Few-sample equipment state prediction method based on transfer learning

Abstract

The invention relates to the field of industrial equipment state monitoring, in particular to a few-sample equipment state prediction method based on transfer learning, which comprises the following steps of: simulating multiple operation state modes of equipment by utilizing PCTran, acquiring simulation data of operation parameters of the equipment in different states, extracting a state mode feature representation mechanism based on time-frequency domain features, and predicting the state of the equipment according to the state mode feature representation mechanism; designing a state prediction model by combining transfer learning and random forest; adopting a nearest class mean classifier as a splitting function of the random forest, obtaining an information gain of each splitting function, and determining an optimal splitting function according to the maximum information gain; and obtaining feature importance and tree structure knowledge of the random forest, dynamically evaluating feature mobility through KL divergence, and migrating the importance feature of the random forest to a target domain. The method has the advantages of model stability and decision interpretability while keeping high classification precision under the condition of few samples, and is particularly suitable for industrial scenes such as nuclear power equipment with extremely high requirements on safety and reliability.

Description

Technical Field

[0001] The present invention relates to the field of industrial equipment status monitoring, and in particular to a few-sample equipment status prediction method based on transfer learning. Background Art

[0002] Condition monitoring and fault prediction technologies are of great significance in ensuring the safe operation of key equipment in high-risk areas such as nuclear power. The coolant system of a pressurized water reactor nuclear power plant is the core unit for maintaining heat extraction and pressure control in the core. If its minor faults are not predicted in time, it may trigger a chain reaction of safety incidents. The traditional regular maintenance model has problems with response lag and resource waste, while predictive maintenance technology based on real-time data analysis can significantly improve fault warning capabilities and reduce losses from unplanned downtime. International practice has shown that advanced fault classification methods can reduce the operation and maintenance costs of nuclear power plants by nearly 30%, while improving the availability of units. However, due to the extremely high cost of fault testing and strict safety constraints of nuclear power equipment, fault samples that can be used for modeling are extremely scarce. The cost of a single start-up and shutdown experiment often reaches tens of millions, which poses a great challenge to the accuracy of condition prediction.

[0003] Current equipment status prediction methods primarily fall into three categories: physical model-based, traditional machine learning, and deep learning. Physical model-based methods rely on precise thermal-hydraulic equations, but the strong nonlinearity and multi-physics coupling of nuclear power systems limit model generalization and increase construction costs. Traditional machine learning methods, such as support vector machines and decision trees, while highly interpretable, require extensive manual feature engineering and are sensitive to data balance, leading to severe overfitting in scenarios with scarce samples. Deep learning methods, while capable of automatically extracting time series features, face significant challenges with their massive data requirements and the varying design parameters across nuclear power units. Cross-domain data offsets caused by differences in the distribution of parameters such as coolant flow and temperature profiles across units often significantly reduce the accuracy of directly migrated models.

[0004] To address the limitations of traditional methods for predicting the status of equipment with few samples, researchers have begun exploring methods such as transfer learning to improve the performance of prediction with few samples. Transfer learning, through a knowledge transfer mechanism, can effectively utilize historical data from other units or simulators, significantly alleviating the dilemma of insufficient samples for the target unit. Its unique feature importance analysis capability can identify key fault characteristics and provide a traceable decision-making basis for diagnostic results, which is particularly important in the nuclear power sector, where safety is paramount. In addition, to ensure the robustness of the status prediction of highly safe and customized equipment, ensemble learning methods, through a multi-model collaborative decision-making mechanism, not only improve the generalization ability of the model, but also enhance the system's robustness to noisy data and missing features. Compared with deep learning methods that require massive amounts of data, this type of method can still maintain stable classification performance when the sample size is limited, while avoiding the overfitting problem that is prone to occur in traditional machine learning methods. Summary of the Invention

[0005] In order to solve the above technical problems existing in the prior art, the present invention proposes a small-sample equipment status prediction method based on transfer learning, and its specific technical solution is as follows: A few-sample equipment status prediction method based on transfer learning, comprising: Step 1: Use the reactor transient and accident simulation software PCTran to simulate various operating modes of pressurized water reactor nuclear power equipment and obtain simulation data of operating parameters under different states; Step 2: extracting time domain features from the simulation data; Step 3: Use the time domain feature data extracted in step 2 to train the source domain random forest model. Use the nearest class mean classifier as the splitting function of the random forest. Obtain the information gain of each splitting function, and determine the optimal classification function based on the maximum information gain. Step 4: Determine the final classification result by maximizing the average class probability of all subtrees in the source domain random forest model; Step 5: Obtain feature importance and tree structure knowledge of the source domain random forest model; Step 6: Calculate the KL divergence between the source domain data and the target domain data, dynamically evaluate the feature transferability through the KL divergence, and automatically filter out mismatched features; Step 7: Based on the feature transferability evaluation results of step 6, transfer the feature importance of the source domain random forest, weight the target domain data, and train the target domain random forest model.

[0006] Furthermore, the step 2 is specifically as follows: taking time series segments of fixed time steps from the simulation data, and calculating the statistical characteristics of each segment of the operating parameter sequence.

[0007] Furthermore, the statistical characteristics include mean, standard deviation, skewness, and kurtosis.

[0008] Furthermore, in step 3, the time domain feature data obtained in step 2 is constructed as a sample and used as an instance pool for training the source domain random forest model; the source domain random forest model is composed of a set of random subtrees, and classification learning is achieved by training subtree bifurcation nodes. Each subtree randomly extracts the same number of samples from the instance pool, and a set of split functions is randomly generated at each subtree bifurcation node. A recursive training process is adopted to determine the optimal splitting function.

[0009] Furthermore, at each subtree bifurcation node, the nearest class mean classifier generates the centroid of each class, where the class closest to the centroid is the class of the sample.

[0010] Furthermore, the information gain of each splitting function is obtained by calculating the entropy of the category.

[0011] Furthermore, in step 4, the prediction results of the nearest class mean classifier are aggregated through the Bagging ensemble learning method, and the final classification result is determined by maximizing the average class probability of all subtrees.

[0012] Furthermore, in step 7, the source domain knowledge and category balance are fused through exponential weighting when weighting the target domain data, and then the target domain random forest model is trained using the weighted weights.

[0013] The beneficial effects of the present invention are: (1) This paper fully utilizes the historical data knowledge of the source domain (other units / simulators) through the fusion architecture of transfer learning and random forest, effectively alleviates the sample scarcity problem of the target unit caused by experimental cost constraints, and improves the state prediction ability of the target equipment under the condition of sample scarcity.

[0014] (2) By introducing a feature importance weighting mechanism, the present invention can automatically identify and enhance key state features that are universal across units, significantly reducing the impact of data distribution offset caused by differences in design parameters. The integration characteristics of random forests further enhance the model's robustness to noisy data and missing features.

[0015] (3) The present invention utilizes the interpretability advantage of the random forest algorithm to provide a transparent decision-making basis for the prediction process. The feature importance ranking generated by the random forest provides an intuitive and traceable technical basis for the prediction results, enabling operation and maintenance personnel to clearly understand the model judgment logic. This feature not only meets the regulatory requirements for the "black box" model of safety-critical systems in the nuclear power field, but also helps engineers quickly locate the root cause of the fault.

[0016] (4) The present invention effectively reduces the reliance of traditional fault diagnosis methods on manual feature engineering and domain expert experience through an automated feature migration and optimization mechanism, making the promotion and application of new technologies in different industrial scenarios more convenient and efficient, and creating favorable conditions for the popularization and application of equipment condition monitoring technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 1 is a flow chart of a method for predicting equipment status using a small number of samples based on transfer learning according to an embodiment of the present invention; Figure 2 is a state prediction model parameter setting diagram of an embodiment of the present invention; Figure 3 This is a schematic diagram showing the performance of the state prediction model according to an embodiment of the present invention; Figure 4 This is a test result diagram of the trained state prediction model according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and technical effect of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0019] like Figure 1 As shown, this embodiment proposes a few-sample equipment status prediction method based on transfer learning. By integrating transfer learning with random forest design state prediction models, it can automatically identify and enhance the portability of key fault features between different units. It can also provide a traceable decision basis for fault diagnosis through feature importance ranking provided by random forest. Specifically, the following steps are included: Step 1: Use the reactor transient and accident simulation software PCTran to simulate various operating state modes such as normal operation control of pressurized water reactor nuclear power equipment, coolant loss accident, steam pipeline rupture, etc., obtain simulation data of operating parameters such as liquid level and temperature under different events, and obtain the time series signal set ,in T is the number of data points, x t For the t Sensor data at a certain point in time.

[0020] Step 2: Extract time domain features from the data.

[0021] Set the fixed time step to N , N < T , reorganize the parameter sequence into T - N +1 time series segment with fixed time step , calculate 16 kinds of statistical characteristics of each running parameter sequence, including mean, standard deviation, skewness, kurtosis, etc., to form T - N +1 sample.

[0022] Step 3: Train the source domain random forest model, using the nearest class mean classifier (NCMC) as the splitting function of the random forest, obtain the information gain of each splitting function, and determine the optimal classification function based on the maximum information gain.

[0023] Specifically, the time domain feature data obtained in step 2 is constructed as a sample and used as an instance pool for training the source domain random forest model.

[0024] The source domain random forest model consists of a set of random subtrees KClassification learning is achieved by training subtree bifurcation nodes. Each subtree randomly extracts the same number of samples from the instance pool and randomly generates a set of split functions at each subtree bifurcation node. , a recursive training process is used to determine the best splitting function f j This step introduces random selection of features and samples, ensuring that the generated subtrees have low correlation.

[0025] NCMC is used as the splitting function of random forest, and at each node j At , NCMC generates the centroid of each class , where the class closest to the centroid is the class of the sample, and the decision rule is: , in i is the pattern category, I is the total number of categories.

[0026] By using NCMC as the splitting function of random forest, only a small number of classes will be used instead of all classes at any specific node, thereby speeding up the calculation process to obtain a weak classifier; by assigning classes to either side, the multi-class output is converted into a binary output: left subclass and right subclass, which can improve the accuracy of the classifier.

[0027] By calculating the entropy of the category H , thereby obtaining the information gain of each splitting function G ( f ), the optimal splitting function is determined based on the maximum information gain.

[0028] Step 4: Determine the final classification result by maximizing the average class probability of all subtrees in the source domain random forest model.

[0029] Training continues recursively until a predefined stopping condition is reached and According to the definition of integrated learning Bagging, the final classification result is determined by maximizing the average class probability of all subtrees, as expressed as follows: , in K is the total number of trees.

[0030] The Bagging method is used to combine multiple decision trees to achieve better generalization ability than a single classifier.

[0031] Step 5: Obtain feature importance and tree structure knowledge of the source domain random forest model.

[0032] Feature importance weight matrix generated by the source domain random forest model W , guide the target domain feature screening and improve the contribution of key features.

[0033] Step 6: Calculate the KL divergence between the source domain data and the target domain data. Use the KL divergence to dynamically evaluate feature transferability and automatically filter out mismatched features, such as invalid features caused by device model differences. The discrete feature formula is: , in is the eigenvalue in the source domain The probability of occurrence, In the target domain The probability of occurrence, Δ is the number of discretized bins of feature values, Used to control migration intensity.

[0034] when hour, α =0, indicating that the feature migration is disabled; hour, α =1, indicating direct migration; hour, α =0.7, indicating attenuated migration.

[0035] Step 7: Transfer the feature importance of the source domain random forest, weight the target domain data, train the target domain random forest model, and obtain the state prediction model.

[0036] Combined with the transferability evaluation results in step 6, the feature importance of the random forest obtained in step 5 is transferred. The feature importance transfer is defined as: , in It is a feature f j In the source domain i The importance of class patterns, D is the feature dimension.

[0037] In one embodiment, the target domain data category weight is calculated based on the coolant system data of a specific three-circuit pressurized water reactor nuclear power plant. , and then the target domain data is weighted, where the source domain knowledge is fused through exponential weighting α Item and Category Balance 1- α Item, alleviates data distribution deviation, the specific expression is as follows: , The obtained weights are then used to train the target domain random forest model to obtain the state prediction model of this embodiment.

[0038] Step 8: Embed the trained state prediction model into the real-time monitoring system. Based on the model output, when the probability of a certain category exceeds the set threshold, the system issues an alarm signal to remind the operator to perform maintenance or repair.

[0039] refer to Figures 2 to 4 The method of this embodiment can maintain high classification accuracy under the condition of few samples while having the dual advantages of model stability and decision interpretability. It is particularly suitable for industrial scenarios such as nuclear power equipment that have extremely high requirements for safety and reliability.

[0040] The above description is only a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the implementation process of the present invention is described in detail above, it is still possible for those familiar with the art to modify the technical solutions described in the above examples or to replace some of the technical features therein with equivalents. Any modifications, equivalent replacements, etc. made within the spirit and principles of the present invention shall be included in the scope of protection of the present invention.

Claims

1. A few-sample equipment status prediction method based on transfer learning, characterized in that: include: Step 1: Use the reactor transient and accident simulation software PCTran to simulate various operating modes of pressurized water reactor nuclear power equipment and obtain simulation data of operating parameters under different states; Step 2: extracting time domain features from the simulation data; Step 3: Use the time domain feature data extracted in step 2 to train the source domain random forest model. Use the nearest class mean classifier as the splitting function of the random forest. Obtain the information gain of each splitting function, and determine the optimal classification function based on the maximum information gain. Step 4: Determine the final classification result by maximizing the average class probability of all subtrees in the source domain random forest model; Step 5: Obtain feature importance and tree structure knowledge of the source domain random forest model; Step 6: Calculate the KL divergence between the source domain data and the target domain data, dynamically evaluate the feature transferability through the KL divergence, and automatically filter out mismatched features; Step 7: Based on the feature transferability evaluation results of step 6, transfer the feature importance of the source domain random forest to the target domain, weight the target domain data, and train the target domain random forest model.

2. The method for predicting equipment status with a small number of samples based on transfer learning according to claim 1, wherein: The step 2 specifically includes: taking time series segments of fixed time steps from the simulation data, and calculating the statistical characteristics of each segment of the operating parameter sequence.

3. The method for predicting equipment status with a small number of samples based on transfer learning according to claim 2, wherein: The statistical characteristics include mean, standard deviation, skewness, and kurtosis.

4. The method for predicting equipment status with a small number of samples based on transfer learning according to claim 1, wherein: In step 3, the time domain feature data obtained in step 2 is constructed as a sample and used as an instance pool for training the source domain random forest model; The source domain random forest model consists of a set of random subtrees, and classification learning is achieved by training subtree bifurcation nodes. Each subtree randomly extracts the same number of samples from the instance pool, and a set of split functions is randomly generated at each subtree bifurcation node. A recursive training process is adopted to determine the optimal split function.

5. The method for predicting equipment status with a small number of samples based on transfer learning according to claim 4, characterized in that: At each subtree bifurcation node, the nearest class mean classifier generates the centroid of each class, where the class closest to the centroid is the category of the sample.

6. The method for predicting equipment status with a small number of samples based on transfer learning according to claim 5, characterized in that: The information gain of each split function is obtained by calculating the entropy of the category.

7. The method for predicting equipment status with a small number of samples based on transfer learning according to claim 1, wherein: In step 4, the prediction results of the nearest class mean classifier are aggregated through the Bagging ensemble learning method, and the final classification result is determined by maximizing the average class probability of all subtrees.

8. The method for predicting equipment status with a small number of samples based on transfer learning according to claim 1, wherein: In step 7, the target domain data is weighted by fusing the source domain knowledge and category balance through exponential weighting, and then the target domain random forest model is trained using the weighted weights.

Images