Life prediction method based on health index construction and neural network fusion

Abstract

The invention discloses a life prediction method based on health index construction and neural network fusion, and belongs to the technical field of equipment state monitoring and predictive maintenance. According to the method, through multi-source degradation feature extraction, common dynamic principal component analysis (CDPCA) dimensionality reduction, health index construction and normalization, deep learning multi-model modeling, integrated learning fusion and Bayesian optimization hyper-parameter optimization, online health assessment and residual life prediction of the equipment part degradation process are realized. Specifically, the method comprises the following steps: firstly, extracting time domain, frequency domain and time-frequency domain features from a sensor acquisition signal, and performing dimension reduction through CDPCA to obtain effective degradation characterization; then, weighting the main features to construct a health index (HI) curve, optimizing the weight through a genetic algorithm, and then performing normalization; a plurality of neural network models such as CNN, Bi-GRU, Bi-RNN, Bi-LSTM and SRNN are constructed based on the normalized HI sequence, and degradation trend modeling is realized; inputting the output results of the neural networks into an integrated learning module for fusion optimization; and finally, carrying out automatic optimization on the key hyper-parameters of the model by utilizing Bayesian optimization. In the equipment operation process, a normalized HI curve can be calculated in real time and input into the fusion model, and the residual life estimation value of the part is dynamically output. According to the method, high-precision, high-robustness and online life prediction can be provided under complex working conditions, the safety and reliability of equipment operation and maintenance are improved, and the method has wide engineering application value.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of equipment state monitoring and predictive maintenance, and in particular to a component-level health assessment and residual life prediction method combining degradation feature extraction, health index construction, deep learning modeling, ensemble learning and Bayesian optimization. BACKGROUND

[0002] With the development of modern industrial equipment towards large-scale, complex and intelligent direction, the health state and life prediction of its key components have become an important research direction for equipment operation and safety guarantee. Traditional life prediction methods mainly rely on modeling methods based on physical models or statistical models, but these methods often have the following shortcomings:

[0003] (1) Strong model dependence: physical models require a deep understanding of the degradation mechanism of the equipment, but the degradation mechanism is difficult to completely grasp under actual complex working conditions;

[0004] (2) Lack of robustness: traditional statistical methods such as regression analysis and Markov model are sensitive to noise, and have limited prediction accuracy under nonlinear degradation and multiple working conditions;

[0005] (3) Difficulty in parameter selection: deep learning methods have been used for life prediction, but their hyperparameter optimization relies on experience or brute force search, which is prone to local optimization and difficult to obtain global optimal solution.

[0006] Therefore, there is an urgent need for a new method that can fuse degradation feature construction to build health index, introduce multiple neural network models and combine ensemble learning and Bayesian optimization to improve prediction accuracy and generalization ability, and realize real-time and high-precision prediction of equipment life. SUMMARY

[0007] The present application aims to provide a life prediction method based on health index construction and neural network fusion, Figure 1 The overall flow framework diagram of the life prediction method of the present application is shown in FIG. Figure 1 The present application includes the following specific steps:

[0008] S1: Degradation feature extraction and dimensionality reduction: extract multi-domain degradation features from sensor signals and use common dynamic principal component analysis (CDPCA) for dimensionality reduction to obtain effective degradation representation.

[0009] S2: Health index construction and normalization, construct health index (HI) curve by weighted fusion of main features, and normalize after optimizing weights with genetic algorithm to ensure consistency between different working conditions.

[0010] S3: Deep learning modeling, based on normalized HI sequence, multiple neural network models such as CNN, Bi-GRU, Bi-RNN, Bi-LSTM and SRNN are constructed respectively to realize degradation trend modeling and life prediction.

[0011] S4: Integrated learning fusion, the prediction results of multiple deep learning sub-models are input into an integrated learning framework (such as AdaBoost or random forest) to obtain more robust prediction output.

[0012] S5: Bayesian optimization of hyperparameters, the Bayesian optimization method of Gaussian process and expected improvement acquisition function is adopted to automatically search and determine the optimal hyperparameter configuration of the model.

[0013] S6: Online health assessment and life prediction, during the operation of the device, the HI curve is calculated and normalized in real time, and the remaining useful life (RUL) of the component is dynamically output by inputting the fusion model.

[0014] Compared with the prior art, the present application has the following significant advantages:

[0015] (1) More reasonable feature construction: through CDPCA dimension reduction and genetic algorithm optimized weight, a health index with monotonicity, robustness, consistency and time correlation is constructed;

[0016] (2) More comprehensive modeling capability: multiple networks such as CNN, Bi-GRU, Bi-RNN, Bi-LSTM and SRNN are complementary, which can accurately capture the features of different stages of the degradation process;

[0017] (3) More robust fusion prediction: the AdaBoost or random forest integrated model is introduced, which significantly improves the prediction accuracy and stability in multiple working conditions and noisy environment;

[0018] (4) More intelligent hyperparameter optimization: through Bayesian optimization combined with Gaussian process and expected improvement acquisition function, the optimal hyperparameter combination is efficiently searched within a limited iteration;

[0019] (5) More practical real-time prediction: the health index curve is updated in real time during the operation of the device, and the remaining useful life is dynamically output, which meets the online application requirements in actual engineering. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 : The overall flow framework diagram of the life prediction method of the present application.

[0021] Figure 2 : Degradation feature extraction and common dynamic principal component analysis (CDPCA) dimension reduction process schematic diagram.

[0022] Figure 3 : Health index construction and normalization process schematic diagram.

[0023] Figure 4 : Deep learning prediction model (CNN, Bi-GRU, Bi-RNN, Bi-LSTM, SRNN) structure diagram.

[0024] Figure 5 : Integrated learning fusion framework diagram.

[0025] Figure 6 : Bayesian optimization hyperparameter optimization process diagram.

[0026] Figure 7 : Online prediction and life estimation operation diagram. DETAILED DESCRIPTION

[0027] In order to realize high-precision health assessment and residual life prediction of equipment components, the present application proposes a life prediction method based on health index construction and neural network fusion. Based on multi-source degradation features, the method sequentially completes feature extraction and dimension reduction, health index construction and normalization, deep learning multi-model modeling, integrated fusion and Bayesian hyperparameter optimization, thereby realizing real-time and accurate prediction of equipment life. Through phased optimization and multi-model collaboration, the method not only significantly improves the prediction accuracy, but also effectively enhances the robustness and generalization ability of the model under different working conditions and equipment individual differences, providing reliable technical support for health management and life prediction of complex industrial systems. The following will be described in detail in conjunction with the drawings.

[0028] S1: Degradation feature extraction and dimension reduction, the flowchart is shown in Figure 2 Firstly, time domain, frequency domain and time-frequency domain degradation features are extracted from the sensor signals collected by the equipment, including: time domain features such as mean, variance, standard deviation, skewness, kurtosis, etc.; frequency domain features such as spectral entropy, main frequency amplitude, frequency band energy, harmonic component, etc.; time-frequency domain features such as wavelet packet energy, Hilbert envelope spectrum, short-time Fourier transform features, etc. Subsequently, the above features are reduced in dimension by using the common dynamic principal component analysis method (CDPCA). First, construct an augmented matrix based on the idea of DPCA, use time delay method to construct an augmented matrix based on the original multi-feature data, and the autocorrelation of multi-features can be directly converted into the intercorrelation between variables that can be processed by PCA method. Then, construct a common projection matrix based on the idea of CPCA, convert the multi-time-frequency domain features to a common low-dimensional space by matrix projection to extract degradation features and obtain characteristic values λ, as shown in formula (1). Finally, principal component screening is performed, and then the value of p is gradually increased. When the variance contribution rate is large enough, the first p principal components are extracted as degradation features, as shown in formula (2). Select features with cumulative contribution rate greater than 95% to ensure that the retained features can fully reflect the degradation trend, while avoiding redundancy and noise interference.

[0029]

[0030] S2: Health index construction and normalization, the flowchart is shown in Figure 3 The main features after screening are obtained by weighted average to obtain the initial health index (HI) curve; then the weighted coefficients of each feature are optimized by using genetic algorithm, and the comprehensive performance of the HI curve in time correlation, monotonicity, robustness and consistency is maximized. Specifically, the four properties are quantitatively calculated by formulas (3)-(6), and a multi-index weighted fusion model is constructed as the fitness function of the genetic algorithm, so as to obtain the optimal feature weight. In the normalization process, firstly, the average HI value of the initial stage of the equipment is extracted from the training data as the upper bound HI max , and the average HI value of the fault stage is extracted as the lower bound HI min , and it is frozen as a unified normalization reference; then according to formula (7), the HI curve is mapped to the interval [0, 1] to ensure the consistency of the HI curve under different equipment or working conditions, and to provide standardized input for equipment remaining life prediction.

[0031]

[0032] S3: Neural network modeling, the flowchart is shown in Figure 4 After obtaining the normalized HI sequence, a plurality of deep learning prediction models are constructed based on the health index sequence for degradation trend modeling and remaining useful life (RUL) prediction. Specifically, the following five types of neural network structures are selected: Convolutional Neural Network (CNN): Taking advantage of its local receptive field and parameter sharing, it can effectively extract local degradation features in the HI sequence, and enhance the recognition ability of the model to early degradation patterns; Bidirectional Gated Recurrent Unit (Bi-GRU): Through the gating mechanism, the historical information and future information are modeled at the same time, which can capture the long-time dependence of the HI sequence, while maintaining high training efficiency; Bidirectional Recurrent Neural Network (Bi-RNN): Through the forward and backward propagation structure, the information before and after the time series is fully utilized to improve the fitting ability of the degradation trend; Bidirectional Long Short-Term Memory Network (Bi-LSTM): Combining the gating mechanism and memory unit, it can effectively solve the long-term dependence problem and perform well in predicting the late degradation trend and the life near the failure stage; Stacked Recurrent Neural Network (SRNN): By introducing multi-layer structured recurrent units (such as stacked LSTM / GRU layers), the complex dynamic characteristics (such as nonlinear trajectory, state mutation) in the equipment degradation process are explicitly modeled, which significantly improves the stability (resistance to noise interference) and robustness (adaptation to working condition fluctuations) of the prediction results.

[0033] The above multiple models have natural complementarity in feature extraction ability and time series modeling ability: (1) spatial-temporal ability complementarity: CNN is good at local spatial feature extraction (such as early degradation signs); the other four neural networks cooperatively strengthen long-term dependence modeling (such as whole-cycle degradation evolution). (2) Adaptability complementarity in degradation stages: in the stable period, CNN and the underlying cycle unit suppress high-frequency noise; in the acceleration period, high-level SRNN and Bi-LSTM capture nonlinear trends; in the failure critical period, Bi-GRU and the bidirectional structure accurately identify the end-of-life features. (3) In terms of joint benefits, by fusing the above models, the limitations of a single structure in a specific degradation stage (such as CNN's insensitivity to long-term trends and RNN's insufficient extraction of early weak features) are systematically overcome, achieving high-precision continuous modeling of the device degradation trajectory and highly reliable prediction of the remaining life.

[0034] Finally, the five neural network models output corresponding life prediction results, forming a vector containing five sub-prediction values. The prediction result vector will be input to the ensemble learning module in the subsequent steps for further fusion and optimization to obtain the final device life prediction result.

[0035] S4: ensemble model fusion, the flowchart is shown in Figure 5 After obtaining the prediction results of multiple deep learning sub-models, to further improve the accuracy and robustness of life prediction, the present application introduces an ensemble learning strategy to fuse the prediction results of the above sub-models. Ensemble learning can establish a complementary relationship between different models, avoid the adverse effects of the performance deficiency of a single model under certain working conditions on the overall prediction result, and thus achieve more robust prediction performance. Specifically, assuming that the life prediction values output by the five neural network models are:

[0036] R = {R CNN ,R BiGRU ,R BiRNN ,R BiLSTM ,R SRNN} (8)

[0037] Where R represents the prediction vector result, and each element corresponds to the predicted life value of a sub-model. The goal of the ensemble model is to learn a mapping function Fusion of multiple sub-model prediction values into the final life prediction result

[0038]

[0039] In terms of fusion methods, the present application preferably uses AdaBoost or Random Forest algorithm to realize model ensemble.

[0040] AdaBoost iteratively trains weak predictors and assigns different weight coefficients, so that each round of iteration pays more attention to difficult-to-predict samples, thereby gradually improving the overall prediction accuracy. The specific fusion form can be expressed as:

[0041]

[0042] Among them, h i (·) represents the i-th weak predictor, α i is its corresponding weight, and M is the total number of weak predictors in the ensemble. This method can dynamically adjust the weight distribution of each sub-model in different prediction scenarios and improve the ability to fit complex degradation laws.

[0043] Random forest ensembles build prediction models by integrating multiple decision trees and introducing randomness in the feature and sample dimensions. The final prediction value is obtained by weighted average or majority vote of the results of multiple trees:

[0044]

[0045] Among them, h t (·) represents the prediction result of the tth tree, where T is the total number of decision trees. This method can effectively reduce the risk of overfitting and enhance the model's generalization ability under different working conditions and noise environments.

[0046] Through this integration strategy, the present invention fully leverages the differentiated advantages of different neural networks, such as CNN, Bi-GRU, Bi-RNN, Bi-LSTM, and SRNN, in feature extraction and time series modeling. As a result, it maintains high prediction accuracy and stability even in complex and changing equipment operating environments, enabling more reliable and robust predictions of equipment lifespan.

[0047] S5: After completing the construction of the integrated model, in order to further improve the prediction performance and generalization ability of the model, the present invention introduces the Bayesian Optimization (BO) method to automatically optimize the key hyperparameters of the model. The flow chart is shown in Figure 6 As shown in Figure 2, traditional hyperparameter tuning methods (such as grid search or random search) not only rely on manual experience and are computationally expensive, but are also prone to local optimality, making it difficult to achieve a global optimal solution with limited resources. In contrast, Bayesian optimization can efficiently search the high-dimensional hyperparameter space within a limited number of iterations, gradually approaching the global optimal configuration, thereby achieving better model performance.

[0048] Specifically, the present application first uses a Gaussian Process (GP) as a surrogate model to probabilistically model the objective function f(θ) (i.e., the validation error or negative log-likelihood value of the model under the hyperparameter configuration θ). Assuming the observed data as Then, under Gaussian Process modeling, we have:

[0049] f(θ) ~ GP(μ(θ), k(θ, θ')) (12)

[0050] where μ(θ) is the mean function and k(θ, θ') is the covariance kernel function (such as the radial basis function RBF kernel).

[0051] Subsequently, the expected improvement is defined as the acquisition function to balance the exploration and exploitation relationship. In the hyperparameter space, the algorithm iteratively updates the surrogate model and the acquisition function to gradually approach the optimal solution. The EI is defined as follows:

[0052] EI(θ) = E[max(0, f best - f(θ))] (13)

[0053] where f best represents the currently known optimal objective function value. According to the predictive distribution N(μ(θ), σ 2 (θ)) of the Gaussian Process, EI can be specifically expressed as:

[0054]

[0055] where Φ(·) and are the cumulative distribution function and probability density function of the standard normal distribution, respectively. The algorithm selects the next set of hyperparameters to be evaluated by maximizing the acquisition function EI(θ) in the hyperparameter space, iteratively updates the surrogate model, and converges until convergence.

[0056] On specific optimization objects, the present application mainly optimizes the following key hyperparameters: the number of neurons, which controls the capacity and expressive ability of the network; the forgetting rate, which suppresses overfitting and improves model generalization ability; the learning rate, which determines the convergence speed and stability of the model; the regularization coefficient, which controls the parameter complexity and avoids overfitting; the batch size, which balances the computational efficiency and stability of model updates; and the training rounds, which ensure sufficient learning without overfitting.

[0057] Bayesian optimization in the present application realizes efficient global search of complex hyperparameter space and can automatically obtain the optimal parameter configuration scheme. Ultimately, the integrated prediction model exhibits stronger generalization ability, prediction accuracy, and stability under different working conditions and diverse data environments, significantly improving the practicality and reliability of equipment life prediction.

[0058] S6: In the actual operation of the device, the multi-source degradation feature signals collected by the sensor in real time are first subjected to health index (HI) calculation via the optimal feature weight obtained and frozen in the training stage. Subsequently, the calculated HI sequence is subjected to normalization processing in combination with the unified normalization upper and lower bounds determined in the training stage, so as to ensure that it is mapped to the interval [0, 1], thereby realizing index unification among different working conditions and different devices.

[0059] The normalized HI curve is sent as input into the deep learning and integrated fusion prediction model proposed in the application. The model comprehensively utilizes the feature extraction and time series modeling capabilities of multiple types of neural networks such as CNN, Bi-GRU, Bi-RNN, Bi-LSTM and SRNN, and realizes fusion and optimization of the prediction results through an integrated learning strategy (such as AdaBoost or random forest). On this basis, the model can dynamically track the degradation process of the device under online operation environment, and output the residual useful life (RUL) estimate value of the component in real time. The entire flowchart is shown in FIG. Figure 7 .

[0060] In summary, the application proposes a life prediction method based on health index construction and neural network fusion, which realizes high-precision, strong-robustness and online life prediction of device components under complex working conditions through feature extraction and dimension reduction, health index construction, deep learning modeling, integrated learning fusion and Bayesian optimization optimization. This method not only improves the practicality and reliability of the prediction, but also provides important technical support for device operation and health management, and has broad engineering application prospects.

Claims

1. A lifespan prediction method based on health index construction and neural network fusion, characterized in that: The following steps are involved: Degradation feature extraction and dimensionality reduction: Extract time-domain, frequency-domain, and time-frequency-domain degradation features from sensor signals and use common dynamic principal component analysis (CDPCA) for dimensionality reduction to obtain device degradation features. Health index construction and normalization: The degradation features are weighted and fused to construct a health index (HI) curve, and the feature weights are optimized through a genetic algorithm, and then the HI curve is normalized; Deep learning modeling: Based on the normalized HI sequence, we construct convolutional neural network (CNN), bidirectional gated recurrent unit (Bi-GRU), bidirectional recurrent neural network (Bi-RNN), bidirectional long short-term memory network (Bi-LSTM), and stacked recurrent neural network (SRNN) models to perform degradation trend modeling and remaining life prediction; Ensemble learning fusion: The prediction results of each neural network model are input into the ensemble learning framework, including AdaBoost or random forest algorithm, to obtain the final life expectancy prediction results; Bayesian optimization hyperparameter optimization: Utilizes the Bayesian optimization method of Gaussian process and expected improved acquisition function to automatically optimize the key hyperparameters of the model; Online health assessment and life prediction: During equipment operation, the HI curve is calculated and normalized in real time, input into the fusion model, and the estimated remaining life of the component is dynamically output.

2. The method according to claim 1, characterized in that The degradation features include time domain features, frequency domain features and time-frequency domain features, wherein the time domain features include mean, variance, standard deviation, skewness and kurtosis; the frequency domain features include spectral entropy, main frequency amplitude, frequency band energy and harmonic components; the time-frequency domain features include wavelet packet energy, Hilbert envelope spectrum and short-time Fourier transform features.

3. The method according to claim 1, characterized in that In the construction of the health index, the goal of the genetic algorithm optimization is to simultaneously maximize the monotonicity, robustness, consistency and time correlation of the HI curve.

4. The method according to claim 1, wherein The five neural networks of the deep learning model complement each other in feature extraction and time series modeling capabilities, and achieve high-precision prediction of different degradation stages through fusion.

5. The method according to claim 1, wherein The ensemble learning method improves the robustness of prediction results under multiple working conditions and noisy environments by dynamically allocating weights to each sub-model or voting average.

6. The method according to claim 1, characterized in that The key hyperparameters optimized by Bayesian optimization include the number of network neurons, learning rate, forgetting rate, regularization coefficient, batch size, and training rounds.

7. The method according to claim 1, characterized in that In the online health assessment and life prediction steps, the normalized HI sequence calculated in real time is input into the integrated fusion model, and the remaining life (RUL) of the component is dynamically output to achieve online monitoring and prediction.

Images