A neural network dynamic load identification method and system based on a particle swarm optimization algorithm

Abstract

The application discloses a neural network dynamic load identification method and system based on a particle swarm optimization algorithm, and the method comprises the following steps: taking the number of hidden layers of a neural network dynamic load identification model and the number of neurons of each hidden layer as optimization targets, constructing particles and initializing related parameters of a particle swarm optimization algorithm; building a corresponding neural network dynamic load identification model for each particle, loading a pre-prepared dynamic load identification calibration training data set to train the model, and updating the model by using the particle swarm optimization algorithm; selecting a particle with the minimum fitness value as a final hyperparameter combination of the neural network dynamic load identification model, building a corresponding neural network dynamic load identification model, training the model, inputting structural dynamic response measurement data, and outputting a load identification result. The application considers an updating strategy of identification precision and identification efficiency, reduces the threshold of the neural network dynamic load identification method, and has engineering practical value.

Description

Technical Field

[0001] This invention belongs to the field of dynamic load identification, specifically relating to a neural network dynamic load identification method and system based on particle swarm optimization algorithm. Background Technology

[0002] Mastering the characteristics of structural dynamic loads plays a crucial role in structural dynamics design, health monitoring, and reliability analysis. Due to limitations in engineering environment and technology, directly acquiring structural dynamic load information by deploying force sensors on the structure is difficult. Therefore, structural dynamic load identification methods offer an indirect approach to obtaining structural dynamic loads based on more easily measurable structural dynamic characteristics and dynamic response information. Originating in the 1970s, these indirect methods have evolved over decades, forming a series of classic approaches, mainly divided into two categories: frequency domain dynamic load identification methods, represented by the direct inversion of the frequency response function, and time domain dynamic load identification methods, represented by the traditional integration method. Both of these typical dynamic load identification methods have certain limitations. For example, frequency domain methods are unsuitable for impact dynamic load identification due to their sampling time requirements; while time domain methods suffer from unstable identification results due to error accumulation and other factors. With the continuous improvement of machine learning, dynamic load identification methods based on deep learning have emerged. Deep learning-based dynamic load identification methods have the advantage of not requiring prior knowledge of structural dynamic characteristics. They directly establish a mapping relationship between structural dynamic response and dynamic load using prior data, thereby enabling the acquisition of unknown external loads on the structure through data-driven approaches. Among numerous deep learning-based dynamic load identification methods, those based on artificial neural networks stand out due to their strong adaptability and wide range of applications.

[0003] Although artificial neural network (ANN) dynamic load identification methods are applicable to various structures and loads, several problems still plague researchers. The construction of an ANN often requires pre-setting hyperparameters for the neural network model. These hyperparameters directly affect the accuracy and efficiency of load identification by the ANN model. However, selecting hyperparameters typically requires researchers to spend considerable time and effort experimenting with different combinations and gradually accumulating experience, significantly limiting the application and widespread adoption of ANN dynamic load identification methods. Therefore, how to quickly and accurately select the hyperparameters of the ANN dynamic load identification model to achieve ANN dynamic load identification is a key research focus. Summary of the Invention

[0004] Purpose of the invention: To address the shortcomings of existing technologies, this invention proposes a neural network dynamic load identification method based on particle swarm optimization algorithm. Through an update strategy that balances identification accuracy and efficiency, the hyperparameter selection of the neural network dynamic load identification model can be completed in iterations, lowering the threshold of the neural network dynamic load method. Moreover, it has high accuracy and stability on various types of loads, meeting theoretical expectations and engineering requirements.

[0005] Another objective of this invention is to provide a neural network dynamic load identification system, electronic device, and computer storage medium based on particle swarm optimization algorithm.

[0006] Technical solution: To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0007] Firstly, a neural network dynamic load identification method based on particle swarm optimization algorithm includes the following steps:

[0008] The number of hidden layers in the neural network dynamic load recognition model and the number of neurons in each hidden layer As the optimization objective, construct particles Establish a particle swarm optimization algorithm and initialize the relevant parameters of the particle swarm optimization algorithm;

[0009] A corresponding neural network dynamic load recognition model is built for each particle. A pre-prepared dynamic load recognition calibration training dataset is loaded to train the model. The model obtains predicted load information based on a set of input structural dynamic responses. The training error between the predicted load information and the calibrated load information is calculated. The fitness of each particle is calculated based on the training error and training time. The optimal position of each particle in one iteration is determined based on the fitness calculation results. and global optimal position The particles are updated at a preset number of iterations until the training of the last generation of particles is completed.

[0010] The particle with the smallest fitness value in all iterations is selected as the final hyperparameter combination of the neural network dynamic load recognition model. The corresponding neural network dynamic load recognition model is built based on the final hyperparameter combination and trained using the training dataset.

[0011] The trained neural network dynamic load identification model is used to identify unknown dynamic loads. The input is the structural dynamic response measurement data, and the output is the load identification result.

[0012] Secondly, a neural network dynamic load identification system based on particle swarm optimization algorithm includes:

[0013] The particle initialization module is used to initialize the number of hidden layers in the neural network dynamic load recognition model. and the number of neurons in each hidden layer As the optimization objective, construct particles Establish a particle swarm optimization algorithm and initialize the relevant parameters of the particle swarm optimization algorithm;

[0014] The hyperparameter optimization module is used to build a corresponding neural network dynamic load recognition model for each particle, load a pre-prepared dynamic load recognition calibration training dataset to train the model, the model obtains predicted load information based on a set of input structural dynamic responses, calculates the training error between the predicted load information and the calibration load information, calculates the fitness of each particle based on the training error and training time, determines the individual optimal particle and the global optimal particle based on the fitness calculation results, and updates the particles at a preset number of iterations until the training of the last generation of particles ends.

[0015] The model determination module is used to select the particle with the smallest fitness value in all iterations as the final hyperparameter combination of the neural network dynamic load recognition model, build the corresponding neural network dynamic load recognition model based on the final hyperparameter combination, and train it with the training dataset.

[0016] The dynamic load identification module is used to identify unknown dynamic loads using a trained neural network dynamic load identification model. It takes structural dynamic response measurement data as input and outputs load identification results.

[0017] Thirdly, an electronic device includes: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the neural network dynamic load identification method based on particle swarm optimization algorithm as described in the first aspect of the invention.

[0018] Fourthly, a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the neural network dynamic load identification method based on particle swarm optimization algorithm as described above.

[0019] Fifthly, a computer program product includes a computer program, characterized in that, when the computer program is executed by a processor, it implements the steps of the neural network dynamic load identification method based on particle swarm optimization algorithm as described above.

[0020] Beneficial Effects: The neural network dynamic load identification method based on particle swarm optimization (PSO) proposed in this invention uses the hyperparameters of the neural network dynamic load identification model (the number of hidden layers and the number of neurons in each hidden layer) as the optimization objective. A particle swarm is constructed, and the PSO algorithm adaptively selects the number of hidden layers and the number of neurons in each layer of the fully connected neural network. This solves the problem of network structure selection relying on manual experience, lowering the application threshold of neural network dynamic load identification methods. In the optimization process, this invention simultaneously considers load identification error and training time. By constructing an objective function that includes normalized training time, it achieves comprehensive optimization of identification accuracy and computational efficiency, improving the overall computational performance of the method. Furthermore, the method proposed in this invention has good noise resistance and numerical robustness, making it suitable for dynamic load identification problems under noisy measurement conditions and possessing high engineering application potential. Attached Figure Description

[0021] Figure 1 This is a flowchart of the neural network dynamic load identification method based on particle swarm optimization algorithm of the present invention;

[0022] Figure 2 This is a schematic diagram of a simply supported beam model under a single-point concentrated load in an embodiment of the present invention;

[0023] Figure 3 This is the impact load identification result in the embodiment of the present invention;

[0024] Figure 4 This embodiment of the invention does not consider the impact load identification result during training time;

[0025] Figure 5 In this embodiment of the invention, the training time impact load identification result is directly introduced;

[0026] Figure 6 This is the sinusoidal load identification result in the embodiment of the present invention. Detailed Implementation

[0027] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0028] This invention proposes a neural network dynamic load recognition method based on particle swarm optimization. This method uses structural dynamics equations as its physical basis and avoids the ill-conditioned problems in traditional inverse recognition by constructing a data-driven mapping of "structural response-external load". At the same time, the particle swarm optimization algorithm is used to optimize the number of hidden layers and the number of neurons in each layer of the neural network, and the training error and training time are jointly considered in the objective function (fitness), thereby improving the recognition efficiency while ensuring recognition accuracy.

[0029] To gain a clearer understanding of the invention, the basic principles of the neural network dynamic load recognition model will be introduced first.

[0030] In dynamic load identification problems, it is often necessary to decouple the structural dynamic equations through modal coordinate transformation and obtain the dynamic equations under a single mode by utilizing the orthogonality of the structure's natural mode shapes. Finally, the relationship between the structural response and the load is established through methods such as convolution integrals and time-domain sampling, as shown in the formula. As shown:

[0031]

[0032] in Represents structural response, The structural dynamic characteristic matrix, This represents dynamic load.

[0033] In traditional dynamic load identification methods, it is necessary to use formulas... The inverse operation is performed to obtain the unknown dynamic load. However, the neural network dynamic load identification model is based on a known set of structural dynamic response information. In this case, it is compared with the load information used for calibration. A direct mapping relationship is established between them, thereby enabling direct control through structural dynamic response. Identifying unknown dynamic loads This process does not require knowledge of the structural dynamic characteristics, thus avoiding ill-posed problems that may occur in the inversion operation.

[0034] A fully connected neural network consists of an input layer, hidden layers, and an output layer. Specifically, in the dynamic load identification problem, the input layer contains structural dynamic response information, and its number of neurons is only related to the number of input dynamic response features. The output layer represents the predicted dynamic load result of the fully connected neural network, and its number of neurons is related to the load to be identified. The neurons in each hidden layer are connected to all neurons in the previous layer, thus forming a fully connected neural network dynamic load identification model.

[0035] Input layer input weight matrix and bias vector Passing it to the first hidden layer, we get the following formula: The output shown :

[0036]

[0037] Then the linear combination of the first hidden layer Through activation function A nonlinear transformation is performed to obtain the activation output as shown in the formula. :

[0038]

[0039] Activate output The input is passed to the second hidden layer and undergoes the same operation until the last layer, the output layer. The operation can be written in the form of a formula:

[0040]

[0041] in Representing the Layer weight matrix, Indicates the first Layer bias matrix, This represents the number of hidden layers.

[0042] Activate output After the final weight matrix and bias vector The output is obtained as shown in the formula. That is, the corresponding load information :

[0043]

[0044] The training process of a fully connected neural network relies on the backpropagation algorithm. In the dynamic load identification task, the mean square error is selected. As the loss function, the gradient of the loss function with respect to each weight and bias is calculated using the gradient descent method. and Then, the weights and biases are updated based on the gradient, as shown in the formula:

[0045]

[0046] in and They represent the first The updated weights and biases of the layer This is the learning rate.

[0047] Repeat the above process until training is complete, then change the input of the fully connected neural network model to the dynamic response information of the structure corresponding to the load to be identified. The output is the predicted payload value obtained by the fully connected neural network. This means that the task of identifying dynamic loads using a fully connected neural network has been completed.

[0048] The following describes the method of the present invention for identifying unknown dynamic loads by optimizing the hyperparameters of a neural network dynamic load identification model based on the particle swarm optimization algorithm.

[0049] The number of hidden layers in the fully connected neural network dynamic load recognition model and the number of neurons contained in each hidden layer. As the optimization objective, and as shown in the formula The particles shown are randomly initialized to generate a finite number of particle populations:

[0050]

[0051] Every particle A corresponding fully connected neural network dynamic load recognition model is constructed. The neural network recognition model is trained using a pre-prepared training set consisting of multiple sets of calibrated loads and their corresponding structural dynamic responses. This allows the mapping relationship between load and response to be obtained. The error between the supervised loads (i.e., calibrated loads) in the training set and the recognition results obtained by each neural network dynamic load recognition model is statistically analyzed. and the training time of each neural network dynamic load recognition model This invention constructs an objective function that comprehensively considers both training accuracy and training efficiency. As shown in the formula:

[0052]

[0053] formula middle This represents the normalized training time, which is achieved by normalizing the training time of each model to the training time of the neural network dynamic load recognition model corresponding to the most complex particle in the search space. The particle swarm optimization algorithm aims to minimize the objective function. To optimize the objective, a high-precision and high-efficiency neural network dynamic load recognition model is constructed. The objective function is also called the fitness function, and the objective function value is also called the fitness value. An iteration is considered complete when the neural network dynamic load recognition model for all particles in the particle swarm has been trained. Based on the differences in objective function values ​​across all iterations, the particle with the smallest objective function value is selected as the individual optimal particle; and based on the differences in objective function values ​​across all iterations, the particle with the smallest objective function value is selected as the swarm optimal particle (global optimal particle), which guides the position updates of particles within the swarm into the next iteration.

[0054] No. The position and velocity of the time-phase particle are respectively and The particle position is the hyperparameter combination, and the particle velocity represents the particle's iteration distance and direction. The optimal individual particle is... The optimal particle in the swarm is Particle updates are performed according to the formula and formula The update proceeds to the next iteration in the manner shown:

[0055]

[0056]

[0057] in Inertial weights; This represents the acceleration constant, which affects the degree to which particles approach their individual optimal position and the group's optimal position. for Random numbers between these ranges are introduced to enhance exploration capabilities by introducing randomness.

[0058] Since the hyperparameters of the neural network dynamic load recognition model targeted in this invention are selected for discrete optimization, in order to obtain Next, a rounding operation is performed to ensure that each particle always corresponds to a model suitable for building a neural network dynamic load recognition system. After training all particles in the final generation, the particle with the smallest objective function value is selected as the final hyperparameter combination for the neural network dynamic load recognition system, and the corresponding neural network model is built. This model is then retrained to obtain the mapping relationship between load and response, and finally, the prediction set is introduced to achieve high-precision and high-efficiency recognition of unknown dynamic loads.

[0059] Figure 1 The flowchart of the neural network dynamic load identification method based on particle swarm optimization algorithm of the present invention is shown below, with reference to... Figure 1 The specific steps are as follows:

[0060] Step S1: Determine the hyperparameter optimization object and the corresponding search space, and set the particle number, search iteration count, inertia weight, and acceleration constant of the particle swarm optimization algorithm; randomly generate parameters as shown in the formula. The particle positions shown are determined by the optimization object, and each position represents a corresponding neural network dynamic load recognition model architecture.

[0061] Step S2: Build a neural network model for each particle, load the pre-prepared dynamic load recognition dataset to train the network model, and record the mean square error and training time of the model; according to the formula... The objective function expression shown calculates the fitness of each particle and determines the optimal position of each particle based on the calculation results. and global optimal position According to the formula and formula Update the particle velocity and position; rebuild the corresponding neural network model based on the updated particle position and train it, repeating the above update operation until the search ends;

[0062] Step S3: Construct the corresponding neural network dynamic load identification model based on the hyperparameter combination corresponding to the optimal particle obtained in step S2, and train it again using the pre-prepared training set data. After training, input the structural dynamic response corresponding to the load to be identified, and the prediction result is the unknown dynamic load identification result.

[0063] To verify the performance of the method of the present invention, the following verification experiments were conducted.

[0064] right Figure 2 The simply supported beam structure shown is subjected to a concentrated load. The parameters of the simply supported beam are shown in Table 1:

[0065] Table 1 Geometric parameters and material properties of simply supported beams

[0066]

[0067] exist Figure 2 On the simply supported beam model shown, An impact load and a sinusoidal load are applied at the point, respectively. The impact load amplitude is 75 N, and the impact duration is 0.01 s. The expression for the sinusoidal load is: Nine points evenly distributed on the beam were selected as response measurement points for dynamic load identification. To demonstrate the stability of the method, 20 dB of Gaussian white noise was added to the response.

[0068] Step 1: Set the hyperparameter objects to be selected in the neural network dynamic load recognition model and their corresponding search space, and set the initial parameters of the particle swarm optimization algorithm. The hyperparameter search space and the initial parameters of the particle swarm optimization algorithm are shown in Tables 2 and 3:

[0069] Table 2 Hyperparameter search space settings

[0070]

[0071] Table 3 Optimization of initial parameter settings

[0072]

[0073] Step 2: Based on the settings in Step 1, randomly generate the positions of 30 different particles in the optimization object search space. Each particle position corresponds to a hyperparameter combination used to build a neural network dynamic load recognition model.

[0074] Step 3: Based on the 30 particle positions generated in Step 2, construct corresponding fully connected neural network dynamic load recognition models, train all neural network dynamic load recognition models using the training set, and record the corresponding training set loss and training time.

[0075] Step 4: Calculate the fitness function value of each particle, and determine the global best particle and the individual best particle for this iteration based on the fitness function value.

[0076] Step 5: Based on the particle position and particle velocity update formula, determine the position and velocity of the 30 particles in the next iteration.

[0077] Step 6: Based on the particle positions determined in the new round of iterations in Step 5, construct the corresponding neural network dynamic load recognition model, and repeat Steps 3-5 until all 10 iterations are completed. The final hyperparameter combination is shown in Table 4.

[0078] Table 4. Results of Hyperparameter Selection

[0079]

[0080] Two comparison schemes were set up: "Ignoring training time" means that the training time is not considered in the objective function (i.e., Equation 8), and hyperparameter optimization is performed only based on the training error (i.e., model recognition loss); "Directly incorporating training time" means that the training time value is directly used for optimization in the objective function of Equation 8.

[0081] Step 7: Construct a neural network model based on the selected hyperparameter combination, and predict the prediction set consisting of the loads to be identified to obtain the final load identification result.

[0082] Figure 3 This is the impact load identification result obtained using the method of this invention in the verification experiment; Figure 4 The impact load identification results do not consider training time; Figure 5 It directly incorporates the training time impact load identification results; Figure 6 This is the sinusoidal load identification result obtained by the method of the present invention.

[0083] The proposed method for impact load identification yields a relative error of 0.97% with a training time of 1482 seconds; the method without considering training time yields a relative error of 1.07% with a training time of 2442 seconds; and the method directly incorporating training time yields a relative error of 3.51% with a training time of 431 seconds. While directly incorporating training time into the objective function significantly shortens the training time of the fully connected neural network model compared to the particle swarm optimization algorithm without considering training time, it drastically reduces the load identification accuracy. This is because after the neural network model corresponding to any particle has been trained, the training error is often relatively small, but the training time is typically tens or even hundreds of times greater. This leads to the objective function value being more dominated by training time. Since the particle swarm optimization algorithm updates the position in the direction of decreasing objective function, the particle update during the hyperparameter optimization process falls into the "misconception" of shortening training time rather than improving identification accuracy. Therefore, this invention introduces normalized training time into the objective function of the particle swarm optimization algorithm, reducing the training time of the resulting fully connected neural network model by 39.31%, while maintaining the dynamic load recognition accuracy essentially unchanged. This achieves hyperparameter optimization that comprehensively considers both recognition accuracy and efficiency. Furthermore, the method proposed in this invention achieves a relative error of 2.21% for sinusoidal load recognition with a training time of 162 seconds, indicating that this invention has high accuracy across different types of loads and meets engineering error requirements.

[0084] According to another embodiment of the present invention, a neural network dynamic load identification system based on particle swarm optimization algorithm is provided, comprising:

[0085] The particle initialization module is used to initialize the number of hidden layers in the neural network dynamic load recognition model. and the number of neurons in each hidden layer As the optimization objective, construct particles Establish a particle swarm optimization algorithm and initialize the relevant parameters of the particle swarm optimization algorithm;

[0086] The hyperparameter optimization module is used to build a corresponding neural network dynamic load recognition model for each particle, load a pre-prepared dynamic load recognition calibration training dataset to train the model, the model obtains predicted load information based on a set of input structural dynamic responses, calculates the training error between the predicted load information and the calibration load information, calculates the fitness of each particle based on the training error and training time, determines the individual optimal particle and the global optimal particle based on the fitness calculation results, and updates the particles at a preset number of iterations until the training of the last generation of particles ends.

[0087] The model determination module is used to select the particle with the smallest fitness value in all iterations as the final hyperparameter combination of the neural network dynamic load recognition model, build the corresponding neural network dynamic load recognition model based on the final hyperparameter combination, and train it with the training dataset.

[0088] The dynamic load identification module is used to identify unknown dynamic loads using a trained neural network dynamic load identification model. It takes structural dynamic response measurement data as input and outputs load identification results.

[0089] It should be understood that the neural network dynamic load identification system based on particle swarm optimization algorithm provided in this embodiment can realize all the technical solutions in the above method embodiments. The functions of each functional module can be specifically implemented according to the methods in the above method embodiments. The specific implementation process can be referred to the relevant descriptions in the above embodiments, which will not be repeated here.

[0090] The present invention also provides an electronic device, comprising: one or more processors; a memory; and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of the neural network dynamic load identification method based on particle swarm optimization algorithm as described above.

[0091] The present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the neural network dynamic load identification method based on particle swarm optimization algorithm as described above.

[0092] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus (systems), computer devices, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0093] This invention is described with reference to a flowchart of a method according to embodiments of the invention. It should be understood that each step in the flowchart and combinations thereof can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A device for a function specified in one or more processes.

[0094] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 The function specified in one or more processes.

[0095] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 Steps of a specified function in one or more processes.

Claims

1. A neural network dynamic load identification method based on particle swarm optimization algorithm, characterized in that, Includes the following steps: The number of hidden layers in the neural network dynamic load recognition model and the number of neurons in each hidden layer As the optimization objective, construct particles Establish a particle swarm optimization algorithm and initialize the relevant parameters of the particle swarm optimization algorithm; A corresponding neural network dynamic load recognition model is built for each particle. The model is trained by loading a pre-prepared dynamic load recognition calibration training dataset. The model obtains predicted load information based on a set of input structural dynamic responses. The training error between the predicted load information and the calibration load information is calculated. The fitness of each particle is calculated based on the training error and training time. The optimal individual particle and the global optimal particle are determined based on the fitness calculation results. The particles are updated at a preset number of iterations until the training of the last generation of particles ends. The particle with the smallest fitness value in all iterations is selected as the final hyperparameter combination of the neural network dynamic load recognition model. The corresponding neural network dynamic load recognition model is built based on the final hyperparameter combination and trained using the training dataset. The trained neural network dynamic load identification model is used to identify unknown dynamic loads. The input is the structural dynamic response measurement data, and the output is the load identification result.

2. The method according to claim 1, characterized in that, The neural network dynamic load recognition model adopts a fully connected neural network, which consists of an input layer, a hidden layer, and an output layer. The input layer contains structural dynamic response information, and the number of its neurons is only related to the number of input dynamic response features. The output layer represents the prediction of dynamic load by the fully connected neural network, and the number of its neurons is related to the load to be identified. The neurons in each hidden layer are connected to all neurons in the previous layer, thus forming a fully connected neural network dynamic load identification model.

3. The method according to claim 1, characterized in that, The training error between the predicted load information and the calibrated load information is calculated using the mean square error.

4. The method according to claim 1, characterized in that, The fitness of each particle is calculated using the following formula: ； in For particles The fitness of; For training error, For training time, This indicates the normalized training time.

5. The method according to claim 1, characterized in that, The particle update formula is as follows: ； ； in, and The first Generational particles The position and velocity of the particle are the hyperparameters, with the particle position representing the particle's iteration distance and direction. For the first The optimal particle for a given time period. For the first The globally optimal particle in time; Inertial weights; Represents the acceleration constant. for A random number between [a certain number of points].

6. The method according to claim 1, characterized in that, During the particle update process, once the neural network dynamic load recognition model for all particles in the particle population has been trained, it is considered that one iteration has been completed. Based on the difference in fitness values ​​of particles in all current iterations, the particle with the smallest fitness value is selected as the individual optimal particle; and based on the difference in fitness values ​​of all particles in all current iterations, the particle with the smallest fitness value is selected as the global optimal particle.

7. A neural network dynamic load identification system based on particle swarm optimization algorithm, characterized in that, include: The particle initialization module is used to initialize the number of hidden layers in the neural network dynamic load recognition model. and the number of neurons in each hidden layer As the optimization objective, construct particles Establish a particle swarm optimization algorithm and initialize the relevant parameters of the particle swarm optimization algorithm; The hyperparameter optimization module is used to build a corresponding neural network dynamic load recognition model for each particle, load a pre-prepared dynamic load recognition calibration training dataset to train the model, the model obtains predicted load information based on a set of input structural dynamic responses, calculates the training error between the predicted load information and the calibration load information, calculates the fitness of each particle based on the training error and training time, determines the individual optimal particle and the global optimal particle based on the fitness calculation results, and updates the particles at a preset number of iterations until the training of the last generation of particles ends. The model determination module is used to select the particle with the smallest fitness value in all iterations as the final hyperparameter combination of the neural network dynamic load recognition model, build the corresponding neural network dynamic load recognition model based on the final hyperparameter combination, and train it with the training dataset. The dynamic load identification module is used to identify unknown dynamic loads using a trained neural network dynamic load identification model. It takes structural dynamic response measurement data as input and outputs load identification results.

8. A computer device, characterized in that, include: One or more processors; Memory; And one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, wherein when the programs are executed by the processors, they implement the steps of neural network dynamic load identification based on particle swarm optimization algorithm as described in any one of claims 1-7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the neural network dynamic load identification method based on the particle swarm optimization algorithm as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the neural network dynamic load identification method based on the particle swarm optimization algorithm as described in any one of claims 1-7.

Images