Superconducting magnet multi-source fusion prediction method based on complementary empirical mode decomposition
By combining complementary empirical mode decomposition and LSTM model, the noise and redundancy problems of multi-source data of superconducting magnets are solved, and high-precision temperature-strain prediction is achieved, supporting damage identification and condition monitoring.
Patent Information
- Application Number
- CN202610595462.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-30
- Publication Date
- 2026-08-25
AI Technical Summary
Existing technologies struggle to effectively handle noise and redundancy issues in multi-source data from superconducting magnets, resulting in low accuracy in temperature and strain predictions, and insufficient generalization ability, especially under complex operating conditions.
Complementary Empirical Mode Decomposition (CEEMD) is used to decompose multi-source signals, extract multi-level intrinsic mode functions (IMFs), calculate mutual information entropy and determine feature weights using an objective weighting method, and combine it with an LSTM hyperparameter adaptive model for temperature-strain prediction.
It improves the data quality of multi-source signals, reduces noise interference, enhances prediction accuracy and model performance, and can accurately capture the effect of temperature on strain, which is helpful for damage identification and condition monitoring.
Smart Images
Figure CN122634473A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of multi-sensor data fusion state prediction technology for complex equipment such as ultra-high-speed low-vacuum maglev transportation systems, and in particular to a multi-source fusion prediction method for superconducting magnets based on complementary empirical mode decomposition. Background Technology
[0002] Superconducting magnets, as typical components of superconducting electromagnetic levitation and propulsion systems, possess numerous advantages that traditional magnets cannot achieve, such as not generating heat, small size, and high magnetic field density. In actual experiments, superconducting magnets utilize a wide variety and large number of surface sensors, and multi-source information, including temperature-induced strain and deflection, is a crucial variable in the cooling process of superconducting magnets. For large devices such as superconducting magnets, the interpeak strain caused by temperature is approximately ten times higher than that of typical structural anomalies. This indicates that even if equipment damage causes abnormal strain responses, these changes may be overwhelmed by temperature-induced strain variations, significantly hindering damage identification and condition assessment of superconducting magnets. Therefore, establishing a multi-source condition prediction method for superconducting magnets is beneficial for providing temperature-strain reference values for superconducting magnet health management, eliminating temperature influences from measured data, better identifying damage and condition anomalies in superconducting magnets, and is of great significance for judging the magnet health during the cooling process.
[0003] For the problem of assessing the condition of superconducting magnets, which is primarily dependent on strain, two main approaches are generally taken: one is to directly extract information such as temperature and strain from measured data using signal processing techniques; the other is to integrate existing sensor information to establish a model of the relationship between temperature and strain, separating the inherent temperature response from the measured data to further uncover anomaly information. Various modeling methods provide solutions for analyzing the mapping relationship between temperature and strain data. On the one hand, multiphysics coupling simulation is an important means of studying the strain of superconducting magnets, and finite element analysis (FEA) is a crucial tool for multiphysics coupling simulation. By coupling multiple physical fields such as electromagnetic, thermal, and mechanical fields through methods like FEA, the strain distribution of superconducting magnets under different operating conditions can be simulated. Combining the simulation results with actual measurement data not only verifies and optimizes the simulation model but also provides more accurate prior knowledge and constraints for strain prediction. However, in real-world operating conditions, the working environment of superconducting magnets is complex, and the quality of multi-source data acquired by sensors is difficult to guarantee. In this context, existing finite element-based simulation methods still need to be improved in terms of generalization ability and adaptability when faced with new structures, new applications, and extreme operating conditions of superconducting magnets.
[0004] On the other hand, some studies choose to process and establish temperature-strain mappings from a signal perspective. Machine learning and deep learning algorithms are widely used for multi-source state prediction, primarily based on strain. Machine learning methods such as Support Vector Machines (SVM) and Random Forests can extract potential patterns and regularities from large amounts of multi-source data, establishing models of the relationship between strain and various influencing factors. Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) in deep learning, and their variants such as Long Short-Term Memory Networks (LSTM), are better suited to handling time-series data and complex nonlinear relationships, and can better capture the dynamic changes in strain of superconducting magnets. However, due to the different specific problems they face, the fusion of multi-source long-term series data is difficult in terms of parameter tuning and training time. Furthermore, how to handle the different amounts of information implied by different sensors to improve prediction accuracy has become an urgent problem to be solved. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a multi-source fusion prediction method for superconducting magnets based on complementary empirical mode decomposition, which can solve the problems in the prior art.
[0006] The technical solution of this invention: a multi-source fusion prediction method for superconducting magnets based on complementary empirical mode decomposition, wherein the method includes:
[0007] Collect raw datasets of multi-source signals, preprocess the raw datasets, and divide the preprocessed datasets into training sets, test sets, and validation sets according to a predetermined ratio;
[0008] The multi-source signals in the original dataset are decomposed by complementary set empirical mode decomposition, multi-level intrinsic mode functions (IMFs) are extracted, and low-frequency signals are reconstructed to extract effective features from the low-frequency signals.
[0009] Calculate the mutual information entropy among the extracted effective features, and calculate the weight of each feature based on the objective weighting method;
[0010] The multi-source signals in the preprocessed dataset are input into the LSTM hyperparameter adaptive model for model training.
[0011] Based on the trained LSTM hyperparameter adaptive model, a temperature-strain model is constructed, and the prediction results of the temperature-strain model for superconducting magnets from multiple sources and multiple sensors are compared with the measured data to verify the accuracy of the prediction results.
[0012] Preferably, the method further includes:
[0013] Use visualization tools to create comparison images based on the comparison results.
[0014] Preferably, preprocessing the original dataset includes: filling in missing values, removing outliers, normalizing input data variables, and windowing the data frame by frame.
[0015] Preferably, calculating the mutual information entropy among the extracted effective features includes:
[0016] Calculate the mutual information between signal features in the original dataset;
[0017] The mutual information entropy between each signal feature in the original dataset is calculated based on the mutual information.
[0018] Preferably, mutual information is calculated using the following formula:
[0019] ,
[0020] in, It is a variable and variables mutual information, It is a variable The x-th eigenvector and variable in The joint probability distribution of the y-th eigenvector in the dataset. and They are variables The x-th eigenvector and variable in The marginal probability distribution of the y-th eigenvector in the dataset.
[0021] Preferably, the mutual information entropy is calculated using the following formula:
[0022] ,
[0023] in It is a variable and variables mutual information, and They are variables and variables entropy, It is a variable and variables The joint entropy.
[0024] Preferably, the model is trained and the model parameters are optimized using the backpropagation algorithm.
[0025] Preferably, the predetermined ratio is 8:1:1.
[0026] The above technical solution allows for the decomposition of multi-source signals using Complementary Empirical Mode Decomposition (CEEMD), ensuring the integrity of multi-source, multi-sensor information from superconducting magnets while increasing the proportion of effective information. The method described in this invention not only processes complex multi-channel signals but also removes noise and improves data quality. Furthermore, this invention employs mutual information entropy and critic weight calculation methods for effective data feature selection and assigns corresponding objective weights to multi-source signals based on the calculation results. This feature selection method reallocates data weights, significantly reducing computational costs while improving model performance and avoiding interference from redundant data in traditional methods. In addition, this invention combines a validated LSTM hyperparameter adaptive model to predict multi-channel information, establishing a relatively accurate temperature-strain state prediction model. This temperature-strain state prediction model can accurately capture the influence of temperature on the strain of superconducting magnets, which is beneficial for further research on damage identification and magnet condition monitoring, and has significant application value. Attached Figure Description
[0027] The accompanying drawings, which form part of this specification, are provided to further illustrate embodiments of the invention and, together with the textual description, explain the principles of the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.
[0028] Figure 1 The flowchart of the multi-source, multi-sensor prediction method for superconducting magnets based on complementary set empirical mode decomposition provided in this embodiment of the invention is shown below.
[0029] Figure 2 This is a flowchart of the empirical mode decomposition of complementary sets in an embodiment of the present invention. Detailed Implementation
[0030] Specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. In the following description, specific details are set forth for purposes of explanation and not limitation, in order to aid in a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention may be practiced in other embodiments departing from these specific details.
[0031] It should be noted that, in order to avoid obscuring the invention with unnecessary details, only the device structure and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.
[0032] like Figure 1-2As shown, this embodiment of the invention provides a multi-source fusion prediction method for superconducting magnets based on complementary empirical mode decomposition, wherein the method includes:
[0033] Multi-source signal acquisition and preprocessing: Acquire raw datasets of multi-source signals, preprocess the raw datasets, and divide the preprocessed datasets into training, testing, and validation sets according to a predetermined ratio;
[0034] For example, multiple sensors can be used to collect multi-source signal data such as temperature and strain of superconducting magnets. The collected raw signals can be preprocessed by framing, windowing, and normalization to ensure the accuracy and effectiveness of the signals and facilitate further signal analysis.
[0035] Signal decomposition: Multi-source signals in the original dataset are decomposed by complementary set empirical mode decomposition, multi-level intrinsic mode functions (IMFs) are extracted, and low-frequency signals are reconstructed to extract effective features from the low-frequency signals.
[0036] This allows for the increase of the proportion of useful information while ensuring signal integrity. Low-frequency noise and other irrelevant signal components are removed, thus improving data quality.
[0037] Feature selection and weight allocation: Calculate the mutual information entropy among the extracted effective features, and calculate the weight of each feature based on the objective weighting method;
[0038] For example, by calculating the mutual information entropy between various signal features, the dependency between different features can be evaluated, the standard deviation and correlation coefficient can be calculated, and the weights can be calculated based on the correlation and importance of the signals using the objective weighting method. Objective weights can be assigned to the signals of each dimension, the most representative data features can be selected, and redundant or unimportant features can be removed, thereby reducing computational complexity and improving model performance.
[0039] Deep learning model training and optimization: The multi-source signals in the preprocessed dataset are input into the LSTM hyperparameter adaptive model for model training; then the LSTM hyperparameter adaptive model can be combined for temperature-strain prediction modeling.
[0040] Temperature-strain model construction: Based on the trained LSTM hyperparameter adaptive model, a temperature-strain model is constructed, and the prediction results of the temperature-strain model for superconducting magnets from multiple sources and multiple sensors are compared with the measured data to verify the accuracy of the prediction results (evaluate the model's performance in multi-source, multi-sensor prediction problems).
[0041] The above technical solution allows for the decomposition of multi-source signals using Complementary Empirical Mode Decomposition (CEEMD), ensuring the integrity of multi-source, multi-sensor information from superconducting magnets while increasing the proportion of effective information. The method described in this invention not only processes complex multi-channel signals but also removes noise and improves data quality. Furthermore, this invention employs mutual information entropy and critic weight calculation methods for effective data feature selection and assigns corresponding objective weights to multi-source signals based on the calculation results. This feature selection method reallocates data weights, significantly reducing computational costs while improving model performance and avoiding interference from redundant data in traditional methods. In addition, this invention combines a validated LSTM hyperparameter adaptive model to predict multi-channel information, establishing a relatively accurate temperature-strain state prediction model. This temperature-strain state prediction model can accurately capture the influence of temperature on the strain of superconducting magnets, which is beneficial for further research on damage identification and magnet condition monitoring, and has significant application value.
[0042] Complementary ensemble empirical mode decomposition (CIMD) is a method proposed to address the problems of mode aliasing and endpoint effects often encountered by traditional CIMD when processing complex non-stationary signals. It involves adding pairs of positive and negative white noise with finite amplitudes to the original signal and performing multiple ensemble averaging operations to effectively improve signal decomposition. This method increases the proportion of useful information while preserving signal integrity, removes low-frequency noise and other irrelevant signal components, and thus improves data quality.
[0043] Specifically, the added paired positive and negative noise can supplement and correct the original signal at different scales, allowing modes that might otherwise overlap to be clearly separated. After multiple ensemble averaging operations, the effects of the noise are canceled out, leaving only the true signal components, thereby greatly improving the precision and accuracy of the decomposition.
[0044] According to one embodiment of the present invention, the method further includes:
[0045] Use visualization tools to create comparison images based on the comparison results.
[0046] Using visualization tools to create comparative images can intuitively demonstrate the model's performance at different time periods, analyze the robustness and applicability of the comparison method, and comprehensively evaluate the model's performance in multi-source, multi-sensor prediction problems.
[0047] According to one embodiment of the present invention, preprocessing the original dataset includes: supplementing missing values, removing outliers, normalizing input data variables, and windowing by frame (i.e., data preparation and data cleaning).
[0048] According to one embodiment of the present invention, calculating the mutual information entropy among the extracted effective features includes:
[0049] Calculate the mutual information between signal features in the original dataset;
[0050] The mutual information entropy between each signal feature in the original dataset is calculated based on the mutual information.
[0051] Mutual information is a measure used to quantify the dependency between two random variables. It measures the amount of information obtained about one variable from another and is commonly used in feature selection and statistical analysis. A higher mutual information value indicates a stronger relationship between the variables; a value of zero indicates that the two variables are independent and unrelated.
[0052] According to one embodiment of the present invention, mutual information is calculated using the following formula:
[0053] ,
[0054] in, It is a variable and variables mutual information, It is a variable The x-th eigenvector and variable in The joint probability distribution of the y-th eigenvector in the dataset. and They are variables The x-th eigenvector and variable in The marginal probability distribution of the y-th eigenvector in the dataset.
[0055] In feature selection, mutual information can help identify features that are highly correlated with the target variable, thereby improving model performance and reducing computational complexity.
[0056] Mutual information entropy is used to quantify the interdependence or information sharing between two random variables. Combining the concept of information entropy, it measures the amount of information obtained from one variable about another, playing a crucial role in many fields. In statistics and machine learning, mutual information entropy is widely used for feature selection and variable correlation analysis. By calculating the mutual information entropy between variables, features with strong correlation to the target variable can be effectively selected, thereby improving model performance and prediction accuracy. For example, in image recognition tasks, using mutual information entropy to select features that significantly contribute to image classification can reduce data dimensionality, lower computational complexity, and improve classification results.
[0057] According to one embodiment of the present invention, mutual information entropy is calculated by the following formula:
[0058] ,
[0059] in It is a variable and variables mutual information, and They are variables and variables entropy, It is a variable and variables The joint entropy.
[0060] By evaluating the dependencies between different features, and using an objective weighting method, weights are calculated by combining the correlation and importance of signals. Objective weights are assigned to signals in each dimension, the most representative data features are selected, and redundant or unimportant features are removed, thereby reducing computational complexity and improving model performance.
[0061] The main steps of the objective empowerment method include:
[0062] a) Calculate the standard deviation of each eigenvector to reflect the degree of dispersion of the eigenvector values.
[0063] b) Calculate the correlation coefficients between each eigenvector to assess the conflict between them.
[0064] c) Determine the information content of each feature vector based on the standard deviation and correlation coefficient.
[0065] d) Calculate the weights of each feature vector based on the amount of information.
[0066] According to one embodiment of the present invention, the model is trained and the model parameters are optimized by using the backpropagation algorithm.
[0067] This ensures the model's ability to generalize across multiple data sources.
[0068] More specifically, the model is trained using the backpropagation algorithm to optimize model parameters, ensure the model's generalization ability on multi-source data, and calculate various performance metrics, such as root mean square error (RMSE), mean absolute error (MAE), and R² coefficient of determination, to quantify the gap between the prediction results and the actual data, provide objective evidence for the model's quality, verify the prediction accuracy, and monitor the performance on the validation set to avoid overfitting.
[0069] According to one embodiment of the present invention, the predetermined ratio is 8:1:1.
[0070] In other words, the data can be divided into training set, test set, and validation set, accounting for 80%, 10%, and 10% of the total dataset, respectively.
[0071] As can be seen from the above embodiments, this invention fuses multi-source information during the cooling process of superconducting magnets and establishes a multi-source, multi-sensor prediction method for superconducting magnets based on complementary set empirical mode decomposition for important state parameters such as temperature and strain of complex equipment. This method makes full use of the information implicit in multiple sources and multiple sensors, providing a high-precision temperature-strain reference value prediction method with short training time for the health management of superconducting magnets. It is beneficial to remove the influence of temperature from the measured data, better identify the damage and abnormal state of superconducting magnets, and is of great significance for judging the health of magnets during the cooling process of superconducting magnets.
[0072] The features described and / or illustrated above with respect to one embodiment may be used in the same or similar manner in one or more other embodiments, and / or in combination with or in lieu of features in other embodiments.
[0073] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, whole, step, or component, but does not exclude the presence or addition of one or more other features, wholes, steps, components, or combinations thereof.
[0074] The apparatus and methods described above can be implemented in hardware or in combination with software. This invention relates to computer-readable programs that, when executed by a logic component, enable that logic component to implement the apparatus or constituent parts described above, or to implement the various methods or steps described above. This invention also relates to storage media for storing the above programs, such as hard disks, magnetic disks, optical disks, DVDs, flash memory, etc.
[0075] Many features and advantages of these embodiments are apparent from this detailed description, and therefore the appended claims are intended to cover all such features and advantages of these embodiments that fall within their true spirit and scope. Furthermore, since many modifications and alterations will readily occur to those skilled in the art, the embodiments of the invention are not intended to be limited to the precise structures and operations illustrated and described, but rather to encompass all suitable modifications and equivalents falling within their scope.
[0076] The parts of this invention not described in detail are techniques known to those skilled in the art.
Claims
1. A multi-source fusion prediction method for superconducting magnets based on complementary empirical mode decomposition, characterized in that, The method includes: Collect raw datasets of multi-source signals, preprocess the raw datasets, and divide the preprocessed datasets into training sets, test sets, and validation sets according to a predetermined ratio; The multi-source signals in the original dataset are decomposed by complementary set empirical mode decomposition, multi-level intrinsic mode functions (IMFs) are extracted, and low-frequency signals are reconstructed to extract effective features from the low-frequency signals. Calculate the mutual information entropy among the extracted effective features, and calculate the weight of each feature based on the objective weighting method; The multi-source signals in the preprocessed dataset are input into the LSTM hyperparameter adaptive model for model training. Based on the trained LSTM hyperparameter adaptive model, a temperature-strain model is constructed, and the prediction results of the temperature-strain model for superconducting magnets from multiple sources and multiple sensors are compared with the measured data to verify the accuracy of the prediction results.
2. The method according to claim 1, characterized in that, The method also includes using visualization tools to draw comparison images based on the comparison results.
3. The method according to claim 1, characterized in that, Preprocessing of the original dataset includes: filling in missing values, removing outliers, normalizing input data variables, and windowing the data frame by frame.
4. The method according to claim 1, characterized in that, The calculation of mutual information entropy among the extracted effective features includes: Calculate the mutual information between each signal feature in the original dataset; calculate the mutual information entropy between each signal feature in the original dataset based on the mutual information.
5. The method according to claim 4, characterized in that, Mutual information is calculated using the following formula: ， in, It is a variable and variables mutual information, It is a variable The x-th eigenvector and variable in The joint probability distribution of the y-th eigenvector in the dataset. and They are variables The x-th eigenvector and variable in The marginal probability distribution of the y-th eigenvector in the dataset.
6. The method according to claim 5, characterized in that, The mutual information entropy is calculated using the following formula: ， in It is a variable and variables mutual information, and They are variables and variables entropy, It is a variable and variables The joint entropy.
7. The method according to any one of claims 1-6, characterized in that, Model parameters are optimized by training the model using the backpropagation algorithm.
8. The method according to any one of claims 1-7, characterized in that, The planned ratio is 8:1:1.