Suspension load identification method, system, medium, and device

Abstract

A hanging load identification method, system, medium and equipment based on transfer learning and a recurrent convolutional network, in the method, a finite element model of a hanging structure is constructed, a simulation vibration response signal is obtained by simulating the vibration response of the hanging structure under different impact load excitations, and a response-load simulation dataset is constructed; a vibration response signal of the hanging structure under concentrated impact load excitation is obtained, and a corresponding concentrated load time history is obtained, a response-load experimental dataset is constructed; a convolutional recurrent block is obtained by combining a convolutional layer and a bidirectional LSTM layer, a deep convolutional recurrent neural network model is constructed by stacking multiple convolutional recurrent blocks, the model is pre-trained on the response-load simulation dataset, and a pre-training model is obtained; parameters of the pre-training model are fine-tuned on the response-load experimental dataset by using transfer learning, and a hanging structure load identification network model is obtained; and the concentrated impact load of the hanging structure is reversely identified by using the hanging structure load identification network model.

Description

Technical Field

[0001] This invention relates to the field of structural health monitoring technology, and in particular to a method, system, medium, and device for identifying suspended loads based on transfer learning and recurrent convolutional networks. Background Technology

[0002] As a critical connection component in modern engineering equipment and machinery, the safety and reliability of suspension structures are paramount. Suspension connections often bear concentrated loads, and accurately identifying these loads is crucial for assessing structural health, optimizing design, and ensuring safety. Directly measuring suspension loads using force sensors is often fraught with difficulties. In recent years, deep neural networks have been widely used in engineering due to their powerful feature extraction and end-to-end processing capabilities. However, in practical applications, obtaining sufficient training data is costly, and often there is a lack of sufficient data for training, which limits the performance of neural networks. Transfer learning can utilize source domain data for pre-training to acquire feature extraction capabilities, and then fine-tuning using a smaller number of samples in the target domain, effectively improving the model's performance in the target domain. Finite element models can obtain a large amount of sample data at a relatively low cost. However, a certain gap in feature distribution between simulation data and experimental data is unavoidable. Therefore, achieving accurate load identification under conditions of small experimental sample size and large simulation sample size is a challenging problem.

[0003] The information disclosed in the background section is only for enhancing the understanding of the background of this invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] This invention provides a method, system, medium, and device for identifying suspended loads based on transfer learning and recurrent convolutional networks, which can accurately identify concentrated impact loads on suspended structures under conditions of small experimental sample size and difficulty in training neural networks.

[0005] To achieve the above objectives, this disclosure provides the following technical solution:

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

[0007] Step S1: Construct a finite element model of the suspended structure, simulate its vibration response under different impact load excitations to obtain the simulated vibration response signal, and construct a response-load simulation dataset.

[0008] Step S2: Obtain the vibration response signal of the suspended structure under concentrated impact load excitation, and obtain the corresponding concentrated load time history to construct a response-load experimental dataset.

[0009] Step S3: Combine convolutional layers and bidirectional LSTM layers to obtain convolutional recurrent blocks. Construct a deep convolutional recurrent neural network model by stacking multiple convolutional recurrent blocks. Pre-train the model on the response-load simulation dataset. After pre-training, the learnable parameters of the network are initially adjusted to obtain the 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 the hanging structure load recognition network model;

[0011] Step S5: Use the suspended structure load identification network model to perform reverse identification of concentrated impact loads on the suspended structure.

[0012] In the method for identifying suspended loads based on transfer learning and recurrent convolutional networks, step S1 includes:

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

[0014] Step S1.2: Obtain the corresponding material properties based on the materials used in the suspension structure;

[0015] Step S1.3: Based on the three-dimensional geometric model and the material properties, establish a finite element model of the hanging structure;

[0016] Step S1.4: Based on the finite element model of the suspended structure, apply concentrated impact loads and the same constraints as the monitored suspended structure, and calculate the simulated vibration response signal at the response point;

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

[0018] In the suspended load identification method based on transfer learning and recurrent convolutional networks, in step S2, the vibration response signal collected by strain gauges or accelerometers is used as the network input feature, and the concentrated impact load time history obtained by force sensors 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 method for identifying suspended loads based on transfer learning and recurrent convolutional networks, step S3 includes:

[0020] Step S3.1: Convolutional recurrent network feature extractor is constructed by stacking multiple convolutional recurrent blocks. Each convolutional recurrent block consists of a convolutional layer, a normalization layer, an activation layer, a pooling layer, a bidirectional LSTM layer, and a normalization layer connected sequentially. Multiple convolutional recurrent blocks are then connected to form the convolutional recurrent network feature extractor. The operations of the convolutional recurrent blocks... It is expressed as follows:

[0021] , where x is the input feature of the convolutional recurrent block, This represents a one-dimensional convolution operation. This represents one-dimensional batch normalization. It is an activation function. This indicates a maximum pooling operation. This indicates a bidirectional LSTM operation. The convolutional and pooling layers in the convolutional loop block expand the model's receptive field. Batch normalization is used to accelerate convergence and ensure training stability. By imbuing the model with nonlinearity, LSTM captures the relationships between 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 dataset 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] in The number of training samples, Indicates the sample length. This represents the predicted value of the i-th sample. This represents the true value of the i-th sample.

[0026] In the suspended load recognition method based on transfer learning and recurrent convolutional networks, in step S4, the learnable parameters of the feature extractor of the pre-trained model are frozen, and the learnable parameters of the predictor are transferred and fine-tuned using the response-load experimental dataset. After the transfer learning is completed, the suspended structure load recognition network model is obtained. Fine-tuning refers to using fewer training epochs and a smaller learning rate than in the pre-training process when transferring and learning the learnable parameters of the predictor using the response-load experimental dataset.

[0027] In the suspended load identification method based on transfer learning and recurrent convolutional networks, 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 concentrated load inverse identification result.

[0028] In the suspended load identification method based on transfer learning and recurrent convolutional networks, the suspended structure includes the suspended object, the suspended inner lining plate, the hanging point, and the connecting parts.

[0029] A suspended load identification system based on transfer learning and recurrent convolutional networks includes,

[0030] Finite element modeling unit, which is used to construct finite element model of suspended structure, simulate its vibration response under different impact load excitation to obtain simulated vibration response signal, and construct response-load simulation dataset;

[0031] The acquisition unit is used to obtain the vibration response signal of the suspended structure under concentrated impact load excitation, as well as the corresponding concentrated load time history, and to construct the response-load experimental dataset.

[0032] The building block is used to combine convolutional layers and bidirectional LSTM layers to obtain convolutional recurrent blocks. By stacking multiple convolutional recurrent blocks, a deep convolutional recurrent neural network model is constructed. It is pre-trained on the response-load simulation dataset. After pre-training, the learnable parameters of the network are initially adjusted to obtain the pre-trained model.

[0033] The fine-tuning unit is used to fine-tune the parameters of the pre-trained model on the response-load experimental dataset using transfer learning to obtain the hanging structure load recognition network model.

[0034] The identification unit is used to reverse identify the concentrated impact load of the suspended structure using the suspended structure load identification network model.

[0035] A computer storage medium including computer instructions that, when run on a computer, cause the computer to perform the method.

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

[0037] Memory, processor, and computer programs stored in memory and executable on the processor, wherein,

[0038] The processor implements the method when executing the program.

[0039] Compared with existing technologies, this invention has the following advantages: Compared with traditional deep learning methods, this invention can make full use of finite element simulation data, reducing the number of experimental samples required and lowering the cost of data acquisition. Utilizing transfer learning technology, it improves the reverse identification accuracy of concentrated impact loads on suspended structures even when experimental data is limited and neural network training is difficult. Attached Figure Description

[0040] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0041] In the attached diagram:

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

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

[0044] Figure 3 This is a schematic diagram illustrating the identification of concentrated impact loads on a simulated suspended structure in one embodiment;

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

[0046] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0047] Specific embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While specific embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0048] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.

[0049] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.

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

[0051] Step S1: Construct a finite element model of the suspended structure, simulate its vibration response under different impact load excitations to obtain the simulated vibration response signal, and construct a response-load simulation dataset.

[0052] Step S2: Obtain the vibration response signal of the suspended structure under concentrated impact load excitation, and obtain the corresponding concentrated load time history to construct a response-load experimental dataset.

[0053] Step S3: Combine convolutional layers and bidirectional LSTM layers to obtain convolutional recurrent blocks. Construct a deep convolutional recurrent neural network model by stacking multiple convolutional recurrent blocks. Pre-train the model on the response-load simulation dataset. After pre-training, the learnable parameters of the network are initially adjusted to obtain the pre-trained model.

[0054] Step S4: Use transfer learning to fine-tune the parameters of the pre-trained model on the response-load experimental dataset to obtain the hanging structure load recognition network model;

[0055] Step S5: Use the suspended structure load identification network model to perform reverse identification of concentrated impact loads on the suspended structure.

[0056] In a preferred embodiment of the suspended load identification method based on transfer learning and recurrent convolutional networks, step S1 includes:

[0057] Step S1.1: Establish a three-dimensional geometric model based on the geometric dimensions of the hanging structure;

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

[0059] Step S1.3: Based on the three-dimensional geometric model and the material properties, establish a finite element model of the hanging structure;

[0060] Step S1.4: Based on the finite element model of the suspended structure, apply concentrated impact loads and the same constraints as the monitored suspended structure, and calculate the simulated vibration response signal at the response point;

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

[0062] In a preferred embodiment of the suspended load identification method based on transfer learning and recurrent convolutional networks, in step S2, the vibration response signal collected by strain gauges or accelerometers is used as the network input feature, and the concentrated impact load time history obtained by force sensors is used as the corresponding sample label to establish a response-load experimental dataset containing experimental vibration response signals and impact load signals.

[0063] In a preferred embodiment of the suspended load identification method based on transfer learning and recurrent convolutional networks, step S3 includes:

[0064] Step S3.1 involves constructing a convolutional recurrent network feature extractor by stacking multiple convolutional recurrent blocks. A convolutional recurrent block is a network structure primarily composed of convolutional layers and recurrent layers, specifically consisting of a convolutional layer, a normalization layer, an activation layer, a pooling layer, a bidirectional LSTM layer (i.e., a recurrent layer), and a normalization layer connected sequentially. Multiple convolutional recurrent blocks are then connected to form the convolutional recurrent network feature extractor. The operation of the convolutional recurrent blocks... It is expressed as follows:

[0065] , where x is the input feature of the convolutional recurrent block, This represents a one-dimensional convolution operation. This represents one-dimensional batch normalization. It is an activation function. This indicates a maximum pooling operation. This indicates a bidirectional LSTM operation. The convolutional and pooling layers in the convolutional loop block expand the model's receptive field. Batch normalization is used to accelerate convergence and ensure training stability. By imbuing the model with nonlinearity, LSTM captures the relationships between 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 dataset 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] in The number of training samples, Indicates the sample length. This represents the predicted value of the i-th sample. This represents the true value of the i-th sample.

[0070] In a preferred embodiment of the suspended load recognition method based on transfer learning and recurrent convolutional networks, in step S4, the learnable parameters of the feature extractor of the pre-trained model are frozen, and the learnable parameters of the predictor are transferred and fine-tuned using the response-load experimental dataset. After the transfer learning is completed, the suspended structure load recognition network model is obtained. Fine-tuning refers to using fewer training epochs and a smaller learning rate than in the pre-training process when transferring and learning the learnable parameters of the predictor using the response-load experimental dataset.

[0071] In a preferred embodiment of the suspended load identification method based on transfer learning and recurrent convolutional networks, 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 concentrated load inverse identification result.

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

[0073] A suspended load identification system based on transfer learning and recurrent convolutional networks includes,

[0074] Finite element modeling unit, which is used to construct finite element model of suspended structure, simulate its vibration response under different impact load excitation to obtain simulated vibration response signal, and construct response-load simulation dataset;

[0075] The acquisition unit is used to obtain the vibration response signal of the suspended structure under concentrated impact load excitation, as well as the corresponding concentrated load time history, and to construct the response-load experimental dataset.

[0076] The building block is used to combine convolutional layers and bidirectional LSTM layers to obtain convolutional recurrent blocks. By stacking multiple convolutional recurrent blocks, a deep convolutional recurrent neural network model is constructed. It is pre-trained on the response-load simulation dataset. After pre-training, the learnable parameters of the network are initially adjusted to obtain the pre-trained model.

[0077] The fine-tuning unit is used to fine-tune the parameters of the pre-trained model on the response-load experimental dataset using transfer learning to obtain the hanging structure load recognition network model.

[0078] The identification unit is used to reverse identify the concentrated impact load of the suspended structure using the suspended structure load identification network model.

[0079] A computer storage medium including computer instructions that, when run on a computer, cause the computer to perform the method.

[0080] An electronic device, the electronic device comprising:

[0081] Memory, processor, and computer programs stored in memory and executable on the processor, wherein,

[0082] The processor implements the method when executing the program.

[0083] In one embodiment, reference Figure 1 This invention provides a method for identifying suspended loads based on transfer learning and recurrent convolutional networks, the method comprising the following steps:

[0084] S1. Construct a high-precision finite element model of the suspended structure, simulate its vibration response under different impact load excitations, obtain the simulated vibration response signal, and construct a response-load simulation dataset.

[0085] S2. Vibration response signals of the suspended structure under concentrated impact load excitation are obtained by using strain gauges, accelerometers and other instruments, and the corresponding concentrated load time history is obtained by using force sensors to construct a response-load experimental dataset.

[0086] S3 combines convolutional layers and bidirectional LSTM layers to obtain convolutional recurrent blocks. By stacking multiple convolutional recurrent blocks, a deep convolutional recurrent neural network model is constructed. Pre-training is performed on a simulation dataset. After pre-training, the learnable parameters of the network are initially adjusted, resulting in a pre-trained model.

[0087] S4. Using transfer learning techniques, the parameters of the pre-trained model are fine-tuned on the experimental dataset to obtain the hanging structure load recognition network model.

[0088] S5 uses a suspended structure load identification network model with finely tuned parameters to perform reverse identification of concentrated impact loads on the suspended structure.

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

[0090] S1.1 Based on the geometric dimensions of each component of the suspension structure (including the suspended object, inner lining plate, hanging points and connectors, etc.), construct an accurate three-dimensional geometric model;

[0091] S1.2 Based on the material list, obtain the material properties of each component of the suspension structure;

[0092] S1.3 Based on the three-dimensional geometric model in S1.1 and the material properties in S1.2, establish the finite element model of the hanging structure;

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

[0094] S1.5 uses the simulated vibration signal obtained in S1.4 as the network input feature and the corresponding concentrated impact load time history as the corresponding sample label to construct the simulation dataset.

[0095] Optionally, in step S2, the vibration response signal of the suspended structure collected by strain gauges, accelerometers, etc., 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 an experimental dataset containing experimental vibration response signal and impact load signal.

[0096] Optionally, in step S3, the convolutional layer and the bidirectional LSTM layer are combined to obtain a convolutional recurrent block. By stacking multiple convolutional recurrent blocks, a deep convolutional recurrent neural network model is constructed, such as... Figure 2 As shown. Pre-training is performed on the simulation dataset to obtain a pre-trained model, specifically including:

[0097] S3.1, a convolutional recurrent network feature extractor is constructed by stacking multiple convolutional recurrent blocks. A convolutional recurrent block is a network structure mainly composed of convolutional layers and recurrent layers, specifically consisting of a convolutional layer, a normalization layer, an activation layer, a pooling layer, a bidirectional LSTM layer (i.e., a recurrent layer), and a normalization layer connected sequentially. Multiple convolutional recurrent blocks are then connected sequentially to form the convolutional recurrent network feature extractor. The operation of the convolutional recurrent blocks... It can be represented as follows:

[0098]

[0099] Where x is the input feature of the convolutional recurrent block, This represents a one-dimensional convolution operation. This represents one-dimensional batch normalization. It is a commonly used activation function. This indicates a maximum pooling operation. This indicates a bidirectional LSTM operation. The convolutional and pooling layers in the convolutional loop block expand the model's receptive field, while batch normalization accelerates convergence and ensures training stability. By giving the model a certain degree of nonlinearity, 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, Based on the actual training results, select an appropriate ratio to divide the simulation dataset into a training set and a validation set for pre-training the deep convolutional recurrent neural network model. The mean squared error (MSE) is used as the loss function during the training process.

[0102]

[0103] in The number of training samples, Indicates the sample length. This represents the predicted value of the i-th sample. This represents the true value of the i-th sample.

[0104] Optionally, in step S4, the pre-trained model from step S3 is further trained using transfer learning. First, the learnable parameters of its feature extractor are frozen, and then the learnable parameters of the predictor are fine-tuned using an experimental dataset. After transfer learning, the suspended structure load recognition network model is obtained. Fine-tuning refers to using fewer training epochs and a smaller learning rate than in the pre-training process when performing transfer learning on the learnable parameters of the predictor using the response-load experimental dataset.

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

[0106] In one embodiment, reference Figure 3 The process of identifying concentrated impact loads using a simulated suspended structure as the research object is as follows:

[0107] 1. The simulated suspension structure mainly consists of two cylindrical parts, simulated suspension points, suspension point liners, and a test platform base plate. The two cylinders are connected by bolts. The inner diameter of each cylinder is 500mm, and the wall thickness is 3mm; the height of the upper cylinder is 500mm, the height of the lower cylinder is 250mm, and the height of the entire cylindrical structure is 750mm. The simulated suspension points and suspension point liners are bolted onto the upper cylinder. Two reinforcing rings are symmetrically arranged inside the upper cylinder, with a height of 20mm and a thickness of 5mm.

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

[0109] Based on the geometric dimensions of each component of the suspension structure (including the suspended object, inner lining plate, hanging points and connectors, etc.), construct an accurate three-dimensional geometric model.

[0110] Based on the bill of materials, obtain the material properties of each component of the suspension structure;

[0111] Based on the three-dimensional geometric model and material properties, a finite element model of the suspended structure is established;

[0112] Based on the established finite element model of the suspended structure, corresponding constraints are added, concentrated impact loads are applied, and the simulated vibration response signal at the response location is calculated.

[0113] The simulated vibration signal calculated from the finite element model of the suspended structure was used as the network input feature, and the corresponding concentrated impact load time history was used as the corresponding sample label to construct a simulation dataset containing 101 simulation samples.

[0114] 3. At the simulated hanging point, a concentrated impact load was applied using a hammer. The vibration response signal of the suspended structure acquired by strain gauges was used as the network input feature, and the time history of the concentrated impact load acquired by the force sensor in the hammer was used as the corresponding sample label to establish an experimental dataset containing experimental vibration response signals and impact load signals, with a total of 60 experimental samples.

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

[0116] A convolutional recurrent network (CRN) feature extractor is constructed by stacking multiple convolutional recurrent blocks. A CRN block is a network structure primarily composed of convolutional and recurrent layers, specifically consisting of a convolutional layer, a normalization layer, an activation layer, a pooling layer, a bidirectional LSTM layer (i.e., a recurrent layer), and a normalization layer connected sequentially. Multiple CRN blocks are then connected to form the CRN feature extractor. The operation of the CRN blocks... It can be represented as follows:

[0117]

[0118] Where x is the input feature of the convolutional recurrent block, This represents a one-dimensional convolution operation. This represents one-dimensional batch normalization. It is a commonly used activation function. This indicates a maximum pooling operation. This indicates a bidirectional LSTM operation. The convolutional and pooling layers in the convolutional loop block expand the model's receptive field, while batch normalization accelerates convergence and ensures training stability. By giving the model a certain degree of nonlinearity, LSTM can capture the relationships in 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] Based on the actual training results, a suitable ratio is selected to divide the simulation dataset into training and validation sets for pre-training of the deep convolutional recurrent neural network model. The mean squared error (MSE) is used as the loss function during the training process.

[0121]

[0122] in The number of training samples, Indicates the sample length. This represents the predicted value of the i-th sample. This 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 dataset. After transfer learning, the suspended structure load recognition network model is obtained. Fine-tuning refers to using fewer training epochs and a smaller learning rate than the pre-training process when transferring learning the learnable parameters of the predictor using the response-load experimental dataset.

[0124] 6. The collected vibration response of the suspended structure under impact load is input into the suspended structure load identification network to obtain the concentrated load inverse identification result.

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

[0126]

[0127]

[0128] The fine-tuned load recognition network for suspended structures was tested using 30 experimental test samples. The pre-trained model had an average RE of 27.12% and an average PRE of 27.10% on the 30 test samples. The deep convolutional recurrent neural network model after transfer learning had an average RE of 10.72% and an average PRE of 7.68% on the test set. Compared to the pre-trained model, the average RE was reduced to 60.47% and the average PRE was reduced to 71.66% of the original. Figures 4(a) to 4(b) The figures illustrate the load recognition results of a deep convolutional recurrent neural network model on the same test sample before and after transfer learning. Figure 4(a) shows the load recognition result of the pre-trained model, and Figure 4(b) shows the load recognition result of the load recognition network for suspended structures after transfer learning. The load recognition accuracy is significantly improved after transfer learning. This invention is applied to the concentrated load recognition of suspended structures. It trains a deep neural network in stages using simulation data and a small amount of experimental data, and uses transfer learning technology to improve the accuracy of inverse load recognition, effectively solving the problems of high cost and limited quantity of experimental samples.

[0129] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.

Claims

1. A method for identifying suspended loads based on transfer learning and recurrent convolutional networks, characterized in that, Includes the following steps: Step S1: Construct a finite element model of the suspended structure, simulate its vibration response under different impact load excitations to obtain the simulated vibration response signal, and construct a response-load simulation dataset. Step S2: Obtain the vibration response signal of the suspended structure under concentrated impact load excitation, and obtain the corresponding concentrated load time history to construct a response-load experimental dataset. Step S3: Combine convolutional layers and bidirectional LSTM layers to obtain convolutional recurrent blocks. Construct a deep convolutional recurrent neural network model by stacking multiple convolutional recurrent blocks. Pre-train the model on the response-load simulation dataset. After pre-training, the learnable parameters of the network are initially adjusted to obtain the pre-trained model. Step S4: Use transfer learning to fine-tune the parameters of the pre-trained model on the response-load experimental dataset to obtain the hanging structure load recognition network model; Step S5: Use the suspended structure load identification network model to perform reverse identification of concentrated impact loads on the suspended structure; Step S3 includes: Step S3.1: Convolutional recurrent network feature extractor is constructed by stacking multiple convolutional recurrent blocks. Each convolutional recurrent block consists of a convolutional layer, a normalization layer, an activation layer, a pooling layer, a bidirectional LSTM layer, and a normalization layer connected sequentially. Multiple convolutional recurrent blocks are then connected to form the convolutional recurrent network feature extractor. The operations of the convolutional recurrent blocks... It is expressed as follows: , where x is the input feature of the convolutional recurrent block. This represents a one-dimensional convolution operation. This represents one-dimensional batch normalization. It is an activation function. This indicates a maximum pooling operation. This indicates a bidirectional LSTM operation. The convolutional and pooling layers in the convolutional loop block expand the model's receptive field. Batch normalization is used to accelerate convergence and ensure training stability. By imbuing the model with nonlinearity, LSTM captures the relationships between time series. 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; Step S3.3: Divide the response-load simulation dataset 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. ， in The number of training samples, Indicates the sample length. This represents the predicted value of the i-th sample. This represents the true value of the i-th sample; In step S4, the learnable parameters of the feature extractor of the pre-trained model are frozen; In step S5, the vibration response of the monitored suspended structure under impact load is collected and input into the load identification network model of the suspended structure to obtain the concentrated load inverse identification result.

2. The method for identifying suspended loads based on transfer learning and recurrent convolutional networks according to claim 1, characterized in that, Step S1 includes, Step S1.1: Establish a three-dimensional geometric model based on the geometric dimensions of the hanging structure; Step S1.2: Obtain the corresponding material properties based on the materials used in the suspension structure; Step S1.3: Based on the three-dimensional geometric model and the material properties, establish a finite element model of the hanging structure; Step S1.4: Based on the finite element model of the suspended structure, apply concentrated impact loads and the same constraints as the monitored suspended structure, and calculate the simulated vibration response signal at the response point; Step S1.5: Using the simulated vibration response signal as the network input feature and the corresponding concentrated impact load time history as the corresponding sample label, construct a response-load simulation dataset.

3. The method for identifying suspended loads based on transfer learning and recurrent convolutional networks according to claim 1, characterized in that, In step S2, the vibration response signal acquired by strain gauges or accelerometers is used as the network input feature, and the concentrated impact load time history obtained by force sensors is used as the corresponding sample label to establish a response-load experimental dataset containing experimental vibration response signals and impact load signals.

4. The method for identifying suspended loads based on transfer learning and recurrent convolutional networks according to claim 1, characterized in that, The hanging structure includes the object to be hung, the inner lining plate, the hanging points, and the connecting parts.

5. A suspended load identification system based on transfer learning and recurrent convolutional networks, used to perform the method according to any one of claims 1-4, characterized in that, It includes, Finite element modeling unit, which is used to construct finite element model of suspended structure, simulate its vibration response under different impact load excitation to obtain simulated vibration response signal, and construct response-load simulation dataset; The acquisition unit is used to obtain the vibration response signal of the suspended structure under concentrated impact load excitation, as well as the corresponding concentrated load time history, and to construct the response-load experimental dataset. The building block is used to combine convolutional layers and bidirectional LSTM layers to obtain convolutional recurrent blocks. By stacking multiple convolutional recurrent blocks, a deep convolutional recurrent neural network model is constructed. It is pre-trained on the response-load simulation dataset. After pre-training, the learnable parameters of the network are initially adjusted to obtain the pre-trained model. The fine-tuning unit is used to fine-tune the parameters of the pre-trained model on the response-load experimental dataset using transfer learning to obtain the hanging structure load recognition network model. The identification unit is used to reverse identify the concentrated impact load of the suspended structure using the suspended structure load identification network model.

6. A computer storage medium, characterized in that, The storage medium includes computer instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1-4.

7. An electronic device, characterized in that, The electronic device includes: Memory, processor, and computer programs stored in memory and executable on the processor, wherein, When the processor executes the program, it implements the method as described in any one of claims 1-4.

Images