Modal recognition system with strong generalization ability

Abstract

The invention relates to the technical field of structural body modal identification, and discloses a modal identification system with strong generalization ability, which is composed of an identification framework and an auxiliary framework. The identification framework comprises a sensitive unit, a data acquisition strategy, a data preprocessing unit, a model generation unit and a model application unit; the auxiliary frame comprises a processor, a memory, a communication unit and a power management unit and assists the identification frame to work. The sensing unit detects the vibration acceleration and temperature of a detected structure body, and the data acquisition strategy acquires data according to the excitation law of the detected structure. The data preprocessing unit carries out Fourier transform, filtering and normalization processing on the acceleration data. And the model generation unit performs training by using a transform model, the model application unit inputs application period data into a trained model to output a modal state score, and early warning is performed if the score is lower than a threshold value. The system can accurately identify the modal change of the detected structural body, has obvious advantages in the aspect of structural body health monitoring, and provides key guidance for structural body maintenance.

Description

Technical Field

[0001] The present invention relates to the technical field of structural modal identification, and specifically provides a modal identification system with strong generalization ability. Background Art

[0002] In the field of structural health monitoring, the safety and stability of structures are of utmost importance. However, traditional monitoring methods have many limitations. In the past, time-domain analysis and manual evaluation were the main means. Time-domain analysis relies on observing time-series data and it is difficult to accurately extract key features from complex vibration data. For signals of structural changes that are weak or hidden in complex backgrounds, time-domain analysis often fails to detect them. Manual evaluation is limited by human subjective experience and fatigue. Different evaluators may have different judgment criteria, and in the face of large-scale and long-term monitoring tasks, manual evaluation is inefficient and prone to omissions.

[0003] With the rapid development of sensing technology, frequency-spectrum based feature analysis has emerged as a new and highly regarded monitoring method. By deeply analyzing the frequency and amplitude of structural vibration data, this method can uncover potential abnormal state information. For example, when there are small cracks or stiffness changes in a structure, the frequency and amplitude of its vibration will change subtly, and spectrum analysis can accurately capture these changes, providing a more precise basis for structural health monitoring. However, in actual application scenarios, which are complex and variable, this method also faces challenges. Under different environmental conditions, interference from factors such as temperature, humidity, and wind can cause fluctuations and noise in the collected spectrum data, affecting the recognition accuracy. Moreover, when faced with abnormal situations that have never been encountered before, models based on frequency-spectrum feature analysis may be unable to accurately identify due to lack of adaptability, resulting in monitoring deviations.

[0004] To effectively solve the above problems, it is urgent to improve the recognition accuracy and enhance the robustness of the model. The modal identification system with strong generalization ability proposed in this paper emerges as the times require. This system organically combines spectrum information and the Transformer model. Through multi-dimensional feature fusion, it integrates various data features such as vibration acceleration and temperature to comprehensively reflect the state of the structure. By using the self-attention mechanism, the model can intelligently focus on key information and enhance its ability to understand complex vibration patterns. This not only greatly improves the recognition accuracy of complex vibration patterns but also enables the system to have excellent generalization performance in diverse actual applications, providing a reliable guarantee for long-term structural health monitoring and evaluation, and strongly promoting the development of structural health monitoring technology. Summary of the Invention

[0005] The purpose of the present invention is to provide a modal identification system with strong generalization ability to solve the problems raised in the above background art.

[0006] To achieve the above object, the present invention provides the following technical solutions: A modal recognition system with strong generalization ability, the system includes: a recognition framework and an auxiliary framework.

[0007] Among them, the recognition framework includes: a sensitive unit, a data acquisition strategy, a data preprocessing unit, a model generation unit, and a model application unit.

[0008] Among them, the auxiliary framework includes: a processor, a memory, a communication unit, and a power management unit.

[0009] The auxiliary framework assists the recognition framework to complete the work of modal recognition.

[0010] The time period of the entire system operation is divided into a model period and an application period.

[0011] The modal states are divided into a first state and a second state, where the first state is a stable state and the second state is a changing state.

[0012] The sensitive unit is used to detect the vibration acceleration and temperature of the measured structure. The data acquisition strategy controls the specific distribution of the time points and time periods for collecting data. The time period of the model period is confirmed according to the fault cycle distribution of the measured structure. All the data collected during the model period is divided into a training set, a validation set, and a test set. The data preprocessing unit performs Fourier transform and filtering on the collected acceleration data. The model generation unit uses the transformer model for training, learns from the training set and the validation set, and trains a model in combination with the test set. The model application unit inputs the data collected during the application period into the trained model, thereby outputting the modal state of the structure. When the output modal state of the structure continues to be the second state, it reminds to perform manual detection on the measured structure.

[0013] Preferably, the sensitive unit is used to detect the vibration acceleration and temperature of the measured structure, and the working frequency range of the sensitive unit is selected according to the inherent properties of the measured structure. The larger the working frequency range of the sensitive unit, the more types of structures can be monitored.

[0014] Preferably, the data acquisition strategy controls the specific distribution of the time points and time periods for collecting data. The basic principle of data sampling is to collect data at a specific sampling frequency within a specific time, that is, one round of data sampling refers to the process of collecting data at a specific sampling frequency within a specific time. Analyze the time law of the excitation of the measured structure, divide the time period into multiple levels according to the density of excitation, and the higher the density, the higher the level assigned. Increase the sampling rounds within the high-level time periods and decrease the sampling rounds within the low-level time periods. Arrange random time periods for sampling outside the time periods of all sampling rounds.

[0015] Preferably, the time period during which the entire system operates is divided into a model period and an application period. The model period is the time period for collecting data for training the model. By default, this time period is at the initial stage of the structural body's life cycle and lasts for more than one temperature change cycle of the environment where the structural body is located. All the data tags collected during the model period are set to the first state. The time after the model period is the application period, during which the trained model is used to output the state.

[0016] Preferably, the modal state is divided into a first state and a second state, where the first state is a stable state and the second state is a changing state. The stable state represents that the dynamic characteristics of the structure have not changed, the structural stiffness of the monitored structural part and its associated parts has not changed, and the reliability of the connection of the structural part has not changed. The changing state represents that the dynamic characteristics of the structure have changed, that is, the structural stiffness of the structural part and its associated parts has changed, and there is an abnormality in the connection of the structural part.

[0017] Preferably, the data preprocessing unit segments the collected acceleration data, with one segment as a group of data, and separately performs Fourier transform and filtering on each group of data. The vibration acceleration of the structural body collected by the sensitive unit is time-domain data, that is, a sequence of accelerations corresponding to a sequence of specific time intervals. After performing Fourier transform on the time-domain data, frequency-domain data is obtained. The acceleration amplitudes corresponding to the frequency points outside the frequency range monitored by the sensitive unit are set to zero, that is, filtering. According to the design form and material of the structural body, the structural body is divided into several categories. The dynamic response range of each category is determined. That is, each category of structural body corresponds to a dynamic response frequency range, and the acceleration amplitudes corresponding to the frequency points outside this frequency range are set to zero. Finally, each type of structural body corresponds to a filter.

[0018] Preferably, after the data preprocessing unit filters the collected acceleration data, the data becomes three-dimensional data consisting of two-dimensional data of frequency points and amplitudes plus temperature data. Among them, the temperature data is global data, which is the average temperature during the time period of this group of data, and the temperatures corresponding to all the frequency points within a group of data are the same. The data is segmented with a basic unit length, the frequency point value selects the first value of each small unit after segmentation, and the amplitude selects the maximum value of the small unit after segmentation. The initial data length is a multiple of the basic unit length of the final data length. Finally, the three-dimensional data is normalized.

[0019] Preferably, the model generation unit is responsible for training the model, and after generating the trained model, outputs the model to the model application unit. The model generation unit uses the transformer architecture to process the normalized frequency point amplitude temperature data.

[0020] The embedding layer converts each pair of frequency point amplitude temperatures into an embedding vector, and the entire input can be represented as a sequence of embedding vectors. The embedding transformation is completed through a multi-layer fully connected network.

[0021] The positional encoding uses the sequential information of the frequencies to generate periodic positional encodings. For each frequency point, corresponding sine and cosine encodings can be generated. It captures the relative positions of the frequency points while retaining the differences of the frequency points that are far apart.

[0022] The data after embedding is passed into the Transformer encoder layer, which has a multi-head attention mechanism, a feed-forward neural network, normalization, and residual connections. The encoding layers are stacked in multiple layers.

[0023] The output layer uses a regression task, and the output result is a value between zero and one, which serves as the modal state score. If the score is greater than or equal to the threshold, it is the first state, and if the score is lower than the threshold, it is the second state.

[0024] Preferably, the model application unit uses the trained model to process the data collected during the application period and outputs the modal state score of the structure. An alarm is issued after the output data continuously falls below the threshold.

[0025] Preferably, the auxiliary framework at least includes: a processor, a graphics processor, a memory, a communication unit, and a power management unit. The processor calculates the input preprocessing of the data, as well as the judgment and alarm instructions after the data output. The graphics processor calculates all the intermediate data during the training and application of the model. The memory stores the data, the communication unit conveys the alarm signal, and the power management unit manages the power of the entire system.

[0026] Compared with the prior art, the beneficial effects of the present invention are:

[0027] In terms of data acquisition, the working frequency range of the sensitive unit of the system can be flexibly selected according to the inherent properties of the structure to be measured, which greatly broadens the types of structures that can be monitored. Whether it is a large building, a bridge, or a precision mechanical component, as long as its vibration frequency is within the working range of the sensitive unit, effective monitoring can be achieved. At the same time, the data acquisition strategy fully considers the time law of the structure to be measured being excited, divides the time period into levels according to the excitation density, increases the sampling rounds in the high-level time periods, reduces the sampling rounds in the low-level time periods, and arranges random sampling at other times. This dynamic sampling method can not only ensure sufficient data acquisition during critical periods, improve the accuracy of monitoring, but also reasonably allocate resources, avoid unnecessary sampling burdens, and save labor and material costs.

[0028] In the data preprocessing stage, the system performs Fourier transform and filtering operations on the collected acceleration data. The Fourier transform converts time-domain data into frequency-domain data, revealing the frequency characteristics hidden in complex vibration signals. Through filtering, not only can interference frequency points outside the monitoring frequency range of the sensitive unit be removed, but also the data can be classified according to the design form and material of the structure, the dynamic response range of each category can be determined, and frequency points outside this range can be further removed, providing purer and more targeted data for subsequent analysis. In addition, the filtered data is normalized to unify the data scale and improve the efficiency and stability of model training.

[0029] The model generation unit adopts the transformer architecture and uses its powerful feature extraction and learning capabilities to process the normalized data. The embedding layer converts the frequency point amplitude temperature data into embedding vectors, the position encoding captures the relative position information of the frequency points, the multi-head attention mechanism deeply learns the relationships between various features, the feed-forward neural network performs feature extraction, and the stacked encoder layers gradually mine deep-level features. This design enables the model to accurately capture the features of the structure in different states and improve the recognition accuracy. At the same time, by introducing simulated unhealthy data and data augmentation techniques, the generalization ability of the model is effectively improved, enabling it to better handle various unknown abnormal situations in actual monitoring.

[0030] In practical applications, the model application unit uses the trained model to process the data collected during the application period and outputs the modal state score of the structure in real time. When the score continuously drops below the threshold, a warning is issued to remind relevant personnel to conduct manual inspections on the structure in a timely manner. This function can promptly detect potential problems of the structure, gain valuable time for the maintenance and repair of the structure, avoid safety accidents caused by structural problems, and ensure the safety of people's lives and property. Brief Description of the Drawings

[0031] Figure 1 It is a schematic diagram of the data flow of the system provided in this article.

[0032] Figure 2 It is a schematic diagram of the frequency distribution of the acquisition strategy of the system provided in this article.

[0033] Figure 3 It is a schematic diagram of the device link of the system provided in this article.

[0034] Figure 4 It is a schematic diagram of the logical link of the system provided in this article.

[0035] Figure 5 It is a block diagram of the device architecture of the system provided in this article.

[0036] Figure 6 It is a data example diagram of the system provided in this article. Detailed Implementation Manner

[0037] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0038] Please refer to Figure 1-6 , the overall implementation solution of a modal recognition system with strong generalization ability of the present invention is as follows:

[0039] The system includes an identification framework and an auxiliary framework. The identification framework includes: a sensitive unit, a data acquisition strategy, a data preprocessing unit, a model generation unit, and a model application unit. The auxiliary framework includes: a processor, a memory, a communication unit, and a power management unit. The auxiliary framework assists the identification framework to complete the work of modal recognition. The sensitive unit is used to detect the vibration acceleration of the structure to be measured. The modal recognition system with strong generalization ability identifies whether the mode of the structure to be measured has changed by using the preprocessed vibration acceleration data, which plays a crucial guiding role in the maintenance of the structure.

[0040] The sensitive unit is used to detect the vibration acceleration and temperature of the structure to be measured. In this paper, a cable-stayed bridge is taken as an example for explanation. Figure 3 A sensitive unit is installed below the bridge deck of the cable-stayed bridge 41 shown. The sensitive unit typically includes a vibration sensor 42. The vibration sensor 42 is installed at key parts of the bridge. Typical positions include the main cable, the suspender, the bridge deck, the bridge tower, and the cross beam. The structures and materials at different positions are different, and their respective modes are also completely different. Vibration sensors with different frequency ranges are selected according to different structures to be measured. The excitation laws at different positions are also different.

[0041] Taking the bridge deck as an example, the excitation frequency is relatively high during the morning and evening rush hours, moderate from six o'clock to twenty-two o'clock except during the morning and evening rush hours, and the lowest from twenty-two o'clock to six o'clock. The data acquisition strategy configuration center is divided into three data acquisition frequency levels according to the high and low excitation frequencies. The unit of data acquisition is 2 minutes. The high-level acquisition frequency acquires data at a high frequency during the peak period. For example, 10 units are acquired per hour. The medium-level acquisition frequency acquires data at a medium frequency from six o'clock to twenty-two o'clock. For example, 5 units are acquired per hour. The low-level acquisition frequency acquires data at a low frequency at other times. For example, 2 units are acquired per hour. As Figure 2As shown in the figure, it is a schematic diagram of the frequency distribution of the acquisition strategy. Among them, the early peak and the late peak are the first interval 31 using a high-level frequency. The interval from 6:00 to 22:00 except for the early and late peaks is the second interval 32 using a medium-level frequency. The interval from 6:00 to 22:00 is the third interval 33 using a low-level frequency.

[0042] The above is the data acquisition strategy after the system is deployed, that is, when the system is in the application period. During the model period of the system, according to the urgency of the implementation project or according to the life cycle of the monitored object, the corresponding acquisition strategy is arranged. A large amount of data needs to be collected during the model period for model training. Data is continuously collected for 24 hours a day during the model period. If the project is not urgent, or the monitored object is a newly built infrastructure project, the data acquisition cycle is extended, and data is collected on an annual basis. The entire temperature cycle of the structure is covered on an annual basis. The data of each day is screened according to the distribution of three intervals. In the first interval 31, 30 units of data are selected per hour. In the second interval 32, 15 units of data are selected per hour. In the third interval 33, 10 units of data are selected per hour.

[0043] The data collected for each unit is Figure 1 the raw data 11 based on the acceleration data [A1, A2,..., A 2n and the initial temperature T within the unit acquisition time. The acceleration sequence is subjected to Fourier transform to generate 2n complex numbers, that is, a sequence of 2n (A + Bi). After taking the modulus of each (A + Bi) and dividing by 2n, that is, select the 1st to the (n + 1)th values of the sequence. For example, if the sequence length is 10, select the 1st to the 6th values. Multiply all the numbers except the first two numbers by 2, and finally obtain the acceleration amplitude corresponding to each frequency point [a1, a2,..., a n+1 .

[0044] The calculation method of the frequency point sequence corresponding to the acceleration amplitude is f s *(0:(n - 1) / (2n), that is, obtain n + 1 frequency points. That is, f1 = 0, where f s is the sampling frequency of the vibration sensor. Thus, n + 1 pairs of frequency point amplitudes are generated. After the above Fourier transform 21 processing, the acceleration time-domain data is transformed into frequency-domain data. The data structure form is as Figure 1 shown in the transformed data 12 as [(f1, a1, T),...,(f n+1 , a n+1, T)]. Exclude the acceleration amplitudes and frequency points corresponding to the frequency points outside the frequency range monitored by the sensitive unit, that is, filtering process 22. Divide the structure into several categories according to the design form and material of the structure. Determine the dynamic response range of each category. That is, each category of structure corresponds to a dynamic response frequency range, and exclude the acceleration amplitudes corresponding to the frequency points outside this frequency range. Finally, each type of structure corresponds to a filter. After filtering process 22, exclude some data of low frequency and high frequency to form filtered data 13, whose form is [(f1, a1, T),...,(f m , a m , T)], where the value of m is less than n + 1. Figure 6 The coordinate diagram 61 in is a schematic diagram of the time-domain acceleration. For convenient viewing, connect the discrete data; the coordinate diagram 62 is a schematic diagram of the frequency-domain acceleration after Fourier transform. For convenient viewing, connect the discrete data. It can be seen from the figure that extract the key information about frequency from the seemingly chaotic data.

[0045] After filtering process 22, the data also needs to be normalized in process 23. Normalize the frequency, amplitude, and temperature data and scale them to the range of [0, 1]. Frequency normalization: where f min is the minimum value among all frequency points, and f max is the maximum value among all frequency points. Amplitude normalization: where a min is the minimum value among all amplitudes, and a max is the maximum value among all amplitudes. Temperature normalization: where T min is the minimum value among all temperature data, and T max is the maximum value among all temperature data. After normalization process 23, organize the data into normalized data 14, whose form is [(f'1, a'1, T'),...,(f' m , a' m , T')], and the temperature values in each data unit are equal.

[0046] The model generation unit is responsible for training the model. After generating the trained model, output the model to the model application unit. The model generation unit uses the transformer architecture to process the normalized data 14. Mainly use the encoder module of the transformer architecture, such as Figure 4 the module 73 shown. Among them, the position encoding 731 uses sine-cosine position encoding, where pos is the position, i is the dimension index, and d is the embedding dimension. The positional encoding will provide relative position information within the frequency range. After the positional encoding 731 is ready, it enters the input embedding layer 732, which maps the sorted and normalized data 14 into a high-dimensional space: Embedding(f' i ,a' i ,T') = W · [f' i ,a' i ,T'] + b. Then the positional encoding is added to the embedding: E(f' i ,a' i ,T') = Embedding(f' i ,a' i ,T') + PE(pos). The Transformer encoder layer 733 is used, and each layer includes a multi-head attention mechanism, a feed-forward neural network, layer normalization, and residual connections. The multi-head self-attention mechanism is used to learn the relationships between frequency points, amplitudes, and temperatures. The feed-forward neural network performs feature extraction on each frequency-amplitude-temperature pair FEN(x) = ReLU(W1x + b1)W2 + b2. Normalization and residual connections help stabilize the training process. The encoder layers are stacked. According to the task requirements, usually 6 or more encoder layers are stacked, layer by layer, to extract features layer by layer. The output layer 734 is designed to predict the health score using a regression task, applying a linear layer y = W · h + b, and the output score 15 result is y as Figure 1 shown.

[0047] All the data collected in the early stage are healthy data, and the score results are all marked as 1. It is feasible to train the model only using healthy data, but it may lead to insufficient ability of the model to recognize abnormal states. To improve the generalization ability and accuracy of the model, it is necessary to include unhealthy data for training. The model needs to understand the characteristics in different states so that it can effectively identify abnormalities in actual monitoring. Using only healthy data may cause the model to overfit and be unable to effectively handle new and unseen abnormal situations. Based on a simulation method, finite element analysis is used, and a structural analysis software is used to simulate different types of damages such as cracks, corrosion, and fatigue damage of the measured structure, the cable-stayed bridge 41, to generate corresponding unhealthy frequency and amplitude data. The data output for cracks is scored 0, and the scores for corrosion and fatigue damage are 0.2. At the same time, data augmentation techniques are also used to generate unhealthy data, adding random noise to the healthy data to simulate possible interferences. The principle for the noise interference score is: the higher the noise intensity, the lower the score of the healthy data; the lower the frequency of the noise, the lower the score of the healthy data.

[0048] Now that there is healthy data and unhealthy data, the built module 73 can be used to train the model. First, the data is divided into a training set, a validation set, and a test set. The training set accounts for 80% of the data, and the validation set and the test set each account for 10% of the data. The mean squared error is used to calculate the loss function, which calculates the error between the predicted output of the model and the true label. The Adam optimizer is used for parameter updates. Dropout and L2 regularization are applied to prevent overfitting. The validation set is used to monitor the model performance and adjust the hyperparameters. The test set is used for the final evaluation, and metrics such as accuracy, precision, and recall are adopted. After the above process, the trained model is output. Thus, the model phase ends.

[0049] After the model phase ends, the application phase begins. The trained model is deployed to the real-time monitoring system 74 to analyze the structural health status of the cable-stayed bridge 41 in real time. The model can make rapid predictions based on new spectrum data and temperature information, thereby improving the timeliness and accuracy of monitoring. During the application phase, the output scores are classified into states. A score greater than or equal to 0.8 is the first state of the mode, and a score lower than 0.8 is the second state of the mode. The first state is the stable state, and the second state is the changing state. As Figure 6 shown in the 65 coordinate diagram, when the state continuously remains in the first state, it represents that the structure of the structure is normal, the dynamic characteristics of the structure have not changed, the structural stiffness of the monitored structural part and its associated parts has not changed, and the reliability of the connection of the structural part has not changed. During the first state, when the second state occasionally appears, it may be due to extreme abnormal conditions such as strong winds and earthquakes. When the second state persists, the dynamic characteristics of the structure change, that is, the structural stiffness of the structural part and its associated parts changes, and the connection of the structural part is abnormal, representing that the structure of the structure has structural abnormalities, such as cracks, fatigue, corrosion and other unhealthy states. The real-time monitoring system issues a warning to the mobile terminals such as the mobile phones of relevant managers after the second state continuously appears according to the score.

[0050] Figure 3 Schematically describes the device link of the entire system. Vibration sensors 42 are installed under the bridge deck of the cable-stayed bridge 41. After the vibration sensors 42 collect the data, the data is sent to a relay 43 near the cable-stayed bridge 41. The relay 43 is communicatively coupled to an access point 44 through an air interface. The air interface can be consistent with a cellular communication protocol, such as an Advanced Long-Term Evolution (LTE-A) protocol, a CDMA network protocol, a Push-to-Talk (PTT) protocol, a Push-to-Talk over Cellular (POC) protocol, Licensed-Assisted Access (LAA) based on LTE, a Universal Mobile Telecommunications System (UMTS) protocol, a 3GPP LTE protocol, a Global System for Mobile Communications (GSM) protocol, a 5G protocol, a New Radio (NR) protocol, a Non-Terrestrial Network (NTN) protocol, a Licensed-Assisted Access (LAA) based on NR, and / or any other communication protocol discussed herein. The access point 44 is coupled to an auxiliary frame 45 through an air interface.

[0051] The auxiliary framework 45 may include one or more graphics processors 55, one or more processors 56, power management 52, communication unit 53, display 54, and bus 57. The bus 57 represents all system buses, peripheral buses, and chipset buses that connect multiple internal devices of the auxiliary framework 45. In some specific implementations, the bus 57 connects one or more processors 56 to the memory 51 for communication. The processor 56 retrieves instructions and data from these memories 51 to perform relevant processing procedures. Depending on different specific implementations, the processor 56 can be a single-core processor or a multi-core processor. The memory 51 is used to store data and instructions required by the processor 56 and other electronic system modules. The processor 56 will retrieve instructions and data from these memories to perform specific processes. In addition, the bus 57 is also connected to the input interface 58 and the output interface 59. The input interface 58 allows users to transmit information and select commands to the auxiliary framework 45, and the input device can be an alphanumeric keyboard or a pointing device (cursor control device). The output interface 59 is used to display images generated by the auxiliary framework 45, and related output devices include the display 54, and may also include printers, indicator lights, projectors, or other devices that output information. Power management 52 plays a crucial role in the auxiliary framework 45, and its main functions include: Power distribution: Power management 52 is responsible for distributing power from a power source (such as a battery or a power adapter) to various parts of the system. It ensures that each module (such as a processor, memory, peripheral, etc.) obtains appropriate voltage and current to ensure normal operation. Voltage regulation: The PMU converts the input voltage into different required output voltages through built-in voltage regulators. These voltage levels are usually used to meet the requirements of different components and ensure that they operate at the optimal voltage. Power consumption management: Power management 52 can monitor and manage the power consumption of the system, and reduce power consumption by dynamically adjusting voltage and frequency. For example, when the system is in a low activity state, power management 52 can reduce the voltage and frequency of the processor to save energy. Standby and wake-up control: Power management 52 controls the standby mode and wake-up process of the system. In the standby mode, the system reduces power consumption, and when needed, the PMU quickly wakes up the system to ensure its quick response. Battery management: In portable devices, power management 52 is also responsible for battery charging and discharging management, ensuring that the battery operates within a safe range, extending battery life, and providing battery status monitoring. Thermal management: Power management 52 can monitor the temperature of the system and take corresponding measures in case of overheating, such as reducing the performance of the processor or shutting down certain modules to protect the safety and stability of the system. System monitoring and feedback: Power management 52 is usually equipped with sensors and feedback mechanisms to monitor parameters such as power status, power consumption, and temperature in real time and report this information to the processing unit for the system to make corresponding adjustments. Most of the training calculations of the model and the calculations for model applications after deployment are performed in the graphics processor 55.The graphics processing unit 55 has parallel computing capabilities: The graphics processing unit has thousands of processing cores and can simultaneously process a large number of computing tasks. This enables it to far outperform traditional central processing units in performing large-scale matrix operations and vector operations, and is particularly suitable for forward propagation and backpropagation in deep learning. The graphics processing unit 55 has high throughput: When training complex models, the graphics processing unit can process more data samples in parallel computing, improving the training efficiency. The graphics processing unit 55 has high memory bandwidth: The memory bandwidth of the graphics processing unit is usually higher than that of the central processing unit, which is very important for processing large amounts of data and high-dimensional features. Especially in deep learning, the model needs to frequently read and update weights. The graphics processing unit 55 is suitable for batch processing: The graphics processing unit can effectively process batch data, enabling multiple samples to be simultaneously transmitted into the network during the training process, further accelerating the training speed. The graphics processing unit 55 is suitable for large-scale data sets: Deep learning models often require a large amount of data for training. The graphics processing unit can process larger data sets, making the training process more efficient. The communication unit 53 can be coupled to the access point 44 through the air interface and send early warning data or evaluation data of the structure to the mobile device 46.

[0052] The above embodiments exemplarily describe the entire process of the recognition system from collecting data, processing data, training data to finally applying the model. The established system can be used to monitor the modal changes of the structure and understand the health status of the structure under test.

[0053] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article or device.

[0054] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A modality recognition system with strong generalization ability, characterized in that The system includes an identification framework and an auxiliary framework; The identification framework includes: a sensitive unit, a data acquisition strategy unit, a data preprocessing unit, a model generation unit, and a model application unit; The auxiliary framework is used to assist the identification framework in completing the modal identification work; the working time period of the system is divided into a model period and an application period; the state of the mode is divided into a first state and a second state, where the first state is a stable state and the second state is a changing state; the sensitive unit is used to detect the vibration acceleration and temperature of the measured structure; the data acquisition strategy unit determines the model period time period according to the fault cycle distribution of the measured structure; the data preprocessing unit performs Fourier transform and filtering on the acceleration data; the model generation unit uses a transformer model to learn from the training set and the validation set and generate a prediction model; the model application unit inputs the application period data into the model to output the modal state, and triggers an artificial detection reminder when the second state is continuously output.

2. The system according to claim 1, wherein The working frequency range of the sensitive unit is selected according to the inherent properties of the measured structure. The larger the working frequency range, the more types of structures can be monitored.

3. The system according to claim 1, wherein The data acquisition strategy unit classifies the time period according to the excitation density. The high-level time period increases the sampling rounds, the low-level time period reduces the sampling rounds, and random sampling is arranged in the non-sampling time period.

4. The system according to claim 1, wherein The model period is set at the initial stage of the structure life cycle and the duration covers the environmental temperature change cycle. The model period data is marked as the first state, and the training model is used for state prediction in the application period.

5. The system according to claim 1, characterized in that, The first state indicates that the structural stiffness and connection reliability have not changed, and the second state indicates that the structural stiffness or the connection has an abnormality.

6. The system according to claim 1, wherein After the vibration time-domain data is Fourier-transformed by the data preprocessing unit, the frequency point amplitudes in the non-feature frequency range are filtered according to the structure type to form a classification filter.

7. The system according to claim 1, characterized in that, The preprocessing unit combines the filtered frequency-domain data with the temperature data into three-dimensional data, extracts the feature values after dividing by the basic unit length, and performs normalization processing.

8. The system according to claim 1, wherein The transformer architecture of the model generation unit includes: an embedding layer that converts the three-dimensional data into an embedding vector sequence, a periodic position encoding layer that generates frequency point position features, a multi-layer encoder that stacks and processes the data, and a regression output layer that generates a modal state score in the range of 0-1.

9. The system according to claim 1, wherein The model application unit issues a warning signal when the modal score is continuously lower than the threshold.

10. The system according to claim 1, wherein The auxiliary framework includes a processor, a memory, a communication unit, a power management unit, and a graphics processor. The processor is responsible for data preprocessing and generating warning instructions. The graphics processor undertakes model training and application calculations. The memory stores data. The communication unit transmits warning signals.

Images