Intelligent diagnosis method for load fault of ship vibration isolation device

Through the neural network model optimized based on Bayesian network and Markov chain-Monte Carlo algorithm, the problem of low load fault detection accuracy of ship vibration isolation devices is solved, and more efficient fault prediction and operation and maintenance management are achieved.

CN119004262APending Publication Date: 2024-11-22JIANGSU YUANWANG INSTR +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411108163.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-11-22

AI Technical Summary

Technical Problem

In the prior art, the fault detection of the ship vibration isolation device fails to predict the load object based on data information, resulting in low prediction accuracy and increasing the time and cost of equipment operation and maintenance.

Method used

Using an intelligent diagnostic method based on vibration isolation device data information, the preset neural network model is trained by obtaining and processing load working parameters, and the Bayesian network and Markov chain-Monte Carlo algorithm are used to optimize the model parameter weights and obtain the load Equipment failure classification and level information.

Benefits of technology

It improves the accuracy and timeliness of fault prediction, ensures that the working performance of ship equipment is fully guaranteed, and reduces operation and maintenance costs.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119004262A_ABST
    Figure CN119004262A_ABST
Patent Text Reader

Abstract

The invention provides an intelligent diagnosis method for a load fault of a ship vibration isolation device, and aims to solve the current situation that fault prediction is not performed on a load object based on data information of the vibration isolation device in the prior art and the problem that an existing neural network model is low in prediction precision due to few fault samples. The method mainly comprises the steps of obtaining a training sample set, optimizing a sample data set and a neural network model parameter weight and the like, the method automatically obtains an initial working parameter sample of the vibration isolation device in real time, optimizes the sample data set and the neural network model parameter weight through a preset Bayesian network and an MCMC algorithm, and obtains the initial working parameter sample of the vibration isolation device. And the preset neural network model perfects the training sample confidence coefficient and the model parameter distribution through the Bayesian network and the MCMC algorithm, so that the fault information of the load equipment of the vibration isolation device is obtained based on the preset neural network model trained by the method, and the fault detection accuracy of the load equipment of the vibration isolation device is improved. The accuracy and timeliness of fault prediction can be effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of fault diagnosis, and in particular to an intelligent diagnosis method for load faults of ship vibration isolation devices. Background Art

[0002] The ship airbag vibration isolation device is an important vibration reduction technology for ship mechanical equipment. It is mainly used to reduce the impact of vibration and noise generated by the operation of large power equipment such as the main and auxiliary engines of the ship on the hull structure and precision instruments on board, thereby improving the comfort, safety and work efficiency of the ship.

[0003] There are defects in the fault detection of ship vibration isolation devices in the prior art, which are specifically manifested in that the fault prediction of the load object is not performed based on the data information of the ship vibration isolation device. When the existing fault prediction neural network model predicts the load fault of the vibration isolation device, the prediction accuracy is low due to the small number of fault samples, which leads to the working performance of the main equipment of the ship cannot be fully guaranteed, and the time and cost investment in equipment operation and maintenance are increased. Summary of the invention

[0004] In response to the shortcomings of the above-mentioned prior art, the present application provides an intelligent diagnosis method for load faults of ship vibration isolation devices, which is a technology for predicting faults of load objects based on vibration isolation device data information, and greatly improves the prediction accuracy of existing neural network models when the number of fault samples is small.

[0005] The technical solution adopted by the present invention to solve the above technical problems is as follows:

[0006] A method for intelligently diagnosing load failure of a ship vibration isolation device comprises the following steps:

[0007] S1. Obtaining the working parameters of the vibration isolation device load;

[0008] S2, obtaining a plurality of training samples including load working parameters of the vibration isolation device and constructing a training sample set;

[0009] S3. According to the training sample set, a preset neural network model is obtained through training;

[0010] S4. Preset a Bayesian network model according to the fault information under actual working conditions, update the training sample set in step S2 to obtain a new preset neural network model, and obtain the parameter weight probability distribution of the updated preset neural network model;

[0011] S5. Optimizing the parameter weights of the updated preset neural network model through the MCMC algorithm;

[0012] S6. Obtain predicted fault information of the load equipment of the vibration isolation device through the final preset neural network model.

[0013] Furthermore, the training sample set in step S2 includes a plurality of training samples including working parameters, each training sample includes labeled sample fault information, and the sample fault information is obtained according to equipment fault classification information and fault level information under preset empirical values.

[0014] Furthermore, the fault classification and fault level information includes equipment fault location, equipment fault category and equipment fault probability.

[0015] Furthermore, the method for obtaining the training sample set in step S2 is: firstly obtaining an initial sample set, processing the initial training sample set using a preset feature engineering method, and using the processed initial training sample set as the training sample set.

[0016] Furthermore, the steps of processing the initial training sample set by the preset feature engineering method include: data preprocessing, feature selection and dimension compression.

[0017] Furthermore, the preset neural network model includes an input layer for inputting a training sample set, a hidden layer for repeatedly training to obtain weights and bias parameters, and an output layer for outputting predicted fault information of a load device of the vibration isolation device.

[0018] Furthermore, the hidden layer repeatedly trains the input data and the output data by means of gradient descent and back propagation.

[0019] Furthermore, in step S5, the method for optimizing the parameter weights of the updated preset neural network model is to calculate the posterior probability distribution of the preset neural network model parameters based on the Bayesian network statistical model, use the MCMC algorithm to obtain samples that obey this distribution, and estimate the expectation and variance values ​​of the posterior distribution. If the expectation and variance values ​​are within the set range, the neural network model parameter weights are determined to be feasible; otherwise, repeat step S4.

[0020] Furthermore, the method for updating the training sample set in step S4 is to perform an inverse probability operation based on the fault information under actual working conditions and a preset Bayesian network model.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] The novel ship intelligent vibration isolation device load fault diagnosis method provided by the present invention automatically obtains the initial working parameter samples of the vibration isolation device in real time, and respectively optimizes the sample data set and the neural network model parameter weights through the preset Bayesian network and MCMC (Markov Chain-Monte Carlo) algorithm, and inputs the updated data set into the updated preset neural network model to obtain the fault classification information and fault level information of the vibration isolation device load equipment; since the preset neural network model is trained based on multiple training samples containing working parameters, and the training sample confidence and model parameter distribution are improved through the Bayesian network and MCMC algorithm, the preset neural network model trained by the above method obtains the fault information of the vibration isolation device load equipment, which can effectively improve the accuracy and timeliness of the load equipment fault prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a flow chart of the present invention;

[0024] Figure 2 It is a schematic diagram of the structure of the preset neural network model in the present invention;

[0025] Figure 3 is a system block diagram of the present invention;

[0026] Figure 4 Schematic diagram of the processing flow of the preset feature engineering method in the present invention. DETAILED DESCRIPTION

[0027] The technical solution of the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0028] The present invention provides an intelligent diagnosis method for load faults of ship vibration isolation devices, aiming to solve the problem that the prior art does not predict faults of load objects based on vibration isolation device data information, and the problem that the existing neural network model has low prediction accuracy due to the small number of fault samples.

[0029] like Figures 1 to 4 As shown, the execution subject of the present invention can be a computer, a server, a processor, a mobile terminal or other device that can perform data processing. The method includes:

[0030] S1. Obtaining the working parameters of the vibration isolation device load;

[0031] S2, obtaining a plurality of training samples including load working parameters of the vibration isolation device and constructing a training sample set;

[0032] S3. According to the training sample set, a preset neural network model is obtained through training;

[0033] S4. Preset a Bayesian network model according to the fault information under actual working conditions, update the training sample set in step S2 to obtain a new preset neural network model, and obtain the parameter weight probability distribution of the updated preset neural network model;

[0034] S5. Optimizing the parameter weights of the updated preset neural network model through the MCMC algorithm;

[0035] S6. Obtain predicted fault information of the load equipment of the vibration isolation device through the final preset neural network model.

[0036] In one embodiment of the present invention, for step S1, the working parameters can be obtained by a signal detection unit installed at a corresponding position of the raft frame of the vibration isolation device, such as a vibration sensor, a displacement sensor, etc.

[0037] In one embodiment of the present invention, for step S2, the training sample set may include: multiple training samples containing working parameters, each training sample may include labeled sample fault information, the sample fault information may include equipment fault classification information and fault level information under preset empirical values, such as vibration data, displacement data and other working parameters and corresponding equipment fault classification information and fault level information.

[0038] Among them, the method for obtaining the training sample set is to first obtain an initial sample set, use a preset feature engineering method to process the initial training sample set, and use the processed initial training sample set as the training sample set; the initial training sample set may include multiple initial training samples, each training sample may include the working parameters as described above, and the present application does not limit the number of training samples included in the initial training sample set, and corresponding settings can be made according to actual application conditions.

[0039] The above-mentioned preset feature engineering method may include: data preprocessing, feature selection, dimension compression and other data processing processes, so that the processed training sample set can be better used for the training of the preset neural network model. Among them, data preprocessing can include the processing processes of dimensional unification, missing value calculation and unique hot code representation of qualitative characteristics. The dimensional unification processing process can include normal distribution standardization method, interval scaling method using the boundary of feature value interval and normalization method of processing data according to the rows of feature matrix, etc. By performing dimensional unification, training samples of different specifications in the initial training sample set can be converted to the same specification; the feature selection processing process can include variance selection method, correlation coefficient method, recursive feature elimination method and model selection method, etc. Through feature selection, meaningful features can be selected from the initial training sample set as training samples of the neural network model. Optionally, whether the features of each sample in the initial training sample set diverge and the correlation between each feature and the target can be considered; dimensionality compression can sample PCA principal component analysis method. Through dimensionality compression, the initial training sample set can be mapped to a sample space with lower dimensionality to avoid the problem that the feature matrix of the initial training sample set is too large, resulting in large calculation amount and long training time.

[0040] In one embodiment of the present invention, for step S3, Figure 2 As shown, the preset neural network model may include an input layer (A0, A1...At), a hidden layer (W0, W1, W2), and an output layer (Z0, Z1...Zt), wherein the input layer is used for inputting a training sample set, and the training sample set may include: a plurality of training samples containing working parameters, each training sample may include labeled sample fault information, and the sample fault information may include equipment fault classification information and fault level information under preset empirical values, such as working parameters such as vibration data and displacement data, and corresponding equipment fault classification information and fault level information.

[0041] Among them, the above-mentioned hidden layer can repeatedly train the input and output through the gradient descent and back propagation methods to make the loss function at the minimum value and obtain parameters such as weights and biases; the output layer can be used to output corresponding fault information, such as faulty components, fault locations and fault classification information.

[0042] In this embodiment, the working parameter is a vibration signal as an example for explanation; for example, the load device of the vibration isolation device is a pump, and the corresponding working parameter is a vibration signal. Optionally, a training sample set can be obtained based on the load device. The training sample set can include vibration signals of different frequencies, such as low frequency 1#, low frequency 2#, low frequency 3#, medium frequency 1#, medium frequency 2#, and high frequency 1#. The vibration signal of each frequency can correspond to the equipment fault classification information and fault level information under the preset experience value. The equipment fault classification information and fault level information under the preset experience value can be parameters set according to existing experience.

[0043] Optionally, the fault classification information can be divided into base looseness, rotor misalignment, rotor mass imbalance, cavitation, etc., and the fault level information can be a fault probability. According to existing experience, the probability of a loose base in the training sample set causing low frequency 1# to occur is 75%, the probability of causing low frequency 2# to occur is 60%, and the probability of causing low frequency 3# to occur is 85%; the probability of a rotor misalignment causing low frequency 2# to occur is 73%, the probability of causing medium frequency 1# to occur is 65%, and the probability of causing medium frequency 2# to occur is 56%; the probability of a rotor mass imbalance causing low frequency 2# to occur is 68%, the probability of causing high frequency 1# to occur is 55%, and the probability of causing low frequency 1# to occur is 83%; the probability of a cavitation causing low frequency 3# to occur is 63%, the probability of causing high frequency 1# to occur is 35%, and the probability of causing medium frequency 1# to occur is 86%. According to the above preset experience values, each training sample in the training sample set can be labeled, but not limited to this.

[0044] After labeling, the labeled training sample set can be used to train and obtain a preset neural network model, so that when the preset neural network model obtained by training is used to predict the fault information of the load equipment, it can be predicted according to the existing experience value, thereby improving the accuracy of predicting the fault information of the load equipment; wherein, the preset neural network model can be trained according to multiple training sample sets containing working parameters, and the working parameters also include the real-time displacement signal and vibration signal of the load; optionally, the preset neural network model can be a recurrent neural network (RNN), a deep neural network (DNN), etc., which is not limited in this application.

[0045] In one embodiment of the present invention, for step S4, a preset Bayesian network model is trained and generated based on a training sample set; the Bayesian network model is preset based on fault information under actual working conditions, and the sample fault information marked by each training sample in the training sample set is updated to obtain an updated training sample set; based on the updated training sample set, the parameter weights of the preset neural network model are updated, and the parameter weight probability distribution of the neural network model is obtained.

[0046] Among them, the fault information of the load equipment under actual working conditions is the fault information corresponding to the load equipment during the actual working process. For example, when a load equipment fails, the staff can know the fault information of the load equipment and obtain the fault information under the actual working conditions. Then, according to the fault information under the actual working conditions and the preset Bayesian network model, the sample fault information marked by each training sample in the training sample set can be updated, so that the preset neural network model can be updated according to the updated training sample set, and the updated preset neural network model can be obtained. When the fault information of the load equipment is obtained by using the updated preset neural network model, it can be closer to the fault information under the actual working conditions of the load equipment, thereby improving the accuracy of the prediction. Compared with directly updating the training sample set according to the fault information of the load equipment under the actual working conditions, due to the actual working The frequency of fault information of the load equipment obtained under actual working conditions is limited. If the preset neural network model is directly trained and updated based on the limited fault information under actual working conditions, it is impossible to more accurately annotate the sample fault information of the training sample set, which will make the predicted fault less consistent with the actual working conditions. However, the present application uses a preset Bayesian network model, which can update the sample fault information annotated by each training sample in the training sample set through a small amount of fault information under actual working conditions, and then update and train the preset neural network model based on the updated training sample set, so that when the updated preset neural network model is used to obtain the fault information of the load equipment, the accuracy of the prediction can be improved, that is, it is more consistent with the fault information of the actual working conditions.

[0047] In one embodiment of the present invention, for step S5, in order to optimize the model parameter weights, reduce the number of training times, and improve the model fit, the method will also calculate the posterior probability distribution of the neural network model parameters based on the Bayesian network statistical model. The MCMC (Markov Chain-Monte Carlo) algorithm is used to obtain samples that obey this distribution and estimate the expectation and variance of the model parameters. Thus, the feasibility of the currently selected parameter weights can be judged. Finally, if the expected value and variance value of the calculated model parameters are within the set range, it is inferred that the weight of the neural network model is ideal. Otherwise, continue to use the Bayesian model to optimize the training set, and combine the means of manual fine-tuning to continue to update the network model parameter distribution.

[0048] For example, assuming the load device is a pump, Figure 3 As shown, the corresponding working parameter is the vibration signal. The working parameters (for example, vibration signals) of the load equipment are obtained under actual working conditions (actual operating state). Since it is an actual working process, the fault information under the actual working conditions corresponding to the working parameters is known. Then, an “inverse probability” operation can be performed based on the fault information under the actual working conditions and the preset Bayesian network model, and the sample fault information marked by each training sample in the training sample set is continuously updated. Since both the preset experience value and the fault information under the actual working conditions are taken into account, when the updated preset neural network model obtained through the updated training is used to detect the fault information of the load equipment, the inferred fault information can be closer to the fault information under the actual working conditions, which can further improve the accuracy of predicting the fault information of the load equipment.

[0049] It should be further explained that, according to the Bayesian network statistical model, the posterior probability distribution of the neural network model parameters can be calculated. The Markov chain-Monte Carlo algorithm (MCMC) is used to obtain samples that obey this distribution, so as to estimate the expectation and variance of the posterior distribution, that is, the fitting value and uncertainty of the neural network model parameters. If the expectation and variance values ​​are within the set range, it can be inferred that the neural network model parameter weights are relatively ideal; otherwise, continue the method described in step S4, and continue to update the neural network model parameter distribution in combination with manual fine-tuning.

[0050] In particular, the fault level information referred to in this application includes one or more combinations of the following: equipment fault location, equipment fault signal range, and equipment failure probability. Among them, the equipment fault location can be used to indicate the specific fault location of the load equipment, such as the above-mentioned base, rotor, etc.; the equipment fault signal range can indicate the range interval of the fault signal when the equipment fails, such as the frequency interval range of the vibration signal, etc.; the equipment failure probability can indicate the equipment failure probability corresponding to each equipment fault location, for example, the probability of base failure is 80%, and the probability of rotor failure is 60%. Of course, it should be noted that this application does not limit the fault level information here, and it can also include the cause of the equipment failure, etc.

[0051] As another implementation scheme for outputting predicted fault information, in addition to inputting the working parameters into the preset neural network model to obtain the fault information of the load equipment; the method also includes: obtaining a 3D model of the load equipment; marking the fault information of the load equipment on the 3D model, that is, after obtaining the fault information of the load equipment, the fault information can be displayed through the 3D model of the load equipment to realize the visualization of the fault information of the load equipment. The visualization method can be displayed through a computer, a mobile terminal, etc. For example, a user can use a mobile handheld device, such as a smart phone, a tablet computer, or other mobile terminal, in a place covered by a local area network WIFI, through an application software program or a WEB browsing method, to view the fault information of the load equipment in 3D visualization, which can greatly enhance the user's human-computer interaction experience. Optionally, a Web Graphics Library (WebGL) can be used to visualize the load equipment, and then the fault information of the load equipment can be further visualized.

[0052] The above description is an explanation of the present invention, not a limitation of the present invention. The scope of the present invention is defined in the claims. Any form of modification may be made within the scope of protection of the present invention.

Claims

1. An intelligent diagnosis method for load fault of ship vibration isolation device, characterized by: The steps include: S1. Obtaining the working parameters of the vibration isolation device load; S2, obtaining a plurality of training samples including load working parameters of the vibration isolation device and constructing a training sample set; S3. According to the training sample set, a preset neural network model is obtained through training; S4. Preset a Bayesian network model according to the fault information under actual working conditions, update the training sample set in step S2 to obtain a new preset neural network model, and obtain the parameter weight probability distribution of the updated preset neural network model; S5. Optimizing the parameter weights of the updated preset neural network model through the MCMC algorithm; S6. Obtain the final preset neural network model to obtain the predicted fault information of the load equipment of the vibration isolation device.

2. The intelligent diagnosis method for load fault of a ship vibration isolation device according to claim 1, characterized in that: The training sample set in step S2 includes a plurality of training samples including working parameters, each training sample includes labeled sample fault information, and the sample fault information is obtained according to equipment fault classification information and fault level information under preset empirical values.

3. The intelligent diagnosis method for load fault of a ship vibration isolation device according to claim 2, characterized in that: The fault classification and fault level information includes equipment fault location, equipment fault category and equipment fault probability.

4. The intelligent diagnosis method for load fault of a ship vibration isolation device according to claim 1, characterized in that: The method for obtaining the training sample set in step S2 is: firstly obtain an initial sample set, process the initial training sample set using a preset feature engineering method, and use the processed initial training sample set as the training sample set.

5. The intelligent diagnosis method for load fault of a ship vibration isolation device according to claim 4, characterized in that: The steps of processing the initial training sample set by the preset feature engineering method include: data preprocessing, feature selection and dimension compression.

6. The intelligent diagnosis method for load fault of a ship vibration isolation device according to claim 1, characterized in that: The preset neural network model includes an input layer for inputting a training sample set, a hidden layer for repeatedly training to obtain weight and bias parameters, and an output layer for outputting predicted fault information of a load device of the vibration isolation device.

7. The intelligent diagnosis method for load fault of a ship vibration isolation device according to claim 6, characterized in that: The hidden layer repeatedly trains the input data and the output data by means of gradient descent and back propagation.

8. The intelligent diagnosis method for load fault of a ship vibration isolation device according to claim 1, characterized in that: In step S5, the method for optimizing the parameter weights of the updated preset neural network model is to calculate the posterior probability distribution of the preset neural network model parameters based on the Bayesian network statistical model, use the MCMC algorithm to obtain samples that obey this distribution, and estimate the expectation and variance values ​​of the posterior distribution. If the expectation and variance values ​​are within the set range, the neural network model parameter weights are determined to be feasible; otherwise, repeat step S4.

9. The intelligent diagnosis method for load fault of a ship vibration isolation device according to claim 1, characterized in that: The method for updating the training sample set in step S4 is to perform inverse probability calculation based on the fault information under actual working conditions and the preset Bayesian network model.