Method for predicting adaptive residual life of mechanical equipment in time-varying working condition domain

Abstract

The invention discloses a time-varying working condition domain adaptive residual life prediction method for mechanical equipment, and the method comprises the steps: firstly, training a basic prediction model through constant working condition full-life degradation data according to the feature distribution difference caused by degradation; secondly, optimizing an optimal domain weight according to time-varying working condition normal stage data, and weighting each working condition domain sample in a degradation stage; and finally, according to an optimal transmission theory, finely adjusting feature distribution of degradation to improve residual life prediction precision. According to the method, a normal-stage sample under the time-varying working condition is taken as a link, and the degradation characteristic distribution obtained under the time-varying working condition and the degradation characteristic distribution obtained under the constant working condition are matched, so that the mechanical equipment degradation rule under the time-varying working condition is modeled by utilizing the existing constant working condition degradation data, and an accurate residual life prediction task is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of remaining life prediction of mechanical equipment, and specifically relates to a method for predicting the remaining life of mechanical equipment in a time-varying operating condition domain. Background Art

[0002] Mechanical equipment is an important support for industrial production, and its safe and stable operation has attracted much attention. Research on the remaining life prediction of mechanical equipment is of great significance to improving equipment operation reliability, reducing operation and maintenance costs, and avoiding catastrophic accidents. However, in engineering practice, the complex and changeable operating environment and diverse working conditions of mechanical equipment lead to the time-varying and uncertainty of degradation data, which makes life prediction face huge challenges.

[0003] Most of the existing data-driven remaining life prediction methods focus on the degradation process under constant operating conditions. These methods can be roughly divided into two categories: the first category is that the training data and test data are in the same operating conditions ([1] B. Wang, Y. Lei, N. Li, W. Wang, Multiscale convolutional attention network for predicting remaining useful life of machinery, IEEE Transactions on Industrial Electronics, 68 (2020) 7496-7504), and the trained network model can be directly applied to the prediction task of the test set. The core is to accurately extract the features that characterize the degradation to construct a mapping relationship between degradation features and remaining life; the second category is the scenario where the training data and test data are in different constant operating conditions ([2] Ding Y, Jia M, Cao Y. Remaining useful life estimation under multiple operating conditions via deep subdomain adaptation [J]. IEEE Transactions on Instrumentation and Measurement, 2021, 70: 1-11), and methods such as transfer learning are required to overcome the differences in feature distribution caused by operating conditions. The accuracy of the prediction results also depends on the effect of feature migration.

[0004] However, during the degradation process of mechanical equipment, working condition conversion often occurs, causing internal components to operate under time-varying working conditions. For example, the speed and load of wind turbines fluctuate in real time due to changes in wind speed; heavy truck gearboxes have 8 to 12 gears and require frequent gear shifting to control the transmission gears; for these time-varying working condition scenarios, the remaining life prediction model established under constant working conditions cannot be directly applied. Therefore, there is an urgent need for a remaining life prediction method for mechanical equipment that is adaptive to the time-varying working condition domain. Summary of the invention

[0005] In order to overcome the defects of the above-mentioned prior art, the purpose of the present invention is to provide a method for predicting the remaining life of mechanical equipment in the time-varying working condition domain. The normal stage samples under the time-varying working condition are used as a link to adapt the degradation characteristic distribution obtained under the time-varying working condition and the constant working condition, so as to use the existing constant working condition degradation data to model the degradation law of mechanical equipment under the time-varying working condition and realize the accurate remaining life prediction task.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is:

[0007] A method for predicting the remaining life of mechanical equipment in a domain-adaptive manner under time-varying working conditions is proposed. Firstly, the basic prediction model is trained using the full-life degradation data under constant working conditions in view of the difference in characteristic distribution caused by degradation. Then, the optimal domain weight is optimized according to the normal stage data of the time-varying working conditions, and the samples of each working condition domain in the degradation stage are weighted. Finally, the characteristic distribution of degradation is fine-tuned according to the optimal transmission theory to improve the accuracy of remaining life prediction.

[0008] A method for predicting the remaining life of mechanical equipment in a time-varying operating condition domain, comprising the following steps:

[0009] Step 1: Obtain a full life sample under constant operating conditions Among them, M C represents the total number of full life samples, C represents the corresponding constant working condition; the normal stage samples are expressed as: Among them, M N is the number of samples in the normal stage, and the corresponding label value is 1; the samples in the degradation stage are represented as: The corresponding label is The degradation samples of time-varying working conditions are composed of samples of various working conditions. The degradation samples under working condition k are expressed as The sample label under working condition k is and Represent samples of its healthy state and degraded state respectively, Assume that their numbers are P N and P D, the total number of samples is P; samples under the same working condition have similar distribution in the feature space, while the distribution characteristics of samples under different working conditions are different, that is, P(X1|k)≈P(X2|k), P(X1|k1)≠P(X2|k2);

[0010] Step 2: Perform constant working condition prediction modeling for the original vibration signal Through multi-layer 1D-CNN feature extraction, a high-dimensional representation of the mechanical health status is obtained, as shown below:

[0011]

[0012] In the above formula, σ(·) is the activation function, S is the one-dimensional convolution kernel size, and W l represents the weight of the convolution kernel of the lth layer, b l is the bias of the lth layer;

[0013] The high-dimensional representations of L samples that are continuous in time are spliced ​​into continuous time series features, and the LSTM module is used to construct the time series relationship and predict the corresponding remaining life. The specific calculation process is as follows:

[0014] g(X) = LSTM(x t ,h t-1 ,U t-1 ),t=1,2,…,L

[0015] In the above formula, LSTM(·) is the use of LSTM module, x t is the spliced ​​continuous time series feature data, h t-1 is the hidden state at the previous moment, U t-1 is the memory state of the previous moment;

[0016] Network model for steady state conditions For example, its loss function is defined as:

[0017]

[0018] Step 3: Weighting the operating condition domain, when the prediction model of the constant operating condition is directly applied to the normal stage of the time-varying operating condition, a set of weights ω k For samples Weighted adjustment is performed to improve fitting accuracy. At this time, the input of the prediction model becomes The output becomes exist The weighted loss function is defined on the sample as:

[0019]

[0020] The above selection weight ω kThe process of improving fitting accuracy is regarded as solving an optimization problem, and the goal is to find the optimal weight sequence ω k , so that the prediction model The fitting error on is the smallest, that is:

[0021]

[0022] Afterwards, the optimal weight sequence ω k Applied to the corresponding working conditions of samples in the degradation stage of time-varying working conditions, the working condition domain of the samples is weighted, and the loss function at this time becomes:

[0023]

[0024] Considering the coupling effect between working condition and degradation, the working condition domain weight in the degradation process is designed to decay gradually, indicating that the influence of working condition on sample distribution gradually decays with degradation, and the domain weight gradually converges to 1. The designed decay function is as follows:

[0025]

[0026] Where ρ represents the attenuation coefficient;

[0027] Step 4: Perform online fine-tuning, adjust the distribution of features extracted by the CNN network according to the predicted value, and use the optimal transmission theory to implement the online fine-tuning process; the optimization process is as follows:

[0028]

[0029] Among them, γ∈Π(μ u ,μ v )={γ∈P(U×V)|p + #γ=μ u ,p - #=μ v}, p + and p - represents the mapping of the joint distribution U×V→Ω, d(u,v) represents a distance metric, and the Euclidean distance is used as the metric;

[0030] The Sinkhorn algorithm is used to solve the above optimization problem. It is necessary to introduce the entropy regularization term H(γ) into the objective function.

[0031]

[0032] Among them, ε is the regularization parameter, which is used to control the degree of entropy regularization;

[0033] According to the optimal transmission theory, the distribution of the extracted high-dimensional representation is adapted to the corresponding high-dimensional representation in the training set, and the feature distribution of the test set is adjusted to be close to the training set; then the adjusted high-dimensional representation is mapped into the adjusted prediction result through LSTM.

[0034] Let the predicted value be y o , at this time the high-dimensional representation of the test set samples is The corresponding high-dimensional representation of the training set with the same label value is ψ, and the γ obtained by optimal transmission is * Make Therefore, the prediction results can be further fine-tuned online during the prediction stage to improve the prediction accuracy.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] Since the present invention takes into account the characteristic distribution differences of full-life data samples under time-varying operating conditions, and attributes the distribution differences to degradation, operating conditions and the nonlinear coupling effects of the two and solves them respectively in a targeted manner, the present invention can effectively reduce the fluctuations in prediction results caused by changes in operating conditions, thereby achieving the advantage of high-precision prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flow chart of an embodiment of the present invention.

[0038] Figure 2 1 and 2 are weighted prediction results of the embodiments of the present invention in the normal stage domain, wherein (a) is the weighted prediction result of the time-varying working condition G_5 in the normal stage domain, and (b) is the weighted prediction result of the time-varying working condition G_6 in the normal stage domain.

[0039] Figure 3 The degradation stage prediction results of the embodiment of the present invention under time-varying working conditions are shown in FIG. 1 , where (a) is the degradation stage prediction result of the time-varying working condition G_5, and (b) is the degradation stage prediction result of the time-varying working condition G_6. DETAILED DESCRIPTION

[0040] The present invention is further described below in conjunction with embodiments and drawings.

[0041] A method for predicting the remaining life of mechanical equipment in a domain-adaptive manner under time-varying working conditions is proposed. The method attributes the differences in the distribution of degradation characteristics under time-varying working conditions to three aspects: degradation, working condition conversion, and the nonlinear coupling between the two. Firstly, the basic prediction model is trained using the full-life degradation data of constant working conditions to target the differences in the distribution of characteristics caused by degradation. Secondly, the optimal domain weights are optimized based on the data of the normal stage of the time-varying working conditions, and the samples of each working condition domain in the degradation stage are weighted to overcome the differences in the distribution of characteristics caused by the working conditions. Finally, the degradation characteristic distribution is fine-tuned based on the optimal transmission theory to further reduce the differences in the distribution of characteristics in the working condition conversion and improve the accuracy of the remaining life prediction.

[0042] Reference Figure 1 A method for predicting the remaining life of mechanical equipment in a time-varying operating condition domain, comprising the following steps:

[0043] Step 1: Obtain a full life sample under constant operating conditions Among them, M C represents the total number of full life samples, C represents the corresponding constant working condition; the normal stage samples are expressed as: Among them, M N is the number of samples in the normal stage, and the corresponding label value is 1; the samples in the degradation stage are represented as: The corresponding label is The degradation samples of time-varying working conditions are composed of samples of various working conditions. The degradation samples under working condition k are expressed as The sample label under working condition k is and Represent samples of its healthy state and degraded state respectively, Assume that their numbers are P N and P D , the total number of samples is P; samples under the same working condition have similar distribution in the feature space, while the distribution characteristics of samples under different working conditions are quite different, that is, P(X1|k)≈P(X2|k), P(X1|k1)≠P(X2|k2);

[0044] Step 2: Perform constant working condition prediction modeling for the original vibration signal Through multi-layer 1D-CNN feature extraction, a high-dimensional representation of the mechanical health status is obtained, as shown below:

[0045]

[0046] In the above formula, σ(·) is the activation function, S is the one-dimensional convolution kernel size, and W l represents the weight of the convolution kernel of the lth layer, b l is the bias of the lth layer;

[0047] The high-dimensional representations of L samples that are continuous in time are spliced ​​into continuous time series features, and the LSTM module is used to construct the time series relationship and predict the corresponding remaining life. The specific calculation process is as follows:

[0048] g(X) = LSTM(x t ,h t-1 ,U t-1 ),t=1,2,…,L

[0049] In the above formula, LSTM(·) is the use of LSTM module, x t is the spliced ​​continuous time series feature data, h t-1is the hidden state at the previous moment, U t-1 is the memory state of the previous moment;

[0050] Network model for steady state conditions For example, its loss function is defined as:

[0051]

[0052] Step 3: Weighting the operating condition domain. When the prediction model of the constant operating condition is directly applied to the normal stage of the time-varying operating condition, the prediction result will deviate from the true value due to the difference in feature distribution caused by the change in the operating condition. k For samples Weighted adjustment is performed to improve fitting accuracy. At this time, the input of the prediction model becomes The output becomes exist The weighted loss function is defined on the sample as:

[0053]

[0054] The above selection weight ω k The process of improving fitting accuracy can be regarded as solving an optimization problem, and the goal is to find the optimal weight sequence ω k , so that the prediction model The fitting error on is the smallest, that is:

[0055]

[0056] Afterwards, the optimal weight sequence ω k Applied to the corresponding working conditions of samples in the degradation stage of time-varying working conditions, the working condition domain of the samples is weighted, and the loss function at this time becomes:

[0057]

[0058] In the normal stage of mechanical equipment, the distribution difference between signals is mainly caused by the change of working conditions, while in the degradation stage, the distribution difference of the collected signals is caused by the degree of degradation, working conditions and their nonlinear coupling effects. Therefore, in order to further consider the coupling effect between working conditions and degradation, the working condition domain weight in the degradation process is designed to gradually decay, indicating that the influence of working conditions on sample distribution gradually decays with degradation, and the domain weight gradually converges to 1. Therefore, the designed attenuation function is as follows:

[0059]

[0060] Where ρ represents the attenuation coefficient;

[0061] Step 4: Perform online fine-tuning. In order to eliminate the nonlinear coupling effect of working conditions and degradation, the distribution of features extracted by the CNN network is adjusted according to the predicted value, and the online fine-tuning process is realized using the optimal transmission theory. The optimization process is as follows:

[0062]

[0063] Among them, γ∈Π(μ u ,μ v )={γ∈P(U×V)|p + #γ=μ u ,p - #=μ v}, p + and p - Represents the mapping of the joint distribution U×V→Ω, d(u,v) represents a distance metric, and the Euclidean distance can generally be used as a metric;

[0064] The Sinkhorn algorithm is used to solve the above optimization problem. The solution of the above optimization process involves complex linear programming operations. In order to speed up the solution and improve the calculation stability, it is necessary to introduce the entropy regularization term H(γ) into the objective function.

[0065]

[0066] Among them, ε is the regularization parameter, which is used to control the degree of entropy regularization;

[0067] According to the optimal transmission theory, the distribution of the extracted high-dimensional representation is adapted to the corresponding high-dimensional representation in the training set, and the feature distribution of the test set is adjusted to be close to the training set; then the adjusted high-dimensional representation is mapped into the adjusted prediction result through LSTM. Specifically, let the prediction value be y o , at this time the high-dimensional representation of the test set samples is The corresponding high-dimensional representation of the training set with the same label value is ψ. The γ obtained by optimal transmission * Can make Therefore, the prediction results can be further fine-tuned online during the prediction stage to improve the prediction accuracy.

[0068] The data used in this embodiment comes from the gearbox fatigue life experiment. During the gear degradation process, a total of 11 channels of data of 5 signal types including vibration, sound, torque, current and speed were collected. The signal sampling frequency is 12.8kHz, the sampling interval is 2min, and the sampling time is 2.56s. The speed and torque of the test gear on the test bench are controlled by the control cabinet. In the algorithm verification, a total of 4 groups of degradation data under constant working conditions are used for prediction model training, and a total of 2 groups of degradation data under time-varying working conditions are used to verify the prediction accuracy of the algorithm. The degradation data conditions and failure conditions under constant working conditions and time-varying working conditions are shown in Table 1.

[0069] Table 1 Bearing vibration signal sample set

[0070]

[0071]

[0072] Reference Figure 2 ,from Figure 2 It can be seen from the figure that for the normal stage data, the prediction results show fluctuations consistent with the operating condition change cycle. Figure 2 The period T marked in t This is the period of working condition conversion, which means that when the trained prediction model is directly applied to time-varying working conditions, the distribution differences caused by the working conditions will directly affect the prediction results. After weighting the samples under each working condition by the obtained working condition domain weight, the prediction results are well consistent with the sample labels, and the prediction error is reduced.

[0073] Reference Figure 3 , Figure 3 It is the degradation stage prediction result of the embodiment under time-varying working conditions. It can be seen that the prediction result is consistent with the actual life in the main trend, but due to the difference in characteristic distribution caused by the working conditions, the prediction result also shows obvious fluctuations with the same working condition change cycle; by weighting the samples in each working condition domain, the volatility of the weighted prediction result is reduced and the error is reduced; then, the optimal transmission theory is used to fine-tune the characteristic distribution of the test set data, and the volatility of the prediction result is further reduced, and the prediction error is also further reduced, indicating that the proposed working condition domain weighting and online fine-tuning algorithm is conducive to overcoming the characteristic distribution differences under time-varying working conditions, thereby improving the prediction accuracy.

Claims

1. A method for predicting the remaining service life of mechanical equipment in a time-varying operating condition domain, characterized in that: Firstly, the basic prediction model is trained using the full-life degradation data of constant working conditions to target the differences in characteristic distribution caused by degradation. Then, the optimal domain weights are optimized according to the normal stage data of time-varying working conditions, and the samples of each working condition domain in the degradation stage are weighted. Finally, the characteristic distribution of degradation is fine-tuned according to the optimal transmission theory to improve the remaining life prediction accuracy.

2. A method for predicting the remaining life of mechanical equipment in a time-varying operating condition domain, characterized in that: The following steps are involved: Step 1: Obtain a full life sample under constant operating conditions Among them, M C represents the total number of full life samples, C represents the corresponding constant working condition; the normal stage samples are expressed as: Among them, M N is the number of samples in the normal stage, and the corresponding label value is 1; the samples in the degradation stage are represented as: The corresponding label is The degradation samples of time-varying working conditions are composed of samples of various working conditions. The degradation samples under working condition k are expressed as The sample label under working condition k is and Represent samples of its healthy state and degraded state respectively, Assume that their numbers are P N and P D , the total number of samples is P; samples under the same working condition have similar distribution in the feature space, while the distribution characteristics of samples under different working conditions are different, that is, P(X1|k)≈P(X2|k), P(X1|k1)≠P(X2|k2); Step 2: Perform constant working condition prediction modeling for the original vibration signal Through multi-layer 1D-CNN feature extraction, a high-dimensional representation of the mechanical health status is obtained, as shown below: In the above formula, σ(·) is the activation function, S is the one-dimensional convolution kernel size, and W l represents the weight of the convolution kernel of the lth layer, b l is the bias of the lth layer; The high-dimensional representations of L samples that are continuous in time are spliced ​​into continuous time series features, and the LSTM module is used to construct the time series relationship and predict the corresponding remaining life. The specific calculation process is as follows: g(X)=LSTM(x t ,h t-1 ,U t-1 ),t=1,2,…,L In the above formula, LSTM(·) is the use of LSTM module, x t is the spliced ​​continuous time series feature data, h t-1 is the hidden state at the previous moment, U t-1 is the memory state of the previous moment; Network model for steady state conditions For example, its loss function is defined as: Step 3: Weighting the operating condition domain, when the prediction model of the constant operating condition is directly applied to the normal stage of the time-varying operating condition, a set of weights ω k For samples Weighted adjustment is performed to improve fitting accuracy. At this time, the input of the prediction model becomes The output becomes exist The weighted loss function is defined on the sample as: The above selection weight ω k The process of improving fitting accuracy is regarded as solving an optimization problem, and the goal is to find the optimal weight sequence ω k , so that the prediction model The fitting error on is the smallest, that is: Afterwards, the optimal weight sequence ω k Applied to the corresponding working conditions of samples in the degradation stage of time-varying working conditions, the working condition domain of the samples is weighted, and the loss function at this time becomes: Considering the coupling effect between working condition and degradation, the working condition domain weight in the degradation process is designed to decay gradually, indicating that the influence of working condition on sample distribution gradually decays with degradation, and the domain weight gradually converges to 1. The designed decay function is as follows: Where ρ represents the attenuation coefficient; Step 4: Perform online fine-tuning, adjust the distribution of features extracted by the CNN network according to the predicted value, and use the optimal transmission theory to implement the online fine-tuning process; the optimization process is as follows: Among them, γ∈Π(μ u ,μ v )={γ∈P(U×V)|p + #γ=μ u ,p - #=μ v }, p + and p - represents the mapping of the joint distribution U×V→Ω, d(u,v) represents a distance metric, and the Euclidean distance is used as the metric; The Sinkhorn algorithm is used to solve the above optimization problem. It is necessary to introduce the entropy regularization term H(γ) into the objective function. Among them, ε is the regularization parameter, which is used to control the degree of entropy regularization; According to the optimal transmission theory, the distribution of the extracted high-dimensional representation is adapted to the corresponding high-dimensional representation in the training set, and the feature distribution of the test set is adjusted to be close to the training set; then the adjusted high-dimensional representation is mapped into the adjusted prediction result through LSTM.

3. The prediction method according to claim 2, characterized in that: Let the predicted value of step 4 be y o , at this time the high-dimensional representation of the test set samples is The corresponding high-dimensional representation of the training set with the same label value is ψ, and the γ obtained by optimal transmission is * Make Therefore, the prediction results can be further fine-tuned online in the prediction stage to improve the prediction accuracy.

Images