Adaptive operation modal analysis method in combination with subspace identification and Mamba network

Abstract

The invention relates to the technical field of structure health monitoring and artificial intelligence, and provides an adaptive operation modal analysis method combining subspace identification and a Mama network, and the method comprises the steps: employing a random subspace method to obtain modal parameters at different time intervals, the modal parameters comprise a modal frequency, a modal damping ratio and a modal shape, and the modal parameters comprise the modal frequency, the modal damping ratio and the modal shape; constructing a training data set; an SSI-Mama neural network model is constructed, and the model is trained; and performing online modal parameter prediction on the new vibration data by using an SSI-Mama neural network model. The method is excellent in modal parameter prediction precision, shows good noise suppression capability and calculation efficiency potential, and provides an effective solution for realizing high-precision and high-efficiency adaptive structure health monitoring.

Description

TECHNICAL FIELD

[0001] Embodiments of the present application generally relate to the field of structural health monitoring and artificial intelligence, and in particular to a method of adaptive operational modal analysis combining subspace identification and Mamba network. BACKGROUND

[0002] Modal parameter identification is the key basic problem for structural damage identification and state assessment in health monitoring systems of engineering structures such as bridges, civil engineering, aerospace, etc. However, due to the influence of the operating environment, it often faces the problem of difficult load measurement, and therefore the development of operational modal analysis methods using only structural output responses has become a major research hotspot in the field of structural dynamics.

[0003] Traditional operational modal analysis methods, such as the classic stochastic subspace method, require the construction and decomposition of large matrices when dealing with large-scale, long-time series of vibration data, which is computationally intensive and time-consuming, and poses challenges for real-time monitoring, long-term continuous monitoring of structural health, and processing of multi-sensor data. In addition, some algorithms rely on user-defined parameters, which also increases the difficulty of accurately and efficiently identifying modal parameters.

[0004] In recent years, deep learning technology has been applied in the field of structural dynamics due to its powerful feature extraction and nonlinear mapping capabilities. Some studies have attempted to use convolutional neural networks (CNN) and deep neural networks (DNN) to directly learn and predict modal parameters or damage indicators from vibration responses. However, the above methods are mainly used to process static data, treating vibration responses as static input patterns, ignoring their inherent correlation and temporal dependence as time series, and are difficult to capture dynamic changes in modal parameters, making them unsuitable for real-time modal analysis. Therefore, existing methods still have deficiencies in practical applications, and further research is needed to use deep learning technology for structural operational modal analysis. SUMMARY

[0005] To solve the above problems, the present application extracts modal parameters from continuous structural vibration responses based on a covariance-driven stochastic subspace algorithm, constructs and trains a deep learning model based on an SSI-Mamba network, and uses its selective state space mechanism to efficiently learn the direct mapping relationship from the original vibration sequence to the modal parameters. The trained SSI-Mamba model can quickly and accurately predict modal parameters from new vibration data, demonstrating good noise suppression capability and computational efficiency potential, providing an effective solution for high-precision and high-efficiency adaptive structural health monitoring.

[0006] According to an embodiment of the present application, a method of adaptive operational modal analysis combining subspace identification and Mamba network is provided.

[0007] In a first aspect of the present application, a method of adaptive operational modal analysis combining subspace identification and Mamba network is provided. The method comprises: Step S01: using a random subspace method to obtain modal parameters at different time intervals, including modal frequency, modal damping ratio and modal shape, and constructing a training data set; Step S02: constructing an SSI-Mamba neural network model and training the model; Step S03: using the SSI-Mamba neural network model to predict the online modal parameters of new vibration data.

[0008] Further, the specific steps of constructing the training data set in step S01 are: Step S011: obtaining the multi-channel acceleration time domain response of the structure through simulation or vibration test, forming an original response matrix with a dimension of ; Step S012: logically dividing the response matrix according to the preset time interval to form several time windows, and cutting the cumulative response data from the start to the end time of each window, identifying the discrete-time state matrix for each cumulative data segment, and obtaining the complex eigenvalues of the discrete-time system by performing eigenvalue decomposition on the state matrix , converting the complex eigenvalues into physical modal parameters of the continuous-time system: modal frequency and modal damping ratio, and calculating the preliminary complex modal shape corresponding to each modal; and constructing a preliminary modal parameter sequence {modal frequency, modal damping ratio, modal shape}; Step S013: performing physical information enhancement on the preliminary modal parameter sequence for screening stable modes and solving the phase ambiguity problem; Step S014: flattening and splicing the real part of the phase-aligned modal shape corresponding to the modal frequency and modal damping ratio of the screened stable mode, constructing a long real value vector, pairing the original vibration data segment with the long real value vector corresponding to the same time window, and forming a final training data set .

[0009] Further, the formula for calculating the complex eigenvalues of the continuous-time system from the complex eigenvalues of the discrete-time system in step S012 is: , wherein, is the th time window identified, the th modal corresponding to the discrete-time system of the subsidiary eigenvalue, is the complex eigenvalue of the continuous-time system converted by , is the signal sampling period.

[0010] Further, the formula for converting the complex eigenvalue into the physical modal parameter of the continuous-time system in step S012 is: , , wherein, is the modal frequency of the th modal, is the modal damping ratio of the th modal.

[0011] Further, the specific steps for screening the stable modal in step S013 are: the stochastic subspace identification is performed under multiple model orders covering a preset range, thereby obtaining the identification results of multiple model orders, and the stable modal is screened out by using the stability diagram criterion.

[0012] Further, the specific steps for solving the phase ambiguity problem in step S013 are: a unique reference mode is determined for each order modal, and the reference mode is used to determine and eliminate any phase in the SSI identification result by using the phase alignment mechanism: , , wherein, is the preset th reference mode; is the th preliminary modal shape identified under the time window , the angle() function is used to extract the phase angle of the complex inner product result, i.e., the phase difference , is the modal shape after phase alignment.

[0013] Further, the process of training the model in step S02 is: taking the vibration response of different time intervals as the input and taking the modal parameter as the output, training the model by minimizing the weighted mean square error loss function , thereby establishing the mapping relationship from the vibration response to the structural modal parameter: , wherein, , , are preset weight coefficients of frequency, damping ratio and modal shape loss respectively, and are the parts of the model prediction vector and the target vector corresponding to the frequency respectively, is the predicted output of the model to the input , represents the neural network model defined by the trainable parameters ; the network weights are iteratively updated using the AdamW optimizer until the model converges.

[0014] In the second aspect of the application, an apparatus for adaptive operational modal analysis combining subspace identification and Mamba network is provided. The apparatus comprises: a training set construction module for obtaining modal parameters including modal frequency, modal damping ratio and modal shape at different time intervals using a stochastic subspace method to construct a training data set; a model construction module for constructing an SSI-Mamba neural network model and training the model; a model prediction module for using the SSI-Mamba neural network model to perform online modal parameter prediction on new vibration data.

[0015] In the third aspect of the application, an electronic device is provided. The electronic device comprises a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to realize the method according to the first aspect of the application.

[0016] In the fourth aspect of the application, a computer readable storage medium is provided, which stores a computer program, and the program is executed by a processor to realize the method according to the first aspect of the application.

[0017] The above-mentioned English explanations are: SSI: Stochastic Subspace Identification Mamba: A novel Selective State Space Model architecture that can efficiently handle long sequence data through an improved structured state space sequence (S4) model The application extracts modal parameters from continuous structural vibration responses based on a covariance-driven random subspace algorithm, constructs and trains a deep learning model based on an SSI-Mamba network, uses the selective state space mechanism thereof to efficiently learn the direct mapping relationship from the original vibration sequence to the modal parameters, and the trained SSI-Mamba model can quickly and accurately predict the modal parameters from new vibration data, has good noise suppression capability and computational efficiency potential, and provides an effective solution for realizing high-precision and high-efficiency adaptive structural health monitoring.

[0018] It should be understood that the content described in the summary section is not intended to limit the key or important features of the embodiments of the application, nor to limit the scope of the application. Other features of the application will become apparent from the following description.

[0019] The beneficial effects of the application are: 1. After training, the SSI-Mamba model can realize end-to-end prediction from vibration signals to modal parameters, bypassing the time-consuming large matrix operation in the traditional SSI method, greatly improving the computational efficiency of online monitoring; 2. The application uses the SSI algorithm to generate high-quality, physically meaningful label data for the deep learning model, and eliminates data ambiguity through post-processing methods such as phase alignment, ensuring the convergence of model training and the accuracy of prediction; 3. The deep learning model can establish a smooth mapping relationship from vibration to modal parameters by learning a large amount of data, and has good noise suppression capability, and the prediction result is more stable than single SSI identification. BRIEF DESCRIPTION OF DRAWINGS

[0020] The above and other features, advantages, and aspects of the embodiments of the application will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings, in which: Figure 1 A method flowchart of adaptive operational modal analysis combining subspace identification and Mamba network according to an embodiment of the application is shown; Figure 2 A five-degree-of-freedom simulation model schematic diagram according to an embodiment of the application is shown; Figure 3 A modal parameter identification flowchart according to an embodiment of the application is shown; Figure 4 A time history tracking diagram of modal frequency identified by different methods according to an embodiment of the application is shown; Figure 5 A modal mode prediction result and subspace method identification result comparison diagram according to an embodiment of the application is shown; Figure 6A modal damping ratio time history tracking plot is shown according to an embodiment of the present application. Figure 7 A device block diagram of adaptive operational modal analysis combining subspace identification and Mamba network is shown according to an embodiment of the present application. Figure 8 A device schematic diagram of adaptive operational modal analysis combining subspace identification and Mamba network is shown according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] To make the objects, technical solutions and advantages of embodiments of the present application clearer, the technical solutions in embodiments of the present application will be described clearly and completely below with reference to the drawings in embodiments of the present application. Obviously, the described embodiments are some but not all of the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0022] According to an embodiment of the present application, a method of adaptive operational modal analysis combining subspace identification and Mamba network is proposed. Modal parameters are extracted from continuous structural vibration responses based on a covariance-driven stochastic subspace algorithm, a deep learning model based on SSI-Mamba network is constructed and trained, and a selective state space mechanism is used to efficiently learn the direct mapping relationship from the original vibration sequence to the modal parameters. The trained SSI-Mamba model can quickly and accurately predict modal parameters from new vibration data, has good noise suppression ability and computational efficiency potential, and provides an effective solution for realizing high-precision and high-efficiency adaptive structural health monitoring.

[0023] The principles and spirits of the present application will be explained in detail below with reference to several representative embodiments of the present application.

[0024] Figure 1 is a method flowchart of adaptive operational modal analysis combining subspace identification and Mamba network according to an embodiment of the present application. The method comprises: Step S01: modal parameters including modal frequency, modal damping ratio and modal shape at different time intervals are obtained by using a stochastic subspace method, and a training data set is constructed; Step S02: an SSI-Mamba neural network model is constructed, and the model is trained; Step S03: the SSI-Mamba neural network model is used to predict online modal parameters for new vibration data.

[0025] It is to be understood that even though operations of the method of the present application are described in a particular, sequential order in the above embodiments and figures, this is not meant to be limiting or to imply that any particular order is required for the operations to be performed, nor that all of the operations are necessarily performed, in order to achieve the desired results. Additionally or alternatively, certain steps can be combined, performed in a different order, or omitted, and / or certain steps can be split into multiple steps, and / or certain steps can be performed concurrently.

[0026] In order to make the method of adaptive operational modal analysis combining subspace identification and Mamba network more clearly explained, a specific embodiment is described below, however, it is worth noting that the embodiment is only to better illustrate the present application, and does not constitute an improper limitation on the present application.

[0027] The method of adaptive operational modal analysis combining subspace identification and Mamba network is further described in detail below with a specific example: Step S01: The random subspace method is used to obtain modal parameters at different time intervals, including modal frequency, modal damping ratio and modal shape, including modal frequency, modal damping ratio and modal shape, and a training data set containing vibration data segments and modal parameters corresponding to different time intervals is constructed. The specific steps are as follows.

[0028] Step S011: Obtain the multi-channel acceleration time domain response of the structure through simulation or vibration test, form an original response matrix with a dimension of as the input of the entire analysis process.

[0029] Step S012: In order to capture the trend of the change of modal parameters over time, the long-term acceleration response matrix is logically divided according to the preset time interval to form a series of time windows. For the ending time of each time window, is the window number. The cumulative response data from the starting time to the ending time of the simulation or test is intercepted. This cumulative processing method can more stably reflect the average dynamic characteristics of the structure in that time period. Subsequently, the random subspace identification-covariance driven method is applied to each cumulative data segment respectively, so as to identify a series of discrete-time state matrices evolving over time.

[0030] By performing eigenvalue decomposition on the state matrix , the complex eigenvalues of the discrete-time system are obtained: , wherein, is the complex eigenvalue of the discrete-time system corresponding to the m-th modal identified in the k-th time window, is the complex eigenvalue of the continuous-time system converted from is the complex eigenvalue of the discrete-time system corresponding to the m-th modal identified in the k-th time window, is the complex eigenvalue of the continuous-time system converted from is the complex eigenvalue of the discrete-time system corresponding to the m-th modal identified in the k-th time window, is the sampling period of the signal.

[0031] Subsequently, the complex eigenvalue is converted into the physical modal parameters of the continuous-time system, i.e., the modal frequency and the modal damping ratio , according to the following formula: , , At the same time, as another output of the stochastic subspace identification process, the preliminary complex modal shape corresponding to each modal is calculated, which is a complex vector with the same dimension as the system degree of freedom.

[0032] The set of modal parameters obtained above in each time window is , , which constitutes the preliminary modal parameter sequence.

[0033] Step S013: The preliminary modal parameter sequence is enhanced with physical information, and stable modes are selected to solve the two core problems faced when it is used as a machine learning label: false modes and phase ambiguity.

[0034] First, the stochastic subspace identification is performed at multiple model orders covering a predetermined range, thereby obtaining identification results of multiple model orders. The stable modes are selected from the identification results using the stability diagram criterion. This step aims to eliminate false modes caused by the algorithm itself or measurement noise, and to ensure that only modal parameters representing the true physical characteristics of the structure are retained.

[0035] Next, the phase ambiguity problem is addressed. As a physical vibration form, the mathematical expression of the modal shape has arbitrary phase rotation invariance in the complex domain, i.e., and describe the same physical mode. This many-to-one supervised learning is fatal, as it will cause the model to learn a non-unique, randomly rotated target when facing the same input.

[0036] To solve this problem, we introduce a phase alignment mechanism, the core purpose of which is to establish a unified and unique mathematical expression benchmark for each physical modal. To this end, we pre-determine a unique reference shape for each modal It is emphasized that the role of the reference mode shape is not to provide a true value to be fitted, but to serve as a fixed phase reference. In simulation, it can be the mode shape obtained from theoretical eigenvalue analysis; in real engineering, it can be the baseline mode shape identified from the initial healthy state of the structure. The phase alignment procedure only utilizes the reference mode shape to determine and eliminate the arbitrary phase in the SSI identification results , while completely preserving the amplitude and shape information reflecting the current true state of the structure. The specific alignment procedure is as follows: , , where, is the pre-set reference mode shape of the th order; is the preliminary modal shape of the th order identified in the time window , and the angle() function is used to extract the phase angle of the complex inner product result, i.e., the phase difference , is the phase-aligned modal shape.

[0037] After this step, the same order modal shape at all time points is converted to a unified, unambiguous coordinate system for expression.

[0038] Step S014: The modal frequencies, modal damping ratios and real parts of the phase-aligned modal shapes corresponding to the screened stable modes are flattened and spliced to construct a stable and unique long real value vector . This vector is the final supervision signal. The present application pairs the original vibration data segment with its corresponding in the same time window to form the final training data set .

[0039] Step S02: Constructing an SSI-Mamba neural network model, taking the vibration responses at different time intervals as input and the modal parameters as output, training the model by minimizing the weighted mean square error loss function to establish the mapping relationship from vibration response to structural modal parameters: , where, , , are the pre-set weight coefficients of frequency, damping ratio and modal shape loss, respectively, and are the corresponding frequency parts of the model prediction vector and the target vector, respectively, It is the model's response to the input The predicted output, Represents the entire set of trainable parameters Define a neural network model. Iterate and update the network weights using the AdamW optimizer. Continue until the model converges.

[0040] Step S03: Use the SSI-Mamba neural network model to predict the modal parameters of the new vibration data online.

[0041] To examine the effectiveness and superiority of the proposed method, a five-degree-of-freedom numerical simulation model was used for verification, such as... Figure 2 As shown. The quality matrix of this model is kg, interlayer stiffness is ( The damping model uses Rayleigh damping. The system time-domain response uses Newmark- The sampling frequency was 100Hz and the sampling duration was 1200s, obtained through numerical integration.

[0042] Adopting such Figure 3 The flowchart shown illustrates the steps for modal parameter identification. In the experiment, 120 datasets with different random excitation sequences were generated for training and testing. Specifically, for the acceleration response of each dataset, modal parameters were identified using a stochastic subspace method with accumulated data at 10-second intervals. The system order was set to 50, and the number of rows in the Hankel matrix was set to 100. For training the SSI-Mamba network, the model input used a vibration response time window of 50 seconds, the network's internal hidden dimension was set to 256, and it consisted of 6 stacked Mamba core modules. The optimization process used the AdamW optimizer with a learning rate of 1e-4 for 60 training epochs.

[0043] Table 1 lists the modal parameter identification results of different methods.

[0044] Table 1

[0045] Figure 4 A comparison chart of time history tracking of the identified modal frequencies. Figure 5 The results show a comparison between the mode shapes predicted by the SSI-Mamba model and the mode shapes identified by the classical SSI-COV method at a selected moment. Figure 6 Time history tracking plots for different methods of identifying modal damping ratios. (From Table 1 and...) Figure 4 As can be seen, compared with the SSI-COV method, the SSI-Mamba method can accurately identify the modal frequencies and achieve results with smaller errors after convergence, demonstrating superior performance. (See Table 1 and...)Figure 5 It can be seen that in the aspect of modal shape identification, the modal confidence (MAC) of each order modal shape predicted by the SSI-Mamba method reaches 1.0000, indicating that the method can accurately predict the modal shape. It is worth noting that from Table 1 and Figure 6 It can be seen that compared with the modal frequency and the modal shape, the damping ratio predicted by the SSI-Mamba method is quite close to the identification result of the classic SSI-COV method, which is mainly due to the inherent uncertainty and instability in identifying small damping using the SSI method, thereby providing inaccurate label signals for the deep learning model.

[0046] Based on the same inventive concept, the application also proposes an apparatus for adaptive operational modal analysis combining subspace identification and Mamba network. The implementation of the apparatus can refer to the implementation of the above-mentioned method, and the repeated parts will not be described here. As Figure 7 shown, the apparatus 100 comprises: a training set construction module 101 for acquiring modal parameters at different time intervals, including modal frequency, modal damping ratio and modal shape, using the random subspace method, and constructing a training data set; a model construction module 102 for constructing an SSI-Mamba neural network model and training the model; a model prediction module 103 for using the SSI-Mamba neural network model to perform online modal parameter prediction on new vibration data.

[0047] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the described modules can refer to the corresponding process in the foregoing method embodiments, which will not be described here.

[0048] As Figure 8 shown, the device includes a central processing unit (CPU) that can perform various appropriate actions and processes according to computer program instructions stored in a read-only memory (ROM) or loaded from a storage unit into a random access memory (RAM). In the RAM, various programs and data required for device operation can also be stored. The CPU, ROM, and RAM are connected to each other through a bus. An input / output (I / O) interface is also connected to the bus.

[0049] A plurality of components in the device are connected to the I / O interface, including: an input unit such as a keyboard, a mouse, etc.; an output unit such as various types of displays, speakers, etc.; a storage unit such as a magnetic disk, an optical disk, etc.; and a communication unit such as a network card, a modem, a wireless communication transceiver, etc. The communication unit allows the device to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunications networks.

[0050] The processing unit performs the various methods and processes described above, such as method steps S01-S03. For example, in some embodiments, the method steps S01-S03 can be implemented as a computer software program tangibly embodied in a machine readable medium, such as the storage unit. In some embodiments, portions or all of the computer program can be loaded and / or installed onto the device via the ROM and / or the communication unit. When the computer program is loaded onto the RAM and executed by the CPU, one or more of the method steps S01-S03 described above can be performed. Alternatively, in other embodiments, the CPU can be configured to perform the method steps S01-S03 by way of other suitable means, such as by way of firmware.

[0051] The functionality described above in this document can be performed, at least in part, by one or more hardware logic components. For example, and without limitation, illustrative types of hardware logic components that can be used include Field-programmable Gate Arrays (FPGAs), Application-specific Integrated Circuits (ASICs), Application-specific Standard Products (ASSPs), System-on-a-chip systems (SOCs), Complex Programmable Logic Devices (CPLDs), etc.

[0052] Program code for carrying out the methods of the present application can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, causes the machine to perform the functions / acts specified in the flow diagrams and / or block diagrams. The program code can be embodied in whole or in part within a machine readable medium, which can be any medium for storing or transmitting the program code. The program code can be transmitted in the form of signals over a transmission medium via a data signal or carrier wave, or it can be provided on a machine readable medium.

[0053] In the context of the present application, a machine-readable medium can be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. The machine-readable medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium will include one or more lines of a computer program code, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0054] Further, although operations are depicted in a particular order, this should not be understood as requiring such order or sequential order of execution, or that all illustrated operations be performed to achieve desirable results. In certain circumstances, multitasking and parallel processing can be advantageous. Likewise, the specific sequence of operations illustrated in the above discussion can not be required in other implementations. Similarly, certain features can also be implemented in a single implementation, while some features can be implemented multiple times in either a single implementation or distributed across many implementations. Conversely, various features could be implemented in a single implementation or across several implementations. It should be understood that the various features could be implemented in hardware, software, firmware, or any combination thereof.

[0055] Although the subject matter has been described in language specific to structural features and / or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims.

Claims

1. A method combining subspace identification and adaptive operational mode analysis of Mamba networks, characterized in that, The method includes: Step S01: Use the random subspace method to obtain modal parameters at different time intervals, including modal frequencies, modal damping ratios, and mode shapes, and construct a training dataset; Step S02: Construct the SSI-Mamba neural network model and train the model; Step S03: Use the SSI-Mamba neural network model to predict the modal parameters of the new vibration data online.

2. The method for combining subspace identification and adaptive operational mode analysis of Mamba networks according to claim 1, characterized in that, The specific steps for constructing the training dataset described in step S01 are as follows: Step S011: Obtain the multi-channel acceleration time-domain response of the structure through simulation or vibration test, forming a dimension of The original response matrix ; Step S012: Receive the response matrix According to the preset time interval The time window is logically divided into several time windows, and each window is captured from the start to the end of the simulation or experiment. The cumulative response data is used to identify the discrete-time state matrix for each cumulative data segment. By analyzing the state matrix Eigenvalue decomposition is performed to obtain complex eigenvalues ​​of the discrete-time system. These complex eigenvalues ​​are then converted into physical modal parameters of the continuous-time system: modal frequencies and modal damping ratios. Preliminary complex mode shapes corresponding to each mode are calculated, and a preliminary modal parameter sequence {modal frequency, modal damping ratio, mode shape} is constructed. Step S013: Perform physical information enhancement on the preliminary modal parameter sequence to screen stable modes and resolve phase ambiguity issues; Step S014: Flatten and stitch together the modal frequencies, modal damping ratios, and real parts of the phase-aligned mode shapes corresponding to the selected stable modes to construct a long real-valued vector. Pair the original vibration data segments with the long real-valued vectors corresponding to them in the same time window to form the final training dataset. .

3. The method for combining subspace identification and adaptive operational mode analysis of Mamba networks according to claim 2, characterized in that, The step S012 described above uses the complex eigenvalues ​​of the discrete-time system. Calculate the complex eigenvalues ​​of a continuous-time system The formula is: ， in, It is the first The first time window identified, The associated eigenvalues ​​of a discrete-time system corresponding to a first-order mode. It is by The complex eigenvalues ​​of the continuous-time system obtained by the transformation It is the signal sampling period.

4. The method for combining subspace identification and adaptive operational mode analysis of Mamba networks according to claim 3, characterized in that, The formula for converting complex eigenvalues ​​into physical modal parameters of a continuous-time system as described in step S012 is: ， ， in, for The modal frequencies of the first mode. for Modal damping ratio of the first mode.

5. The method for combining subspace identification and adaptive operational mode analysis of Mamba networks according to claim 2, characterized in that, The specific steps for screening stable modes in step S013 are as follows: random subspace identification is performed under multiple model orders covering a preset range to obtain identification results of multiple model orders, and stable modes are screened out from them using the stability graph criterion.

6. The method for combining subspace identification and adaptive operational mode analysis of Mamba networks according to claim 2, characterized in that, The specific steps for resolving the phase ambiguity problem described in step S013 are as follows: A unique reference mode shape is pre-determined for each mode. The phase alignment mechanism utilizes this reference mode shape to determine and eliminate arbitrary phases in the SSI identification results. : ， ， in, For the preset first First reference mode shape; In the time window The first one identified The preliminary mode shape is obtained using the angle() function, which extracts the phase angle, i.e., the phase difference, from the complex inner product result. , These are the mode shapes after phase alignment.

7. The method for combining subspace identification and adaptive operational mode analysis of Mamba networks according to claim 1, characterized in that, The training process described in step S02 is as follows: taking the vibration response at different time intervals as input and the modal parameters as output, the model is trained by minimizing the weighted mean square error loss function. To train the model, thereby establishing the mapping relationship between vibration response and structural modal parameters: ， in, , , These are preset weighting coefficients for frequency, damping ratio, and modal loss, respectively. and These are the frequency components of the model's predicted vector and the target vector, respectively. It is the model's response to the input The predicted output, Represents the entire set of trainable parameters Define a neural network model; iteratively update the network weights using the AdamW optimizer. Continue until the model converges.

8. An apparatus combining subspace identification and adaptive operational mode analysis of Mamba networks, characterized in that, The device implements the method as described in any one of claims 1 to 7, comprising: Training set construction module: used to obtain modal parameters at different time intervals using the random subspace method, including modal frequencies, modal damping ratios and mode shapes, and to construct the training dataset; Model building module: Used to build and train the SSI-Mamba neural network model; Model prediction module: Used to predict modal parameters of new vibration data online using the SSI-Mamba neural network model.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 7.

Images