Structural body health state monitoring method based on multi-scale cascade graph neural network

Abstract

The invention relates to the field of structural body health state monitoring. The method comprises the following steps: constructing a training data set; building a graph neural network, enabling feature extraction modules on different branches of the graph neural network to extract features of the training data set under different scales, and carrying out feature fusion on the features of the multiple branches to obtain fused features; reconstructing and predicting the fusion features, and training the graph neural network until network convergence to obtain a pre-training model; the method comprises the following steps: monitoring a non-vulnerable area and a vulnerable area of a structural body in real time through a plurality of sensors, and recording a data true value of the non-vulnerable area and a data true value of the vulnerable area; inputting the real value of the data of the non-vulnerable area into a pre-training model to obtain a predicted value of the vulnerable area; comparing the deviation between the real value of the vulnerable area data and the predicted value of the vulnerable area, and judging the health state of the structural body, thereby achieving the precise evaluation of the health state of the structural body.

Description

Technical Field

[0001] The present invention relates to the field of structural health status monitoring, and in particular to a structural health status monitoring method based on a multi-scale cascade graph neural network. Background Art

[0002] Health status monitoring of structures is crucial to ensure structural safety. With technological advances, sensor-based data acquisition and intelligent analysis have gradually become the main trend of modern structural health monitoring. These technologies not only improve detection efficiency and accuracy, but also provide an important basis for the maintenance and management of engineering structures. Especially in the fields of aerospace, bridge construction and energy industries, the application of health status monitoring systems improves the safety and reliability of structures, reduces potential accident risks, and increases the feasibility of large-scale deployment and long-term operation of the system.

[0003] Existing structural health status monitoring mainly relies on traditional monitoring methods, such as statistical-based status monitoring methods, model-based status monitoring methods, and signal processing-based status monitoring methods. Statistical methods predict future status changes through statistical analysis of historical data, model methods simulate the real-time operating status of mechanical structures or components by establishing physical models, and signal processing focuses on extracting useful information from a large amount of sensor data to determine the current health status of the structure. The above methods can achieve effective monitoring of structural health status to a certain extent.

[0004] However, existing health status monitoring methods are not sensitive enough to changes in vulnerable areas of structures, making it difficult to accurately monitor the health status of structures in all directions. Especially in the context of complex working conditions and large amounts of sensor data, how to effectively integrate multi-scale information has become a technical problem that needs to be solved urgently. Summary of the invention

[0005] The main purpose of the present invention is to provide a structural health status monitoring method, system and computer equipment based on a multi-scale cascaded graph neural network.

[0006] To achieve the above object, the present invention provides a method for monitoring the health status of a structure based on a multi-scale cascaded graph neural network, comprising the following steps:

[0007] Build a training dataset;

[0008] Build a graph neural network so that the feature extraction modules on different branches of the graph neural network can extract features of training data sets at different scales, and fuse the features of multiple branches to obtain fused features.

[0009] Reconstruct and predict the fused features, and train the graph neural network until the network converges to obtain a pre-trained model;

[0010] Use multiple sensors to monitor the non-vulnerable area and the vulnerable area of ​​the structure in real time, and record the true value of the data in the non-vulnerable area and the true value of the data in the vulnerable area;

[0011] The real value of the non-vulnerable area data is input into the pre-trained model to obtain the predicted value of the vulnerable area;

[0012] Compare the deviation between the actual value of the vulnerable area data and the predicted value of the vulnerable area to determine the health status of the structure.

[0013] Optionally, constructing a training data set specifically includes the following steps:

[0014] Based on the morphology of the structure, a scaled model is constructed;

[0015] Perform stress simulation analysis on the scaled model and design the sensor layout plan based on the stress distribution obtained from the stress simulation analysis;

[0016] Make a scaled model sample and install sensors on the scaled model sample according to the sensor layout plan;

[0017] The scaled model sample is loaded with load, and multi-channel node data are collected through sensors to form a training data set.

[0018] Optionally, the feature extractors on different branches of the graph neural network are used to extract features of training data sets at different scales, and the features of multiple branches are fused to obtain fused features, which specifically includes the following steps:

[0019] Perform feature extraction on the first original input data with a length of l and a shape of l×d in the training data set to obtain the first output matrix feature with a shape of l×2d;

[0020] Another segment of original input data with a length of l is divided into two equal parts in the time series dimension to obtain two second segments of original input data with a length of l / 2; feature extraction is performed on the two segments of time series data with a resolution of l / 2 respectively, and each second output matrix feature with a shape of l / 2×2d is obtained; the two second output matrix features are concatenated in the time series dimension to form a third output matrix feature with a shape of l×2d;

[0021] Perform feature fusion, add the first output matrix feature to the third output matrix feature, and obtain a fused feature with a shape of l×5d.

[0022] Optionally, the fused features are reconstructed and predicted, specifically including:

[0023] The fused features are input into the first processing branch for decoding to generate a fourth output matrix with a shape of l×d; the fourth output matrix is ​​used as the reconstruction of the original input data by the graph neural network, and the understanding and expression ability of the graph neural network for the original input data is evaluated by comparing the fourth output matrix with the original input data;

[0024] The fused feature input and the second processed feature are fitted to generate a fifth output matrix with a shape of l×d; the fifth output matrix is ​​used as the information of the l+1th time point predicted using the first l time point information; by comparing the fifth output matrix with the actual l+1th information, the performance of the network on the time series prediction task can be evaluated.

[0025] Optionally, the fourth output matrix is ​​used as the reconstruction loss function of the graph neural network for the original input data:

[0026]

[0027] In the formula, x i,j is the real data in the input window, To reconstruct data;

[0028] The fifth output matrix is ​​the prediction loss function of the information of the l+1th time point predicted using the information of the first l time points:

[0029]

[0030] In the formula, x l+1,i is the real data of the l+1th row, For forecast data;

[0031] The total loss function of the graph neural network is: L = L pre +L rec .

[0032] Optionally, after obtaining the pre-trained model, the training data set is input again to calculate the anomaly threshold, and the test data is input again to perform anomaly detection;

[0033] Input the training data set again to calculate the anomaly threshold, and input the test data for anomaly detection, which specifically includes the following steps:

[0034] Get the predicted value and reconstructed value of each observation position, calculate the score of each observation position; average the scores of each row to get the global anomaly score of each row; form all the global anomaly scores into a set {S1,S2,...,S N′};

[0035] Set the abnormal threshold th F , set the first hyperparameter q and the second hyperparameter p;

[0036] The extreme value t is calculated by the first hyperparameter q, and the initial threshold th is calculated by the second hyperparameter p;

[0037] Set the initial window to N′ / 12, set the data in the initial window that accounts for the first hyperparameter q to be less than or equal to the extreme value t, and set the data in the initial window that accounts for the second hyperparameter p to be less than or equal to the initial threshold th;

[0038] Input test data and calculate the anomaly score of each test data. If the anomaly score of the test data is greater than the anomaly threshold th F , then the data at this time point is judged to be abnormal; the prediction output requires a window of data, and the row-by-row comparison starts from the l+1th data.

[0039] Optionally, the score for each observation position is calculated using the formula:

[0040]

[0041] Where (x per ―x i ) 2 is the predicted value x pre With the actual value x i The square error, (x rec ―x i ) 2 is the reconstructed value x rec With the actual value x i The squared error.

[0042] Optionally, it also includes updating the abnormal threshold th F , specifically including the following steps:

[0043] Extend the initial window, if the global anomaly score S i ≤ extreme value t or global anomaly score S i ≥ initial threshold th, no processing;

[0044] If the extreme value t<global anomaly score S i <initial threshold th, then update the abnormal threshold th F , update the abnormal threshold th F The calculation formula used is:

[0045]

[0046] Where N is the number of observations, N th is the global anomaly score S i <The number of initial thresholds th, the values ​​of parameters γ and β are estimated using maximum likelihood.

[0047] The present invention also discloses a structural health status monitoring system based on a multi-scale cascade graph neural network. The system applies the structural health status monitoring method based on a multi-scale cascade graph neural network as described above. The system includes:

[0048] Sensors are used to monitor the non-vulnerable area and the vulnerable area of ​​the structure in real time, and record the true value of the data in the non-vulnerable area and the true value of the data in the vulnerable area;

[0049] The graph neural network module is used to extract the connection between the data of each sensor. Each sensor is regarded as a node of the graph neural network. The relationship between a certain point and all adjacent nodes is modeled through the attention layer, so that the information transmitted by the node to the next layer of the network aggregates the information of all its adjacent nodes;

[0050] Feature extraction module, used to extract features of training data sets at different scales;

[0051] The feature fusion module is used to fuse the features of training data sets at different scales to obtain fused features;

[0052] The comparison module compares the deviation between the actual value of the vulnerable area data and the predicted value of the vulnerable area to determine the health status of the structure.

[0053] The present invention also discloses a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and the steps of the above method are implemented when the processor executes the computer program.

[0054] The present invention also provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the above method are implemented.

[0055] The structural health status monitoring method based on multi-scale cascade graph neural network provided by the present invention can extract the features of the structure from multiple scales, effectively improving the accuracy of monitoring the health status of complex structures. First, by constructing a training data set and using different branches of the graph neural network for feature extraction, the data features of the structure can be fully captured at different scales. Secondly, the application of feature fusion technology enables the multi-scale features to be effectively integrated, further improving the data expression ability and the generalization ability of the model. In addition, through the reconstruction and prediction mechanism, not only the understanding and expression ability of the model for the input data is evaluated, but also the time series prediction performance of the model is verified by predicting the information at the next time point. Finally, the non-vulnerable area of ​​the structure is monitored in real time by the sensor, and the predicted value of the vulnerable area is obtained by combining the pre-trained model, and then the health status of the structure is accurately judged by comparing the predicted value with the true value. The vulnerable area data obtained by predicting the non-vulnerable area data can avoid the negative impact of the abnormal vulnerable area data on the monitoring, and the data of the vulnerable area is easier to obtain than the data of the non-vulnerable area, which is convenient for long-term data acquisition, and ensures that the sensor that obtains the data will not be abnormal due to the sudden damage of the structure, which significantly improves the reliability and real-time performance of the monitoring system. Overall, the synergy between the various features in the above technical solutions ensures the high efficiency and accuracy of structural health status monitoring. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a flow chart of the structural health status monitoring method based on multi-scale cascaded graph neural network of the present invention;

[0057] Figure 2 is a schematic diagram of the structure of the computer device of the present invention;

[0058] Figure 3 This is the static tensile loading response curve of the scaled model sample of the present invention.

[0059] Figure 4 is the offset static tensile loading response curve of the scaled model sample of the present invention;

[0060] Figure 5 It is the response curve of the scaled model sample of the present invention to the central triangle wave loading;

[0061] Figure 6 It is the sinusoidal wave loading response curve of the scaled model sample of the present invention;

[0062] Figure 7 This is the square wave loading response curve of the scaled model sample of the present invention;

[0063] Figure 8 is the biased triangle wave loading response curve of the scaled model sample of the present invention;

[0064] Fig. 9 is the biased sine wave loading response curve of the scaled model sample of the present invention;

[0065] Fig.10 is the bias square wave loading response curve of the scaled model sample of the present invention;

[0066] Fig.11 is a healthy point response curve of the scaled model sample of the present invention;

[0067] Fig.12 is a fatigue failure response curve of a vulnerable point of a scaled model sample of the present invention;

[0068] Fig.13 is a cross-correlation analysis diagram of each data channel of the present invention;

[0069] Fig.14 is the health status classification result of the single vulnerable channel of channel 4 of the present invention;

[0070] Fig.15 It is the classification result of the health status of the single vulnerable channel of channel 13 of the present invention;

[0071] Fig.16 This is the classification result of the health status of the dual vulnerable channels of channel 4 and channel 13 of the present invention. DETAILED DESCRIPTION

[0072] The following will be combined with the accompanying drawings to clearly and completely describe the technical solutions in the embodiments of the present invention. The described embodiments are only possible technical implementations of the present invention, not all possible implementations. Those skilled in the art can completely combine the embodiments of the present invention to obtain other embodiments without creative work, and these embodiments are also within the scope of protection of the present invention.

[0073] Embodiment 1

[0074] Reference Figure 1 In this embodiment, a method for monitoring the health status of a structure based on a multi-scale cascaded graph neural network is provided, comprising the following steps:

[0075] Step 1, construct a training data set; Constructing a training data set specifically includes the following steps: construct a scaled model based on the morphology of the structure; perform stress simulation analysis on the scaled model, and design a sensor layout scheme based on the stress distribution obtained by the stress simulation analysis; Specifically, the simulation experiment of this embodiment is carried out using ANSYS software. After the model built in SolidWorks is imported into ANSYS, the established typical hull structure model is meshed, and bearing loads are added to the center ear plate position of the model, fixed support constraints are added to the connection holes at both ends, and total deformation and equivalent stress solutions are added, and then the solution is started. The static load starts from 500N and gradually increases with a gradient of 500N until multiple high stress / crack danger zones are found. Stop. By reducing the stress display range of the stress cloud map of the simulation result to a higher stress range, it is possible to clearly see the parts with higher stress in the model. These parts with higher stress are the key monitoring points that need to be focused on in the subsequent sample verification test, that is, the points where sensors need to be arranged for monitoring.

[0076] Then make a scaled model sample, install sensors on the scaled model sample according to the sensor layout plan; load the scaled model sample, the loading device is mainly composed of the test bench frame, servo loading cylinder, servo controller, force sensor, position sensor and control host computer; the servo loading cylinder can realize force loading within the frequency of 3-4Hz and the amplitude of 100 tons, and the position control can be accurate to 0.01 mm, which can realize two loading modes: force loading and displacement loading. The force loading connection ear plate at the end of the principle test model is connected to the end of the loading cylinder through a pin, so as to apply an in-plane load or bending moment on the scaled model sample. Arrange the points for monitoring by sensors according to the needs, and perform strain gauge patching at the corresponding positions of the scaled model sample. When installing the strain gauge, it is pasted on the surface of the structural monitoring point with 502 glue. Before pasting, the monitoring point is wiped and polished to ensure neatness, and finally coated with silicone on the surface to achieve the effect of waterproofing and moisture-proofing. After arranging the strain gauge, the sample needs to be installed on the test bench frame through the test fixture. Use bolts to connect and tighten the two, and then use pins to connect the force loading connection ear plate at the top of the sample to the servo cylinder. After ensuring the connection is firm, use anchor bolts to install the test fixture on the test bench frame. The test fixture has sufficient strength and rigidity to ensure that the deformation of the test fixture structure caused by the cylinder loading force and preload during the test is much smaller than the deformation of the sample during the formal test.

[0077] Finally, the sensors are connected to the channels of the data acquisition system in turn according to the corresponding relationship, and the data acquisition instrument is connected to the PC with the configured environment. The monitoring software is used to check each channel one by one to ensure that the measurement system is in normal working condition. At the same time, a reasonable amplification factor or attenuation factor is selected through debugging.

[0078] Static tension is applied to the center ear plate of the scaled model sample. The levels of static tension are loaded from 0.5 tons to 5.5 tons, and then unloaded to 0.5 tons. The scale of each loading / unloading change is 1 ton. After the static tension is loaded at each level, the cylinder is loaded and maintained for 1.5 minutes to ensure that the structure reaches a steady state, and a 0.5-minute measurement interval is set. The response of each monitoring point of the sample under different static load sizes is obtained through the strain sensor, and the data of channels 3 and 4 are selected to draw the response diagram as shown in the figure. Figure 3 As shown, the response curve is basically consistent with the step response of the second-order system.

[0079] Then, static tension is applied to the left ear plate of the sample. The static tension level is loaded from 0.5 tons to 5.5 tons, and then unloaded to 0.5 tons. The scale of each loading / unloading change is 1 ton. After the static tension is loaded at each level, the cylinder is loaded and maintained for 1.5 minutes to ensure that the structure reaches a steady state, and a 0.5-minute measurement interval is set. The response of each monitoring point of the sample under different static load sizes is obtained through the strain sensor, and the data of channels 3 and 4 are selected to draw the response diagram as shown in the figure. Figure 4 As shown, the response curve is basically consistent with the step response of the second-order system.

[0080] Then, a dynamic load with a triangular waveform is applied to the center ear plate of the sample. The loading amplitude is between 0.5 tons and 5 tons, and the loading frequency is between 0.5Hz and 3Hz. The response of each monitoring point of the sample under different amplitude and frequency conditions is obtained through the strain sensor. The number of loading times for each waveform is 1000 waveform cycles. In the test with a loading amplitude of 2 tons and a loading frequency of 0.5Hz, some data from channels 3 and 4 are selected to draw the response diagram as shown in the figure below. Figure 5 As shown, the response curve is basically consistent with the triangle wave response of the second-order system.

[0081] Then, a dynamic load with a sinusoidal waveform is applied to the center ear plate of the sample. The loading amplitude is between 0.5 tons and 5 tons, and the loading frequency is between 0.5Hz and 3Hz. The response of each monitoring point of the sample under different amplitude and frequency conditions is obtained through the strain sensor. The number of loading times for each waveform is 1000 waveform cycles. In the test with a loading amplitude of 2 tons and a loading frequency of 1Hz, some data from channels 3 and 4 are selected to draw the response diagram as shown in the figure below. Figure 6 As shown, the response curve basically conforms to the sinusoidal wave response of the second-order system.

[0082] Then, a square wave dynamic load is applied to the center ear plate of the sample. The loading amplitude is between 0.5 tons and 5 tons, and the loading frequency is between 0.5Hz and 3Hz. The response of each monitoring point of the sample under different amplitude and frequency conditions is obtained through the strain sensor. The number of loading times for each waveform is 1000 waveform cycles. In the test with a loading amplitude of 2 tons and a loading frequency of 0.25Hz, partial data from channels 3 and 4 are selected to draw the response diagram as shown in the figure below. Figure 7 As shown, the response curve is basically consistent with the square wave response of the second-order system.

[0083] Then, a dynamic load with a triangular waveform is applied to the left ear plate of the sample. The loading amplitude is between 0.5 tons and 5 tons, and the loading frequency is between 0.5Hz and 3Hz. The response of each monitoring point of the sample under different amplitude and frequency conditions is obtained through the strain sensor. The number of loading times for each waveform is 1000 waveform cycles. In the test with a loading amplitude of 2 tons and a loading frequency of 0.5Hz, some data from channels 3 and 4 are selected to draw the response diagram as shown in the figure below. Figure 8 As shown, the response curve is basically consistent with the triangle wave response of the second-order system.

[0084] Then, a dynamic load with a sinusoidal waveform is applied to the left ear plate of the sample. The loading amplitude is between 0.5 tons and 5 tons, and the loading frequency is between 0.5Hz and 3Hz. The response of each monitoring point of the sample under different amplitude and frequency conditions is obtained through the strain sensor. The number of loading times for each waveform is 1000 waveform cycles. In the test with a loading amplitude of 2 tons and a loading frequency of 1Hz, partial data from channels 3 and 4 are selected to draw the response diagram as shown in the figure below. Fig. 9 As shown, the response curve basically conforms to the sinusoidal wave response of the second-order system.

[0085] Then, a square wave dynamic load is applied to the left ear plate of the sample. The loading amplitude is between 0.5 tons and 5 tons, and the loading frequency is between 0.5Hz and 3Hz. The response of each monitoring point of the sample under different amplitude and frequency conditions is obtained through the strain sensor. The number of loading times for each waveform is 1000 waveform cycles. In the test with a loading amplitude of 2 tons and a loading frequency of 0.25Hz, some data from channels 3 and 4 are selected to draw the response diagram as shown in the figure below. Fig.10 As shown, the response curve is basically consistent with the square wave response of the second-order system.

[0086] A dynamic load with a sinusoidal waveform was applied to the center ear plate of the sample, with a loading amplitude between 0.5 tons and 10.5 tons and a loading frequency of 1 Hz. The appearance of fatigue cracks was detected when the dynamic load was applied to the high stress area, and the response of each monitoring point of the sample under different amplitude conditions was obtained through the strain sensor. Fig.11 As shown in Figure 2, the response curve of the healthy point remains basically stable. Fig.12As shown in the figure, the vulnerable point will experience amplitude narrowing or even fatigue failure in the later stage of the experiment.

[0087] In this step, all experimental data are collected through sensors to form a training data set; specifically, each batch of experimental sensor data of each loading type is divided into data sets with a length of 100 sampling time points, and 80% of the data is used as a training set, 10% of the data is used as a validation set, and 10% of the data is used as a test set in units of test batches, ensuring that the test set data is data that has not been used in the training set and validation set to verify that the model is not overfitting.

[0088] Step 2: Build a graph neural network so that the feature extraction modules on different branches of the graph neural network can extract features of training data sets at different scales, and fuse the features of multiple branches to obtain fused features. In this step, in order to process all sensor data as a whole, a graph attention network GAT is introduced to extract the connection between the sensor data, and a gated recurrent unit GRU is introduced to process time series data. Use the GAT and GRU networks to extract features, and then use the RNN decoder to decode the extracted features. Specifically, the feature extractors on different branches of the graph neural network extract features of training data sets at different scales. The overall implementation process is as follows:

[0089] The shape of the input data is l×d, and then dilated convolution is performed for each dimension. A convolution kernel of size 7 is used, and the dilation rate is set to 2. The dilated convolution expands the receptive field, allowing the model to capture a wider range of contextual information, so 3 zeros need to be padded on both sides of the input sequence. After the convolution operation is completed, this output sequence is added to the original input data to achieve residual connection. The convolved data is then modeled using the GAT-V2 network. The dimensional feature map and the temporal feature map are constructed by the GAT-V2 network, respectively. These two feature maps capture the correlation between different dimensions and the dynamic changes of the time series. Finally, the idea of ​​skip connection is used again to splice the dimensional feature map, the temporal feature map, and the original input. This step integrates information from different sources and enriches the feature representation. The shape of the spliced ​​matrix is ​​l×3d, which not only retains the structural information of the original data, but also incorporates the advanced features extracted by the convolution and GAT networks, providing strong support for subsequent tasks.

[0090] Then, the features of multiple branches are fused to obtain the fused features. The specific steps are as follows:

[0091] While extracting features of time series data with a window length of l, features with smaller window lengths, such as l / 2, l / 4, l / 8, etc., are also extracted. After fusing features of different scales, anomaly detection is performed, so that different contextual information can be extracted. After the data is processed by the feature extractor, its length does not change. Therefore, when the time series data with a window length of l / 2 is extracted, the length of its output data is still l / 2, which is different from the length of the features extracted from the data with a window length of l, resulting in the inability to directly fuse. This inconsistency in length makes it impossible for the network to synchronously update the loss, making it impossible to perform effective training.

[0092] Therefore, it is necessary to construct two parallel branches in the graph neural network to process data of different resolutions at the same time to solve the problem of data misalignment. This embodiment adopts the following solution: the first branch uses a larger window of size l for feature extraction; the second branch divides the larger input window into two smaller windows of size l / 2, and extracts features for each small window. Subsequently, the features of the two small windows are spliced ​​so that their total length reaches l, which is consistent with the output length of the first branch. This method not only ensures data alignment, but also achieves more comprehensive and detailed feature extraction. In this way, the two branches can be trained simultaneously. After the data is aligned, a feature fusion operation is further performed to make full use of the feature information extracted by different branches. At the same time, the sliding window input ensures that all time points can be covered.

[0093] The specific implementation process is as follows: First, feature extraction is performed on the first original input data with a length of l and a shape of l×d in the training data set to obtain the first output matrix feature with a shape of l×2d;

[0094] Another segment of original input data with a length of l is divided into two equal parts in the time series dimension to obtain two second segments of original input data with a length of l / 2; feature extraction is performed on the two segments of time series data with a resolution of l / 2 respectively, and each second output matrix feature with a shape of l / 2×2d is obtained; the two second output matrix features are concatenated in the time series dimension to form a third output matrix feature with a shape of l×2d;

[0095] Perform feature fusion, add the first output matrix feature to the third output matrix feature, and obtain a fused feature with a shape of l×5d. The multi-scale cascade strategy effectively combines the feature information of time series data of different input scales, providing a more comprehensive and rich data foundation for subsequent analysis and processing.

[0096] Step 3: Reconstruct and predict the fused features, and train the graph neural network until the network converges to obtain a pre-trained model;

[0097] Reconstructing and predicting the fusion features specifically includes:

[0098] The fused features are input into the first processing branch for decoding to generate a fourth output matrix with a shape of l×d; the fourth output matrix is ​​used as the reconstruction of the original input data by the graph neural network, and the understanding and expression ability of the graph neural network for the original input data is evaluated by comparing the fourth output matrix with the original input data;

[0099] The fused feature input and the second processed feature are fitted to generate a fifth output matrix with a shape of l×d; the fifth output matrix is ​​used as the information of the l+1th time point predicted using the first l time point information; by comparing the fifth output matrix with the actual l+1th information, the performance of the network on the time series prediction task can be evaluated.

[0100] Furthermore, the fourth output matrix of this embodiment is used as the reconstruction loss function of the graph neural network for the original input data:

[0101]

[0102] In the formula, x i,j is the real data in the input window, To reconstruct data;

[0103] The fifth output matrix is ​​the prediction loss function of the information of the l+1th time point predicted using the information of the first l time points:

[0104]

[0105] In the formula, x l+1,i is the real data of the l+1th row, For forecast data;

[0106] The total loss function of the graph neural network is: L = L pre +L rec .

[0107] Furthermore, in this embodiment, after obtaining the pre-trained model, it also includes inputting the training data set again to calculate the anomaly threshold, and inputting the test data to perform anomaly detection;

[0108] Input the training data set again to calculate the anomaly threshold, and input the test data for anomaly detection. The specific steps include the following:

[0109] Get the predicted value and reconstructed value of each observation position, calculate the score of each observation position; average the scores of each row to get the global anomaly score of each row; form all the global anomaly scores into a set {S1,S2,...,S N′};

[0110] Set the abnormal threshold th F , set the first hyperparameter q and the second hyperparameter p;

[0111] The extreme value t is calculated by the first hyperparameter q, and the initial threshold th is calculated by the second hyperparameter p;

[0112] Set the initial window to N′ / 12, set the data in the initial window that accounts for the first hyperparameter q to be less than or equal to the extreme value t, and set the data in the initial window that accounts for the second hyperparameter p to be less than or equal to the initial threshold th;

[0113] Input test data and calculate the anomaly score of each test data. If the anomaly score of the test data is greater than the anomaly threshold th F , then the data at this time point is judged to be abnormal; the prediction output requires a window of data, and the row-by-row comparison starts from the l+1th data.

[0114] Furthermore, the calculation formula used in this embodiment to calculate the score of each observation position is:

[0115]

[0116] Where (x pre ―x i ) 2 is the predicted value x pre With the actual value x i The square error, (x rec ―x i ) 2 is the reconstructed value x rec With the actual value x i The squared error.

[0117] Optionally, it also includes updating the abnormal threshold th F , specifically including the following steps:

[0118] Extend the initial window, if the global anomaly score S i ≤ extreme value t or global anomaly score S i ≥ initial threshold th, no processing;

[0119] If the extreme value t<global anomaly score S i <initial threshold th, then update the abnormal threshold th F , update the abnormal threshold th F The calculation formula used is:

[0120]

[0121] Where N is the number of observations, N th is the global anomaly score Si <The number of initial thresholds th, the values ​​of parameters γ and β are estimated using maximum likelihood. For the input test data, compare row by row. If the abnormal score of the test data is greater than the threshold, that is, score test >th F , then the data at this time point is judged to be abnormal. Since the prediction output requires a window of data, the row-by-row comparison starts from the l+1th data.

[0122] Step 4: monitor the non-vulnerable area and the vulnerable area of ​​the structure in real time through multiple sensors, and record the true value of the data in the non-vulnerable area and the true value of the data in the vulnerable area;

[0123] Specifically, the data used for the correlation of health data between the non-vulnerable area and the vulnerable area in this embodiment is the data of 30 channels collected under healthy conditions. The correlation coefficient is calculated for every two channels. The value is between 0 and 1. The closer to 1, the stronger the correlation, and the closer to 0, the weaker the correlation. In order to verify the effectiveness of the reference channel selection and its independence from a single sample, the correlation coefficient is calculated for the data of three samples respectively, and the value greater than 0.99 is set to 1, and the value less than or equal to 0.99 is set to 0. Then the intersection of the results of the three samples is obtained as follows: Fig.13 The results are shown. The gray area in the figure calculates the correlation coefficient for each channel, which is the diagonal position of the matrix. From the figure below, it can be seen that there is a strong correlation between channels 1, 2, 3, and 9 in the non-destructive channels and vulnerable channels 4 and 13, and this is the case on all three samples. Therefore, it can be proved that it is completely feasible to select channels 1, 2, 3, and 9 in this experiment to predict the health values ​​of channels 4 and 13, and then classify their health conditions.

[0124] Therefore, in the aforementioned training process, the training data set of this embodiment is collected from a single sample under a fatigue cycle, with a total of 6 channels of data, namely 1, 2, 3, 9, 4, and 13. Among them, channels 1, 2, 3, and 9 are known non-fragile and reliable channels, and channels 4 and 13 are fragile channels. The entire data is a sample that is continuously loaded from a healthy state to allow it to slowly wear until it is completely destroyed to obtain the stress value collected by the sensor in a continuous time. The frequency used is 10Hz. There are a total of 143,361 continuous time points in the data, and every 50 consecutive points are taken as a sample to construct a data set. 85% of the data are used as the training set, 10% of the data are used as the validation set, and 5% of the data are used as the test set. Finally, the training set has 123,109 samples, the validation set has 14,483 samples, and the test set has 7,241 samples.

[0125] The real labels of the dataset are given artificially. The classification of the health conditions of the 13 channels is shown in the figure below. The first 12,000 time point data are regarded as the health condition of the vulnerable point. 12,000-37,000 is a mild fatigue condition, 37,000-58,500 is a moderate fatigue condition, and 58,500-77,000 is a severe fatigue condition. After 77,000, it can be seen that the amplitude changes significantly and deforms, which is classified as complete damage.

[0126] Step 5: Input the real value of the non-vulnerable area data into the pre-trained model to obtain the predicted value of the vulnerable area;

[0127] Step 6: Compare the deviation between the actual value of the vulnerable area data and the predicted value of the vulnerable area to determine the health status of the structure.

[0128] like Fig.14 The confusion matrix and classification report of the health status classification of channel 4 single vulnerable channel are shown, 0-4 represents health, mild fatigue, moderate fatigue, severe fatigue and complete damage respectively. From the confusion matrix, it can be seen that the number of samples of the five categories randomly sampled in the test set are 588, 1293, 1103, 878, 3379, a total of 7241 samples, and all are classified correctly. From the classification report in the figure below, it can be seen that the accuracy of health status classification is 95.61%, and the misclassified data are basically at the edge of two adjacent categories. This error may be caused by the fact that the edges of the training data are not fine enough.

[0129] like Fig.15 The confusion matrix and classification report of the single vulnerable channel health classification of channel 13 are shown. From the confusion matrix, it can be seen that the number of samples of the five categories randomly sampled in the test set are 625, 1221, 1073, 924, and 3398, a total of 7241 samples, and the average accuracy is 97.83%.

[0130] like Fig.16 The confusion matrix and classification report of the health status classification of the dual vulnerable channels of channel 4 and channel 13 are shown in the figure. The left figure is channel 4 and the right figure is channel 13. From the confusion matrix, it can be seen that the number of samples of the five categories randomly sampled in the test set are 561, 1296, 1110, 924, and 3350, respectively, with a total of 7241 samples. The average classification accuracy of the two channels is 94.30%. It can be seen that the anomaly detection effect in a single vulnerable area is relatively ideal, and there are also good results when anomaly detection is performed on multiple vulnerable areas at the same time. When the amount of data is sufficient and the abnormal conditions are divided more finely, it can be applied to actual application scenarios. This method can detect anomalies in multiple vulnerable areas at the same time.

[0131] Furthermore, an embodiment of the present invention also provides a structural health status monitoring system based on a multi-scale cascaded graph neural network. The system applies the structural health status monitoring method based on a multi-scale cascaded graph neural network as described above, and the system includes:

[0132] Sensors are used to monitor the non-vulnerable area and the vulnerable area of ​​the structure in real time, and record the true value of the data in the non-vulnerable area and the true value of the data in the vulnerable area;

[0133] The graph neural network module is used to extract the connection between the data of each sensor. Each sensor is regarded as a node of the graph neural network. The relationship between a certain point and all adjacent nodes is modeled through the attention layer, so that the information transmitted by the node to the next layer of the network aggregates the information of all its adjacent nodes;

[0134] Feature extraction module, used to extract features of training data sets at different scales;

[0135] The feature fusion module is used to fuse the features of training data sets at different scales to obtain fused features;

[0136] The comparison module compares the deviation between the actual value of the vulnerable area data and the predicted value of the vulnerable area to determine the health status of the structure.

[0137] Reference Figure 2 In an embodiment of the present invention, a computer device is also provided. The computer device may be a server, and its internal structure may be as follows: Figure 2 As shown. The computer device includes a processor, a memory, a display screen, an input device, a network interface and a database connected through a system bus. Among them, the processor designed by the computer is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store the corresponding data in this embodiment. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the above method is implemented.

[0138] Those skilled in the art will understand that Figure 2 The structure shown in the figure is merely a block diagram of a portion of the structure related to the solution of the present invention, and does not constitute a limitation on the computer device to which the solution of the present invention is applied.

[0139] An embodiment of the present invention further provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented. It can be understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0140] In summary, the method, system and device for monitoring the health status of a structure based on a multi-scale cascaded graph neural network are provided in an embodiment of the present invention. The method includes: constructing a training data set; building a graph neural network, so that the feature extraction modules on different branches of the graph neural network extract features of the training data set at different scales, and fuse the features of multiple branches to obtain fused features; reconstructing and predicting the fused features, and training the graph neural network until the network converges to obtain a pre-trained model; real-time monitoring of non-vulnerable areas and vulnerable areas of the structure through multiple sensors, recording the true value of the non-vulnerable area data and the true value of the vulnerable area data; inputting the true value of the non-vulnerable area data into the pre-trained model to obtain the predicted value of the vulnerable area; comparing the deviation between the true value of the vulnerable area data and the predicted value of the vulnerable area to judge the health status of the structure.

[0141] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media provided by the present invention and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double-speed data rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM.

[0142] It should be noted that, in this article, the terms "include", "comprises" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "includes a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.

[0143] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A structural health status monitoring method based on a multi-scale cascaded graph neural network, characterized in that: The following steps are involved: Build a training dataset; Building a graph neural network, so that feature extraction modules on different branches of the graph neural network extract features of the training data set at different scales, and fuse features of multiple branches to obtain fused features; Reconstructing and predicting the fused features, and training the graph neural network until the network converges to obtain a pre-trained model; Use multiple sensors to monitor the non-vulnerable area and the vulnerable area of ​​the structure in real time, and record the true value of the data in the non-vulnerable area and the true value of the data in the vulnerable area; Inputting the true value of the non-vulnerable area data into the pre-trained model to obtain a predicted value of the vulnerable area; The deviation between the actual value of the vulnerable area data and the predicted value of the vulnerable area is compared to determine the health status of the structure.

2. The method for monitoring the structural health status according to claim 1, characterized in that: The construction of the training data set specifically includes the following steps: constructing a scaled model based on the morphology of the structure; Performing stress simulation analysis on the scaled model, and designing a sensor layout plan according to stress distribution obtained from the stress simulation analysis; Making the scaled model sample, and installing sensors on the scaled model sample according to the sensor layout plan; A load is applied to the scaled model sample, and multi-channel node data is collected through the sensor to form the training data set.

3. The method for monitoring the structural health status according to claim 1, characterized in that: The step of enabling the feature extractors on different branches of the graph neural network to extract features of the training data set at different scales, and fusing the features of multiple branches to obtain fused features specifically includes the following steps: Performing feature extraction on first original input data with a length of l and a shape of l×d in the training data set to obtain first output matrix features with a shape of l×2d; Another segment of original input data with a length of l is divided into two equal parts in the time series dimension to obtain two second segments of original input data with a length of l / 2; feature extraction is performed on the two segments of the time series data with a resolution of l / 2 respectively to obtain second output matrix features with a shape of l / 2×2d for each segment; Concatenate the two second output matrix features in the time dimension to form a third output matrix feature with a shape of l×2d; Perform feature fusion, add the first output matrix feature to the third output matrix feature, and obtain a fused feature with a shape of l×5d.

4. The method for monitoring the structural health status according to claim 3, characterized in that: The reconstructing and predicting the fusion features specifically includes: Input the fused features into the first processing branch for decoding to generate a fourth output matrix with a shape of l×d; the fourth output matrix is ​​used as the reconstruction of the original input data by the graph neural network, and by comparing the fourth output matrix with the original input data, the understanding and expression ability of the graph neural network for the original input data is evaluated; The fused feature input and the second processing feature are fitted to generate a fifth output matrix with a shape of l×d; the fifth output matrix is ​​used as the information of the l+1th time point predicted using the first l time point information; by comparing the fifth output matrix with the actual l+1th information, the performance of the network on the time series prediction task can be evaluated.

5. The method for monitoring the health status of a structure according to claim 4, characterized in that: The fourth output matrix as the reconstruction loss function of the graph neural network for the original input data is: In the formula, x i,j is the real data in the input window, To reconstruct data; The fifth output matrix is ​​a prediction loss function of the information of the l+1th time point predicted using the information of the first l time points: In the formula, x l+1,i is the real data of the l+1th row, For forecast data; The total loss function of the graph neural network is: L = L pre +L rec .

6. The method for monitoring the structural health status according to claim 1, characterized in that: After obtaining the pre-trained model, the method further includes inputting the training data set again to calculate an anomaly threshold, and inputting the test data to perform anomaly detection; The step of inputting the training data set again to calculate the anomaly threshold and inputting the test data for anomaly detection specifically comprises the following steps: Get the predicted value and reconstructed value of each observation position, calculate the score of each observation position; average the scores of each row to get the global anomaly score of each row; form all the global anomaly scores into a set {S1,S2,...,S N′ }; Set the abnormal threshold th F , set the first hyperparameter q and the second hyperparameter p; The extreme value t is calculated by the first hyperparameter q, and the initial threshold th is calculated by the second hyperparameter p; Set the initial window to N′ / 12, set the data in the initial window that accounts for the first hyperparameter q to be less than or equal to the extreme value t, and set the data in the initial window that accounts for the second hyperparameter p to be less than or equal to the initial threshold th; Input test data and calculate the anomaly score of each test data. If the anomaly score of the test data is greater than the anomaly threshold th F , then the data at this time point is judged to be abnormal; The prediction output requires a window of data, and the row-by-row comparison starts from the l+1th data.

7. The method for monitoring the structural health status according to claim 6, characterized in that: The calculation formula used to calculate the score of each observation position is: Where (x pre ―x i ) 2 is the predicted value x pre With the actual value x i The square error, (x rec ―x i ) 2 is the reconstructed value x rec With the actual value x i The squared error.

8. The method for monitoring the structural health status according to claim 6, characterized in that: It also includes updating the abnormal threshold th F , specifically including the following steps: Extend the initial window, if the global anomaly score S i ≤ the extreme value t or the global anomaly score S i ≥ the initial threshold th, no processing; If the extreme value t< the global anomaly score S i < the initial threshold th, then update the abnormal threshold th F , update the abnormal threshold th F The calculation formula used is: Where N is the number of observations, N th is the global anomaly score S i <The number of the initial threshold th, the values ​​of the parameters γ and β are estimated using maximum likelihood.

9. A structural health status monitoring system based on a multi-scale cascaded graph neural network, characterized in that: The system applies the structural health status monitoring method based on a multi-scale cascaded graph neural network as described in any one of claims 1 to 8 above, and the system comprises: Sensors are used to monitor the non-vulnerable area and the vulnerable area of ​​the structure in real time, and record the true value of the data in the non-vulnerable area and the true value of the data in the vulnerable area; The graph neural network module is used to extract the connection between the data of each sensor. Each sensor is regarded as a node of the graph neural network. The relationship between a certain point and all adjacent nodes is modeled through the attention layer, so that the information transmitted by the node to the next layer of the network aggregates the information of all its adjacent nodes; Feature extraction module, used to extract features of training data sets at different scales; The feature fusion module is used to fuse the features of training data sets at different scales to obtain fused features; The comparison module compares the deviation between the actual value of the vulnerable area data and the predicted value of the vulnerable area to determine the health status of the structure.

10. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 8 are implemented.

Images