A method and device for identifying abnormal data of structure monitoring with parameter self-adaptive adjustment

Abstract

The present application belongs to the technical field of engineering structure monitoring data analysis, and proposes a kind of parameter self-adapting adjustment structure monitoring abnormal data identification method and device.Bridge monitoring data is converted into image by time domain conversion.According to the contour feature of the image, the bridge monitoring data is classified.Then, by using the method of transfer learning, a pre-trained model is selected to train the classified bridge monitoring data.In the data training process, parameter self-adapting stable convergence judgment is carried out.If the stable convergence judgment condition is met, the multiple transfer learning can be terminated, the parameters are optimally configured, the model obtains the best identification performance, and the identification accuracy of each type of image reaches the best.The present application can make the transfer learning parameter configuration optimal without fine adjustment of the parameters, greatly reduce the time and labor cost, and make the pre-trained model reach the best performance, so as to obtain better identification accuracy in data identification.

Description

Technical Field

[0001] This invention relates to the field of engineering structure monitoring data analysis technology, and in particular to a method and apparatus for identifying abnormal data in structural monitoring with adaptive parameter adjustment. Background Technology

[0002] Structural health monitoring is a crucial means of ensuring structural safety. However, environmental factors can cause anomalies in the data collected by structural health monitoring sensors. A large number of anomalous data points can lead to inaccurate assessments of structural safety, potentially resulting in false warnings, unnecessary structural maintenance losses, or an increased probability of accidents. Therefore, effective detection of anomalous data is essential for obtaining reliable structural health monitoring results.

[0003] Traditional manual inspection can easily identify data anomalies such as missing data and skipped points. However, due to the massive amount of data being monitored, the efficiency of manual anomaly identification is very limited, resulting in a significant waste of time and labor costs. Therefore, adopting automated methods for anomaly identification can greatly reduce time and labor costs, and is more beneficial for practical engineering applications.

[0004] Research on structural anomaly detection has made some progress. “BAO Y, TANG Z, LI H, et al. 2019. Computer vision and deep learning–based data anomaly detection method for structural health monitoring [J / OL]. Structural Health Monitoring, 18(2): 401-421. https: / / doi.org / 10.1177 / 1475921718757405.” In 2019, Bao et al. used a trained deep neural network to identify multi-class anomaly data. The effectiveness of the strategy was verified by evaluating the measured data of a long-span bridge in China. “LIU G, NIU Y, ZHAO W, et al. 2022. Data anomaly detection for structural health monitoring using a combination network of GANomaly and CNN [J / OL]. Smart Structures and Systems, 29(1): 53-62. https: / / doi.org / 10.12989 / SSS.2022.29.1.053.” In 2019, Tang et al. visualized the monitoring data in the time and frequency domains, converting it into time-frequency images. They then used a trained convolutional neural network to automatically detect the data images and identify multiple categories of abnormal data. "TANGZ, CHEN Z, BAO Y, et al. 2019. Convolutional neural network-based data anomaly detection method using multiple information for structural health monitoring [J / OL]. Structural Control and Health Monitoring, 26(1):e2296. https: / / doi.org / 10.1002 / stc.2296." In 2022, Liu et al. proposed a method for detecting anomalies in structural health monitoring data based on generative adversarial networks and convolutional neural networks, and tested the proposed method using acceleration data from a structural health monitoring system for long-span bridges. However, this deep learning-based detection method is data-dependent; if the amount of data for a certain type is insufficient, the accuracy of the neural network model in identifying anomalies of that type will decrease.Therefore, the introduction of transfer learning is necessary, as it only requires a relatively small dataset to adapt existing deep networks to bridge anomaly detection tasks.

[0005] However, the effectiveness of transfer learning is inextricably linked to parameter settings; different combinations will yield varying classification accuracies on different datasets. Obtaining the optimal model requires careful determination of each parameter, and adjusting network parameters is a time-consuming process. Therefore, it is crucial to propose an adaptive parameter tuning method for structural anomaly detection to automate anomaly detection and achieve high accuracy in anomaly classification. Summary of the Invention

[0006] The present invention aims to provide a method and apparatus for identifying abnormal data in structural monitoring with adaptive parameter adjustment, thereby solving the problem of parameter setting in the transfer learning process.

[0007] The technical solution of the present invention is as follows:

[0008] A method for identifying anomaly data in structural monitoring with adaptive parameter adjustment is proposed. First, the raw bridge monitoring data is converted into images in the time domain. Based on the contour features of the converted images, the bridge monitoring data is classified. A transfer learning method is adopted, and a pre-trained model is selected to train the classified bridge monitoring data. During the data training process, parameter adaptive stability convergence judgment is performed. When the stability convergence judgment condition is met, multiple transfer learning operations are terminated, the parameters are optimally configured, the pre-trained model obtains the best recognition performance, and the recognition accuracy for each category of images reaches the best.

[0009] The specific steps are as follows:

[0010] Step 1: Image conversion;

[0011] The accelerometers installed on the bridge have a sampling frequency of f = x Hz, collecting data x times per second, and collecting the acceleration at that point 3600*x times per hour. Using time as the horizontal axis and the acceleration value at that time point as the vertical axis, Matlab is used to connect the data points collected by the accelerometers in one hour to draw a two-dimensional image. The contour features of this image represent the main features of the acceleration data in one hour.

[0012] Step 2: Image classification;

[0013] Based on image contour features, bridge monitoring data are divided into six categories: normal, drift, local gain, missing, noise, and jump points.

[0014] Step 3: Use transfer learning to train the pre-trained model on the bridge monitoring dataset classified in Step 2;

[0015] (3.1) First Transfer Learning

[0016] Adjust the pre-trained ResNet50 model by adding a new fully connected layer and a classification layer. Set the InputSize parameter of the new fully connected layer and the OutputSize parameter of the new classification layer to d. Initialize the pre-trained model using the weights of the ResNet50 model instead of starting training from the pre-trained model, allowing the network's weight parameters to gradually adapt to the characteristics of different categories of bridge monitoring data. Train the pre-trained model using the bridge monitoring dataset classified in step two, retaining the obtained weights after training. Where d = 6.

[0017] (3.2) Second Transfer Learning

[0018] Using the weights obtained from the first training of the pre-trained model, the network model obtained in step (3.1) is initialized again; the bridge monitoring dataset after classification in step two is input to train the network model obtained in step (3.1) a second time, and the weights of the network model are retained after training.

[0019] (3.3) Nth Transfer Learning

[0020] In round N-1, weights are obtained and the network model is initialized. The bridge monitoring dataset classified in step two is input, and the network model obtained from the previous transfer learning is trained for the Nth time. The accuracy curves on the Nth validation set and the test set of the Nth transfer learning round are used for evaluation. At this point, the network weights have adapted to the characteristics of the highway bridge dataset, the pre-trained model can provide optimal feedback, and the model reaches a stable convergence state. Its classification accuracy for abnormal data meets the identification requirements of massive, multi-anomaly health monitoring data.

[0021] The criteria for determining the six data categories in step two are as follows: In a normal image, each data point fluctuates significantly within a range near the central axis of the image, and the overall image features are regular. Drift data may be caused by calibration failure of internal sensor components or temperature compensation failure, resulting in an overall tilt or vertical fluctuation of the central axis of the image. Local gain data may be caused by the sensor being excited by traffic flow on a bridge or by wind vibration disturbance for a certain period, resulting in an acceleration amplitude higher than other time periods for a period exceeding 20%. Missing data may be caused by a sudden power outage or malfunction of the sensor, resulting in blank data for one or more time periods in the image, displayed as only a straight line. Noise data may be caused by electromagnetic interference from the surrounding environment or the acquisition equipment not being grounded, resulting in insignificant jumps in the data points within the image, with the overall feature being a large black rectangular block. Jumping point data may be caused by a sudden, strong impact on the sensor. This results in isolated data points with large amplitudes within the image, a reduction in the overall image features, or an overall deviation of the image data from the center line by 1 / 3.

[0022] The determination steps in step (3.3) are as follows:

[0023] (3.3.1) Data point fitting;

[0024] A fourth-order polynomial is used to fit the discrete points of the validation accuracy. The four-order polynomial fitting curve can more closely match the discrete data points and reflect the true distribution of the data. On the other hand, it can fully demonstrate the characteristic that the validation accuracy fluctuates within a certain range after a certain number of iterations.

[0025] (3.3.2) Finding the optimal point of the curve;

[0026] The optimal point S is found on the accuracy curve of the test set after multiple transfer learning iterations, and transfer learning stops at point S. Even after multiple more transfer learning iterations, the recognition accuracy is unlikely to improve significantly. At this point, the model can be considered to have been sufficiently trained, converged stably, and achieved a high recognition accuracy, meeting the requirements for anomaly detection.

[0027] The S-point meets the following characteristics: the peak of the test set validation accuracy curve after multiple transfer learnings is within the range of the number of transfer learnings [m,n], and the validation accuracy shows a downward trend after n points; the number of transfer learnings at point S is within the range of [m,n], and the slope of the line connecting any two points within this range is approximately zero; point S is closest to the test set validation accuracy curve after multiple transfer learnings.

[0028] A structural monitoring anomaly data identification device based on parameter adaptive adjustment, comprising:

[0029] The data acquisition module is used to obtain raw data from bridge monitoring.

[0030] The memory is used to store the acquired bridge monitoring data and computer programs; the computer programs include image conversion programs, multiple migration programs, and data fitting programs.

[0031] A processor is configured to execute a computer program stored in the memory, wherein, when the computer program is executed, the processor is configured to:

[0032] Read the stored bridge monitoring data, which is collected and stored at the same sampling frequency; based on the acquired bridge monitoring data, use an image conversion program to convert the bridge monitoring data into images; the images are identified and classified using an image processing program, and then the identified and classified image dataset is trained using a multiple transfer learning program; the model's stability and convergence are judged using a data fitting program.

[0033] The beneficial effects of this invention are as follows: The method and apparatus proposed in this invention can optimize the configuration of transfer learning parameters without fine-tuning them. This can significantly reduce time and labor costs, while simultaneously enabling the pre-trained model to achieve optimal performance and obtain better recognition accuracy for data identification. Attached Figure Description

[0034] Figure 1 This is the time-domain image of the converted data;

[0035] Figures 2(a)-2(h) These are the main characteristics of normal time-domain image data;

[0036] Figures 3(a)-3(h) The main characteristics of time-domain image drift data;

[0037] Figures 4(a)-4(h) The main features of local gain data in the time domain image;

[0038] Figures 5(a)-5(h) The main characteristics of missing data in time-domain images;

[0039] Figures 6(a)-6(h) The main characteristics of time-domain image noise data;

[0040] Figures 7(a)-7(h) These are the main features of time-domain image jump point data;

[0041] Figure 8 This is a diagram illustrating the image recognition process of the pre-trained ResNet 50 model.

[0042] Figure 9 Flowchart for determining model stability and convergence;

[0043] Figure 10This is a fitted curve showing the change in test set accuracy with the number of transitions. Detailed Implementation

[0044] The following describes the embodiments of the present invention in conjunction with the technical solution.

[0045] Taking a certain bridge as an example, the pre-trained model ResNet50 is used to detect anomalies in its acceleration data;

[0046] Step 1: Image conversion;

[0047] The accelerometers deployed on the bridge have a sampling frequency of f = 100Hz, collecting data 100 times per second, for a total of 3600 * 100 = 360,000 acceleration data points per hour. Using time as the x-axis and the acceleration value at that time point as the y-axis, Matlab was used to connect and plot the 360,000 data points collected by the sensor in one hour into a two-dimensional image. The features of this image roughly represent the main characteristics of the acceleration data within one hour. The converted image is shown below. Figure 1 As shown.

[0048] Step 2: Classification of Abnormal Data Features

[0049] Based on the contour features of the 8250 converted images, the images were divided into six data categories: normal, drift, local gain, missing, noisy, and jump points, and stored in different folders. The main features of each data category are shown in Figures 2-7.

[0050] Step 3: Use transfer learning to train the dataset using a pre-trained model.

[0051] Using the pre-trained ResNet50 model from the MATLAB toolbox, add a new fully connected layer and a classification layer. Set the InputSize parameter of the new fully connected layer to 6, and the OutputSize parameter of the new classification layer to 6. Perform initial hyperparameter settings for the convolutional neural network: initialLearnRate - 0.01, optimizer - SGD, MiniBitchsize - 32, MaxEpochs - 1. The ResNet50 pre-trained model recognition process is as follows: Figure 8 As shown.

[0052] Step 4: Determine if the model is stable and convergent

[0053] The process for determining model stability and convergence is as follows: Figure 9 As shown.

[0054] (1) The model obtained in the previous round is used to train the dataset, and the data points of transfer learning times and recognition accuracy are obtained. The data points are then fitted using a fourth-order polynomial. The fitted data point graph is shown below. Figure 10 As shown, after the first transfer learning, the model's accuracy in recognizing images of each category was 64.6%, which is very low and cannot effectively classify the data. The transfer learning parameters were not optimally configured, so multiple transfer learning sessions were conducted to adaptively optimize the parameter configuration.

[0055] (2) During the 20th transfer learning iteration, the final accuracy rate for data image recognition was 94.1%, meeting the preliminary criteria for model stable convergence, and model stable convergence can be determined. A fitting curve of the test set accuracy changing with the number of transfer learning iterations is plotted. Figure 5 shows that this curve fits the discrete data points well, consistent with the data distribution characteristics. The slope of the curve changes significantly between the [1, 7] transfer learning iterations, indicating a rapid increase in verification accuracy. Between the [7, 20] transfer learning iterations, the verification accuracy shows a fluctuating trend with minimal amplitude changes. It is considered that the model has been sufficiently trained at this stage. Within the [7-10] transfer learning iterations, the verification accuracy reaches a peak, and the verification accuracy of the ninth transfer learning iteration is closer to the fitting curve, indicating that the model has been sufficiently trained at this point and has achieved optimal performance. Therefore, the model obtained from the 9th transfer learning iteration is considered the optimal model (9 being point S). Using the model trained at this point, accurate classification of abnormal data is achieved.

Claims

1. A method for identifying structural monitoring anomaly data with adaptive parameter adjustment, characterized in that, First, the raw bridge monitoring data is converted into images in the time domain. Based on the contour features of the converted images, the bridge monitoring data is classified. A transfer learning method is adopted, and a pre-trained model is selected to train the classified bridge monitoring data. During the data training process, parameter adaptive stability convergence judgment is performed. When the stability convergence judgment condition is met, multiple transfer learnings are terminated, the parameters are optimally configured, the pre-trained model obtains the best recognition performance, and the recognition accuracy of each category of images reaches the best. The specific steps are as follows: Step 1: Image conversion; The sampling frequency of the accelerometers deployed on the bridge is f= Hz, sampling per second This data point collects acceleration data for the bridge at 3600* per hour. Next, using time as the horizontal axis and the acceleration value at the acquisition time point as the vertical axis, Matlab is used to connect the data points collected by the accelerometer in one hour to draw a two-dimensional image. The contour features of this image represent the main features of the acceleration data in one hour. Step 2: Image classification; Based on image contour features, bridge monitoring data are divided into six categories: normal, drift, local gain, missing, noise, and jump points. Step 3: Use transfer learning to train the pre-trained model on the bridge monitoring dataset classified in Step 2; (3.1) First Transfer Learning Adjust the pre-trained ResNet50 model by setting a new fully connected layer and a classification layer. Set the InputSize of the new fully connected layer and the OutputSize of the new classification layer to d. Initialize the pre-trained model using the weights of the ResNet50 model. Train the pre-trained model using the bridge monitoring dataset classified in step two, and retain the obtained weights after training. (3.2) Second Transfer Learning Using the weights obtained from the first training of the pre-trained model, the network model obtained in step (3.1) is initialized again; the bridge monitoring dataset after classification in step two is input to train the network model obtained in step (3.1) a second time, and the weights of the network model are retained after training. (3.3) Nth Transfer Learning In round N-1, weights are obtained and the network model is initialized. The bridge monitoring dataset after classification in step two is input, and the network model obtained in the previous transfer learning is trained for the Nth time. The accuracy curve of the Nth validation set and the accuracy curve of the Nth transfer learning test set are used to make a judgment. The determination steps in step (3.3) are as follows: (3.3.1) Data point fitting; A fourth-order polynomial is used to fit the discrete points of the verification accuracy. (3.3.2) Finding the optimal point of the curve; Find the optimal point S on the accuracy curve of the test set after multiple transfer learning sessions, and stop transfer learning when point S is found. The S-point meets the following characteristics: the peak of the test set validation accuracy curve after multiple transfer learnings is within the range of the number of transfer learnings [m,n], and the validation accuracy shows a downward trend after n points; the number of transfer learnings at point S is within the range of [m,n], and the slope of the line connecting any two points within this range is approximately zero; point S is closest to the test set validation accuracy curve after multiple transfer learnings.

2. The method for identifying structural monitoring anomaly data with adaptive parameter adjustment according to claim 1, characterized in that, The criteria for determining the six data categories in step two are as follows: In images of normal data, each data point fluctuates up and down along the central axis of the image, and the fluctuation is obvious, with regular overall image features; in images of drift data, the central axis is tilted or fluctuates up and down; in images of local gain data, the acceleration amplitude is higher than that of other time periods within a certain time period, and this time period accounts for more than 20%; in images of missing data, data in a certain time period or several time periods are blank, and only a straight line is displayed; in images of noisy data, the data points do not jump significantly, and the overall feature is a black rectangular block with width; in images of jump point data, there are isolated data points with large amplitudes, and the overall image features are reduced or the overall image data deviates from the center line by 1 / 3.

3. The method for identifying structural monitoring anomaly data with adaptive parameter adjustment according to claim 1, characterized in that, The value of d is 6.

Images