A method for predicting the remaining useful life of a rotating machine under variable working conditions

Abstract

The application discloses a residual service life prediction method of rotating machinery under variable working conditions, divides full-life vibration signals of the rotating machinery into a source domain data set and a target domain data set, respectively extracts a plurality of features from time domain, frequency domain and time-frequency domain for the source domain data set and the target domain data set, and filters the extracted features from two aspects of feature trend and feature similarity in turn; the data set after the feature filtering is divided into a training set and a test set, wherein the training set contains labeled samples in the source domain and part of unlabeled samples in the target domain, and the test set is part of the unlabeled samples in the target domain; a transfer prediction model is trained by using the training set, and then the test set is input into the trained model to predict the residual service life of the rotating machinery. The application can accurately predict the residual service life of the rotating machinery under the target working condition.

Description

Technical Field

[0001] This invention relates to a method for predicting the remaining service life of rotating machinery under varying operating conditions based on deep migration. Background Technology

[0002] Traditional methods for predicting the remaining useful life (RUL) of rotating machinery typically rely on manually extracted vibration signal features to establish a model of the machinery's life degradation mechanism. However, these traditional methods are time-consuming and labor-intensive, relying heavily on expert knowledge and experience, and are no longer sufficient to meet researchers' needs. While machine learning and deep learning have demonstrated significant advantages in RUL prediction for rotating machinery, most training and testing data are derived from the same operating conditions. In real-world applications, different rotating machines operate under varying conditions, leading to changes in the distribution of training and testing data. This can cause trained models to fail to accurately predict the RUL under the current operating conditions. Transfer learning effectively addresses this problem.

[0003] Transfer learning can achieve cross-domain learning by reducing the differences in data distribution under different operating conditions. Therefore, through transfer learning, the RUL of rotating machinery in the target domain operating conditions can be accurately predicted using a large amount of source domain data readily available in the laboratory. However, since the extracted vibration signals usually contain a large amount of useless and noise signals, this may have a certain impact on the final prediction effect of the above methods. Furthermore, the transferability of data feature distribution and its sensitivity to degradation trends directly determine the accuracy of the model's transfer prediction, but this factor is not currently considered in mainstream methods. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method for predicting the remaining service life of rotating machinery under varying operating conditions, thereby achieving the goal of accurately predicting the remaining service life of rotating machinery under target operating conditions.

[0005] The present invention achieves the above-mentioned technical objectives through the following technical means.

[0006] A method for predicting the remaining service life of rotating machinery under varying operating conditions:

[0007] The vibration signal of rotating machinery throughout its entire lifespan is divided into a source domain dataset and a target domain dataset. Multiple features are extracted from the source domain dataset and the target domain dataset in the time domain, frequency domain, and time-frequency domain, respectively. The extracted features are then selected based on two aspects: feature trend and feature similarity.

[0008] The dataset after feature filtering is divided into a training set and a test set. The training set contains labeled samples from the source domain and some unlabeled samples from the target domain, while the test set contains some unlabeled samples from the target domain.

[0009] The training set is used to train a transfer prediction model, and then the test set is input into the trained model to predict the remaining service life of rotating machinery.

[0010] A further technical solution involves filtering the extracted features based on feature similarity. Specifically, this involves using a feature similarity evaluation index Y to further filter the features extracted based on feature trends.

[0011]

[0012] Among them: WDTW lq The WDTW distance represents a certain feature between the l-th rotating machine in the source domain and the q-th rotating machine in the target domain, where u and v represent the number of rotating machines in the source domain and the target domain, respectively.

[0013] A further technical solution involves the following process for acquiring the WDTW:

[0014] Assume Q = q1, q2, q3, ..., q n C = c1, c2, c3, ..., c m Given two feature sequences of length n and m from the source and target domains respectively, construct an n*m distance matrix. Then, use this distance matrix to find the shortest normalized path.

[0015]

[0016] Where: K is the path length, w k It is the path regularized in the distance matrix, and WDTW is the minimum distance between the features of the source domain and the target domain.

[0017] A further technical solution is that the element in the i-th row and j-th column of the distance matrix is ​​the weighted Euclidean distance between the i-th data point in the source domain dataset and the j-th data point in the target domain dataset.

[0018] d(q i ,c j )={w |i-j| (q i ,c j )} 2

[0019] Where: q i Let c be the i-th point of the source domain feature sequence Q. j Let q be the j-th point of the target domain feature sequence C. ic j Distance weight w between |i-j| With q i c j It depends on the distance of time:

[0020]

[0021] In a further technical solution, the migration prediction model consists of a feature extractor, a predictor, and a domain classifier. The feature extractor includes two layers of bidirectional long short-term memory network and one fully connected layer. The predictor and the domain classifier are each composed of three fully connected layers with different parameters.

[0022] A further technical solution involves optimizing the migration prediction model to reduce the feature extractor G. f Loss, reduce predictor G y Loss and increase domain classifier G d The loss.

[0023] A further technical solution involves using two optimizers in the optimization process.

[0024] A further technical solution, the optimization objective of the first optimizer is:

[0025] L loss1 (θ f ,θ y ) = L f (θ f )+L y (θ f ,θ y )

[0026] Where: L f For the feature extractor loss, L y For the predictor loss, θ f θ is the parameter of the feature extractor loss. y These are the parameters of the predictor.

[0027] A further technical solution involves the optimization objective of the second optimizer:

[0028] L loss2 (θ f ,θ d ) = L f (θ f )-λL d (θ f ,θ d )

[0029] Where: L d For the loss of the domain classifier, θ d λ represents the parameters of the domain classifier, and λ is the hyperparameter of the balancing loss term.

[0030] The beneficial effects of this invention are as follows: This invention utilizes Spearman correlation coefficient and improved dynamic time warping as feature evaluation indicators, which can select features with good trends and transferability as training samples, enabling the transfer prediction model to better extract domain-invariant features in the source and target domains, and use the degradation information contained in the features to predict the remaining service life; at the same time, it uses a bidirectional long short-term memory network as a feature extractor combined with a domain adversarial network to build a transfer prediction model, and adds the maximum mean difference between feature extractors to the model optimization objective to enhance the model's transfer learning ability. For multi-task optimization objectives, the network parameters of each part of the model are assigned to different optimization tasks, eliminating the interference of the optimization objective on irrelevant network parameters during the optimization process, improving the model optimization efficiency, and improving the prediction accuracy of the remaining service life of rotating machinery. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method for predicting the remaining service life of rotating machinery under varying operating conditions as described in this invention.

[0032] Figure 2 This is a diagram of the migration prediction network model of the present invention;

[0033] Figure 3 This is a graph showing the variation trend of parameter λ in the gradient inversion layer of this invention;

[0034] Figure 4(a) shows the prediction results of bearings 1-7 in an embodiment of the present invention;

[0035] Figure 4(b) shows the prediction results for bearings 2-6 in an embodiment of the present invention;

[0036] Figure 4(c) shows the prediction results of bearing 3-3 in the embodiment of the present invention. Detailed Implementation

[0037] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention. The invention will be further described below with reference to the accompanying drawings and an embodiment taking a rolling bearing as an example.

[0038] The data for the prediction method in this invention comes from bearing life-cycle vibration data provided by the IEEE PHM Challenge 2012. This data records all bearing data from normal operation to failure in chronological order, and includes life-cycle vibration data under three different operating conditions: Condition 1 has a load of 4000 N and a speed of 1800 rpm; Condition 2 has a load of 4200 N and a speed of 1650 rpm; and Condition 3 has a load of 5000 N and a speed of 1500 rpm. Conditions 1 and 2 each contain seven bearings with seven different failure types, while Condition 3 contains three bearings with three different failure types. These failure types represent common failures during bearing operation. The vibration data is collected by two 25.6 kHz accelerometers installed in different directions (lateral and longitudinal) of the bearing, and the bearing degradation process is accelerated by increasing the radial load. The sampling interval is 10 seconds, and each acquisition time is 0.1 seconds. Each acquisition results in a sample containing 2560 points in two different directions. The signal acquisition is recorded from the start until the bearing fails, and a file is generated containing sample data of the entire degradation process of the bearing. Among them, operating conditions one, two, and three can be used as both source and target domains, and the source and target domains are different.

[0039] Figure 1 The remaining lifetime prediction method based on deep migration of this invention mainly includes the following steps:

[0040] S1. Feature extraction is performed on the sample data collected from two directions for each bearing. The extracted features include 15 time-domain features, 10 frequency-domain features, and 8 time-frequency-domain features, for a total of 66 features, as shown in Table 1.

[0041] Table 1 Feature Extraction

[0042]

[0043] S2, Filter the features extracted from the vibration signal throughout the bearing's lifespan.

[0044] S2.1, using the Spearman correlation coefficient as a characteristic trend indicator, the above 66 characteristics of each bearing are screened, retaining the characteristics with the greater overall trend for each bearing. The screening formula is as follows:

[0045]

[0046] In the formula: x i Let t represent the feature value at time i in a file. i x represents i The corresponding time scale, where N is the length of the feature sequence; These represent N characteristic values ​​and their corresponding mean values ​​at different times; ρ(x,t) represents the trend value of a certain characteristic.

[0047] The trend evaluation index Z, derived from the Spearman correlation coefficient, is given by the following formula:

[0048]

[0049] In the formula: ρ l Let h be the trend value of a certain characteristic of the l-th bearing, and h be the number of bearings under this operating condition;

[0050] The trend of the sample data features is calculated using the Z-value, and features with a large overall trend are selected for each bearing.

[0051] S2.2, the retained features are further screened using the feature similarity evaluation index Y, and the formula is as follows:

[0052]

[0053] In the formula: WDTW lq The WDTW distance represents a certain feature between the l-th bearing in the source domain and the q-th bearing in the target domain, where u and v represent the number of bearings in the source domain and the target domain, respectively.

[0054] Transferability can be determined by calculating the similarity of a feature between the source and target domains using the Y-value. Features with high similarity are then further filtered using the Y-value, ultimately retaining 30 features with strong trends and high transferability.

[0055] The process of obtaining WDTW is as follows:

[0056] (1) When calculating the similarity of features extracted from the vibration signal of the bearing throughout its entire lifespan, assume that Q = q1, q2, q3, ..., q n C = c1, c2, c3, ..., c m Given two feature sequences of length n and m from the source and target domains respectively, construct an n*m distance matrix. The element in the i-th row and j-th column of the distance matrix is ​​the weighted Euclidean distance between the i-th data point in the source domain and the j-th data point in the target domain dataset, calculated using the following formula:

[0057] d(q i ,c j )={w |i-j| (q i ,c j )} 2

[0058] In the formula: q i Let c be the i-th point of the source domain feature sequence Q. j For the j-th point of the target domain feature sequence C, w |i-j| With q i c jIt is related to the distance at any given time, and its calculation formula is as follows:

[0059]

[0060] In the formula: n is the length of the source domain feature sequence Q, m is the length of the target domain feature sequence C, and w |i-j| For q i c j Distance weights between them.

[0061] (2) Using the n*m ​​distance matrix constructed above, the shortest regularized path is obtained as follows:

[0062]

[0063] In the formula: K is the path length, which can eliminate the influence of paths of different lengths, w k It is the path normalized in the distance matrix (the specific normalization method is the existing technology), while WDTW is the minimum distance between the features of the source domain and the target domain.

[0064] S3, the features selected from each bearing are grouped into a sample based on the feature values ​​at every ten consecutive time points. The data in the sample is arranged as follows:

[0065]

[0066] In the formula: Let X be the eigenvalue at time s, t. s Let be the feature set at time s.

[0067] The feature sets of every ten consecutive time points are arranged as a group, as shown in the following formula:

[0068]

[0069] During training, each column of data in the above matrix is ​​used as a sample input into the model for time-series feature extraction.

[0070] S4. A domain adversarial network (i.e., a transfer prediction model) is built using a bidirectional long short-term memory network as a feature extractor, and the maximum mean difference is incorporated into the model's optimization process. The network structure diagram is attached. Figure 2 As shown, the feature extractor consists of two bidirectional long short-term memory (LSTM) layers and one fully connected layer, while the predictor and domain classifier each consist of three fully connected layers with different parameters. The optimization objective includes three parts: reducing the feature extractor G... f Loss, reduce predictor G y Loss and increase domain classifier G d To minimize network loss, the optimization process employs two optimizers with different network parameters. The first optimizer's objective is:

[0071] L loss1 (θ f ,θ y ) = L f (θ f )+L y (θ f ,θ y )

[0072] In the formula: L f L is the feature extractor loss (i.e., the maximum mean difference between feature extractors). y For the predictor loss, θ f θ is the parameter of the feature extractor loss. y These are the parameters of the predictor;

[0073] The optimization objective of the second optimizer is:

[0074] L loss2 (θ f ,θ d ) = L f (θ f )-λL d (θ f ,θ d )

[0075] In the formula: L d For the loss of the domain classifier, θ d Let be the parameters of the domain classifier, and λ be a hyperparameter of the manually set balancing loss term, calculated as follows:

[0076]

[0077] In the formula: γ is a positive value set manually; p is the ratio of the current training steps to the total training steps, changing from 0 to 1, gradually enabling the model to extract features that can confuse the domain classifier. The trend of this change with the number of training steps is shown in the attached figure. Figure 3 As shown.

[0078] S5. Before training, set the source domain and target domain labels to 0 and 1 respectively. Then, use the remaining lifetime percentage at the current time as the label. The calculation formula is as follows:

[0079]

[0080] In the formula: t i y is the value at the current time. i The remaining life tag value is at the current moment, and T is the bearing failure time.

[0081] Arrange the remaining lifetime label values ​​in the same way as S3, and take the last remaining lifetime label value in each column as the label of the corresponding sample in S3.

[0082] S6. Divide the datasets under three different working conditions into training sets and test sets. The training set contains labeled samples from the source domain and some unlabeled samples from the target domain. The test set contains some unlabeled samples from the target domain. There are three groups of experiments in total, as shown in the table below:

[0083] Table 2 Dataset Division for Transfer Learning Experiments

[0084]

[0085] S8. Input the three sets of training data into the transfer prediction model built in S4 for training, and put the corresponding test sets into the trained transfer prediction model for prediction. The prediction results are shown in Figures 4(a), (b), and (c). It is easy to see that in the three sets of experiments, the prediction curves of the remaining service life of the tested bearings can fit the real bearing degradation curves well.

[0086] In summary, the present invention provides a method for predicting the remaining service life of rotating machinery based on deep migration, which can effectively select features with good trend and transferability, and extract domain-invariant features from the source and target domain data to complete the prediction of the remaining service life of rotating machinery under varying operating conditions.

[0087] The terms "an embodiment," "example," "specific instance," etc., used in this specification refer to specific features, materials, structures, or characteristics incorporated in that embodiment that are included in at least one embodiment or example of the invention. In this specification, the illustrative use of the above terms does not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0088] The embodiments described above are preferred embodiments of the present invention, but the present invention is not limited to the above embodiments. Any obvious improvements, substitutions or modifications that can be made by those skilled in the art without departing from the essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for predicting the remaining service life of rotating machinery under varying operating conditions, characterized in that: The vibration signal of rotating machinery throughout its entire lifespan is divided into a source domain dataset and a target domain dataset. Multiple features are extracted from the source domain dataset and the target domain dataset in the time domain, frequency domain, and time-frequency domain, respectively. The extracted features are then selected based on two aspects: feature trend and feature similarity. The dataset after feature filtering is divided into a training set and a test set. The training set contains labeled samples from the source domain and some unlabeled samples from the target domain, while the test set contains some unlabeled samples from the target domain. The training set is used to train a transfer prediction model, and then the test set is input into the trained model to predict the remaining service life of the rotating machinery. The process of filtering extracted features based on feature similarity specifically involves using feature similarity evaluation metrics. The features extracted based on their trend characteristics are then further filtered, including: in: Indicates the first in the source domain The first rotating machine and the target domain Between a certain feature of a rotating machine distance, , These represent the number of rotating machines in the source and target domains, respectively. The The acquisition process is as follows: Assumption C Given two feature sequences of length n and m from the source and target domains respectively, construct a... Given the distance matrix, find the shortest normalized path by regularizing the distance matrix: Where: K is the path length. It is the path regularized in the distance matrix, and WDTW is the minimum distance between the features of the source domain and the target domain. The element in the i-th row and j-th column of the distance matrix is ​​the weighted Euclidean distance between the i-th data point in the source domain dataset and the j-th data point in the target domain dataset. in: Source domain feature sequence The One point, For target domain feature sequences The One point, , Distance weights between and , It depends on the distance of time: ； The migration prediction model consists of a feature extractor, a predictor, and a domain classifier. The feature extractor includes two layers of bidirectional long short-term memory network and one fully connected layer. The predictor and the domain classifier are each composed of three fully connected layers with different parameters. The optimization process uses two optimizers; The optimization objective of the first optimizer is: in: For feature extractor loss, For predictor loss, These are the parameters of the feature extractor loss. These are the parameters of the predictor; The optimization objective of the second optimizer is: in: For the domain classifier loss, For the parameters of the domain classifier, The hyperparameters are used to balance the loss term.

Images