A method for diagnosing faults in a water jet propulsion device

By using a hybrid migration method combining simulation data and model parameters, the problem of low fault diagnosis accuracy in waterjet propulsion devices under varying operating conditions was solved, achieving high-precision fault identification and low-cost diagnostic results.

CN117520884BActive Publication Date: 2026-11-17RES INST 708 OF CHINA STATE SHIPBUILDING CORP
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202311444913.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-01
Publication Date
2026-11-17
Estimated Expiration
2043-11-01

AI Technical Summary

Technical Problem

Water jet propulsion devices have low fault diagnosis accuracy under varying operating conditions, especially due to the scarcity of fault data samples for impeller bearings and blade scraping, making diagnosis difficult. Existing transfer learning methods are insufficient in terms of experimental cost and diagnostic accuracy.

Method used

A hybrid transfer method combining simulation data sample transfer and model parameter transfer is adopted. High-confidence simulation data samples are selected based on similarity to pre-train the DBN network. The parameters of the lower-level network are locked, and only the parameters of the higher-level network are fine-tuned to form a hybrid transfer DBN network model for diagnosis.

Benefits of technology

It improves the fault diagnosis accuracy of water jet propulsion devices under varying operating conditions, achieves efficient fault identification, is superior to single migration diagnosis methods, and reduces test costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117520884B_ABST
    Figure CN117520884B_ABST
Patent Text Reader

Abstract

This invention relates to a fault diagnosis method for a waterjet propulsion device, comprising the following steps: obtaining simulated vibration data samples of the waterjet pump through a dynamic simulation model of the pump, dividing them into training and testing sets; selecting simulation data samples with high confidence from the simulation data samples; establishing a deep learning network (DBN); pre-training the DBN network using the selected simulation data samples, and then using a parameter transfer mechanism to lock the low-level network parameters of the pre-trained DBN network without updating them, and fine-tuning the high-level network parameters of the pre-trained DBN network using data samples from the target domain, forming a trained hybrid transfer DBN network model, and performing diagnostic analysis on the test samples. This method solves the problem of low fault diagnosis accuracy in waterjet propulsion devices under varying operating conditions and with few samples. By combining simulation data sample transfer and model parameter transfer to construct a hybrid transfer mechanism, and applying it to waterjet pump fault diagnosis, the accuracy of fault diagnosis is effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a fault diagnosis method for a water jet propulsion device, and more particularly to a fault diagnosis method and a data migration method for a water jet propulsion device. Background Technology

[0002] Waterjet propulsion systems are prone to frequent malfunctions due to the influence of seawater, sewage, algae, and varying operating conditions during long-term operation, severely impacting their safe and stable operation. Faults such as impeller bearing defects, blade scraping, and impeller imbalance are particularly challenging because data samples are difficult to obtain, resulting in a limited number of fault samples and hindering diagnosis. Common methods for addressing this diagnostic problem with limited fault samples include data augmentation techniques such as resampling, generative adversarial networks (GANs), and transfer learning. However, data augmentation methods often result in insufficient diversity of newly added data samples, while transfer learning methods are widely used because they can transfer knowledge learned from the source domain to the target domain for diagnostic tasks.

[0003] Current transfer learning diagnostic methods typically transfer measured data from one operating condition or one piece of equipment to another operating condition or other similar equipment for diagnostic analysis. However, this transfer diagnostic method requires training with fault data samples under multiple operating conditions, which undoubtedly increases experimental costs. Later, simulation data sample-based transfer methods were also developed and applied to actual mechanical equipment fault diagnosis. This method not only reduces experimental costs but can also simulate simulation data samples of any operating condition and any fault type. However, due to the complexity of influencing factors in mechanical equipment simulation models, especially water jet propulsion devices, simplification during the construction process leads to low consistency between the simulation data samples and measured data of their dynamic response, resulting in insufficient accuracy of transfer diagnosis. Summary of the Invention

[0004] To address the issue of low fault diagnosis accuracy in waterjet propulsion devices under varying operating conditions with limited sample data, a fault diagnosis method for waterjet propulsion devices using a hybrid transfer of simulation data and parameters is proposed. This method combines simulation data sample transfer and model parameter transfer to construct a hybrid transfer mechanism, which is then applied to the fault diagnosis of the waterjet pump (i.e., the waterjet propulsion device). This effectively improves the accuracy of fault diagnosis and solves the problem of limited actual fault samples for the waterjet pump.

[0005] The technical solution of this invention is as follows:

[0006] A method for diagnosing faults in a water jet propulsion device includes the following steps:

[0007] Step 1: Obtain simulated vibration data samples of the spray pump through the dynamic simulation model of the spray pump, and obtain measured vibration data samples of the spray pump using experimental testing methods. Divide the simulated data samples and measured data samples into training sets and test sets respectively. Using the sample transfer mechanism, select simulated data samples with high confidence from the simulated data samples using the similarity method. Establish a DBN network.

[0008] Step 2: Pre-train the DBN network using selected simulation data samples. Then, using the parameter transfer mechanism, lock the low-level network parameters of the pre-trained DBN network without updating them, and fine-tune the high-level network parameters of the pre-trained DBN network using data samples from the target domain to form a trained hybrid transfer DBN network model.

[0009] Step 3: Perform diagnostic analysis on the test samples using the trained DBN hybrid transfer network model.

[0010] Furthermore, the specific steps for selecting simulation data samples with high confidence levels in step one are as follows:

[0011] Vibration signals obtained under different fault types and rotational speeds were obtained using a fluid-structure interaction dynamics simulation model of a jet pump. The confidence level between the simulation data samples in the source domain and the measured fault data samples of the water jet propulsion device was calculated, and the simulation data samples with high confidence were selected as transferable data samples.

[0012] Furthermore, step one also includes randomly sampling the screened simulation data samples in order to obtain sufficient simulation data samples, and using these resampled simulation data samples as source domain data samples for parameter migration.

[0013] Furthermore, the confidence level of the simulation data samples is calculated using the cosine angle method. The formula for calculating the cosine angle between data samples in the source and target domains is:

[0014]

[0015] In the formula Indicates the confidence level. and These are the target domain sample data and the source domain sample data, respectively. The larger S is, the more similar the source domain sample data and the target domain sample data are.

[0016] Furthermore, the feature is that the fine-tuning training of the high-level network parameters of the pre-trained DBN network in step two specifically involves:

[0017] Step 2.1: Initialize the DBN neural network model, where the DBN network model outputs N nodes, representing the number of different fault types;

[0018] Step 2.2: The screened resampled simulation data samples are decomposed using Integrated Empirical Mode Decomposition (EEMD). The M intrinsic mode functions (IMF) signals after decomposition are input into M DBN network models for pre-training and to obtain M feature parameter sets. These sets are then fused in series and input into Softmax for further training.

[0019] Step 2.3: Lock the lower layers of the pre-trained multi-scale DBN network model, i.e. do not update the parameters, and then use the target domain training data samples to fine-tune the parameters of the higher layers of the DBN network, i.e., Softmax, until training is complete.

[0020] Preferably, the fault type in step 2.1 can be one of three fault types: normal, minor fault, and severe fault, then N = 3.

[0021] Preferably, the fault type in step 2.1 can be one of the following fault types: normal, minor fault, moderate fault, and severe fault, then N = 4.

[0022] Preferably, the diagnosis in step three can be clearly and effectively distinguished by establishing a confusion matrix for fault diagnosis and identification.

[0023] Preferably, the simulation data samples in step one are sourced from the following sources:

[0024] Establishing a dynamic simulation model of the spray pump involves creating a fluid-structure interaction dynamic simulation model of the spray pump. Three fault states—normal, moderate wear, and severe wear—are set in the simulation model, and the operating parameters of the simulation model are set according to the actual working conditions to obtain vibration simulation signals under the three fault states.

[0025] Furthermore, the specific sources of the measured samples in step one are as follows:

[0026] Three fault states—normal, moderate wear, and heavy wear—were set in the physical model of the pump. Vibration signals under these three fault states were obtained using a vibration testing system as measured data samples.

[0027] The beneficial effects of this invention are as follows:

[0028] By employing a hybrid migration method that combines simulation data sample migration and model parameter migration, the problem of low diagnostic accuracy caused by the limited number of actual pump fault samples is solved. This method achieves high-precision diagnostic analysis of pump faults, and the diagnostic results are superior to those of single migration diagnostic mechanisms that combine simulation data sample migration and model parameter migration. Attached Figure Description

[0029] Figure 1 This is a flowchart of the migration fault diagnosis process using a mixture of simulation data samples and parameters, as described in this invention.

[0030] Figure 2 This is a diagram showing the fine-tuning of DBN parameters for pre-trained multi-scale feature fusion in this invention.

[0031] Figure 3 This is a confusion matrix diagram for identifying bearing outer ring wear faults according to the present invention. Detailed Implementation

[0032] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.

[0033] To address the problem of low fault diagnosis accuracy in waterjet propulsion devices under varying operating conditions and with a limited sample size, a new method was invented. Figure 1 The method described is a deep belief network (DBN) fault diagnosis approach that combines simulation data and parameter transfer. First, a sample transfer mechanism is used to select high-confidence simulation data samples from the simulation data sample using a similarity method. The DBN network is then pre-trained using these selected simulation data samples. Next, a parameter transfer mechanism is used to lock the low-level network parameters of the pre-trained DBN network, and the high-level network parameters are trained using data samples from the target domain, forming a trained hybrid transfer network model. This model is then used to diagnose and analyze test samples.

[0034] The specific algorithm is as follows:

[0035] (1) Source Domain Data Sample Screening. First, the confidence level between the simulated data samples in the source domain (obtained using the pump-jet fluid-structure interaction dynamics simulation model under different fault types (normal, moderate wear of the bearing outer ring, and severe wear of the bearing outer ring) and different speeds) and the measured data samples in the target domain (measured fault data samples of the water jet propulsion device) is calculated. Simulated data samples with a confidence level greater than 0.8 are selected as transferable data samples. Here, the cosine angle method is used to calculate the confidence level of the simulated data samples. That is, the formula for calculating the cosine angle between the data samples in the source domain and the target domain is:

[0036]

[0037] In the formula Indicates the confidence level. and These are the target domain sample data and the source domain sample data, respectively. The larger S is, the more similar the source domain sample data and the target domain sample data are.

[0038] And establish a DBN network;

[0039] (2) Random sampling. Random sampling is performed on the screened simulation data samples in order to obtain sufficient simulation data samples, and these resampled simulation data samples are used as source domain data samples for parameter migration.

[0040] (3) Model pre-training and parameter fine-tuning. The multi-scale DBN network model is pre-trained using resampled simulation data samples to obtain the initial network model parameters. Then, based on the parameter transfer mechanism, the shallow layers of the pre-trained DBN network model are locked and no parameter updates are performed. The high-level parameters of the DBN network model, namely Softmax, are updated and learned only using the training data samples of the target domain until training is complete.

[0041] (4) Fault diagnosis and analysis. The trained DBN neural network model is used to perform diagnostic analysis on the test data samples in the target domain to verify the effectiveness of the diagnostic model.

[0042] The algorithm flow for fine-tuning the pre-trained network model is as follows: Figure 2 As shown, the specific process is as follows:

[0043] ① Initialize the DBN neural network model, where the number of output nodes of the DBN network model is N, representing the number of different fault types. For example, if there are three fault types, such as impeller normal, impeller slight scratch, and impeller heavy scratch, then N=3; other fault types can be deduced in the same way, such as bearing normal, outer ring moderate wear, and outer ring heavy wear, then N=3.

[0044] ② The screened resampled simulation data samples are decomposed using Ensemble Empirical Mode Decomposition (EEMD). The M intrinsic mode function (IMF) signals after decomposition are input into M DBN network models for pre-training and to obtain M feature parameter sets. Then, they are fused in series and input into Softmax for training.

[0045] ③ Lock the lower layers of the pre-trained multi-scale DBN network model, i.e. do not update the parameters, and then use the target domain training data samples to fine-tune the parameters of the higher layers of the DBN network, i.e., Softmax, until training is complete.

[0046] To verify the effectiveness of the fault diagnosis method of this invention, a fault diagnosis analysis was performed on the bearing at the impeller of the spray pump. Table 1 below shows the statistical data of three types of wear faults on the outer ring of the impeller bearing:

[0047] Table 1. Statistical data of three types of wear failures on the outer ring of impeller bearings.

[0048]

[0049] In the source domain (simulation data), there are 600 data samples for each fault type (three fault types of the bearing outer ring), 10 training samples for each fault type in the target domain, and 200 data samples for each fault type in the target domain test sample set. Figure 3 The confusion matrix of the multi-scale DBN diagnostic method using simulated data samples and parameter hybrid migration for the diagnosis and identification of impeller bearing outer ring wear faults is presented. As can be seen from the figure, the method can identify the three types of bearing faults with high accuracy, thus demonstrating that the invention can effectively solve the problem of low diagnostic accuracy caused by the scarcity of actual fault samples by utilizing simulated data samples.

[0050] The above-described embodiments are merely one implementation of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention should be determined by the appended claims.

Claims

1. A method for diagnosing faults in a waterjet propulsion device, characterized in that, Includes the following steps: Step 1: Obtain simulated vibration data samples of the spray pump through the dynamic simulation model of the spray pump, and obtain measured vibration data samples of the spray pump using experimental testing methods. Divide the simulated data samples and measured data samples into training sets and test sets respectively. Using the sample transfer mechanism, select simulated data samples with high confidence from the simulated data samples using the similarity method. Establish a DBN network. Step 2: Pre-train the DBN network using selected simulation data samples. Then, using the parameter transfer mechanism, lock the low-level network parameters of the pre-trained DBN network without updating them, and fine-tune the high-level network parameters of the pre-trained DBN network using data samples from the target domain to form a trained hybrid transfer DBN network model. Step 3: Perform diagnostic analysis on the test samples using the trained DBN hybrid transfer network model; Specifically, the fine-tuning of the high-level network parameters of the pre-trained DBN network in step two involves: Step 2.1: Initialize the DBN neural network model, where the DBN network model outputs N nodes, representing the number of different fault types; Step 2.2: The screened resampled simulation data samples are decomposed using Integrated Empirical Mode Decomposition (EEMD). The M intrinsic mode functions (IMF) signals after decomposition are input into M DBN network models for pre-training and to obtain M feature parameter sets. These sets are then fused in series and input into Softmax for further training. Step 2.3: Lock the lower layers of the pre-trained multi-scale DBN network model, i.e. do not update the parameters, and then use the target domain training data samples to fine-tune the parameters of the higher layers of the DBN network, i.e., Softmax, until training is complete.

2. The fault diagnosis method for a waterjet propulsion device according to claim 1, characterized in that, The specific steps for selecting simulation data samples with high confidence in step one are as follows: Vibration signals obtained under different fault types and rotational speeds were obtained using a fluid-structure interaction dynamics simulation model of a jet pump. The confidence level between the simulation data samples in the source domain and the measured fault data samples of the water jet propulsion device was calculated, and the simulation data samples with high confidence were selected as transferable data samples.

3. The fault diagnosis method for a waterjet propulsion device according to claim 2, characterized in that, Step one also includes randomly sampling the filtered simulation data samples in order to obtain sufficient simulation data samples, and using these resampled simulation data samples as source domain data samples for parameter migration.

4. The fault diagnosis method for a waterjet propulsion device according to claim 2, characterized in that, The confidence level of the simulation data samples is calculated using the cosine angle method. Specifically, the formula for calculating the cosine angle between data samples in the source and target domains is: (1) In the formula Indicates the confidence level. and These are the target domain sample data and the source domain sample data, respectively. The larger S is, the more similar the source domain sample data and the target domain sample data are.

5. The fault diagnosis method for a waterjet propulsion device according to claim 1, characterized in that, The fault type in step 2.1 can be one of three types: normal, minor fault, or severe fault, so N=3.

6. The fault diagnosis method for a waterjet propulsion device according to claim 1, characterized in that, The fault type in step 2.1 can be one of four types: normal, minor fault, moderate fault, and severe fault, so N=4.

7. The fault diagnosis method for a waterjet propulsion device according to claim 1, characterized in that, The diagnosis in step three can be clearly and effectively distinguished by establishing a confusion matrix for fault diagnosis and identification.

8. The fault diagnosis method for a waterjet propulsion device according to claim 1, characterized in that, The simulation data samples in step one specifically come from the following sources: Establishing a dynamic simulation model of the spray pump involves creating a fluid-structure interaction dynamic simulation model of the spray pump. Three fault states—normal, moderate wear, and severe wear—are set in the simulation model, and the operating parameters of the simulation model are set according to the actual working conditions to obtain vibration simulation signals under the three fault states.

9. The fault diagnosis method for a waterjet propulsion device according to claim 1, characterized in that, The specific sources of the measured samples in Step 1 are as follows: Three fault states—normal, moderate wear, and heavy wear—were set in the physical model of the pump. Vibration signals under these three fault states were obtained using a vibration testing system as measured data samples.

Citation Information

Patent Citations

  • Bearing fault diagnosis method based on EEMD-AR model and DBN

    CN109141884A

  • Induction motor fault diagnosis system based on finite element simulation and symmetric feature migration

    CN114118149A