Rotating machine fault diagnosis method based on feature transfer learning

Abstract

The invention discloses a rotating machine fault diagnosis method based on feature transfer learning, and the method comprises the steps: constructing a data set through a multi-sensor fault frequency spectrum, minimizing the difference between main data of a source domain and a target domain through employing the idea of feature transfer learning, and learning a main classifier on the source domain, and learning an auxiliary classifier on an auxiliary data set of the target domain according to a prediction result of the main classifier for the main data of the target domain, and finally carrying out weighted fusion on the two classifiers to obtain a final prediction model. The method still has relatively high diagnosis precision under the condition of rare target working condition samples.

Description

Technical Field

[0001] The invention relates to a rotating machinery fault diagnosis method based on feature transfer learning, and belongs to the technical field of artificial intelligence fault diagnosis. Background Art

[0002] Rotating machinery plays an indispensable role in modern industry, and its operation covers many fields, including energy production, chemical industry, manufacturing industry and transportation. However, with long-term operation, rotating machinery is subjected to huge loads and motion stress during operation, which will inevitably lead to wear, deformation and deterioration of mechanical parts. Long-term use will threaten its stability, safety and efficiency. Therefore, it is very necessary to carry out research on intelligent fault diagnosis of rotating machinery, and monitor the operating status of the equipment at all times, so as to better predict the equipment before failure.

[0003] Traditional rotating machinery fault diagnosis usually only imports a single sensor data sample, which has certain limitations and one-sidedness. The present invention patent designs a fault diagnosis method for feature transfer learning of multiple sensors. The technical methods involved in the present invention patent include fast Fourier transform, subspace alignment, feature transfer learning, distribution adaptation, and coupled classifier learning. Summary of the invention

[0004] The purpose of the present invention is to solve the problems of single sensor sample limitation and inability to generalize across working conditions in existing rotating machinery fault diagnosis methods, and to provide a rotating machinery fault diagnosis method based on feature transfer learning.

[0005] The objective of the present invention is achieved through the following technical solutions:

[0006] A method for diagnosing rotating machinery faults based on feature transfer learning of the present invention specifically comprises the following steps:

[0007] Step 1: First, obtain the vibration sample data of the centrifugal pump driving end perpendicular to the bearing axis with a known label as the main sensor signal with a known source working condition label;

[0008] Step 2: Change the working condition of the centrifugal pump, obtain the vibration sample data of the centrifugal pump driving end perpendicular to the bearing axis direction and the non-driving end perpendicular to the bearing axis direction with unknown labels, and use them as the main sensor signal of the target working condition label unknown and the auxiliary sensor signal of the target working condition label unknown respectively;

[0009] Step 3: Smoothing, filtering, sampling, fast Fourier transform, and standardization are performed on the source working condition main sensor signal with known labels and the target working condition main sensor and auxiliary sensor signals with unknown labels, thereby generating the source domain main data X, the target domain main data X1, and the target domain auxiliary data X2 respectively;

[0010] Step 4: Map the source domain main data X and the target domain main data X1 to a common subspace through a subspace alignment method, reduce the dimensions to Z and Z1, and reduce the auxiliary data set X2 of the target domain to Z2;

[0011] Step 5: Use Z and the labeled classifier to predict Z1 to obtain the label of the target domain sample, select the matching kernel function to perform nonlinear prediction on Z, Z1, and Z2 to obtain the kernel matrices K and K';

[0012] Step 6: Minimize the difference between Z and Z1 through conditional distribution adaptiveness, and learn the main classifier f on the source domain according to the principle of structural risk minimization, whose parameter vector is a, and learn the auxiliary classifier f' on the target domain, whose parameter vector is b;

[0013] Step 7: Use an alternating iterative strategy to continuously update a and b until the maximum number of iterations is reached, and perform weighted fusion on the predictions of the main classifier and the auxiliary classifier to obtain the final result.

[0014] The sample interception in step three is completed by using the overlapping sampling method.

[0015] The standardization in step three is accomplished using the z-score method.

[0016] The dimension reduction method in step 4 adopts the principal component analysis (PCA) method.

[0017] Beneficial Effects

[0018] (1) Overcoming the limitations of traditional supervised learning. Since supervised learning relies on a large amount of labeled training data, it means that the data set needs to be manually labeled, which consumes time and resources. When there is a difference between the training data and the actual application scenario, the model may not perform well. The supervised learning model also has the disadvantage of being prone to overfitting, that is, it performs well on the training data but poorly on new data. Transfer learning is a kind of unsupervised learning. The training data used does not need to be manually labeled, and it can be applied to complex environments with environmental noise, which is more convenient for practical application.

[0019] (2) Subspace alignment is used to project the main data of the source domain and the target domain into a low-dimensional common subspace, which can obtain better feature representation, improve discriminability while reducing feature dimensionality, and reduce distribution differences to a certain extent.

[0020] (3) This method uses a deep learning method of feature transfer learning, which can realize fault reuse under different working conditions, and provides value for future research on equipment fault diagnosis of different models and different environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 The overall framework diagram of the diagnostic method of the present invention;

[0022] Figure 2 This is a diagram showing the effect of the method of the present invention being applied to a centrifugal pump bearing. DETAILED DESCRIPTION

[0023] The present invention is further described below with reference to the accompanying drawings and embodiments.

[0024] Example

[0025] The present invention provides a method for diagnosing rotating machinery faults based on feature transfer learning, and the overall architecture diagram is as follows: Figure 1 As shown, the specific steps include:

[0026] Step 1: First, conduct a fault experiment and set the bearing to have 10 states: Normal, inner ring fault 1 (IR007), inner ring fault 2 (IR0014), inner ring fault 3 (IR0021), outer ring fault 1 (OR007@6), outer ring fault 2 (OR0014@6), outer ring fault 3 (OR0021@6), rolling element fault 1 (B007), rolling element fault 2 (B0014), rolling element fault 3 (B0021), see Table 1 for details, obtain the driving end vibration data of the centrifugal pump at a motor speed of 960 rpm, as the source working condition label known main sensor signal;

[0027] Step 2: A main sensor is arranged at the driving end of the centrifugal pump perpendicular to the bearing axis, and an auxiliary sensor is arranged at the non-driving end of the centrifugal pump perpendicular to the bearing axis. Nine bearings in different states are randomly selected at the driving end to collect the bearing vibration data of the driving end and non-driving end at a motor speed of 1278 rpm, which are respectively used as the target working condition label unknown main sensor signal and the target working condition label unknown auxiliary sensor signal;

[0028] Step 3: Filter and denoise the main sensor signals of the source working condition with known labels and the main sensor and auxiliary sensor signals of the target working condition with unknown labels to remove environmental interference, then perform sample interception in an overlapping sampling manner, and then perform fast Fourier transform to obtain the corresponding spectrum, and use the z-score method to standardize it, and finally obtain the main data set X of the source domain, the main data set X1 of the target domain, and the auxiliary data set X2 of the target domain;

[0029] Step 3: Map the source domain main data X and the target domain main data X1 to a common subspace through a subspace alignment method, reduce the dimensions to Z and Z1, and reduce the dimension of the auxiliary data set X2 of the target domain to Z2. The dimension reduction method here adopts the principal component analysis (PCA) method;

[0030] Step 4: Use Z and the labeled classifier to predict Z1 to obtain the label of the target domain sample, select the matching kernel function to perform nonlinear prediction on Z, Z1, and Z2 to obtain the kernel matrices K and K';

[0031] Step 5: Minimize the difference between Z and Z1 through conditional distribution adaptiveness, and learn the main classifier f on the source domain according to the principle of structural risk minimization, whose parameter vector is a, and learn the auxiliary classifier f' on the target domain, whose parameter vector is b;

[0032] Step 6: Use the alternating iteration strategy to continuously update a and b until the maximum number of iterations is reached, and perform weighted fusion on the predictions of the main classifier and the auxiliary classifier to obtain the final result. The final prediction result of the target domain is as follows: Figure 2 As shown, the effect is better.

[0033] Table 1 Fault category classification table

[0034] Classification Category Fault location Fault diameter / mm Normal healthy / IR007 Inner Circle 0.1778 IR014 Inner Circle 0.3556 IR021 Inner Circle 0.5534 OR007@6 Outer ring 0.1778 OR014@6 Outer ring 0.3556 OR021@6 Outer ring 0.5534 B007 Rolling element 0.1778 B014 Rolling element 0.3556 B021 Rolling element 0.5534

Claims

1. A method for fault diagnosis of rotating machinery based on feature transfer learning, characterized in that: The specific steps include: Step 1: First, obtain the vibration sample data of the centrifugal pump driving end perpendicular to the bearing axis with a known label as the main sensor signal with a known source working condition label; Step 2: Change the working condition of the centrifugal pump, obtain the vibration sample data of the centrifugal pump driving end perpendicular to the bearing axis direction and the non-driving end perpendicular to the bearing axis direction with unknown labels, and use them as the main sensor signal of the target working condition label unknown and the auxiliary sensor signal of the target working condition label unknown respectively; Step 3: Smoothing, filtering, sampling, fast Fourier transform, and standardization are performed on the source working condition main sensor signal with known labels and the target working condition main sensor and auxiliary sensor signals with unknown labels, thereby generating the source domain main data X, the target domain main data X1, and the target domain auxiliary data X2 respectively; Step 4: Map the source domain main data X and the target domain main data X1 to a common subspace through a subspace alignment method, reduce the dimensions to Z and Z1, and reduce the auxiliary data set X2 of the target domain to Z2; Step 5: Use Z and the labeled classifier to predict Z1 to obtain the label of the target domain sample, select the matching kernel function to perform nonlinear prediction on Z, Z1, and Z2 to obtain the kernel matrices K and K'; Step 6: Minimize the difference between Z and Z1 through conditional distribution adaptiveness, and learn the main classifier f on the source domain according to the principle of structural risk minimization, whose parameter vector is a, and learn the auxiliary classifier f' on the target domain, whose parameter vector is b; Step 7: Use an alternating iterative strategy to continuously update a and b until the maximum number of iterations is reached, and perform weighted fusion on the predictions of the main classifier and the auxiliary classifier to obtain the final result.

2. A method for diagnosing rotating machinery faults based on feature transfer learning as claimed in claim 1, characterized in that: The sample interception in step three is completed by using the overlapping sampling method.

3. A method for diagnosing rotating machinery faults based on feature transfer learning as claimed in claim 1, characterized in that: The standardization in step three is accomplished using the z-score method.

4. A method for diagnosing rotating machinery faults based on feature transfer learning as claimed in claim 1, characterized in that: The dimension reduction method in step 4 adopts the principal component analysis method.

Images