CN116431966A - Reactor core temperature anomaly detection method of incremental characteristic decoupling self-encoder - Google Patents

Abstract

The invention discloses a core temperature anomaly detection method of an incremental feature decoupling self-encoder. Aiming at the problem that traditional decoupling representation learning is difficult to pre-designate feature dimensions, the present invention designs a feature increment strategy, which generates latent space features step by step during autoencoder model training and adaptively determines feature dimensions. At the same time, an iterative training strategy based on dual performance indicators is proposed for model training, so that the features extracted by the autoencoder model have a strong ability to reconstruct data and meet the requirements of latent space feature decoupling. Finally, the feature space and residual space of the core temperature data are described respectively by statistics, so as to realize the comprehensive anomaly detection of the reactor core temperature. The method of the invention can effectively reduce the false alarm rate of faults and improve the fault detection rate in the abnormal detection task of the temperature data of multiple measuring points in the nuclear reactor core, and provide practical help for the safe and stable operation and intelligent operation and maintenance of the nuclear reactor.

Description

An Incremental Feature Decoupling Autoencoder Core Temperature Abnormal Detection Method

technical field

The invention discloses a core temperature anomaly detection method of an incremental feature decoupling self-encoder. The invention belongs to the field of industrial fault detection, in particular to the abnormal detection of core temperature of nuclear reactors.

Background technique

Nuclear power generation is a clean, economical and efficient power generation method, which puts forward higher requirements for the continuous, safe and stable operation of nuclear power equipment. The reactor core is the energy core of the entire nuclear power system, and the core temperature is the most direct indicator of the health status of the core. If abnormalities in the temperature distribution cannot be detected in time, it is very likely to cause a series of major accidents such as core meltdown. , causing serious casualties and economic losses. Therefore, it is of great significance for the production safety of nuclear power plants to carry out abnormal detection work for nuclear reactor core temperature, find faults in time, and avoid major accidents. The traditional core temperature anomaly detection is generally carried out with the idea of post-accident monitoring, relying on the professional experience and mechanism knowledge of the operators to judge and summarize the core temperature change trend during and after the accident, the labor cost is high, and the detection efficiency is low. And the real-time performance is poor. In recent years, with the development of machine learning and artificial intelligence technology, data-driven anomaly detection methods have been greatly developed and widely used in industrial process fault detection tasks with good results. Therefore, it is very urgent to develop nuclear reactor fault detection based on data-driven methods combined with the characteristics of nuclear power core temperature data.

The data-driven anomaly detection method does not rely on mechanism-related expertise, but only relies on a large amount of data collected during system operation to capture information such as potential coupling relationships between variables, and can achieve efficient and real-time anomaly detection. Among them, the more commonly used methods are principal component analysis (principle component analysis, PCA), partial least square (partial least square, PLS), independent component analysis (independent component analysis, ICA) and other multivariate statistical analysis methods, and autoencoders (auto -encoder, AE), convolutional neural network (convolutional neural network, CNN) and other deep learning methods. However, there is a complex nonlinear coupling relationship between different measurement points of the core temperature, and multivariate statistical analysis methods are generally difficult to effectively capture the nonlinear characteristics of the data, and it is difficult to achieve high-accuracy anomaly detection. Although the deep learning method can use the nonlinear mapping between neurons to learn the hidden nonlinear features in the data, the features extracted in the hidden space of the model may have a strong coupling relationship, that is, there is a problem of information redundancy. The monitoring statistics calculated based on redundant features may not be able to accurately describe the degree of deviation of the test data from the model in the feature space in the anomaly detection task, resulting in false negatives and false positives.

Anomaly detection methods based on decoupled representation learning design decoupling constraints on latent space generation based on traditional deep learning models to overcome the problem of feature redundancy, such as variational auto-encoders (VAE), etc. method. However, such methods often need to artificially preset and fix the dimension of the latent space of the model before training, and the choice of the hidden space dimension has a great impact on the performance of the model. The application in the modeling process poses difficulties.

Contents of the invention

The purpose of the present invention is to solve the problem that traditional decoupling representation learning is difficult to specify the feature dimension in advance, and propose a core temperature anomaly detection method of incremental feature decoupling autoencoder, which generates the latent space of autoencoder model step by step feature and adaptively determine the feature dimension to meet the dual requirements of the features extracted by the model, which have a strong reconstruction ability for the data and fully decouple the latent space features of the model, and construct the latent space and residual space based on the core temperature data respectively Statistics are used for anomaly detection.

The purpose of the present invention is achieved through the following technical solutions:

A core temperature anomaly detection method for an incremental feature decoupling autoencoder, specifically:

Input the real-time collected nuclear reactor core temperature data samples into the trained incremental feature decoupling autoencoder model to obtain feature vectors and reconstructed samples, and calculate statistics based on the feature vectors and reconstructed samples to perform anomalies on the core temperature detection;

The incremental feature decoupling autoencoder model is obtained by training as follows:

Build a training set, each sample of the training set is nuclear reactor core temperature data collected when the nuclear reactor is in normal operation;

Build an incremental feature decoupling autoencoder, the incremental feature decoupling autoencoder is composed of an input layer I, an adjustment layer P, a feature layer F and an output layer O; set the adjustment layer P, feature layer F The number of initial neurons;

Input the samples of the training set to the incremental feature decoupling autoencoder to obtain feature vectors and reconstructed samples, and perform neuron incremental iterative training on the incremental feature decoupling autoencoder based on the loss function until the model performance index meets Requirements; the loss function includes: a first loss function, composed of reconstruction loss; a second loss function, composed of the sum of reconstruction loss and hidden space decoupling loss; the model performance index includes: reconstruction error index R , is numerically the same as the reconstruction loss; the latent space feature correlation index C is numerically the same as the latent space decoupling loss; where:

If C<C _th and R<R _th , it is considered that the current model reconstruction ability and hidden space decoupling degree meet the requirements, and the model training is completed;

If C<C _th and R>R _th , it is considered that the degree of decoupling of the hidden space of the current model meets the requirements, but the reconstruction ability is insufficient, and neurons need to be incremented and trained in the feature layer F;

If C>C _th and R<R _th , it is considered that the reconstruction ability of the current model meets the requirements, but the degree of decoupling of the hidden space is insufficient, and neurons need to be incremented and trained in the adjustment layer P;

If C>C _th and R>R _th , it is considered that the current model reconstruction ability and latent space decoupling degree are not up to standard, and it is necessary to give priority to meeting the requirements for latent space decoupling degree, and perform neuron increment and training in the adjustment layer P ;

Among them, C _th and R _th represent the thresholds of latent space feature correlation index C and reconstruction error index R, respectively.

Further, after the neuron increment in the feature layer F, during training, the network parameters of the mapping between the input layer I and the adjustment layer P remain unchanged, and the mapping between the adjustment layer P and the feature layer F is fixed. The network parameters of the round of training and update the network parameters of the new neurons used to generate the new feature vector f _k in this round. The subscript k represents the hidden space dimension of the current model after the addition of neurons, the feature layer F and the output layer O The mapping between network parameters is completely updated.

Further, after the adjustment layer P performs neuron increment, during training, the mapping between the input layer I and the adjustment layer P fixes the network parameters participating in the previous round of training and updates the new neurons in this round to generate The network parameters of the new vector p _j+1 , the subscript j represents the number of neurons in the regulation layer P that participated in the previous round of training, the mapping between the regulation layer P and the feature layer F fixes the network parameters that participated in the previous round of training and Update the network parameters of the new feature vector generated based on the matrix containing the new vector p _j+1 in this round, and the network parameters of the mapping between the feature layer F and the output layer O are completely updated.

Further, the hidden space decoupling loss Loss _C is expressed as:

Cov(f _k ,f _i )＝E(f _k ·f _i )-E(f _k )E(f _i )

where k is the hidden space dimension of the current model, f _k is the k-th dimension feature vector output by the feature layer F, and f _i (i=1,2,...,k-1) is the k- 1 eigenvector, the function E calculates the mean of the eigenvectors.

Further, the reconstruction loss Loss _R is expressed as:

是重构样本矩阵，n是样本数。Where ||·|| _F represents the Frobenius norm, X is the nuclear reactor core temperature data sample matrix input into the model,

is the reconstructed sample matrix, and n is the number of samples.

Further, the second loss function is expressed as:

Loss _total ＝Loss _R +βLoss _C

Among them, β is the model hyperparameter, Loss _R is the reconstruction loss, and Loss _C is the hidden space decoupling loss.

Further, the statistics include ^T2 and SPE statistics.

Further, based on the eigenvector and the reconstructed sample calculation statistics, the abnormal detection of the core temperature is carried out, specifically:

Statistics are calculated based on eigenvectors and reconstructed samples. If any of the calculated statistics exceeds the control limit, it indicates that the nuclear reactor has failed during operation.

Further, the control limit of the statistic is calculated by means of kernel density estimation.

Further, the nuclear reactor core temperature data includes core temperatures collected by sensors distributed at different positions of the reactor core.

The beneficial effect of the present invention is: the present invention aims at the problem that the traditional decoupling representation learning is difficult to pre-designate the feature dimension, and proposes an incremental feature decoupling self-encoder core temperature anomaly detection method. The present invention designs a feature increment strategy, which generates latent space features of the self-encoder model step by step and adaptively determines the feature dimension when training the autoencoder model. At the same time, an iterative training strategy based on dual performance indicators is proposed for the training of the autoencoder model, so that the features extracted by the model have a strong ability to reconstruct the data and meet the requirements of latent space feature decoupling. Finally, the present invention respectively constructs monitoring statistics based on the feature space and residual error of the data, and realizes comprehensive anomaly detection of core temperature data. Compared with the anomaly detection method based on traditional deep learning, the present invention realizes the decoupling of latent space features of the model, and builds statistics that can capture data anomalies more accurately based on the latent space, reducing false positives and missed negatives in fault detection . Compared with the anomaly detection method based on traditional decoupling representation learning, the present invention can realize self-adaptive determination of latent space feature dimension, effectively reducing the application difficulty of the method in the actual modeling process.

Description of drawings

Fig. 1 is the overall framework flowchart of the present invention

Fig. 2 is the fault detection result diagram of the method proposed by the present invention on the test set 1, wherein the dotted line is the control limit, and the solid line is the statistical calculation result of the test set sample;

Fig. 3 is the diagram of the fault detection results of the principal component analysis method on the test set 1, where the dotted line is the control limit, and the solid line is the statistical calculation result of the test set samples;

Figure 4 is a diagram of the fault detection results of the traditional autoencoder method on the test set 1, where the dotted line is the control limit, and the solid line is the statistical calculation result of the test set samples;

Figure 5 is a diagram of the fault detection results of the direct decoupling autoencoder method on the test set 1, where the dotted line is the control limit, and the solid line is the statistical calculation result of the test set samples;

Fig. 6 is the fault detection result diagram on the test set 2 of the proposed method of the present invention, wherein the dotted line is the control limit, and the solid line is the statistical calculation result of the test set sample;

Fig. 7 is the diagram of the fault detection results of the principal component analysis method on the test set 2, where the dotted line is the control limit, and the solid line is the statistical calculation result of the test set samples;

Fig. 8 is a diagram of the fault detection results of the traditional autoencoder method on the test set 2, wherein the dotted line is the control limit, and the solid line is the statistical calculation result of the test set samples;

Figure 9 is a diagram of the fault detection results of the direct decoupling autoencoder method on the test set 2, where the dotted line is the control limit, and the solid line is the statistical calculation result of the test set samples;

Detailed ways

In this embodiment, the effectiveness of the method is verified by taking the core temperature data of a real nuclear reactor as an example. The present invention will be described in further detail below in conjunction with the accompanying drawings and specific embodiments.

The core temperature abnormal detection method of an incremental feature decoupling self-encoder of the present invention is to input the nuclear reactor core temperature data samples collected in real time into the trained incremental feature decoupling self-encoder model, and obtain Eigenvectors and reconstructed samples are used to calculate statistics based on eigenvectors and reconstructed samples to detect anomalies in core temperature. Wherein, the training of the incremental feature decoupling autoencoder model is the focus of the present invention, as shown in Figure 1, the offline modeling training includes the following steps:

Step 1: constructing a training set, each sample of the training set is nuclear reactor core temperature data collected during normal operation of the nuclear reactor;

In this embodiment, 6,000 normal samples are collected as a training set to establish a model, and 6,000 samples are taken for two different types of faults that occur in the process as two test sets for anomaly detection testing, each sample contains 40 test samples. The temperature variables of the points are collected by sensors axially distributed in the top, middle, and bottom parts of the nuclear reactor core, and the sampling interval of the samples is 60 seconds. The faults of the two test sets are caused by sudden and slow changes of the core temperature distribution respectively, and the faults begin to appear from the 2000th sample.

Further, standardize the collected data, as follows:

标准化的计算公式为：For the collected normal core temperature data

The standardized calculation formula is:

是所有样本中第r个过程变量的均值；s_r表示所有样本中第r个过程变量的标准差；x_i,r则是对应样本、对应变量经过标准化后的数值。对数据X进行标准化后可得数据矩阵/>

Among them, n represents the number of samples, m represents the number of variables; x _{i, r} represents the element in the i-th row and r-th column in the data matrix X, that is, the value of the r-th temperature variable in the i-th sample collected;

is the mean value of the r-th process variable in all samples; s _r represents the standard deviation of the r-th process variable in all samples; x _i,r is the standardized value of the corresponding sample and corresponding variable. After normalizing the data X , the data matrix can be obtained />

Step 2: Build an incremental feature decoupling autoencoder, model the normal data X, obtain feature vectors and reconstructed samples, and perform neuron incremental iterative training on the incremental feature decoupling autoencoder based on the loss function Until the model performance indicators meet the requirements; specifically include the following sub-steps:

Step 2.1: The incremental feature decoupling autoencoder model consists of an input layer I, an adjustment layer P, a feature layer F and an output layer O, where the feature layer F is defined as the model latent space, and the number of neurons in the layer corresponds to The feature dimension of the latent space. In the present invention, the feature layer F is designed in a form in which neurons can be increased. When the model has a poor ability to reconstruct data, it is necessary to add new neurons to the feature layer F for training. In the present invention, the adjustment layer P is also designed in the form of neurons that can be increased. If the decoupling degree of the feature layer F is insufficient during the training process, it is necessary to add new neurons to the adjustment layer P for training.

分别表示模型在生成一维特征时，输入层I与调节层P之间、调节层P与特征层F之间、特征层F与输出层O之间的映射，则信息在模型中的传递可以总结为：输入数据/>

的全部信息将先被映射至调节层P的j个神经元中得到矩阵

即/>

然后P所包含的信息再进一步被压缩至特征层F的一个神经元中，得到特征向量/>

即/>

最后，特征f₁被映射至输出层O，得到重构数据/>

即/>

记映射/>

包含的网络参数分别为/>

模型在初始状态下训练的第一损失函数与传统自编码器相同，为重构损失Loss_R，本实施例采用的计算方法为：Step 2.2: Set the number of neurons in the feature layer F and adjustment layer P of the incremental feature decoupled autoencoder model in the initial state; in general, the initial number of neurons in the feature layer F is set to 1, which can be Maximum adaptive increment. In this embodiment, the number of neurons in the feature layer F and the adjustment layer P of the incremental feature decoupled autoencoder model in the initial state is set to 1 and 3, respectively. use

Respectively represent the mapping between the input layer I and the adjustment layer P, between the adjustment layer P and the feature layer F, and between the feature layer F and the output layer O when the model generates one-dimensional features, then the transfer of information in the model can be Summarized as: input data />

All the information of will be mapped to the j neurons of the adjustment layer P to obtain the matrix

i.e. />

Then the information contained in P is further compressed into a neuron in the feature layer F to obtain the feature vector />

i.e. />

Finally, the feature _f1 is mapped to the output layer O to obtain the reconstructed data />

i.e. />

note mapping />

The network parameters included are />

The first loss function of the model trained in the initial state is the same as the traditional autoencoder, which is the reconstruction loss Loss _R , and the calculation method used in this embodiment is:

where ||·|| _F represents the Frobenius norm.

表示模型在生成二维特征时，输入层I与调节层P之间、调节层P与特征层F之间、特征层F与输出层O之间的映射，记上述映射包含的网络参数分别为/>

在本次训练中，对于映射/>

模型固定其参数/>

与上一轮的参数/>

相同且不更新；对于映射/>

其参数/>

由需要固定和更新的两部分参数组成：需要固定的部分与上一轮的参数/>

相同且不更新，需要更新的部分是在本轮中新增的、用于生成f₂的新映射的参数，记为/>

即/>

对于映射/>

模型将其参数/>

进行完全更新，即舍弃上一轮的参数/>

重新训练。在本次训练生成f₂的过程中，模型的损失函数在原有重构损失Loss_R的基础上，添加了隐空间解耦损失项Loss_C，可以采用如下形式：Step 2.3: Set the reconstruction error index R=Loss _R , and set the corresponding threshold R _th . If the model has R>R _th after training in the initial state, add a new neuron to the model feature layer F and retrain to generate the second dimension feature vector f ₂ , then combine f ₁ and f ₂ and map to the output Layer O. use

Indicates the mapping between the input layer I and the adjustment layer P, between the adjustment layer P and the feature layer F, and between the feature layer F and the output layer O when the model generates two-dimensional features. Note that the network parameters included in the above mapping are respectively />

In this training, for the mapping />

The model fixes its parameters />

with the parameters of the previous round />

Same and do not update; for mapping />

its parameters />

It consists of two parts of parameters that need to be fixed and updated: the part that needs to be fixed and the parameters of the previous round />

The same and not updated, the part that needs to be updated is the parameter that is added in this round and used to generate the new mapping of f ₂ , denoted as />

i.e. />

for mapping />

The model will have its parameters />

Perform a complete update, i.e. discard the parameters of the previous round />

Retrain. In the process of generating f ₂ in this training, the loss function of the model is based on the original reconstruction loss Loss _R , and the hidden space decoupling loss item Loss _C is added, which can be in the following form:

Cov(f _k ,f _i )＝E(f _k ·f _i )-E(f _k )E(f _i )

Where k is the hidden space dimension of the current model, f _k is the k-th dimension feature vector, f _i (i=1,2,...,k-1) is the k-1 feature vectors extracted in the previous round, Function E computes the mean of the eigenvectors. The second loss function Loss _total consists of the sum of the reconstruction loss Loss _R and the hidden space decoupling loss Loss _C :

Loss _total ＝Loss _R +βLoss _C

Among them, β is the model hyperparameter. In the subsequent training process, the model will continue to use this loss function for training.

即：将调节层P中新生成的向量p_j+1与矩阵P合并得到矩阵/>

再映射至特征层F以重新生成第二维特征f₂。在本次训练中，对于映射/>

其参数/>

由需要固定和更新的两部分参数组成：需要固定的部分与参数/>

相同且不更新，需要更新的部分是在本轮中新增的、用于生成p_j+1的新映射的参数，记为/>

即/>

对于映射/>

其参数/>

由需要固定和更新的两部分参数组成：需要固定的部分与参数/>

相同且不更新，需要更新的部分是对上一轮中的参数/>

进行更新，用于重新生成f₂，即/>

对于映射

模型将其参数/>

进行完全更新。Step 2.4: Set the latent spatial feature correlation index C=Loss _C , and set the corresponding threshold C _th . If the model has C>C _th after the last round of training, then perform neuron increments to the adjustment layer P and retrain the model mapping

That is: merge the newly generated vector p _j+1 in the adjustment layer P with the matrix P to obtain the matrix />

Then map to the feature layer F to regenerate the second dimension feature f ₂ . In this training, for the mapping />

its parameters />

It consists of two parts of parameters that need to be fixed and updated: the part that needs to be fixed and the parameter />

The same and not updated, the part that needs to be updated is the new parameter used to generate the new mapping of p _j+1 in this round, denoted as />

i.e. />

for mapping />

its parameters />

It consists of two parts of parameters that need to be fixed and updated: the part that needs to be fixed and the parameter />

The same and not updated, the part that needs to be updated is the parameters in the previous round />

is updated to regenerate f ₂ , i.e. />

for mapping

The model will have its parameters />

Do a full update.

Step 2.5: Based on the specific operation of neuron increment in step 2.3 and step 2.4, and so on, based on the loss function, perform neuron incremental iterative training on the incremental feature decoupled autoencoder until the model performance index meets the requirements , the complete model iterative training strategy is summarized as follows:

Judgment condition 1: If C<C _th and R<R _th , it is considered that the current model reconstruction ability and hidden space decoupling degree meet the requirements, and the model training is completed;

Judgment condition 2: If C<C _th and R>R _th , it is considered that the degree of decoupling of the hidden space of the current model meets the requirements, but the reconstruction ability is insufficient, and neurons need to be incremented and trained in the feature layer F.

Judgment condition 3: If C>C _th and R<R _th , it is considered that the reconstruction ability of the current model meets the requirements, but the degree of decoupling of the latent space is insufficient, and neurons need to be incremented and trained in the adjustment layer P.

Judgment condition 4: If C>C _th and R>R _th , it is considered that the current model reconstruction ability and latent space decoupling degree are not up to standard, and the requirements for latent space decoupling degree should be met first, and the operation is the same as condition 3.

Among them, C _th and R _th represent the thresholds of latent space feature correlation index C and reconstruction error index R, respectively. In this example, they are 0.8 and 0.1, respectively.

模型训练完毕。In this embodiment, when the incremental feature decoupled autoencoder model is trained according to the above strategy until the iteration termination condition is met, the number of neurons in the feature layer F and adjustment layer P of the model is 6 and 9, respectively. The model structure is fixed, and the model mapping obtained in the last round of training is denoted as

The model is trained.

其中的每一个样本可表示为/>

使用增量式特征解耦自编码器模型提取其特征向量

同时得到重构输出/>

并计算重构残差项/>

基于上述向量构造T²和SPE统计量如下：For the feature space and residual space of normal core temperature data, statistics are designed for anomaly detection. In this embodiment, ^T2 and SPE are taken as examples, and the calculation method is: for normal data

Each of these samples can be expressed as />

Extract feature vectors from encoder models using incremental feature decoupling

Also get the refactored output />

and calculate the reconstruction residual term />

Construct ^T2 and SPE statistics based on the above vectors as follows:

T ² ＝f _t ^T Σ ^-1 f _t

SPE＝e ^T e

Among them, Σ is the covariance matrix of the features extracted by the model based on normal data. The control limits Ctr _T2 and Ctr _SPE of T ² and SPE statistics were calculated by means of kernel density estimation.

输入至训练好的增量式特征解耦自编码器模型中即可实现故障检测。使用训练好的增量式特征解耦自编码器模型提取其特征向量/>

同时得到重构输出/>

计算重构残差项/>

计算T²和SPE统计量对堆芯温度进行异常检测：So far, two statistics and their control limits have been obtained. Fault detection can be achieved by decoupling the autoencoder model with the trained incremental features. Specifically, the real-time collected nuclear reactor core temperature data sample

Fault detection can be achieved by inputting into the trained incremental feature decoupled autoencoder model. Use the trained incremental feature decoupled autoencoder model to extract its feature vector />

Also get the refactored output />

Compute Reconstruction Residuals />

Compute ^T2 and SPE statistics for core temperature anomaly detection:

T ² ＝ f _new ^T Σ ^-1 f _new

SPE＝e _new ^T e _new

If any statistic of T ² and SPE exceeds the control limit, it indicates that the nuclear reactor is malfunctioning during operation. The overall framework flow chart of the present invention is shown in Fig. 1 . The following uses the samples of the two test sets to perform fault detection and analysis to illustrate the effect of the present invention.

In terms of fault detection, the present invention performs fault detection on test sets of two types of faults, sudden temperature change and slow change, and the results are shown in Fig. 2 and Fig. 6 respectively. It can be found that under sudden faults, the two statistics of this method obviously exceed the control limit at the moment of fault occurrence, and the occurrence of faults can be detected sensitively in time. Under slow-changing faults, the two statistics of this method can also capture slowly and gradually changing abnormalities in the data, and detect the occurrence of faults more accurately.

In order to more clearly reflect the superiority of the feature incremental strategy and iterative training strategy designed in the present invention in anomaly detection tasks, the incremental strategy and iterative training strategy of the incremental feature decoupling self-encoder in the present invention are all discarded. The resulting model is named Directly Decoupled Autoencoder (constrained training with only the second loss function) for experimental comparison. Here the principal component analysis PCA (W.Svante, K.Esbensen, and P.Geladi. "Principalcomponent analysis," Chemom.Intell.Lab.Syst., vol.2, no.1-3, pp.37- 52, 1987), traditional autoencoder AE (Sakurada, Mayu, and Takehisa Yairi. "Anomaly detection using autoencoders with nonlinear dimensionality reduction." Proceedings of the MLSDA 2014 2nd workshop on machine learning for sensory data analysis. 2014.) and direct decoupling from encoders for comparison. The fault detection effects of the above method on the two test sets are shown in Figure 3 and Figure 7, Figure 4 and Figure 8, Figure 5 and Figure 9, respectively. It can be found that most of the above methods can detect the occurrence of faults relatively accurately in the experiments of sudden fault types. However, in the experiments of slowly changing faults, the above method produced a large number of false alarms before the fault occurred, and at the same time, the time to detect the fault was also greatly delayed, and the fault detection rate was low.

For a more intuitive comparison, Table 1 and Table 2 list the fault detection rate (FDR) and fault false alarm rate (FAR) of different methods on the test set, respectively. The calculation expressions of FDR and FAR are:

Among them, N _MAE , N _abnormal , N _FAE , and N _normal respectively represent the number of missed events, abnormal samples, false positive events and normal samples.

It can be found that the fault false alarm rate of the proposed method of the present invention is the lowest among the four methods on the two test sets, and the fault detection rate is also generally at the highest level among the four methods. The superiority of the proposed method compared to other methods is most obvious on the test set 2 with slow-varying faults. This proves the feasibility and effectiveness of the proposed method of the present invention.

Table 1 Comparison of fault detection rate (FDR) of different methods (%)

Table 2 Comparison of False Alarm Rate (FAR) of Different Methods (%)

Apparently, the above-mentioned embodiments are only examples for clear description, rather than limiting the implementation. For those of ordinary skill in the art, other changes or changes in different forms can be made on the basis of the above description. It is not necessary and impossible to exhaustively list all implementation modes here. However, the obvious changes or variations derived therefrom still fall within the protection scope of the present invention.

Claims (10)

1. The method for detecting the abnormal temperature of the reactor core of the incremental characteristic decoupling self-encoder is characterized by comprising the following steps of:

inputting a nuclear reactor core temperature data sample acquired in real time into a trained incremental characteristic decoupling self-encoder model to obtain a characteristic vector and a reconstructed sample, and calculating statistics based on the characteristic vector and the reconstructed sample to perform anomaly detection on the core temperature;

the incremental characteristic decoupling self-encoder model is obtained through training by the following method:

constructing a training set, wherein each sample of the training set is nuclear reactor core temperature data acquired during normal operation of a nuclear reactor;

constructing an incremental characteristic decoupling self-encoder, wherein the incremental characteristic decoupling self-encoder consists of an input layer I, an adjusting layer P, a characteristic layer F and an output layer O; setting the initial neuron numbers of an adjusting layer P and a characteristic layer F;

inputting samples of the training set to an incremental characteristic decoupling self-encoder to obtain feature vectors and reconstructed samples, and performing neuron incremental iterative training on the incremental characteristic decoupling self-encoder based on a loss function until the model performance index meets the requirement; the loss function includes: a first loss function consisting of reconstruction losses; a second loss function consisting of a sum of reconstruction loss and hidden space decoupling loss; the model performance index comprises: the reconstruction error index R is the same as the reconstruction loss in value; the hidden space characteristic correlation index C is the same as the hidden space decoupling loss in value; wherein:

if C < C _th And R < R _th Considering that the reconstruction capability and the hidden space decoupling degree of the current model reach the requirements, and finishing model training;

if C < C _th And R > R _th Considering that the hidden space decoupling degree of the current model meets the requirement, but the reconstruction capability is insufficient, and neuron increment and training are needed in a feature layer F;

if C > C _th And R < R _th Considering that the reconstruction capability of the current model meets the requirement, but the hidden space decoupling degree is insufficient, and neuron increment and training are needed in the adjusting layer P;

if C > C _th And R > R _th Considering that the reconstruction capability and the hidden space decoupling degree of the current model are not up to the standard, and preferentially meeting the requirement on the hidden space decoupling degree is required, and performing neuron increment and training on an adjusting layer P;

wherein C is _th 、R _th The thresholds of the hidden space feature correlation index C and the reconstruction error index R are respectively represented.

2. The method according to claim 1, wherein after the feature layer F performs the neuron increment, the network parameters of the mapping between the input layer I and the adjustment layer P are unchanged, the mapping between the adjustment layer P and the feature layer F fixes the network parameters in which the training of the previous round is participated and updates the newly added neurons in the present round for generating the new feature vector F _k The subscript k represents the hidden space dimension of the current model after the neuron is newly added, and the mapped network parameters between the feature layer F and the output layer O are completely updated.

3. The method according to claim 1, wherein after the adjustment layer P performs the neuron increment, the mapping between the input layer I and the adjustment layer P fixes the network parameters involved in the previous training and updates the newly added neurons in the current training round for generating the new vector P _j+1 The subscript j indicates the number of neurons of the adjustment layer P participating in the previous round of training, the mapping between the adjustment layer P and the feature layer F fixes the network parameters participating in the previous round of training therein and updates the network parameters in the current round based on the new vector P _j+1 The matrix of which generates new feature vectors, the mapped network parameters between feature layer F and output layer O are fully updated.

4. The method of claim 1, wherein the implicit spatial decoupling Loss _C Expressed as:

Cov(f _k ,f _i )＝E(f _k ·f _i )-E(f _k )E(f _i )

where k is the hidden space dimension of the current model, f _k The kth dimension feature vector F output for feature layer F _i (i=1, 2,., k-1) calculating the mean of the feature vectors for k-1 feature vectors extracted for the previous round, function E.

5. The method of claim 1, wherein the reconstruction Loss _R Expressed as:

wherein I II _F Representing the Frobenius norm, X is a matrix of nuclear reactor core temperature data samples of the input model,

is the reconstructed sample matrix and n is the number of samples.

6. The method of claim 1, wherein the second loss function is expressed as:

Loss _total ＝Loss _R +βLoss _C

wherein beta is a model hyper-parameter, loss _R Is reconstruction Loss, loss _C Is the hidden space decoupling loss.

7. The method of claim 1, wherein the statistic comprises T ² And SPE statistics.

8. The method of claim 1, wherein the anomaly detection of core temperature is performed based on the eigenvector and the reconstructed sample calculation statistic, in particular:

and calculating statistics based on the feature vector and the reconstructed sample, and if any calculated statistics exceed a control limit, indicating that the nuclear reactor operation process is faulty.

9. The method of claim 8, wherein the control limit of the statistic is calculated by a method of nuclear density estimation.

10. The method of claim 1, wherein the nuclear reactor core temperature data comprises core temperatures at a plurality of stations acquired by sensors distributed at different locations in a reactor core.

Images