A method for core temperature anomaly detection of an incremental feature decoupled autoencoder
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2023-03-16
- Publication Date
- 2026-06-23
Smart Images
Figure CN116431966B_ABST
Abstract
Description
Technical Field
[0001] This invention discloses a method for detecting core temperature anomalies using an incremental feature-decoupled self-encoder. This invention belongs to the field of industrial fault detection, specifically targeting the detection of core temperature anomalies in nuclear reactors. Background Technology
[0002] Nuclear power generation is a clean, economical, and efficient method of power generation, which places higher demands on the continuous, safe, and stable operation of nuclear power equipment. The reactor core is the energy core of the entire nuclear power system, and core temperature is the most direct indicator of core health. If anomalies in temperature distribution are not detected in time, they can potentially trigger a series of major accidents, such as core meltdown, causing serious casualties and economic losses. Therefore, conducting anomaly detection of nuclear reactor core temperatures to promptly identify faults and prevent major accidents is of great significance to the production safety of nuclear power plants. Traditional core temperature anomaly detection generally adopts a post-accident monitoring approach, relying on the professional experience and mechanistic knowledge of operators to judge and summarize the core temperature change trends during and after an accident. This method is labor-intensive, inefficient, and lacks real-time performance. In recent years, with the development of machine learning and artificial intelligence technologies, data-driven anomaly detection methods have made significant progress and have been widely applied in industrial process fault detection tasks, achieving good results. Therefore, it is urgent to conduct nuclear reactor fault detection work based on data-driven methods, taking into account the characteristics of nuclear power core temperature data.
[0003] Data-driven anomaly detection methods do not rely on mechanistic expertise. They achieve efficient, real-time anomaly detection solely by leveraging the vast amounts of data collected during system operation and capturing potential coupling relationships between variables. Commonly used methods include multivariate statistical analysis such as Principal Component Analysis (PCA), Partial Least Squares (PLS), and Independent Component Analysis (ICA), as well as deep learning methods such as auto-encoders (AE) and convolutional neural networks (CNNs). However, complex nonlinear coupling relationships exist between different core temperature measurement points, making it difficult for multivariate statistical analysis methods to effectively capture the nonlinear characteristics of the data and achieve high-accuracy anomaly detection. While deep learning methods can utilize nonlinear mappings between neurons to learn the implicit nonlinear features in the data, the features extracted from the model's latent space may exhibit strong coupling relationships, resulting in information redundancy. Monitoring statistics calculated based on redundant features may not accurately describe the degree to which test data deviates from the model in the feature space, leading to missed and false alarms.
[0004] Anomaly detection methods based on decoupled representation learning overcome feature redundancy by designing decoupled constraints on latent space generation on top of traditional deep learning models, such as variational autoencoders (VAEs). However, these methods often require pre-setting and fixing the dimensions of the model's latent space before training, and the choice of latent space dimensions has a significant impact on model performance. This poses a challenge to the application of anomaly detection methods based on decoupled representation learning in practical modeling processes. Summary of the Invention
[0005] The purpose of this invention is to address the problem that traditional decoupled representation learning is difficult to pre-specify feature dimensions. It proposes an incremental feature decoupled autoencoder for core temperature anomaly detection, which generates features in the latent space of the autoencoder model in a step-by-step manner and adaptively determines the feature dimensions. This satisfies the dual requirements that the features extracted by the model have a strong ability to reconstruct data and that the latent space features of the model are fully decoupled. Furthermore, it constructs statistics based on the latent space and residual space of the core temperature data for anomaly detection.
[0006] The objective of this invention is achieved through the following technical solution:
[0007] A method for detecting core temperature anomalies using an incremental feature-decoupled self-encoder is as follows:
[0008] The real-time collected nuclear reactor core temperature data samples are input into the trained incremental feature decoupling autoencoder model to obtain feature vectors and reconstructed samples. Based on the feature vectors and reconstructed samples, statistics are calculated to detect anomalies in the core temperature.
[0009] The incremental feature decoupling autoencoder model is trained using the following method:
[0010] Construct a training set, wherein each sample in the training set is nuclear reactor core temperature data collected during normal operation of the nuclear reactor;
[0011] An incremental feature decoupling autoencoder is constructed, which consists of an input layer I, a regulation layer P, a feature layer F, and an output layer O; the initial number of neurons in the regulation layer P and the feature layer F is set.
[0012] The training set samples are input into an incremental feature decoupling autoencoder to obtain feature vectors and reconstructed samples. The incremental feature decoupling autoencoder undergoes incremental iterative training of neurons based on a loss function until the model performance metrics meet the requirements. The loss function includes: a first loss function, composed of the reconstruction loss; and a second loss function, composed of the sum of the reconstruction loss and the latent space decoupling loss. The model performance metrics include: a reconstruction error metric R, numerically the same as the reconstruction loss; and a latent space feature correlation metric C, numerically the same as the latent space decoupling loss.
[0013] If C < C th And R < R th It is believed that the current model's reconstruction capability and latent space decoupling degree have met the requirements, and the model training is complete.
[0014] If C < C th And R > R th It is believed that the current model has achieved the required degree of decoupling in the latent space, but its reconstruction ability is insufficient. Therefore, it is necessary to increment and train neurons in the feature layer F.
[0015] If C > C th And R < R th It is believed that the current model's reconstruction capability meets the requirements, but the degree of decoupling in the latent space is insufficient, and neuron increments and training are needed in the regulation layer P.
[0016] If C > C th And R > R th It is believed that the current model's reconstruction capability and the degree of decoupling of the latent space are not up to standard, and the requirement for the degree of decoupling of the latent space needs to be met first. Neuron increments are performed and trained in the regulation layer P.
[0017] Among them, C th R threpresents the threshold values of the latent space feature correlation index C and the reconstruction error index R, respectively.
[0018] Furthermore, after incrementing the number of neurons in the feature layer F, during training, the network parameters of the mapping between the input layer I and the regulation layer P remain unchanged. The mapping between the regulation layer P and the feature layer F is fixed, retaining the network parameters that participated in the previous training round, and updating the newly added neurons in this round to generate the new feature vector f. k The network parameters, where the subscript k represents the latent space dimension of the current model after adding a new neuron, are completely updated, and the network parameters mapping between the feature layer F and the output layer O are fully updated.
[0019] Furthermore, after incrementing the number of neurons in the regulation layer P, during training, the mapping between the input layer I and the regulation layer P is fixed, including the network parameters that participated in the previous training round, and the newly added neurons in this round are updated to generate the new vector p. j+1 The network parameters are defined as follows: the subscript j represents the number of neurons in the regulation layer P that participated in the previous training round; the mapping between the regulation layer P and the feature layer F is fixed; and the network parameters that participated in the previous training round are updated in this round based on the new vector p. j+1 The network parameters that generate new feature vectors from the matrix are completely updated, and the network parameters that map between the feature layer F and the output layer O are fully updated.
[0020] Furthermore, the latent space decoupling loss Loss C , represented as:
[0021]
[0022] Cov(f k ,f i )=E(f k ·f i )-E(f k )E(f i )
[0023] Where k is the latent space dimension of the current model, f k Let f be the k-th eigenvector output by the feature layer F. i (i = 1, 2, ..., k-1) are the k-1 feature vectors extracted in the previous rounds, and the function E calculates the mean of the feature vectors.
[0024] Furthermore, the reconstruction loss Loss R , represented as:
[0025]
[0026] Among them ||·|| F Let X represent the Frobenius norm, and X be the sample matrix of nuclear reactor core temperature data input to the model. It is the reconstructed sample matrix, where n is the number of samples.
[0027] Furthermore, the second loss function is expressed as:
[0028] Loss total =Loss R +βLoss C
[0029] Where β is the model hyperparameter, and Loss R It is the reconstruction loss, Loss C It is the latent space decoupling loss.
[0030] Furthermore, the statistics include T 2 There are two types: SPE statistics and SPE statistics.
[0031] Furthermore, anomaly detection of core temperature is performed based on eigenvectors and reconstructed samples to calculate statistics, specifically:
[0032] Based on the eigenvectors and reconstructed samples, statistics are calculated. If any of the calculated statistics exceeds the control limit, it indicates that a malfunction has occurred in the nuclear reactor operation.
[0033] Furthermore, the control limits of the statistic are calculated using a kernel density estimation method.
[0034] Furthermore, the nuclear reactor core temperature data includes core temperatures collected from multiple measuring points by sensors distributed at different locations within the reactor core.
[0035] The beneficial effects of this invention are as follows: Addressing the problem of traditional decoupled representation learning's difficulty in pre-specifying feature dimensions, this invention proposes an incremental feature-decoupled autoencoder method for core temperature anomaly detection. This invention designs a feature increment strategy, progressively generating latent space features and adaptively determining feature dimensions during autoencoder model training. Simultaneously, it proposes an iterative training strategy based on dual performance indicators for autoencoder model training, enabling the extracted features to have strong data reconstruction capabilities and satisfying the latent space feature decoupling requirements. Finally, this invention constructs monitoring statistics based on the data's feature space and residuals, achieving comprehensive anomaly detection for core temperature data. Compared to traditional deep learning-based anomaly detection methods, this invention achieves decoupling of the model's latent space features and constructs statistics based on the latent space that can more accurately capture data anomalies, reducing false positives and false negatives in fault detection. Compared to traditional decoupled representation learning-based anomaly detection methods, this invention can adaptively determine the latent space feature dimensions, effectively reducing the application difficulty of the method in actual modeling processes. Attached Figure Description
[0036] Figure 1 This is a flowchart illustrating the overall framework of the present invention.
[0037] Figure 2 The diagram shows the fault detection results of the method proposed in this invention on test set 1, where the dashed line represents the control limits and the solid line represents the statistical calculation results of the test set samples.
[0038] Figure 3 The graph shows the fault detection results of the principal component analysis method on test set 1, where the dashed line represents the control limits and the solid line represents the statistical calculation results of the test set samples.
[0039] Figure 4 The graph shows the fault detection results of the traditional autoencoder method on test set 1, where the dashed line represents the control limits and the solid line represents the statistical calculation results of the test set samples.
[0040] Figure 5 The graph shows the fault detection results of the direct decoupling autoencoder method on test set 1, where the dashed line represents the control limits and the solid line represents the statistical calculation results of the test set samples.
[0041] Figure 6 The graph shows the fault detection results of the method proposed in this invention on test set 2, where the dashed line represents the control limit and the solid line represents the statistical calculation results of the test set samples.
[0042] Figure 7 The graph shows the fault detection results of the principal component analysis method on test set 2, where the dashed line represents the control limits and the solid line represents the statistical calculation results of the test set samples.
[0043] Figure 8 The graph shows the fault detection results of the traditional autoencoder method on test set 2, where the dashed line represents the control limits and the solid line represents the statistical calculation results of the test set samples.
[0044] Figure 9 The graph shows the fault detection results of the direct decoupling autoencoder method on test set 2, where the dashed line represents the control limits and the solid line represents the statistical calculation results of the test set samples. Detailed Implementation
[0045] This embodiment uses real nuclear reactor core temperature data as an example to verify the effectiveness of the method. The invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0046] This invention discloses a method for detecting core temperature anomalies using an incremental feature-decoupled autoencoder. The method involves inputting real-time collected nuclear reactor core temperature data samples into a trained incremental feature-decoupled autoencoder model to obtain feature vectors and reconstructed samples. Based on these feature vectors and reconstructed samples, statistical measures are calculated to detect core temperature anomalies. The training of the incremental feature-decoupled autoencoder model is a key aspect of this invention. Figure 1 As shown, offline modeling training includes the following steps:
[0047] Step 1: Construct a training set, where each sample in the training set is the core temperature data of the nuclear reactor collected during normal operation of the nuclear reactor;
[0048] In this embodiment, 6000 normal samples were collected as a training set to build the model. Two test sets of 6000 samples each were taken for two different fault types occurring during the process to conduct anomaly detection tests. Each sample contained temperature variables at 40 measurement points, collected by sensors axially distributed in the top, middle, and bottom layers of the nuclear reactor core. The sampling interval was 60 seconds. The faults in the two test sets were caused by abrupt changes and gradual changes in the core temperature distribution, respectively, and these faults began to appear from the 2000th sample.
[0049] Furthermore, the collected data is standardized, as follows:
[0050] Regarding the collected normal core temperature data The standardized calculation formula is as follows:
[0051]
[0052] Where n represents the number of samples, and m represents the number of variables; x i,r This represents the element in the i-th row and r-th column of the data matrix X, that is, the value of the r-th temperature variable in the i-th sample collected; It is the mean of the r-th process variable in all samples; s r x represents the standard deviation of the r-th process variable in all samples; i,r These are the standardized values of the corresponding samples and variables. (Regarding the data...) X After standardization, a data matrix can be obtained.
[0053] Step 2: Construct an incremental feature decoupling autoencoder, model normal data X, obtain feature vectors and reconstructed samples, and perform incremental iterative training of neurons in the incremental feature decoupling autoencoder based on the loss function until the model performance metrics meet the requirements; specifically including the following sub-steps:
[0054] Step 2.1: The incremental feature decoupling autoencoder model consists of an input layer I, a regulation layer P, a feature layer F, and an output layer O. The feature layer F is defined as the model's latent space, and the number of neurons within the layer corresponds to the feature dimension of the latent space. This invention designs the feature layer F with incrementally increasing neurons. When the model's ability to reconstruct data is poor, new neurons need to be added to the feature layer F for training. Similarly, this invention designs the regulation layer P with incrementally increasing neurons. If the feature decoupling degree of the feature layer F is insufficient during training, new neurons need to be added to the regulation layer P for training.
[0055] Step 2.2: Set the number of neurons in the feature layer F and the regulation layer P of the incremental feature decoupling autoencoder model in the initial state. Generally, the initial number of neurons in the feature layer F is set to 1 to maximize adaptive incrementing. In this embodiment, the number of neurons in the feature layer F and the regulation layer P of the incremental feature decoupling autoencoder model in the initial state are set to 1 and 3, respectively. Let I represent the mappings between the input layer I and the adjustment layer P, the adjustment layer P and the feature layer F, and the feature layer F and the output layer O, respectively, when the model generates one-dimensional features. The information transfer within the model can then be summarized as: input data... All the information will first be mapped to j neurons in the regulatory layer P to obtain a matrix. Right now The information contained in P is then further compressed into a neuron in the feature layer F to obtain the feature vector. Right now Finally, feature f1 is mapped to output layer O to obtain reconstructed data. Right now Mapping The network parameters included are as follows: The first loss function trained on the model in its initial state is the same as that of a traditional autoencoder, which is the reconstruction loss. R The calculation method used in this embodiment is as follows:
[0056]
[0057] Among them ||·|| F This represents the Frobenius norm.
[0058] Step 2.3: Let the reconstruction error index R = Loss R And set the corresponding threshold R th If the model is trained under the initial conditions and has R > R, then... th Then, a new neuron is added to the model's feature layer F and retrained to generate a second-dimensional feature vector f2. Then, f1 and f2 are combined and mapped to the output layer O. This represents 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. Let the network parameters included in the above mappings be denoted as follows: In this training, for the mapping The model has its parameters fixed. Parameters compared to the previous round Same and not updated; for mappings Its parameters It consists of two parts of parameters that need to be fixed and updated: the part that needs to be fixed is the same as the parameters from the previous round. The parameters that are the same and not updated, and the parameters that need to be updated, are the new parameters added in this round to generate the new mapping for f2, denoted as... Right now For mapping The model sets its parameters Perform a full update, which means discarding the parameters from the previous round. Retrain. During this training process to generate f2, the model's loss function is adjusted based on the original reconstruction loss. R Based on this, a latent space decoupling loss term, Loss, was added. C It can be in the following form:
[0059]
[0060] Cov(f k ,f i )=E(f k ·f i )-E(f k )E(f i )
[0061] Where k is the latent space dimension of the current model, f k Let f be the k-th eigenvector. i (i = 1, 2, ..., k-1) represent the k-1 feature vectors extracted in previous rounds, and function E calculates the mean of these feature vectors. This second loss function, Loss... total Loss from reconstruction R Loss on decoupling from latent space C The sum is composed of:
[0062] Loss total =Loss R +βLoss C
[0063] Here, β is a model hyperparameter. The model will continue to use this loss function during subsequent training.
[0064] Step 2.4: Let the correlation index of latent space features be C = Loss C And set the corresponding threshold C th If the model has C > C after the previous training round. th Then, neuron increments are performed on the regulation layer P and the model mapping is retrained. That is: the newly generated vector p in the adjustment layer P j+1 Combined with matrix P, we obtain matrix This is then mapped back to the feature layer F to regenerate the second-dimensional feature f2. In this training process, for the mapping... Its parameters It consists of two parts of parameters that need to be fixed and those that need to be updated: the part that needs to be fixed and the parameter. The parts that are the same and not updated are those added in this round and used to generate p. j+1 The parameters of the new mapping are denoted as Right now For mapping Its parameters It consists of two parts of parameters that need to be fixed and those that need to be updated: the part that needs to be fixed and the parameter. The parameters are the same and will not be updated; the parts that need to be updated are the parameters from the previous round. An update is performed to regenerate f2, i.e. For mapping The model sets its parameters Perform a full update.
[0065] Step 2.5: Based on the specific operations of neuron increment in Steps 2.3 and 2.4, and so on, perform incremental iterative training of neurons on the incremental feature decoupled autoencoder based on the loss function until the model performance metrics meet the requirements. The complete model iterative training strategy is summarized as follows:
[0066] Condition 1: If C < C th And R < R th It is believed that the current model's reconstruction capability and latent space decoupling degree have met the requirements, and the model training is complete.
[0067] Condition 2: If C < C th And R > R th It is believed that the current model has achieved the required degree of decoupling in the latent space, but its reconstruction capability is insufficient. Therefore, it is necessary to increment and train neurons in the feature layer F.
[0068] Condition 3: If C > C th And R < R th It is believed that the current model has sufficient reconstruction capability, but the degree of decoupling in the latent space is insufficient, and neuron increments and training are needed in the regulation layer P.
[0069] Condition 4: If C > C th And R > R th It is believed that the current model reconstruction capability and the degree of decoupling of the latent space are not up to standard, and the requirement for the degree of decoupling of the latent space needs to be prioritized. The operation is the same as that of condition 3.
[0070] Among them, C th R th represents the threshold values for the latent space feature correlation index C and the reconstruction error index R, respectively. In this embodiment, they are 0.8 and 0.1, respectively.
[0071] In this embodiment, the incremental feature decoupling autoencoder model is trained according to the above strategy until the iteration termination condition is met. At this point, the number of neurons in the feature layer F and the regulation layer P are 6 and 9, respectively. The model structure is fixed, and the model mapping obtained in the last round of training is denoted as... Model training is complete.
[0072] Statistics are designed for anomaly detection based on the characteristic space and residual space of normal core temperature data. In this embodiment, T is used as the statistical measure. 2 Taking SPE as an example, the calculation method is as follows: for normal data Each of these samples can be represented as The feature vector is extracted using an incremental feature decoupling autoencoder model. Simultaneously obtain the reconstructed output And calculate the reconstructed residual term T is constructed based on the above vectors. 2 The SPE statistics are as follows:
[0073] T 2 =f t T Σ -1 f t
[0074] SPE = e T e
[0075] Where Σ is the covariance matrix of features extracted from normal data by the model. T is calculated using kernel density estimation. 2 Control limit Ctr of SPE statistic T2 and Ctr SPE .
[0076] Thus, two statistics and their control limits have been obtained. Fault detection can be achieved using the trained incremental feature-decoupled autoencoder model. Specifically, the real-time collected nuclear reactor core temperature data samples... Fault detection can be achieved by inputting the data into a trained incremental feature decoupling autoencoder model. The trained incremental feature decoupling autoencoder model is then used to extract its feature vectors. Simultaneously obtain the reconstructed output Calculate the reconstructed residual term Calculate T 2 Using SPE statistics to detect anomalies in core temperature:
[0077] T 2 =f new T Σ -1 f new
[0078] SPE = e new T e new
[0079] If T 2 If any statistic in SPE exceeds the control limit, it indicates a malfunction in the nuclear reactor operation. The overall framework flowchart of this invention is as follows: Figure 1 As shown below, fault detection analysis is performed using samples from two test sets to illustrate the effectiveness of this invention.
[0080] In terms of fault detection, this invention performs fault detection on test sets for two fault types: sudden temperature changes and gradual temperature changes. The results are as follows: Figure 2 and Figure 6 As shown, under abrupt fault conditions, both statistics of this method significantly exceeded the control limits at the time of fault occurrence, demonstrating timely and sensitive fault detection. Under slowly changing fault conditions, the two statistics of this method were also able to capture the slowly changing anomalies in the data, detecting fault occurrence with relatively high accuracy.
[0081] To more clearly demonstrate the superiority of the feature increment strategy and iterative training strategy designed in this invention in anomaly detection tasks, the model obtained by discarding all the incremental and iterative training strategies of the incremental feature decoupling autoencoder in this invention is named the direct decoupling autoencoder (using only the second loss function for constrained training) for experimental comparison. Principal Component Analysis (PCA) (W. Svante, K. Esbensen, and P. Gelabi. "Principal component 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 the direct decoupling autoencoder are compared. The fault detection performance of the above methods on the two test sets is as follows: Figure 3 and Figure 7 , Figure 4 and Figure 8 , Figure 5 and Figure 9 As shown, in experiments involving sudden-change faults, the above methods were generally able to detect the fault occurrence relatively accurately. However, in experiments involving slowly changing faults, the above methods generated a large number of false alarms before the fault occurred, and the fault detection time was also significantly delayed, resulting in a low fault detection rate.
[0082] For a more intuitive comparison, Tables 1 and 2 list the Fault Detection Rate (FDR) and Fault False Alarm Rate (FAR) of different methods on the test set, respectively. The formulas for calculating FDR and FAR are as follows:
[0083]
[0084]
[0085] Where N MAE N abnormal N FAE N normal These represent the number of missed events, anomalous samples, false alarm events, and normal samples, respectively.
[0086] It can be observed that the false alarm rate of the method proposed in this invention is the lowest among the four methods on both test sets, and the fault detection rate is also the highest among the four methods overall. Among them, the superiority of the method proposed in this invention compared with other methods is most significant on test set 2 containing slowly changing fault types. This proves the feasibility and effectiveness of the method proposed in this invention.
[0087] Table 1 Comparison of Fault Detection Rate (FDR) of Different Methods (%)
[0088]
[0089] Table 2 Comparison of Fault False Alarm Rates (FAR) of Different Methods (%)
[0090]
[0091] Obviously, the above embodiments are merely illustrative examples for clear explanation and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations here. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.
Claims
1. A method for detecting core temperature anomalies using an incremental feature-decoupled self-encoder, characterized in that, Specifically: The real-time collected nuclear reactor core temperature data samples are input into the trained incremental feature decoupling autoencoder model to obtain feature vectors and reconstructed samples. Based on the feature vectors and reconstructed samples, statistics are calculated to detect anomalies in the core temperature. The incremental feature decoupling autoencoder model is trained using the following method: Construct a training set, wherein each sample in the training set is nuclear reactor core temperature data collected during normal operation of the nuclear reactor; An incremental feature decoupling autoencoder is constructed, which consists of an input layer I, a regulation layer P, a feature layer F, and an output layer O; the initial number of neurons in the regulation layer P and the feature layer F is set. The training set samples are input into an incremental feature decoupling autoencoder to obtain feature vectors and reconstructed samples. The incremental feature decoupling autoencoder undergoes incremental iterative training of neurons based on a loss function until the model performance metrics meet the requirements. The loss function includes: a first loss function, composed of the reconstruction loss; and a second loss function, composed of the sum of the reconstruction loss and the latent space decoupling loss. The model performance metrics include: a reconstruction error metric R, numerically the same as the reconstruction loss; and a latent space feature correlation metric C, numerically the same as the latent space decoupling loss. If C < C th And R < R th It is believed that the current model's reconstruction capability and latent space decoupling degree have met the requirements, and the model training is complete. If C < C th And R > R th It is believed that the current model has achieved the required degree of decoupling in the latent space, but its reconstruction ability is insufficient. Therefore, it is necessary to increment and train neurons in the feature layer F. If C > C th And R < R th It is believed that the current model's reconstruction capability meets the requirements, but the degree of decoupling in the latent space is insufficient, and neuron increments and training are needed in the regulation layer P. If C > C th And R > R th It is believed that the current model's reconstruction capability and the degree of decoupling of the latent space are not up to standard, and the requirement for the degree of decoupling of the latent space needs to be met first. Neuron increments are performed and trained in the regulation layer P. Among them, C th R th represents the threshold values of the latent space feature correlation index C and the reconstruction error index R, respectively.
2. The method according to claim 1, characterized in that, After incrementing the number of neurons in the feature layer F, during training, the network parameters mapping between the input layer I and the regulation layer P remain unchanged. The mapping between the regulation layer P and the feature layer F is fixed, including the network parameters that participated in the previous training round, and the newly added neurons in this round are updated to generate the new feature vector f. k The network parameters, where the subscript k represents the latent space dimension of the current model after adding a new neuron, are completely updated, and the network parameters mapping between the feature layer F and the output layer O are fully updated.
3. The method according to claim 1, characterized in that, After incrementing neurons in the regulation layer P, during training, the mapping between the input layer I and the regulation layer P is fixed, including the network parameters that participated in the previous training round, and the newly added neurons in the current round are updated to generate the new vector p. j+1 The network parameters are defined as follows: the subscript j represents the number of neurons in the regulation layer P that participated in the previous training round; the mapping between the regulation layer P and the feature layer F is fixed; and the network parameters that participated in the previous training round are updated in this round based on the new vector p. j+1 The network parameters that generate new feature vectors from the matrix are completely updated, and the network parameters that map between the feature layer F and the output layer O are fully updated.
4. The method according to claim 1, characterized in that, The hidden space decoupling loss Loss C , represented as: The (f k ,f i )=E(f k ·f i )-E(f k )E(f i ) Where k is the latent space dimension of the current model, f k Let f be the k-th eigenvector output by the feature layer F. i (i = 1, 2, ..., k-1) are the k-1 feature vectors extracted in the previous rounds, and the function E calculates the mean of the feature vectors.
5. The method according to claim 1, characterized in that, The reconstruction loss Loss R , represented as: Among them ||·|| F Let X represent the Frobenius norm, and X be the sample matrix of nuclear reactor core temperature data input to the model. It is the reconstructed sample matrix, where n is the number of samples.
6. The method according to claim 1, characterized in that, The second loss function is expressed as: Loss total =Loss R +βLoss C Where β is the model hyperparameter, and Loss R It is the reconstruction loss, Loss C It is the latent space decoupling loss.
7. The method according to claim 1, characterized in that, Statistics include T 2 There are two types: SPE statistics and SPE statistics.
8. The method according to claim 1, characterized in that, Anomaly detection of core temperature is performed based on eigenvectors and reconstructed sample statistics. Based on the eigenvectors and reconstructed samples, statistics are calculated. If any of the calculated statistics exceeds the control limit, it indicates that a malfunction has occurred in the nuclear reactor operation.
9. The method according to claim 8, characterized in that, The control limits of the statistic are calculated using kernel density estimation.
10. The method according to claim 1, characterized in that, The nuclear reactor core temperature data includes core temperatures collected from multiple measuring points by sensors located at different locations within the reactor core.