electrical quantity time series data; the statistical features include at least one of mean, standard deviation and rate of change.

3. The monitoring method for the converter valve submodule based on status words according to claim 1, characterized in that, The feature set also includes trend features for each data segment, which is generated based on electrical quantity time series data. The calculation method for the trend features includes: comparing the average value of the previous data segment with the average value of the current data segment; if the difference between the average value of the current data segment and the average value of the previous data segment is greater than the upper limit of the trend feature threshold, the trend feature of the current data segment is marked as "rising"; if the difference between the average value of the current data segment and the average value of the previous data segment is less than the lower limit of the trend feature threshold, the trend feature of the current data segment is marked as "falling"; if the difference between the average value of the current data segment and the average value of the previous data segment is less than or equal to the upper limit of the trend feature threshold and greater than or equal to the lower limit of the trend feature threshold, the trend feature of the current data segment is marked as "stable".

4. The monitoring method for the converter valve submodule based on status words according to claim 1, characterized in that, The feature set also includes status word change features for each data segment, the data segments being generated based on a status word sequence; the calculation method for the status word change features includes: comparing the status word trigger percentage of the previous data segment with the status word trigger percentage of the current data segment, the status word trigger percentage being the percentage of status words with a value of 1 in the status word sequence; if the difference between the status word trigger percentage of the current data segment and the status word trigger percentage of the previous data segment is greater than the upper limit of the status word change feature threshold, the status word change feature of the current data segment is marked as "increasing"; if the difference between the status word trigger percentage of the current data segment and the status word trigger percentage of the previous data segment is less than the lower limit of the status word change feature threshold, the status word change feature of the current data segment is marked as "decreasing"; if the difference between the status word trigger percentage of the current data segment and the status word trigger percentage of the previous data segment is less than or equal to the upper limit of the status word change feature threshold and greater than or equal to the lower limit of the status word change feature threshold, the status word change feature of the current data segment is marked as "stable".

5. The monitoring method for the converter valve submodule based on status words according to claim 2 or 3, characterized in that, Methods for generating data segments from electrical quantity time series data include: electrical quantity time series data is represented as {d1, d2, ..., d...} n }, where d i This is the electrical data at the i-th time point. The segment length is set to K and the step size to L. Data segment P is generated using the sliding window method. j ={d j ,d j+1 ,...,d j+K-1 }; where j is the starting position of the sliding window, and the generated fragment P j It contains data at K consecutive time points.

6. The monitoring method for the converter valve submodule based on status words according to claim 1, characterized in that, The electrical quantity time series data of the submodule is obtained by preprocessing the original electrical quantity time series data of the submodule. The preprocessing method includes removing outliers and / or processing missing values ​​using interpolation techniques based on time series models. The method for identifying outliers is the interquartile range method.

7. The monitoring method for the converter valve submodule based on status words according to claim 1, characterized in that, The status word of the submodule is obtained by preprocessing the original submodule status word sequence. The preprocessing method includes automatically identifying and marking abnormal bits in the status word through a self-supervised learning model after training, removing the marked abnormal values, encoding the status word after removing the abnormal values, and encoding each bit of the status word as a separate column.

8. The monitoring method for the converter valve submodule based on status words according to claim 1, characterized in that, The sub-module fault prediction model includes a convolutional neural network and a gated recurrent unit. The convolutional neural network is used to extract spatial feature information, and the gated recurrent unit is used to capture the dependencies of time-series data based on the spatial feature information.

9. The monitoring method for the converter valve submodule based on status words according to claim 1, characterized in that, The feature set also includes environmental factors and / or historical operation records, wherein the environmental factors include temperature and / or humidity, and the historical operation records include at least one of the following: submodule component replacement time, number of historical submodule failures, and number of submodule maintenance.

10. The monitoring method for the converter valve submodule based on status words according to claim 1, characterized in that, The method for monitoring the status of the converter valve submodule based on the submodule status word of the predicted time period includes: determining the fault point based on the predicted submodule status word, calculating the fault duration of the fault point, and triggering an alarm if the fault duration is greater than a set time threshold.

11. The monitoring method for the converter valve submodule based on status words according to claim 10, characterized in that, After an alarm is triggered, the mean square error between the predicted submodule status word and the corresponding true status word is calculated. If the mean square error is greater than the set mean square error threshold, the true status value is added to the offline sample library. The submodule fault prediction model is retrained using the offline sample library, and the retrained submodule fault prediction model is used to predict the submodule status.

12. The monitoring method for the converter valve submodule based on status words according to claim 1, characterized in that, The electrical quantity time series data of the submodule includes: physical quantity data, operating status data and control command data of the converter valve; the physical quantity data of the converter valve includes at least one of the following: submodule capacitor voltage, bridge arm current, capacitance value and bridge arm current direction.

13. A computer system, comprising a processor, characterized in that, The processor is used to execute a computer program to implement the steps of the status word-based converter valve submodule monitoring method as described in any one of claims 1 to 12.

Images