A structural health prediction method based on rat cage strain measurement

Abstract

This invention discloses a structural health prediction method based on squirrel cage strain measurement. A simulation model of a squirrel cage elastic support structure is established, and simulation analysis is performed on the structure. An XGBoost model from machine learning is used to construct a forward prediction model from the load on the squirrel cage elastic support structure to the overall stress / deformation. Tensor decomposition technology is selected to compress the model parameter scale, enabling the prediction of the overall stress / deformation of the squirrel cage elastic support structure after input load. Experimental data on the load and strain on the ribs of the squirrel cage elastic support structure under working conditions are obtained, and a deep convolutional neural network model embedded with LSTM units is constructed and trained to achieve the inverse mapping from the strain on the ribs to the load on the squirrel cage elastic support structure. This method overcomes the core problems of low efficiency in traditional finite element simulation, difficulty in directly measuring the load on the squirrel cage support, difficulty in modeling high-dimensional data, and lag in the detection of hidden damage, providing a low-cost, high-precision solution for the health management of aerospace power equipment.

Description

Technical Field

[0001] This invention relates to the technical field of elastic support systems for rat cages, and specifically to a structural health prediction method based on rat cage strain measurement. Background Technology

[0002] As a core support component of rotor systems in aero-engines and gas turbines such as turboshafts, turbojet engines, and turbofans, the performance of squirrel-cage elastic supports directly determines the dynamic stability and service reliability of rotating machinery. This structure achieves stiffness adjustment and vibration isolation of the rotor system through elastic deformation, and must simultaneously meet stiffness accuracy requirements (error controlled within 1%) and deflection limitations (to avoid stress concentration leading to fatigue fracture) under extreme operating conditions (high temperature, high pressure, high speed). However, in actual operation, squirrel-cage elastic supports are subjected to complex effects such as alternating loads, thermal deformation, and residual assembly stress over long periods, which can easily lead to local stiffness degradation, cracks at the root of the squirrel cage bars, and other latent damage, resulting in serious failures such as rotor critical speed deviation and excessive vibration amplitude.

[0003] Traditional methods for designing and maintaining squirrel cage elastic supports have certain limitations. On the one hand, the design process often relies on periodic maintenance and limited data sources, making it difficult to detect potential structural problems in a timely manner. On the other hand, while existing optimization design methods have achieved some success in improving stiffness design accuracy and reducing the number of design iterations, they are still insufficient in real-time monitoring and early warning.

[0004] Against this backdrop, structural health prediction methods based on strain measurement of squirrel cages have become a research hotspot. By monitoring the strain distribution of key components such as cage arms and bars in real time, the changes in support stiffness and stress concentration can be accurately inverted, providing key technical support for the following objectives: Stiffness degradation assessment: By combining squirrel cage structural parameters (bar length, cross-sectional dimensions, etc.) with strain data, a stiffness-strain mapping model can be established to achieve quantitative characterization of nonlinear stiffness degradation; Fatigue life prediction: Based on strain time series data and fatigue stress conditions, a prediction model for crack initiation and propagation can be constructed, overcoming the limitations of traditional empirical formulas; Early fault warning: By detecting strain anomalies (such as local strain mutations), typical faults such as root cracks and loose assembly can be identified, avoiding catastrophic failures caused by damage accumulation; Optimized design verification: Real-time feedback can be provided for optimizing squirrel cage structural parameters (such as the number of arms and geometry), shortening the design iteration cycle and improving reliability.

[0005] Currently, the aircraft engine health management (EHM) system has higher requirements for real-time performance and accuracy. The squirrel cage strain measurement technology, combined with digital twins and deep learning algorithms, can overcome the limitations of traditional offline detection, promote the transformation of aircraft power equipment to predictive maintenance mode, and has significant economic benefits and engineering application value. Summary of the Invention

[0006] This invention belongs to the field of health monitoring technology for aero-engine and gas turbine rotor systems, specifically relating to a health prediction method and system for squirrel cage elastic support structures based on strain-load inverse mapping and machine learning-accelerated simulation calculation.

[0007] This method overcomes the efficiency bottleneck of traditional finite element simulation by constructing a dual-channel neural network architecture and a multi-scale data fusion mechanism, achieving closed-loop monitoring of dynamic load identification, rapid deformation prediction, and real-time health status assessment of squirrel-cage supported structures. Since traditional finite element simulation is inefficient and time-consuming, this invention employs the XGBoost model from machine learning to construct a forward mapping model from the load on the squirrel-cage elastic support structure to the overall strain / deformation. This model can efficiently process high-dimensional data and uses tensor decomposition technology to compress the model parameter scale, resulting in a single deformation prediction time that is much shorter than normal simulation calculations. In actual working conditions, the load on the squirrel-cage elastic support structure is difficult to detect. To obtain a more accurate load on the squirrel-cage elastic support structure, a deep convolutional neural network (DCNN) is constructed to achieve an inverse mapping from the strain on the ribs to the load on the squirrel-cage elastic support structure. LSTM units are embedded into the deep convolutional neural network to construct a prediction model, extracting shallow features, deep features, and time-related feature information to achieve a load recognition rate ≥95%. This invention overcomes core problems in health monitoring of squirrel cage support structures, such as the difficulty of decoupling multi-source signals, low efficiency of high-precision simulation, and lag in the identification of hidden damage, through a technical chain of "physical mechanism modeling - data-driven optimization - virtual-real interaction verification". It provides an innovative solution for predictive maintenance of aerospace power equipment.

[0008] To achieve the above objectives, this application discloses a structural health prediction method based on strain measurement of a rat cage, the core technical solution of which is as follows:

[0009] S1. Establish a simulation model of the elastic support structure of the squirrel cage and conduct simulation analysis of the elastic support structure of the squirrel cage;

[0010] S2. The XGBoost model in machine learning is used to construct a positive prediction model from the load of the squirrel cage spring support structure to the overall stress / deformation. This model can efficiently process high-dimensional data and compress the model parameter scale by selecting tensor decomposition technology. After inputting the load, the overall stress / deformation of the squirrel cage spring support structure can be predicted with a prediction accuracy of 95%.

[0011] S3. Obtain experimental data on the load and strain on the ribs of the squirrel cage spring support structure under working conditions, construct and train a deep convolutional neural network model embedded with LSTM units, realize the inverse mapping from the strain on the ribs to the load on the squirrel cage spring support structure, and continuously adjust the model parameters to make the load recognition rate ≥95%.

[0012] S1 includes the following steps:

[0013] S1.1 Measure the actual dimensions of the squirrel cage spring support structure, establish a simulation model in ANSYS, apply loads to the squirrel cage spring support structure according to the actual load-bearing position, obtain the overall strain / deformation, and also obtain the position of maximum strain on the ribs, providing ideas for attaching strain gauges in subsequent experiments.

[0014] In S1.2, using the custom design in the ANSYS Workbench response surface design module, multiple sets of loads with different magnitudes and directions are set to obtain the corresponding stress / deformation, providing training data for subsequent training of the forward mapping model.

[0015] S2 includes the following steps:

[0016] S2.1 Based on S1.2, a large amount of data corresponding to load and strain is obtained, and the data is preprocessed;

[0017] S2.2 Construct and train the XGboost model, with the load on the squirrel cage elastic support system as input and the overall stress / deformation of the squirrel cage elastic support system as output. Establish a positive mapping model from the load of the squirrel cage elastic support structure to the overall stress / deformation. By continuously adjusting the parameters, the prediction accuracy is made ≥95%.

[0018] S2.3 Tensor decomposition technology is selected to compress the model parameter scale, reducing the number of tens of thousands of nodes, accelerating the model's response time, and reducing the time required for single deformation prediction.

[0019] S3 includes the following steps:

[0020] Based on the location of maximum rib strain obtained from the simulation in S1.1, fiber optic gratings or strain gauges are pasted in the form of strain rosettes in key areas such as the root of the ribs of the squirrel cage and the transition fillet of the support arm. The sampling frequency is ≥1kHz (matching the dynamic load frequency) to collect multi-channel strain signals.

[0021] S3.2 By applying excitation forces of different magnitudes and orientations to the shaft through an electromagnetic vibrator, the shaft will be displaced after being subjected to the excitation force, and the bearing will also be subjected to force, which will be transmitted to the squirrel cage elastic support system. The force sensor in the outer ring clamp can detect the load data.

[0022] S3.3 Cleans and preprocesses the acquired strain signals to obtain data that can be input into the neural network;

[0023] S3.4 Construct and train a deep convolutional neural network model embedded with LSTM units, where the strain signal is used as input and the load is used as output. By continuously adjusting the model parameters, the load recognition rate is made ≥95%, thus completing the inverse mapping from the strain on the rib to the load on the squirrel cage spring support structure.

[0024] The present invention has the following advantages:

[0025] This invention addresses the challenge of health monitoring of squirrel-cage elastic support structures in aero-engine and gas turbine rotor systems. It proposes an innovative method based on strain-load inverse mapping and machine learning to accelerate finite element analysis. By combining XGBoost forward mapping and tensor decomposition techniques, and integrating deep convolutional neural networks (DCNN) and LSTM units to construct an inverse mapping model, it achieves rapid and high-precision load identification and deformation prediction. This method overcomes core problems such as low efficiency of traditional finite element simulation, difficulty in directly measuring the load on squirrel-cage supports, challenges in high-dimensional data modeling, and lag in the detection of latent damage. Through a combination of physical mechanism modeling and data-driven optimization, it significantly improves the real-time performance and reliability of predictive maintenance, providing a low-cost, high-precision solution for the health management of aero-engine power equipment. Attached Figure Description

[0026] Figure 1 This is a technical roadmap for an embodiment of the present invention.

[0027] Figure 2 This is a three-dimensional model of the elastic support of the squirrel cage established in an embodiment of the present invention.

[0028] Figure 3 The mesh has been generated for the elastic support model of the squirrel cage in this embodiment of the invention.

[0029] Figure 4 The prediction results in this embodiment of the invention are visualized as a model diagram.

[0030] Figure 5 This is a schematic diagram of the force ring clamp in an embodiment of the present invention.

[0031] Figure 6-1 This is a diagram showing the accuracy of load identification in the Y direction.

[0032] Figure 6-2 This is a diagram showing the accuracy of load identification in the Z-direction. Detailed Implementation

[0033] like Figure 1 As shown, this invention implements a structural health prediction method based on rat cage strain measurement:

[0034] This invention addresses the challenge of health monitoring of squirrel cage elastic support structures in aero-engine and gas turbine rotor systems. It proposes an innovative method based on strain-load inverse mapping and machine learning to accelerate finite element method calculations. This method combines XGBoost forward mapping and tensor decomposition techniques, fusing deep convolutional neural networks (DCNN) and LSTM units to construct an inverse mapping model. The specific calculation steps are as follows:

[0035] The scope of protection of this invention is not limited to the description of this embodiment.

[0036] The first step is to establish a simulation model of the elastic support structure of the squirrel cage and conduct simulation analysis on the elastic support structure of the squirrel cage.

[0037] 1) Based on actual size measurements, reverse modeling is performed using ANSYS SpaceClaim to reconstruct a 3D solid model, such as... Figure 2 As shown, the geometric reconstruction error is controlled to a very small extent. A hybrid meshing technique is used to create a high-precision hexahedral mesh for the rib section, while other areas are meshed with tetrahedral meshes. The overall number of nodes is controlled to within 50,000. Hm adaptive mesh refinement technology is used to ensure the accuracy of the stress gradient region, such as... Figure 3 As shown. Boundary constraints and forces were simulated to reflect the actual constraints and loads experienced by the squirrel cage elastic support system. Dynamic load simulation employed sine and cosine sweep frequency loads, solved using the transient dynamics module. The overall deformation and equivalent stress / strain were viewed using the ANSYS Mechanical post-processing module. The location of maximum strain in the ribs was identified using strain contour maps, and the experimental patch layout was optimized based on the strain gradient distribution.

[0038] 2) Regarding the generation of multi-condition response surface data, the ANSYS response surface design module was used to plan load parameters through experimental design methods. A large number of samples were generated using Latin hypercube sampling to cover the design space and avoid parameter redundancy. ANSYS Workbench was automatically invoked via ACT scripts to execute multiple load condition simulations, saving the maximum strain value of the ribs and the overall deformation.

[0039] This approach addresses the strain prediction bias caused by geometric simplification in traditional simulations through high-fidelity modeling across the entire geometry-materials-load chain. Furthermore, it incorporates data-driven experimental design optimization to improve training data coverage and support the generalization capability of lightweight models. Strain hotspot localization results directly guide experimental patch placement, reducing trial-and-error costs and providing high-quality datasets for subsequent neural network model training.

[0040] The second step involves using the XGBoost model in machine learning to construct an end-to-end nonlinear mapping relationship from structural load input to overall deformation output. This model can efficiently process massive amounts of data. Tensor decomposition technology is selected to compress the model parameter scale, enabling the prediction of the overall stress / deformation of the squirrel cage spring support structure after inputting the load, with a prediction accuracy of 99.95%.

[0041] 1) Multi-source heterogeneous data acquisition and preprocessing: This step constructs a data-driven model input through high-fidelity simulation and intelligent feature extraction. The specific implementation process includes:

[0042] Data Acquisition: A multiphysics finite element model of the squirrel-cage spring support structure was established based on ANSYS Workbench, and a dynamic load spectrum was set, including impact loads, random vibrations, and alternating conditions. Using the modal superposition method and transient dynamic analysis module, time-series simulation data of nodal stresses (σxx / σxy / σxz / σyy / σyz / σzz), strains (ξxx / ξxy / ξxz / ξyy / ξyz / ξzz), deformations (x / y / z), and environmental parameters were generated. The sampling frequency was set to 10kHz to capture high-frequency dynamic responses.

[0043] Data cleaning: A sliding window mechanism was used to segment the time series data, with a window length of 1024 points and a step size of 512 points. The 3δ criterion was used to identify and remove stress mutation outliers, with a threshold set at the mean ± 3 times the standard deviation.

[0044] Feature engineering and extraction: Construct a multi-dimensional normalization processing flow. Min-Max normalization is applied to the mechanical signal, with the range set to [0,1].

[0045]

[0046] Where X = {x1, x2, ..., x} n}

[0047] Key features with a correlation coefficient greater than 0.8 with deformation were selected by calculating mutual information entropy. A deep autoencoder network was constructed, with a 5-layer fully connected encoder structure, LeakyReLU activation function, and a compression dimension of 64; the decoder adopted a symmetric structure.

[0048]

[0049] Where x is the input value; α is a hyperparameter called the "leakage function", which is usually 0.01 or less.

[0050] The nonlinear feature extraction capability is enhanced by an adversarial training strategy. The discriminator adopts the PatchGAN architecture and takes the reconstruction error (mean squared error MSE less than 1e-4) as the optimization objective. The final output is a low-dimensional representation vector with a dimension of 64.

[0051]

[0052] Where the actual values ​​are y1, y2, ..., y n The predicted value is

[0053] This process uses a data-driven approach that augments physical information, preserving over 92% of the dynamic nonlinear features compared to traditional PCA dimensionality reduction, thus providing high-quality input data for subsequent XGBoost model training.

[0054] 2) XGBoost gradient boosting decision tree modeling and optimization: This step uses an ensemble learning framework to build a high-precision structural load-deformation mapping model. The specific implementation process is as follows:

[0055] Model Building and Training: During the model building and training phase, the XGBoost model is first initialized, and relevant parameters such as objective, n_estimators, learning_rate, and max_depth are set. `tree_method='hist'` is used for large-scale data, `early_stopping_rounds=50` is set, and `scale_pos_weight` is adjusted based on class imbalance. During training, GPU acceleration is enabled, `tree_method='gpu_hist'` is set, and CUDA parallel computing is used to improve efficiency. Simultaneously, explicit regularization (controlling model complexity through `reg_alpha` and `reg_lambda`) and implicit regularization (limiting the minimum number of samples in leaf nodes to `min_child_weight=5`) strategies are employed to prevent overfitting and ensure the model's generalization ability and performance.

[0056] Hyperparameter optimization: In the hyperparameter optimization stage, the parameter search space is first determined: learning_rate ranges from [0.01, 0.3], max_depth ranges from [3, 10], subsample ranges from [0.6, 1.0], and reg_alpha ranges from [0, 10]. A random search optimization method is used for different parameters. The optimization process adopts a two-stage approach. First, initial parameters are set based on domain experience. Then, the Optuna framework is used to perform Bayesian optimization for 50 iterations with the MSE of 5-fold cross-validation as the optimization objective. An early stopping mechanism is set, and training is terminated when the MSE of the validation set does not decrease for 50 consecutive rounds, thereby finding the optimal combination of hyperparameters to improve model performance.

[0057] Model Evaluation and Validation: In the model evaluation and validation phase, R is used... 2 The coefficient of determination is used as the core evaluation indicator, while also ensuring that the maximum absolute error (MAE) is less than 5% of the allowable deformation.

[0058]

[0059] Where the actual values ​​are y1, y2, ..., y n The predicted value is

[0060] To verify the model's robustness, adversarial example testing was conducted by adding Gaussian noise (σ = 0.1) to the input features to evaluate model stability. Furthermore, the model's cross-condition generalization ability was verified using the CLUE benchmark set, ensuring its reliability and accuracy across different scenarios and meeting the needs of practical engineering applications.

[0061] 3) Tensor decomposition parameter compression: This step adopts a model compression strategy based on tensor encoding and CP decomposition to further optimize model parameters and improve the model's computation speed. The specific implementation process is as follows:

[0062] First, the tree structure weights and leaf node scores in the XGBoost model are systematically tensor-encoded, and these parameters are organized and transformed into a third-order tensor, where the dimensions correspond to the number of trees, the feature dimension, and the output dimension, respectively, thus laying the foundation for subsequent tensor decomposition operations. Then, using the CP decomposition technique, the constructed third-order parameter tensor is decomposed into a combination of several rank-1 tensors. During this process, the factor matrix is ​​iteratively optimized using Alternating Least Squares (ALS), gradually approximating the structure and information of the original tensor. Ultimately, the model parameter size is successfully compressed to 1 / 10 of its original size, significantly reducing the model's storage space requirements and computational complexity, and significantly improving the model's running efficiency in practical applications.

[0063] Meanwhile, to compensate for the potential loss in prediction accuracy due to model parameter compression, a residual compensation mechanism was cleverly introduced. Specifically, an LSTM-based prediction network was constructed to dynamically compensate for the residuals between the original XGBoost model output and the compressed lightweight tensor model output. This LSTM network can fully mine and learn the time-series features and latent patterns in the residuals, thereby effectively restoring and improving the accuracy of the prediction results while ensuring the model's fast response. This reduces the time required for a single deformation prediction, meeting the dual requirements of real-time performance and accuracy in practical engineering applications.

[0064] After obtaining the predicted stress and deformation values, code is written to call the VTK library (Visualization Toolkit). Utilizing its visualization rendering capabilities, the discrete stress and deformation data are intelligently mapped and refined according to the spatial distribution patterns of the physical field. The code drives the VTK library to build data-color and shape association rules, transforming abstract values ​​into intuitive stress and deformation cloud maps by using specific color gradients corresponding to different numerical ranges. In the cloud maps, the color hierarchy clearly presents the differences in stress and deformation at various parts of the structure. Engineers no longer need to tediously interpret complex data; they can perceive the changes in the structure's mechanical response under load in a comprehensive and immersive way simply through the visualized color distribution and shape changes. This provides intuitive and convincing visual evidence for subsequent design optimization and performance evaluation, transforming abstract mechanical analysis results into vivid and perceptible visual presentations. This effectively assists in the verification of technical solutions and the iteration of innovative designs. Specific effects are shown below. Figure 4 As shown.

[0065] The third step is to obtain experimental data on the load and strain on the ribs of the squirrel cage spring support structure under working conditions, construct and train a deep convolutional neural network model with embedded LSTM units, realize the inverse mapping from the strain on the ribs to the load on the squirrel cage spring support structure, and continuously adjust the model parameters to make the load recognition rate ≥95%.

[0066] 1) Construction of the multimodal experimental data acquisition system. This step prepares the neural network training data by acquiring data from the force sensor in the ring fixture and the strain gauges on the ribs of the squirrel cage. The specific implementation process includes:

[0067] Hardware preparation and installation: Install force sensors at key locations on the ring clamp, such as... Figure 5 As shown, strain gauges were attached to the ribs of the squirrel cage to complete the sensor deployment. Subsequently, a multi-channel synchronous acquisition card and signal conditioning module were configured to ensure correct connection between the sensor and the acquisition equipment. Then, the fixture was fixed and the loading system was calibrated to eliminate initial deformation interference.

[0068] Data Synchronization Acquisition: Time synchronization of force sensor and strain gauge data is achieved through a trigger signal from the main control unit. Then, an appropriate sampling rate is set to cover the highest signal frequency, while channel gain and filtering range are configured. Signal stability is monitored in real time to eliminate the influence of environmental noise such as electromagnetic interference.

[0069] Data preprocessing: The raw data is filtered and denoised to remove high-frequency noise and baseline drift. Outliers such as sensor over-range and signal disconnection are removed. The cross-modal data (force and strain) are also time-axis aligned to ensure data synchronization.

[0070] Database construction: The raw data is archived by time or experiment number, and features are extracted and stored in the database. At the same time, metadata information such as experimental conditions, loading parameters, and equipment status is saved, and key events such as buckling points and fatigue thresholds are labeled manually or semi-automatically.

[0071] System verification and optimization: Functional testing is completed by verifying sensor accuracy, such as force sensor calibration, and checking data synchronization errors. The acquired data is used to train a neural network, the effectiveness of features is evaluated, and then the sensor layout or acquisition parameters are iteratively improved based on the verification results.

[0072] 2) Construction and training of the deep convolutional neural network model: This step involves building a deep convolutional neural network model embedding LSTM units and using the processed experimental data as training data to train the model. The specific implementation process includes:

[0073] Data preparation: The preprocessed data from step 1) is divided into training, validation, and test sets using stratified sampling, with the training and validation sets comprising 80% and the test set comprising 20%. Similarly, stratified sampling is used again to further divide the training and validation sets, with the training set comprising 80% and the test set comprising 20%. Generally, the training set is used for learning model parameters, the validation set for hyperparameter tuning and model structure selection, and the test set for final evaluation of the model's generalization ability.

[0074] Model architecture design: First, a CNN feature extraction part is constructed, taking the processed strain data as input (expanded by time steps, e.g., time steps = 100) and the load data as output. Local features are extracted using 1D convolutional layers, and dimensionality is reduced using pooling layers. Next, LSTM units are embedded for temporal modeling. The feature sequence output by the CNN is input into the LSTM layer, and the return sequences parameter is set to preserve intermediate temporal states. An attention mechanism can be optionally added to strengthen the weights of key time steps. Finally, the output layer is designed, using a fully connected layer to output the predicted value. During implementation, it is important to add a Dropout layer after the convolutional layers for regularization to prevent overfitting; gradient clipping is used during the training phase to limit gradient explosion.

[0075] Model training configuration: This task uses Huber Loss to improve robustness; the optimizer is Adam, the learning rate is set to lr = 1e-4 and weight decay is added; in terms of training parameters, the batch size is set to 64~128, the number of training epochs is set to 100~200 and the early stopping mechanism is enabled, and training is terminated when the validation set loss does not decrease for 5 consecutive epochs.

[0076] Model Validation and Iteration: This task uses MAE (Mean Absolute Error) as the evaluation metric. Based on the validation results, the model can be optimized and iterated in ways such as adjusting the number of LSTM layers or hidden units, increasing the time step, and improving the sensor fusion strategy to improve model performance.

[0077]

[0078] Where the actual values ​​are y1, y2, ..., y n The predicted value is

[0079] 3) Validation of the deep convolutional neural network model: This step verifies the quality of the trained model by testing it, and continuously refines the model based on the training results. The specific implementation process includes:

[0080] Preparation before verification: Divide the data reasonably to ensure that the test set data is independent and has not been used for training and parameter tuning. Divide the time series data in chronological order, and at the same time ensure that the test set covers different working conditions (such as different loads and ambient temperatures).

[0081] Model performance evaluation: During quantitative evaluation, test data is input into the model to obtain predicted values, evaluation indicators are calculated, and the dynamic alignment between the model's predicted time series curve and the true value is compared using the DTW dynamic time warping distance. Since the attention mechanism is included, the weight allocation of LSTM for key time steps is shown. At the same time, a prediction error histogram is plotted to check for systematic biases. During robustness testing, Gaussian noise (SNR = 10dB) is added to the test data or time shift is performed to evaluate the model's anti-interference ability. Adversarial examples (such as FGSM attacks) can also be generated to test the model's sensitivity to small perturbations.

[0082] Iterative optimization process: When the validation set error is high, check if the model is underfitting or if the data distribution is skewed; if the test set generalization ability is poor, increase data diversity. Online incremental training is used, adding newly collected data to the training set and periodically fine-tuning the model. During this process, some layers need to be frozen to prevent overfitting. The execution order is: data partitioning → indicator definition → quantitative testing → qualitative analysis → error localization → model tuning → re-validation. Through this systematic validation and targeted optimization, the reliability of the model in real-world scenarios is ensured, especially strengthening the evaluation of time-series prediction capabilities (such as the dynamic response characteristics of LSTM). Through continuous model adjustments, the predicted load accuracy after input strain reaches over 95%, as shown in the prediction results. Figure 6-1 , Figure 6-2 As shown.

[0083] The above description is a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Those skilled in the art can make appropriate adjustments without departing from the principle of the present invention, and these adjustments should all be included within the scope of protection of the present invention.

Claims

1. A structural health prediction method based on strain measurement of a rat cage, characterized in that, The implementation steps of this method include the following: S1. Establish a simulation model of the elastic support structure of the squirrel cage and conduct simulation analysis of the elastic support structure of the squirrel cage; S2. A positive prediction model for the load to overall stress / deformation of the squirrel cage spring support structure is constructed using the XGBoost model in machine learning to process high-dimensional data. Tensor decomposition technology is selected to compress the model parameter scale, so as to predict the overall stress / deformation of the squirrel cage spring support structure after inputting the load. S3. Obtain experimental data on the load and strain on the ribs of the squirrel cage spring support structure under working conditions, construct and train a deep convolutional neural network model embedded with LSTM units, realize the inverse mapping from the strain on the ribs to the load on the squirrel cage spring support structure, and adjust the model parameters to make the load recognition rate ≥95%; S3 includes the following steps: S3.1 Based on the location of maximum rib strain obtained from the simulation in S1.1, fiber optic gratings or strain gauges are pasted in the form of strain rosettes at the root of the ribs of the cage and the key area of ​​the transition rounded corner of the support arm. The sampling frequency is ≥1kHz to match the dynamic load frequency, and multi-channel strain signals are collected. S3.2, by applying excitation forces of different magnitudes and orientations to the shaft through an electromagnetic vibrator, the shaft will be displaced after being subjected to the excitation force, and the bearing will also be subjected to force, which will be transmitted to the squirrel cage elastic support system. The force sensor in the outer ring clamp can detect the load data. S3.3, The collected strain signals are cleaned and preprocessed to obtain the data input into the neural network; S3.4 Construct and train a deep convolutional neural network model embedded with LSTM units, where the strain signal is the input and the load is the output; by continuously adjusting the parameters of the model, the load recognition rate is made ≥95%, and the inverse mapping from the strain on the rib to the load on the squirrel cage spring support structure is completed.

2. The structural health prediction method based on rat cage strain measurement according to claim 1, characterized in that, S1 includes the following steps: S1.1 Measure the actual dimensions of the squirrel cage spring support structure, establish a simulation model in ANSYS, apply loads to the squirrel cage spring support structure according to the actual load-bearing position, obtain the overall strain / deformation, and at the same time obtain the position of maximum strain on the rib, providing ideas for attaching strain gauges in subsequent experiments. S1.2 In the response surface design module of ANSYS Workbench, using the custom design in the experimental design, multiple sets of loads with different magnitudes and directions are set to obtain the corresponding stress / deformation, providing training data for subsequent training of the forward mapping model.

3. The structural health prediction method based on rat cage strain measurement according to claim 1, characterized in that, S2 includes the following steps: S2.1, Based on S1.2, obtain a large amount of force and strain data, and preprocess the data; S2.2, Construct and train the XGboost model, using the load on the squirrel cage elastic support system as input and the overall stress / deformation of the squirrel cage elastic support system as output, establish a positive mapping model from the load of the squirrel cage elastic support structure to the overall stress / deformation, and continuously adjust the parameters to achieve a prediction accuracy of ≥95%; S2.3, Tensor decomposition technology is selected to compress the model parameter scale, compressing the number of tens of thousands of nodes, speeding up the model's response time, and reducing the time required for single deformation prediction.

4. The structural health prediction method based on rat cage strain measurement according to claim 1, characterized in that, The process of constructing the end-to-end nonlinear mapping relationship from structural load input to overall deformation output using the XGBoost model in machine learning is as follows: 1) Multi-source heterogeneous data acquisition and preprocessing includes: Data Acquisition: A multiphysics finite element model of the squirrel cage spring support structure was established using ANSYS Workbench, and a dynamic load spectrum was set, including impact loads, random vibrations, and alternating load conditions. Nodal stresses were generated using the modal superposition method and transient dynamics analysis module. ,strain: Deformation amount: time-series simulation data of x / y / z and environmental parameters, with the sampling frequency set to 10kHz to capture high-frequency dynamic response; Data cleaning: A sliding window mechanism was used to segment the time series data, with a window length of 1024 points and a step size of 512 points; 3 The criteria identify and eliminate stress mutation outliers, with a threshold set at the mean ± 3 times the standard deviation. Feature engineering and extraction: Construct a multi-dimensional normalization processing flow; perform Min-Max normalization on the mechanical signal, with the range set to [0,1]; ； in, By calculating mutual information entropy, key features with a correlation degree greater than 0.8 with deformation were selected; a deep autoencoder network was constructed, with an encoder structure of 5 fully connected layers, LeakyReLU activation function, and a compression dimension of 64; the decoder adopted a symmetric structure. ； Where x is the input value; It is a hyperparameter, called the "leaking function"; The nonlinear feature extraction capability is enhanced by an adversarial training strategy. The discriminator adopts the PatchGAN architecture and takes the reconstruction error as the optimization objective. The final output is a low-dimensional representation vector with a dimension of 64. ； The actual value is The predicted value is ; 2) XGBoost gradient boosting decision tree modeling and optimization: A high-precision structural load-deformation mapping model is constructed through an ensemble learning framework. The specific implementation process is as follows: Model Building and Training: During the model building and training phase, the XGBoost model is first initialized, and relevant parameters such as objective, n_estimators, learning_rate, and max_depth are set. tree_method=hist is used to suit large-scale data, early_stopping_rounds=50 is set, and scale_pos_weight is set according to the class imbalance. During training, GPU acceleration is enabled, tree_method=gpu_hist is set, and CUDA parallel computing is used to improve efficiency. At the same time, explicit and implicit regularization strategies are used to prevent overfitting and ensure the model's generalization ability and performance. Hyperparameter optimization: In the hyperparameter optimization stage, the parameter search space is first determined: learning_rate ranges from [0.01, 0.3], max_depth ranges from [3, 10], subsample ranges from [0.6, 1.0], and reg_alpha ranges from [0, 10]. A random search optimization method is used for different parameters. The optimization process adopts a two-stage approach. First, initial parameters are set based on domain experience. Then, the Optuna framework is used to perform Bayesian optimization for 50 iterations with the MSE of 5-fold cross-validation as the optimization objective. An early stopping mechanism is set, and training is terminated when the MSE of the validation set does not decrease for 50 consecutive rounds. The optimal combination of hyperparameters is found to improve model performance. Model Evaluation and Validation: In the model evaluation and validation phase, the coefficient of determination R² is used as the core evaluation indicator, while also ensuring that the maximum absolute error (MAE) is less than 5% of the allowable deformation. ； The actual value is The predicted value is To verify the robustness of the model, adversarial example testing was conducted. Gaussian noise was added to the input features to evaluate the model's stability. The generalization ability across different working conditions was verified through the CLUE benchmark set to ensure the reliability and accuracy of the model in different scenarios and meet the needs of practical engineering applications. 3) Tensor decomposition parameter compression: A model compression strategy based on tensor encoding and CP decomposition is adopted to optimize model parameters and improve the model's computation speed. The specific implementation process is as follows: First, the tree structure weights and leaf node scores in the XGBoost model are systematically tensor-encoded, and these parameters are organized and transformed into a third-order tensor, where the dimensions correspond to the number of trees, the feature dimension, and the output dimension, respectively, thus laying the foundation for subsequent tensor decomposition operations. Then, using the CP decomposition technique, the constructed third-order parameter tensor is decomposed into a combination of several rank-1 tensors. The factor matrix is ​​iteratively optimized using the alternating least squares method to gradually approximate the structure and information of the original tensor, and finally the model parameter size is successfully compressed to 1 / 10 of the original size. An LSTM-based prediction network is constructed to dynamically compensate for the residuals between the output of the original XGBoost model and the output of the compressed lightweight tensor model. After obtaining the predicted stress and deformation values, code is written to call the VTK library and utilize its visualization rendering function to intelligently map and refine the discrete stress and deformation data according to the spatial distribution law of the physical field. The code drives the VTK library to build data-color and shape association rules, transforming abstract values ​​into intuitive stress and deformation cloud maps by using specific color gradients corresponding to different numerical ranges.

5. The structural health prediction method based on rat cage strain measurement according to claim 1, characterized in that, Experimental data on the loads and strains on the ribs of the squirrel cage spring support structure under working conditions were obtained. A deep convolutional neural network model with embedded LSTM units was constructed and trained to achieve the inverse mapping from the strain on the ribs to the loads on the squirrel cage spring support structure. By continuously adjusting the model parameters, the load recognition rate was made ≥95%. 1) Construction of a multimodal experimental data acquisition system: This involves preparing neural network training data by acquiring data from force sensors in a ring fixture and strain gauges on the ribs of the mouse cage. The specific implementation process includes: Hardware preparation and installation: Install force sensors at key locations on the ring fixture and attach strain gauges to the ribs of the squirrel cage to complete the sensor deployment; then configure a multi-channel synchronous acquisition card and signal conditioning module to ensure that the sensors and acquisition equipment are correctly connected, then fix the fixture and calibrate the loading system to eliminate initial deformation interference; Data Synchronization Acquisition: The main control unit triggers a signal to achieve time synchronization of force sensor and strain gauge data; then, an appropriate sampling rate is set to cover the highest frequency of the signal, while the channel gain and filtering range are configured, and the signal stability is monitored in real time to eliminate the influence of environmental noise such as electromagnetic interference. Data preprocessing: The raw data is filtered and denoised to remove high-frequency noise and baseline drift; outliers such as sensor over-range and signal disconnection are removed; and the force and strain data of cross-modal data are aligned on the time axis to ensure data synchronization. Database construction: Archive the raw data by time or experiment number, extract features and store them in the database; at the same time, save the experimental conditions such as temperature, loading parameters and equipment status metadata information, and use manual or semi-automatic annotation of key events, including buckling points and fatigue thresholds; System verification and optimization: Functional testing is completed by verifying sensor accuracy, such as force sensor calibration and checking data synchronization errors; neural networks are trained using the collected data to evaluate the effectiveness of features, and then the sensor layout or acquisition parameters are iteratively improved based on the verification results; 2) Construction and training of deep convolutional neural network models: A deep convolutional neural network model embedding LSTM units is built, and the processed experimental data is used as training data to complete the model training. The specific implementation process includes: Data preparation: The preprocessed data from step 1) is divided into training, validation, and test sets using stratified sampling, with the training and validation sets comprising 80% and the test set comprising 20%. Then, the training and validation sets are further divided within the training and validation sets using stratified sampling, with the training set comprising 80% and the test set comprising 20%. Model architecture design: First, a CNN feature extraction part is constructed, using processed strain data as input and load data as output. Local features are extracted using 1D convolutional layers, and dimensionality is reduced using pooling layers. Next, LSTM units are embedded for temporal modeling. The feature sequence output by the CNN is input into the LSTM layer, and the return sequences parameter is set to preserve intermediate temporal states. An attention mechanism can be optionally added to strengthen the weights of key time steps. Finally, the output layer is designed, using a fully connected layer to output the predicted value. During implementation, it is important to add a Dropout layer after the convolutional layer for regularization to prevent overfitting. Gradient clipping is used during the training phase to limit gradient explosion. Model training configuration: Huber Loss is used to improve robustness; Adam is selected as the optimizer, with a learning rate of lr=1e-4 and weight decay added; in terms of training parameters, the batch size is set to 64~128, the number of training epochs is set to 100~200 and an early stopping mechanism is enabled, and training is terminated when the validation set loss does not decrease for 5 consecutive epochs. Model Validation and Iteration: Using the mean absolute error (MAE) as the evaluation metric, the model is optimized and iterated based on the validation results by adjusting the number of LSTM layers or hidden units, increasing the time step, and improving the sensor fusion strategy to improve model performance. ； The actual value is The predicted value is ; 3) Validation of deep convolutional neural network models: This involves testing the trained model to verify its quality and continuously refining it based on the training results. The specific implementation process includes: Preparation before verification: Divide the data reasonably to ensure that the test set data is independent and has not been used for training and parameter tuning. Divide the time series data according to the time order, and at the same time ensure that the test set covers different working conditions. Model performance evaluation: When conducting quantitative evaluation, test data is input into the model to obtain predicted values, evaluation indicators are calculated, and the dynamic alignment between the model's predicted time series curve and the actual value is compared using the DTW dynamic time warping distance. Iterative optimization process: When the validation set error is high, check whether the model is underfitting or the data distribution is off-target; if the test set has poor generalization ability, increase data diversity; adopt online incremental training, add newly collected data to the training set, fine-tune the model periodically, freeze some layers to prevent overfitting, and execute in the following order: data partitioning → indicator definition → quantitative testing → qualitative analysis → error localization → model tuning → re-validation.

Images