A large-span bridge expansion device monitoring data anomaly intelligent diagnosis and processing method
By converting monitoring data of long-span bridge expansion joints into two-dimensional image samples and constructing a deep learning model, the problem of insufficient adaptability and generalization ability in the anomaly diagnosis of long-span bridge expansion joint monitoring data is solved, achieving efficient anomaly identification and processing, and obtaining high-quality monitoring datasets.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- RAILWAY CONSTR RES INST OF CHINA ACAD OF RAILWAY SCI CO LTD
- Filing Date
- 2026-05-14
- Publication Date
- 2026-06-19
Smart Images
Figure CN122241010A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method, and more particularly to an intelligent diagnosis and processing method for abnormal monitoring data of expansion joints in long-span bridges. Background Technology
[0002] With the increasing demands on bridge span capacity for high-speed railways in my country, long-span and ultra-long-span railway bridges, as well as dual-purpose road-rail bridges, are emerging in large numbers. Several high-speed railway bridges with spans exceeding 1,000 meters, such as the Wufengshan Yangtze River Bridge and the Shanghai-Suzhou-Tongzhou Yangtze River Bridge, have been completed. Long-span bridges are characterized by high flexibility and high load sensitivity, especially under temperature loads, where the beam ends experience significant expansion and contraction. To ensure continuous track support stiffness at the beam joints, long-span bridges typically install expansion joints at the beam ends, along with rail expansion joint adjusters. These expansion joints must work in tandem with the rail expansion joint adjusters and the beam body to ensure coordinated deformation of the track and bridge, guaranteeing smooth train passage. However, due to the complex structure of the expansion joints, which include various special components such as steel sleepers, steel longitudinal beams, and scissor beams, and numerous sliding surfaces, the coordination between components is challenging, making it prone to slippage problems. This can lead to sleeper misalignment and cracking, scissor beam jamming, and other defects. Therefore, close monitoring of their service status is crucial for timely and targeted maintenance.
[0003] In recent years, many long-span bridges have installed monitoring systems in the expansion joint areas to monitor the operational status of various components of the expansion joints in real time. However, as the monitoring systems have been in operation for an extended period, the performance of some monitoring equipment has degraded, and their stability and reliability have significantly decreased. Coupled with interference from external factors such as severe weather, transmission failures, and track maintenance, the raw monitoring data exhibits various forms of anomalies. These anomalies may alter the original numerical range and trend of the monitoring data, interfering with the acquisition of accurate service information of the expansion joints, thus affecting the accurate assessment of the structural condition of the expansion joints themselves, leading to frequent false alarms and missed alarms. Therefore, there is an urgent need to develop effective diagnostic and processing methods for anomalies in the monitoring data of expansion joints in order to eliminate monitoring anomaly interference and provide a sound data foundation for the condition analysis of the expansion joints.
[0004] Analysis of long-term monitoring data from on-site expansion joints revealed that the raw data primarily exhibited anomalies such as missing data, outliers, noise, gradual drift, and sudden drift. These anomalies varied across different monitoring indicators, and the same data segment could contain multiple anomalies. Therefore, the key lies in how to reasonably categorize anomaly samples without compromising anomaly handling, and in establishing data anomaly diagnosis methods applicable to different monitoring indicators.
[0005] In the field of anomaly identification technology for expansion joint monitoring data, there are currently no publicly available research results. However, for other civil engineering structures such as bridges, buildings, and dams, relevant research has proposed some methods for identifying anomalies in monitoring data, which can provide some reference for this invention. Among them, local outlier factor and clustering algorithms screen data anomalies in an unsupervised manner; machine learning methods such as Bayesian filtering, random forest, support vector machine, and pattern recognition neural networks classify and identify data anomalies by manually designing and extracting feature engineering; and some studies are closest to the technical solution of this invention. For a certain monitoring indicator, the time series is first plotted as image samples, and then deep learning techniques such as convolutional neural networks are used to automatically learn the features of abnormal images, thereby achieving the classification and identification of various data anomalies.
[0006] Currently, in the actual operation of expansion joint monitoring systems, some systems use consecutive null values to determine missing data; for values exceeding limits, they use threshold-based alarm methods, but since thresholds are mostly set based on experience, their rationality needs to be verified, and they cannot identify data jumps within the threshold; and no effective diagnostic methods have been proposed for other data anomalies.
[0007] Currently, existing technologies mainly have the following problems: (1) Unsupervised methods can only determine whether the data is abnormal, but cannot further clarify the type of abnormality.
[0008] (2) Machine learning methods rely on complex feature engineering. For large-scale monitoring data and diverse monitoring indicators, it is difficult to guarantee the sufficiency of the extracted abnormal features and the generalization ability of the model.
[0009] (3) Image-based deep learning methods have low accuracy in identifying some anomalies, and most of them only target one monitoring indicator without taking into account the characteristics of different monitoring indicators. They are not adaptable to other monitoring indicators outside the research object. Considering that the time-frequency domain characteristics and anomaly classification standards of other structural monitoring indicators are different from those of this study, the relevant image sample construction methods and anomaly diagnosis models are difficult to apply directly.
[0010] (4) There is a lack of standardized procedures for handling abnormal data after diagnosis, and the methods for handling different types of abnormal monitoring data of telescopic devices are not yet clear. Summary of the Invention
[0011] To address the shortcomings of existing technologies, this invention discloses an intelligent diagnosis and processing method for abnormal monitoring data of expansion joints in long-span bridges, the technical solution of which is as follows:
[0012] A method for intelligent diagnosis and processing of monitoring data anomalies is disclosed. This method converts the original time series data into two-dimensional image samples, utilizes graphic features to quickly screen for anomalies in the series data, and performs targeted processing according to the anomaly category. Its features include:
[0013] (1) Image sample library construction: Based on a self-developed algorithm in Python, the original sequence segments are captured using a sliding window and two-dimensional image samples are drawn adaptively; the image samples are classified, labeled and stored based on the anomaly classification criteria, and the anomaly samples are expanded through data augmentation algorithms;
[0014] (2) Data anomaly diagnosis: Based on the PyTorch deep learning framework, a convolutional neural network anomaly diagnosis model based on improved training methods is constructed. The model architecture and parameters are determined through parameter optimization and repeated training, and the optimal model is used to complete the classification diagnosis of samples.
[0015] (3) Data anomaly handling: Based on the sample diagnosis results, clarify the data anomaly handling process and methods; for the data anomaly sections, adopt the corresponding methods to remove and correct anomalies, and obtain high-quality datasets.
[0016] This invention also discloses an intelligent diagnostic and processing device for abnormal monitoring data of long-span bridge expansion joints. This device converts the original time series into two-dimensional image samples, utilizes graphic features to quickly screen for anomalies in the sequence data, and performs targeted processing according to the anomaly category. Its features include:
[0017] (1) Image sample library construction module: Based on a self-developed Python algorithm, the original sequence segments are captured using a sliding window and two-dimensional image samples are drawn adaptively; the image samples are classified, labeled and stored based on the anomaly classification criteria, and the anomaly samples are expanded through data augmentation algorithms;
[0018] (2) Data anomaly diagnosis module: Based on the PyTorch deep learning framework, a convolutional neural network anomaly diagnosis model based on improved training methods is constructed. The model architecture and parameters are determined through parameter optimization and repeated training, and the optimal model is used to complete the classification diagnosis of samples.
[0019] (3) Data anomaly handling module: Based on the sample diagnosis results, clarify the data anomaly handling process and methods; for the data anomaly section, adopt the corresponding methods to remove and correct the anomalies and obtain a high-quality dataset.
[0020] The present invention also discloses a non-volatile storage medium, characterized in that the non-volatile storage medium includes a stored program, wherein the program, when running, controls the device where the non-volatile storage medium is located to execute the above-described method.
[0021] The present invention also discloses a terminal device, characterized in that the terminal device includes: a processor, a memory, a communication interface, and a bus; the processor, the memory, and the communication interface are connected through the bus and communicate with each other; the memory stores executable program code; the processor reads the executable program code stored in the memory to run a program corresponding to the executable program code, so as to execute the above-described method.
[0022] Beneficial effects
[0023] (1) The proposed image sample library construction method has good adaptability to different monitoring indicators and different abnormal forms. It can manage different indicator abnormalities in a unified manner without adding extra interference information, which is convenient for subsequent abnormal diagnosis using the same model.
[0024] (2) The constructed data anomaly diagnosis model does not require separate design of model parameters for different monitoring indicators. Based on the same model architecture, it can classify and diagnose data anomalies for each indicator, has strong generalization ability, and the diagnostic performance of each anomaly category is significantly improved compared with the existing technical solutions.
[0025] (3) The proposed standardized processing flow and method for abnormal monitoring data can effectively process abnormal data of each category in sequence, improve the quality and efficiency of abnormal monitoring data processing, and thus obtain a high-quality monitoring dataset. Attached Figure Description
[0026] Figure 1 This is a schematic diagram of the overall technical route of the present invention;
[0027] Figure 2 This is a schematic diagram of image samples labeled with different monitoring indicators of the present invention;
[0028] Figure 3 This is a schematic diagram of the data anomaly diagnosis model architecture of the present invention;
[0029] Figure 4 This is a schematic diagram of the abnormal data handling process of the present invention. Detailed Implementation
[0030] This invention proposes an intelligent diagnostic and processing device for monitoring data anomalies in various indicators of expansion joints in long-span bridges. It converts the original time series data into two-dimensional image samples, utilizes graphic features to quickly screen for anomalies in the series data, and performs targeted processing according to the anomaly category. The overall technical approach is as follows: Figure 1 As shown, it mainly includes three modules: image sample library construction, data anomaly diagnosis, and data anomaly handling.
[0031] Image Sample Library Construction Module
[0032] This program, developed independently using Python, constructs an image sample library through sequence-to-image conversion. The main processes include raw sequence data extraction, adaptive generation of 2D images, sample classification and labeling, and anomalous sample augmentation. The specific workflow is as follows:
[0033] (1) Raw sequence data capture: Construct a sliding window and extract date-indicator sequence segments segment by segment along the monitoring date. Since all the monitoring indicators of the telescopic device have obvious daily cycle characteristics, the sliding window length is set to 24 hours and the translation step size is set to 1 hour. The setting of the sliding window step size in this patent takes into account the timeliness of data screening, the diversity of sample characteristics and the redundancy of sample storage in the on-site monitoring system. In practical applications, it can be adjusted according to the frequency requirements of abnormal diagnosis.
[0034] (2) Adaptive Generation of Two-Dimensional Images: Based on the extracted sequence segments, a two-dimensional image is drawn with the date as the horizontal axis and the index value as the vertical axis. The image is set to RGB format, the aspect ratio is set to 1, the coordinate axes are set to not be displayed, and the line margin is set to 0 (i.e., the line fully covers the paper). In the image, to better distinguish the graphic features of different abnormal categories, red dots and blue lines are used to represent the sequence waveform; the horizontal axis range is set to (start date of the extracted segment, start date of the extracted segment + 24 hours); the vertical axis range is set to (normal minimum value of the measuring point, normal maximum value of the measuring point), that is, the normal distribution range of the measuring point values. The set vertical axis range can adaptively change with the change of the measuring points. In this technical solution, the box plot method is used to calculate and determine (based on the historical monitoring data of each measuring point (requiring more than one year, covering the complete high and low temperature cycle) for statistical analysis, the lower limit of the vertical axis is Q1-1.5×IQR, the upper limit is Q3+1.5×IQR, Q3 and Q1 are the data quartiles, IQR=Q3-Q1), which can be adjusted according to the needs in actual applications. This plotting method takes into account the differences in the data distribution of various monitoring indicators of the expansion joint, establishing a unified measurement scale for different indicators. This ensures that anomalies are detected while avoiding additional interference caused by excessive data scaling. Furthermore, to facilitate the recording and retrieval of specific information about data anomalies, images are saved using the format "Measurement Point_Start Date_Start Time". For example, "01D1_20240620_15:00" indicates that this sample represents data from the left end measurement point of the distance between the fixed steel sleepers on both sides of the No. 1 expansion joint area, from 15:00 on June 20, 2024 to 15:00 on June 21, 2024.
[0035] (3) Sample classification and labeling: Since the abnormal forms of different monitoring indicators of the telescopic device are different (e.g., the distance between the two fixed steel sleepers often shows data noise, that is, the data points are significantly deviated from the adjacent data; the distance between the movable steel sleeper and the fixed steel sleeper and the distance between the movable steel sleepers often show an overall change in the data baseline due to the operation of the manual square sleeper), an abnormal classification criterion is established based on the abnormal graphic characteristics of the telescopic device monitoring data, and the abnormal forms of different monitoring indicators are unified and divided into three categories: missing (blank or broken line segments in the figure), jump point (some data points are significantly deviated from the normal range or adjacent data), and drift (the trend of the sequence waveform deviates from the original baseline). With the help of manual screening, the generated samples are marked as "normal (0)", "missing (1)", "jump point (2)" and "drift (3)" according to the abnormal classification criterion. For samples containing multiple abnormal categories, they are labeled according to the priority of "drift > jump point > missing", with drift having the highest priority and missing having the lowest priority. According to the classification criterion, the image samples of each category labeled by different monitoring indicators are as follows Figure 2 As shown in the figure, the monitoring indicators used include the distance between the fixed steel sleepers on both sides of the telescopic device (sampling frequency of 1 min / time), the distance between the fixed steel sleeper and the movable steel sleeper (sampling frequency of 10 min / time), and the distance between the movable steel sleepers (sampling frequency of 10 min / time). It can be seen that the established classification criteria can effectively distinguish samples of different categories, and the proposed sample construction method can better unify the scale of different monitoring indicators, ensure the consistency of the characteristics of samples of the same category, and avoid interference caused by differences in monitoring indicator values, sampling frequencies, etc.
[0036] (4) Abnormal Sample Expansion: To avoid an imbalance between the number of normal and abnormal categories, which could lead to overfitting of the diagnostic model to the features of normal samples and underfitting of the features of abnormal samples, horizontal and vertical flipping data augmentation methods were used after labeling to increase the number of abnormal samples. This further expanded the diversity of abnormal sample features without adding interfering features, and covered as much as possible the possible locations and forms of abnormalities in reality. Among them, horizontally flipped samples were named with the original sample name plus "-h" suffix, and vertically flipped samples were named with the original sample name plus "-v" suffix, thus completing the construction of the image sample library.
[0037] Data Anomaly Diagnosis Module
[0038] Based on the PyTorch deep learning framework, this patent constructs a convolutional neural network data anomaly diagnosis model, mainly comprising an input layer, convolutional layers, pooling layers, and fully connected layers. By optimizing structural parameters such as input sample size, convolutional kernel size, number of convolutional-pooling layers, and fully connected layer structure, the optimal diagnostic model architecture is obtained. Figure 3 As shown.
[0039] (1) Input layer: The image samples are processed in batches using a data loader and then input into the model for training and testing. The batch size is set to 128. In the batch processing stage, each dataset sample is set to 64 pixels × 64 pixels × 3 channels and normalized. The training set samples need to be randomly shuffled.
[0040] (2) Convolutional-pooling layers: Four sets of convolutional-pooling layers are used to automatically learn and extract features from the input samples. The kernel size of the convolutional layer is 9×9, the padding value is 5, and the stride is 1 to ensure that the size of the feature map before and after input remains unchanged. The ReLU function is used to activate the output of the convolutional layer. The number of channels of each set of convolutional layers is set to 6, 12, 24, and 48, respectively. The filter size of the pooling layer is 2×2, the stride is 2, and max pooling is used for processing.
[0041] (3) Fully connected layer
[0042] The feature map output from the last pooling layer is flattened into a 1×768 vector and fed into a fully connected layer with 128 nodes. The ReLU function is used to activate the output, and the Dropout probability is set to 0.25 (randomly losing 25% of node information during training) to avoid overfitting. The processed result is then passed through a fully connected layer with 4 nodes to obtain a four-dimensional vector representing the probability of the input sample relative to each anomaly category. The anomaly category of the sample can be determined by searching for the index of the maximum value in the output vector.
[0043] Compared to traditional convolutional neural networks, this model uses the AdamW optimizer to decouple weight decay from gradient calculation, and employs the OneCycleLR learning rate scheduler to dynamically adjust the learning rate, thereby accelerating model convergence, preventing the model from getting stuck in local optima, and improving the model's diagnostic performance and generalization ability. Through parameter optimization training, the maximum learning rate and weight decay were determined to be 1e-4, and a linear annealing learning rate variation strategy was adopted. The learning rate variation and optimizer parameter update process are shown in equations (1) and (2).
[0044] (1)
[0045] In the formula, α t Let α0 and αt be the learning rate for the t-th iteration max and α T These are the initial learning rate, maximum learning rate, and final learning rate, respectively, with α0 = α by default. max / 25、α T =α max / 10000; T, T rise and T fallThese represent the total number of iterations, the number of iterations during the learning rate rise phase, and the number of iterations during the learning rate fall phase, respectively, with a default value of T. rise / T=0.3.
[0046] (2)
[0047] In the formula, θ t Here are the updated model parameters, and λ is the weight decay coefficient. and For the estimation of the first and second moments after bias correction, ε is a minimal constant.
[0048] During the training process, the number of training cycles was set to 2000. An automatic mixed precision training method was adopted, which uses autocast to automatically select the data type and GradScaler to dynamically scale the gradient value. The two work together to ensure the efficiency and stability of the training. The cross-entropy loss function (nn.functional.Cross entropy) was used to calculate the loss. The Softmax operation can be applied automatically, and the negative log-likelihood loss can be calculated, as shown in Equation (3).
[0049] (3)
[0050] In the formula, N is the batch size, C is the number of sample categories, and z j y is the original output vector of the model. n The true labels for the samples.
[0051] Based on 40,000 image samples with multiple monitoring indicators (training:validation:test = 6:2:2), the proposed diagnostic model was trained and tested. The F1-score of each anomaly category on the test set exceeded 99%, and the average diagnostic accuracy of all categories reached 99.6%. The training time for a single session was 1.77 seconds, which well balances computational performance and efficiency, and effectively solves the problem of extremely low accuracy in identifying some anomaly categories in existing technical solutions.
[0052] Data anomaly handling module
[0053] Based on the anomaly classification and diagnosis results of monitoring data from telescopic devices, a set of monitoring data anomaly handling procedures and methods are proposed. For anomaly segments, corresponding methods are used for anomaly removal and correction, thereby obtaining a high-quality dataset that can be used for subsequent structural state analysis and evaluation. The specific process is as follows: Figure 4 As shown.
[0054] (1) Data jump point processing: including outliers and noise. ① Outlier data: Based on historical monitoring data of each measuring point (requiring more than one year and covering the complete high and low temperature cycle), the box plot method is used to calculate the upper and lower quartiles Q3, Q1 and interquartile range IQR, thereby determining the outlier boundary [Q1-1.5×IQR, Q3+1.5×IQR]. That is, when the data point y < Q1-1.5×IQR or y > Q3+1.5×IQR, it needs to be removed, and supplemented and corrected according to the data missing processing method according to actual needs. ② Noise data: Discrete wavelet transform is used to decompose and reconstruct the original data, filter out high-frequency noise details, and retain only the low-frequency approximate signal. Through experiments, the sym8 wavelet is selected as the wavelet basis, and the number of decomposition levels needs to be determined according to the Nyquist sampling frequency and the sampling frequency of the monitoring index. The monitoring indicators used in this patent mainly include two sampling frequencies: 1 minute and 10 minutes. For monitoring indicators with a sampling frequency of 1 minute, the number of decomposition layers is set to 7; for monitoring indicators with a sampling frequency of 10 minutes, the number of decomposition layers is set to 3.
[0055] (2) Data drift handling: Generally, data drift is caused by monitoring equipment failure, and the structural state has not changed. After removing the abnormal sections, the data can be corrected according to the data missing handling method. It should be noted that due to the special structure of the telescopic device, jamming problems often occur, requiring manual maintenance during maintenance windows. This may force adjustments to the shape and position of components such as movable steel sleepers, resulting in changes to the structural state and corresponding drift anomalies in the monitoring data. This type of drift is not significantly different from the drift caused by monitoring equipment failure. Therefore, based on the time of the anomaly (0:00~6:00) and the on-site personnel's work records, this type of anomaly is screened a second time. Only the anomaly is marked but not processed, and the abnormal data is retained.
[0056] (3) Data missing handling: This includes diagnosed missing data and missing data caused by outlier and drift data removal. Measurement points x with strong correlation to the missing measurement point y (should be the same monitoring indicator, with a Pearson correlation coefficient exceeding 0.95) and good data quality (no anomalies within the missing time period) are selected. Based on the linear correlation between measurement points, a linear regression model y=kx+b is constructed. The model is fitted using historical data from more than one year of complete high and low temperature cycles to obtain the fitting parameters k and b. Thus, the measurement point y data is supplemented using measurement point x from the missing time period. Verification using normal segment data shows that the supplemented data based on the regression model has the same trend and amplitude as the original data, which can meet the needs of anomaly correction.
[0057] Example 2
[0058] Based on Embodiment 1, this embodiment also discloses an intelligent diagnosis and processing method for abnormal monitoring data of long-span bridge expansion joints. This method converts the original time series into two-dimensional image samples, utilizes graphic features to quickly screen for anomalies in the series data, and performs targeted processing according to the anomaly category. Its features are as follows:
[0059] (1) Image sample library construction: Based on a self-developed algorithm in Python, the original sequence segments are captured using a sliding window and two-dimensional image samples are drawn adaptively; the image samples are classified, labeled and stored based on the anomaly classification criteria, and the anomaly samples are expanded through data augmentation algorithms;
[0060] (2) Data anomaly diagnosis: Based on the PyTorch deep learning framework, a convolutional neural network anomaly diagnosis model based on improved training methods is constructed. The model architecture and parameters are determined through parameter optimization and repeated training, and the optimal model is used to complete the classification diagnosis of samples.
[0061] (3) Data anomaly handling: Based on the sample diagnosis results, clarify the data anomaly handling process and methods; for the data anomaly sections, adopt the corresponding methods to remove and correct anomalies, and obtain high-quality datasets.
[0062] This invention, based on the abnormal characteristics of monitoring indicators of expansion joints, formulates a classification criterion for data anomalies in the monitoring indicators of long-span bridge expansion joints. It proposes a sample library construction method for sequence-image adaptive conversion, which adaptively changes the coordinate range according to different monitoring indicators to automatically generate image samples. Furthermore, it increases the diversity of abnormal sample features through sliding window overlap and data augmentation, thereby comprehensively considering the features of different monitoring indicators and various anomaly forms, and preserving the most complete anomaly information. An improved convolutional neural network anomaly diagnosis model is established, introducing advanced technologies such as the AdamW optimizer, OnecycleLR learning rate scheduler, and automatic mixed-precision training algorithm into the traditional convolutional neural network. This comprehensively improves the model's diagnostic accuracy, computational efficiency, and generalization ability, achieving accurate and rapid classification and diagnosis of anomalies in various monitoring indicators of expansion joints.
[0063] This invention addresses data outliers and noise using box plots and discrete wavelet transforms respectively, and supplements missing data based on the linear correlation between measurement points, effectively improving the quality of monitoring data and meeting the needs of subsequent structural state analysis.
[0064] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.
Claims
1. A method for intelligent diagnosis and processing of anomalies in monitoring data of expansion joints in long-span bridges, wherein the method converts the original time series into two-dimensional image samples, utilizes graphic features to quickly screen for anomalies in the series data, and performs targeted processing according to the anomaly category; characterized in that, Includes the following steps: (1) Image sample library construction: Based on a self-developed Python algorithm, the original sequence segments are captured using a sliding window and two-dimensional image samples are drawn adaptively; the image samples are classified, labeled and stored based on the anomaly classification criteria, and the anomaly samples are expanded through data augmentation algorithms; (2) Data anomaly diagnosis: Based on the PyTorch deep learning framework, a convolutional neural network anomaly diagnosis model based on improved training method is constructed. The model architecture and parameters are determined through parameter optimization and repeated training, and the optimal model is used to complete the classification diagnosis of samples. (3) Data anomaly handling: Based on the sample diagnosis results, clarify the data anomaly handling process and methods; for the data anomaly sections, adopt the corresponding methods to remove and correct anomalies, and obtain high-quality datasets.
2. The method according to claim 1, characterized in that, In step (1): The adaptive two-dimensional image sample is plotted as an RGB format image with date as the horizontal axis and index value as the vertical axis. The image aspect ratio is 1, the coordinate axes are not displayed, and the lines fill the entire paper. The range of the vertical axis is adaptively determined according to the normal value distribution range of the measuring points. Based on the historical monitoring data of each measuring point, the box plot method is used for calculation. The lower limit of the vertical axis is Q1-1.5×IQR, and the upper limit is Q3+1.5×IQR. Q3 and Q1 are the data quartiles, and IQR=Q3-Q1. The anomaly classification criteria divide anomalies into three categories: missing, skip points, and drift, and label samples as four categories: "normal", "missing", "skip points", and "drift". For samples containing multiple anomaly categories, they are labeled according to the priority of "drift > skip points > missing". The data augmentation algorithm employs horizontal and vertical flipping methods.
3. The method according to claim 1, characterized in that, In step (2), the convolutional neural network anomaly diagnosis model includes an input layer, a convolutional layer, a pooling layer, and a fully connected layer; wherein: The input layer performs batch processing on the image samples, with the batch processing size set to 128 and the sample size set to 64 pixels × 64 pixels × 3 channels, and performs normalization processing. Four groups of convolutional and pooling layers are used. The convolutional kernel size of the convolutional layer is 9×9, the padding value is 5, the stride is 1, and the ReLU activation function is used. The number of channels in each group of convolutional layers are 6, 12, 24, and 48, respectively. The filter size of the pooling layer is 2×2, the stride is 2, and max pooling is used. The fully connected layer includes a fully connected layer with 128 nodes and a fully connected layer with 4 nodes. The fully connected layer with 128 nodes has a Dropout probability of 0.
25.
4. The method according to claim 1, characterized in that, In step (2), the improved training method includes: using the AdamW optimizer with a weight decay coefficient of 1e-4; The OneCycleLR learning rate scheduler is used to dynamically adjust the learning rate, with the maximum learning rate set to 1e-4. An automatic mixed-precision training method is adopted, utilizing the autocast and GradScaler components to work together. The cross-entropy loss function is used for loss calculation.
5. The method according to claim 2, characterized in that, In step (3), the data anomaly handling specifically includes: For outliers in the data jump points, the box plot method is used to determine the outlier boundary [Q1-1.5×IQR, Q3+1.5×IQR]. Data points outside the boundary are removed, and missing data handling methods are used to supplement them as needed. For noisy data in the data jump points, discrete wavelet transform is used for filtering. The sym8 wavelet is selected as the wavelet basis, and the corresponding decomposition level is determined according to the sampling frequency of the monitoring index based on the Nyquist sampling theorem. For data drift, the abnormal segment is determined to be removed or retained depending on whether its corresponding structural state has changed. After removal, the missing data is supplemented as needed using data missing handling methods. For missing data, another measurement point with strong correlation to the missing measurement point and good data quality is selected, and the missing data is supplemented based on a linear regression fitting model.
6. The method according to claim 1, characterized in that, The method is applied to the monitoring data of the expansion joint of a long-span bridge, and the monitoring data comes from the monitoring points in the expansion joint area.
7. A smart diagnostic and processing system for abnormal monitoring data of expansion joints in long-span bridges, characterized in that, The system is used to implement the method as described in any one of claims 1 to 6, comprising: The image sample library construction module is used to capture original sequence segments based on a sliding window, adaptively draw two-dimensional image samples, and perform classification labeling and sample expansion. The data anomaly diagnosis module is used to perform anomaly classification and diagnosis on image samples based on a convolutional neural network model with improved training methods. The data anomaly handling module is used to remove and correct abnormal data segments based on the diagnostic results.
8. The system according to claim 7, characterized in that, The image sample library construction module is configured as follows: Capture sequence segments using a sliding window with a length of 24 hours and a step size of 1 hour; Using the date as the horizontal axis, and adaptively determining the range of the vertical axis based on the box plot method, a two-dimensional image in RGB format is drawn. The samples are labeled according to the classification criteria of missing, skip points, and drift, as well as priority rules; Data augmentation is performed using horizontal and vertical flipping techniques.
9. The system according to claim 7, characterized in that, The data anomaly diagnosis module includes a convolutional neural network, which comprises four sets of convolutional-pooling layers and fully connected layers; and uses the AdamW optimizer, OneCycleLR learning rate scheduler, and automatic mixed-precision training method for model training.
10. The system according to claim 7, characterized in that, The data anomaly handling module is configured as follows: Outlier data is processed using the box plot method, and noisy data is processed using discrete wavelet transform based on sym8 wavelet. Process drift data based on the results of structural state change assessment; Missing data were supplemented based on a linear regression model between measurement points.
11. A non-volatile storage medium, characterized in that, The non-volatile storage medium includes a stored program, wherein the program, when executed, controls the device where the non-volatile storage medium is located to perform the method described in any one of claims 1 to 6.
12. A terminal device, characterized in that, The terminal device includes: a processor, a memory, a communication interface, and a bus; the processor, the memory, and the communication interface are connected through the bus and communicate with each other; the memory stores executable program code; the processor reads the executable program code stored in the memory to run a program corresponding to the executable program code, so as to perform the method as described in any one of claims 1-6 above.