A few-shot equipment state prediction method based on transfer learning

Abstract

The application 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: PCTran is used to simulate various running state modes of equipment, simulation data of running parameters of the equipment in different states are obtained, then a state mode feature representation mechanism is extracted based on time-frequency domain features, a state prediction model is designed by combining transfer learning and a random forest, a nearest class mean classifier is used as a splitting function of the random forest, information gain of each splitting function is obtained, an optimal splitting function is determined according to maximum information gain, feature importance and tree structure knowledge of the random forest are obtained, feature transferability is dynamically evaluated through KL divergence, and important features of the random forest are transferred to a target domain. The application can keep high classification precision under the condition of few samples, has the dual advantages of model stability and decision explainability, and is particularly suitable for nuclear power equipment and other industrial scenes with extremely high requirements for safety and reliability.

Description

Technical Field

[0001] This invention relates to the field of industrial equipment condition monitoring, specifically a few-sample equipment condition prediction method based on transfer learning. Background Technology

[0002] Condition monitoring and fault prediction technologies are of great significance for ensuring the safe operation of critical equipment in high-risk fields such as nuclear power. As the core unit for maintaining core heat removal and pressure control, the coolant system of a pressurized water reactor nuclear power plant is highly susceptible to cascading safety events if even minor faults are not predicted in time. Traditional periodic maintenance methods suffer from response delays and resource waste, while predictive maintenance technologies based on real-time data analysis can significantly improve fault early warning capabilities and reduce losses from unplanned downtime. International practice shows that advanced fault classification methods can reduce nuclear power plant operation and maintenance costs by nearly 30% while improving unit availability. However, due to the extremely high cost of fault testing and strict safety constraints of nuclear power equipment, fault samples suitable for modeling are extremely scarce. A single start-up and shutdown test often costs tens of millions of dollars, posing a significant challenge to the accuracy of condition prediction.

[0003] Current equipment condition prediction methods mainly fall into three categories: physical model-based, traditional machine learning, and deep learning. Physical model methods rely on accurate thermo-hydraulic equations, but the strong nonlinearity and multi-physics coupling of nuclear power systems limit the generalization ability of these models and result in high construction costs. Traditional machine learning methods, such as support vector machines and decision trees, while possessing strong interpretability, require extensive manual feature engineering and are sensitive to data balance, easily leading to severe overfitting in scenarios with scarce samples. Deep learning methods can automatically extract time-series features, but their massive data requirements create a sharp contradiction with the differences in design parameters among nuclear power units. Cross-domain data shifts caused by differences in the distribution of parameters such as coolant flow and temperature profiles among different units often significantly reduce the accuracy of directly transferred models.

[0004] To address the limitations of traditional methods in predicting equipment status with limited sample sizes, researchers have begun exploring methods such as transfer learning to improve performance. Transfer learning, through its knowledge transfer mechanism, can effectively utilize historical data from other units or simulators, significantly alleviating the challenge of insufficient samples for the target unit. Its unique feature importance analysis capability can identify key fault characteristics, providing traceable decision-making basis for diagnostic results, which is particularly important in the safety-critical nuclear power field. Furthermore, to ensure the robustness of highly customized equipment status prediction, ensemble learning methods, through multi-model collaborative decision-making mechanisms, not only improve the model's generalization ability but also enhance the system's robustness to noisy data and missing features. Compared to deep learning methods that require massive amounts of data, these methods can maintain stable classification performance even with limited sample sizes, while avoiding the overfitting problem that is common in traditional machine learning methods. Summary of the Invention

[0005] To address the aforementioned technical problems in the existing technology, this invention proposes a few-sample equipment state prediction method based on transfer learning, the specific technical solution of which is as follows:

[0006] A few-shot equipment state prediction method based on transfer learning, comprising:

[0007] Step 1: Use the reactor transient and accident simulation software PCTran to simulate various operating states of pressurized water reactor nuclear power equipment and obtain simulation data of operating parameters under different states;

[0008] Step 2: Extract time-domain features from the simulated data;

[0009] Step 3: Train the source domain random forest model using the time-domain feature data extracted in Step 2. Use the nearest-neighbor 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.

[0010] Step 4: Determine the final classification result by maximizing the average class probability of all subtrees in the source domain random forest model;

[0011] Step 5: Obtain feature importance and tree structure knowledge of the source domain random forest model;

[0012] Step 6: Calculate the KL divergence between the source domain data and the target domain data, dynamically evaluate feature transferability through KL divergence, and automatically filter out mismatched features;

[0013] Step 7: Based on the feature transferability evaluation results in 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.

[0014] Furthermore, step 2 specifically involves taking time-series segments of the simulation data with a fixed time step and calculating the statistical characteristics of each segment of the running parameter sequence.

[0015] Furthermore, the statistical characteristics include mean, standard deviation, skewness, and kurtosis.

[0016] Furthermore, in step 3, the temporal feature data obtained in step 2 is used to construct samples and serve 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. Classification learning is achieved by training the subtree branch nodes. Each subtree randomly draws the same number of samples from the instance pool. A set of splitting functions is randomly generated at each subtree branch node. A recursive training process is adopted to determine the optimal splitting function.

[0017] Furthermore, at each subtree branch node, the nearest-neighbor classifier generates the centroid of each class, where the class closest to the centroid is the class of the sample.

[0018] Furthermore, the information gain of each splitting function is obtained by calculating the entropy of the categories.

[0019] Furthermore, in step 4, the prediction results of the nearest class mean classifier are aggregated by the Bagging ensemble learning method, and the final classification result is determined by maximizing the average class probability of all subtrees.

[0020] Furthermore, in step 7, when weighting the target domain data, source domain knowledge and category balance are fused through exponential weighting, and then the target domain random forest model is trained using the weighted weights.

[0021] The beneficial effects of this invention are:

[0022] (1) This invention utilizes the fusion architecture of transfer learning and random forest to make full use of the historical data knowledge of the source domain (other units / simulators), effectively alleviate the problem of sample scarcity caused by experimental cost limitations of the target unit, and improve the state prediction capability of the target equipment under sample scarcity conditions.

[0023] (2) By introducing a feature importance weighting mechanism, this invention can automatically identify and strengthen key state features that are universal across units, significantly reduce the impact of data distribution shift caused by differences in design parameters, and the integration characteristics of random forest further improve the robustness of the model to noisy data and missing features.

[0024] (3) This invention utilizes the interpretability advantage of the random forest algorithm to provide a transparent decision basis for the prediction process. The ranking of the importance of features 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's judgment logic. This feature not only meets the regulatory requirements of the nuclear power field for the "black box" model of safety-critical systems, but also helps engineers quickly locate the root cause of the fault.

[0025] (4) This invention effectively reduces the reliance of traditional fault diagnosis methods on manual feature engineering and domain expert experience through an automated feature transfer 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. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the few-sample equipment state prediction method based on transfer learning according to an embodiment of the present invention.

[0027] Figure 2This is a parameter setting diagram of the state prediction model according to an embodiment of the present invention;

[0028] Figure 3 This is a schematic representation of the performance of the state prediction model in an embodiment of the present invention;

[0029] Figure 4 This is a test result diagram of the trained state prediction model according to an embodiment of the present invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and technical effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0031] like Figure 1 As shown, this embodiment proposes a few-sample equipment status prediction method based on transfer learning. By integrating transfer learning and random forest to design a status prediction model, it can automatically identify and enhance the transferability of key fault features across different units. Furthermore, it can provide traceable decision-making basis for fault diagnosis through the feature importance ranking provided by random forest. Specifically, it includes the following steps:

[0032] Step 1: Using the reactor transient and accident simulation software PCTran, simulate various operating conditions of pressurized water reactor nuclear power equipment, including normal operation control, coolant loss accidents, and steam pipeline ruptures. Obtain simulation data of operating parameters such as liquid level and temperature under different events to obtain a time-series signal set. ,in T The number of data points, x t For the first t Sensor data at various time points.

[0033] Step 2: Extract time-domain features from the data.

[0034] Set a fixed time step as N , N < T Reorganize the parameter sequence into T - N +1 time sequence segment with a fixed time step The system calculates 16 statistical characteristics for each sequence of operating parameters, including mean, standard deviation, skewness, and kurtosis, to form... T - N +1 sample.

[0035] Step 3: Train the source domain random forest model, using the Nearest Class Mean Classifiers (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.

[0036] Specifically, the temporal feature data obtained in step 2 is used to construct samples and serve as an instance pool for training the source domain random forest model.

[0037] The source domain random forest model consists of random subtree sets. K The system employs a classification learning approach by training subtree branch nodes. Each subtree randomly draws the same number of samples from the instance pool, and a set of splitting functions is randomly generated at each subtree branch node. A recursive training process is used to determine the optimal splitting function. f j This step introduces random selection of features and samples to ensure low correlation between the generated subtrees.

[0038] Using NCMC as the splitting function for the random forest, in each node j At this point, NCMC generates the centroid of each class. The class of the sample is the nearest centroid, and the decision rule is:

[0039] ,

[0040] in i It is a pattern category. I It represents the total number of categories.

[0041] Using NCMC as the splitting function of the random forest, only a small subset of classes are used at any given node instead of all classes, thus speeding up the computation process to obtain a weak classifier; by assigning classes to either side, the multi-class output is converted into binary output: left subclass and right subclass, which can improve the accuracy of the classifier.

[0042] By calculating the entropy of the category H This allows us to obtain the information gain of each splitting function. G ( f The optimal splitting function is determined based on the maximum information gain.

[0043] Step 4: Determine the final classification result by maximizing the average class probability of all subtrees in the source domain random forest model.

[0044] Training continues recursively until a predefined stopping condition is met. and According to the definition of ensemble learning (Bagging), the final classification result is determined by maximizing the average class probability of all subtrees, as shown in the following expression:

[0045] ,

[0046] in K The total number of trees.

[0047] The Bagging method combines multiple decision trees to achieve better generalization ability than a single classifier.

[0048] Step 5: Obtain feature importance and tree structure knowledge of the source domain random forest model.

[0049] Feature importance weight matrix generated by source domain random forest model W This guides the selection of target domain features and enhances the contribution of key features.

[0050] Step 6: Calculate the KL divergence between the source domain data and the target domain data. Dynamically evaluate feature transferability using the KL divergence and automatically filter out mismatched features, such as invalid features caused by differences in device models. The formula for discrete features is:

[0051] ,

[0052] in Eigenvalues ​​in the source domain The probability of its occurrence, For the target domain The probability of occurrence of is given by Δ, where Δ is the number of bins for discretizing the feature value. Used to control migration intensity.

[0053] when hour, α =0 indicates that feature transfer is disabled; hour, α =1 indicates direct migration; hour, α =0.7 indicates decay migration.

[0054] 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.

[0055] Based on the transferability evaluation results in step 6, the feature importance of the random forest obtained in step 5 is transferred. Feature importance transfer is defined as:

[0056] ,

[0057] in It is a featuref j In the source domain i The importance of class patterns D For feature dimensions.

[0058] In one embodiment, the data category weight of the target domain is calculated based on the coolant system data of a specific triple-loop pressurized water reactor nuclear power unit. Then, the target domain data is weighted, with source domain knowledge being fused through exponential weighting. α Item and Category Balance 1- α This item helps mitigate data distribution skew; the specific expression is as follows:

[0059] ,

[0060] Then, using the obtained weights, the random forest model for the target domain is trained to obtain the state prediction model of this embodiment.

[0061] Step 8: Embed the trained state prediction model into the real-time monitoring system. Based on the model's output, when the probability of a certain category exceeds a set threshold, the system issues an alarm signal to remind the operator to perform maintenance or repair.

[0062] refer to Figures 2 to 4 The method in this embodiment can maintain high classification accuracy under conditions of few samples, while also having the dual advantages of model stability and decision interpretability. It is particularly suitable for industrial scenarios such as nuclear power equipment, which have extremely high requirements for safety and reliability.

[0063] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the implementation process of the present invention has been described in detail above, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting equipment state using a small number of samples based on transfer learning, characterized in that, include: Step 1: Use the reactor transient and accident simulation software PCTran to simulate various operating states of pressurized water reactor nuclear power equipment and obtain simulation data of operating parameters under different states; Step 2: Extract time-domain features from the simulated data; Step 3: Train the source domain random forest model using the time-domain feature data extracted in Step 2. Use the nearest-neighbor class mean classifier as the splitting function of the random forest, obtain the information gain of each splitting function, and determine the optimal splitting 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 feature transferability through KL divergence, and automatically filter out mismatched features; Step 7: Based on the feature transferability evaluation results in Step 6, transfer the feature importance of the source domain random forest to the target domain, weight the data in the target domain, and train the target domain random forest model.

2. The method for predicting equipment state based on transfer learning with few samples as described in claim 1, characterized in that, Step 2 specifically involves taking time-series segments of the simulation data with a fixed time step and calculating the statistical characteristics of each segment of the running parameter sequence.

3. The method for predicting equipment state based on transfer learning with few samples as described in claim 2, characterized in that, The statistical characteristics include mean, standard deviation, skewness, and kurtosis.

4. The method for predicting equipment state based on transfer learning with few samples as described in claim 1, characterized in that, In step 3, the temporal feature data obtained in step 2 is used to construct samples and serve 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. Classification learning is achieved by training the subtree branch nodes. Each subtree randomly draws the same number of samples from the instance pool. A set of splitting functions is randomly generated at each subtree branch node. The optimal splitting function is determined by a recursive training process.

5. The method for predicting equipment state based on transfer learning with few samples as described in claim 4, characterized in that, At each subtree branch 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.

6. The method for predicting equipment state based on transfer learning with few samples as described in claim 5, characterized in that, The information gain of each splitting function is obtained by calculating the entropy of the categories.

7. The method for predicting equipment state based on transfer learning using few samples as described in claim 1, characterized in that, In step 4, the prediction results of the nearest class mean classifier are aggregated by 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 state based on transfer learning with few samples as described in claim 1, characterized in that, In step 7, when weighting the target domain data, source domain knowledge and category balance are fused through exponential weighting, and then the target domain random forest model is trained using the weighted weights.

Images