Multi-support shaft bearing displacement value calculation method based on pre-training-fine-tuning algorithm

Abstract

The application discloses a kind of based on pre-training-fine-tuning algorithm's multiple support shaft bearing displacement value calculation method, including the following steps: constructing ideal shafting model, according to the requirement of shafting design calculation and shafting design drawing to ideal shafting model adds constraint condition and is trained to ideal shafting model by simulation data;Actual shafting neural network model is obtained by fine-tuning to ideal shafting model;Obtain several groups of bearing load measured data using iterative method;Several groups of measured data are combined respectively, expand training data set, and adjust actual shafting neural network model;The current load of intermediate bearing is input into the adjusted actual shafting neural network model with the target load in the shafting design book, to obtain accurate displacement value.The application combines a large number of simulation data and measured data, and considers the problem that it is difficult to obtain height reference, to obtain actual shafting neural network, improve the quality and efficiency of shafting installation.

Description

Technical Field

[0001] This invention belongs to the field of intermediate bearing installation displacement technology, specifically involving a method for calculating the displacement value of multi-support shaft bearings based on a pre-training-fine-tuning algorithm. Background Technology

[0002] As a crucial component of a ship's power plant, the alignment quality of the propulsion shafting directly impacts the ship's propulsion performance. A well-aligned shafting system is essential for safe navigation, while poor alignment can lead to excessive bearing stress, abnormal wear, and even shafting failure, severely impacting the ship's power and navigability. Therefore, shipyards adhere to stringent shafting installation standards to ensure installation quality. In engineering, bearing load measurements are often used to verify shafting installation quality. However, the coupled bearing loads in multi-support shafting systems increase the difficulty of alignment, resulting in long installation and adjustment periods and low alignment efficiency. To address this issue, some scholars have proposed a method to optimize a shafting neural network based on training samples with different confidence levels. This method trains the neural network model using simulated data representing the shafting design model, and then further refines the model using test data representing the actual shafting adjustment process. A method for calculating the displacement value of intermediate bearings based on training samples with different confidence levels is proposed. However, since a height benchmark cannot be found at the dock, the bearing height defined by this method is difficult to obtain in practical engineering.

[0003] The theoretical centerline determined on the slipway deviates due to various factors after the ship is launched, causing elevation deviations in the shafting bearings. Subsequent bearing repositioning adjustments lose their height reference, making it difficult to meet the specified load requirements for each bearing using traditional trial-and-error methods, often requiring numerous attempts. The coupled load characteristics of multi-support shafting bearings increase the difficulty of shafting installation and alignment. Shafting assembly and adjustment personnel often need to perform multiple repeated adjustments, easily leading to difficulties in finding adjustment patterns, losing direction, or even increasing bearing load deviations with each adjustment. This results in long shafting assembly and adjustment periods and low alignment efficiency. To obtain accurate bearing displacement values, the key lies in obtaining an actual shafting model.

[0004] Using neural networks to train and learn from measured data to obtain a neural network model of the actual shafting system, and then using this model to calculate the precise displacement values ​​of each bearing, is an effective solution. However, a large amount of measured data is crucial for obtaining more accurate load displacement relationships between shafting bearings. During actual shipboard installation, the environment and human factors are more complex, and unlike ideal conditions, measured data is often difficult to obtain. Summary of the Invention

[0005] The purpose of this invention is to provide a method for calculating bearing displacement values ​​in multi-support shaft systems based on a pre-trained and fine-tuned algorithm. This method combines a large amount of simulation data and measured data, while also considering the difficulty in obtaining a height reference, to obtain an actual shaft system neural network, thereby improving the quality and efficiency of shaft system installation.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is: a method for calculating the displacement value of bearings in multi-support shaft systems based on a pre-training-fine-tuning algorithm, comprising the following steps:

[0007] The model building module is used to construct an ideal shaft system neural network. It adds constraints to the shaft system simulation model to obtain the ideal shaft system model. Simulation data is calculated using the ideal shaft system model, and the neural network is trained to obtain the ideal shaft system neural network and the ideal shaft system model. The design requirements for the ideal shaft system model are: the forces on each intermediate bearing in the model meet the requirements of the shaft system design calculation book and shaft system design drawings.

[0008] To obtain a real-world axonal neural network, the following steps are taken: First, construct a blank neural network model. Then, copy all network designs and parameters from the ideal axonal neural network except for the connection layers and import them into the neural network model. Next, add an output layer with the same size and structure as the ideal axonal model to the neural network model and randomly initialize its parameters. Finally, use real-world data as training data to obtain new fully connected layer weights. In the neural network model, adjust the parameters of all remaining layers except for the connection layers according to the parameters of the ideal axonal model. The adjusted neural network model is then used as the real-world axonal neural network model.

[0009] Several sets of measured bearing load data were obtained using an iterative method; these sets of measured data were then combined to expand the training dataset, and the actual shaft system neural network model was adjusted.

[0010] The current load of the intermediate bearing and the target load in the shaft system design are input into the adjusted actual shaft system neural network model to obtain the precise relative displacement value, that is, the displacement value required to adjust the current position to the ideal state; the height of each intermediate bearing is adjusted according to the relative displacement value, and the bearing load of each intermediate bearing after the height adjustment is measured. If it does not meet the specification requirements, the fine-tuning training and iterative adjustment of the neural network are repeated; where the relative displacement value is the difference between the load of the intermediate bearing after displacement and the load of the intermediate bearing before displacement in one adjustment of the shaft system.

[0011] In the ideal shaft system neural network, the input of the BP neural network is set as the load before and after the intermediate bearing displacement in one adjustment of the shaft system, and the output is set as the displacement value.

[0012] The method for measuring bearing load simulation data is as follows: the theoretical centerline of the shaft system is taken as the initial state of the shaft system, that is, all intermediate bearings are at the same height, and the bearing height is recorded as zero; the intermediate bearings are displaced step by step according to the bearing displacement value, while the height of other bearings is kept constant, and the load simulation data of each intermediate bearing is calculated.

[0013] The BP neural network selects the logarithmic sigmoid function as the transfer function for the hidden layer nodes and sets the linear purelin function as the transfer function for the output layer nodes.

[0014] A multi-support shaft bearing displacement calculation system based on a pre-training-fine-tuning algorithm is also provided, including a model building module, a fine-tuning module, and a measurement calculation module; wherein,

[0015] The model building module is used to construct an ideal shaft system neural network. It adds constraints to the shaft system simulation model to obtain the ideal shaft system model. Simulation data is calculated using the ideal shaft system model, and the neural network is trained to obtain the ideal shaft system neural network and the ideal shaft system model. The design requirements for the ideal shaft system model are: the forces on each intermediate bearing in the model meet the requirements of the shaft system design calculation book and shaft system design drawings.

[0016] The fine-tuning module is used to fine-tune the ideal axisymmetric model to obtain the actual axisymmetric neural network model. The specific steps of fine-tuning are as follows: construct a blank neural network model, copy all network designs and parameters except for the connection layers in the ideal axisymmetric model and import them into the neural network model; add an output layer with the same size and structure as the output layer of the ideal axisymmetric model to the neural network model, and randomly initialize the model parameters of this layer; use the measured data as the training data for the neural network model, and train the neural network with the measured dataset to obtain new fully connected layer weights; in the neural network model, adjust the parameters of all remaining layers except for the connection layers according to the parameters of the ideal axisymmetric model, and use the adjusted neural network model as the actual axisymmetric neural network model.

[0017] The measurement and calculation module uses an iterative method to obtain several sets of measured bearing load data. These sets of measured data are then combined to expand the training dataset and adjust the actual shaft system neural network model. The current load of the intermediate bearings and the target load in the shaft system design are input into the adjusted actual shaft system neural network model to obtain precise displacement values, i.e., the displacement values ​​required to adjust the current position to the ideal state. The height of each intermediate bearing is adjusted according to the displacement values, and the bearing load of each intermediate bearing after the height adjustment is measured. If the load does not meet the specifications, the fine-tuning training and iterative adjustment of the neural network are repeated. The displacement value is the difference between the load of the intermediate bearing after displacement and the load before displacement in one adjustment of the shaft system.

[0018] In the ideal shaft system model, the input of the BP neural network is set as the load before and after the intermediate bearing displacement in one adjustment of the shaft system, and the output is set as the displacement value.

[0019] The method for measuring bearing load simulation data is as follows: taking the theoretical centerline of the shaft system as the initial state of the shaft system, the intermediate bearings are sequentially displaced according to the bearing displacement value step size, while keeping the height of other bearings constant, and the load simulation data of each intermediate bearing is calculated.

[0020] The BP neural network selects the logarithmic sigmoid function as the transfer function for the hidden layer nodes and sets the linear purelin function as the transfer function for the output layer nodes.

[0021] A computer device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, performs the steps of the method as described in any of the preceding claims.

[0022] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method as described in any of the preceding claims.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] (1) This invention solves the problem that it is difficult to define the bearing height in actual engineering because the dock cannot find a height benchmark. At the same time, due to the special nature of the neural network input and output method, the training dataset can be quickly expanded by permutation and combination under the condition of a small amount of measured data, which reduces the dependence on the amount of measured data.

[0025] (2) The present invention can obtain a relatively accurate intermediate bearing displacement value with only a few iterations, which further improves the installation efficiency and installation quality of ship shafting. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating an embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of the pre-training-fine-tuning training method in an embodiment of the present invention;

[0028] Figure 3 This is an iterative diagram of the prediction error of the axis system neural network in an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0030] like Figure 1 Based on the shaft system design document, an ideal shaft system model is established and simulation data is calculated. The simulation dataset is used to train an ideal-state shaft system neural network to obtain the ideal-state shaft system neural network. A new neural network model, which closely approximates the actual shaft system, is then established, replicating all network designs and parameters of the ideal shaft system neural network model except for the output layer. The output layer, with an output size equal to the number of classes in the measured dataset, is added to the target model, and its parameters are randomly initialized. The new shaft system neural network model is trained using an iterative method to obtain the actual shaft system neural network. The steps for constructing a method for calculating the displacement value of multi-support shaft system bearings based on a pre-training-fine-tuning algorithm are as follows:

[0031] (1) Constructing an ideal axis neural network

[0032] According to the requirements of the shaft system design calculation sheet and shaft system design drawings, add constraints to the shaft system simulation model, evaluate the model accuracy, and ensure that the force on each intermediate bearing meets the design requirements. If so, the simulation model is considered to have met the design requirements and can be used as a reliable model for calculation simulation data.

[0033] Using the theoretical centerline of the shaft system as the initial state, individual intermediate bearings are sequentially displaced according to their displacement values, while keeping the heights of other bearings constant. Load data for each intermediate bearing is then calculated. A data set is created using four loads at a given state before and after displacement, along with the displacement values ​​for both states. The intermediate bearing heights and corresponding load data calculated by the simulation model can be randomly combined to expand the training dataset.

[0034] The input to the BP neural network is set as the pre-displacement and post-displacement loads during a single shaft system adjustment, and the output is set as the relative displacement value. For a propulsion shaft system with n intermediate bearings, vector P represents the initial load of the intermediate bearings, and vector P' represents the post-displacement load of the intermediate bearings. These two vectors together constitute the input of the neural network, which contains 2^n elements. Vector T represents the output value of the intermediate bearing displacement, which contains n elements. The BP algorithm selects the logarithmic sigmoid function as the transfer function for the hidden layer nodes and sets the linear purelin function as the transfer function for the output layer nodes. The ideal shaft system neural network is trained using simulation data, and the accuracy of the trained neural network is verified as a basis for fine-tuning.

[0035] (2) Fine-tuning of neural networks

[0036] The pre-training-fine-tuning method aims to find shared parameter information between the source and target domains to achieve transfer learning. This transfer method requires that the data in the source and target domains share some model parameters. In the shaft system assembly and adjustment problem, due to the correlation between the ideal and actual shaft system models, this method can be used to share some parameters.

[0037] A new neural network model, namely the actual axisymmetric neural network model, is established, replicating all network designs and parameters of the ideal axisymmetric neural network model except for the fully connected layers. These model parameters contain knowledge learned from the simulation dataset, and some of this knowledge will also apply to the experimental dataset. However, because the output layers of the ideal axisymmetric neural network are closely related to the labels in the simulation dataset, they are not used in the actual axisymmetric model.

[0038] Add an output layer whose output size is equal to the number of categories in the actual test dataset to the actual axial neural network model, and randomly initialize the model parameters of this layer.

[0039] The actual measured data was used as the training data for the real-world axisymmetric neural network. The network was trained using the measured dataset to obtain new weights for the fully connected layers. Simultaneously, the parameters of all remaining layers were fine-tuned based on the parameters of the ideal axisymmetric model. The overall structure is as follows: Figure 2 As shown, the actual axisymmetric neural network model is finally obtained.

[0040] (3) Obtaining measured data

[0041] An iterative method is employed. First, the bearing load of each intermediate bearing in its initial state is measured, and this data is recorded as displacement values ​​and corresponding bearing loads. The first set of relative displacement values ​​for the shaft system is predicted using a shaft system neural network, and each intermediate bearing is adjusted accordingly. For example, if the shaft system neural network predicts relative displacement values ​​of +0.1mm, +0.09mm, -0.08mm, and +0.08mm, intermediate bearing #1 is adjusted to +0.1mm, intermediate bearing #2 to +0.09mm, intermediate bearing #3 to -0.08mm, and intermediate bearing #4 to +0.08mm. The resulting second set of data is recorded as displacement values ​​and corresponding bearing loads. These two sets of measured data constitute a training dataset for fine-tuning the neural network. The fine-tuned shaft system neural network guides the next installation of the intermediate bearings, simultaneously obtaining a new set of measured data. The measured data is obtained iteratively, with each installation and adjustment yielding a new set of measured data. Through permutations and combinations, the new measured data is combined with existing measured data to rapidly expand the training dataset and re-fine-tune the neural network.

[0042] (4) Calculate displacement value

[0043] After fine-tuning the shaft system neural network, the current load of the intermediate bearings and the target load in the shaft system design are input into the adjusted neural network to obtain a set of precise displacement values, i.e., the displacement values ​​required to adjust the current position to the ideal state. The height of each intermediate bearing is adjusted according to this set of displacement values, and then the bearing load of each intermediate bearing is measured. If it does not meet the specifications, the iteration and repetition of the neural network fine-tuning training and prediction steps are required.

[0044] The method presented in this paper is verified using the shafting of a large cruise rescue vessel as an example, and the results are as follows: Figure 3 As shown, as the assembly and adjustment process continues, although the error fluctuates slightly, the error of the load of each intermediate bearing shows a gradual decreasing trend, eventually reaching a stable state, and the error is significantly reduced.

[0045] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for calculating bearing displacement values ​​in multi-support shaft systems based on a pre-training-fine-tuning algorithm, characterized in that... Includes the following steps: An ideal shaft system neural network is constructed, and constraints are added to the shaft system simulation model to obtain the ideal shaft system model. The neural network is trained using simulation data to obtain the ideal shaft system neural network and the ideal shaft system model. The design requirements for the ideal shaft system model are: the forces on each intermediate bearing in the model meet the requirements of the shaft system design calculation and shaft system design drawings. To obtain an actual axonal neural network model, the following steps are taken: First, construct a blank neural network model. Copy all network designs and parameters from the ideal axonal neural network except for the output layer and import them into the neural network model. Second, add an output layer with the same size and structure as the output layer of the ideal axonal neural network to the neural network model and randomly initialize the model parameters of this layer. Third, use actual test data as the training data for the neural network model to obtain new fully connected layer weights. Fourth, adjust the parameters of all remaining layers except the connected layers according to the parameters of the ideal axonal model. Finally, use the adjusted neural network model as the actual axonal neural network model. Several sets of measured bearing load data were obtained using an iterative method; these sets of measured data were then combined to expand the training dataset, and the actual shaft system neural network model was adjusted. The current load of the intermediate bearing and the target load in the shaft system design are input into the adjusted actual shaft system neural network model to obtain the precise displacement value, that is, the displacement value required to adjust the current position to the ideal state; the height of each intermediate bearing is adjusted according to the displacement value, and the bearing load of each intermediate bearing after the height adjustment is measured. If it does not meet the specification requirements, the fine-tuning training and iterative adjustment of the neural network are repeated; where the displacement value is the difference between the load of the intermediate bearing after displacement and the load of the intermediate bearing before displacement in one adjustment of the shaft system.

2. The method for calculating bearing displacement values ​​in multi-support shaft systems based on a pre-training-fine-tuning algorithm according to claim 1, characterized in that, In the ideal shaft system model, the input of the BP neural network is set as the load before and after the intermediate bearing displacement in one adjustment of the shaft system, and the output is set as the displacement value.

3. The method for calculating bearing displacement values ​​in multi-support shaft systems based on a pre-training-fine-tuning algorithm according to claim 1, characterized in that, The calculation method for bearing load simulation data is as follows: the theoretical centerline of the shaft system is taken as the initial state of the simulated shaft system, that is, at this time all intermediate bearings are at the same height, and the bearing height is recorded as zero; the intermediate bearings are displaced step by step according to the bearing displacement value, while the height of other bearings is kept constant, and the load simulation data of each intermediate bearing is calculated.

4. The method for calculating bearing displacement values ​​in a multi-support shaft system based on a pre-training-fine-tuning algorithm according to claim 2, characterized in that, The BP neural network selects the logarithmic sigmoid function as the transfer function for the hidden layer nodes and sets the linear purelin function as the transfer function for the output layer nodes.

5. A system using the pre-training-fine-tuning algorithm-based method for calculating bearing displacement values ​​in a multi-support shaft system as described in claim 1, characterized in that, It includes a model building module, a fine-tuning module, and a measurement calculation module; among which, The model building module is used to construct an ideal shaft system neural network. It adds constraints to the shaft system simulation model to obtain the ideal shaft system model. Simulation data is calculated using the ideal shaft system model, and the neural network is trained to obtain the ideal shaft system neural network and the ideal shaft system model. The design requirements for the ideal shaft system model are: the forces on each intermediate bearing in the model meet the requirements of the shaft system design calculation book and shaft system design drawings. The fine-tuning module is used to fine-tune the ideal axial neural network to obtain the actual axial neural network model. The specific steps of fine-tuning are as follows: construct a blank neural network model, copy all network designs and parameters of the ideal axial neural network except for the connection layers, and import them into the neural network model; add an output layer with the same size and structure as the output layer of the ideal axial neural network to the neural network model, and randomly initialize the model parameters of this layer; use the actual test data as the training data for the neural network model, and train the neural network with the actual test dataset to obtain new fully connected layer weights; in the neural network model, adjust the parameters of all remaining layers except for the connection layers according to the parameters of the ideal axial model, and use the adjusted neural network model as the actual axial neural network model. The measurement and calculation module uses an iterative method to obtain several sets of measured bearing load data. These sets of measured data are then combined to expand the measured dataset, and the actual shaft system neural network model is adjusted. The current load of the intermediate bearings and the target load in the shaft system design are input into the adjusted actual shaft system neural network model to obtain precise displacement values, i.e., the displacement values ​​required to adjust the current position to the ideal state. The height of each intermediate bearing is adjusted according to the displacement values, and the bearing load of each intermediate bearing after height adjustment is measured. If the load does not meet the specifications, the fine-tuning training and iterative adjustment of the neural network are repeated. The displacement value represents the change in height of the intermediate bearings in the shaft system before and after displacement during a single adjustment.

6. The system according to claim 5, characterized in that, In the ideal shaft system model, the input of the BP neural network is set as the load before and after the intermediate bearing displacement in one adjustment of the shaft system, and the output is set as the displacement value.

7. The system according to claim 5, characterized in that, The calculation method for bearing load simulation data is as follows: the theoretical center line of the shaft system is taken as the initial state of the shaft system, that is, all intermediate bearings are at the same height, and the bearing height is recorded as zero; the intermediate bearings are displaced step by step according to the bearing displacement value, while the height of other bearings is kept constant, and the load simulation data of each intermediate bearing is calculated.

8. The system according to claim 6, characterized in that, The BP neural network selects the logarithmic sigmoid function as the transfer function for the hidden layer nodes and sets the linear purelin function as the transfer function for the output layer nodes.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method as described in any one of claims 1-4.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-4.

Images