A steel structure monitoring data edge processing method, system, medium and equipment

Abstract

The present invention relates to the technical field of steel structure monitoring, and discloses a method, system, medium and equipment for edge processing of steel structure monitoring data, including: for spectral data of a single parameter, after noise reduction and smoothing processing, fitting is performed through different functions respectively to obtain a variety of fitting results, and the multiple fitting results are weighted by the entropy weight method. After the weighted fitting results are obtained, the center wavelength is identified by the relative position addressing algorithm; for spectral data of multiple parameters, a demodulation and identification algorithm based on a convolutional neural network is used to obtain the center wavelength; based on the center wavelength, multimodal data is calculated; multimodal data collected by electrical sensors is obtained, combined with the multimodal data obtained by demodulation of spectral data, abnormal data is diagnosed and repaired, and data compression is realized by an adaptive data compression algorithm. Not only the comprehensive performance and anti-interference ability of the demodulation process are enhanced; but also the data integrity and credibility are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of steel structure monitoring, and in particular to a method, system, medium and equipment for edge processing of steel structure monitoring data. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Large-scale steel structure health monitoring systems have many types of sensors and usually work in a harsh environment. They face practical challenges such as frequent abnormal and missing measurement data, huge data storage and transmission loads, and heavy cloud computing burdens.

[0004] At present, large-scale steel structure health monitoring usually uses multiple independent electrical sensors to monitor physical quantities such as strain, stress, temperature, and wind speed. As the monitoring requirements increase, large-scale steel structure health monitoring systems contain a large number of sensors of different types, facing the problems of difficult demodulation and low precision of complex multi-parameter spectral data, as well as frequent measurement data anomalies and missing, huge data storage and transmission load, and heavy cloud computing burden. Therefore, it is particularly important to pre-process a large amount of sensor data on the edge.

[0005] In the health monitoring of large steel structures, demodulation is the core link of fiber optic sensor signal processing, and its performance is crucial to data accuracy and system stability. Traditional fiber wavelength fitting methods are usually difficult to apply to multiple types of fiber grating sensors at the same time, and it is difficult to achieve a balance between accuracy, computational efficiency and robustness, especially in complex monitoring environments, which are easily affected by external interference and abnormal data.

[0006] In addition, the demodulated data is often affected by environmental interference or equipment limitations in practical applications, resulting in data loss, high noise and other problems, which poses a great challenge to subsequent analysis and processing. Traditional data processing methods often rely on a single rule, lack the comprehensive ability of multi-source heterogeneous data, and lack real-time and reliability, making it difficult to meet complex monitoring needs. Summary of the invention

[0007] In order to solve the above problems, the present invention provides a method, system, medium and equipment for edge processing of steel structure monitoring data, adopts different demodulation methods for single-parameter spectral data and multi-parameter spectral data, and proposes a demodulation strategy based on the entropy weight method for the demodulation of single-parameter spectral data. By adaptively allocating weights to the importance of evaluation indicators, while improving the fitting accuracy, the calculation efficiency is taken into account, and the comprehensive performance and anti-interference ability of the demodulation process are significantly enhanced; moreover, multi-sensor multi-source heterogeneous fusion data is used to realize real-time diagnosis and repair of data, thereby improving data integrity and credibility.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] A first aspect of the present invention provides a method for edge processing of steel structure monitoring data, comprising:

[0010] Acquire the spectral data of the steel structure transmitted by the optical fiber, wherein different types of fiber Bragg grating sensors are arranged at different key positions of the steel structure, and different types of fiber Bragg grating sensors are connected in different forms through optical fibers; for the spectral data of a single parameter, after noise reduction and smoothing, fit them respectively through different functions to obtain a variety of fitting results, weight the multiple fitting results through the entropy weight method, and after obtaining the weighted fitting results, identify the central wavelength through the relative position addressing algorithm; for the spectral data of multiple parameters, adopt a demodulation and recognition algorithm based on a convolutional neural network to obtain the central wavelength; based on the central wavelength, calculate and obtain multimodal data;

[0011] The multimodal data collected by electrical sensors is obtained and combined with the multimodal data obtained by spectral data demodulation. After abnormal data diagnosis and repair, data compression is achieved through an adaptive data compression algorithm.

[0012] Furthermore, the step of weighting the multiple fitting results includes:

[0013] For each spectral data, three fitting results were obtained by fitting with Lorentz, Gaussian and Voigt functions respectively;

[0014] Based on the fitting results, the evaluation indicators of each function are calculated respectively, and the evaluation indicator matrix is ​​obtained and normalized;

[0015] Based on the evaluation index matrix, the information entropy of each evaluation index is calculated, and based on the information entropy, the weight of each evaluation index is calculated;

[0016] Based on the weight, each evaluation index of each function is weighted to obtain the comprehensive score of each function;

[0017] Based on the comprehensive score of each function, calculate the weight of each function;

[0018] Determine whether the change value of the comprehensive score is within the set range. If so, obtain the weight of each function, weight the three fitting results, and output the weighted fitting result; otherwise, use the gradient descent method to adjust and optimize the parameters of the three fitting functions, and then return to refit the spectral data using the three functions.

[0019] Furthermore, the multimodal data obtained by demodulating the spectral data includes pressure, strain, temperature and vibration data.

[0020] Furthermore, the multimodal data collected by the electrical sensor includes: infrared images of multiple key parts collected by an infrared thermal imager, temperatures of multiple key parts collected by a thermometer, wind speeds of multiple key parts collected by an anemometer, and inclination angles of multiple key parts collected by an inclination sensor.

[0021] Furthermore, the abnormal data diagnosis step includes:

[0022] Each modal data is organized into a matrix form with the time dimension as the row and the space dimension as the column;

[0023] The matrix of each modality data is visualized in the time domain and frequency domain to generate grayscale images;

[0024] The time domain grayscale image and frequency domain grayscale image of each modality data are superimposed to form a multi-channel composite image;

[0025] Based on the multi-channel composite image, the location of abnormal data in each matrix is ​​obtained through a convolutional neural network with multi-channel input.

[0026] Furthermore, the abnormal data is repaired using a generative adversarial network, and the input of the generative adversarial network is a matrix in which the positions of the abnormal data are marked.

[0027] Furthermore, the adaptive data compression algorithm adopts a fusion algorithm of three models: long short-term memory network, β-total correlation variational autoencoder and lossless compression algorithm.

[0028] A second aspect of the present invention provides a steel structure monitoring data edge processing system, comprising:

[0029] The optical fiber edge processing module is configured to: obtain the spectral data of the steel structure transmitted by the optical fiber, wherein different types of optical fiber Bragg grating sensors are arranged at different key positions of the steel structure, and different types of optical fiber Bragg grating sensors are connected in different forms through optical fibers; for the spectral data of a single parameter, after noise reduction and smoothing processing, fit them respectively through different functions to obtain a variety of fitting results, weight the multiple fitting results through the entropy weight method, and after obtaining the weighted fitting results, identify the central wavelength through the relative position addressing algorithm; for the spectral data of multiple parameters, adopt a demodulation and recognition algorithm based on a convolutional neural network to obtain the central wavelength; based on the central wavelength, calculate and obtain multimodal data;

[0030] The multi-sensor multi-heterogeneous data edge processing module is configured to: obtain the multi-modal data collected by electrical sensors, combine it with the multi-modal data obtained by spectral data demodulation, diagnose and repair abnormal data, and then realize data compression through an adaptive data compression algorithm.

[0031] A third aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, the program being executed by a processor, and the program, when executed by the processor, implementing the steps in the above-mentioned method for edge processing of steel structure monitoring data.

[0032] A fourth aspect of the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the program, the steps in the edge processing method for steel structure monitoring data as described above are implemented.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention provides an edge processing method for steel structure monitoring data, which adopts different demodulation methods for single-parameter spectral data and multi-parameter spectral data. For the demodulation of single-parameter spectral data, a demodulation strategy based on the entropy weight method is proposed. By adaptively allocating weights to the importance of evaluation indicators, while improving the fitting accuracy, the calculation efficiency is taken into account, and the comprehensive performance and anti-interference ability of the demodulation process are significantly enhanced.

[0035] The present invention provides a steel structure monitoring data edge processing method, which adopts multi-sensor multi-source heterogeneous fusion data to achieve real-time data diagnosis and repair, and improve data integrity and credibility.

[0036] The present invention provides a steel structure monitoring data edge processing method, which reduces the transmission and storage pressure based on a multi-sensor data compression sampling algorithm based on time-space correlation, adapts to complex environments, and provides a reliable and efficient monitoring solution. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings, which constitute a part of the specification of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention, but do not constitute limitations of the present invention.

[0038] Figure 1 This is an architecture diagram of a steel structure monitoring data edge processing method according to the first embodiment of the present invention;

[0039] Figure 2 This is a convolutional neural network framework diagram of the first embodiment of the present invention;

[0040] Figure 3 This is a network framework diagram for abnormal data diagnosis and repair according to the first embodiment of the present invention;

[0041] Figure 4 This is a flow chart of an adaptive data compression algorithm according to the first embodiment of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0043] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0044] In the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other. The present invention is further described below with reference to the accompanying drawings and embodiments.

[0045] Embodiment 1

[0046] The purpose of this first embodiment is to provide a method for edge processing of steel structure monitoring data.

[0047] Aiming at the problem of multi-source heterogeneity of large steel structures, edge computing methods for preprocessing multi-source heterogeneous data are studied to disperse the computing burden of the cloud and reduce data transmission between "end-cloud". Focus on the high-precision demodulation method of optical fiber multi-parameter sensor array based on tunable ultra-narrow linewidth laser to solve the problem of high-precision, high-efficiency and accurate acquisition of large-scale optical fiber sensor data; study the abnormal diagnosis method of multi-source heterogeneous data for structural health monitoring, develop multi-dimensional feature mining methods such as data time domain, frequency domain and image representation, and analyze, diagnose and repair sensor abnormal data in real time; study the compression sampling technology of massive data under high-frequency sampling of structural health monitoring, establish a compressed sensing mechanism based on multi-sensor spatiotemporal correlation, and achieve efficient data compression and accurate reconstruction; finally, establish a real-time online intelligent processing technology system for perception data based on edge computing, and create a full-process edge data processing chain of multi-sensor "multi-parameter data demodulation-abnormal data diagnosis-missing data repair-redundant data compression" to provide sufficient, complete and accurate data foundation for health monitoring of large steel structures.

[0048] This embodiment provides a method for edge processing of steel structure monitoring data, further developing a multi-parameter and multi-functional optical fiber sensor to perform time-space synchronous integrated monitoring of multiple indicators such as structural vibration, ambient temperature, stress and strain, and adding an infrared thermal imaging sensor to monitor areas prone to leakage and moisture, thereby realizing the upgrade of health monitoring of large steel structures.

[0049] This embodiment provides a steel structure monitoring data edge processing method, such as Figure 1 As shown, edge computing exists at two levels:

[0050] First, the real-time and efficient demodulation processing of the multi-parameter and multi-functional optical fiber sensor's own data: a demodulation strategy based on the entropy weight method is proposed. By adaptively assigning weights to the importance of the evaluation indicators, the fitting accuracy is improved while taking into account the calculation efficiency, which significantly enhances the comprehensive performance and anti-interference ability of the demodulation process.

[0051] Second, after the multi-type and multi-number sensor data are collected, the abnormal data is repaired and compressed based on the related properties of multi-source heterogeneous data: This embodiment proposes to use multi-sensor multi-source heterogeneous fusion data, combining the advantages of CNN (convolutional neural network) and GAN (generative adversarial network) to achieve real-time data diagnosis and repair, improve data integrity and credibility, and at the same time, the multi-sensor data compression sampling algorithm based on spatiotemporal correlation reduces the transmission and storage pressure, adapts to complex environments, and provides reliable and efficient monitoring solutions.

[0052] According to the above ideas, this embodiment provides a steel structure monitoring data edge processing method, which revolves around "large-scale fiber Bragg grating data demodulation → abnormal data diagnosis and repair → redundant data compression".

[0053] This embodiment provides a steel structure monitoring data edge processing method, including:

[0054] Step 1: Large-scale fiber Bragg grating network demodulation.

[0055] In order to solve the problem of accurate data collection of multi-parameter and multi-functional fiber optic sensor in health monitoring of large-scale infrastructure projects, we study high-sampling and high-resolution large-scale fiber Bragg grating network demodulation technology, implement the demodulation idea around the reflection spectrum fitting algorithm, construct a demodulation processing flow including "noise reduction and smoothing processing → peak fitting algorithm → addressing algorithm positioning", and propose a demodulation recognition algorithm based on convolutional neural network. All of them are deployed in the fiber edge processing unit to demodulate the spectral data.

[0056] Step 101 : The optical fiber edge processing unit receives the spectral data of the steel structure transmitted by the optical fiber. If the spectral data only contains a single parameter, the process proceeds to step 102 ; if the spectral data contains multiple parameters, the process proceeds to step 103 .

[0057] The optical fiber is laid along the steel structure at key locations (e.g., stress-concentrated areas, specific beams or columns, or key connection nodes), and can be fixed to the surface or interior of the steel structure by bonding, tying, or embedding. For example, a highly durable adhesive is used to directly adhere the optical fiber to the surface of the steel structure, or a special clamp is used to ensure that the optical fiber is tightly combined with the steel structure, so that it can accurately capture changes in stress, strain, and other parameters.

[0058] In this embodiment, pressure fiber grating sensors are arranged in the stress-concentrated areas of the steel structure, and each pressure fiber grating sensor is transmitted to the fiber edge processing unit through a separate optical fiber. The spectral data about pressure received by the fiber edge processing unit is single parameter data, and the pressure data of the corresponding area can be obtained by demodulating the data.

[0059] In this embodiment, strain fiber grating sensors are arranged on specific beams or columns of the steel structure, and each strain fiber grating sensor is transmitted to the fiber edge processing unit through a separate optical fiber. The spectral data about strain received by the fiber edge processing unit is single parameter data, and the strain data of the corresponding beam or column can be obtained by demodulating the data.

[0060] In this embodiment, temperature grating sensors are set at multiple key nodes of the steel structure, and all the temperature grating sensors are connected in series through an optical fiber and then connected to a fiber edge processing unit. The spectral data about temperature received by the fiber edge processing unit is multi-parameter data, and the temperature data of all key nodes can be obtained by demodulating the data.

[0061] In this embodiment, vibration grating sensors are set at multiple key nodes of the steel structure, and all temperature grating sensors are connected in series through an optical fiber and then connected to a fiber edge processing unit. The spectral data about vibration received by the fiber edge processing unit is multi-parameter data, and the vibration data of all key nodes can be obtained by demodulating the data.

[0062] Step 102: For the spectral data of a single parameter, a demodulation process of "noise reduction and smoothing processing → peak fitting algorithm → addressing algorithm positioning" is adopted. The specific steps are as follows:

[0063] Step 1021: For the spectral data of a single parameter, perform noise reduction and smoothing processing through wavelet transform or Butterworth filter to obtain the spectral data after noise reduction.

[0064] Step 1022: The spectral data after noise reduction is fitted by the Lorentz, Gaussian and Voigt functions respectively to obtain fitting results; then, the fiber grating wavelengths obtained by fitting the three fitting functions (Lorentz, Gaussian and Voigt) are weighted by the entropy weight method, and the weights of the three fitting functions are adaptively assigned by using the size of information entropy, so as to improve the accuracy and precision of the fitting.

[0065] In spectral data fitting, the Lorentz, Gaussian, and Voigt functions each have different characteristics and are suitable for different spectral line shapes. Among them, the Lorentz function is suitable for describing spectral lines with wide line width and strong edges; the Gaussian function is suitable for describing spectra with less noise influence and central symmetry; the Voigt function combines the characteristics of the Lorentz and Gaussian functions and can handle more complex spectral line structures. Different types of fiber Bragg grating signals have different wavelength characteristics, and a single function may not be able to perfectly fit all signals.

[0066] Therefore, in this embodiment, the entropy weight method is introduced to select and optimize the weights among the Lorentz, Gaussian and Voigt functions, which can reduce subjective bias and improve the objectivity and robustness of spectral data fitting. The entropy weight method dynamically adjusts the combined weights of the Lorentz, Gaussian and Voigt functions according to the fitting index (such as fitting error, residual distribution), so that the fitting algorithm can adapt to different spectral characteristics and improve the generalization of fitting. The entropy weight method can automatically calculate the weights without manual adjustment of parameters, which is convenient for implementation in large-scale data processing or online processing.

[0067] Therefore, optimizing the weight selection of the fitting function by the entropy weight method can provide a more accurate and adaptable algorithm framework for spectral data fitting, effectively improving the accuracy and robustness of the fitting results.

[0068] The peak fitting algorithm of step 1022 has the following specific steps:

[0069] (1) Initialize the number of iterations t=1 and initialize the parameters of the Lorentz, Gaussian, and Voigt functions.

[0070] (2) For each denoised spectral data, three fitting functions, namely, Lorentz, Gaussian and Voigt functions, are used to fit the data, and three fitting results are obtained. Based on all the fitting results obtained by fitting the Lorentz, Gaussian and Voigt functions, various evaluation indicators are calculated to obtain the evaluation indicator matrix: , where R G , MSE G 、STD G , C G They represent the goodness of fit, mean square error, standard deviation and computational complexity of the Gaussian function respectively; R L , MSE L 、STD L , C L They represent the goodness of fit, mean square error, standard deviation and computational complexity of the Lorentz function respectively; R V , MSE V 、STD V , C V They represent the goodness of fit, mean square error, standard deviation and computational complexity of the Voigt function respectively.

[0071] The calculation formula for goodness of fit is: ; The calculation formula of mean square error is ; The formula for calculating the standard deviation is ; The computational complexity C refers to the number of free parameters in the fitting function; where, is the original data (the spectral data after denoising of the i-th spectrum), is the fitted data, and N is the number of spectral data.

[0072] (3) For each evaluation indicator, the sample data is standardized so that the value range of each evaluation indicator is the same, which is convenient for subsequent calculations. The positive indicator is used: , ; In the positive indicator formula, Refers to the x The value of the evaluation index on the y-th fitting function, that is, The data in the xth column and yth row, ; express The minimum value in column x of .

[0073] (4) For each evaluation indicator, calculate its information entropy. Information entropy reflects the discrete degree or uncertainty of the indicator. The calculation formula is as follows: In the formula, is the information entropy of the evaluation index x, and n is the number of fitting functions.

[0074] (5) Calculate the weight of each evaluation indicator. The information entropy weight obtained by calculating the information entropy value can determine the weight distribution of each evaluation indicator in the comprehensive evaluation. The weight calculation formula is as follows: ;in, Indicates x The weight of the evaluation index, Indicates x The information entropy of the evaluation indicators; through this formula, the relative importance of each evaluation indicator can be obtained, and the sum of the weights is 1.

[0075] (6) Comprehensive evaluation calculation. Calculate the comprehensive score of each fitting function S y : , m is the number of evaluation indicators x, where y=0, 1, 2 represent Lorentz L, Gauss G or Voigt function V respectively.

[0076] (7) Adjust the combination of the three fitting functions according to the weights calculated by the information entropy method F(x) , so that the loss function takes into account the comprehensive influence of different fitting functions. The weighted fitting result is as follows: , where are the weights of the Gaussian, Lorentzian and Voigt functions, respectively, , , and represent the fiber Bragg grating wavelengths obtained by Gaussian, Lorentzian and Voigt functions respectively.

[0077] (8) Determine the difference between the sum of the comprehensive scores of the three fitting functions obtained in the tth iteration and the t-1th iteration (comprehensive score change value). If the difference is within the set range, output the weighted fitting result. ; Otherwise, the gradient descent method is used to adjust and optimize the parameters of the three fitting functions, and after the number of iterations t is increased by 1, return to step (2).

[0078] According to the Gaussian function formula For example, the parameters that need to be adjusted and optimized include peak height , peak position and curve width .

[0079] Step 1023: Based on the weighted fitting result, the center wavelength is accurately identified through a relative position addressing algorithm.

[0080] Step 103: For the multi-parameter spectral data, a demodulation recognition algorithm based on a convolutional neural network is used to obtain the central wavelength.

[0081] In order to solve the problem of difficult demodulation and low precision of complex multi-parameter spectral data, a demodulation recognition algorithm based on convolutional neural network is studied, including:

[0082] (1) In the pre-training stage, the collected spectral data is preprocessed to prevent the demodulation performance from being weakened due to the inconsistency of the data distribution space, while improving the convergence speed, reliability and accuracy of the neural network model.

[0083] For example, Min-Max normalization is used to linearly transform the original data so that their normalized results fall into within the range.

[0084] (2) During the training phase, the normalized data (one-dimensional signal) is input into the neural network for training, and the training parameters are updated to establish a demodulation model.

[0085] like Figure 2 As shown in the figure, the neural network uses convolution and pooling operations to extract the features of the spectral signal, and obtains the recognition result, that is, the central wavelength position, through the Softmax (normalized exponential) function. Specifically, the neural network includes two convolutional layers connected in sequence, and each convolutional layer is connected to a pooling layer, and the last pooling layer is connected to a fully connected layer and a Softmax function in sequence.

[0086] (3) During the testing phase, the demodulation model is used to test the spectral signal and obtain the central wavelength.

[0087] Step 104: Based on the central wavelength, calculate and obtain the pressure, strain, temperature or vibration data of each key part at each moment.

[0088] It overcomes the difficulty of demodulating and low-precision spectral data from multi-parameter sensors, and enables accurate acquisition of optical fiber sensing data with different parameters, types, and frequencies.

[0089] Step 2: Diagnose and repair abnormal data.

[0090] During the health monitoring of large steel structures, the data collected by sensors will be affected by factors such as environmental changes and equipment aging, resulting in abnormal data. Abnormal data not only affects the accuracy of monitoring results, but may also lead to misjudgments and omissions. Therefore, how to effectively diagnose and repair abnormal data and ensure the integrity and reliability of data has become an urgent problem to be solved.

[0091] In traditional data anomaly diagnosis and repair, usually only a single type of data source is used. This embodiment introduces multimodal data, which can utilize complementary information of different sensor types, effectively reduce the impact of single-modal noise or sensor failure, enhance stability in complex environments, and improve the accuracy of anomaly detection and the reliability of repair.

[0092] In this embodiment, the pressure, strain, temperature and vibration data demodulated from the spectral data are combined with the infrared images of multiple key parts collected by the infrared thermal imager, the temperatures of multiple key parts collected by the thermometer, the wind speeds of multiple key parts collected by the anemometer, and the inclinations of multiple key parts collected by the inclination sensor; through the multimodal deep learning model, it is possible to analyze the correlation between different modes and capture abnormal features that cannot be identified in a single data modality.

[0093] like Figure 3 As shown in the figure, in order to solve the above problems, a method for abnormal data diagnosis and repair based on deep learning is proposed. The convolutional neural network (CNN) model is deployed through the edge computing node (i.e., the multi-sensor multi-source heterogeneous data edge processing module) to analyze and diagnose the abnormal state of sensor data in real time, and mark and process the abnormal data. It includes:

[0094] Step 201 : pre-process the raw data collected by different types of sensors, convert the multi-modal data into a standard format, and make the data of different modalities have a unified structure and range.

[0095] Specifically, the infrared image data will be represented by a gray value matrix; while the temperature, acceleration and other data will be represented in matrix form and mapped to the same value range through normalization.

[0096] Among them, the matrix representation of temperature, acceleration and other data is: ,in, represents the matrix of the ath modal data, Represents the data of the jth key part at the ith moment, the number of moments is m, and the number of key parts is n.

[0097] Step 202: visualize the matrix of each modal data in the time domain and frequency domain respectively to generate a grayscale image; then, superimpose the time domain and frequency domain grayscale images of each modality to form a multi-channel composite image, which can capture the joint features of multi-modal data and improve the accuracy and robustness of fault detection.

[0098] Assume that there are A types of modal data, each modality generates time domain and frequency domain images, and the shape of the input data is (height, width, A×2), where height and width represent the height and width of the image, respectively, in pixels; A×2 channels represent one channel each for the time domain and frequency domain images of each modality.

[0099] Step 203: construct a CNN model with multi-channel input to perform abnormal data training, so that the model can learn the feature description of multi-modal abnormal data. The trained CNN can automatically diagnose potential multiple abnormal data in the sensor data set, thereby improving the accuracy and automation level of abnormality detection.

[0100] Step 204: For the detected abnormal data, a generative adversarial network (GAN) is used to repair it.

[0101] In the GAN generator, the input is a complete data matrix containing outliers, but the locations of these outliers have been marked by the detection network. The generator will fill and repair according to the outlier locations to make the generated data consistent with the distribution of normal data. The discriminator compares the generated data (repaired data matrix) with the real data (normal value data matrix) to obtain the authenticity loss to distinguish whether the data is real, thereby guiding the generator to improve the quality of the data.

[0102] GAN consists of a generator and a multi-level discriminator. The generator is responsible for generating simulated data consistent with the normal data distribution, while the discriminator is used to detect local features and global features respectively to ensure the integrity of the generated data and the fidelity of the details, and to distinguish the generated data from the real data. Through iterative training, GAN can generate high-quality repair data, effectively compensate and correct abnormal data, and ensure the integrity and reliability of the final data.

[0103] On this basis, combined with the feedback data in the actual monitoring environment, the accuracy and efficiency of abnormal data detection and repair are continuously improved by optimizing and iterating the deep learning model. The overall abnormal data diagnosis and repair network framework is as follows: Figure 3As shown in Figure 1. First, the input data is processed by “time-frequency compounding” to generate a dual-channel image. This information is then input into the abnormal data detection network. The abnormal data detection network detects abnormal points in the data based on these image features and generates an abnormal labeling matrix. Figure 3 The abnormal data positions are shown as red marks in the figure, and then the abnormal marking matrix is ​​input into the GAN network to fill and repair the abnormal positions.

[0104] Step 3: Redundant data compression.

[0105] In the process of health monitoring of large steel structures, multi-source heterogeneous sensors generate a large amount of data every moment, which needs to be processed and analyzed in real time to ensure the safety and integrity of the structure. However, the network bandwidth resources in the IoT environment are limited, and the transmission of large amounts of data will cause bandwidth congestion, which in turn affects the real-time and reliability of the data; the storage resources of edge computing nodes are limited, and large amounts of historical data cannot be stored for a long time. These problems significantly affect the performance and efficiency of the monitoring system.

[0106] Faced with the problems of huge amounts of data from multi-source heterogeneous sensors, limited transmission bandwidth and restricted storage resources, this paper studies the massive data compression sampling technology under high-frequency sampling of structural health monitoring, and establishes a compressed sensing mechanism based on the spatiotemporal correlation of multiple sensors.

[0107] In this embodiment, an adaptive data compression algorithm is used to establish the spatiotemporal correlation of multi-sensor data, and the sensor data is predicted and encoded through a training model to achieve efficient data compression. The specific steps are as follows:

[0108] Assume that the matrix of the ath modal data is repaired and expressed as , Represents the repaired data of the j-th key part at the i-th moment.

[0109] Step 301: feature extraction and pattern recognition technology. First, for , extract features through linear discriminant analysis technology, and extract useful features from sensor data (m ≥ n), The feature of the repaired data of the jth key part at the i-th moment is expressed, which reduces the complexity of data processing in subsequent analysis; then, through pattern recognition techniques such as cluster analysis, the repeated patterns and trends in the data are identified, the repeated patterns in the data are detected, the redundant information in the data is eliminated, and the accuracy and effectiveness of the data are improved; finally, the noise filtering technology is applied to remove the high-frequency noise in the data through the adaptive filter, thereby improving the stability and reliability of the data, reducing data redundancy, and obtaining the data X ’ = , where s is the number of moments remaining after filtering.

[0110] Step 302, adaptive data compression algorithm. The adaptive data compression algorithm is deployed on the edge computing node, and uses a fusion algorithm based on LSTM (Long Short-Term Memory Network), β-TCVAE (β-Total Correlation Variational Autoencoder), and LZW (Lossless Compression Algorithm) to achieve hybrid data compression. The compression flow chart is as follows: Figure 4 As shown:

[0111] First, the data of each key point in the data X' (i.e., each row of data) is captured through the LSTM network (i.e., LSTM encoder, which contains several LSTM cells, i.e., long short-term memory units) to capture the long-term dependencies and correlations of multi-sensor data and generate lower-dimensional feature representations. H=LSTM(X j ’)= [ h 1 h 2 … h s ] to achieve preliminary data compression, where each h s is a d-dimensional matrix representing the hidden state at the s-th time step, X j ’ For data X ’ All data in row j of ;

[0112] Then, use β -TCVAE compressor maximizes the total correlation of latent variables and reduces redundant information in multi-sensor spatial data. Specifically: for each All the values ​​are output from the neural network through the encoder (the neural network encoder output layer has three layers, so a group of 3) {μ i , σ i}, thus generating the latent space representation Z=[ z 1 z 2 … z t ], (t≤s) to improve the data compression effect, z i The three latent variables {μ i , σ i}constitute;

[0113] Finally, the LZW encoding algorithm is used to detect the repeated patterns in the data, and Z is encoded and compressed, and the repeated patterns are replaced with dictionary indexes, thereby reducing the total size of the data and obtaining the final compression result L = [ l 1 l 2 … l s ].

[0114] The data before compression is lengthy, high-dimensional, noisy, and complex raw data, while the data after compression is streamlined, structured, non-redundant, and more informative "core data", which is suitable for subsequent fast calculations and real-time analysis. The compression result is L, l s Represents the LZW code after data compression at time s.

[0115] LSTM focuses on the long-term and short-term dependencies of time series and generates low-dimensional features in the time dimension; β -TCVAE focuses on eliminating redundancy in spatial data and compressing data through latent space representation, while LZW coding is based on the statistical properties of repeated patterns. These models have different objectives and characteristics, and direct fusion may lead to information loss or inconsistent optimization objectives.

[0116] In this embodiment, the time feature extraction and β -TCVAE’s spatial redundancy compression and LZW’s repetitive pattern encoding are performed sequentially, ensuring that the characteristics of each model are fully utilized in an independent stage, and the functions of each part complement each other when the final compression result is generated.

[0117] In addition, the data formats output by different models may be incompatible, especially the temporal features of LSTM output and β -TCVAE’s latent space features are difficult to directly interface with. β -Introduce a feature conversion layer between TCVAE to convert the time series output of LSTM into β - Input format for TCVAE.

[0118] For the coordination of training strategies, a joint loss function is designed to balance the prediction performance of LSTM and β -TCVAE's compression performance allows the two models to be optimized collaboratively to ensure the overall model performance. The loss function expression is as follows , where are the encoder and decoder parameters, is the input data, z is the latent space representation, They are the balance weight and constant offset for adjusting the Kullback-Leibler (KL, relative entropy) divergence, is the KL divergence, is the reconstruction loss, which represents the difference between the output generated by the model and the input.

[0119] Through this data compression strategy, efficient compression and transmission of multi-source heterogeneous sensor data are achieved, which significantly reduces the cost of data storage and transmission, improves the utilization efficiency of edge computing resources, and provides an effective solution for large-scale data processing in the edge computing environment of large steel structure monitoring.

[0120] In order to solve the problem of accurate data collection of optical fiber sensors in health monitoring systems of large-scale infrastructure projects, this embodiment studies high-sampling and high-resolution large-scale fiber Bragg grating network demodulation technology, implements the demodulation idea around the reflection spectrum fitting algorithm, and constructs a demodulation processing flow including "noise reduction and smoothing processing → peak fitting algorithm → addressing algorithm positioning"; and in order to solve the problem of difficult demodulation and low precision of complex multi-parameter spectral data, this embodiment studies a demodulation recognition algorithm based on convolutional neural network.

[0121] In view of the problem that sensor data is easily affected by environmental changes and equipment aging and becomes abnormal during the health monitoring of large steel structures, this embodiment studies the abnormal data diagnosis and repair technology based on deep learning. By introducing multimodal data and deep learning models, a data processing flow including "multimodal feature extraction → abnormal data diagnosis → multimodal association repair" is constructed to achieve efficient diagnosis and accurate repair of abnormal data in complex environments.

[0122] To address the problems of huge amount of data from multi-source heterogeneous sensors, much redundant information, and low transmission and storage efficiency, this embodiment studies a multi-sensor data compression sampling algorithm based on spatiotemporal correlation.

[0123] Embodiment 2

[0124] The purpose of this second embodiment is to provide a steel structure monitoring data edge processing system, including:

[0125] The optical fiber edge processing module is configured to: obtain the spectral data of the steel structure transmitted by the optical fiber, wherein different types of optical fiber Bragg grating sensors are arranged at different key positions of the steel structure, and different types of optical fiber Bragg grating sensors are connected in different forms through optical fibers; for the spectral data of a single parameter, after noise reduction and smoothing processing, fit them respectively through different functions to obtain a variety of fitting results, weight the multiple fitting results through the entropy weight method, and after obtaining the weighted fitting results, identify the central wavelength through the relative position addressing algorithm; for the spectral data of multiple parameters, adopt a demodulation and recognition algorithm based on a convolutional neural network to obtain the central wavelength; based on the central wavelength, calculate and obtain multimodal data;

[0126] The multi-sensor multi-heterogeneous data edge processing module is configured to: obtain the multi-modal data collected by electrical sensors, combine it with the multi-modal data obtained by spectral data demodulation, diagnose and repair abnormal data, and then realize data compression through an adaptive data compression algorithm.

[0127] It should be noted here that each module in this embodiment corresponds to each step in Example 1 one by one, and the specific implementation process is the same, which will not be repeated here.

[0128] Embodiment 3

[0129] This embodiment provides a computer-readable storage medium on which a computer program is stored. The program is executed by a processor. When the program is executed by the processor, the steps in the edge processing method of steel structure monitoring data described in the above-mentioned embodiment 1 are implemented.

[0130] Embodiment 4

[0131] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the program, the steps in the edge processing method of steel structure monitoring data as described in the above-mentioned embodiment 1 are implemented.

[0132] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

[0133] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A steel structure monitoring data edge processing method, characterized in that: include: Acquire the spectral data of the steel structure transmitted by the optical fiber, wherein different types of fiber Bragg grating sensors are arranged at different key positions of the steel structure, and different types of fiber Bragg grating sensors are connected in different forms through optical fibers; if the spectral data only contains a single parameter, the spectral data of the single parameter is subjected to noise reduction and smoothing processing, and then fitted by different functions to obtain a plurality of fitting results, and the plurality of fitting results are weighted by an entropy weight method, and after obtaining the weighted fitting results, the central wavelength is identified by a relative position addressing algorithm; If the spectral data contains multiple parameters, a demodulation recognition algorithm based on a convolutional neural network is used to obtain the central wavelength for the multi-parameter spectral data; Based on the central wavelength, multimodal data is calculated; The multimodal data collected by electrical sensors is obtained and combined with the multimodal data obtained by spectral data demodulation. After abnormal data diagnosis and repair, data compression is achieved through an adaptive data compression algorithm.

2. A steel structure monitoring data edge processing method as claimed in claim 1, characterized in that: The step of weighting the multiple fitting results comprises: For each spectral data, three fitting results were obtained by fitting with Lorentz, Gaussian and Voigt functions respectively; Based on the fitting results, the evaluation indicators of each function are calculated respectively, and the evaluation indicator matrix is ​​obtained and normalized; Based on the evaluation index matrix, the information entropy of each evaluation index is calculated, and based on the information entropy, the weight of each evaluation index is calculated; Based on the weight, each evaluation index of each function is weighted to obtain the comprehensive score of each function; Based on the comprehensive score of each function, calculate the weight of each function; Determine whether the change value of the comprehensive score is within the set range. If so, obtain the weight of each function, weight the three fitting results, and output the weighted fitting result; otherwise, use the gradient descent method to adjust and optimize the parameters of the three fitting functions, and then return to refit the spectral data using the three functions.

3. A steel structure monitoring data edge processing method as claimed in claim 1, characterized in that: The multimodal data obtained by demodulating the spectral data includes pressure, strain, temperature and vibration data.

4. A steel structure monitoring data edge processing method as claimed in claim 1, characterized in that: The multimodal data collected by the electrical sensor includes: infrared images of multiple key parts collected by an infrared thermal imager, temperatures of multiple key parts collected by a thermometer, wind speeds of multiple key parts collected by an anemometer, and inclination angles of multiple key parts collected by an inclination sensor.

5. The edge processing method for steel structure monitoring data according to claim 1, characterized in that: The abnormal data diagnosis step comprises: Each modal data is organized into a matrix form with the time dimension as the row and the space dimension as the column; The matrix of each modality data is visualized in the time domain and frequency domain to generate grayscale images; The time domain grayscale image and frequency domain grayscale image of each modality data are superimposed to form a multi-channel composite image; Based on the multi-channel composite image, the location of abnormal data in each matrix is ​​obtained through a convolutional neural network with multi-channel input.

6. A steel structure monitoring data edge processing method as claimed in claim 5, characterized in that: The abnormal data is repaired by using a generative adversarial network, and the input of the generative adversarial network is a matrix in which the positions of the abnormal data are marked.

7. A steel structure monitoring data edge processing method as claimed in claim 1, characterized in that: The adaptive data compression algorithm adopts a fusion algorithm of three models: long short-term memory network, beta-total correlation variational autoencoder and lossless compression algorithm.

8. A steel structure monitoring data edge processing system, characterized in that: include: The optical fiber edge processing module is configured to: obtain spectral data of a steel structure transmitted by an optical fiber, wherein different types of optical fiber Bragg grating sensors are arranged at different key positions of the steel structure, and different types of optical fiber Bragg grating sensors are connected in different forms through optical fibers; if the spectral data only contains a single parameter, the spectral data of the single parameter is subjected to noise reduction and smoothing processing, and then fitted by different functions to obtain a plurality of fitting results, and the plurality of fitting results are weighted by an entropy weight method, and after obtaining the weighted fitting results, the central wavelength is identified by a relative position addressing algorithm; If the spectral data contains multiple parameters, a demodulation recognition algorithm based on a convolutional neural network is used to obtain the central wavelength for the multi-parameter spectral data; Based on the central wavelength, multimodal data is calculated; The multi-sensor multi-heterogeneous data edge processing module is configured to: obtain the multi-modal data collected by electrical sensors, combine it with the multi-modal data obtained by spectral data demodulation, diagnose and repair abnormal data, and then realize data compression through an adaptive data compression algorithm.

9. A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor, characterized in that: When the program is executed by a processor, the steps in the edge processing method for steel structure monitoring data as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps in the edge processing method for steel structure monitoring data as described in any one of claims 1-7 are implemented.

Images