Rotor hoisting load prediction method based on virtual and real data fusion neural network load prediction model

Abstract

A rotor hoisting load prediction method based on a virtual and real data fusion neural network load prediction model comprises the steps that strain-load data under different load working conditions are obtained through finite element simulation, actual strain-load data are obtained through experiments, and training data samples are obtained; and training the convolutional neural network through the strain-load data sample to obtain a load neural network prediction model, taking the strain as the input of the convolutional neural network, outputting the load borne by the structure, and realizing the online prediction of the load. According to the method, actual data and finite element simulation data are fused, and deep feature information is effectively extracted from the structural strain data of the hoisting equipment, so that online prediction of the load is realized, and the safety of the hoisting load is ensured.

Description

Technical Field

[0001] The invention belongs to the technical field of rotor lifting weight prediction, and in particular relates to a rotor lifting weight prediction method based on a virtual-real data fusion neural network load prediction model. Background Art

[0002] The rotor of a large hydropower station is relatively large. In this example, the rotor of the hydropower station has a diameter of 15m, a height of 3m, and a weight of more than 1000t. During the hoisting process of the hydropower station rotor, a large bridge crane is required to lift the rotor with a special hoisting device, and then lift the rotor from the maintenance room to the top of the foundation pit to complete the assembly of the rotor and stator of the hydropower generator.

[0003] The rotor balance beam is located at the interface between the lifting equipment and the lifting object, connecting the overhead crane and the rotor, and plays an important role in the rotor lifting operation. Its main functions include: ensuring that the rotor remains balanced and stable during the lifting process, and avoiding damage to the rotor due to tilting or rotation, which is crucial to ensuring the safety of the lifting operation and the integrity of the equipment. Reasonably distribute or balance the load of each lifting point. In the case of multiple machines lifting, the rotor balance beam reasonably distributes or balances the load of each lifting point to ensure the stability and safety of the lifting operation. Therefore, it is very necessary to use a rotor balance beam to hang the rotor when performing rotor lifting operations. When the rotor balance beam is lifting heavy objects, it is easy to be dangerous because the load exceeds the rotor balance beam. The safety risks involved cannot be ignored. Therefore, it is very necessary to measure the load of the rotor balance beam to ensure the normal lifting.

[0004] Therefore, the technical problem to be solved by the present invention is: how to predict the load size in real time online, so as to issue a warning in time before the load approaches or reaches the bearing limit of the rotor balance beam, so that the operator can take measures in advance according to the prediction results, such as adjusting the lifting method, checking the equipment status, etc., to effectively reduce the safety risks in the lifting process. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a rotor lifting weight prediction method based on a virtual-real data fusion neural network load prediction model, which can predict the load size in real time online and issue an early warning in time before the load approaches or reaches the bearing limit of the rotor balance beam. The operator can take measures in advance according to the prediction results, such as adjusting the lifting method, checking the equipment status, etc., effectively reducing the safety risks in the lifting process.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A rotor lifting weight prediction method based on a virtual-real data fusion neural network load prediction model, the steps are:

[0008] Determine stress measurement points of the rotor balance beam structure, the stress measurement points include the middle position of the rotor balance beam, the left head connector of the rotor balance beam, the right head connector of the rotor balance beam, the front of the left head of the rotor balance beam, the back of the left head of the rotor balance beam, the front of the right head connector of the rotor balance beam, the back of the right head connector of the rotor balance beam, the left side of the middle connector of the rotor balance beam, and the right side of the middle connector of the rotor balance beam; obtain strain data of the structure under different loads in actual conditions through strain sensors arranged at the stress measurement points, and obtain virtual strain load data through finite element simulation;

[0009] The actual strain data is used as the input of the neural network, and the load size under the corresponding working condition is used as the output to obtain the strain-load training data sample;

[0010] The strain-load training data samples are used as model training data samples to construct a convolutional neural network; the training parameters are set, and the convolutional neural network is trained through the strain-load data samples to obtain a neural network load prediction model;

[0011] In the process of determining the service life of the rotor balance beam, strain sensors are arranged at the stress measurement points of the rotor balance beam to obtain strain data at the same position at the same time intervals. The obtained strain data is used as the input of the neural network load prediction model, and the load borne by the structure is output to realize online prediction of the load.

[0012] Preferably, the specific process of determining the stress measurement point of the rotor balance beam structure is:

[0013] Through finite element simulation of the overall structure of the rotor balance beam, it is determined that the strain of the hoisting device under the hoisting conditions with acceleration and inclination is greater than |1.680×10 -4 | parts, and arrange strain sensors at these parts;

[0014] The working condition is set as N; the strain data under each working condition is collected and recorded as x, including the strain value of each point measured by the strain bridge, and the strain vector x = [ε1ε2ε3...ε k ], where ε k Represents the strain value measured at the strain measuring point numbered k.

[0015] Preferably, when using a neural network to obtain strain-load training data samples, in order to facilitate the neural network to extract the characteristics of the parameters themselves and avoid the influence of various strain measurement points and load parameters on the training of the neural network, all variables need to be normalized to the same interval;

[0016] The normalization formula is:

[0017]

[0018] where ε max and ε min are the maximum and minimum strain values ​​of the measuring point in all calibration conditions, respectively, and the normalized strain vector x'=[ε'1ε'2ε'3...ε' k ].

[0019] Preferably, the number of the middle positions of the rotor balance beam is two, and the two middle positions of the rotor balance beam are located on both sides of the rotor;

[0020] The number of connecting parts on the left head of the rotor balance beam is two;

[0021] The number of the right head connecting parts of the rotor balance beam is two;

[0022] One each at the front of the left head of the rotor balance beam, the rear of the left head of the rotor balance beam, the front of the connecting piece of the right head of the rotor balance beam, the rear of the connecting piece of the right head of the rotor balance beam, the left side of the middle connecting piece of the rotor balance beam, and the right side of the middle connecting piece of the rotor balance beam;

[0023] The total number of stress measurement points of the rotor balance beam structure is 12, and the 12 stress measurement points are symmetrically distributed based on the longitudinal and transverse planes where the rotor axis is located.

[0024] Preferably, the strain-load training data samples based on the neural network load prediction model consist of the strain vector x and load parameter y of the measuring point under the same calibration condition.

[0025] Preferably, the neural network-based load prediction model consists of a load prediction module and a data compensation module.

[0026] Preferably, the neural network-based load prediction model is obtained by training the load prediction module with a virtual data set obtained by finite element simulation, and the data compensation module is obtained by training with actual data.

[0027] Preferably, the load prediction module parameters are updated using the gradient descent method, and the loss function during the training process is:

[0028]

[0029] in: represents the sum of the predicted values ​​of the sample numbered i in the load prediction module; N represents the experimental sample load parameter of the finite element simulation virtual data numbered i; L is the number of finite element simulation virtual calibration experimental samples used for neural network training; σ p is a learnable parameter of the load prediction module; β is the regularization coefficient.

[0030] Preferably, the gradient descent method is the Adam optimization algorithm, and the specific steps are as follows:

[0031] (1) Calculate the first-order moment estimate of the gradient (momentum):

[0032]

[0033] (2) Calculate the second-order moment estimate of the gradient (RMSProp):

[0034]

[0035] (3) Correct the deviation of the first-order moment and the second-order moment:

[0036]

[0037] (4) Update parameters:

[0038]

[0039] Among them, β1 and β2 are the decay rates, usually with values ​​of 0.9 and 0.999, and ∈ is a small constant used to prevent division by zero errors.

[0040] A rotor lifting weight prediction system based on a virtual-real data fusion neural network load prediction model adopts the rotor lifting weight prediction method based on a virtual-real data fusion neural network load prediction model.

[0041] The present invention can achieve the following beneficial effects:

[0042] 1. The present invention can realize online prediction of the load borne by the rotor balance beam. During the rotor hoisting process of a hydropower station, the rotor balance beam plays a key role in connecting the overhead crane and the rotor and ensuring the balance and stability of the hoisting. Due to the large size and heavy weight of the rotor (such as the diameter of 15m, the height of 3m, and the weight of over 1000t in the example), once the load exceeds the load borne by the rotor balance beam, it is easy to be dangerous. By accurately predicting the load, safety accidents caused by overloading and other situations can be avoided, and the safety of the hoisting operation and the integrity of the equipment can be guaranteed.

[0043] 2. The present invention predicts the load size in real time online and can issue a warning in time before the load approaches or reaches the bearing limit of the rotor balance beam. The operator can take measures in advance according to the prediction results, such as adjusting the lifting method, checking the equipment status, etc., which effectively reduces the safety risks during the lifting process.

[0044] 3. The present invention combines actual strain data and virtual strain load data obtained by finite element simulation to construct strain-load training data samples. Virtual data can simulate various complex working conditions and make up for the shortcomings of actual data in some special cases, thereby enriching the diversity of training data, enabling the neural network to learn a wider range of characteristics and laws, and improving the generalization ability and prediction accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The present invention will be further described below in conjunction with the accompanying drawings and embodiments:

[0046] Figure 1 Flow chart of the method of the present invention.

[0047] Figure 2 Schematic diagram of the strain sensor layout of the lifting equipment in the example.

[0048] Figure 3 The figure is a comparison between the predicted results and the actual results of the load in the example.

[0049] Figure 4 This is the relative error diagram of the load prediction results. DETAILED DESCRIPTION

[0050] Embodiment 1:

[0051] The preferred solution is Figures 1 to 4 As shown in FIG. 1 , a rotor lifting weight prediction method based on a virtual-real data fusion neural network load prediction model is shown. The specific method is as follows:

[0052] The force analysis of the structure is carried out in finite element software such as ANSYS, and the corresponding load simulation results are obtained by setting the load conditions and working conditions. According to the simulation results, the positions where the structural strain changes greatly are determined and set as strain monitoring points; strain sensors are pasted on the surface of these positions, and the number of pasted sensors is 12. Then the structural loading experiment is started, and N different load conditions are set according to the actual service conditions of the structure. The strain data under each working condition is collected, recorded as x, including the strain values ​​of each point measured by the strain bridge, and the strain vector x = [ε1ε2ε3...ε k ], where ε k Represents the strain value measured at the strain measuring point numbered k.

[0053] In order to facilitate the neural network to extract the characteristics of the parameters themselves and avoid the influence of various strain measurement points and load parameters on the training of the neural network, all variables need to be normalized to the same interval. The normalization formula is:

[0054]

[0055] where ε max and ε min are the maximum and minimum strain values ​​of the measuring point in all calibration conditions, respectively, and the normalized strain vector x'=[ε'1ε'2ε'3...ε' k ];

[0056] Obtain the load vector y. The normalized strain vector of the measuring point and the load parameter vector are used for neural network training.

[0057] The input data is the strain vector of the corresponding point under different working conditions, and the output data is the force data of the structure under different working conditions.

[0058] Each neural network module of the neural network load prediction model is a fully connected neural network, with 3 hidden layers and 80 neuron nodes in each hidden layer. The training adopts the Adam optimization algorithm, and the program is run in the NVIDIA RTX 4060 GPU and PyTorch2.1.1 environment.

[0059] The neural network load prediction model consists of a load prediction module and a data compensation module.

[0060] The load prediction model based on the neural network is obtained by training the load prediction module with the virtual data set obtained by the finite element simulation, and the data compensation module is obtained by training with the actual data.

[0061] The learning rate adjustment method in the neural network training setting is piecewise constant decay, which selects to modify the learning rate by multiplying the learning rate by a factor after each certain number of rounds. This allows a larger learning rate to be selected at the beginning of training, and the value to be gradually reduced during the optimization process, which can shorten the training time. After training, the connection weights and biases between adjacent layers of the network are updated, and the mapping relationship between input data and output data is learned, so that the load data borne by the structure can be obtained after calculating the strain data.

[0062] During training, we first calibrate the experimental samples based on virtual data, taking the strain vector x of a point as input and the load parameter y as the prediction label. We train the prediction module, calculate the loss function as follows, and use the gradient descent method to update the neural network parameters.

[0063] The loss function during training is:

[0064]

[0065] in: represents the sum of the predicted values ​​of the sample numbered i in the load prediction module; N represents the experimental sample load parameter of the finite element simulation virtual data numbered i; L is the number of finite element simulation virtual calibration experimental samples used for neural network training; σ p is a learnable parameter of the load prediction module; β is the regularization coefficient.

[0066] After training is completed, for the strain x measured at a certain test condition in the ground calibration experiment, the load prediction module can give the corresponding low-fidelity prediction value Compare x with A new feature vector is formed as the input of the data compensation module. The actual load parameters corresponding to the ground calibration experimental conditions are used as prediction labels, and the neural network parameters are updated using the gradient descent method.

[0067] The loss function during training is:

[0068]

[0069] in: Represents the current and predicted values ​​of the sample number i in the data compensation module; represents the actual experimental sample load parameter numbered i; N H is the number of actual experimental sample data used for neural network training; σ H is the learnable parameter of the neural network of the data compensation module; β is the regularization coefficient.

[0070] After completing the training of the data compensation module, the neural network load prediction model can predict the strain information of any input measuring point and obtain the corresponding load parameters.

[0071] The specific prediction process is as follows: For a certain working condition, the load prediction module first gives its predicted value Compare x with By forming a new eigenvector input data compensation module, the final load parameter prediction value can be obtained.

[0072] In this embodiment, the load estimation of a scaled-down rotor balance beam structure is taken as an example to illustrate the specific implementation process of the method. The rotor balance beam has a body weight of about 160 t and a size of 17000 mm×5346 mm×3200 mm.

[0073] As shown in the steps, finite element simulation analysis is first performed, different load loading methods are set, and the strain gauge layout positions are selected. There are a total of 12 strain sensors, such as Figure 4 In order to simulate the load effect of the rotor balance beam during actual hoisting, the experiment collected strain data based on strain sensors, and set up 200 groups of working conditions. The strain data under different working conditions were collected, and the load size of the rotor balance beam was obtained by calculation.

[0074] The data from twelve measuring points were used for estimation. 1 / 2 of the samples in the experimental data set were selected for neural network training, 1 / 6 of the samples were used for neural network verification, and 1 / 3 of the samples were used to test the training effect of the neural network. The maximum training rounds were set to 60, and the learning rate adjustment method was piecewise constant decay. The initial learning rate during training was set to 0.02. After every 10 rounds of training, the learning rate was reduced to 0.2 times the previous one. The batch size was 10, and the regularization coefficient was 0.0001. Finally, the training converged to obtain a multilayer perceptron neural network load prediction model.

[0075] A loading experiment was carried out on it. The strain data collected by the strain sensor was normalized and input into the trained neural network load prediction model. The results were compared with the actual results. The comparison chart is as follows Figure 3 , the relative error is as follows Figure 4 As shown, it can be seen that the difference between the load prediction result and the true value is very small, basically within 10%, which shows that the proposed method can accurately evaluate the load condition of the target result.

[0076] The above embodiments are only preferred technical solutions of the present invention and should not be regarded as limiting the present invention. The protection scope of the present invention shall be the technical solutions recorded in the claims, including equivalent replacement solutions of the technical features in the technical solutions recorded in the claims. That is, equivalent replacement improvements within this scope are also within the protection scope of the present invention.

Claims

1. A rotor lifting weight prediction method based on a virtual-real data fusion neural network load prediction model, characterized in that The following steps are involved: Determine the stress measurement points of the rotor balance beam structure, the stress measurement points include the middle position of the rotor balance beam, the left head connecting piece of the rotor balance beam, the right head connecting piece of the rotor balance beam, the front of the left head of the rotor balance beam, the rear of the left head of the rotor balance beam, the front of the right head connecting piece of the rotor balance beam, the rear of the right head connecting piece of the rotor balance beam, the left side of the middle connecting piece of the rotor balance beam, and the right side of the middle connecting piece of the rotor balance beam; The strain data of the structure under different loads in actual conditions are obtained by strain sensors arranged at stress measurement points, and virtual strain load data are obtained by finite element simulation; The actual strain data is used as the input of the neural network, and the load size under the corresponding working condition is used as the output to obtain the strain-load training data sample; The strain-load training data samples are used as model training data samples to construct a convolutional neural network; Set training parameters, train the convolutional neural network through strain-load data samples, and obtain a neural network load prediction model; In the process of determining the service life of the rotor balance beam, strain sensors are arranged at the stress measurement points of the rotor balance beam to obtain strain data at the same position at the same time intervals. The obtained strain data is used as the input of the neural network load prediction model, and the load borne by the structure is output to realize online prediction of the load.

2. The rotor lifting weight prediction method based on the virtual-real data fusion neural network load prediction model according to claim 1 is characterized in that: The specific process of determining the stress measurement points of the rotor balance beam structure is as follows: Through finite element simulation of the overall structure of the rotor balance beam, it is determined that the strain of the hoisting device under the hoisting conditions with acceleration and inclination is greater than |1.680×10 -4 | parts, and arrange strain sensors at these parts; The working condition is set as N; the strain data under each working condition is collected and recorded as x, including the strain value of each point measured by the strain bridge, and the strain vector x = [ε1ε2ε3...ε k ], where ε k Represents the strain value measured at the strain measuring point numbered k.

3. The rotor lifting weight prediction method based on the virtual-real data fusion neural network load prediction model according to claim 1 is characterized in that: When using a neural network to obtain strain-load training data samples, in order to facilitate the neural network to extract the characteristics of the parameters themselves and avoid the influence of various strain measurement points and load parameters on the training of the neural network, all variables need to be normalized to the same interval; The normalization formula is: where ε max and ε min are the maximum and minimum strain values ​​of the measuring point in all calibration conditions, respectively, and the normalized strain vector x of the point is obtained. ‘ =[ε ’ 1ε ’ 2ε ’ 3...ε ’ k ].

4. The rotor lifting weight prediction method based on the virtual-real data fusion neural network load prediction model according to claim 3 is characterized by: There are two rotor balance beam middle positions, and the two rotor balance beam middle positions are located on both sides of the rotor; The number of connecting parts on the left head of the rotor balance beam is two; The number of the right head connecting parts of the rotor balance beam is two; One each at the front of the left head of the rotor balance beam, the rear of the left head of the rotor balance beam, the front of the connecting piece of the right head of the rotor balance beam, the rear of the connecting piece of the right head of the rotor balance beam, the left side of the middle connecting piece of the rotor balance beam, and the right side of the middle connecting piece of the rotor balance beam; The total number of stress measurement points of the rotor balance beam structure is 12, and the 12 stress measurement points are symmetrically distributed based on the longitudinal and transverse planes where the rotor axis is located.

5. The rotor lifting weight prediction method based on the virtual-real data fusion neural network load prediction model according to claim 4 is characterized in that: The strain-load training data samples based on the neural network load prediction model consist of the strain vector x and load parameter y of the measuring point under the same calibration condition.

6. The rotor lifting weight prediction method based on the virtual-real data fusion neural network load prediction model according to claim 1 is characterized by: The load prediction model based on neural network consists of a load prediction module and a data compensation module.

7. The rotor lifting weight prediction method based on the virtual-real data fusion neural network load prediction model according to claim 1 is characterized by: The load prediction model based on the neural network is obtained by training the load prediction module with the virtual data set obtained by the finite element simulation, and the data compensation module is obtained by training with the actual data.

8. The rotor lifting weight prediction method based on the virtual-real data fusion neural network load prediction model according to claim 7 is characterized in that: The gradient descent method is used to update the parameters of the load prediction module. The loss function during the training process is: in: represents the sum of the predicted values ​​of the sample numbered i in the load prediction module; N represents the experimental sample load parameter of the finite element simulation virtual data numbered i; L is the number of finite element simulation virtual calibration experimental samples used for neural network training; σ p is a learnable parameter of the load prediction module; β is the regularization coefficient.

9. The rotor lifting weight prediction method based on the virtual-real data fusion neural network load prediction model according to claim 8 is characterized by: The gradient descent method is the Adam optimization algorithm. The specific steps are as follows: (1) Calculate the first-order moment estimate of the gradient: m t =β1·m t-1 +(1-β1)·▽ θ J(θ); (2) Calculate the second-order moment estimate of the gradient: m t =β2·v t-1 +(1-β2)·(▽ θ J(θ)) 2 ; (3) Correct the deviation of the first-order moment and the second-order moment: (4) Update parameters: Where β1 and β2 are the decay rates and ∈ is a constant used to prevent division by zero errors.

10. A rotor lifting weight prediction system based on a virtual-real data fusion neural network load prediction model, characterized in that: A rotor lifting weight prediction method based on a virtual-real data fusion neural network load prediction model according to any one of claims 1 to 9 is adopted.

Images