A bearing cross-operating condition fault prediction method based on adversarial transfer learning

Abstract

The present invention discloses a bearing cross-operating condition fault prediction method based on adversarial transfer learning, the method comprising: setting different operating conditions and collecting bearing signals to obtain vibration signals; identifying vibration signals based on a continuous abnormal point detection method of peak measurement to obtain healthy stage signals and degradation stage signals; preprocessing degradation stage signals and extracting fault features to obtain a fault feature set; inputting the fault feature set into a domain adversarial transfer learning two-branch neural network and performing update training to obtain a prediction model; obtaining data to be tested and inputting it into the prediction model to obtain a fault prediction result. By using the present invention, accurate life prediction of high-speed shaft ball bearings of large-scale wind turbine gearboxes under different operating conditions is achieved based on actual operating condition characteristics. As a bearing cross-operating condition fault prediction method based on adversarial transfer learning, the present invention can be widely used in the field of engineering component life prediction.

Description

Technical Field

[0001] The present invention relates to the field of life prediction of engineering components, and in particular to a bearing cross-operating condition fault prediction method based on adversarial transfer learning. Background Art

[0002] Wind power development has led to the vigorous development of the wind power industry. Gearboxes are important mechanical components in the transmission chain of large wind turbines. They are characterized by complex structure, low speed, heavy load, and long service life. Their key components (such as the front bearing at the input end, the front and rear bearings of the high-speed shaft, planetary gears, sun gears, and ring gears) frequently fail under the influence of complex and variable loads and extremely strong instantaneous impacts. Due to the severity of the consequences of their failure, gearboxes have always been the key components and weak links of large wind turbines. Gearbox failures have typical chain characteristics, that is, minor failures can easily expand into catastrophic failures in a short period of time, generally leading directly to the shutdown of large wind turbines. Secondly, the difficulty in repairing gearbox failures will result in longer downtime, causing huge economic losses, and thus significantly reducing the availability and economy of large wind turbines throughout their life cycle.

[0003] Large wind turbine gearboxes differ from ordinary rotating mechanical systems due to their complex structure, harsh working environment, and sophisticated assembly process. As a result, their remaining life prediction methods for engineering practice face many challenges: fault characteristics are difficult to effectively extract, the strong randomness of the degradation process makes it difficult to robustly predict the remaining life, and the implementation steps are often limited by expert experience, resulting in a low level of intelligence in life prediction research. Summary of the Invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide a bearing cross-operating condition fault prediction method based on adversarial transfer learning, which has good high precision and domain adversarial unsupervised deep transfer learning network to realize cross-operating condition fault prediction of high-speed shaft ball bearings of large wind turbine gearboxes.

[0005] The first technical solution adopted by the present invention is: a bearing cross-operating condition fault prediction method based on adversarial transfer learning, comprising the following steps:

[0006] Set different working conditions and collect bearing signals to obtain vibration signals;

[0007] The continuous abnormal point detection method based on peak measurement is used to identify the vibration signal and obtain the healthy stage signal and the degradation stage signal;

[0008] Preprocess the degradation stage signal and extract fault features to obtain the fault feature set;

[0009] The fault feature set is input into the domain adversarial transfer learning two-branch neural network and updated and trained to obtain a prediction model;

[0010] Obtain the data to be tested and input it into the prediction model to obtain the fault prediction results.

[0011] Furthermore, the step of setting different working conditions and collecting bearing signals to obtain vibration signals specifically includes:

[0012] Collect the bearing vibration signal when the equipment runs to the cutoff point to obtain the target domain fault feature data;

[0013] The same equipment is run from the starting point to failure in a controlled environment, and the bearing vibration signal is collected to obtain source domain fault feature data;

[0014] The target domain fault feature data and the source domain fault feature data are integrated to obtain the vibration signal.

[0015] Furthermore, the continuous abnormal point detection method based on peak measurement identifies the vibration signal to obtain the healthy stage signal and the degraded stage signal, which specifically includes:

[0016] The continuous abnormal point detection method based on peak measurement is used to identify the vibration signal and obtain the state inflection point;

[0017] The vibration signal is divided into a healthy stage and a degraded stage according to the state inflection point, and a healthy stage signal and a degraded stage signal are obtained.

[0018] Furthermore, the step of preprocessing the degradation stage signal and extracting fault features to obtain a fault feature set specifically includes:

[0019] Extract features of degradation stage signals based on one-dimensional time series to obtain the first fault feature;

[0020] Extract features of degradation stage signals based on two-dimensional time spectrum to obtain the second fault feature;

[0021] The first fault feature is converted into a normalized scale and dimension reduction is performed using bilinear interpolation to obtain the preprocessed first fault feature;

[0022] The preprocessed first fault feature and the second fault feature are integrated to obtain a fault feature set.

[0023] Furthermore, the prediction model includes a feature extractor module, a remaining life predictor module and a domain adversarial module.

[0024] Furthermore, the step of inputting the fault feature set into the domain adversarial transfer learning two-branch neural network and performing update training to obtain a prediction model specifically includes:

[0025] The first fault feature and the second fault feature preprocessed in the fault feature set are input into the first channel and the second channel of the domain adversarial transfer learning two-branch neural network respectively;

[0026] Extract high-level fusion features based on the feature extractor and input them into the remaining life prediction module and the domain adversarial module;

[0027] Predicting the remaining life of the bearing based on the remaining life predictor module;

[0028] Perform domain classification judgment on high-level fusion features based on the domain adversarial module;

[0029] Based on the adaptive optimization algorithm, combined with the true label back propagation, the training network parameters are iteratively updated to obtain the prediction model.

[0030] Furthermore, the expression of the remaining life predictor module is as follows:

[0031]

[0032] In the above formula, Indicates the percentage of the predicted remaining useful life, G p (·;θ p ) represents the remaining life predictor module composed of fully connected layers, σ(·) represents the sigmoid activation function, θ p Represents the parameter set of the remaining life prediction module.

[0033] Furthermore, the optimization objectives of the iterative update training network parameters include the regression loss of the remaining life predictor module and the domain classification loss of the domain adversarial module.

[0034] The beneficial effects of the method of the present invention are as follows: the present invention utilizes the dual-channel heterogeneous input of the dual-branch neural network of domain adversarial transfer learning to improve the accuracy of transmission fault prediction. Through domain adversarial training in an unsupervised domain adaptation framework, the problem of significant distribution differences can be effectively solved, and accurate life prediction of high-speed shaft ball bearings of large wind turbine gearboxes under different working conditions is achieved based on actual working condition characteristics. The method has excellent transfer prediction performance and strong robustness and generalization capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is a flowchart of the steps of a bearing cross-operating condition fault prediction method based on adversarial transfer learning of the present invention;

[0036] Figure 2 1 is a flow chart of a prediction method according to a specific embodiment of the present invention;

[0037] Figure 3 It is a structural diagram of a prediction model according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0038] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are provided for ease of description only and do not limit the order of the steps. The order of execution of the steps in the embodiments can be adaptively adjusted based on the understanding of those skilled in the art.

[0039] Reference Figure 1 and Figure 2 The present invention provides a bearing cross-operating condition fault prediction method based on adversarial transfer learning, which includes the following steps:

[0040] S1. Set different working conditions and collect bearing signals to obtain vibration signals;

[0041] S1.1. Collect bearing vibration signals when the equipment runs to the cutoff point to obtain target domain fault feature data;

[0042] S1.2. Run the same equipment from the starting point to the failure in a controlled environment and collect bearing vibration signals to obtain source domain fault feature data;

[0043] S1.3. Integrate the target domain fault feature data and the source domain fault feature data to obtain a vibration signal.

[0044] Specifically, the data in the source domain are bearing vibration signals collected from similar equipment running from the starting point to failure in a controlled laboratory environment, while the data in the target domain are vibration signals collected from the equipment running to the truncation point, which are small sample monitoring data with a small number of labels or no labels.

[0045] S2. A continuous outlier detection method based on peak measurement is used to identify the vibration signal and obtain healthy stage signals and degradation stage signals;

[0046] S2.1. Identify the vibration signal using a continuous outlier detection method based on peak measurement to obtain the state inflection point;

[0047] S2.2. Divide the vibration signal into a healthy stage and a degraded stage according to the state inflection point, and obtain a healthy stage signal and a degraded stage signal.

[0048] Specifically, the first prediction time (FPT) of the transferable fault prediction method is further determined. After the FPT is determined, the fault prediction algorithm is triggered to predict the remaining useful life (RUL) of the bearing in the target domain. The FPT discriminant function is expressed as:

[0049] {|k t-j -μ|>3σ} i＝0,1,2,3,4

[0050] Among them, k t-j represents the original peak metric sequence (tj), μ and σ represent the mean and variance of the peak metric respectively, i = 0, 1, 2, 3, 4 represent 5 consecutive monitoring points, and the time parameter t is defined as the FPT that triggers the fault prediction algorithm.

[0051] S3. Preprocess the degradation stage signal and extract fault features to obtain a fault feature set;

[0052] S3.1. Extract features of degradation stage signals based on a one-dimensional time series to obtain a first fault feature.

[0053] S3.2. Extract features of the degradation stage signal based on the two-dimensional time spectrum to obtain the second fault feature;

[0054] Specifically, the first input channel is the fault features based on one-dimensional time series extracted from the time and frequency domains of the original vibration signal. The second input channel is the fault features based on two-dimensional time-frequency representations (TFRs), which is a two-dimensional coefficient matrix obtained by converting the one-dimensional time series through the continuous wavelet transform algorithm. These dual-channel heterogeneous fault features are extracted from the original vibration subsequence between the FPT point and the end point. In the source domain, there are fault features with RUL labels, and their extraction end point is the end of life (EoL), while in the target domain, there are fault features with a small number of RUL labels or no labels, and their extraction end point is the truncation point of the bearing.

[0055] S3.3. Convert the first fault feature into a normalized scale and perform dimensionality reduction using bilinear interpolation to obtain a preprocessed first fault feature;

[0056] Specifically, Max-Min normalization is used to transform the fault features of the first input channel into a normalized scale, and bilinear interpolation is adopted to reduce the high dimension of the original TFRs and eliminate the adverse effects of signal differences.

[0057] S3.4. Integrate the pre-processed first fault feature and the second fault feature to obtain a fault feature set.

[0058] Specifically, Indicates that it contains training samples N S The source domain, χ S is the feature space of the source domain, represents the RUL label value corresponding to the i-th training sample in the source domain; Indicates that it contains test sample N T The target domain, Indicates that from χT The j-th test sample is extracted from the target feature space, but the corresponding RUL label value is not available.

[0059] Specifically, the association between the first fault feature and the second fault feature and the source domain fault feature and the target fault feature is that the source domain fault feature is divided into one-dimensional and two-dimensional, that is, the source domain fault feature includes the fault feature of the one-dimensional time series and the fault feature of the two-dimensional time spectrum, and the same applies to the target fault feature.

[0060] S4. Input the fault feature set into the domain adversarial transfer learning two-branch neural network and perform update training to obtain a prediction model;

[0061] S4.1. Input the first fault feature and the second fault feature after centralized preprocessing into the first channel and the second channel of the domain adversarial transfer learning two-branch neural network respectively;

[0062] Specifically, the two-channel heterogeneous original fault features, namely the source domain fault features and target domain fault characteristics Input to the domain adversarial transfer learning dual-branch neural network (DBNN-DA), refer to Figure 3 ,The DBNN-DA network model consists of three modules: feature extractor module, RUL predictor module and domain adversarial ,module.

[0063] S4.2. Extract high-level fusion features based on the feature extractor and input them into the remaining life prediction module and the domain adversarial module;

[0064] Specifically, the feature extractor module of the DBNN-DA network model automatically extracts high-level fusion features from the two-channel heterogeneous raw fault features in the labeled source domain and the unlabeled target domain, and the output is:

[0065] h FL =G f (x S_1D ,x S_2D ,x T_1D ,x T_2D θ f )

[0066] In the above formula, h FL represents the extracted high-level fusion features, G f (·;θ f ) represents the feature extractor module, i.e., a two-branch network, θ f is the parameter set of the feature extractor.

[0067] S4.3. Predicting the remaining life of the bearing based on the remaining life predictor module;

[0068] Specifically, the expression of the remaining life predictor module is as follows:

[0069]

[0070] In the above formula, Indicates the percentage of the predicted remaining useful life, G p (·;θ p ) represents the remaining life predictor module composed of fully connected layers, σ(·) represents the sigmoid activation function, θ p Represents the parameter set of the remaining life prediction module.

[0071] S4.4, domain classification judgment of high-level fusion features based on domain adversarial module;

[0072] Specifically, the high-level fusion features are simultaneously input into the domain adversarial module to achieve unsupervised domain adaptation. The domain adversarial module consists of a domain classifier that can determine whether a sample is from the source domain or the target domain. The domain classifier is defined as:

[0073]

[0074] In the above formula, G d (·;θ d ) is a domain classifier composed of several fully connected layers, and the corresponding last layer is a binary classification layer, θ d is the parameter set of the domain classifier, γ(·) represents the logarithmic softmax activation function, is the predicted domain label (source domain: 0, target domain: 1).

[0075] S4.5. Based on the adaptive optimization algorithm, combined with the true label back propagation, the training network parameters are iteratively updated to obtain the prediction model.

[0076] Specifically, the optimization target of the iterative update training network parameters includes the regression loss L of the remaining life predictor module p and the domain classification loss L of the domain adversarial module d .

[0077] Minimize the first loss term L p , feature extractor G f (·;θ f ) and RUL predictor G p (·;θ p ) are jointly optimized to minimize the empirical regression loss of source domain samples and ensure that the high-level fusion feature h FL The discriminability of , and the overall good RUL prediction accuracy of the feature extractor and RUL predictor on the source domain. The loss function L is defined as the mean square error p for:

[0078]

[0079] In the above formula, N batch is the batch size of labeled samples in the source domain, y i and are the actual RULP labels and predicted RULP values, respectively.

[0080] Maximize the second loss term L d Because we hope that the high-level fusion features are domain-invariant, so that the domain classifier cannot distinguish between samples in the source domain and the target domain, this process is called domain adversarial training. Its basic idea is to use the domain classifier and feature extractor to align the distribution in an adversarial manner. The domain classifier is used to distinguish the domain labels of the high-level fusion features generated by the feature extractor, while the feature extractor is trained to deceive the domain classifier. The loss function L is defined as the cross entropy loss d for:

[0081]

[0082] In the above formula, l i is the actual field label, is the predicted domain label generated by the domain classifier.

[0083] The final optimization objective function of the DBNN-DA model is summarized as:

[0084]

[0085] In the above formula, λ is the control L p and L d A non-negative hyperparameter that trades off between and is the corresponding loss function for the i-th training sample.

[0086] By finding the parameter set of the saddle point of the objective function Achieve optimization.

[0087]

[0088]

[0089] At the saddle point, the parameter set θ of the domain classifier d Minimize the domain classification loss (with a minus sign in the optimization objective function), the parameter set θ of the RUL predictor p Minimize the RUL regression loss, the parameter set θ of the feature extractor f Minimize the RUL regression loss while maximizing the domain classification loss.

[0090] Through the Adam optimizer, the network parameter set Θ = (θf ,θ p ,θ d ) is updated to:

[0091]

[0092]

[0093]

[0094] Where ε represents the learning rate.

[0095] S5. Obtain the data to be tested and input it into the prediction model to obtain the fault prediction result.

[0096] Specifically, the real-time detection data of the high-speed shaft ball bearing of a large wind turbine gearbox under different working conditions in the target domain are input into the trained DBNN-DA network model for accurate online fault prediction.

[0097] A bearing cross-operating condition fault prediction device based on adversarial transfer learning:

[0098] at least one processor;

[0099] at least one memory for storing at least one program;

[0100] When the at least one program is executed by the at least one processor, the at least one processor implements the above-mentioned bearing cross-operating condition fault prediction method based on adversarial transfer learning.

[0101] The contents of the above method embodiments are all applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0102] A storage medium storing processor-executable instructions, characterized in that the processor-executable instructions, when executed by the processor, are used to implement the above-mentioned bearing cross-operating condition fault prediction method based on adversarial transfer learning.

[0103] The contents of the above method embodiments are all applicable to the present storage medium embodiment. The functions specifically implemented by the present storage medium embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0104] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A bearing cross-operating condition fault prediction method based on adversarial transfer learning, characterized in that: The following steps are involved: Set different working conditions and collect bearing signals to obtain vibration signals; The continuous abnormal point detection method based on peak measurement is used to identify the vibration signal and obtain the healthy stage signal and the degradation stage signal; Preprocess the degradation stage signal and extract fault features to obtain the fault feature set; The fault feature set is input into the domain adversarial transfer learning two-branch neural network and updated and trained to obtain a prediction model; Obtain the data to be tested and input it into the prediction model to obtain the fault prediction results; The prediction model includes a feature extractor module, a remaining life predictor module and a domain adversarial module; The step of inputting the fault feature set into the domain adversarial transfer learning two-branch neural network and performing update training to obtain a prediction model specifically includes: The first fault feature and the second fault feature preprocessed in the fault feature set are input into the first channel and the second channel of the domain adversarial transfer learning two-branch neural network respectively; Extract high-level fusion features based on the feature extractor and input them into the remaining life prediction module and the domain adversarial module; Predicting the remaining life of the bearing based on the remaining life predictor module; Perform domain classification judgment on high-level fusion features based on the domain adversarial module; Based on the adaptive optimization algorithm, combined with the real label back propagation, the training network parameters are iteratively updated to obtain the prediction model; The output expression of the feature extractor module is as follows: h FL =G f (x S_1D ,x S_2D ,x T_1D ,x T_2D ;θ f ) In the above formula, h FL represents the extracted high-level fusion features, G f (·;θ f ) represents the feature extractor module, i.e., a two-branch network, θ f is the parameter set of the feature extractor; The expression of the remaining life predictor module is as follows: In the above formula, Indicates the percentage of the predicted remaining useful life, G p (·;θ p ) represents the remaining life predictor module composed of fully connected layers, σ(·) represents the sigmoid activation function, θ p Represents the parameter set of the remaining life prediction module; The domain adversarial module consists of a domain classifier, which is defined as: In the above formula, G d (·;θ d ) is a domain classifier composed of several fully connected layers, and the corresponding last layer is a binary classification layer, θ d is the parameter set of the domain classifier, γ(·) represents the logarithmic softmax activation function, is the predicted domain label; The optimization objectives of the iterative updating training network parameters include the regression loss of the remaining life predictor module and the domain classification loss of the domain adversarial module; The final optimization objective function of the prediction model is summarized as: In the above formula, L p represents the regression loss of the remaining life predictor module, L d represents the domain classification loss of the domain adversarial module, N batch represents the batch size of labeled samples in the source domain, and λ is the control L p and L d A non-negative hyperparameter that trades off between and is the corresponding loss function for the i-th training sample.

2. A bearing cross-operating condition fault prediction method based on adversarial transfer learning according to claim 1, characterized in that: The step of setting different working conditions and collecting bearing signals to obtain vibration signals specifically includes: Collect the bearing vibration signal when the equipment runs to the cutoff point to obtain the target domain fault feature data; The same equipment is run from the starting point to failure in a controlled environment, and the bearing vibration signal is collected to obtain source domain fault feature data; The target domain fault feature data and the source domain fault feature data are integrated to obtain the vibration signal.

3. The bearing cross-operating condition fault prediction method based on adversarial transfer learning according to claim 2 is characterized in that: The continuous abnormal point detection method based on peak measurement identifies the vibration signal to obtain the healthy stage signal and the degraded stage signal, which specifically includes: The continuous abnormal point detection method based on peak measurement is used to identify the vibration signal and obtain the state inflection point; The vibration signal is divided into a healthy stage and a degraded stage according to the state inflection point, and a healthy stage signal and a degraded stage signal are obtained.

4. The bearing cross-operating condition fault prediction method based on adversarial transfer learning according to claim 3 is characterized in that: The step of preprocessing the degradation stage signal and extracting fault features to obtain a fault feature set specifically includes: Extract features of degradation stage signals based on one-dimensional time series to obtain the first fault feature; Extract features of degradation stage signals based on two-dimensional time spectrum to obtain the second fault feature; The first fault feature is converted into a normalized scale and dimension reduction is performed using bilinear interpolation to obtain the preprocessed first fault feature; The preprocessed first fault feature and the second fault feature are integrated to obtain a fault feature set.

Images