Hanging load identification method and system, medium and equipment

Abstract

The invention discloses a hanging load identification method and system based on transfer learning and a cyclic convolutional network, a medium and equipment, and the method comprises the steps: constructing a hanging structure finite element model, simulating the vibration response of the hanging structure finite element model under the excitation of different impact loads to obtain a simulation vibration response signal, and constructing a response-load simulation data set; obtaining a vibration response signal of the hanging structure under concentrated impact load excitation, obtaining a corresponding concentrated load time history, and constructing a response-load experiment data set; combining the convolution layer and the bidirectional LSTM layer to obtain convolution circulation blocks, constructing a deep convolution circulation neural network model by stacking a plurality of convolution circulation blocks, and pre-training the deep convolution circulation neural network model on the response-load simulation data set to obtain a pre-training model; carrying out parameter fine tuning on the pre-training model on the response-load experiment data set by adopting transfer learning to obtain a hanging structure load identification network model; and carrying out reverse identification on the concentrated impact load of the hanging structure by using the hanging structure load identification network model.

Description

Technical Field

[0001] The present invention relates to the technical field of structural health monitoring, and particularly to a method, system, medium and device for identifying hanging loads based on transfer learning and recurrent convolutional network. Background Art

[0002] As a key connecting part in modern engineering equipment and machinery, the safety and reliability of the hanging structure are of crucial importance. The hanging connection often bears concentrated loads, and accurately identifying these loads is crucial for evaluating structural health, optimizing design, and ensuring safety. It is often extremely difficult to directly measure the hanging load with a force sensor. In recent years, due to its powerful feature extraction and end-to-end processing capabilities, deep neural networks have been widely used in the engineering field. However, in practical applications, it is costly to obtain sufficient training data, and there is often a lack of sufficient data for training, which limits the performance of neural networks. Transfer learning can use the source domain data for pre-training to obtain the feature extraction ability, and then use fewer samples in the target domain for fine-tuning, which can effectively improve the performance of the model in the target domain. A large number of sample data can be obtained at a relatively low cost through the finite element model. However, there are inevitably certain differences in the feature distributions between the simulation data and the experimental data. Therefore, how to accurately identify the load under the conditions of a small number of experimental samples and a large number of simulation samples is a challenging problem.

[0003] The above information disclosed in the background art section is only used to enhance the understanding of the background of the present invention, and thus may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0004] The present invention provides a method, system, medium and device for identifying hanging loads based on transfer learning and recurrent convolutional network, which can accurately identify the concentrated impact load of the hanging structure under the conditions of a small number of experimental samples and difficulty in training the neural network.

[0005] To achieve the above object, the present disclosure provides the following technical solutions:

[0006] A method for identifying hanging loads based on transfer learning and recurrent convolutional network includes:

[0007] Step S1, constructing a finite element model of the hanging structure, simulating its vibration response under different impact load excitations to obtain a simulation vibration response signal, and constructing a response-load simulation data set;

[0008] Step S2, obtaining the vibration response signal of the hanging structure under the excitation of the concentrated impact load, and obtaining the corresponding concentrated load time history, and constructing a response-load experimental data set;

[0009] Step S3: Combine the convolutional layer and the bidirectional LSTM layer to obtain a convolutional recurrent block, and construct a deep convolutional recurrent neural network model by stacking multiple convolutional recurrent blocks. It is pre-trained on the response-load simulation dataset. After pre-training, the learnable parameters of the network are preliminarily adjusted to obtain a pre-trained model.

[0010] Step S4: Use transfer learning to fine-tune the parameters of the pre-trained model on the response-load experimental dataset to obtain a hanging structure load identification network model.

[0011] Step S5: Use the hanging structure load identification network model to inversely identify the concentrated impact load of the hanging structure.

[0012] In the described hanging load identification method based on transfer learning and recurrent convolutional network, Step S1 includes:

[0013] Step S1.1: Establish a three-dimensional geometric model according to the geometric dimensions of the hanging structure.

[0014] Step S1.2: Obtain the corresponding material properties according to the materials used in the hanging structure.

[0015] Step S1.3: Establish a finite element model of the hanging structure according to the three-dimensional geometric model and the material properties.

[0016] Step S1.4: According to the finite element model of the hanging structure, apply a concentrated impact load and the same constraint conditions as the monitored hanging structure, and calculate the simulated vibration response signal of the response point.

[0017] Step S1.5: Use the simulated vibration response signal as the network input feature, and use the corresponding concentrated impact load time history as the corresponding sample label to construct a response-load simulation dataset.

[0018] In the described hanging load identification method based on transfer learning and recurrent convolutional network, in Step S2, the vibration response signal collected by a strain gauge or an acceleration sensor is used as the network input feature, and the concentrated impact load time history obtained by a force sensor is used as the corresponding sample label to establish a response-load experimental dataset containing experimental vibration response signals and impact load signals.

[0019] In the described hanging load identification method based on transfer learning and recurrent convolutional network, Step S3 includes:

[0020] Step S3.1. Construct a convolutional recurrent network feature extractor by stacking multiple convolutional recurrent blocks. A convolutional recurrent block is composed of a convolutional layer, a normalization layer, an activation layer, a pooling layer, a bidirectional LSTM layer, and a normalization layer connected in sequence. Connect multiple convolutional recurrent blocks in sequence to form a convolutional recurrent network feature extractor. The operations of the convolutional recurrent block are as follows: are as follows:

[0021] , where x is the input feature of the convolutional recurrent block, represents a one-dimensional convolutional operation, represents one-dimensional batch normalization, is an activation function, represents a max pooling operation, represents a bidirectional LSTM operation. The convolutional layer and pooling layer in the convolutional recurrent block expand the receptive field of the model. Batch normalization is used to accelerate convergence and ensure stable training, endows the model with nonlinearity, and the LSTM captures the relationships in the time series;

[0022] Step S3.2. Construct a regressor, and use a fully connected layer to map the features extracted by the feature extractor to the corresponding load time series;

[0023] Step S3.3. Divide the response-load simulation data set into a training set and a validation set, and pre-train the deep convolutional recurrent neural network model. The mean squared error (MSE) is used as the loss function during the training process:

[0024] ,

[0025] where is the number of training samples, represents the sample length, represents the predicted value of the i-th sample, represents the true value of the i-th sample.

[0026] In the described method for identifying hanging load based on transfer learning and recurrent convolutional network, in step S4, freeze the learnable parameters of the feature extractor of the pre-trained model, use the response-load experimental data set to perform transfer learning on the learnable parameters of the predictor, and fine-tune its parameter values. After the transfer learning is completed, a hanging structure load identification network model is obtained. Fine-tuning means that when performing transfer learning on the learnable parameters of the predictor using the response-load experimental data set, use fewer training epochs and a smaller learning rate than in the pre-training process.

[0027] In the described method for identifying suspended load based on transfer learning and recurrent convolutional network, in step S5, the vibration response of the monitored suspended structure under impact load is collected and input into the suspended structure load identification network model to obtain the reverse identification result of the concentrated load.

[0028] In the described method for identifying suspended load based on transfer learning and recurrent convolutional network, the suspended structure includes a suspended object, a suspended inner lining plate, a hanging point, and a connecting member.

[0029] A system for identifying suspended load based on transfer learning and recurrent convolutional network includes

[0030] A finite element modeling unit, which is used to construct a finite element model of the suspended structure, simulate its vibration response under different impact load excitations to obtain a simulated vibration response signal, and construct a response-load simulation data set;

[0031] An acquisition unit, which is used to obtain the vibration response signal of the suspended structure under concentrated impact load excitation, and obtain the corresponding concentrated load time history, and construct a response-load experimental data set;

[0032] A construction unit, which is used to combine a convolutional layer and a bidirectional LSTM layer to obtain a convolutional recurrent block, and construct a deep convolutional recurrent neural network model by stacking multiple convolutional recurrent blocks. It is pre-trained on the response-load simulation data set. After pre-training, the learnable parameters of the network are initially adjusted to obtain a pre-trained model;

[0033] A fine-tuning unit, which is used to perform parameter fine-tuning on the pre-trained model on the response-load experimental data set by using transfer learning to obtain a suspended structure load identification network model;

[0034] An identification unit, which is used to perform reverse identification of the concentrated impact load of the suspended structure by using the suspended structure load identification network model.

[0035] A computer storage medium, the storage medium includes computer instructions, when it runs on a computer, it causes the computer to execute the described method.

[0036] An electronic device, the electronic device includes:

[0037] A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein

[0038] When the processor executes the program, it implements the described method.

[0039] Compared with the prior art, the present invention has the following advantages: Compared with traditional deep learning methods, the present invention can make full use of finite element simulation data, reduce the required number of experimental samples, and lower the cost of obtaining data. By using transfer learning technology, the inverse identification accuracy of the concentrated impact load of the hanging structure can be improved in the case of few experimental data and difficult neural network training. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] By reading the following detailed description of the preferred specific embodiments, various other advantages and benefits of the present invention will become clear to those of ordinary skill in the art. The accompanying drawings in the specification are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the present invention. Obviously, the following described drawings are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts. Moreover, throughout the drawings, the same reference numerals are used to represent the same components.

[0041] In the drawings:

[0042] Figure 1 is the flowchart of the method of the present invention;

[0043] Figure 2 is the overall structural schematic diagram of the deep convolutional recurrent neural network model;

[0044] Figure 3 is a schematic diagram of identifying the concentrated impact load on the simulated hanging structure in an embodiment;

[0045] FIG. 4(a) and FIG. 4(b) are schematic diagrams of the load identification results of the deep convolutional recurrent neural network model for the same test sample before and after transfer learning; wherein FIG. 4(a) is the load identification result of the pre-trained model, and FIG. 4(b) is the load identification result of the load identification network of the hanging structure after transfer learning.

[0046] The present invention will be further explained below in conjunction with the drawings and embodiments. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] The specific embodiments of the present invention will be described in more detail below with reference to the drawings. Although the specific embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present invention and to convey the full scope of the present invention to those skilled in the art.

[0048] It should be noted that in the specification and claims, certain terms are used to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. The specification and claims do not distinguish components by the difference in nouns, but by the difference in the functions of components. For example, the terms "comprising" or "including" mentioned throughout the specification and claims are open-ended terms, so they should be interpreted as "including but not limited to". The subsequent description in the specification is the preferred implementation manner for implementing the present invention, but the description is for the purpose of the general principles of the specification and is not used to limit the scope of the present invention. The protection scope of the present invention shall be subject to what is defined by the appended claims.

[0049] For the convenience of understanding the embodiments of the present invention, the following will further explain with specific embodiments in conjunction with the accompanying drawings, and the accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0050] As Figures 1 to 4(b) shown, the hanging load identification method based on transfer learning and recurrent convolutional network includes the following steps:

[0051] Step S1, construct a finite element model of the hanging structure, simulate its vibration response under different impact load excitations to obtain a simulated vibration response signal, and construct a response-load simulation data set;

[0052] Step S2, obtain the vibration response signal of the hanging structure under the excitation of a concentrated impact load, and obtain the corresponding concentrated load time history, and construct a response-load experimental data set;

[0053] Step S3, combine a convolutional layer and a bidirectional LSTM layer to obtain a convolutional recurrent block, and construct a deep convolutional recurrent neural network model by stacking multiple convolutional recurrent blocks. It is pre-trained on the response-load simulation data set. After pre-training, the learnable parameters of the network are initially adjusted to obtain a pre-trained model;

[0054] Step S4, use transfer learning to fine-tune the parameters of the pre-trained model on the response-load experimental data set to obtain a hanging structure load identification network model;

[0055] Step S5, use the hanging structure load identification network model to inversely identify the concentrated impact load of the hanging structure.

[0056] In the preferred implementation manner of the above-mentioned hanging load identification method based on transfer learning and recurrent convolutional network, step S1 includes,

[0057] Step S1.1, establish a three-dimensional geometric model according to the geometric dimensions of the hanging structure;

[0058] Step S1.2: Obtain the corresponding material properties according to the materials used in the suspension structure;

[0059] Step S1.3: Establish a finite element model of the suspension structure according to the three-dimensional geometric model and the material properties;

[0060] Step S1.4: Apply a concentrated impact load and the same constraint conditions as the monitored suspension structure according to the finite element model of the suspension structure, and calculate the simulated vibration response signal of the response point;

[0061] Step S1.5: Use the simulated vibration response signal as the network input feature, and use the corresponding concentrated impact load time history as the corresponding sample label to construct a response-load simulation data set.

[0062] In the preferred implementation manner of the suspension load identification method based on transfer learning and recurrent convolutional network, in step S2, the vibration response signal collected by the strain gauge or acceleration sensor is used as the network input feature, and the concentrated impact load time history obtained by the force sensor is used as the corresponding sample label to establish a response-load experimental data set including the experimental vibration response signal and the impact load signal.

[0063] In the preferred implementation manner of the suspension load identification method based on transfer learning and recurrent convolutional network, step S3 includes:

[0064] Step S3.1: Construct a recurrent convolutional network feature extractor by stacking multiple convolutional recurrent blocks. The convolutional recurrent block is a network structure mainly composed of a convolutional layer and a recurrent layer. Specifically, it is composed of a convolutional layer, a normalization layer, an activation layer, a pooling layer, a bidirectional LSTM layer (i.e., the recurrent layer) and a normalization layer connected in sequence. Connect multiple convolutional recurrent blocks in sequence to form a recurrent convolutional network feature extractor. The operation of the convolutional recurrent block is expressed as follows:

[0065] , where x is the input feature of the convolutional recurrent block, represents a one-dimensional convolutional operation, represents one-dimensional batch normalization, is an activation function, represents a max pooling operation, represents a bidirectional LSTM operation. The convolutional layer and the pooling layer in the convolutional recurrent block expand the receptive field of the model. Batch normalization is used to accelerate convergence and ensure training stability, endows the model with nonlinearity, and LSTM captures the connections in the time series;

[0066] Step S3.2: Construct a regressor, and use a fully connected layer to map the features extracted by the feature extractor to the corresponding load time series;

[0067] Step S3.3: Divide the response-load simulation data set into a training set and a validation set, and pre-train the deep convolutional recurrent neural network model. The mean squared error (MSE) is used as the loss function during the training process:

[0068] ,

[0069] where is the number of training samples, represents the sample length, represents the predicted value of the i-th sample, represents the true value of the i-th sample.

[0070] In the preferred embodiment of the hanging load identification method based on transfer learning and recurrent convolutional network, in step S4, freeze the learnable parameters of the feature extractor of the pre-trained model, use the response-load experimental data set to perform transfer learning on the learnable parameters of the predictor, and fine-tune its parameter values. After the transfer learning is completed, a hanging structure load identification network model is obtained. Fine-tuning means that when using the response-load experimental data set to perform transfer learning on the learnable parameters of the predictor, a smaller number of training epochs and a smaller learning rate are adopted compared with the pre-training process.

[0071] In the preferred embodiment of the hanging load identification method based on transfer learning and recurrent convolutional network, in step S5, collect the vibration response of the monitored hanging structure under the action of impact load, input it into the hanging structure load identification network model, and obtain the reverse identification result of the concentrated load.

[0072] In the preferred embodiment of the hanging load identification method based on transfer learning and recurrent convolutional network, the hanging structure includes a hanging object, a hanging inner lining plate, a hanging point, and a connecting piece.

[0073] A hanging load identification system based on transfer learning and recurrent convolutional network includes

[0074] A finite element modeling unit, which is used to construct a finite element model of the hanging structure, simulate its vibration response under different impact load excitations to obtain a simulated vibration response signal, and construct a response-load simulation data set;

[0075] A collection unit, which is used to obtain the vibration response signal of the hanging structure under the excitation of a concentrated impact load, and obtain the corresponding concentrated load time history, and construct a response-load experimental data set;

[0076] A building unit, which is used to combine a convolutional layer and a bidirectional LSTM layer to obtain a convolutional recurrent block, and build a deep convolutional recurrent neural network model by stacking multiple convolutional recurrent blocks. It is pre-trained on a response-load simulation dataset. After pre-training, the learnable parameters of the network are preliminarily adjusted to obtain a pre-trained model;

[0077] A fine-tuning unit, which is used to fine-tune the parameters of the pre-trained model on the response-load experimental dataset by using transfer learning to obtain a suspension structure load identification network model;

[0078] An identification unit, which is used to inversely identify the concentrated impact load of the suspension structure by using the suspension structure load identification network model.

[0079] A computer storage medium, which includes computer instructions. When it runs on a computer, it enables the computer to execute the method described above.

[0080] An electronic device, which includes:

[0081] A memory, a processor, and a computer program stored on the memory and executable on the processor. Among them,

[0082] When the processor executes the program, it implements the method described above.

[0083] In one embodiment, referring to Figure 1 , the present invention provides a suspension load identification method based on transfer learning and recurrent convolutional network. The method includes the following steps:

[0084] S1. Build a high-precision finite element model of the suspension structure, simulate its vibration response under different impact load excitations, obtain a simulation vibration response signal, and build a response-load simulation dataset;

[0085] S2. Use strain gauges, accelerometers, etc. to obtain the vibration response signal of the suspension structure under the concentrated impact load excitation, use a force sensor to obtain the corresponding concentrated load time history, and build a response-load experimental dataset;

[0086] S3. Combine the convolutional layer and the bidirectional LSTM layer to obtain a convolutional recurrent block, and build a deep convolutional recurrent neural network model by stacking multiple convolutional recurrent blocks. Pre-train on the simulation dataset. After pre-training, the learnable parameters of the network are preliminarily adjusted to obtain a pre-trained model;

[0087] S4. Use transfer learning technology to fine-tune the parameters of the pre-trained model on the experimental dataset to obtain a suspension structure load identification network model;

[0088] S5. Use the load identification network model of the suspension structure with fine-tuned parameters to inversely identify the concentrated impact load of the suspension structure.

[0089] Optionally, in step S1, construct a high-precision finite element model of the suspension structure, simulate its vibration response under different impact load excitations, obtain the simulated vibration response signal, and construct a response-load simulation data set, specifically including:

[0090] S1.1. According to the geometric dimensions of each component of the suspension structure (including the suspended object, inner lining board, hanging point, and connecting parts, etc.), construct an accurate three-dimensional geometric model;

[0091] S1.2. Obtain the material properties of each component of the suspension structure according to the material list;

[0092] S1.3. Establish a finite element model of the suspension structure according to the three-dimensional geometric model in S1.1 and the material properties in S1.2;

[0093] S1.4. Add corresponding constraint conditions to the model established in S1.3, apply a concentrated impact load, and calculate the simulated vibration response signal at the response position;

[0094] S1.5. Use the simulated vibration signal obtained in S1.4 as the network input feature, and use the corresponding concentrated impact load time history as the corresponding sample label to construct a simulation data set.

[0095] Optionally, in step S2, use the vibration response signal of the suspension structure collected by strain gauges, accelerometers, etc. as the network input feature, and use the concentrated impact load time history obtained by the force sensor as the corresponding sample label to establish an experimental data set containing experimental vibration response signals and impact load signals.

[0096] Optionally, in step S3, combine the convolutional layer and the bidirectional LSTM layer to obtain a convolutional recurrent block, and construct a deep convolutional recurrent neural network model by stacking multiple convolutional recurrent blocks, as Figure 2 shown. Perform pre-training on the simulation data set to obtain a pre-trained model, specifically including:

[0097] S3.1. Construct a convolutional recurrent network feature extractor by stacking multiple convolutional recurrent blocks. The convolutional recurrent block is a network structure mainly composed of a convolutional layer and a recurrent layer, specifically composed of a convolutional layer, a normalization layer, an activation layer, a pooling layer, a bidirectional LSTM layer (i.e., the recurrent layer), and a normalization layer connected in sequence. Connect multiple convolutional recurrent blocks in sequence to form a convolutional recurrent network feature extractor. The operation of the convolutional recurrent block can be expressed as follows:

[0098]

[0099] where \(x\) is the input feature of the convolutional recurrent block, represents a one-dimensional convolutional operation, represents one-dimensional batch normalization, is a commonly used activation function, represents a max pooling operation, represents a bidirectional LSTM operation. The convolutional layer and pooling layer in the convolutional recurrent block can expand the receptive field of the model, and batch normalization is used to accelerate convergence and ensure stable training, endows the model with a certain degree of non-linearity, and LSTM can capture the relationships in time series.

[0100] S3.2. Construct a regressor, and use a fully connected layer to map the features extracted by the feature extractor to the corresponding load time series.

[0101] S3.3. According to the actual training effect, select a suitable ratio to divide the simulation data set into a training set and a validation set, and pre-train the deep convolutional recurrent neural network model. The mean squared error (MSE) is used as the loss function during the training process:

[0102]

[0103] where is the number of training samples, represents the sample length, represents the predicted value of the \(i\)-th sample, represents the true value of the \(i\)-th sample.

[0104] Optionally, in step S4, for the pre-trained model in step S3, the transfer learning method is used for further training. First, freeze the learnable parameters of its feature extractor, and use the experimental data set to fine-tune the learnable parameters of the predictor. After the transfer learning, the hanging structure load identification network model is obtained. Fine-tuning means that when using the response-load experimental data set to perform transfer learning on the learnable parameters of the predictor, fewer training epochs and a smaller learning rate are used than in the pre-training process.

[0105] Optionally, in step S5, the vibration response of the hanging structure collected under the impact load is input into the hanging structure load identification network to obtain the inverse identification result of the concentrated load.

[0106] In one embodiment, referring to Figure 3 , using a simulated hanging structure as the research object, the process of identifying the concentrated impact load is as follows:

[0107] 1. The simulated hanging structure mainly includes two cylindrical parts, simulated hanging points, hanging point liners, and the test bench bottom plate, etc. The two cylinders are connected by bolts. The inner diameter of the cylinder is 500 mm and the wall thickness is 3 mm; among them, the height of the upper cylinder is 500 mm, the height of the lower cylinder is 250 mm, and the height of the entire cylindrical structure is 750 mm. The simulated hanging point and the hanging point liner are connected and installed on the upper cylinder by bolts. Two reinforcing rings are symmetrically arranged inside the upper cylinder, with a height of 20 mm and a thickness of 5 mm.

[0108] 2. Build a high-precision finite element model of the hanging structure, simulate its vibration response under different impact load excitations, obtain the simulated vibration response signal, and construct a response-load simulation data set, specifically including:

[0109] According to the geometric dimensions of each component of the hanging structure (including the hanging object, inner liner, hanging point, and connecting parts, etc.), construct an accurate three-dimensional geometric model;

[0110] According to the material list, obtain the material properties of each component of the hanging structure;

[0111] According to the three-dimensional geometric model and material properties, establish a finite element model of the hanging structure;

[0112] According to the established finite element model of the hanging structure, add corresponding constraint conditions, apply a concentrated impact load, and calculate the simulated vibration response signal at the response position;

[0113] Take the simulated vibration signal calculated by the finite element model of the hanging structure as the network input feature, and take the corresponding concentrated impact load time history as the corresponding sample label to construct a simulation data set, which contains 101 simulation samples in total.

[0114] 3. At the simulated hanging point, apply a concentrated impact load with a force hammer. Take the vibration response signal of the hanging structure collected by the strain gauge as the network input feature, and take the concentrated impact load time history obtained by the force sensor in the force hammer as the corresponding sample label to establish an experimental data set containing experimental vibration response signals and impact load signals, which contains 60 experimental samples in total.

[0115] 4. Build a deep convolutional recurrent neural network model, specifically including:

[0116] Build a convolutional recurrent network feature extractor by stacking multiple convolutional recurrent blocks. The convolutional recurrent block is a network structure mainly composed of a convolutional layer and a recurrent layer, specifically composed of a convolutional layer, a normalization layer, an activation layer, a pooling layer, a bidirectional LSTM layer (i.e., the recurrent layer), and a normalization layer connected in sequence. Connect multiple convolutional recurrent blocks in sequence to form a convolutional recurrent network feature extractor. The operation of the convolutional recurrent block can be expressed as follows:

[0117]

[0118] Where x is the input feature of the convolutional recurrent block, represents a one-dimensional convolutional operation, represents one-dimensional batch normalization, is a commonly used activation function, represents a max pooling operation, represents a bidirectional LSTM operation. The convolutional layer and pooling layer in the convolutional recurrent block can expand the receptive field of the model, and batch normalization is used to accelerate convergence and ensure stable training, endows the model with a certain degree of non-linearity, and LSTM can capture the relationships in the time series.

[0119] Construct a regressor, and use a fully connected layer to map the features extracted by the feature extractor to the corresponding load time series.

[0120] According to the actual training effect, select a suitable ratio to divide the simulation data set into a training set and a validation set, and pre-train the deep convolutional recurrent neural network model. The mean squared error MSE is used as the loss function during the training process:

[0121]

[0122] Where is the number of training samples, represents the sample length, represents the predicted value of the i-th sample, represents the true value of the i-th sample.

[0123] 5. Further train the pre-trained model using transfer learning. First, freeze the learnable parameters of its feature extractor, and fine-tune the learnable parameters of the predictor using the experimental data set. After transfer learning, the hanging structure load identification network model is obtained. Fine-tuning means that when using the response-load experimental data set to perform transfer learning on the learnable parameters of the predictor, a smaller number of training epochs and a smaller learning rate are used compared to the pre-training process.

[0124] 6. Input the vibration response of the hanging structure under the impact load collected into the hanging structure load identification network to obtain the reverse identification result of the concentrated load.

[0125] To quantitatively evaluate the performance of the hanging load identification method based on transfer learning and recurrent convolutional network in identifying concentrated impact loads, the following relative error and peak relative error are defined:

[0126]

[0127]

[0128] The fine-tuned hanging structure load identification network was tested using 30 experimental test samples. The average RE of the pre-trained model on the 30 test samples was 27.12%, and the average PRE was 27.10%. The average RE of the deep convolutional recurrent neural network model after transfer learning on the test set was 10.72%, and the average PRE was 7.68%. Compared with the pre-trained model, the average RE was reduced to 60.47% of the original, and the average PRE was reduced to 71.66% of the original. Figures 4(a) to 4(b) The load identification results of the deep convolutional recurrent neural network model for the same test sample before and after transfer learning are shown. Among them, Fig. 4(a) is the load identification result of the pre-trained model, and Fig. 4(b) is the load identification result of the hanging structure load identification network after transfer learning. The load identification accuracy after transfer learning has been significantly improved. The present invention is applied to the centralized load identification of the hanging structure. By training the deep neural network in stages with simulation data and a small amount of experimental data, and using the transfer learning technology to improve the load reverse identification accuracy, the problems of high cost and small quantity of experimental sample acquisition are effectively solved.

[0129] Although the embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the above specific embodiments and application fields. The above specific embodiments are merely illustrative and guiding, rather than restrictive. Those of ordinary skill in the art can also make many forms under the inspiration of this specification and without departing from the scope protected by the claims of the present invention, and these all belong to the scope of protection of the present invention.

Claims

1. A method for identifying suspended loads based on transfer learning and recurrent convolutional network, characterized in that, It includes the following steps: Step S1: Construct a finite element model of the hanging structure, simulate its vibration response under different impact load excitations to obtain a simulated vibration response signal, and construct a response-load simulation data set; Step S2: Obtain the vibration response signal of the hanging structure under the excitation of a concentrated impact load, and obtain the corresponding time history of the concentrated load, and construct a response-load experimental data set; Step S3: Combine the convolutional layer and the bidirectional LSTM layer to obtain a convolutional recurrent block, and construct a deep convolutional recurrent neural network model by stacking multiple convolutional recurrent blocks. It is pre-trained on the response-load simulation data set. After pre-training, the learnable parameters of the network are preliminarily adjusted to obtain a pre-trained model; Step S4: Use transfer learning to fine-tune the parameters of the pre-trained model on the response-load experimental data set to obtain a load identification network model for the hanging structure; Step S5: Use the load identification network model of the hanging structure to inversely identify the concentrated impact load of the hanging structure.

2. The method for identifying suspended load based on transfer learning and recurrent convolutional network according to claim 1, wherein Preferably, step S1 includes Step S1.1: Establish a three-dimensional geometric model according to the geometric dimensions of the hanging structure; Step S1.2: Obtain the corresponding material properties according to the materials used in the hanging structure; Step S1.3: Establish a finite element model of the hanging structure according to the three-dimensional geometric model and the material properties; Step S1.4: Apply a concentrated impact load and the same constraint conditions as the monitored hanging structure to the finite element model of the hanging structure, and calculate the simulated vibration response signal of the response point; Step S1.5: Use the simulated vibration response signal as the network input feature, and use the corresponding time history of the concentrated impact load as the corresponding sample label to construct a response-load simulation data set.

3. A method for identifying suspended loads based on transfer learning and recurrent convolutional network according to claim 1, characterized in that, In step S2, the vibration response signal collected by the strain gauge or accelerometer is used as the network input feature, and the time history of the concentrated impact load obtained by the force sensor is used as the corresponding sample label to establish a response-load experimental data set containing the experimental vibration response signal and the impact load signal.

4. The hanging load recognition method based on transfer learning and recurrent convolutional network according to claim 1, characterized in that Step S3 includes: Step S3.1, construct a convolutional recurrent network feature extractor by stacking multiple convolutional recurrent blocks. The convolutional recurrent block is composed of a convolutional layer, a normalization layer, an activation layer, a pooling layer, a bidirectional LSTM layer, and a normalization layer connected in sequence. Connect multiple convolutional recurrent blocks in sequence to form a convolutional recurrent network feature extractor. The operations of the convolutional recurrent block are expressed as follows: , where x is the input feature of the convolutional recurrent block, represents a one-dimensional convolutional operation, represents one-dimensional batch normalization, is an activation function, represents a max pooling operation, represents a bidirectional LSTM operation. The convolutional layer and the pooling layer in the convolutional recurrent block expand the receptive field of the model. Batch normalization is used to accelerate convergence and ensure stable training, endows the model with non-linearity, and the LSTM captures the connections in the time series; Step S3.2: Construct a regressor, and use a fully connected layer to map the features extracted by the feature extractor into the corresponding load time series; Step S3.3: Divide the response-load simulation data set into a training set and a validation set, and pre-train the deep convolutional recurrent neural network model. The mean square error MSE is used as the loss function during the training process: ， where is the number of training samples, represents the sample length, represents the predicted value of the i-th sample, represents the true value of the i-th sample.

5. A method for identifying suspended loads based on transfer learning and recurrent convolutional network according to claim 1, characterized in that In step S4, freeze the pre-trained model The learnable parameters of its feature extractor.

6. The method for identifying a suspended load based on transfer learning and recurrent convolutional network according to claim 1, wherein In step S5, collect the vibration response of the monitored hanging structure under the impact load, input it into the load identification network model of the hanging structure, and obtain the inverse identification result of the concentrated load.

7. A method for identifying hanging loads based on transfer learning and recurrent convolutional network according to claim 1, characterized in that, The hanging structure includes a hanging object, a hanging inner lining plate, a hanging point, and a connecting piece.

8. A suspension load recognition system based on transfer learning and recurrent convolutional network, characterized in that, It includes A finite element modeling unit, which is used to construct a finite element model of the hanging structure, simulate its vibration response under different impact load excitations to obtain a simulated vibration response signal, and construct a response-load simulation data set; An acquisition unit, which is used to obtain the vibration response signal of the hanging structure under the excitation of a concentrated impact load, and obtain the corresponding time history of the concentrated load, and construct a response-load experimental data set; A construction unit, which is used to combine a convolutional layer and a bidirectional LSTM layer to obtain a convolutional recurrent block, and construct a deep convolutional recurrent neural network model by stacking multiple convolutional recurrent blocks, which is pre-trained on the response-load simulation data set. After pre-training, the learnable parameters of the network are preliminarily adjusted to obtain a pre-trained model; A fine-tuning unit, which is used to fine-tune the parameters of the pre-trained model on the response-load experimental data set by using transfer learning to obtain a load identification network model for the hanging structure; An identification unit, which is used to inversely identify the concentrated impact load of the hanging structure by using the load identification network model of the hanging structure; 9. A computer storage medium, characterized in that, The storage medium includes computer instructions, which when running on a computer, cause the computer to execute the method according to any one of claims 1-7; 10. An electronic device, characterized in that, The electronic device includes: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein When the processor executes the program, it implements the method according to any one of claims 1-7.

Images