Model training method, modal parameter identification method, electronic device and storage medium

Abstract

The invention discloses a model training method, a modal parameter identification method, a model training device, a modal parameter identification device, an electronic device and a nonvolatile computer readable storage medium. The model training method comprises the steps that a plurality of stability diagram samples are obtained, each stability diagram sample comprises a real mode and a false mode, the frequencies of stable points forming the real modes are the same, and the frequencies of at least part of the stable points and / or unstable points forming the false modes are different; and training to obtain a modal parameter identification model based on the plurality of stability map samples. It can be understood that in the training process, the training model can learn based on the real modalities in the multiple stability diagrams, and the real modalities of the multiple stability diagrams are different, so that the modal parameter recognition model obtained after training has high real modal recognition ability, and the modal parameter recognition efficiency is improved after the modal parameter recognition model is put into use. The modal parameter identification model can rapidly and accurately identify modal parameters.

Description

Technical Field

[0001] The present application relates to the technical field of building modal recognition, and more specifically, to a model training method, a modal parameter recognition method, a model training device, a modal parameter recognition device, an electronic device, and a non-volatile computer-readable storage medium. Background Art

[0002] Modes refer to the natural vibration characteristics of mechanical structures. Each mode has specific modal parameters, which can include natural frequency, damping ratio, and mode shape. Modal parameter analysis can be used to analyze the state of mechanical mechanisms and diagnose equipment faults. Therefore, accurately identifying the modal parameters of various buildings has become an urgent problem. Summary of the Invention

[0003] Embodiments of the present application provide a model training method, a modal parameter identification method, a model training device, a modal parameter identification device, an electronic device, and a non-volatile computer-readable storage medium.

[0004] The model training method of the embodiment of the present application is applied to training a model for identifying modal parameters, and the method includes: obtaining multiple stability graph samples, the stability graph samples including real modes and false modes, the stable points constituting the real modes having the same frequency, and at least some of the stable points and / or unstable points constituting the false modes having different frequencies; based on the multiple stability graph samples, training a modal parameter identification model.

[0005] The modal parameter identification method of the embodiment of the present application includes obtaining a stability diagram of a target object, the stability diagram including stable points and unstable points, and the stable points and the unstable points are generated based on vibration data of the target object; based on a preset modal parameter identification model, identifying the stable points of the stability diagram, and determining the true mode based on the stable points; and outputting the modal parameters of the target object based on the true mode.

[0006] The model training device according to an embodiment of the present application is used to train a model for identifying modal parameters. The device includes a first acquisition module and a training module. The first acquisition module is configured to acquire multiple stabilization diagram samples, wherein the stabilization diagram samples include true modes and false modes, wherein the stable points constituting the true modes have the same frequency, and at least some of the stable points and / or unstable points constituting the false modes have different frequencies. The training module is configured to train a modal parameter identification model based on the multiple stabilization diagram samples.

[0007] The modal parameter identification device according to an embodiment of the present application includes a second acquisition module, an identification module, and an output module. The second acquisition module is configured to acquire a stability diagram of a target object, wherein the stability diagram includes stable points and unstable points, and the stable points and unstable points are generated based on the vibration data of the target object. The identification module is configured to identify the stable points of the stability diagram based on a preset modal parameter identification model and determine the true mode based on the stable points. The output module is configured to output the modal parameters of the target object based on the true mode.

[0008] The electronic device of the embodiment of the present application includes a processor, a memory and a computer program, wherein the computer program is stored in the memory and executed by the processor, and the computer program includes instructions for executing a model training method or a modal parameter identification method. The model training method is applied to train a model for identifying modal parameters, and the method includes: obtaining a plurality of stability diagram samples, the stability diagram samples including real modes and false modes, the stable points constituting the real modes having the same frequency, and at least some of the stable points and / or unstable points constituting the false modes having different frequencies; based on the plurality of stability diagram samples, training a modal parameter identification model. The modal parameter identification method includes obtaining a stability diagram of a target object, the stability diagram including stable points and unstable points, the stable points and the unstable points being generated based on vibration data of the target object; identifying the stable points of the stability diagram based on a preset modal parameter identification model, and determining the real mode based on the stable points; and outputting the modal parameters of the target object based on the real mode.

[0009] The non-volatile computer-readable storage medium of the embodiment of the present application includes a computer program, and when the computer program is executed by a processor, the processor executes a model training method or a modal parameter identification method. The model training method is applied to a model for training and identifying modal parameters, and the method includes: obtaining a plurality of stability diagram samples, the stability diagram samples include real modes and false modes, the stable points constituting the real modes have the same frequency, and at least some of the stable points and / or unstable points constituting the false modes have different frequencies; based on the plurality of stability diagram samples, a modal parameter identification model is trained. The modal parameter identification method includes obtaining a stability diagram of a target object, the stability diagram includes stable points and unstable points, and the stable points and the unstable points are generated based on the vibration data of the target object; based on a preset modal parameter identification model, the stable points of the stability diagram are identified, and the real mode is determined based on the stable points; based on the real mode, the modal parameters of the target object are output.

[0010] The model training method, modal parameter identification method, model training device, modal parameter identification device, electronic device and non-volatile computer-readable storage medium of the embodiments of the present application can obtain multiple stability graph samples, wherein the stability graph samples include real modes and false modes, and each stability graph sample is different, so as to improve the sample diversity of the stability graph samples. Then, the training model can be trained based on the multiple stability graph samples to obtain a modal parameter identification model. It can be understood that during the training process, the training model can learn based on the real modes in the multiple stability graphs, and the real modes of the multiple stability graphs are different, so that the modal parameter identification model obtained after the training has a high real mode identification ability, so that after the modal parameter identification model is put into use, the modal parameter identification model can effectively remove the false modes in the stability graph and accurately calibrate the real modes, thereby enabling the modal parameter identification model to quickly and accurately identify the modal parameters.

[0011] Additional aspects and advantages of the embodiments of the present application will be given in part in the description below, and in part will become obvious from the description below, or will be learned through practice of the embodiments of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which:

[0013] Figure 1 is a flowchart of a model training method according to certain embodiments of the present application;

[0014] Figure 2 is a schematic diagram of a scenario of a model training method in certain embodiments of the present application;

[0015] Figure 3 is a flowchart of a model training method according to certain embodiments of the present application;

[0016] Figure 4 is a schematic diagram of a scenario of a model training method in certain embodiments of the present application;

[0017] Figure 5 Schematic diagram of the structure of a modal parameter identification model of a model training method in certain embodiments of the present application;

[0018] Figure 6 is a flowchart of a model training method according to certain embodiments of the present application;

[0019] Figure 7 is a flowchart of a model training method according to certain embodiments of the present application;

[0020] Figure 8is a flowchart of a model training method according to certain embodiments of the present application;

[0021] Figure 9 is a schematic diagram of a scenario of a model training method in certain embodiments of the present application;

[0022] Figure 10 is a flowchart of a model training method according to certain embodiments of the present application;

[0023] Figure 11 is a flowchart of a model training method according to certain embodiments of the present application;

[0024] Figure 12 is a flow chart of a modal parameter identification method according to certain embodiments of the present application;

[0025] Figure 13 is a schematic diagram of a module of a model training device according to certain embodiments of the present application;

[0026] Figure 14 is a module schematic diagram of a modal parameter identification device in certain embodiments of the present application;

[0027] Figure 15 is a schematic diagram of a module of an electronic device according to certain embodiments of the present application;

[0028] Figure 16 It is a schematic diagram of the connection status of a non-volatile computer-readable storage medium and a processor in certain embodiments of the present application. DETAILED DESCRIPTION

[0029] The embodiments of the present application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the embodiments of the present application, and should not be understood as limiting the embodiments of the present application.

[0030] The present invention provides a model training method and a modal parameter identification method. The model training method and modal parameter identification method of the present invention are described in detail below:

[0031] See also Figure 1 The present application provides a model training method, which is applied to train a model for identifying modal parameters. The method includes:

[0032] Step 01: Acquire multiple stability diagram samples, where the stability diagram samples include real modes and false modes, where the stable points constituting the real modes have the same frequency, and where at least some of the stable points and / or unstable points constituting the false modes have different frequencies;

[0033] Specifically, modal analysis is a method for studying the dynamic characteristics of structures, generally used in the field of engineering vibration. Modes refer to the natural vibration characteristics of mechanical structures. Each mode has a specific natural frequency, damping ratio, and modal shape, which are referred to as modal parameters. Modal analysis can determine the characteristics of each major mode within a certain susceptible frequency range, thereby predicting the actual vibration response of the structure to various external or internal vibration sources within this frequency range. Therefore, modal analysis is a key method for structural dynamic design and equipment fault diagnosis.

[0034] The stability diagram is an effective tool for eliminating false modes. The stability diagram can be generated based on the vibration data of the building. The vibration data may include the vibration amplitude, vibration velocity and vibration acceleration of each location of the building at different vibration frequencies. The vibration data can be obtained through sensors installed on the building. After obtaining the vibration data, it can be processed by the corresponding algorithm (covariance subspace method) to obtain the stability diagram. The points on the stability diagram usually have three kinds of information: frequency, damping ratio and modal vibration shape. They are obtained by setting different model orders and then calculating. Then, according to the stability judgment conditions, these poles are classified according to stability, and different icons and colors are used to distinguish stable points from unstable points. For example Figure 2 , Figure 2 The left figure is a stability diagram. The black solid circles in the stability diagram are stable points, and the gray hollow circles are unstable points. The horizontal axis is frequency, the horizontal axis on the left is order, and the horizontal axis on the right is amplitude.

[0035] The true mode will form a stable and clear stable band at its corresponding frequency, such as Figure 2 As shown on the left, as the assumed modal order changes, the poles representing the system's true modes align at the natural frequencies (vertical lines in the figure), while spurious modes do not always exist. True modes represent the actual physical modes of a structure or system. The multiple stable points that comprise these modes exhibit stability and consistency across multiple calculations or analyses of different orders. In contrast to true modes, spurious modes appear unstable and random in stability diagrams. They are typically caused by noise, computational errors, model inaccuracies, or external excitations that do not qualify as white noise. The location and parameters of spurious modes may change across multiple calculations or analyses of different orders.

[0036] In other words, only when there are multiple stable points in a certain frequency band can these multiple stable points be considered to represent the true mode. This can be understood as multiple stable points forming a vertical line in the stability diagram, which means that these multiple stable points can represent the true mode, and the frequency corresponding to this vertical line can represent the true frequency of the structure at a certain order. Among them, the same frequency of multiple stable points can be considered to mean that the frequencies of the multiple stable points are exactly the same, or the frequencies of the multiple stable points are roughly the same. In this case, a preset frequency difference range can be set. The preset frequency difference range is the maximum difference in the frequencies of the multiple stable points when multiple stable points can constitute the true mode. If the difference in the frequencies of multiple stable points is within the preset frequency difference range, the frequencies of the multiple stable points can be considered to be the same.

[0037] On the contrary, if there are unstable points in a certain frequency band, or the existing stable points are connected to stable points of other frequencies, that is, among the multiple stable points in the multiple connected data points in the stability graph, at least some of the stable points have different frequencies, or the multiple connected data points include unstable points, then these connected data points constitute false modes. As long as the data points include unstable points, regardless of whether the frequencies of the multiple unstable points are the same, and whether the frequencies corresponding to the unstable points and the stable points are the same, the multiple connected data points can be directly considered to constitute false modes. For example Figure 2 The right picture is the output image of the model, where the three vertical lines correspond to Figure 2 The three true modes in the left image.

[0038] Therefore, when performing modal analysis, it is necessary to identify the true modes in the stability diagram and determine the modal parameters of the structure based on the true modes. To improve the accuracy and speed of modal identification, modal identification can be performed by training a model to identify modal parameters (i.e., the modal parameter identification model described below). In this case, it is necessary to obtain multiple stability diagram samples, where each stability diagram sample includes true modes and false modes, and each stability diagram sample corresponds to different true modes and false modes, thereby ensuring sample diversity.

[0039] Step 02: Based on multiple stability diagram samples, a modal parameter identification model is trained.

[0040] Specifically, after obtaining multiple stabilization graph samples, a pre-set training model can be trained using these stabilization graph samples to obtain a modal parameter recognition model. For example, a label image corresponding to each stabilization graph sample can be obtained, and then the network parameters of the training model can be adjusted based on the output image and label image of the training model. By continuously adjusting the network parameters of the training model, a modal parameter recognition model can be obtained.

[0041] Each stability diagram sample corresponds to a different true mode and false mode, which ensures that the training model has enough samples to learn from, allowing it to capture more complex patterns and relationships. Therefore, the training model can learn how to accurately identify true modes during continuous training, making the modal parameter identification model obtained after training more capable of identifying true modes. This allows the modal parameter identification model to accurately and quickly identify true modes when real vibration data is subsequently input into the modal parameter identification model. At the same time, increasing sample diversity helps reduce the risk of overfitting and enables the model to perform better on new data. In this way, the subsequent modal parameter identification model can automatically identify the modal parameters of theoretical models and engineering structures with high recognition efficiency.

[0042] Existing modal identification methods identify structural modal parameters based on hierarchical clustering methods, but clustering methods have the same principle in terms of the identification process. They all need to set appropriate thresholds to achieve clustering, and existing methods basically use fixed thresholds. Fixed thresholds are usually obtained based on experience. In actual situations, different structures require different thresholds. Therefore, the modal parameters obtained by fixed threshold identification are usually not very accurate. In order to obtain accurate modal parameters, other methods must be sought. However, the present application uses multiple stability diagram samples to train the training model to obtain a modal parameter identification model that can accurately and quickly identify the true mode. After the subsequent modal parameter identification model is officially put into use, it does not require human participation, but automatically identifies the modal parameters. Therefore, compared to the existing technology using clustering methods, the present application can realize automatic identification of modal parameters without the need for human parameters, making the modal parameter identification accuracy of the present application higher.

[0043] The model training method of the embodiment of the present application can obtain multiple stable graph samples, which include real modes and false modes, and each stable graph sample is different, so as to improve the sample diversity of the stable graph samples. Then, the training model can be trained based on the multiple stable graph samples to obtain a modal parameter identification model. It can be understood that during the training process, the training model can learn based on the real modes in the multiple stable graphs, and the real modes of the multiple stable graphs are different, so that the modal parameter identification model obtained after the training has a high real mode identification ability, so that after the modal parameter identification model is put into use, the modal parameter identification model can effectively remove the false modes in the stable graph and accurately calibrate the real modes, thereby enabling the modal parameter identification model to quickly and accurately identify the modal parameters.

[0044] See also Figure 3 In some embodiments, step 01: obtaining a plurality of stabilization map samples comprises:

[0045] Step 011: Generate multiple stability diagram samples based on randomly generated sample generation parameters, where the sample generation parameters include at least parameters such as the true modal order, the proportion of stable points in the true mode, and the position of unstable points.

[0046] Specifically, the stability diagram samples can be automatically generated by setting the corresponding sample generation parameters. The sample generation parameters include at least parameters such as the true modal order, the proportion of stable points in the true mode, and the position of unstable points. The sample generation parameters can be automatically generated by setting the code, for example, by writing a loop python code to automatically set the sample generation parameters, that is, automatically generate the various values of the true modal order, the proportion of stable points in the true mode, and the position of unstable points. Then, the stability diagram samples are generated based on the randomly generated sample generation parameters. It can be understood that there will be at least a plurality of different sample generation parameters in the automatically generated sample generation parameters, so that at least some of the generated stability diagram samples are different. In this way, a plurality of stability diagram samples with random distribution of true and false modes can be automatically generated to improve the diversity of the samples and ensure that there are enough stability diagram samples to train the training model, so that the modal parameter identification model obtained after training can accurately identify the stability diagrams of various different structures and has a certain degree of versatility.

[0047] In addition, since stability diagram samples can be generated automatically, a sufficient number of stability diagram samples can be generated, for example, 10,000 stability diagram samples with random distribution of real and false modes, to ensure that the training model can be trained based on massive stability diagram samples, making up for the defect that it is difficult to obtain massive real stability diagrams to train the training model in real life, and ensuring that the modal parameter identification model obtained after training can accurately identify the stability diagrams of various different structures.

[0048] In other embodiments, the stability diagram samples may also be stability diagrams corresponding to real structures, so that the training model can be trained using the real stability diagram samples, ensuring that the modal parameter identification model obtained after training can accurately identify the stability diagrams of the real structure, thereby ensuring the accuracy of the modal parameter identification model in actual use. In still other embodiments, the stability diagram samples may include stability diagrams corresponding to real structures, and may also include automatically randomly generated stability diagrams. This ensures that the training model can be trained using real stability diagram samples, and that the number of samples in the training process is sufficient, so that the modal parameter identification model obtained after training can accurately identify the stability diagrams of a variety of different structures.

[0049] See also Figure 5In some embodiments, the modal parameter identification model includes an encoder 30 and a decoder 40, the encoder 30 is connected to the decoder 40, the encoder 30 includes a spatial pyramid module 31 and multiple feature extraction modules 32, the multiple feature extraction modules 32 are connected in sequence, the last feature extraction module 32 is connected to the spatial pyramid module 31, the feature extraction module 32 includes a feature extraction submodule 321 and a residual module 322, the feature extraction submodule 321 is connected to the residual module 322, the input end of the encoder 30 is connected to the first feature extraction module 32, and the output end of the spatial pyramid module 31 is connected to the input end of the decoder 40.

[0050] Specifically, the modal parameter identification model may include an encoder 30 and a decoder 40. Decoder 40 (Decoder) and encoder 30 (Encoder) are two concepts often encountered in the fields of digital signal processing and multimedia. The encoder 30 converts the original data into a compressed form to reduce the amount of data and improve transmission efficiency. It converts the original data into a specific coding format, such as audio coding (such as MP3, AAC) or video coding (such as H.264, HEVC) so that it takes up less space and bandwidth during storage or transmission. The decoder 40 performs the opposite operation and parses the encoded data into the original data form. The decoder 40 receives the encoded data stream and restores it to the original audio, video or image data for playback or further processing. The encoding formats between the encoder 30 and the decoder 40 must match, otherwise the decoder 40 cannot correctly decode the data generated by the encoder 30. The encoder 30 is connected to the decoder 40 so that under the joint action of the two, an output image corresponding to the stabilization map sample can be obtained.

[0051] The encoder 30 includes a spatial pyramid module 31 and multiple feature extraction modules 32. The multiple feature extraction modules 32 are connected in sequence, with the last feature extraction module 32 connected to the spatial pyramid module 31. The feature extraction module 32 includes a feature extraction submodule 321 and a residual module 322. The feature extraction submodule 321 is connected to the residual module 322. The input of the encoder 30 is connected to the first feature extraction module 32, and the output of the spatial pyramid module 31 is connected to the input of the decoder 40. That is, the stabilization map samples input to the encoder 30 first undergo feature extraction by the multiple feature extraction modules 32, and then extract features by the spatial pyramid module 31. The feature map output by the spatial pyramid module 31 is input to the decoder 40 for data reconstruction.

[0052] The modal parameter identification model consists of convolutional layers, pooling layers, activation layers, and batch normalization layers. The convolutional layer, composed of learnable convolution kernels (variable weight matrices), is used to extract low-dimensional feature information from the data and construct high-dimensional features. The pooling layer compresses the input data, reducing the number of parameters and, in turn, the data dimension to prevent overfitting. The activation layer determines whether the signal is passed to the next layer, which helps improve the representation and training capabilities of the neural network. The batch normalization layer (BN) ensures that the mean of each dimension of the output signal is 0 and the variance is 1 during batch stochastic gradient descent training, increasing the model's capacity.

[0053] The feature extraction submodule 321 is composed of a convolution layer, a pooling layer, an activation layer, and a batch normalization layer. The residual module 322 is composed of a convolution layer and a BN layer. In particular, the convolution kernel size is 1×1, that is, an identity mapping is established between the input and output of the layer, so that information can be directly transmitted. The residual module 322 can effectively avoid phenomena such as gradient explosion and gradient disappearance in the deep learning training process. A feature extraction submodule 321 and a residual module 322 are connected to form a feature extraction module 32. The stabilization map sample input to the encoder 30 will be input into the feature extraction submodule 321 and the residual module 322 of the feature extraction module 32. Both the feature extraction submodule 321 and the residual module 322 can extract features from the image and output the corresponding feature map. Then, the output feature maps of the residual module 322 and the feature extraction submodule 321 in the same feature extraction module 32 are fused and passed to the next feature extraction module 32 for deeper feature extraction. In this process, the size of the feature map will gradually decrease and the number of channels will gradually increase. Figure 5 In the figure shown, there are four feature extraction modules 32. The stabilization map sample input to the encoder 30 is a color stabilization map of size 240×320×3. The first feature extraction module 32 outputs a feature map of size 240×320×16. The second feature extraction module 32 outputs a feature map of size 120×160×32. The third feature extraction module 32 outputs a feature map of size 60×80×64. The final feature extraction module 32 outputs a feature map of size 30×40×128. The number of feature extraction modules 32 can be multiple, and the specific number can be determined according to needs. For example, as shown in the figure, the number of feature extraction modules 32 is four. In this case, the modal parameter identification model can perform four convolution pooling operations on the stabilization map samples.

[0054] The last feature extraction module 32 is connected to the spatial pyramid module 31, so the feature map output by the last feature extraction module 32 can be input into the spatial pyramid module 31. The feature map output by the spatial pyramid module 31 is then input into the decoder 40, which can reconstruct features based on the feature map and generate an output image.

[0055] The spatial pyramid module 31 proposes a convolutional approach to address the problem of downsampling (pooling) during feature extraction, which reduces image resolution and causes information loss. By setting different convolution dilation rates and using convolution kernels of different scales to extract features across pixels, it obtains information of different dimensions. For example, assuming there are four feature extraction modules 32, the feature map size is 30×40 after four convolution and pooling operations. The convolution dilation rates are set to 1, 2, 3, and 4, respectively, to perform feature extraction of four different dimensions. The four output feature maps are then fused (concatenated), and the channels are adjusted through a single convolution to obtain a 15×20×256 feature map. This is then fed into the decoder 40 simultaneously with the 15×20×256 feature map obtained after conventional convolution for data reconstruction and feature restoration. The former is upsampled by a factor of 4, while the latter undergoes two convolution upsampling operations, and the two feature maps are then fused and the number of channels is adjusted. Then, the two are upsampled twice, and the number of channels is adjusted to 1, resulting in an output image of 240×320×1.

[0056] In this way, a feature extraction module 32 and a spatial pyramid module 31 can be set in the encoder 30, wherein the feature extraction module 32 includes a feature extraction sub-module 321 and a residual module 322, so that the encoder 30 can accurately extract features from the stable map samples, so that the subsequent decoder 40 can accurately reconstruct features based on the feature map output by the encoder 30, thereby helping the decoder 40 to output an output image with higher accuracy.

[0057] See also Figure 5 In some embodiments, the decoder 40 includes multiple decoding modules 41, and the multiple decoding modules 41 are connected in sequence. The output end of the spatial pyramid module 31 is connected to the input end of the first decoding module 41, and the input end of the target decoding module 41 is connected to the output end of the corresponding feature extraction module 32. The target decoding module 41 has the same feature map size as the connected feature extraction module 32. The target decoding module 41 is a module among the multiple decoding modules 41, and the size of the feature map is the same as the feature map size of any feature extraction module 32 in the encoder 30.

[0058] Specifically, the encoder 30 includes multiple decoding modules 41, which are connected sequentially, and the feature maps corresponding to the multiple decoders 41 have different sizes. The output of the pyramid module is connected to the input of the first decoding module 41. Therefore, the feature map output by the spatial pyramid module 31 can be input to the first decoding module 41. After the first decoding module 41 processes the feature map output by the spatial pyramid module 31, it can generate an output image. The output image is then input to the next decoding module 41, and so on until each decoding module 41 has completed data reconstruction and feature restoration. The output image generated by the last decoding module 41 is the output image of the training module.

[0059] At this time, a cascade structure can also be designed, that is, the input end of the decoding module 41 is connected to the output end of the feature extraction module 32 with the same feature map size. At this time, the feature map size corresponding to each decoding module 41 and the feature map size corresponding to each feature extraction module 32 can be obtained. If the feature map size of a decoding module 41 is the same as the feature map size corresponding to a feature extraction module 32, then the decoding module 41 is identified as the target decoding module 41 and connected to the feature extraction module 32 with the same feature map size. For example Figure 2 The connection relationship shown by the three arrows at the top is Figure 2 The feature extraction module 32 and the target decoding module 41 are connected with a feature map size of 120×160×32, the feature extraction module 32 and the target decoding module 41 are connected with a feature map size of 60×80×64, and the feature extraction module 32 and the target decoding module 41 are connected with a feature map size of 30×40×128.

[0060] The purpose of designing a cascade structure is to pass the feature map of corresponding size in the feature extraction module 32 into the decoding module 41 for feature superposition (Add) operation. When the decoding module 41 reconstructs the features, the prior knowledge obtained by the feature extraction module 32 is added to help the decoding module 41 restore the output image.

[0061] In this way, when the decoder 40 performs data reconstruction and feature restoration, the accuracy of the output image output by the decoder 40 can be improved through the cascade structure.

[0062] Furthermore, a modal parameter identification model can be constructed based on a suitable framework and corresponding modules can be added to the framework. For example, the framework can be TensorFlow, PyTorch, or CNTK. For example, a lightweight encoder-decoder deep learning network including a cascade structure, a residual module 322, and a spatial pyramid module 31 can be constructed based on the TensorFlow framework.

[0063] During training, a training model based on the above encoder and decoder can be constructed. After training, a modal parameter recognition model with the structure as described above and relatively appropriate network parameters can be obtained.

[0064] In summary, the present invention constructs a lightweight encoder-decoder deep learning network (i.e., a training model and a modal parameter identification model) including a cascade structure, a residual module 322, and a spatial pyramid module 31, so as to improve the recognition accuracy of the modal parameter identification model for stable graph samples.

[0065] See also Figure 4 and Figure 6 In some embodiments, step 02: training a modal parameter identification model based on multiple stability diagram samples includes:

[0066] Step 021: Generate a label image based on the true mode in the stabilization map sample;

[0067] Step 022: Generate a training dataset based on the labeled image and the stabilization map samples, and divide the training dataset into a training set, a test set, and a validation set;

[0068] Step 023: Train the preset training model based on the training set, test set and validation set to obtain a modal parameter identification model.

[0069] Specifically, if the stabilization image sample is a real image, then the true mode in the middle is known, so the label image can be generated based on the true mode in the stabilization image sample. If the stabilization image sample is an automatically generated random sample, then the true mode in the middle is known, so the label image can be generated based on the true mode in the stabilization image sample. In this case, the pixel values of the stable points corresponding to the true mode in the stabilization image sample can be adjusted to 1, and the pixel values of all other parts can be adjusted to 0 to generate the label image.

[0070] Stability graph samples can be divided into test sets, training sets, and validation sets to facilitate the construction of more robust models with enhanced generalization capabilities. The training set is the dataset used to train machine learning models. During training, the trained model uses samples from the training set to learn features and patterns to make predictions or classifications. The validation set is the dataset used to adjust the trained model's hyperparameters and evaluate its performance. During training, the validation set is used to adjust the model's parameters to prevent the model from overfitting to the training set. The performance of the validation set helps select optimal model parameters. The test set is the dataset used to evaluate the performance of the trained model. After the model is trained, the test set is used to verify the model's ability to generalize to unseen data. The performance of the trained model on the test set helps evaluate the model's accuracy and performance.

[0071] In this case, a stabilization map sample and its corresponding labeled image are combined into a pair of sample images. Multiple pairs of sample images form a training dataset, where the labeled image is an image that only includes the real modal image from the corresponding stabilization map sample. Within the training dataset, a test set, training set, and validation set can be generated according to a preset ratio or number. For example, 1% of the sample images from the training dataset can be selected as the test set, and the remaining images can be divided into the training set and validation set in an 8:2 ratio. Alternatively, the number of training datasets can be fixed: 100 pairs of sample images can be selected as the test set, 1000 pairs as the training set, and 200 pairs as the validation set.

[0072] In this way, corresponding labeled images can be generated based on the stabilization map samples. A training dataset can then be generated based on the labeled images and stabilization map samples, ensuring that the training dataset contains the sample data required for model training. The training dataset can then be divided into training, test, and validation sets, and the model trained based on these sets to ensure that a modal parameter recognition model with strong recognition capabilities and high recognition accuracy is ultimately obtained.

[0073] See also Figure 4 and Figure 7 In some embodiments, step 023: training a preset training model based on the training set, the test set, and the validation set to obtain a modal parameter identification model includes:

[0074] Step 0231: Training the training model based on the training set and the validation set to determine an intermediate model that meets a first preset condition from the training models corresponding to the multiple trainings. The first preset condition at least includes that the loss value obtained after the training model processes the validation set is the smallest;

[0075] Step 0232: Verify the intermediate model based on the test set. If the intermediate model meets the second preset condition, determine that the intermediate model is a modal parameter identification model. The second preset condition at least includes that the loss value corresponding to the intermediate model is less than the preset loss value threshold.

[0076] Specifically, during the training process, the training set can be imported into the training model for training, and parameters can be adjusted based on the training results of the training set. Simultaneously, the validation set can be used for reverse parameter adjustment during the training process. For example, by inputting stabilization map samples from the validation set into the training model and comparing the output images obtained at this time with the corresponding labeled images, the hyperparameters of the training model can be adjusted based on the comparison results. Furthermore, the output results of the validation set can be used to evaluate the recognition ability of the training model at each iteration, thereby facilitating the subsequent screening of suitable intermediate models.

[0077] After training based on the training set and validation set, an intermediate model that meets the first preset condition can be selected from the trained models from multiple iterations. For example, the first preset condition includes at least the minimum loss value obtained by the trained model after processing the validation set. In this case, the validation set can be used to calculate the loss value corresponding to each iteration of the trained model, and the trained model with the minimum loss value can be determined as the intermediate model. It can be understood that the intermediate model has been evaluated on the validation set, so the recognition accuracy of the intermediate model is also higher.

[0078] After training, the intermediate model can be verified based on the test set. If the intermediate model meets the second preset condition, the trained model that meets the second preset condition is determined to be a modal parameter recognition model. The second preset condition at least includes that the loss value corresponding to the intermediate model is less than a preset loss value threshold. The preset loss value threshold is the maximum loss value when the trained model can accurately identify the true mode.

[0079] When verifying the intermediate model based on the test set, the stabilization graph sample of the test set can be input into the intermediate model to obtain a test output sample. Then, the loss value is determined based on the test output sample and its corresponding label image, as well as the loss value function. Among them, the loss value function includes Mean Squared Error (MSE), Cross-Entropy Loss or Cosine Similarity Loss. When the loss value is less than or equal to the preset loss value threshold, it can be considered that the recognition ability of the intermediate model is better. At this time, the training can be stopped, and the modal parameter recognition model is generated based on the intermediate model that meets the second preset condition.

[0080] If the intermediate model does not meet the second preset condition, the parameters of the training model are readjusted, and the training model is trained again based on the training set and the validation set until an intermediate model that meets the second preset condition can be obtained.

[0081] In this way, the training model can be trained using the training set, test set, and validation set respectively. During model development and iteration, the validation and test sets help determine whether further adjustments or improvements are needed. By continuously evaluating model performance on the validation set, the training model can be gradually optimized, and the improved results can ultimately be verified on the test set, resulting in a modal parameter identification model with strong recognition capabilities and high recognition accuracy.

[0082] See also Figure 4 and Figure 8 In some embodiments, step 0231: training the training model based on the training set and the validation set includes:

[0083] Step 02311: Process the training set based on the training model to obtain a loss value corresponding to the training set;

[0084] Step 02312: Process the validation set based on the training model to obtain an evaluation score corresponding to the validation set;

[0085] Step 02313: When the loss value is greater than a preset loss value threshold, adjust the network parameters of the training model according to the learning rate, where the learning rate is adjusted based on the number of training times and / or the performance index of the training model, and the performance index is determined based on the loss value and / or the evaluation score;

[0086] Step 02314: Based on the training model after parameter adjustment, enter the training model again to process the training set to obtain the loss value corresponding to the training set until the training is completed;

[0087] Step 02315: Based on the first preset condition, determine an intermediate model among the training models corresponding to the multiple trainings.

[0088] Specifically, the network parameters of the training model need to be adjusted continuously during network training. Figure 9 The weights W and bias b are network parameters. These parameters are not set in advance but are derived from the training data by the model. During network training, they are continuously updated based on the input data. Hyperparameters are manually set before training begins, such as the size and number of convolution kernels, network architecture, and learning rate. The network model can be adjusted based on the training and validation sets.

[0089] The training set's stabilization graph samples are fed into the training model. After processing the stabilization graph samples, the training model outputs a training output image. This training output image includes the true modality recognized by the training model at that point. The loss value is then calculated based on the training output image and its corresponding labeled image.

[0090] The stabilization graph samples from the validation set are then fed into the training model. After processing the stabilization graph samples, the training model outputs a validation output image. This validation output image, understood as including the true modality recognized by the training model at that point, can then be used to calculate an evaluation score based on the validation output image and its corresponding label image. For example, the output image can be evaluated using the coefficient of determination (R2) and the label image to obtain the corresponding evaluation score. A larger R2 indicates a stronger recognition capability of the trained model.

[0091] When the loss value is less than or equal to the preset loss value threshold, it can be considered that the recognition ability of the training model is better. At this time, training can be stopped and a modal parameter recognition model can be generated based on the training model that meets the second preset condition.

[0092] If the loss value exceeds the preset loss threshold, the trained model's recognition capability is considered insufficient, and the network parameters of the trained model still need to be adjusted. Therefore, the network parameters of the trained model can be adjusted based on the current learning rate to continuously optimize the error value, allowing the trained model to train to the optimal solution. For example, the mean square error (MSE) between the output image and the labeled image can be calculated. This is done by calculating the average square of the difference between the predicted value y' and the actual value y. The mean square error can then be continuously optimized to train the network to the optimal solution.

[0093] A learning rate decay strategy can be used to adjust the network parameters of the training model. During parameter adjustment, an optimizer algorithm can be employed. The optimizer has two key parameters: the gradient and the learning rate. The gradient determines the direction of the parameter update, while the learning rate determines the extent of the parameter update. For example, in one embodiment, the Adam optimization algorithm is selected as the optimizer to comprehensively consider the first- and second-order moment estimates of the gradient, thereby enabling adjustment of the learning rate for each parameter.

[0094] The loss value is a metric that measures the difference between the model's predicted results and the actual results. The learning rate determines the step size of the model's parameter updates in each iteration. An appropriate learning rate can significantly improve model training speed. In the early stages of training, a larger learning rate allows the model parameters to quickly approach the optimal value, thereby accelerating convergence. As training progresses, the loss value gradually decreases. At this time, appropriately reducing the learning rate can help the model adjust parameters more finely, further reducing the loss value and improving training efficiency. Therefore, the learning rate can be adjusted during training to achieve an appropriate value to improve training efficiency.

[0095] The learning rate can be adjusted based on the number of training runs or the performance index of the trained model. The performance index is determined based on the loss value and / or the evaluation score. For example, the performance index can be determined based only on the loss value, only on the evaluation score, or based on both the loss value and the evaluation score. A larger loss value results in a smaller performance index, and a larger evaluation score results in a larger performance index.

[0096] In one embodiment, the learning rate is adjusted only according to the number of trainings. In this case, a training number threshold is set. The training number threshold is the maximum number of trainings when the current learning rate can well adjust the network parameters of the training model so that the performance index of the training model gradually decreases. In the case where the number of trainings is greater than the training number threshold, it can be considered that the current learning rate is difficult to gradually reduce the performance index of the training model, so the learning rate can be reduced. In another embodiment, the learning rate is adjusted only according to the performance index of the training model. If the performance index of the training model does not improve within the preset time, the learning rate is reduced. In another embodiment, a first preset number threshold can be set. The first preset number threshold is the maximum number of trainings when the current learning rate can well adjust the network parameters of the training model so that the performance index of the training model gradually decreases. In the case where the loss value is greater than the preset loss value threshold and the number of trainings reaches the first preset number threshold, if the performance index of the training model does not increase during the training process corresponding to the first preset number threshold, it can be considered that the current learning rate is difficult to gradually reduce the performance index of the training model, so the learning rate can be reduced. For example, the first preset number threshold is 15 times, and the initial learning rate can be set to 10 -5 After 15 iterations, if the network model performance has not improved, the learning rate is multiplied by a factor of 0.5. At this time, you can also set a minimum learning rate, for example, the minimum learning rate is limited to 10 -6 , the minimum learning rate can only be reduced to the minimum learning rate to reduce the risk of overfitting.

[0097] In this way, the learning rate decay strategy can be used to smooth the training process, that is, the learning rate is adjusted during the training process, so as to improve the efficiency of model training and ensure that a modal parameter recognition model with high recognition ability can be quickly trained.

[0098] Then, based on the training model with adjusted parameters, step 02311 is re-entered to enter the next training cycle to train again until the training is completed. The condition for the completion of training includes at least that the loss value obtained by the training model after processing the validation set is less than a preset loss value threshold.

[0099] Alternatively, training can terminate when the weight parameters (.h5) and network training metrics (CSV) meet certain requirements. Weight parameters (.h5) are core components of a neural network model, determining how the model extracts information from input features and generates predictions. Their values (or changes) can indirectly reflect the model's training effectiveness. Network training metrics are a series of performance metrics recorded during the training process, such as loss function value, precision, recall, and F1 score. These metrics directly reflect the model's performance on the training and validation sets. Each iteration saves a file containing the weight parameters (.h5) and network training metrics (CSV) corresponding to the current iteration. The performance of the trained model can then be evaluated based on the weight parameters (.h5) and network training metrics (CSV). For example, a table can be plotted, or predictions can be made on data outside the training set based on the weight parameters and network training metrics to generate performance graphs. The results can then be evaluated based on the table or graphs. When the training model meets the requirements, training is terminated. After training, the trained model for each iteration can be saved. Then, an intermediate model that meets the first preset condition can be selected from the training models corresponding to each iteration.

[0100] In this way, the training model can be trained multiple times based on the training set and the validation set, and the training model that meets the first preset condition is determined as the intermediate model, thereby ensuring that a modal parameter recognition model with higher recognition ability can be obtained after the training is completed.

[0101] See also Figure 4 In some embodiments, the condition for the end of training also includes that the loss value does not decrease when the number of iterations reaches a second preset number threshold or the number of iterations reaches a third preset number threshold.

[0102] Specifically, the second preset number threshold is the maximum number of iterations, representing the maximum number of times the dataset will be completely traversed and learned during the entire training process. The second preset number threshold can be set manually or based on empirical values. When the number of iterations reaches the second preset number threshold, the dataset can be considered to have been completely learned, and training can be terminated at this point.

[0103] Alternatively, an early stopping mechanism (i.e. Figure 4The early stopping algorithm in (in) determines the time to end training. The third preset number threshold represents the maximum number of training times that the training process can continue when the performance has not improved. Therefore, when the number of iterations reaches the third preset number threshold, if the performance indicators of the training model have not improved in the multiple training processes corresponding to the third preset number threshold, for example, the loss value (or loss function curve) has not decreased, or the evaluation score has not improved, the model training is stopped. For example, if the third preset number threshold is 30 times, after 30 iterations, if the loss function curve of the validation set has not decreased, the network training is stopped.

[0104] There may be multiple training end conditions. In actual use, the training end condition is selected as needed, and then the model training process is set according to the selected training end condition. The training end time will be determined based on the selected training end condition later.

[0105] In this way, you can set different training end conditions to ensure that training ends under appropriate circumstances and ensure that the model can achieve the best training results. For example, you can use the early stopping mechanism to smooth the training process and thus improve training stability.

[0106] See also Figure 4 In some embodiments, the first preset condition is that the evaluation score of the training model is the highest.

[0107] Specifically, the training model with the smallest loss value may not have the strongest modal parameter recognition capability. Therefore, an evaluation score can be introduced as an evaluation criterion to test the training models. It can be understood that the training model with the highest evaluation score is the one with the strongest parameter recognition capability. Therefore, an intermediate model can be determined based on the training model with the highest evaluation score. Specifically, the training model with the best reconstructed image performance can be determined as the intermediate model to ensure high recognition capability of the subsequent modal parameter recognition model.

[0108] In this way, the evaluation score can be used to evaluate the modal parameter recognition capability of the training model of each iteration, and the training model with the highest evaluation score can be determined as the intermediate model, thereby ensuring that the modal parameter recognition capability of the modal parameter recognition model that is subsequently put into formal use is strong.

[0109] It should be noted that the first preset condition may include multiple conditions. In actual use, the first preset condition is selected as needed, and then the model training process is set according to the selected first preset condition. Subsequently, the training model will be trained based on the selected first preset condition.

[0110] See also Figure 4 and Figure 10 In some embodiments, step 0232: determining that the training model that meets the second preset condition is a modal parameter identification model includes:

[0111] Step 02321: Input the stable map sample of the test set into the intermediate model to obtain the test output image;

[0112] Step 02322: Determine an evaluation value corresponding to the intermediate model based on the image similarity measurement index, the test output image, and the corresponding label image;

[0113] Step 02323: when the evaluation value is greater than or equal to the preset evaluation value threshold, determining that the intermediate model is a modal parameter identification model, the second preset condition further includes that the evaluation value is greater than or equal to the preset evaluation value threshold;

[0114] Model training methods also include:

[0115] Step 03: When the evaluation value is less than the preset evaluation value threshold, the parameters of the training model are readjusted, and the training model is trained again based on the training set and the validation set to obtain an intermediate model that meets the first preset condition.

[0116] Specifically, the image similarity measurement index measures similarity based on three dimensions: luminance l(x,y), contrast c(x,y), and structure s(x,y). Therefore, the evaluation value tested by the image similarity measurement index can be more consistent with the evaluation of the human eye.

[0117] Therefore, the training model that satisfies the first pre-set condition can be first determined as the intermediate model. The stabilization map samples from the test set are then input into the intermediate model to obtain a test output image. An evaluation value corresponding to the intermediate model is then determined using an image similarity measurement metric, the test output image, and the corresponding labeled image, thereby further evaluating the intermediate model. This evaluation of the intermediate model can be combined with the evaluation of staff members, who can evaluate the intermediate model based on their relevant knowledge.

[0118] The preset evaluation value threshold is the minimum value of the evaluation value when the ability of the intermediate model to identify the real mode is relatively consistent with the recognition ability of the human eye. Therefore, only when the evaluation value is greater than or equal to the preset evaluation value threshold, the ability of the intermediate model to identify the real mode is considered to be relatively consistent with the recognition ability of the human eye, and the intermediate model is determined to be a modal parameter recognition model. If the evaluation value is less than the preset evaluation value threshold, it is considered that the ability of the intermediate model to identify the real mode is still insufficient, and the parameters of the training model are readjusted, and the training model is trained based on the training set and the validation set again to obtain a step (i.e., step 0231) of an intermediate model that meets the first preset condition, that is, when the evaluation value is less than the preset evaluation value threshold, the parameters of the training model are readjusted, and training is performed again until an intermediate model with an evaluation value greater than or equal to the preset evaluation value threshold is obtained. That is, the second preset condition also includes an evaluation value greater than or equal to the preset evaluation value threshold.

[0119] Therefore, after training is complete, the recognition performance of the trained intermediate model can be further evaluated using image similarity metrics to determine whether the trained intermediate model meets the required recognition performance. If so, it is considered a modal parameter recognition model and applied to the stability diagram corresponding to the actual scene for true modal recognition, thereby ensuring the recognition capability of the modal parameter recognition model. If not, the network architecture and hyperparameters are adjusted until it meets the requirements of the engineering application.

[0120] It should be noted that the second preset condition may include multiple types. In actual use, the second preset condition is selected as needed, and then the model training process is set according to the selected second preset condition. Subsequently, the training model will be trained based on the selected second preset condition.

[0121] See also Figure 4 and Figure 11 In some embodiments, before step 02: training a modal parameter identification model based on a plurality of stability diagram samples, the model training method further includes:

[0122] Step 04: Preprocessing the stabilized map samples to update the stabilized map samples. The preprocessing includes adjusting the number of channels of the plurality of stabilized map samples to a preset number of channels, adjusting the resolution of the plurality of stabilized map samples to a preset resolution, and normalizing at least one of the pixel values of the stabilized map samples.

[0123] Specifically, before using the stabilized image samples to train the training model, the stabilized image samples can also be preprocessed. The preprocessing includes adjusting the number of channels of the plurality of stabilized image samples to a preset number of channels, adjusting the resolution of the plurality of stabilized image samples to a preset resolution, and normalizing at least one of the pixel values of the stabilized image samples. All of these preprocessing steps can be performed using OPEN CV. Open CV, short for Open Source Computer Vision Library, is a cross-platform computer vision and machine learning software library released under the Apache 2.0 license (open source).

[0124] For example, you can use OpenCV to adjust the number of channels of multiple stabilized image samples and convert them into 3-channel RGB images. Then, change the resolution of the stabilized image samples to a uniform size. Finally, use a normalization algorithm to reduce the RGB value range of each stabilized image sample to the range of 0-1.

[0125] By standardizing the number of channels and resolution, we can simplify the image preprocessing steps for subsequent training models, reducing data conversion and adaptation time, thereby improving the data processing efficiency of the training models. Furthermore, since the input data format and dimensions are consistent, the training models can process data more efficiently during training, reducing the waste of computing resources and thus accelerating the training process. Reducing the RGB value range of the stabilized image can also help reduce computing resources.

[0126] In addition, during preprocessing, the stable map samples can be converted into numpy arrays to improve the efficiency of subsequent training models in processing stable map samples.

[0127] In order to better illustrate the effect, the modal parameter identification model trained according to the above method is tested by taking the modal parameter identification of a bridge as an example. The results are as follows: Figure 2 shown. Figure 2 The data is the bridge data, the upper left corner is the stability diagram, and the upper right corner is the output image of the modal parameter identification model. The output image includes the real mode identified by the modal parameter identification model. The horizontal axis of the real mode is the frequency of the bridge at each order. The specific results are as follows Figure 2As shown in the table, the monitoring system in the table is the actual frequency of the bridge, and the data corresponding to deep learning is the frequency output by the modal parameter identification model of this application. The numbers in the table represent the frequency of the bridge at each order. It can be found that after identification by the modal parameter identification model, the false modes in the stability diagram are effectively removed, and the real modes are accurately calibrated in the image. The frequency corresponding to the real mode is obtained by pixel conversion, and the error of the result is within 1%. The points on the stability diagram usually have three kinds of information: frequency, damping ratio and modal vibration shape. Therefore, after obtaining the frequency, the damping ratio and modal vibration shape of the bridge can be calculated according to the corresponding algorithm to obtain the modal parameters of the bridge. Among them, when identifying the bridge mode, most of the time it is only necessary to identify the first two frequencies. Therefore, although the output image has three straight lines, the table only has two frequencies.

[0128] Therefore, the modal parameter identification model trained by the model training method of the present application has a high recognition accuracy, a small positioning frequency error, a high degree of automation, and a fast separation speed, achieving the requirements of de-manualization, strong real-time performance, and high precision of structural modal parameter identification, which is beneficial to the analysis of data by structural (such as building) health monitoring personnel.

[0129] See also Figure 12 , the modal parameter identification method of the embodiment of the present application includes:

[0130] Step 05: Obtain a stability map of the target object. The stability map includes stable points and unstable points, which are generated based on the vibration data of the target object.

[0131] Step 06: Based on the preset modal parameter identification model, identify the stable point of the stability diagram and determine the true mode based on the stable point;

[0132] Step 07: Based on the true mode, output the modal parameters of the target object.

[0133] Specifically, during modal parameter identification, vibration data of the target object can be obtained. This data can be acquired using sensors mounted on the target object. A stability map corresponding to the target object can then be generated based on the vibration data. The stability map includes stable and unstable points. For example, the stability map can be generated using the covariance subspace method and the vibration data.

[0134] The stabilization diagram is then input into a pre-trained modal parameter identification model. The modal parameter identification model can identify the stabilization points of the stabilization diagram and determine the true mode based on the stabilization points with the same frequency, thereby outputting the modal parameters of the target object based on the true mode. In certain embodiments, the modal parameter identification model can be obtained by obtaining multiple stabilization diagram samples and training based on the multiple stabilization diagram samples. The stabilization diagram samples include true modes and false modes, where the stabilization points constituting the true mode have the same frequency, and at least some of the stabilization points and / or unstable points constituting the false mode have different frequencies. That is, the modal parameter identification model used in the modal parameter identification method can be obtained based on steps 01-04 above. For the sake of brevity, the training process of the modal parameter identification model will not be further described here.

[0135] The generated initial stable map includes pixels corresponding to stable points and unstable points. Subsequently, it is necessary to convert pixels into corresponding frequencies based on a certain algorithm and the position of pixels to generate the final version of the stable map, such as Figure 2 Left image. The modal parameter identification model can determine whether there are pixels with the same horizontal coordinate among multiple stable points based on the position of the pixels corresponding to each stable point. If there are multiple pixels with the same horizontal coordinate, it can be considered that the stable points corresponding to these multiple pixels constitute the true mode. Then, the corresponding algorithm can be used to convert the pixels corresponding to the true mode into frequency. At this time, the modal parameter identification model can generate an output image based on the true mode and its frequency, such as Figure 2 Finally, the damping ratio and modal vibration shape of the target object can be obtained based on the frequency, thereby obtaining the modal parameters of the target object.

[0136] In this way, the present application can use the trained and best-performing network (i.e., the modal parameter identification model) to locate the true mode and remove the false mode of the target object, so as to automatically and accurately identify and locate the true mode in the stability diagram in real time and filter out the false mode. At the same time, the present application can also obtain the frequency corresponding to each mode of the target object through pixel index conversion. After the network training is completed, there is no need to manually select thresholds and adjust parameters, thereby realizing automatic recognition of modal parameters, making the recognition accuracy of modal parameters higher and the recognition speed faster.

[0137] The modal parameter identification method of the present application embodiment can obtain a stabilization diagram of a target object and input the stabilization diagram into a preset modal parameter identification model, which is a network with strong recognition capabilities. Therefore, the modal parameter identification model can accurately locate the true mode in the stabilization diagram and accurately identify the modal parameters of the target object based on the true mode, thereby facilitating personnel to analyze the target object.

[0138] See also Figure 13In order to facilitate better implementation of the model training method of the embodiment of the present application, the embodiment of the present application also provides a model training device 10. The model training device 10 is used to train a model for identifying modal parameters. The model training device 10 may include a first acquisition module 11 and a training module 12. The first acquisition module 11 is used to obtain a plurality of stable graph samples, the stable graph samples include real modes and false modes, the stable points constituting the real modes have the same frequency, and the stable points and / or unstable points constituting the false modes have different frequencies; the training module 12 is used to train a modal parameter identification model based on the plurality of stable graph samples.

[0139] The first acquisition module 11 is specifically used to generate a plurality of stability diagram samples based on randomly generated sample generation parameters, where the sample generation parameters include at least parameters such as the true modal order, the proportion of stable points in the true mode, and the position of unstable points.

[0140] The training module 12 is specifically used to generate a label image based on the real mode in the stable map sample; generate a training data set based on the label image and the stable map sample, and divide the training data set into a training set, a test set and a validation set; train a preset training model based on the training set, the test set and the validation set to obtain a modal parameter recognition model.

[0141] The training module 12 is specifically used to train the training model based on the training set and the validation set to obtain an intermediate model that meets the first preset condition, and the first preset condition at least includes that the loss value obtained after the training model processes the validation set is the smallest; verify the intermediate model based on the test set, and when the intermediate model meets the second preset condition, determine that the intermediate model is a modal parameter identification model, and the second preset condition at least includes that the loss value corresponding to the intermediate model is less than the preset loss value threshold.

[0142] The training module 12 is specifically used to process the training set based on the training model to obtain the loss value corresponding to the training set; process the validation set based on the training model to obtain the evaluation score corresponding to the validation set; when the loss value is greater than the preset loss value threshold, adjust the network parameters of the training model according to the learning rate, and the learning rate is adjusted based on the number of training times and / or the performance indicators of the training model, and the performance indicators are determined according to the loss value and / or the evaluation score; based on the training model after parameter adjustment, enter the training model again to process the training set to obtain the loss value corresponding to the training set, until the training is completed, and the condition for the end of training at least includes that the loss value obtained after the training model processes the validation set is less than the preset loss value threshold; based on the first preset condition, determine the intermediate model in the training models corresponding to multiple trainings.

[0143] The training module 12 is specifically used to reduce the learning rate when the loss value is greater than the preset loss value threshold and the number of training times reaches the first preset number threshold, if the performance index of the training model does not increase during the training process corresponding to the first preset number threshold.

[0144] The training module 12 is specifically used to input the stability map samples of the test set into the intermediate model to obtain a test output image; determine the evaluation value corresponding to the intermediate model based on the image similarity measurement index, the test output image and the corresponding label image; when the evaluation value is greater than or equal to the preset evaluation value threshold, determine that the intermediate model is a modal parameter recognition model, and the second preset condition also includes that the evaluation value is greater than or equal to the preset evaluation value threshold.

[0145] The model training device 10 further includes an adjustment module 13. The adjustment module 13 is configured to, when the evaluation value is less than a preset evaluation value threshold, readjust the parameters of the training model and re-enter the step of training the training model based on the training set and the validation set to obtain an intermediate model that meets the first preset condition.

[0146] The model training device 10 further includes a preprocessing module 14. The preprocessing module 14 is configured to preprocess the stable map samples to update the stable map samples. The preprocessing includes adjusting the number of channels of the stable map samples to a preset number of channels, adjusting the resolution of the stable map samples to a preset resolution, and normalizing at least one of the pixel values of the stable map samples.

[0147] See also Figure 14 In order to better implement the model training method of the embodiment of the present application, the embodiment of the present application also provides a modal parameter identification device 20. The modal parameter identification device 20 may include a second acquisition module 21, an identification module 22 and an output module 23. The second acquisition module 21 is used to obtain a stability diagram of the target object, where the stability diagram includes stable points and unstable points, and the stable points and unstable points are generated based on the vibration data of the target object. The identification module 22 is used to identify the stable points of the stability diagram based on a preset modal parameter identification model, and determine the true mode based on the stable points. The output module 23 is used to output the modal parameters of the target object based on the true mode.

[0148] The above describes the model training device 10 and the modal parameter identification device 20 from the perspective of functional modules in conjunction with the accompanying drawings. The functional modules can be implemented in hardware form, can be implemented by instructions in software form, or can be implemented by a combination of hardware and software modules. Specifically, the steps of the method embodiment in the embodiment of the present application can be completed by the hardware integrated logic circuit and / or software form instructions in the processor. The steps of the method disclosed in the embodiment of the present application can be directly reflected as being executed by a hardware coding processor, or can be executed by a combination of hardware and software modules in the coding processor. Optionally, the software module can be located in a mature storage medium in the field such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps in the above method embodiment in conjunction with its hardware.

[0149] See also Figure 15 The electronic device 100 of an embodiment of the present application includes a processor 50, a memory 60 and a computer program, wherein the computer program is stored in the memory 50 and executed by the processor 60, and the computer program includes instructions for executing the model training method or modal parameter identification method of any of the above embodiments.

[0150] See also Figure 16 The embodiment of the present application also provides a computer-readable storage medium 200 on which a computer program 210 is stored. When the computer program 210 is executed by the processor 220, the steps of the model training method or modal parameter identification method of any of the above-mentioned embodiments are implemented. For the sake of brevity, they are not repeated here.

[0151] In the description of this specification, the reference terms "certain embodiments", "in an example", "exemplarily", etc. mean that the specific features, structures, materials or characteristics described in conjunction with the embodiments or examples are included in at least one embodiment or example of the present application. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are mutually inconsistent.

[0152] Any process or method description in a flowchart or otherwise described herein may be understood to represent a module, segment or portion of code comprising one or more executable instructions for implementing the steps of a specific logical function or process, and the scope of the preferred embodiments of the present application includes alternative implementations in which functions may be performed out of the order shown or discussed, including performing functions in a substantially simultaneous manner or in the reverse order depending on the functions involved, which should be understood by those skilled in the art to which the embodiments of the present application belong.

[0153] Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and cannot be understood as limitations on the present application. Ordinary technicians in this field can change, modify, replace and modify the above embodiments within the scope of the present application.

Claims

1. A model training method, characterized in that: The method is applied to a model for training and identifying modal parameters, and includes: Acquire a plurality of stabilization graph samples, the stabilization graph samples including real modes and false modes, wherein stable points constituting the real modes have the same frequency, and at least some of the stable points and / or unstable points constituting the false modes have different frequencies; A modal parameter identification model is trained based on the multiple stability diagram samples.

2. The model training method according to claim 1, characterized in that The obtaining of a plurality of stabilization graph samples comprises: A plurality of stability diagram samples are generated based on randomly generated sample generation parameters, wherein the sample generation parameters include at least parameters such as the true modal order, the proportion of stable points in the true mode, and the position of unstable points.

3. The model training method according to claim 1, characterized in that The modal parameter identification model includes an encoder and a decoder, the encoder is connected to the decoder, the encoder includes a spatial pyramid module and multiple feature extraction modules, the multiple feature extraction modules are connected in sequence, the last feature extraction module is connected to the spatial pyramid module, the feature extraction module includes a feature extraction submodule and a residual module, the feature extraction submodule is connected to the residual module, the input end of the encoder is connected to the first feature extraction module, and the output end of the spatial pyramid module is connected to the input end of the decoder.

4. The model training method according to claim 3, characterized in that The decoder includes multiple decoding modules, which are connected in sequence. The output end of the spatial pyramid module is connected to the input end of the first decoding module, and the input end of the target decoding module is connected to the output end of the corresponding feature extraction module. The target decoding module has the same feature map size as the connected feature extraction module. The target decoding module is a module among the multiple decoding modules whose feature map size is the same as the feature map size of any feature extraction module in the encoder.

5. The model training method according to claim 1, characterized in that The training of a modal parameter identification model based on the plurality of stability diagram samples includes: generating a label image based on the true mode in the stabilization map sample; generating a training data set based on the label image and the stabilization map sample, and dividing the training data set into a training set, a test set, and a validation set; The preset training model is trained based on the training set, the test set and the validation set to obtain a modal parameter identification model.

6. The model training method according to claim 5, characterized in that The step of training a preset training model based on the training set, the test set, and the validation set to obtain a modal parameter identification model includes: Training the training model based on the training set and the validation set to determine an intermediate model that meets a first preset condition from the training models corresponding to the multiple trainings, wherein the first preset condition at least includes that a loss value obtained by the training model after processing the validation set is minimized; The intermediate model is verified based on the test set. When the intermediate model meets a second preset condition, the intermediate model is determined to be the modal parameter identification model. The second preset condition at least includes that the loss value corresponding to the intermediate model is less than a preset loss value threshold.

7. The model training method according to claim 6, characterized in that The training model based on the training set and the validation set includes: Processing the training set based on the training model to obtain a loss value corresponding to the training set; Processing the validation set based on the training model to obtain an evaluation score corresponding to the validation set; When the loss value is greater than a preset loss value threshold, adjusting the network parameters of the training model according to a learning rate, wherein the learning rate is adjusted based on the number of training times and / or a performance indicator of the training model, wherein the performance indicator is determined according to the loss value and / or the evaluation score; Based on the training model after parameter adjustment, the training model is entered again to process the training set to obtain a loss value corresponding to the training set, until the training is completed, and the condition for the completion of the training includes at least that the loss value obtained after the training model processes the validation set is less than a preset loss value threshold; Based on the first preset condition, the intermediate model is determined in the training models corresponding to the multiple trainings.

8. The model training method according to claim 7, characterized in that: The training model based on the training set and the validation set further includes: When the loss value is greater than the preset loss value threshold and the number of training times reaches a first preset number threshold, if the performance index of the training model does not increase during the training process corresponding to the first preset number threshold, the learning rate is reduced.

9. The model training method according to claim 7, characterized in that: The condition for the training model to be trained to end also includes: when the number of iterations reaches a second preset number threshold or the number of iterations reaches a third preset number threshold, the loss value corresponding to the training model does not decrease.

10. The model training method according to claim 6, characterized in that: The first preset condition also includes that the evaluation score of the training model is the highest.

11. The model training method according to claim 6, characterized in that: The verifying the intermediate model based on the test set, and determining that the intermediate model is the modal parameter identification model when the intermediate model meets a second preset condition, includes: Inputting the stabilization map samples of the test set into the intermediate model to obtain a test output image; Determining an evaluation value corresponding to the intermediate model based on an image similarity measurement index, the test output image, and the corresponding label image; In the case where the evaluation value is greater than or equal to a preset evaluation value threshold, determining that the intermediate model is a modal parameter identification model, the second preset condition further comprising that the evaluation value is greater than or equal to the preset evaluation value threshold; The method further comprises: When the evaluation value is less than the preset evaluation value threshold, the parameters of the training model are readjusted, and the step of training the training model based on the training set and the validation set is entered again to obtain an intermediate model that meets the first preset condition.

12. The model training method according to claim 1, characterized in that Before training the modal parameter identification model based on the multiple stability diagram samples, the method further includes: The stabilized map samples are preprocessed to update the stabilized map samples, wherein the preprocessing includes adjusting the number of channels of the plurality of stabilized map samples to a preset number of channels, adjusting the resolution of the plurality of stabilized map samples to a preset resolution, and normalizing at least one of the pixel values of the stabilized map samples.

13. A modal parameter identification method, characterized in that: include: Acquire a stability map of a target object, the stability map including stable points and unstable points, the stable points and the unstable points being generated based on vibration data of the target object; Identifying a stable point of the stability diagram based on a preset modal parameter identification model, and determining a true mode based on the stable point; Based on the true mode, modal parameters of the target object are output.

14. The modal parameter identification method according to claim 13, characterized in that: The modal parameter identification model is obtained by acquiring a plurality of stabilization graph samples and training based on the plurality of stabilization graph samples. The stabilization graph samples include true modes and false modes. The frequencies of the stabilization points constituting the true modes are the same, and the frequencies of at least some of the stabilization points and / or unstable points constituting the false modes are different.

15. A model training device, characterized in that: A model for training and identifying modal parameters, the device comprising: a first acquisition module, configured to acquire a plurality of stabilization graph samples, the stabilization graph samples including real modes and false modes, wherein the stabilization points constituting the real modes have the same frequency, and at least some of the stabilization points and / or unstable points constituting the false modes have different frequencies; The training module is used to train a modal parameter identification model based on the multiple stability diagram samples.

16. A modal parameter identification device, characterized in that: include: a second acquisition module, configured to acquire a stability map of the target object, the stability map including stable points and unstable points, the stable points and the unstable points being generated based on vibration data of the target object; an identification module, configured to identify a stable point of the stability diagram based on a preset modal parameter identification model, and determine a true mode based on the stable point; An output module is used to output modal parameters of the target object based on the real mode.

17. An electronic device, characterized in that: include: Processor, memory; and A computer program, wherein the computer program is stored in the memory and executed by the processor, and the computer program includes instructions for executing the model training method described in any one of claims 1 to 12, or instructions for the modal parameter identification method described in any one of claims 13 to 14.

18. A non-volatile computer-readable storage medium containing a computer program, characterized in that When the computer program is executed by a processor, the processor executes the model training method described in any one of claims 1 to 12 or the modal parameter identification method described in any one of claims 13 to 14.

Images