Abnormality detection method for state information of control rod driving mechanism

Abstract

The invention belongs to the technical field of anomaly detection of state information of a control rod driving mechanism, and particularly relates to an anomaly detection method of the state information of the control rod driving mechanism. Acquiring coil current historical data of the control rod driving mechanism; the data is cleaned and normalized; constructing an SAE model, inputting the coil current normal data of the control rod driving mechanism into the SAE model for feature learning, learning the original features of the coil current data of the control rod driving mechanism, and enabling the SAE model to learn a normal operation model of the control rod driving mechanism; coil current data of the control rod driving mechanism in actual operation are put into the model to obtain a reconstruction error; and constructing a random forest algorithm, taking the reconstruction error as the input of the algorithm, and completing the fault detection of the coil current of the control rod driving mechanism. According to the invention, long-sequence and periodic abnormal events can be processed; compared with a traditional fault detection algorithm, the provided SAE and random forest combined method has higher optimization capability and better detection effect.

Description

Technical Field

[0001] The present invention belongs to the technical field of abnormal detection of control rod drive mechanism status information, and particularly relates to an abnormal detection method for control rod drive mechanism status information. Background Art

[0002] The control rod drive mechanism is the only movable equipment unit inside the reactor pressure vessel and is the actuator of the reactor control and nuclear safety protection system. The control rod can control the normal operation of the reactor in a pressurized water reactor nuclear power plant. Therefore, the normal operation of the control rod should be ensured under two conditions. During protection actions, it is necessary to ensure that when the control rod drive structure (Control Rod Drive Mechanism, CRDM) loses power, the claw releases the control rod and the control rod falls into the reactor by gravity. The reaction of the reactor is stopped by the lowering of the control rod to protect the reactor. During normal operation, it is necessary to ensure the action of the claw transferring the magnetic pole during the normal timing process to prevent the reactor from shutting down or reducing power due to abnormal action conditions and ensure the correct action of the CRDM. The reactor relies on the control rod drive mechanism to tow the control rod to achieve reactor startup, power regulation, reactivity compensation, and safe shutdown. Nuclear power plants adopt a magnetic lifting type drive mechanism that relies on coil energization to achieve control rod movement. The control rod relies on the alternating action of 3 groups of coils in the drive mechanism driven by current signals to perform a stepping motion. Each movement requires the close cooperation of 3 groups of coils and the claw. Slight errors may cause faults such as pulling, rod dropping, or lifting failure. Once the control rod fails, it will cause the reactor's heat output to exceed the limit, which may lead to a reactor explosion and cause serious nuclear leakage. The current information of the control rod drive mechanism coil contains the action information of the control rod, and the action condition and status information of the control rod can be judged based on the change of the current. Fault detection of the coil current of the control rod drive mechanism can achieve timely and accurate identification of control rod drive mechanism faults. Therefore, it is of great significance to detect the coil current faults of the control rod drive mechanism in nuclear power plants.

[0003] Some scholars at home and abroad mostly use machine learning for fault detection of the coil current of the control rod drive mechanism, such as traditional abnormal detection algorithms like Gaussian anomaly detection, SVM, KNN, PCA, etc. Most of the existing abnormal detection methods are unable to handle long sequences and periodic abnormal events, and it is difficult to extract features from these data and accurately detect faults. Summary of the Invention

[0004] The object of the present invention is to provide an abnormal detection method for the state information of a control rod drive mechanism. By using an artificial intelligence algorithm and deeply mining the coil current data of the control rod drive mechanism, the faults of the control rod drive mechanism can be identified in a timely and accurate manner. The stacked autoencoder network of SAE (autoencoding algorithm) can automatically and effectively extract the original features of the CRDM coil current, and then construct the output value of the current signal based on these features. The random forest algorithm is used to monitor the error and identify the abnormal values of the coil current.

[0005] To achieve the above object, the technical solution adopted by the present invention is as follows:

[0006] An abnormal detection method for the state information of a control rod drive mechanism,

[0007] Step 1: Obtain the historical data of the coil current of the control rod drive mechanism;

[0008] Step 2: Clean and normalize the data;

[0009] Step 3: Construct an SAE model, input the normal data of the coil current of the control rod drive mechanism into the SAE model for feature learning, learn the original features of the coil current data of the control rod drive mechanism, and let the SAE model learn the normal operation model of the control rod drive mechanism;

[0010] Step 4: Put the coil current data of the actually operating control rod drive mechanism into the model to obtain the reconstruction error;

[0011] Step 5: Construct a random forest algorithm, use the reconstruction error as the input of the algorithm, and complete the fault detection of the coil current of the control rod drive mechanism.

[0012] Step 2: Use a data cleaning algorithm to preprocess the normal data of the coil current of the control rod drive mechanism and normalize it to between (0, 1).

[0013] Step 3: SAE, that is, the stacked autoencoder neural network is a deep neural network with hidden layers. Each autoencoder in the SAE is pre-trained independently, using the original input data in an unsupervised manner. For each autoencoder, its purpose is to compress the input features into a higher-level and more abstract representation form, which is the output of the encoder. Then, the decoder part is used to decompress the encoded vector into the original input. Each autoencoder repeats this unsupervised training process until the reconstruction meets certain conditions. By continuously stacking different autoencoders, a deeper SAE network is formed. The most important thing about the SAE neural network is that the reconstructed data should be as close as possible to the original data, that is, the error between them needs to be minimized. The reconstruction error is calculated as follows:

[0014]

[0015] Among them, N and M represent the dimension and the number of samples of the original data respectively. The goal of model training is to find w and b that minimize the loss function. After training with the SAE neural network, a fault detection model is obtained.

[0016] The fifth step: The random forest consists of many decision trees, which are independent of each other. Each tree will obtain an independent classification result for a sample. After voting and selection, the random forest will finally take the one with the most votes as the classification result. The hyperparameters adjusted by the random forest include the number of decision trees and the number of classifier samplings. Among them, adjusting the number of classifier samplings is to increase randomness and reduce the correlation between different trees. Assuming that the feature dimension of the dataset is M, then m features are randomly and with replacement selected as samplings each time, and m is the number of classifier samplings, M = m 2 。

[0017] The beneficial effects achieved by the present invention are as follows:

[0018] Aiming at the long-sequence periodicity of the control rod drive mechanism coil current data, the SAE model is used to learn the original features of the normal operation data, and there will be a large reconstruction error when the fault data is input. The random forest can be used to process high-dimensional data, is relatively robust to missing values and outliers, and performs feature selection. Using the random forest algorithm to detect faults in the reconstruction error constructed by the SAE model has a better fault detection effect compared with the threshold. It can handle long-sequence and periodic abnormal events; it can learn the non-linear features and time-series features of the control rod drive mechanism coil current; the proposed method combining SAE and random forest has stronger optimization ability and better detection effect than traditional fault detection algorithms. Description of the Drawings

[0019] Figure 1 It is a flowchart of an abnormal detection method for the state information of the control rod drive mechanism. Detailed Embodiment

[0020] The present invention will be described in detail below with reference to the drawings and specific embodiments.

[0021] The first step: Obtain the historical data of the control rod drive mechanism coil current.

[0022] The second step: Use the data cleaning algorithm to preprocess the normal data of the control rod drive mechanism coil current and normalize it to between (0, 1).

[0023] The third step: Build an SAE model, input the normal data of the control rod drive mechanism coil current into the SAE model for feature learning, learn the original features of the control rod drive mechanism coil current data, and let the SAE model learn the normal operation model of the control rod drive mechanism.

[0024] The stacked autoencoder neural network (SAE) is a deep neural network with hidden layers. Each autoencoder in the SAE is pre-trained independently and uses the original input data in an unsupervised manner. For each autoencoder, its purpose is to compress the input features into a higher-level and more abstract representation, which is the output of the encoder, and then use the decoder part to decompress the encoded vector into the original input. Each autoencoder repeats this unsupervised training process until the reconstruction meets certain conditions. By continuously stacking different autoencoders, a deeper SAE network is formed. Although training a deep neural network is more complex, by stacking layers, the problem of gradient disappearance can be avoided, and the stability and reliability of the model can be improved.

[0025] The most important thing for the SAE neural network is that the reconstructed data should be as close as possible to the original data; that is, the error between them needs to be minimized. Therefore, the reconstruction error can be calculated as follows:

[0026]

[0027] where N and M represent the dimensions and the number of samples of the original data respectively. The goal of model training is to find w and b that minimize the loss function.

[0028] After training with the SAE neural network, a fault detection model is obtained.

[0029] Step 4: Put the coil current data of the actual operating control rod drive mechanism into the model to obtain the reconstruction error.

[0030] Step 5: Construct a random forest algorithm, use the reconstruction error as the input of the algorithm, and complete the fault detection of the control rod coil current. The main idea of the random forest is ensemble learning, which consists of many decision trees. Different trees are independent of each other, and each tree will get an independent classification result for a sample. After voting and selection, the random forest will finally take the result with the most votes as the classification result. The main hyperparameters adjusted by the random forest include the number of decision trees and the number of classifier samplings. Among them, adjusting the number of classifier samplings is to increase randomness and reduce the correlation between different trees. Suppose the feature dimension of the dataset is M, then each time m features are randomly selected with replacement as the sampling, and m is the number of classifier samplings. Generally, M = m 2 .

[0031] Step 1: Obtain the historical data of the coil current of the control rod drive mechanism;

[0032] Step 2: Clean and normalize the data;

[0033] Step 3: Construct an SAE model to learn the normal operation mode;

[0034] Step 4: Input actual data to obtain the reconstruction error;

[0035] Step 5: Construct a random forest algorithm for fault detection.

Claims

1. An abnormal detection method for the state information of a control rod drive mechanism, characterized in that: The first step: Obtain the historical data of the coil current of the control rod drive mechanism; The second step: Clean and normalize the data; The third step: Construct an SAE model, input the normal data of the coil current of the control rod drive mechanism into the SAE model for feature learning, learn the original features of the coil current data of the control rod drive mechanism, and let the SAE model learn the normal operation model of the control rod drive mechanism; The fourth step: Put the coil current data of the actually operating control rod drive mechanism into the model to obtain the reconstruction error; The fifth step: Construct a random forest algorithm, use the reconstruction error as the input of the algorithm, and complete the fault detection of the coil current of the control rod drive mechanism.

2. The abnormal detection method for the control rod drive mechanism status information according to claim 1, characterized in that: The second step: Use a data cleaning algorithm to preprocess the normal data of the coil current of the control rod drive mechanism and normalize it to between (0, 1).

3. The abnormal detection method for the control rod drive mechanism status information according to claim 1, characterized in that: The third step: SAE, that is, the stacked autoencoder neural network is a deep neural network with hidden layers. Each autoencoder in the SAE is pre-trained independently and uses the original input data in an unsupervised manner. For each autoencoder, its purpose is to compress the input features into a higher-level and more abstract representation form, which is the output of the encoder, and then use the decoder part to decompress the encoded vector into the original input. Each autoencoder repeats this unsupervised training process until the reconstruction meets certain conditions. By continuously stacking different autoencoders, a deeper SAE network is formed; the most important thing about the SAE neural network is that the reconstructed data should be as close as possible to the original data, that is, the error between them needs to be minimized. The reconstruction error is calculated as follows: Where N and M represent the dimensions and the number of samples of the original data respectively. The goal of model training is to find w and b that minimize the loss function; after training with the SAE neural network, a fault detection model is obtained.

4. The abnormal detection method for the state information of the control rod drive mechanism according to claim 1, characterized in that: Step 5: A random forest consists of many decision trees. These trees are independent of each other. For each sample, each tree will obtain an independent classification result. After voting, the random forest will finally take the result with the most votes as the classification result. The hyperparameters adjusted by the random forest include the number of decision trees and the number of classifier samplings. Among them, adjusting the number of classifier samplings is to increase randomness and reduce the correlation between different trees. Assuming that the feature dimension of the dataset is M, then each time m features are randomly selected with replacement as the sampling, and m is the number of classifier samplings, where M = m 2 。

Images