Data analysis method and device applied to large-span rigid structure

Abstract

The invention relates to a data analysis method and device applied to a large-span rigid structure. The method comprises the steps that a sensor data set captured by multiple types of sensors in the large-span rigid structure is acquired; dividing the sensor data set into a first sensor data subset and a second sensor data subset according to a sensor type corresponding to each piece of sensor data in the sensor data set; extracting data features of each piece of sensor data in the first sensor data subset, and generating a fusion feature vector through a feature fusion technology; correcting the fusion feature vector according to the sensor data in the second sensor data subset to obtain a target feature vector; and determining the damage degree of the large-span rigid structure according to the target feature vector. The high-dimensional features are automatically extracted from the time sequence data of the sensor through the convolutional feature extraction network, the workload of manual feature selection is reduced, and the representativeness and reliability of the features are improved.

Description

Technical Field

[0001] The present application relates to the field of data processing technology, and in particular to a data analysis method and device applied to large-span rigid structures. Background Art

[0002] At present, the health monitoring and damage identification technology of large-span steel structures mainly relies on various types of sensors to collect the response data of the structure under different conditions, and then analyze its health status; in order to extract valuable information from massive amounts of raw data, the commonly used methods are model-based methods. This method establishes a dynamic model of the structure and combines it with actual measured data to identify changes in model parameters and then infer the damage of the structure.

[0003] However, existing monitoring methods have several significant limitations. In practical applications, structural health monitoring data usually contains a lot of noise and redundant information. Traditional feature selection methods often require manual intervention to select appropriate feature combinations, which is not only time-consuming and labor-intensive, but also prone to missing important feature information. Moreover, existing technologies are insufficient in processing multimodal data. Health monitoring of large-span steel structures usually involves multiple different types of data sources, such as strain, acceleration, displacement, and temperature. Each data source can provide a different perspective on the state of the structure, but existing methods can often only process a certain type of data separately, or simply splice data of different modes together for analysis, lacking an effective data fusion mechanism.

[0004] Therefore, there is an urgent need to develop a data analysis method and device for large-span rigid structures to solve one or more of the above-mentioned problems. Summary of the invention

[0005] In view of this, in order to solve the above technical problems or part of the technical problems, an embodiment of the present invention provides a data analysis method and device applied to a large-span rigid structure.

[0006] In a first aspect, the present application provides a data analysis method applied to a large-span rigid structure, the method comprising:

[0007] Obtain a set of sensor data captured by multiple types of sensors in a long-span rigid structure;

[0008] Dividing the sensor data set into a first sensor data subset and a second sensor data subset according to a sensor type corresponding to each sensor data in the sensor data set;

[0009] Extracting data features of each sensor data in the first sensor data subset, and generating a fused feature vector by a feature fusion technology;

[0010] Correcting the fused feature vector according to the sensor data in the second sensor data subset to obtain a target feature vector;

[0011] The damage degree of the long-span rigid structure is determined according to the target eigenvector.

[0012] In a possible implementation, extracting data features of each sensor data in the first sensor data subset includes:

[0013] Inputting the sensor data in the first sensor data subset into a pre-trained convolutional neural network to obtain an initial feature map of each sensor data;

[0014] Perform a multi-scale convolution operation on the initial feature map of each sensor data to obtain a multi-scale feature map set of each sensor data;

[0015] Perform a corresponding pooling operation on each scale feature map in the multi-scale feature map set of each sensor data to obtain a multi-scale feature vector set of each sensor data;

[0016] The multi-scale feature vector set of each sensor data is processed by feature concatenation to obtain the data features of each sensor data.

[0017] In a possible implementation, before performing a corresponding pooling operation on each scale feature map in the multi-scale feature map set of each sensor data, the method further includes:

[0018] A feature map of a preset size in a multi-scale feature map set is input into a feature enhancement model for processing; the feature enhancement model includes at least one hole convolution layer.

[0019] In a possible implementation, before performing feature splicing processing on the multi-scale feature vector set of each sensor data, the method further includes:

[0020] Upsampling is performed on the feature vectors in the multi-scale feature vector set so that the vector length of the feature vector after processing meets a preset length requirement.

[0021] In a possible implementation, the generating a fused feature vector by using a feature fusion technology includes:

[0022] Obtaining a spatial resolution of a data feature corresponding to each sensor data in the first sensor data subset;

[0023] Determine the standard resolution based on the spatial resolution of the data features corresponding to each sensor data;

[0024] The spatial resolution of the data feature corresponding to each sensor is adjusted to the standard resolution in turn to obtain the target data feature of each sensor;

[0025] Obtaining a feature weight of each sensor data in the first sensor data subset;

[0026] Based on the feature weights, feature fusion is performed on the target data features of all sensor data in the first sensor data subset to obtain a fused feature vector.

[0027] In a possible implementation manner, obtaining a feature weight of each sensor data in the first sensor data subset includes:

[0028] Obtain a weight parameter of each sensor type corresponding to the first sensor subset;

[0029] A feature weight of each sensor data in the first sensor data subset is determined according to the weight parameter of each sensor type.

[0030] In a possible implementation, the calculation formula of the feature vector is as follows:

[0031]

[0032] Among them, α i is the feature weight of the i-th sensor data, λ i is the weight parameter of the i-th sensor type, and N is the total number of sensor types.

[0033] In a possible implementation, the correcting the fused feature vector according to the sensor data in the second sensor data subset to obtain the target feature vector includes:

[0034] Obtaining environmental parameters according to the sensor data in the second sensor data subset;

[0035] Establishing a correction matrix based on the environmental parameters;

[0036] The correction matrix is ​​used to correct the fused features to obtain a target feature vector.

[0037] In a possible implementation manner, determining the damage degree of the large-span rigid structure according to the target characteristic vector includes:

[0038] Retrieving a damage degree assessment model for the large-span rigid structure;

[0039] The target feature vector is input into the damage degree assessment model to obtain the damage degree of the long-span rigid structure.

[0040] In a second aspect, the present application provides a data analysis device applied to a large-span rigid structure, the device comprising:

[0041] An acquisition module is used to acquire a sensor data set captured by multiple types of sensors in a large-span rigid structure;

[0042] a partitioning module, configured to partition the sensor data set into a first sensor data subset and a second sensor data subset according to a sensor type corresponding to each sensor data in the sensor data set;

[0043] An extraction module, used for extracting data features of each sensor data in the first sensor data subset, and generating a fused feature vector by a feature fusion technology;

[0044] A correction module, used for correcting the fused feature vector according to the sensor data in the second sensor data subset to obtain a target feature vector;

[0045] A determination module is used to determine the damage degree of the long-span rigid structure according to the target characteristic vector.

[0046] The above-mentioned technical solution provided by the embodiment of the present application has the following advantages over the prior art: the method provided by the embodiment of the present application automatically extracts high-dimensional features from the time series data of the sensor through a convolutional feature extraction network, reduces the workload of manual feature selection, and improves the representativeness and reliability of the features; and automatically generates a fused feature vector to realize the correction of the fused feature vector, which can flexibly cope with the response differences of large-span steel structures at different positions and improve the accuracy and reliability of data fusion. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0049] One or more embodiments are exemplarily described by pictures in the corresponding drawings, and these exemplified descriptions do not constitute limitations on the embodiments. Elements with the same reference numerals in the drawings represent similar elements, and unless otherwise stated, the figures in the drawings do not constitute proportional limitations.

[0050] Figure 1A schematic flow chart of a data analysis method applied to a large-span rigid structure provided in an embodiment of the present application;

[0051] Figure 2 A schematic flow chart of another data analysis method applied to a large-span rigid structure provided in an embodiment of the present application;

[0052] Figure 3 A flow chart of another data analysis method applied to a large-span rigid structure provided in an embodiment of the present application;

[0053] Figure 4 A schematic diagram of the steps of a data analysis method applied to a large-span rigid structure provided in an embodiment of the present application;

[0054] Figure 5 A structural schematic diagram of a data analysis device applied to a large-span rigid structure provided in an embodiment of the present application. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0056] The disclosure below provides many different embodiments or examples to implement different structures of the present invention. In order to simplify the disclosure of the present invention, the parts and settings of specific examples are described below. Of course, they are only examples, and the purpose is not to limit the present invention. In addition, the present invention can repeat reference numbers and / or letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed.

[0057] The health monitoring and damage identification technology of large-span steel structures mainly relies on various types of sensors to collect the response data of the structure under different conditions and then analyze its health status. Common sensor types include but are not limited to strain sensors, acceleration sensors, displacement sensors, and temperature sensors. The core of these technologies lies in how to extract useful feature information from a large amount of monitoring data to support subsequent analysis and decision-making processes.

[0058] In order to extract valuable information from massive amounts of raw data, the commonly used methods currently include model-based methods, which establish a dynamic model of the structure and combine it with actual measured data to identify changes in model parameters and then infer the damage of the structure. For example, modal analysis is a widely used model-based method that identifies the location and extent of damage by comparing changes in modal parameters such as the natural frequency and vibration mode of the structure before and after damage. However, this method has high requirements for the accuracy of the model, and it is difficult to establish an accurate mathematical model for complex large structures.

[0059] Another important method is based on signal processing technology, which mainly uses various digital signal processing algorithms to extract features. Common algorithms include principal component analysis (PCA), wavelet transform, Fourier transform, etc. PCA reduces the redundancy of data through dimensionality reduction technology while maintaining the main features of the data, and is suitable for processing high-dimensional data sets. Wavelet transform can provide time and frequency information of the signal, which is suitable for the analysis of non-stationary signals and can effectively capture transient characteristics in structural response. Fourier transform converts time domain data into frequency domain data, which helps to identify periodic signal characteristics.

[0060] However, most existing methods rely on pre-defined features and do not work well for heterogeneous data fusion, so continuous innovation and improvement are needed.

[0061] In addition, existing technologies are insufficient in processing multimodal data. Health monitoring of large-span steel structures usually involves multiple different types of data sources, such as strain, acceleration, displacement, and temperature. Each data source can provide a different perspective on the state of the structure, but existing methods can often only process a certain type of data separately, or simply splice data from different modes together for analysis, lacking an effective data fusion mechanism.

[0062] In order to solve several significant limitations in the prior art, in practical applications, structural health monitoring data usually contains a large amount of noise and redundant information, and feature selection methods often require manual intervention. Selecting a suitable feature combination is not only time-consuming and labor-intensive, but also easy to miss important feature information; the isolation and waste of information cannot fully utilize the complementarity between multimodal data, thereby affecting the overall monitoring effect. Technical problems, the present application provides a data analysis method and device for large-span rigid structures, which focuses on solving the shortcomings of the prior art in adaptive feature extraction and multimodal data analysis fusion, and improves the comprehensiveness and accuracy of feature extraction and enhances the processing capabilities of multimodal data by introducing advanced convolutional feature extraction networks (CFEN) and adaptive spatial feature fusion (ASFF) methods.

[0063] Figure 1 A flow chart of a data analysis method for a large-span rigid structure provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method specifically includes:

[0064] S101, obtaining a sensor data set captured by multiple types of sensors in a large-span rigid structure;

[0065] In this embodiment, strain sensors, acceleration sensors, displacement sensors, temperature sensors and wind speed sensors are deployed at key locations of the large-span steel structure to collect data in real time. The data from these sensors reflect the local stress state, overall dynamic response, displacement change, temperature change and wind speed change of the structure respectively; the data sets collected by the sensors are collected and integrated from various key locations to ensure that the operating status of the system can be fully monitored and analyzed.

[0066] S102, dividing the sensor data set into a first sensor data subset and a second sensor data subset according to a sensor type corresponding to each sensor data in the sensor data set;

[0067] In this embodiment, the strain sensor, acceleration sensor, and displacement sensor among the multiple types of sensors are regarded as the first type of sensors, and the temperature sensor and wind speed sensor are regarded as the second type of sensors.

[0068] Strain sensors, acceleration sensors, and displacement sensors used to measure the physical deformation and motion state of an object are classified into one category of sensors, and temperature sensors and wind speed sensors used to monitor the temperature changes and wind speed conditions of the environment are classified into another category of sensors;

[0069] Furthermore, the multimodal data of strain sensors, acceleration sensors, and displacement sensors are used for feature extraction, and the temperature sensor and wind speed sensor data will participate in the environmental correction calculation.

[0070] S103, extracting data features of each sensor data in the first sensor data subset, and generating a fused feature vector by using a feature fusion technology;

[0071] In this embodiment, data features provided by each sensor are extracted from the first sensor data subset, and then, feature fusion technology is used to integrate these scattered feature information into a comprehensive feature vector. The fused feature vector can more comprehensively reflect the overall characteristics of the first sensor data subset, and provide richer and more accurate information for subsequent data processing and analysis.

[0072] S104, correcting the fused feature vector according to the sensor data in the second sensor data subset to obtain a target feature vector;

[0073] In this embodiment, the environmental characteristics of the large-span rigid structure are extracted from the second sensor data subset, and the characteristic vector of the large-span rigid structure is corrected using the environmental characteristics to obtain an accurate target characteristic vector after correction. Through this correction, the accuracy of the data analysis results can be ensured, thereby providing reliable data support for the health monitoring and maintenance of the large-span rigid structure, which not only improves the efficiency of data processing, but also enhances the accuracy of the status assessment of the large-span rigid structure, and provides a strong technical guarantee for engineering safety.

[0074] S105. Determine the damage degree of the long-span rigid structure according to the target eigenvector.

[0075] In this embodiment, the damage degree of the large-span rigid structure is evaluated through in-depth analysis of the target feature vector; in the evaluation process, the evaluation model is generally used to perform quantitative analysis of the damage degree. The evaluation model is usually based on a machine learning algorithm, which can learn and identify characteristic patterns under different damage states based on historical data. By comparing the target feature vector with the damage pattern stored in the model, the current damage level of the large-span rigid structure can be accurately determined.

[0076] The data analysis method applied to the large-span rigid structure provided in the embodiment of the present application obtains a sensor data set captured by the sensor in the large-span rigid structure, and then uses a type of data in the sensor data set to determine the data characteristics of the large-span rigid structure, and then corrects the data characteristics based on the second type of data in the sensor data set to obtain a target feature vector, and finally analyzes the target feature vector to determine the degree of damage to the large-span rigid structure.

[0077] In the monitoring process of large-span rigid structures, data sets from sensors are first collected. These data sets contain data information collected by various types of sensors. By analyzing these data, the key data features of large-span rigid structures are determined. Then, these features are corrected and optimized using another type of data in the data set to obtain more accurate target feature vectors. The target feature vectors can accurately reflect the true condition of the structure. Finally, through in-depth analysis of the target feature vectors, the degree of damage to the large-span rigid structure is evaluated and determined, thereby providing a data basis for the maintenance of the structure.

[0078] Figure 2 A flow chart of another data analysis method for a large-span rigid structure provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the method specifically includes:

[0079] S201, inputting the sensor data in the first sensor data subset into a pre-trained convolutional neural network to obtain an initial feature map of each sensor data;

[0080] Convolutional feature extraction network is a deep learning model used to extract features from graphs or time series data. Through multi-scale convolution kernels, dilated convolutions, and pyramid pooling, CFEN can effectively capture multi-scale features and global information in time series, improving the accuracy and robustness of feature extraction; the application of CFEN can help large-span steel structures identify potential structural fault characteristics in health monitoring.

[0081] In this embodiment, the monitoring data of each sensor in the first group of sensor data subset, i.e., the multimodal data of strain sensors, acceleration sensors, and displacement sensors, is input into a pre-trained convolutional neural network model for feature extraction to obtain an initial feature map corresponding to each individual sensor data.

[0082] S202, performing a multi-scale convolution operation on an initial feature map of each sensor data to obtain a multi-scale feature map set of each sensor data;

[0083] Multi-Scale Convolutional Kernels. Use convolution kernels of different sizes in the same layer to extract features of different scales of time series, improving the richness and diversity of features. Convolution operation:

[0084] F k =K k *X

[0085] Among them, F k It is a convolution kernel of size k*1, where k can be 3, 5, 7, etc.

[0086] In this embodiment, when processing each sensor data, it is first necessary to perform a series of multi-scale convolution operations on the initial feature map of each sensor data to obtain a multi-scale feature map set of each sensor data. The feature map set can capture information at different scales, thereby providing a richer and more detailed feature description for subsequent data analysis and processing.

[0087] For example, we obtain the input time series X, taking the shape (100,3) as an example, which represents the data collected from multiple sensors. We use multi-scale convolution kernels to perform convolution operations using 3*1, 5*1 and 7*1 convolution kernels respectively to obtain the feature map

[0088] F3=K3*X

[0089] F5=K5*X

[0090] F7=K7*X

[0091] Concatenate these feature maps in the channel dimension:

[0092] F multi =concat(F3,F5,F7)

[0093] S203, performing a corresponding pooling operation on each scale feature map in the multi-scale feature map set of each sensor data to obtain a multi-scale feature vector set of each sensor data;

[0094] Pyramid Pooling: Pooling is performed at different scales and then the results are merged to enhance the perception of global information and help capture the features of different time scales in the time series. Pooling operation:

[0095] P l =Pool(F dilated ,s l )

[0096] Among them, s l is the pooling window size of layer l.

[0097] In this embodiment, feature maps of different scales are subjected to dimensionality reduction processing through pooling technology to extract key information and obtain a set of multi-scale feature vector sets representing each sensor data. The multi-scale feature vector set can effectively capture the feature information of the data at different scales.

[0098] For example, for F dilated Pooling operations are performed at different scales, with pooling window sizes of 10, 20, and 50, respectively, to obtain the pooling results:

[0099] P1=Pool(Fdilated ,10)

[0100] P2=Pool(F dilated ,20)

[0101] P3=Pool(F dilated ,50)

[0102] S204 , performing feature concatenation processing on the multi-scale feature vector set of each sensor data to obtain data features of each sensor data.

[0103] In this embodiment, these feature vectors of different scales are comprehensively processed and integrated together through feature splicing to obtain a comprehensive and rich data feature representation, which not only contains the detailed information of the original data, but also integrates high-level abstract features, and can more accurately capture the key information in the sensor data, providing a solid foundation for subsequent data analysis and processing.

[0104] For example, these pooling results are upsampled to restore to the same level as F dilated Same length:

[0105] P'1=Upsample(P1,100)

[0106] P'2=Upsample(P1,100)

[0107] P'3=Upsample(P1,100)

[0108] Concatenate in the channel dimension:

[0109] F final =concat(F dilated ,P'1,P'2,P'3)

[0110] The data analysis for large-span rigid structures provided in the embodiment of the present application first inputs each sensor data in a first group of sensor data subsets into a convolutional neural network, and extracts an initial feature map of each sensor data through the convolutional neural network; then, in order to capture features at different scales, multi-scale convolution processing is performed on the initial feature map to obtain a multi-scale feature map set for each sensor data; after obtaining the multi-scale feature map set, a pooling operation is performed on each size feature map therein to obtain a set of multi-scale feature vector sets; finally, the multi-scale feature vector sets of each sensor data are spliced ​​to form a comprehensive data feature of each sensor data, thereby providing strong data support for fault detection of large-span rigid structures.

[0111] In an optional solution of the embodiment of the present invention, before performing a corresponding pooling operation on each scale feature map in the multi-scale feature map set of each sensor data, the method further includes:

[0112] A feature map of a preset size in a multi-scale feature map set is input into a feature enhancement model for processing; the feature enhancement model includes at least one hole convolution layer.

[0113] The dilated convolution increases the spacing between convolution kernels and expands the receptive field, thereby capturing features of a longer time span, which helps capture long-term dependencies in time series.

[0114] F dilated =K dilated *F multi

[0115] Among them, K dilated It is a 3*1 convolution kernel with a dilation rate of d.

[0116] In this embodiment, a feature map of a preset size in a set of feature maps of different scales is input into a feature enhancement model for further processing; the feature enhancement model includes at least one dilated convolutional layer to achieve deep learning processing and feature extraction of the feature map.

[0117] For example, for F multi Use a 3*1 hole convolution kernel with a dilation rate of 2 to perform a convolution operation and obtain the feature map F dilated :

[0118] F dilated =K dilated *F multi

[0119] In an optional solution of the embodiment of the present invention, before performing feature splicing processing on the multi-scale feature vector set of each sensor data, the method further includes:

[0120] Upsampling is performed on the feature vectors in the multi-scale feature vector set so that the vector length of the feature vector after processing meets a preset length requirement.

[0121] In this embodiment, when processing a multi-scale feature vector set, an upsampling operation is first performed on the feature vectors to adjust the length of the feature vectors to ensure that after the upsampling process, the length of each feature vector can reach a preset standard length requirement. In this way, the consistency and accuracy of the data in the subsequent processing steps are guaranteed, thereby improving the performance and efficiency of the entire system.

[0122] For example, these pooling results are upsampled to restore to the same level as Fdilated Same length:

[0123] P'1=Upsample(P1,100)

[0124] P'2=Upsample(P1,100)

[0125] P'3=Upsample(P1,100)

[0126] The adaptive spatial feature fusion (ASFF) strategy mainly consists of three parts: feature extraction, feature adjustment, and adaptive weight assignment. First, high-dimensional features are extracted from the time series data of strain sensors, accelerometers, and displacement sensors through a convolutional feature extraction network (CFEN). Then, the extracted feature vectors are adjusted to the same dimension to ensure their compatibility before fusion. Finally, the weights of each feature vector are dynamically adjusted through a learning mechanism to achieve adaptive fusion and generate a comprehensive feature vector.

[0127] Figure 3 A flow chart of another data analysis method for a large-span rigid structure provided in an embodiment of the present application is shown in FIG. Figure 3 As shown, the method specifically includes:

[0128] S301, obtaining the spatial resolution of the data feature corresponding to each sensor data in the first sensor data subset;

[0129] In this embodiment, the spatial resolution refers to how small a spatial unit the sensor data can distinguish. In this embodiment, the spatial resolution of the data features corresponding to each sensor data in the first sensor data subset is obtained to better understand the details of the sensor data, thereby providing a more solid foundation for subsequent data processing and decision-making.

[0130] S302, determining a standard resolution based on the spatial resolution of the data feature corresponding to each sensor data;

[0131] In this embodiment, after obtaining the spatial resolution of the data feature corresponding to each sensor data, since the spatial resolution is not uniform, the spatial resolution of the data feature corresponding to each sensor data needs to be adjusted to the same spatial resolution; therefore, a standard resolution is determined for the spatial resolution of the data feature corresponding to each sensor data.

[0132] Feature adjustment can enhance the model's perception of structural information at different scales, improve the detection accuracy of small and large-scale damage, and adjust feature maps of different resolutions to the same resolution to facilitate subsequent fusion.

[0133] S303, sequentially adjusting the spatial resolution of the data feature corresponding to each sensor to the standard resolution, to obtain the target data feature of each sensor;

[0134] In this embodiment, the spatial resolution of the data features corresponding to each sensor is adjusted to ensure that the spatial resolution of the data features reaches a unified standard resolution. Through this adjustment process, the target data features corresponding to each sensor are obtained, laying a solid foundation for subsequent data processing and analysis.

[0135] S304, obtaining a feature weight of each sensor data in the first sensor data subset;

[0136] S305 . Based on the feature weights, perform feature fusion on target data features of all sensor data in the first sensor data subset to obtain a fused feature vector.

[0137] In this embodiment, the feature weight of each sensor data in the first sensor data subset is first obtained to evaluate the importance of the data, thereby providing a basis for subsequent data fusion. Then, based on the feature weight calculated previously, the target data features of all sensor data in the first sensor data subset are feature fused, and the data features of different sensors are integrated together to form a comprehensive fusion feature vector, which helps to understand and analyze the data more comprehensively.

[0138] For example, let x ∈ 、x a 、x d They represent the data of strain sensor, acceleration sensor and displacement sensor respectively. The output feature of the feature extraction network is f ∈ 、f a 、f d :

[0139] f ∈ =CFEN(x ∈ )

[0140] f a =CFEN(x a )

[0141] f d =CFEN(x d )

[0142] The feature maps of different resolutions are adjusted to the same spatial resolution. The adjusted feature maps are f′ ∈ ,f′ a ,f′ d :

[0143] f′ ∈ =Resize(f∈ )

[0144] f' a =Resize(f a )

[0145] f' d =Resize(f d )

[0146] At each spatial location, different types of features are adaptively fused using learnable weights.

[0147] The final fused feature vector f fusion Calculated by the following formula:

[0148] f fusion (i,j)=α ij ·f′ ∈ (i,j)+β ij ·f′ a (i,j)+γ ij ·f′ d (i,j)

[0149] The data analysis method adaptive spatial feature fusion (ASFF) strategy applied to large-span rigid structures provided in the embodiment of the present application mainly consists of three parts: feature extraction, feature adjustment and adaptive weight allocation. First, high-dimensional features are extracted from the time series data of strain sensors, acceleration sensors and displacement sensors through a convolutional feature extraction network (CFEN); then, the extracted feature vectors are adjusted to the same dimension to ensure their compatibility before fusion. Finally, the weights of each feature vector are dynamically adjusted through a learning mechanism to achieve adaptive fusion and generate a comprehensive feature vector.

[0150] In an optional solution of the embodiment of the present invention, the acquiring the feature weight of each sensor data in the first sensor data subset includes:

[0151] Obtain a weight parameter of each sensor type corresponding to the first sensor subset;

[0152] A feature weight of each sensor data in the first sensor data subset is determined according to the weight parameter of each sensor type.

[0153] In this embodiment, adaptive weight allocation is used. According to the importance and relevance of different sensor data, the weight of feature fusion is dynamically adjusted, so that the model can better adapt to different monitoring scenarios. The weight of each feature map is calculated through a learning mechanism to ensure that important features occupy a larger proportion in the fusion process.

[0154] Specifically, it is necessary to obtain weight parameters of each sensor type corresponding to the first sensor subset; then, based on the obtained weight parameters of each sensor type, further determine the feature weight of each sensor data in the first sensor data subset.

[0155] Exemplarily, the weights are determined through a training process to maximize the effectiveness of the features. ij , β ij , γ ij They represent the weights of strain feature, acceleration feature and displacement feature at position (i, j) respectively and are calculated by the following formula:

[0156]

[0157] Among them, λ α (i,j),λ β (i,j),λ γ (i,j) is the weight scalar map learned from the resized feature map through a 1×1 convolutional layer.

[0158] In an optional solution of the embodiment of the present invention, the calculation formula of the feature vector is as follows:

[0159]

[0160] Among them, α i is the feature weight of the i-th sensor data, λ i is the weight parameter of the i-th sensor type, and N is the total number of sensor types.

[0161] In an optional solution of the embodiment of the present invention, the step of correcting the fused feature vector according to the sensor data in the second sensor data subset to obtain the target feature vector includes:

[0162] Obtaining environmental parameters according to the sensor data in the second sensor data subset;

[0163] Establishing a correction matrix based on the environmental parameters;

[0164] The correction matrix is ​​used to correct the fused features to obtain a target feature vector.

[0165] In this embodiment, various parameters of the current environment are obtained by analyzing the sensor data contained in the second sensor data subset, namely, the data of the temperature sensor and the wind speed sensor; after obtaining these environmental parameters, a correction matrix is ​​constructed based on the environmental parameters; and the correction matrix is ​​used to effectively correct the fused features, and finally the target feature vector is obtained.

[0166] For example, in order to reduce the impact of environmental factors, the data of the temperature sensor and the wind speed sensor are introduced to correct the sensor data. Let e ​​represent the environmental parameters (temperature is T, wind speed is V), and the corrected feature vector f correction Calculation formula:

[0167] f correction =f fusion +Ae

[0168] A is the correction matrix, which is learned by the relationship between historical data and environmental parameters. The correction matrix can be learned by the following linear regression model:

[0169] A=(X T X) -1 X T Y

[0170] Where X is the environmental parameter matrix and Y is the uncorrected eigenvector matrix.

[0171] In an optional solution of the embodiment of the present invention, determining the damage degree of the large-span rigid structure according to the target characteristic vector includes:

[0172] Retrieving a damage degree assessment model for the large-span rigid structure;

[0173] The target feature vector is input into the damage degree assessment model to obtain the damage degree of the long-span rigid structure.

[0174] In this embodiment, the previously established large-span rigid structure damage degree assessment model is retrieved and activated; then the preprocessed target feature vector is input into the damage degree assessment model, and the current damage degree assessment result of the large-span rigid structure is obtained through calculation and analysis of the model.

[0175] For example, the main task of ASFF is to extract and fuse features from multimodal sensor data to generate high-quality feature vectors, which can be input into other models to support high-quality training of other models. Taking the embedded diagnostic classifier as an example, the ASFF module is embedded into the classifier to form an end-to-end model that can identify different states of large-span steel structures, such as normal state, minor damage, and severe damage. The overall structure of the model can be expressed as:

[0176] Model(x ∈ ,x a ,x d )=Classifier(f fusion )

[0177] Figure 4A schematic diagram of the steps of a data analysis method for a large-span rigid structure provided in an embodiment of the present application is shown in FIG. Figure 4 As shown, specifically including:

[0178] Step 1: Initialize the parameters of the ASFF module and classifier

[0179] Step 2: Forward propagation: extract features through CFEN; adjust the dimension of the feature vector; calculate adaptive weights and fuse features;

[0180] Step 3: Use the correction matrix A and environmental parameter e to perform environmental correction;

[0181] Step 4: Input the fused and corrected vector into the diagnostic classifier and use the classification loss function Calculate model loss;

[0182] Step 5: Back propagation, using stochastic gradient descent (SGD) to minimize the loss function;

[0183] Step 6: Verify the model performance. If the performance does not meet the requirements, try to adjust the hyperparameters and repeat the training. Otherwise, output the diagnostic results.

[0184] The data analysis method for large-span rigid structures provided in the embodiment of the present application has significant advantages for the multimodal data analysis of large-span steel structures. On the one hand, high-dimensional features are automatically extracted from the time series data of strain sensors, acceleration sensors and displacement sensors through the convolutional feature extraction network (CFEN), which reduces the workload of manual feature selection and improves the representativeness and reliability of the features. This automated feature extraction method can more comprehensively capture the response of the structure under different conditions; on the other hand, the adaptive spatial feature fusion (ASFF) strategy dynamically adjusts the weights of each sensor feature through an adaptive weight mechanism, effectively handles the redundancy and conflict between different sensor data, and ensures the efficient fusion of multimodal data. This strategy is particularly suitable for large-span steel structures, because the response differences of such structures at different positions and time points are large. ASFF can flexibly deal with the response differences of large-span steel structures at different positions and time points, and improve the accuracy and reliability of data fusion.

[0185] In summary, through efficient feature extraction and adaptive feature fusion, high-quality input data can be provided for the subsequent fault diagnosis model, significantly improving the prediction ability and robustness of the model.

[0186] Figure 5 A schematic diagram of a data analysis device for a large-span rigid structure provided in an embodiment of the present application is shown in FIG. Figure 5 As shown, the device specifically includes:

[0187] An acquisition module 501 is used to acquire a sensor data set captured by multiple types of sensors in a long-span rigid structure;

[0188] A division module 502, configured to divide the sensor data set into a first sensor data subset and a second sensor data subset according to a sensor type corresponding to each sensor data in the sensor data set;

[0189] An extraction module 503, configured to extract data features of each sensor data in the first sensor data subset, and generate a fused feature vector by a feature fusion technology;

[0190] A correction module 504, configured to correct the fused feature vector according to the sensor data in the second sensor data subset to obtain a target feature vector;

[0191] The determination module 505 is used to determine the damage degree of the long-span rigid structure according to the target characteristic vector.

[0192] In one possible implementation, the extraction module 503 is further used to input the sensor data in the first sensor data subset into a pre-trained convolutional neural network to obtain an initial feature map of each sensor data; perform a multi-scale convolution operation on the initial feature map of each sensor data to obtain a multi-scale feature map set of each sensor data; perform a corresponding pooling operation on each scale feature map in the multi-scale feature map set of each sensor data to obtain a multi-scale feature vector set of each sensor data; perform feature splicing processing on the multi-scale feature vector set of each sensor data to obtain data features of each sensor data.

[0193] In a possible implementation, the extraction module 503 is further used to input a feature map of a preset size in the multi-scale feature map set into a feature enhancement model for processing; the feature enhancement model includes at least one dilated convolution layer.

[0194] In a possible implementation, the extraction module 503 is further configured to perform upsampling processing on the feature vectors in the multi-scale feature vector set, so that the vector length of the feature vector after processing meets a preset length requirement.

[0195] In one possible implementation, the extraction module 503 is further used to obtain the spatial resolution of the data features corresponding to each sensor data in the first sensor data subset; determine the standard resolution based on the spatial resolution of the data features corresponding to each sensor data; adjust the spatial resolution of the data features corresponding to each sensor to the standard resolution in turn to obtain the target data features of each sensor; obtain the feature weight of each sensor data in the first sensor data subset; and based on the feature weight, perform feature fusion on the target data features of all sensor data in the first sensor data subset to obtain a fused feature vector.

[0196] In a possible implementation, the extraction module 503 is further configured to obtain a weight parameter of each sensor type corresponding to the first sensor data subset; and determine a feature weight of each sensor data in the first sensor data subset according to the weight parameter of each sensor type.

[0197] In a possible implementation, the calculation formula of the feature vector is as follows:

[0198]

[0199] Among them, α i is the feature weight of the i-th sensor data, λ i is the weight parameter of the i-th sensor type, and N is the total number of sensor types.

[0200] In a possible implementation, the correction module 504 is further used to obtain environmental parameters according to the sensor data in the second sensor data subset; establish a correction matrix based on the environmental parameters; and use the correction matrix to correct the fused features to obtain a target feature vector.

[0201] In a possible implementation, the determination module 505 is further used to retrieve a damage degree assessment model of the long-span rigid structure; input the target feature vector into the damage degree assessment model to obtain the damage degree of the long-span rigid structure.

[0202] The data analysis device for large-span rigid structures provided in this embodiment can be as follows: Figure 5 The data analysis device for large-span rigid structures shown in FIG. can be executed as follows: Figure 1-4 All the steps in the data analysis of large-span rigid structures are implemented in Figure 1-4 The technical effect of data analysis applied to large-span rigid structures is shown in the figure. For details, please refer to Figure 1-4 For the sake of brevity, the relevant description is not repeated here.

[0203] The device embodiments described above are merely illustrative, wherein the device components described as separate components may or may not be physically separated, and the components displayed as device components may or may not be physical device components, that is, they may be located in one place, or may be distributed on multiple network device components. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0204] The device embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0205] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a general hardware platform, and of course, by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the relevant technology can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0206] It should be understood that the terms used herein are only for the purpose of describing specific example embodiments and are not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms "one", "an" and "said" as used herein may also be meant to include plural forms. The terms "include", "comprise", "contain", and "have" are inclusive, and therefore specify the existence of stated features, steps, operations, elements and / or parts, but do not exclude the existence or addition of one or more other features, steps, operations, elements, parts, and / or combinations thereof. The method steps, processes, and operations described herein are not interpreted as necessarily requiring them to be performed in the specific order described or illustrated, unless the execution order is clearly indicated. It should also be understood that additional or alternative steps may be used.

[0207] The foregoing is merely a specific embodiment of the present invention, which enables those skilled in the art to understand or implement the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A data analysis method applied to large-span rigid structures, characterized in that: include: Obtain a set of sensor data captured by multiple types of sensors in a long-span rigid structure; Dividing the sensor data set into a first sensor data subset and a second sensor data subset according to a sensor type corresponding to each sensor data in the sensor data set; Extracting data features of each sensor data in the first sensor data subset, and generating a fused feature vector by a feature fusion technology; Correcting the fused feature vector according to the sensor data in the second sensor data subset to obtain a target feature vector; The damage degree of the long-span rigid structure is determined according to the target eigenvector.

2. The method according to claim 1, characterized in that The extracting the data feature of each sensor data in the first sensor data subset includes: Inputting the sensor data in the first sensor data subset into a pre-trained convolutional neural network to obtain an initial feature map of each sensor data; Perform a multi-scale convolution operation on the initial feature map of each sensor data to obtain a multi-scale feature map set of each sensor data; Perform a corresponding pooling operation on each scale feature map in the multi-scale feature map set of each sensor data to obtain a multi-scale feature vector set of each sensor data; The multi-scale feature vector set of each sensor data is processed by feature concatenation to obtain the data features of each sensor data.

3. The method according to claim 2, characterized in that Before performing a corresponding pooling operation on each scale feature map in the multi-scale feature map set of each sensor data, the method further includes: A feature map of a preset size in a multi-scale feature map set is input into a feature enhancement model for processing; the feature enhancement model includes at least one hole convolution layer.

4. The method according to claim 2, characterized in that Before performing feature splicing processing on the multi-scale feature vector set of each sensor data, the method further includes: Upsampling is performed on the feature vectors in the multi-scale feature vector set so that the vector length of the feature vector after processing meets a preset length requirement.

5. The method according to claim 1, characterized in that: The generating of the fused feature vector by the feature fusion technology includes: Obtaining a spatial resolution of a data feature corresponding to each sensor data in the first sensor data subset; Determine the standard resolution based on the spatial resolution of the data features corresponding to each sensor data; The spatial resolution of the data feature corresponding to each sensor is adjusted to the standard resolution in turn to obtain the target data feature of each sensor; Obtaining a feature weight of each sensor data in the first sensor data subset; Based on the feature weights, feature fusion is performed on the target data features of all sensor data in the first sensor data subset to obtain a fused feature vector.

6. The method according to claim 5, characterized in that The obtaining of the feature weight of each sensor data in the first sensor data subset includes: Obtain a weight parameter of each sensor type corresponding to the first sensor subset; A feature weight of each sensor data in the first sensor data subset is determined according to the weight parameter of each sensor type.

7. The method according to claim 1, characterized in that The calculation formula of the feature vector is as follows: Among them, α i is the feature weight of the i-th sensor data, λ i is the weight parameter of the i-th sensor type, and N is the total number of sensor types.

8. The method according to claim 1, characterized in that: The step of correcting the fused feature vector according to the sensor data in the second sensor data subset to obtain the target feature vector includes: Obtaining environmental parameters according to the sensor data in the second sensor data subset; Establishing a correction matrix based on the environmental parameters; The correction matrix is ​​used to correct the fused features to obtain a target feature vector.

9. The method according to claim 2, characterized in that: Determining the damage degree of the large-span rigid structure according to the target characteristic vector includes: Retrieving a damage degree assessment model for the large-span rigid structure; The target feature vector is input into the damage degree assessment model to obtain the damage degree of the long-span rigid structure.

10. A data analysis device applied to a large-span rigid structure, characterized in that: include: An acquisition module is used to acquire a sensor data set captured by multiple types of sensors in a large-span rigid structure; a partitioning module, configured to partition the sensor data set into a first sensor data subset and a second sensor data subset according to a sensor type corresponding to each sensor data in the sensor data set; An extraction module, used for extracting data features of each sensor data in the first sensor data subset, and generating a fused feature vector by a feature fusion technology; A correction module, used for correcting the fused feature vector according to the sensor data in the second sensor data subset to obtain a target feature vector; A determination module is used to determine the damage degree of the long-span rigid structure according to the target characteristic vector.

Images