A method for determining the operating status of a converter valve

Abstract

This invention relates to a method for determining the working status of a converter valve, comprising the following steps: S1: Data acquisition; the inspection robot measures and acquires various data from the valve module of the converter valve; S2: Obtaining raw data; the inspection robot transmits the acquired detection data to the control center to form raw data; S3: Data integration; the control center integrates the data; the data reconstruction system includes at least two data input modules and one integration module; the control center selects from the raw data, transmitting each data item to a corresponding data input module for reconstruction, and judging the abnormality of the data item based on the reconstruction error and intermediate variable error; the reconstruction error corresponding to each data input module is finally integrated in the integration module, and an overall risk score is obtained based on the various reconstruction errors. Finally, the risk score is obtained by weighting the overall risk score with the risk score corresponding to each data input module, thus completing the analysis and evaluation of the working status of the converter valve module.

Description

Technical Field

[0001] This invention relates to a method for determining the working status of a converter valve, belonging to the technical field of power grid transmission and transformation equipment status monitoring systems. Background Technology

[0002] Flexible DC transmission technology (VSC-HVDC) is a new generation of DC transmission system based on voltage source converters. The core equipment of VSC-HVDC is the VSC converter valve. With the gradual development and promotion of VSC-HVDC technology, its voltage level and transmission capacity are constantly improving, which places higher demands on the reliability of the VSC converter valve.

[0003] Currently, there are two main methods for determining the working status of valve modules in flexible DC converter stations:

[0004] First, online thermal imaging temperature measurement systems are installed in all flexible DC converter valve halls. These systems automatically measure the surface temperature of the valve modules, allowing inspection personnel to determine the operational status of the flexible DC converter valves based on the surface temperature of the valve modules. Figure 1 As shown.

[0005] Second, UHF partial discharge detection uses a UHF sensor to measure the UHF (300~3000MHz) electromagnetic wave signal excited by partial discharge, thereby achieving partial discharge measurement.

[0006] The Chinese invention patent with patent application number CN202010864496.2, entitled "Method for Judging the Working Status of Valve Module by Surface Temperature and UHF Data", includes the following steps: setting a timed inspection workflow for an inspection robot, periodically measuring the surface temperature and UHF partial discharge data of valve modules within a set range, obtaining the valve module's average surface temperature Tave, maximum temperature Tmax, temperature difference Tdet, and the valve module's measured UHF partial discharge data P and historical average data Pave; judging whether the valve module is malfunctioning: if Tdet≥5℃ or (((P Pave) / Pave)*100%).20% or k≥threshold K, then the valve module is judged to be malfunctioning; k=k1×(Tmax Tave) / Tave+k2×(1 (5 Tdet) / 5)+k3×(PPave) / Pave. This method uses an inspection robot to simultaneously collect surface temperature and UHF partial discharge data of valve modules within a set range, and combines the two data to determine the working status of the converter valve module. It overcomes the limitations of using a single judgment method, improves the effectiveness and accuracy of the judgment, and has strong practicality and public application prospects.

[0007] However, the aforementioned invention patent mainly uses surface temperature and ultra-high frequency data to determine the working status of the valve module. In operation, the converter valve also generates electric fields, magnetic fields, ultrasonic partial discharge, ultra-high frequency partial discharge, audio, and gas. By comprehensively detecting and judging these data, the operating conditions of the converter valve can be more accurately and comprehensively grasped, and equipment abnormalities can be detected as early as possible. Summary of the Invention

[0008] To overcome the above problems, this invention provides a method for determining the operating status of a converter valve. This method uses data collected from a linear infrared temperature measurement system, as well as electric field strength, magnetic field strength, ultrasonic partial discharge data, ultra-high frequency partial discharge data, audio data, and gas concentration data to determine the operating status of the converter valve. By comprehensively detecting and judging these data, the operating conditions of the converter valve can be more accurately and comprehensively understood, and equipment abnormalities can be detected early.

[0009] The technical solution of the present invention is as follows:

[0010] A method for determining the operating status of a converter valve includes the following steps:

[0011] S1: Data Acquisition;

[0012] The inspection robot measures and collects various data of the valve module of the converter valve within a set range at the predetermined inspection points;

[0013] S2: Obtain raw data;

[0014] The inspection robot transmits the collected detection data to the control center, forming raw data, which is then recorded in the control center.

[0015] S3: Data integration;

[0016] The control center transmits the raw data to the data reconstruction system for integration;

[0017] The data reconstruction system includes at least two data input modules and one integration module. The control center selects data from the original data, and each data item is transmitted to a corresponding data input module for reconstruction. The system determines the abnormality of the data item based on the reconstruction error and intermediate variable error. The reconstruction error corresponding to each data input module is finally integrated in the integration module. An overall risk score is obtained based on the reconstruction errors. Finally, the overall risk score is obtained by weighting the overall risk score with the risk score corresponding to each data input module, thus completing the analysis and evaluation of the working status of the converter valve module.

[0018] S4: Optimize training;

[0019] Based on the different data collected by each of the aforementioned data input modules, optimization training is performed separately;

[0020] S5: Risk Prediction;

[0021] After the optimization training is completed, each of the data input modules collects a new set of data. The control center then judges the new set of data and predicts its risk score.

[0022] Furthermore, the data input module includes a PCA preprocessing unit and a generative adversarial network based on a variational autoencoder architecture; the generative adversarial network includes a variational autoencoder network, a discriminator, and a verification encoder; the variational autoencoder network includes an input encoder and an input decoder; the PCA preprocessing unit processes the data from the control center as input data; the input encoder encodes the input data, mapping high-dimensional data to low-dimensional data; the input decoder restores the encoded low-dimensional data back to high-dimensional data through an inverse transformation; the verification encoder re-encodes the high-dimensional data after the inverse transformation of the input decoder into low-dimensional data; the discriminator is used to determine the consistency between the low-dimensional data re-encoded by the verification encoder and the original low-dimensional data input by the input encoder.

[0023] Furthermore, the integration module includes an autoencoder network; the autoencoder network includes an integrated encoder and an integrated decoder; the integrated encoder is used to merge and encode the set of data generated by the verification encoder in the multiple data input modules, and then input it as input data into the integrated decoder.

[0024] Furthermore, in S3: data integration, a low-dimensional vector is defined after processing by the input encoder. As shown in equation (1):

[0025] (1)

[0026] in, The input data is the raw data processed by the PCA preprocessing unit; the output of the input encoder is the mean of a Gaussian distribution. and variance ;

[0027] Define the reconstructed data after inverse transformation by the input decoder. As shown in equation (2):

[0028] (2)

[0029] Define the reconstruction loss after inverse transformation by the input decoder. As shown in equation (3):

[0030] (3)

[0031] Define the recoding loss after recoding by the verified encoder. As shown in equation (4):

[0032] (4)

[0033] in, A low-dimensional vector re-encoded by a verified encoder;

[0034] Define the encoding distribution as a standard normal distribution. Loss results during divergence As shown in equation (5):

[0035] (5)

[0036] in, Let be the dimension of the Gaussian distribution, and be the hyperparameter. , These are the Gaussian distributions predicted by the variational autoencoder. and The The mean and standard deviation of the dimension;

[0037] Define the original input data With reconstructing data Adversarial loss after input discriminator As shown in equation (6):

[0038] (6)

[0039] in, , All are the outputs of the discriminator (12); This represents the total number of data items for the valve module of the converter valve.

[0040] Define the final loss of the data input module. As shown in equation (7):

[0041] (7)

[0042] in, , , , These are hyperparameters, representing the weights of each loss.

[0043] Furthermore, in S3: data integration, a low-dimensional vector is defined after processing by the ensemble encoder. As shown in equation (8):

[0044] (8)

[0045] in, A collection of data generated by multiple verification encoders;

[0046] Define the reconstruction loss after inverse transformation by the integrated decoder. As shown in equation (9):

[0047] (9)

[0048] in, It is a high-dimensional vector that has been remapped by the integrated decoder.

[0049] Furthermore, the inspection robot can collect data including temperature data, electric field strength data, magnetic field strength data, ultrasonic partial discharge data, ultra-high frequency partial discharge data, audio data, and gas concentration data.

[0050] The present invention has the following beneficial effects:

[0051] 1. This method collects temperature data, electric field strength data, magnetic field strength data, ultrasonic partial discharge data, ultra-high frequency partial discharge data, audio data, and gas concentration data to determine the working status data of the converter valve. Through comprehensive detection and judgment of these data, the operating conditions of the converter valve can be more accurately and comprehensively grasped, and equipment abnormalities can be detected as early as possible. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of an existing infrared thermal imaging analysis system.

[0053] Figure 2 This is a flowchart of the discrimination method of the present invention.

[0054] Figure 3 This is a schematic diagram of the data reconstruction system of the present invention.

[0055] Figure 4 This is a structural diagram of the data input module of the present invention.

[0056] Figure 5 This is a structural diagram of the integrated module of the present invention.

[0057] The reference numerals in the figure are as follows:

[0058] 1. Data input module; 11. PCA preprocessing unit; 12. Discriminator; 13. Input encoder; 14. Input decoder; 15. Verification encoder; 2. Integrated module; 21. Autoencoder; 22. Integrated encoder; 23. Integrated decoder. Detailed Implementation

[0059] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0060] This assessment method, in conjunction with a dedicated valve hall inspection robot system, collects relevant information and transmits the monitored data to the control center via network communication, completing the collection, transmission, and recording of the working status of the flexible vertical valve hall equipment. Based on this data, the inspection robot system, combined with PCA (principal components analysis), generative adversarial networks, and autoencoder technology, proposes a general and scalable model architecture for risk prediction and assessment of the flexible vertical valve hall and the data collected by the inspection robot. This model receives data in the following categories: temperature data, electric field strength data, magnetic field strength data, ultrasonic partial discharge data, ultra-high frequency partial discharge data, audio data, and gas concentration data, etc. It predicts risk scores for different categories of data separately, and finally integrates the data to obtain an overall risk score. A weighted average is then used to obtain the final expected result, completing the analysis and evaluation of the valve component's working status.

[0061] See Figure 1-5 A method for determining the operating status of a converter valve includes the following steps:

[0062] S1: Data Acquisition;

[0063] The inspection robot measures and collects various data of the valve module of the converter valve within a set range at the predetermined inspection points;

[0064] S2: Obtain raw data;

[0065] The inspection robot transmits the collected detection data to the control center, forming raw data, which is then recorded in the control center.

[0066] S3: Data integration;

[0067] The control center transmits the raw data to the data reconstruction system for integration;

[0068] The data reconstruction system includes at least two data input modules 1 and one integration module 2. The control center selects data from the original data, and each data item is transmitted to a corresponding data input module 1 for reconstruction. The abnormality of the data item is judged based on the reconstruction error and intermediate variable error. The reconstruction data corresponding to each data input module 1 is finally integrated in the integration module 2. The overall risk score is obtained based on the reconstruction errors. Finally, the risk score is obtained by weighting the overall risk score with the risk score corresponding to each data input module 1, thus completing the analysis and evaluation of the working status of the valve module of the converter valve.

[0069] The framework of the data reconstruction system is as follows Figure 2 As shown, multiple data input modules 1 model different types of input data (such as temperature data, electric field strength data, magnetic field strength data, ultrasonic partial discharge data, UHF partial discharge data, audio data, and gas concentration data). Each data input module 1 includes a generative adversarial network based on a variational autoencoder architecture, capable of reconstructing the input data and determining anomalies based on reconstruction errors and intermediate variable errors. An integration module 2 integrates the reconstructed data from the above modules and inputs it into an autoencoder for final reconstruction. An overall risk score is obtained based on the reconstruction error, and finally, the overall risk score is obtained by weighting the risk scores from each module.

[0070] The data input module 1 includes a PCA preprocessing unit 11 and a generative adversarial network based on a variational autoencoder architecture. The generative adversarial network includes a variational autoencoder network, a discriminator 12, and a verification encoder 15. The variational autoencoder network includes an input encoder 13 and an input decoder 14. The PCA preprocessing unit 11 processes the data from the control center as input data. The input encoder 13 encodes the input data, mapping high-dimensional data to low-dimensional data. The input decoder 14 restores the encoded low-dimensional data back to high-dimensional data through an inverse transformation. The verification encoder 15 re-encodes the high-dimensional data after the inverse transformation of the input decoder 14 into low-dimensional data. The discriminator 12 is used to determine the consistency between the low-dimensional data re-encoded by the verification encoder 15 and the original low-dimensional data input by the input encoder 13.

[0071] The integrated module 2 includes an autoencoder network 21; the autoencoder network 21 includes an integrated encoder 22 and an integrated decoder 23; the integrated encoder 22 is used to process the data (i.e., reconstructed data) generated by the input decoders 14 in the multiple data input modules 1. After being merged and encoded, the data is input into the integrated decoder 23.

[0072] Specifically, the data input module 1 employs a Variational Autoencoder-based Generative Adversarial Network (VAE-GAN). VAE-GAN combines a variational autoencoder with a generative adversarial network. The VAE (Variational Auto-Encoder) consists of an input encoder 13 and an input decoder 14. The input encoder 13 encodes the data input from the PCA preprocessing unit 11, mapping high-dimensional data to low-dimensional data. The input decoder 14 maps the encoded low-dimensional vector back to a high-dimensional vector, ensuring the generated data is as similar as possible to the input data. Unlike ordinary autoencoders, VAE incorporates a standard normal distribution into the encoding process. The divergence loss makes the encoded result more continuous. The GAN (Generative Adversarial Network) part includes a generator and a discriminator 12. The generator is the input decoder 14 part in the aforementioned VAE, which maps low-dimensional vectors to high-dimensional vectors, while the discriminator 12 is used to judge whether the generated data is true or false and give a "score" for the judgment. In order to enable the encoder to perform mapping better, an additional verification encoder 15 is added to re-encode the generated data, so that the encoded result is as close as possible to the initial encoded result.

[0073] The integrated module 2 is composed of a common autoencoder, which is simpler than the variational autoencoder.

[0074] Specifically, the process of data integration in S3 is as follows:

[0075] The control center selects from the raw data, and each data item is transmitted to a corresponding data input module 1 for reconstruction.

[0076] Define the low-dimensional vector obtained after processing by input encoder 13. As shown in equation (1):

[0077] (1)

[0078] in, The input data is the raw data processed by the PCA preprocessing unit 11; the output of the input encoder 13 is the mean of a Gaussian distribution. and variance ;

[0079] Input encoder 13 receives raw input data processed by PCA preprocessing unit 11. As input, the low-dimensional vector is obtained after passing through input encoder 13. The output of the input encoder 13 is a Gaussian distribution with a mean value. and variance After resampling, a low-dimensional vector is obtained. .

[0080] Define the reconstructed data after inverse transformation by input decoder 14 As shown in equation (2):

[0081] (2)

[0082] low-dimensional vectors After the inverse transformation by the input decoder 14, the reconstructed data can be obtained. .

[0083] Define the reconstruction loss after inverse transformation of input decoder 14. As shown in equation (3):

[0084] (3)

[0085] And obviously, reconstructing the data Compared with the original input data The closer the better, i.e., reconstruction loss The smaller the better.

[0086] Define the recoding loss after recoding by the verified encoder 15. As shown in equation (4):

[0087] (4)

[0088] in, A low-dimensional vector re-encoded by the verified encoder 15;

[0089] To make the encoding results more accurate, an additional verification encoder 15 was added to the reconstructed data. Re-encode to obtain a re-encoded low-dimensional vector. And obviously, the re-encoded low-dimensional vector With low-dimensional vectors The closer the better, i.e., recoding loss (i.e., the loss of intermediate variables) should be as small as possible.

[0090] Define the encoding distribution as a standard normal distribution. Loss results during divergence As shown in equation (5):

[0091] (5)

[0092] in, Let be the dimension of the Gaussian distribution, and be the hyperparameter. , These are the Gaussian distributions predicted by the variational autoencoder. and The The mean and standard deviation of the dimension;

[0093] Because of the requirements of VAE-based encoders, the encoded vectors Approximately a standard normal distribution Therefore, the encoding distribution here is based on the standard normal distribution. Divergence, i.e. To account for the loss.

[0094] Define the original input data With reconstructing data Adversarial loss after input discriminator 12 As shown in equation (6):

[0095] (6)

[0096] in, , All are outputs of discriminator 12; This refers to the total number of data items in the valve module of the converter valve, such as temperature data, electric field strength data, magnetic field strength data, ultrasonic partial discharge data, ultra-high frequency partial discharge data, audio data, and gas concentration data.

[0097] Define the final loss of data input module 1 As shown in equation (7):

[0098] (7)

[0099] in, , , , These are hyperparameters, representing the weights of each loss.

[0100] In this embodiment, take , , , During the training phase, the final loss can be calculated according to the values ​​mentioned above and equation (7). Train and update the system. Once the training is stable, discard discriminator 12 and use only the reconstruction loss. and recoding loss This serves as a standard for risk prediction.

[0101] Define the low-dimensional vector after processing by the integrated encoder 22 As shown in equation (8):

[0102] (8)

[0103] in, The data generated by the multiple input decoders 14 (i.e., the reconstructed data) A collection of )

[0104] The integrated encoder 22 of the integrated module 2 receives data generated by the previous modules (i.e., multiple data input modules 1), merges them together as input data, and obtains a low-dimensional vector through the integrated encoder 22. The reconstructed data is then obtained after inverse transformation by the integrated decoder 23. .

[0105] Define the reconstruction loss after inverse transformation by integrated decoder 23. As shown in equation (9):

[0106] (9)

[0107] in, It is a high-dimensional vector that has been remapped by the integrated decoder 23.

[0108] Then, the reconstruction loss can be calculated. To train the data reconstruction system. When the reconstruction loss... A larger value indicates that the input data to the data reconstruction system is more likely to be abnormal data.

[0109] S4: Optimize training;

[0110] Based on the different data collected by each of the data input modules 1, optimization training is performed separately;

[0111] If the current training is for data input module 1, multiple data input modules 1 can be trained in parallel. Each data input module 1 is trained in the manner of generative adversarial network training, that is, when training the generator (input decoder 14), the discriminator 12 is frozen; when training the discriminator 12, the generator (input decoder 14) is frozen.

[0112] If the current training is of data input module 1, then freeze ensemble module 2; if the current training is of ensemble module 2, then freeze data input module 1.

[0113] The input data for the data reconstruction system consists of different types of data collected by the flexible valve hall and the inspection robot. Considering that the data reconstruction system is composed of multiple modules, and each module contains a corresponding neural network, in order to reduce the computational load, the collected data is first subjected to dimensionality reduction processing to obtain low-dimensional data, which is then used as the input to each network, enabling faster network training. Multiple data input modules 1 can be trained in parallel, and each module is trained according to the generative adversarial network training method. When training the generator, the discriminator 12 is frozen; when training the discriminator 12, the generator is frozen.

[0114] For the entire data reconstruction system, when training ensemble module 2, other modules are frozen; and when training other modules, ensemble module 2 is frozen.

[0115] Once the specified number of training iterations is reached, training is stopped, and the entire data reconstruction system can then be used for risk prediction.

[0116] S5: Risk Prediction;

[0117] After the optimization training is completed, each of the data input modules 1 collects a new set of data. The control center then judges the new set of data and predicts its risk score.

[0118] If the risk score is within the training loss range of the data reconstruction system, the data is considered to have no risk or low risk; if the risk score exceeds or is close to the maximum loss value, the data is considered to have risk and should be examined.

[0119] Based on the methods used in the data reconstruction system, the reconstruction loss can be directly calculated. The training loss is used as the final risk prediction score. During training, the network records the range of training loss for the data reconstruction system. After training, when a new set of data is collected, the data reconstruction system can provide risk scores for each module. If the obtained score is within the loss range of the data reconstruction system, we consider the data to have no risk or low risk. If the obtained score exceeds or approaches the maximum loss value, it indicates that the data set does not conform to the "pattern" of previously collected data, i.e., it has not occurred before. Therefore, the data set is considered to have risk and should be examined. The magnitude of the value represents the degree of risk.

[0120] For data that has already been used in training, it can still be re-input into the already trained data reconstruction system. When the trained data obtains a risk score close to the highest value after being processed by the data reconstruction system, it can also be regarded as potentially risky data. Therefore, the final data reconstruction system can not only determine the risk level of the subsequently generated data, but also play an important role in our specified risk evaluation criteria.

[0121] Furthermore, the inspection robot can collect data including temperature data, electric field strength data, magnetic field strength data, ultrasonic partial discharge data, ultra-high frequency partial discharge data, audio data, and gas concentration data.

[0122] Specifically, the work of the valve hall inspection robot system is as follows:

[0123] 1. The electric field strength sensor, magnetic field strength sensor, ultrasonic partial discharge sensor, ultra-high frequency partial discharge sensor, audio sensor, gas concentration sensor, and infrared temperature measurement system are connected to the control center via network cable. The data monitored by the sensors is transmitted to the robot's main control server. The following text will use the infrared temperature measurement system as an example for explanation.

[0124] 2. The control center obtains the raw data collected by the sensors through the network cable, which can be used for post-processing and analysis, and stores it in the control center.

[0125] 3. All control commands for the robot are sent from the control center to the robot via a WIFI line.

[0126] 4. Users can set any number of monitoring locations on the robot, which facilitates automatic or manual monitoring of the working status of valve hall equipment.

[0127] 5. Users can set multiple temperature measurement targets for each monitoring location and use various methods such as the highest temperature in the area, the average temperature in the area, the lowest temperature in the area, and point temperature to obtain the surface temperature of the valve module and measure the partial discharge parameters of the valve module.

[0128] 6. The system allows setting multiple alarm states, such as over-temperature alarm and temperature difference alarm, and appropriate alarm parameters can be set for each temperature measurement target of each preset position.

[0129] 7. Allow users to set the system's automatic inspection mode. For example, the robot can be set to stay at the preset position 1 for 10 minutes at 10:00 AM to measure the temperature of all preset temperature measurement targets and store an infrared thermal image. At 10:10 AM, it can move to the inspection position 3 and stay there for 10 minutes.

[0130] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for determining the operating status of a converter valve, characterized in that: Includes the following steps: S1: Data Acquisition; The inspection robot measures and collects various data of the valve module of the converter valve within a set range at the predetermined inspection points; S2: Obtain raw data; The inspection robot transmits the collected detection data to the control center, forming raw data, which is then recorded in the control center. S3: Data integration; The control center transmits the raw data to the data reconstruction system for integration; The data reconstruction system includes at least two data input modules (1) and an integration module (2). The control center selects data from the original data, and each data item is transmitted to a data input module (1) for reconstruction. The abnormality of the data item is judged based on the reconstruction error and the intermediate variable error. The reconstruction data corresponding to each data input module (1) is finally integrated in the integration module (2) to obtain the overall risk score. Finally, the risk score is obtained by weighting the reconstruction error corresponding to each data input module (1) to complete the analysis and evaluation of the working status of the valve module of the converter valve. S4: Optimize training; Based on the different data collected by each of the data input modules (1), the data input modules (1) are optimized and trained respectively; or, the integration module (2) is optimized and trained. S5: Risk Prediction; After the optimization training is completed, each of the data input modules (1) collects a new set of data. The control center then judges the new set of data and predicts its risk score. The inspection robot can collect data including temperature data, electric field strength data, magnetic field strength data, ultrasonic partial discharge data, ultra-high frequency partial discharge data, audio data, and gas concentration data.

2. The method for determining the operating state of a converter valve according to claim 1, characterized in that: The data input module (1) includes a PCA preprocessing unit (11) and a generative adversarial network based on a variational autoencoder architecture; the generative adversarial network includes a variational autoencoder network, a discriminator (12), and a verification encoder (15); the variational autoencoder network includes an input encoder (13) and an input decoder (14); the PCA preprocessing unit (11) processes the data from the control center as input data; the input encoder (13) encodes the input data, mapping high-dimensional data to low-dimensional data; the input decoder (14) restores the encoded low-dimensional data back to high-dimensional data through inverse transformation; the verification encoder (15) re-encodes the high-dimensional data after the inverse transformation of the input decoder (14) into low-dimensional data; the discriminator (12) is used to determine the consistency between the low-dimensional data re-encoded by the verification encoder (15) and the original low-dimensional data input by the input encoder (13).

3. The method for determining the operating state of a converter valve according to claim 2, characterized in that: The integrated module (2) includes an autoencoder network (21); the autoencoder network (21) includes an integrated encoder (22) and an integrated decoder (23); the integrated encoder (22) is used to encode the set of data generated by the input decoder (14) in the multiple data input modules (1) and then input it as input data into the integrated decoder (23).

4. The method for determining the operating status of a converter valve according to claim 3, Its features include: S3: In the data integration, a low-dimensional vector is defined after processing by the input encoder (13). As shown in equation (1): （1） in, The original input data is processed by the PCA preprocessing unit (11); the output of the input encoder (13) is the mean of a Gaussian distribution. and variance ; Define the reconstructed data after inverse transformation by the input decoder (14). As shown in equation (2): （2） The reconstruction error corresponding to the data input module (1) after the inverse transformation by the input decoder (14) is defined as the reconstruction loss. Then the reconstruction loss As shown in equation (3): （3） Define the recoding loss after recoding by the verified encoder (15). As shown in equation (4): （4） in, The low-dimensional vector is re-encoded by the verified encoder (15); Define the encoding distribution as a standard normal distribution. Loss results during divergence As shown in equation (5): （5） in, Let be the dimension of the Gaussian distribution, and be the hyperparameter. , These are the Gaussian distributions predicted by the variational autoencoder. and The The mean and standard deviation of the dimension; Define the original input data With reconstructing data Adversarial loss after input discriminator (12) As shown in equation (6): （6） in, , All of these are the outputs of the discriminator; This represents the total number of data items for the valve module of the converter valve. Define the final loss of the data input module (1) As shown in equation (7): （7） in, , , , These are hyperparameters, representing the weights assigned to each loss function; The intermediate variable loss corresponding to the data input module (1) is the recoding loss. Loss Combating losses A collection of [items / items].

5. A method for determining the working state of a converter valve according to claim 4, characterized in that: S3: In the data integration, a low-dimensional vector is defined after processing by the integrated encoder (22). As shown in equation (8): （8） in, A collection of data generated by multiple input decoders (14); The reconstruction loss is defined as the reconstruction error corresponding to the integrated module (2) after the inverse transformation by the integrated decoder (23). Then the reconstruction loss As shown in equation (9): （9） in, It is a high-dimensional vector that has been remapped by the integrated decoder (23).

Images