Metal rubber residual life prediction method based on physical information-time sequence convolutional network

Abstract

The invention relates to a metal rubber residual life prediction method based on a physical information-time sequence convolutional network, and in the proposed method, physical degradation information such as a metal rubber structure loss factor and kinetic parameters can effectively reflect the characteristics of a metal rubber degradation trend. The quality of a network input space can be improved by fusing the data into a time sequence convolutional network, and meanwhile, a mixed loss function of a network model is constructed by adopting a metal rubber comprehensive damage Weibull equation, so that the correlation of a network feature layer is improved. A time sequence convolutional network method of physical information is expanded to life prediction of a material with a nonlinear hysteresis characteristic, and meanwhile, a new thought is provided for residual life prediction modeling of metal rubber.

Description

Technical Field

[0001] The invention relates to the technical field of metal rubber, and in particular to a method for predicting the remaining service life of metal rubber based on a physical information-temporal convolutional network. Background Art

[0002] Metal rubber is a key vibration damping material for high-end equipment. However, due to its complex preparation process, the internal spatial structure of each sample is different. Traditional life prediction methods often use experiments and empirical formulas, which have problems such as insufficient universality and poor prediction accuracy. The life prediction method of the time series convolutional network can couple the comprehensive characteristics of fatigue life data of multiple metal rubbers into the prediction model, which can solve the problem of insufficient universality of metal rubber life prediction to a certain extent. However, the life prediction accuracy of the time series convolutional network has strict requirements on the amount of data from fatigue tests. Insufficient data often leads to a significant decrease in prediction accuracy. It is urgent to combine the physical information of the metal rubber degradation process with the time series convolutional network, and construct a method for predicting the remaining life of metal rubber based on the time series convolutional network (PI-TCN) of the physical information of metal rubber fatigue damage. Summary of the invention

[0003] In view of this, the purpose of the present invention is to provide a metal rubber remaining life prediction method based on physical information-temporal convolutional network, which utilizes the physical information of the metal rubber degradation process to enhance the input space of the temporal convolutional network and optimize the gradient direction, and utilizes the ability of the temporal convolutional network (TCN) to couple multiple metal rubber degradation characteristics, thereby improving the problem of insufficient versatility of traditional metal rubber life prediction and improving the accuracy of metal rubber life prediction.

[0004] To achieve the above object, the present invention adopts the following technical solution: a method for predicting the remaining life of metal rubber based on physical information-temporal convolutional network, comprising the following steps:

[0005] Step S1: Degradation information of the metal rubber structure loss factor, the formula is as follows:

[0006]

[0007] In the formula, ΔW is the energy consumption of the metal rubber in one stress cycle, and W is the maximum energy storage of the metal rubber;

[0008] Step S2: constructing a nonlinear elastic restoring force model, a nonlinear damping force model, and a hysteresis damping force model according to a traditional dynamics model, and decomposing the hysteresis damping force using Chebyshev polynomials;

[0009] Step S3: using a damage factor to characterize the damage degree of the metal rubber dynamic parameters;

[0010] Step S4: Multiple dynamic parameter damage factors are coupled to construct a metal rubber dynamic comprehensive damage factor, the formula of which is as follows:

[0011] d s =P i d i (2)

[0012] Where, d s is the comprehensive damage factor, d i is the damage factor of each kinetic parameter, P i is the contribution rate of each dynamic damage factor;

[0013] Step S5: Use a two-parameter Weibull function to perform curve fitting on the mean value of the metal rubber comprehensive damage factor, and the formula is as follows:

[0014]

[0015] Where: N is the number of cycles, λ and β are parameters related only to the loading conditions and the properties of the metal rubber itself, and λ and β are regarded as parameters to be identified;

[0016] Step S6: Based on the Weibull damage function and the mean square error function, a physical information loss function is constructed, and its formula is as follows:

[0017]

[0018] In the formula, y i Is the real life label, is the predicted lifespan label;

[0019] Step S7: construct the remaining service life RUL to characterize the fatigue life of the metal rubber, and the formula is as follows:

[0020] RUL(N)=N 0 -N (5)

[0021] Where λ is a hyperparameter. i Is the real life label, is the predicted lifespan label;

[0022] Step S8: Introduce the physical information of metal rubber degradation constructed in steps 1 and 2 into the input layer of the physical information-temporal convolutional network, and use the physical information loss function constructed in step 6 as the mixed loss function of the network to finally complete the construction of the physical information-temporal convolutional network.

[0023] In a preferred embodiment: the formula for decomposing the hysteresis damping force using Chebyshev polynomials in step S2 is as follows;

[0024]

[0025] In the formula, k 2i-1 is the nonlinear elastic restoring force F k (y) 2i-1 order stiffness coefficient, y 0 is the initial preload; c 2i-1 is the nonlinear damping force The 2i-1 order damping coefficient, N 2 is the maximum order of the nonlinear damping force. k s is the equivalent linear stiffness of the hysteresis damping force, y s is the equivalent maximum slip distance, z s is the memory restoring force when the metal wire undergoes relative slip, and sgn(*) is a symbolic equation.

[0026] In a preferred embodiment: in step S3, a damage factor is used to characterize the damage degree of the metal rubber dynamic parameter, and the formula is as follows:

[0027]

[0028] In the formula, is the tangent modulus of the metal rubber at the Nth cycle, is the initial tangent modulus of the metal rubber.

[0029] In a preferred embodiment: the damage equation of each parameter of metal rubber dynamics is:

[0030]

[0031] In the formula, k i (0) is the initial stiffness coefficient of the metal rubber at each order; k i (N) is the stiffness coefficient of each order after N cycles of loading, c i (0) is the initial damping coefficient of the metal rubber; c i (N) is the stiffness coefficient of each order after N cycles of loading, k s (0) is the equivalent linear stiffness of the initial hysteresis damping force of the metal rubber; z s (0) is the memory recovery force of the metal rubber during initial relative sliding; k s (N) is the equivalent linear stiffness of the hysteresis damping force after N cycles of loading; z s (N) is the memory recovery force during relative slip after N cycles of loading;

[0032] The construction of the comprehensive damage factor mainly adopts PCA principal component analysis:

[0033]

[0034] Where: x iExpressed as the kinetic parameter damage factor, μ i and σ i Represent the mean and standard deviation of the column respectively, X is the sample matrix of principal component analysis, R is the covariance matrix, X T is the transpose of X;

[0035] The mean square error loss function:

[0036]

[0037] In the formula, y i Is the real life label, is the predicted lifespan label;

[0038] Construction of the hybrid loss function:

[0039]

[0040] Where λ is a hyperparameter.

[0041] Compared with the prior art, the present invention has the following beneficial effects: the present invention expands the input space of the physical-based temporal convolutional network by using the physical information of the energy dissipation and dynamic parameter degradation in the fatigue damage process of metal rubber, and further constructs the mixed loss function of the network of the metal rubber comprehensive damage factor Weibull equation to constrain the optimization boundary of the physical-based temporal convolutional network (PI-TCN). Finally, a life prediction method of the physical-based temporal convolutional network (PI-TCN) is proposed to achieve effective prediction of the remaining life of metal rubber under finite samples. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Schematic diagram of a system according to a preferred embodiment of the present invention.

[0043] Figure 2 A test tool diagram of a preferred embodiment of the present invention;

[0044] Figure 3 A schematic diagram of the upper and lower ranges of the predicted true value of the PI-TCN model in the preferred embodiment of the present invention;

[0045] Figure 4 A schematic diagram of channel correlation in a preferred embodiment of the present invention;

[0046] Figure 5 This is a heat map of the convolution dimension and channel of a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0047] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0048] It should be noted that the following detailed descriptions are illustrative and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0049] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application; as used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or their combinations.

[0050] This implementation provides a life prediction method based on a physical-based temporal convolutional network, such as Figure 1 Said, comprising the following steps:

[0051] Step S1: Degradation information of the metal rubber structure loss factor, the formula is as follows:

[0052]

[0053] Where ΔW is the energy consumption of the metal rubber in one stress cycle, W is the maximum energy storage of the metal rubber,

[0054] Step S2: construct a nonlinear elastic restoring force model, a nonlinear damping force model, and a hysteresis damping force model based on the traditional dynamics model, and decompose the hysteresis damping force using Chebyshev polynomials, the formula of which is as follows;

[0055]

[0056] In the formula, k 2i-1 is the nonlinear elastic restoring force F k (y) 2i-1 order stiffness coefficient, y 0 is the initial preload. 2i-1 is the nonlinear damping force The 2i-1 order damping coefficient, N 2 is the maximum order of the nonlinear damping force. k s is the equivalent linear stiffness of the hysteresis damping force, y s is the equivalent maximum slip distance, z s is the memory restoring force when the metal wire undergoes relative slip, and sgn(*) is a symbolic equation.

[0057] Step S3: using a damage factor to characterize the damage degree of the metal rubber dynamic parameters, the formula is as follows;

[0058]

[0059] In the formula, is the tangent modulus of the metal rubber at the Nth cycle, is the initial tangent modulus of the metal rubber.

[0060] Step S4: Since the dynamic damage of metal rubber is the coupling effect of multiple dynamic parameter damage factors, only considering the change trend of a single dynamic damage factor cannot effectively reflect the degree of fatigue damage of metal rubber. Therefore, it is necessary to couple multiple dynamic parameter damage factors to construct a comprehensive dynamic damage factor of metal rubber, and its formula is as follows:

[0061] d s =P i d i (4)

[0062] Where, d s is the comprehensive damage factor, d i is the damage factor of each kinetic parameter, P i is the contribution rate of each dynamic damage factor.

[0063] Step S5: Use a two-parameter Weibull function to perform curve fitting on the mean value of the metal rubber comprehensive damage factor, and the formula is as follows:

[0064]

[0065] Where: N is the number of cycles, λ and β are parameters that are only related to the loading conditions and the properties of the metal rubber itself, and λ and β are regarded as parameters to be identified.

[0066] Step S6: Based on the Weibull damage function and the mean square error function, a physical information loss function is constructed, and its formula is as follows:

[0067]

[0068] In the formula, y i Is the real life label, is the predicted lifespan label.

[0069] Step S7: Construct the remaining useful life (RUL) to characterize the fatigue life of the metal rubber (data label), and the formula is as follows:

[0070] RUL(N)=N 0 -N (7)

[0071] Where λ is a hyperparameter. i Is the real life label, is the predicted lifespan label.

[0072] Step S8: Introduce the physical information of metal rubber degradation constructed in steps 1 and 2 into the input layer of the physical information-temporal convolutional network, and use the physical information loss function constructed in step 6 as the mixed loss function of the network to finally complete the construction of the physical information-temporal convolutional network.

[0073] In this embodiment, 10 groups of metal rubber materials are subjected to cyclic load fatigue tests. The test system is mainly composed of a hydraulic station, a universal testing machine and a data acquisition system. The test tooling is as follows: Figure 2 As shown. Made of 304 stainless steel with an outer diameter of 20mm, an inner diameter of 9mm, a height of 16mm, and a density of 1.25g / cm 3 The preparation process parameters of the annular metal rubber sample are shown in Table 1. The upper chuck is driven to perform fatigue test on the metal rubber sample with a sinusoidal displacement excitation displacement with an input frequency of 10 Hz and an amplitude of 1 mm. The data acquisition adopts the interval sampling method, and the test data is collected every 2000 stress cycles. The sampling frequency of the test is set to 2560.

[0074] Table 1 Characteristics and preparation parameters of metal rubber

[0075]

[0076] Substitute the test set data into the three models for prediction comparison. The prediction accuracy of the Weibull model is R 2 =0.405, while the R 2 The prediction accuracy of the TCN model is only 0.2181, and the prediction accuracy of the TCN model is reduced by 18.7%. This is because the use of the TCN model to predict the remaining life of metal rubber is a pure data prediction method. This method has a great dependence on the quality and quantity of data. In the case of a small data set (set has only 10 groups), the prediction of the TCN model is difficult to achieve the desired effect, and its fitting effect is even worse than that of the traditional model. PI-TCN model R 2 =0.967. Compared with the traditional Weibull model, the prediction accuracy of the PI-TCN model is improved by 56.2%.

[0077] Table 2 Comparison of training results of three prediction methods

[0078]

[0079] The PI-TCN model can significantly reduce the training loss value of the model within 50 steps, while the TCN model needs more than 50 steps before the model training will show a downward trend. In the initial cyclic load period, the fitting results of the Weibull model are closer to the true value, but as the cyclic load time increases, the fitting results begin to gradually deviate from the true value, and the overall prediction results of the Weibull model are higher than the actual value. There is also a significant deviation between the prediction results of the TCN model and the true value of the metal rubber damage. The prediction accuracy of the PI-TCN model is higher, and it fluctuates around the upper and lower range of the true value, such as Figure 3 shown.

[0080] The significance of the physical information features in the PI-TCN model is analyzed, a channel information model is established in the training network model, and a channel-dimensional heat map is generated. The heat map can effectively characterize the correlation between the dimensional information and the channel. When the dimensional information is more correlated in the channel, the normalized value of the area is closer to 1, and when the dimensional information is less correlated in the channel, the normalized value of the area is closer to 0. Among the physical information feature information, only feature 26 has a low correlation to all channels, and the remaining physical features have a high correlation in the channel, among which the first-order stiffness degradation information (feature 21) has the highest correlation, such as Figure 4 shown.

[0081] To further study the influence of physical information in the convolution layer, we selected the input layer as the input and the middle convolution layer as the output, trained the model and drew the heat map of the convolution dimension and channel. The PI-TCN model improved both the convolution feature dimension and the number of channels. The number of correlated convolution features increased from 7 dimensions to 8 dimensions, and the number of channels also increased by one (channel 1). Figure 5 This is because the hybrid loss function can increase the spatial boundary constraints after convolution processing, optimize the convolution features, increase the correlation between the convolution features and the channels, and finally optimize the output space and improve the prediction results of the model.

[0082] The above description is only a preferred embodiment of the present invention. All equivalent changes and modifications made according to the scope of the patent application of the present invention should fall within the scope of the present invention.

Claims

1. A method for predicting the remaining life of metal rubber based on physical information-temporal convolutional network, characterized by: The following steps are involved: Step S1: Degradation information of the metal rubber structure loss factor, the formula is as follows: In the formula, ΔW is the energy consumption of the metal rubber in one stress cycle, and W is the maximum energy storage of the metal rubber; Step S2: constructing a nonlinear elastic restoring force model, a nonlinear damping force model, and a hysteresis damping force model according to a traditional dynamics model, and decomposing the hysteresis damping force using Chebyshev polynomials; Step S3: using a damage factor to characterize the damage degree of the metal rubber dynamic parameters; Step S4: Multiple dynamic parameter damage factors are coupled to construct a metal rubber dynamic comprehensive damage factor, the formula of which is as follows: d s =P i d i (2) Where, d s is the comprehensive damage factor, d i is the damage factor of each kinetic parameter, P i is the contribution rate of each dynamic damage factor; Step S5: Use a two-parameter Weibull function to perform curve fitting on the mean value of the metal rubber comprehensive damage factor, and the formula is as follows: Where: N is the number of cycles, λ and β are parameters related only to the loading conditions and the properties of the metal rubber itself, and λ and β are regarded as parameters to be identified; Step S6: Based on the Weibull damage function and the mean square error function, a physical information loss function is constructed, and its formula is as follows: In the formula, y i Is the real life label, is the predicted lifespan label; Step S7: construct the remaining service life RUL to characterize the fatigue life of the metal rubber, and the formula is as follows: RUL(N)=N0-N (5) Where λ is a hyperparameter. i Is the real life label, is the predicted lifespan label; Step S8: Introduce the physical information of metal rubber degradation constructed in steps 1 and 2 into the input layer of the physical information-temporal convolutional network, and use the physical information loss function constructed in step 6 as the mixed loss function of the network to finally complete the construction of the physical information-temporal convolutional network.

2. The method for predicting the remaining service life of metal rubber based on physical information-temporal convolutional network according to claim 1 is characterized by: The formula for decomposing the hysteresis damping force using Chebyshev polynomials in step S2 is as follows: In the formula, k 2i-1 is the nonlinear elastic restoring force F k (y) is the 2i-1 order stiffness coefficient, y0 is the initial preload; c 2i-1 is the nonlinear damping force The 2i-1 order damping coefficient, N2 is the maximum order of the nonlinear damping force. s is the equivalent linear stiffness of the hysteresis damping force, y s is the equivalent maximum slip distance, z s is the memory restoring force when the metal wire undergoes relative slip, and sgn(*) is a symbolic equation.

3. The method for predicting the remaining service life of metal rubber based on physical information-temporal convolutional network according to claim 1 is characterized by: In step S3, a damage factor is used to characterize the damage degree of the metal rubber dynamic parameters, and the formula is as follows: In the formula, is the tangent modulus of the metal rubber at the Nth cycle, is the initial tangent modulus of the metal rubber.

4. The method for predicting the remaining service life of metal rubber based on physical information-temporal convolutional network according to claim 1 is characterized in that: Damage equations for various parameters of metal rubber dynamics: In the formula, k i (0) is the initial stiffness coefficient of the metal rubber at each order; k i (N) is the stiffness coefficient of each order after N cycles of loading, c i (0) is the initial damping coefficient of the metal rubber; c i (N) is the stiffness coefficient of each order after N cycles of loading, k s (0) is the equivalent linear stiffness of the initial hysteresis damping force of the metal rubber; z s (0) is the memory recovery force of the metal rubber during initial relative sliding; k s (N) is the equivalent linear stiffness of the hysteresis damping force after N cycles of loading; z s (N) is the memory recovery force during relative slip after N cycles of loading; The construction of the comprehensive damage factor mainly adopts PCA principal component analysis: Where: x i Expressed as the kinetic parameter damage factor, μ i and σ i Represent the mean and standard deviation of the column respectively, X is the sample matrix of principal component analysis, R is the covariance matrix, X T is the transpose of X; The mean square error loss function: In the formula, y i Is the real life label, is the predicted lifespan label; Construction of the hybrid loss function: Where λ is a hyperparameter.

Images