Cable temperature missing data interpolation and long-period prediction method and system based on multi-dimensional meteorological features

Abstract

The method and system for cable temperature missing data interpolation and long-period prediction based on multi-dimensional meteorological characteristics relate to the technical field of bridge structure health monitoring and time series data processing. The method comprises the following steps: obtaining cable temperature and multi-dimensional meteorological data and performing cleaning; constructing a pre-processed time series data set and generating a missing label matrix; extracting features by using a sliding window technique and constructing a fusion feature vector; dividing samples into a complete sample set and a sample set to be interpolated; constructing and training a Transformer time series prediction model; inputting the sample to be interpolated into the trained model to output the temperature interpolation value of the missing period, thereby obtaining a complete cable temperature time series data set. In addition, the complete cable temperature time series data set is used as a historical input sequence, and the trained model is inputted, so that the temperature change trend prediction result of the cable structure in a future target period can be outputted. The method and system can realize missing data repair and temperature change trend prediction of cable structure temperature monitoring data.

Description

Technical Field

[0001] This invention relates to the field of bridge structural health monitoring and time-series data processing technology, and in particular to a method and system for interpolating missing cable temperature data and predicting long-term data based on multidimensional meteorological characteristics. Background Technology

[0002] Structural health monitoring technology collects real-time response signals by deploying sensors on engineering structures and combines them with intelligent algorithms to extract structural status information contained in the signals. It has significant advantages such as strong real-time performance, low operation and maintenance costs, and low manpower input, and has been widely used in bridges, buildings, wind power and other fields.

[0003] With the continuous advancement of transportation infrastructure construction in my country, long-span cable-stayed bridges, as a crucial component of key transportation hubs, are of paramount importance for their structural safety and long-term stable operation. As the core load-bearing component of cable-stayed bridges, the temperature changes of the cables directly trigger thermal expansion and contraction of the materials, thus affecting the overall stress balance and deformation state of the structure. This is a key factor leading to temperature stress in the bridge structure, impacting track smoothness and traffic safety. Currently, research on the temperature characteristics of cable structures largely relies on traditional thermodynamic calculations or finite element simulation methods. However, due to the slender structure of cables, the complex service environment, and the difficulty in accurately modeling the dynamic changes of actual meteorological conditions (such as temperature, humidity, and radiation), traditional methods struggle to accurately reproduce the time-varying patterns of cable temperature and cannot effectively address the issue of missing monitoring data.

[0004] In practical engineering scenarios, cable-stayed bridge temperature monitoring data often suffers from varying degrees of loss due to factors such as sensor failure, data transmission interruption, and harsh environmental interference. This results in incomplete temperature time-series data, directly impacting the accuracy of structural health assessments and the reliability of temperature effect predictions. Meanwhile, the long-term operation of structural health monitoring systems has accumulated massive amounts of cable-stayed bridge temperature and multi-dimensional meteorological monitoring data. This data contains the coupling patterns between cable-stayed bridge temperature and environmental factors, as well as diurnal and seasonal variation characteristics. Therefore, from a "data-driven" perspective, there is an urgent need to develop an efficient method for interpolating missing cable-stayed bridge temperature data and performing long-term predictions. This would solve the problem of monitoring data integrity, accurately capture the patterns of cable-stayed bridge temperature changes, provide reliable data support for early warning and operation and maintenance decisions for cable-stayed bridge structures, and ensure the long-term stable operation of bridges. Summary of the Invention

[0005] In addressing the problem that temperature monitoring data is often missing in existing cable-stayed structure health monitoring due to factors such as sensor failure, signal transmission interruption, and extreme environmental interference, thus affecting the accuracy of structural condition assessment, and in order to overcome the shortcomings of traditional time series analysis methods in accurately capturing complex time-varying patterns and failing to achieve long-term prediction, this invention provides a method for intelligent interpolation and long-term prediction of missing cable temperature data based on multidimensional meteorological characteristics.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] In a first aspect, the present invention provides a method for interpolating missing cable temperature data based on multidimensional meteorological characteristics, so as to realize the repair of missing monitoring data of cable structure temperature.

[0008] Specifically, the method for imputing missing temperature data in LAC based on multidimensional meteorological characteristics includes the following steps:

[0009] Step 1: Data Acquisition and Preprocessing

[0010] Acquire cable temperature data and multidimensional meteorological data, wherein the multidimensional meteorological data includes environmental parameters that affect cable temperature changes;

[0011] The raw data is cleaned to construct a preprocessed time-series dataset.

[0012] Step 2: Identification and Marking of Missing Data

[0013] Missing data detection is performed on the preprocessed time series dataset to generate a missing data marker matrix;

[0014] Step 3: Feature Sample Construction

[0015] The sliding window technique is used to extract features from time-series datasets and construct fused feature vectors.

[0016] Based on data integrity, the extracted samples are divided into a complete sample set for model training and a sample set to be imputed for missing data imputation.

[0017] Step 4: Transformer Model Training

[0018] Construct a Transformer time series prediction model;

[0019] Using the model training sample set, the model is iteratively trained with the goal of minimizing the prediction error until the model meets the preset convergence condition;

[0020] Step 5: Imputing missing data:

[0021] Input the sample set to be interpolated into the trained Transformer time series prediction model, output the temperature interpolation values ​​for the missing time periods, and combine them with the original data to obtain the complete Lasso temperature time series dataset.

[0022] Furthermore, in step one, the environmental parameters include atmospheric temperature, atmospheric humidity, wind speed, wind direction, and radiation (usually referring to total radiation).

[0023] Furthermore, in step one, cleaning the original data includes: performing validity screening and removing outliers and duplicate data.

[0024] Furthermore, in step one, the preprocessing also includes:

[0025] The collected raw data is sorted by timestamp to ensure the temporal continuity of the data.

[0026] Furthermore, in step one, the preprocessing also includes:

[0027] The multidimensional meteorological feature data and Lasso temperature data were processed using the Z-score normalization method (e.g., normalization) to eliminate dimensional differences.

[0028] Furthermore, in step two, based on the principle of temporal continuity, missing data detection is performed on the preprocessed temporal dataset.

[0029] Furthermore, in step two, the missing detection includes: determining the missing time period, missing duration, and missing location.

[0030] Furthermore, in step three, the fused feature vector includes time dimension features, meteorological environment features, and historical temperature features.

[0031] Furthermore, in step three, the sliding window technology is an adaptive sliding window technology.

[0032] Furthermore, in step three, the window size of the sliding window technology is dynamically adjusted according to the missing duration, wherein the window size is positively correlated with the missing duration, and the window size is greater than the missing duration.

[0033] The dynamic adjustment of the window size based on the duration of the missing information includes:

[0034] When the missing duration is less than or equal to the first threshold (i.e., when the missing duration is short), the window size is set to the first preset value;

[0035] When the missing duration is greater than the first threshold and less than or equal to the second threshold (i.e., when the missing duration is moderate), the window size is set to the second preset value.

[0036] When the missing duration exceeds the second threshold (i.e., when the missing duration is relatively long), the window size is set to the third preset value.

[0037] For example, the dynamic adjustment of the window size based on the duration of the missing information includes:

[0038] When the missing duration is ≤12 hours, the window size is set to 24 hours;

[0039] When 12 hours < missing duration ≤ 48 hours, the window size is set to 72 hours;

[0040] When the missing duration is greater than 48 hours, the window size is set to 168 hours.

[0041] Furthermore, in step three, the sample division includes: for data segments containing missing positions, extracting a sample set to be imputed centered on the missing time period; for complete data segments without missing positions, extracting a sample set to be used for model training.

[0042] It is worth mentioning that by dividing the training set and the set to be interpolated, the model's interpolation logic can be made more rigorous and its generalization ability stronger.

[0043] Furthermore, in step four, the Transformer time series prediction model is a position-encoded Transformer time series prediction model.

[0044] Further, in step four, the step of using the model training sample set to iteratively train the model with the goal of minimizing the prediction error until the model meets the preset convergence condition includes:

[0045] The model training sample set is divided into a training set and a validation set;

[0046] The model is iteratively trained using an adaptive learning rate scheduling strategy.

[0047] The iterative training uses the Huber loss function as the optimization objective and takes the prediction error on the validation set being less than a preset threshold as the convergence condition.

[0048] It is worth mentioning that by introducing the Transformer model, which has strong time series modeling capabilities, and using the Huber loss function to replace the traditional mean square error, the interference of extreme weather or sensor outliers on model training can be effectively overcome.

[0049] Furthermore, the iterative training of the model with the goal of minimizing the prediction error specifically refers to constructing a self-supervised loss function based on Huber loss, the calculation formula of which is:

[0050] .

[0051] Furthermore, the Huber function is specifically defined as (where... correspond , correspond ):

[0052] .

[0053] Furthermore, in step five, after inputting the sample set to be interpolated into the trained Transformer time series prediction model, the model extracts the feature patterns (such as the captured diurnal temperature periodicity, monthly seasonality, and the coupling relationship between meteorological features and Lasso temperature) to output the temperature interpolation values ​​for the missing time period.

[0054] Secondly, this invention provides a long-term prediction method for cable temperature based on multi-dimensional meteorological characteristics, so as to predict the trend of temperature change in cable structures.

[0055] Specifically, the long-term prediction method for cable temperature based on multidimensional meteorological characteristics includes the steps in the cable temperature missing data imputation method based on multidimensional meteorological characteristics; it also includes the following steps:

[0056] Step Six: Long-Term Temperature Prediction

[0057] The complete time-series dataset of cable temperature is used as the historical input sequence and input into the trained Transformer time-series prediction model. The output is the predicted result of the temperature change trend of the cable structure in the future target period.

[0058] Furthermore, the duration of the future target period is 720 to 1440 hours (1 to 2 months), thus achieving long-term forecasting.

[0059] Furthermore, the present invention also provides a cable temperature monitoring and data processing system, which includes:

[0060] The data acquisition layer includes a cable temperature sensor deployed on the cable to be tested, and a meteorological parameter sensor for collecting multi-dimensional meteorological parameters that affect the temperature change of the cable.

[0061] The data transmission and storage layer includes a data acquisition module that samples at a preset period (e.g., 30 minutes / time), as well as a data transmission module and a data storage module for transmitting the acquired data to a remote location;

[0062] The data analysis layer includes a computer analysis module with embedded computer programs;

[0063] The computer analysis module is configured to execute the cable temperature missing data interpolation method based on multidimensional meteorological characteristics, or to execute the cable temperature long-term prediction method based on multidimensional meteorological characteristics.

[0064] Furthermore, the cable temperature sensor is deployed at a critical section of the cable under test (cable-stayed bridge cable).

[0065] Furthermore, the multidimensional meteorological parameters affecting cable temperature changes include atmospheric temperature, atmospheric humidity, wind speed, wind direction, and radiation (usually referring to total radiation).

[0066] Furthermore, the data transmission module is a wireless transmission module responsible for uploading data in real time.

[0067] Furthermore, the data storage module is a cloud storage module for centralized management.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] (1) Breakthrough solution to the problem of missing data: This invention adopts a "data-driven" model and innovatively introduces multi-dimensional meteorological features as contextual information. By constructing a fused feature vector through sliding window technology, meteorological factors are deeply coupled with the historical time-series features of LAC temperature. Even when there is a long period of missing LAC temperature data (forming samples to be interpolated), this strong coupling relationship can still be used to achieve high-precision intelligent interpolation.

[0070] (2) Achieving a leap from "historical restoration" to "future early warning": This invention not only solves the problem of missing historical data, but also provides a high-quality data foundation for long-term prediction. Through the high-quality interpolation of this invention, it can support subsequent long-term prediction tasks, enabling it to accurately output the cable temperature evolution trend within the future target period (such as 720 to 1440 hours), thereby providing reliable data support for the health assessment, fatigue life prediction and active safety early warning of cable-stayed bridge cable structures. Attached Figure Description

[0071] Figure 1 This is a schematic diagram of the monitoring system of the present invention;

[0072] Figure 2 This is a schematic flowchart of the method of the present invention;

[0073] Figure 3 This is a schematic diagram of a cable-stayed bridge structure based on the method of this invention.

[0074] Figure 4 This is a schematic diagram of the cable temperature monitoring device in the embodiment; where the unit is cm.

[0075] Figure 5 This is a schematic diagram illustrating the loss changes during the training process based on the method of this invention;

[0076] Figure 6 This is a comparison curve of the predicted temperature of the cable structure based on the method of this invention on the validation set with the actual monitored values;

[0077] Figure 7 This is a schematic diagram of the temperature interpolation and prediction results of a cable structure on a prediction set containing missing values ​​based on the method of this invention;

[0078] Figure 8 The image shown is a physical diagram of the cable segment test device in the embodiment. Detailed Implementation

[0079] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto.

[0080] In summary, this invention provides an intelligent interpolation and long-term prediction method for missing cable temperature data based on multi-dimensional meteorological characteristics. This method relies on a cable temperature monitoring system and remote data transmission technology. First, it filters and removes outliers from the collected cable temperature and multi-dimensional meteorological time-series data. Then, based on the Transformer time-series prediction model and combined with location coding to capture the diurnal and monthly seasonal patterns of temperature, it performs high-precision intelligent interpolation of cable temperature data for missing periods. Finally, based on the interpolated complete time-series data, it further predicts the future long-term (monthly / yearly) structural temperature change trend of the cable. This solution solves the problem of missing values ​​caused by sensor failure and data transmission interruption in cable temperature monitoring. It offers high interpolation accuracy and reliable prediction results, completely restoring the full-cycle temperature change characteristics of the cable. It provides accurate data support for cable structure health status assessment, safety early warning, and operation and maintenance decisions, and has the advantages of wide applicability, high interpolation efficiency, and stable prediction results.

[0081] The implementation of the method of this invention relies on a cable structure health monitoring system (i.e., a cable temperature monitoring system). For example... Figure 1 As shown, the monitoring system includes: cable temperature sensors deployed at key sections of the cable-stayed bridge cables; meteorological parameter sensors that simultaneously collect multi-dimensional environmental characteristic data such as atmospheric temperature, atmospheric humidity, wind speed, wind direction, and total radiation; a data acquisition module that samples at a preset cycle (e.g., 30 minutes / time); a wireless transmission module responsible for uploading data in real time; a cloud storage module for centralized management; and a computer analysis module with the intelligent algorithm of this invention embedded within it.

[0082] The computer analysis module accesses data from the cloud storage module and specifically executes the following steps (i.e., the method provided by this invention, the process of which can be found in [reference]). Figure 2 ):

[0083] Step 1: Data acquisition and preprocessing. The collected raw data of cable structure temperature and multidimensional meteorological characteristics are screened for validity, outliers and duplicate data are removed, sorted by timestamp and Z-score standardized to obtain the preprocessed time series dataset.

[0084] Step 2: Missing data identification and labeling. Based on the principle of temporal continuity, missing data detection is performed on the cable temperature data in the time series dataset to determine the missing time period, missing duration and missing location, and a missing labeling matrix is ​​generated.

[0085] Step 3: Feature sample construction. An adaptive sliding window technique is used (the window size is dynamically adjusted according to the duration of the missing data). Feature vectors are extracted and integrated based on the time dimension, meteorological environment, and historical temperature, centered on the missing time period. A sample set to be imputed is constructed for data segments containing missing locations, and a model training sample set is constructed for complete data segments without missing locations.

[0086] The window size is dynamically adjusted based on the duration of the missing information, including:

[0087] When the missing duration is ≤12 hours, the window size is set to 24 hours;

[0088] When 12 hours < missing duration ≤ 48 hours, the window size is set to 72 hours;

[0089] When the missing duration is greater than 48 hours, the window size is set to 168 hours;

[0090] Step 4: Transformer model training. Construct a Transformer time-series prediction model based on position encoding. Using the Huber loss function as the optimization objective, train the training set and validation set divided from the model training sample set through an adaptive learning rate scheduling strategy until the root mean square error of temperature prediction on the validation set is less than a preset threshold.

[0091] Step 5: Intelligent imputation of missing data. Input the sample set to be imputed into the trained Transformer model. Utilize the temperature periodicity captured by the model and the coupling relationship between meteorology and temperature to output the cable temperature imputation value for the missing period, thereby achieving data repair.

[0092] Step Six: Long-term temperature prediction. Using the complete cable temperature time series dataset obtained after interpolation as the historical input sequence, the Transformer model is used to set the prediction step size to 720 hours to 1440 hours, and the prediction results of the future long-term temperature change trend of the cable structure are output.

[0093] This invention employs the Transformer model as the core prediction and interpolation architecture. The core of this model lies in its self-attention mechanism, which can process the entire time-series input in parallel and effectively capture long-distance temporal dependencies (such as diurnal periodicity and monthly seasonality). Specifically, this mechanism can accurately extract the diurnal periodicity and seasonal trend features inherent in the data, particularly relevant to the temperature variation characteristics of Lasso.

[0094] The processing flow provided in this embodiment is as follows:

[0095] (1) Input construction, self-supervised mask and feature mapping (corresponding to step three)

[0096] Based on the adaptive sliding window technique, let the current processing window length be L (based on the duration of missing data). During the training phase, complete data segments without missing data positions are selected. Multidimensional meteorological features (atmospheric temperature, humidity, wind speed, wind direction, total radiation) and historical temperatures of Lasso constitute an input matrix. ,in The original feature dimension (e.g., 5 dimensions).

[0097] This method introduces a self-supervised masking mechanism, as detailed below:

[0098] To simulate missing data, a binary mask matrix with the same dimensions as X is generated. M t,d =1 indicates that the data at that moment is preserved (visible), M t,d =0 indicates that the data at that moment was artificially masked (set to 0 or replaced with a special token), resulting in the masked input matrix. :

[0099] (1)

[0100] in, This represents the Hadamard product (element-by-element multiplication).

[0101] Subsequently, the masked original features are projected onto the hidden layer dimension of the Transformer through a linear mapping layer. :

[0102] (2)

[0103] in, For learnable weight matrix, For bias terms, , This is the final output embedding feature matrix.

[0104] (2) Location coding introduction (corresponding to step four)

[0105] Since self-attention mechanisms themselves do not contain positional information of the sequence, positional encoding (PE) must be added to maintain temporal continuity. This method uses sine and cosine functions for absolute positional encoding:

[0106] (3)

[0107] (4)

[0108] Where pos is the position of the time step within the window ( ), where i is the dimension index.

[0109] in, It is the basic model dimension of Transformer. It is a sinusoidal positional encoding with an even-numbered dimension. It is a cosine positional encoding for odd-dimensional dimensions.

[0110] Adding the positional encoding to the input mapping yields the hidden state of the final input Transformer:

[0111] (5)

[0112] in, This is the initial hidden state matrix that will ultimately be input to the Transformer encoder.

[0113] (3) Multi-head self-attention calculation (extracting spatiotemporal coupling relationship, corresponding to step five)

[0114] For the l-th layer Transformer, the input is H (l-1) First, a linear transformation is used to obtain the query, key, and value matrix:

[0115] (6)

[0116] in, , , , These are the learnable linear projection weight matrices corresponding to the query, key, and value, respectively. It is the basic model dimension of Transformer; It is the key vector dimension; Q, K, and V refer to the query matrix, key matrix, and value matrix, respectively; H (l-1) Let be the input hidden state matrix of the l-th layer Transformer encoder.

[0117] Calculate the attention score (Scaled Dot-Product Attention):

[0118] (7)

[0119] Physical meaning: QK T The feature similarity between any two time steps within the window was calculated. Divided by... This is to prevent the gradient from entering the saturation region of softmax. Softmax normalizes the correlation into weights. Finally, multiplying by V means that the feature at the current time is a weighted sum of the features at other times within the window, thus perfectly capturing the diurnal and seasonal periodicity of temperature.

[0120] A multi-head mechanism is used to concatenate multiple attention mechanisms:

[0121] (8)

[0122] in, This is the projection weight matrix for multi-head self-attention output.

[0123] (4) Residual connection, normalization and feedforward network

[0124] Perform residual connection and normalization on the attention output:

[0125] (9)

[0126] in, This is the output feature matrix of the l-th Transformer encoder after multi-head self-attention layer + residual connection + layer normalization (Layer Norm).

[0127] Next, a two-layer fully connected feedforward network is used (introducing ReLU or GELU activation functions to extract nonlinear features):

[0128] (10)

[0129] in, This is the learnable weight matrix for the first fully connected layer; For the learnable bias term of the first fully connected layer; This is the learnable weight matrix for the second fully connected layer; This is the learnable bias term for the second fully connected layer.

[0130] Finally, perform residual connection and normalization again to obtain the output of this layer:

[0131] (11)

[0132] in, Let be the input hidden state matrix of the (l+1)th layer Transformer encoder.

[0133] (5) Self-supervised reconstruction loss function and adaptive training (corresponding to step four)

[0134] After feature extraction via multiple Transformer layers, the hidden state is mapped to the reconstructed prediction matrix through a linear output layer. In the self-supervised learning paradigm, the loss function is only calculated at the artificially masked locations to test the model's ability to imputate data "out of thin air." To address the interference of meteorological or sensor outliers on model training, this method combines the mask matrix M with the Huber loss function as the optimization objective.

[0135] (12)

[0136] in, The self-supervised total loss function; It is a mask matrix; These are actual observations; Reconstruct the predicted values ​​for the model; t represents all time steps; d represents all features.

[0137] The Huber function is specifically defined as follows:

[0138] (13)

[0139] in, The actual observed values ​​of the monitoring data; These are the reconstructed predicted values ​​of the model; This is the threshold hyperparameter.

[0140] Physical meaning: In the formula This ensures that the model only calculates errors for "pseudo-missing values" that originally existed but were manually removed. When the prediction error is small (less than a threshold)... The model employs squared error; when the prediction error is large, absolute error is used to effectively prevent gradient explosion caused by outliers. Combined with the AdamW optimizer and the Cosine Annealing adaptive learning rate scheduling strategy, the model parameters are continuously optimized.

[0141] (6) Imputation of missing data and long-term prediction (corresponding to steps five and six)

[0142] a. Imputation of truly missing data (Inference Phase):

[0143] The input at this point is the actual sample set X to be imputed, which contains missing values. miss Simply input it into the trained model, and it will output a prediction matrix. Extract the predicted values ​​corresponding to the missing locations. The data is then filled into the original data, thus completing the data repair and obtaining a complete time-series dataset.

[0144] b. For long-term forecasts (taking 720–1440 hours as an example):

[0145] The complete dataset after interpolation is used as input for the historical sequence. To predict future long-term trends, the model's prediction strategy is configured as either Direct Multi-step Forecasting or an Auto-regressive mechanism, continuously generating future LAC temperature trends based on historical sequence features. (Where H is the prediction step size, such as 720 or 1440).

[0146] Compared with the prior art, the method provided by the present invention has the following beneficial effects:

[0147] (1) Breakthrough solution to the problem of missing data: This invention adopts a "data-driven" mode and innovatively introduces multi-dimensional meteorological features as contextual information. Combined with adaptive sliding window technology, even when there is a long period of missing temperature data, it can still achieve high-precision intelligent interpolation by utilizing the strong coupling relationship between meteorology and temperature.

[0148] (2) Significantly improve the robustness and accuracy of the model: The Transformer model with strong time series modeling capabilities is introduced, and the Huber loss function is used to replace the traditional mean square error, which effectively overcomes the interference of extreme weather or sensor outliers on model training; by dividing the training set and the interpolation set, the model interpolation logic is more rigorous and the generalization ability is stronger.

[0149] (3) Achieving a leap from “historical repair” to “future early warning”: Based on the completion of high-quality data interpolation, this invention seamlessly connects to long-cycle prediction tasks and can accurately output the temperature evolution trend of the cable in the next 1 to 2 months (720 to 1440 hours), providing reliable data support for the health assessment, fatigue life prediction and active safety early warning of cable-stayed bridge cable structures.

[0150] The technical solution of the present invention will be described in detail below using a long-span cable-stayed bridge structure as an example, in conjunction with the accompanying drawings.

[0151] 1) The structural characteristics of cable-stayed bridges are as follows:

[0152] The Quanzhou Bay Bridge is 20.287 km long, located at 118°42′13″E, 24°49′13″N, with approximately 8.5 km spanning the sea. The main bridge is 800 m long with a span arrangement of (70+130+400+130+70) m. It is a double-tower, double-cable-stayed steel-concrete composite beam (concrete bridge deck + trough steel box girder) cable-stayed bridge with a semi-floating system. Figure 3 As shown.

[0153] The bridge's main girder employs a steel-concrete composite structure, while the bridge deck utilizes a concrete structure. The main body consists of channel-shaped steel beams, with the bridge deck and channel-shaped steel beams combining a single-box, three-cell box girder cross-section, a design that balances aesthetics and practicality. The standard cross-section of the main girder measures 4.25m in height and 17m in width (excluding wind deflectors). The concrete bridge deck is 30cm thick, the steel beam bottom plate thickness ranges from 14mm to 32mm, the web plate thickness is 24mm, the steel plate thickness in the cable anchorage zone is reinforced to 40mm, while the steel plate thickness in other areas is 30mm.

[0154] 2) The layout of measuring points and real-time signal acquisition are as follows:

[0155] The temperature sensor used in this experiment was a PT100, with a sampling interval of 0.5 hours, collecting 48 data points per day (2 data points per hour over 24 hours). The experiment employed a ZY-108 data acquisition instrument, which connected to the server via a ZY-34G communication module through a 4G network, enabling remote data acquisition and control for monitoring and recording of the cable temperature.

[0156] Temperature sensors are distributed within the test section, divided into two groups. One group has one measuring point, with the temperature sensor located at the center of the section, 56.5 mm from the outer edge of the cable. The other group has four measuring points, with the temperature sensors located on the outer layer of the section, 28.5 mm from the outer edge of the cable, and the outer layer is covered by a 9 mm thick HDPE protective layer. The locations of the temperature sensors on the cable section are shown in the diagram. Figure 4 , 8 As shown.

[0157] 3) Data updates and calculations / predictions are as follows:

[0158] Download the current time and temperature data of all cable measuring points collected and stored on the cloud platform, as well as multivariate meteorological data, and analyze and output the prediction analysis results based on the method described above.

[0159] The application results of the intelligent interpolation and long-term prediction method for missing cable temperature data based on multidimensional meteorological characteristics proposed in this invention in this example are analyzed as follows.

[0160] Figure 6 The figure shows a comparison curve between the predicted temperature of the cable-stayed bridge structure and the actual monitored temperature on the validation set based on the method of this invention. As can be seen from the figure, the predicted temperature (orange dashed line) output by the method of this invention is highly consistent with the fluctuation trend and periodicity of the actual monitored temperature (blue solid line). The prediction deviation is extremely small throughout the entire time period, and the overall root mean square error (RMSE) of the validation set is only 0.625℃. This fully verifies the high accuracy and high reliability of the method of this invention in the task of predicting cable temperature, and can effectively meet the engineering application requirements of health monitoring of cable-stayed bridge cable structures.

[0161] Figure 7 This figure compares the results of temperature interpolation and prediction for cable-stayed bridge structures using the method of this invention on a prediction set containing missing values. As shown in the figure, for the period of missing cable temperature data from November 2nd to November 4th (marked by the purple dashed box), the method of this invention, based on multidimensional meteorological features and the Transformer time-series model, accurately reconstructs the actual temperature variation pattern of the missing cable segments, with the interpolated values ​​closely matching the actual monitored values. Furthermore, in the subsequent complete period, the predicted temperature (orange dashed line) output by the method of this invention completely matches the fluctuation trend and periodic pattern of the actual temperature (blue solid line), with an overall root mean square error (RMSE) of only 0.631℃ for the prediction set. These results fully verify the high accuracy and high reliability of the method of this invention in tasks involving missing cable temperature data interpolation and long-term prediction. It can effectively solve the pain point of missing monitoring data in engineering projects, providing complete and accurate data support for the health monitoring of cable-stayed bridge cable structures.

[0162] The above results demonstrate that the intelligent interpolation and long-term prediction method for cable temperature based on multi-dimensional meteorological characteristics proposed in this invention effectively solves the serious data loss problem caused by sensor failure and data transmission interruption in cable-stayed bridge cable temperature monitoring, while achieving accurate long-term prediction of cable structure temperature. This method does not rely on complex structural mechanics modeling; it can achieve intelligent repair of missing data and long-term temperature prediction solely through monitoring data. It possesses advantages such as high interpolation efficiency, stable prediction accuracy, strong generalization, and wide engineering applicability. It can completely reconstruct the full-cycle variation characteristics of cable temperature, providing accurate and complete data support for cable-stayed bridge cable structure health status assessment, safety early warning, and operation and maintenance decisions. It has significant engineering application value and promotional significance for ensuring the long-term safe operation of large-span cable-stayed bridges.

[0163] This invention is not limited to the above embodiments. Those skilled in the art will understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of this invention. The scope of this invention is defined by the claims and their equivalents.

Claims

1. A method for interpolating missing cable temperature data based on multidimensional meteorological characteristics, characterized in that, Includes the following steps: Step 1: Data Acquisition and Preprocessing Acquire cable temperature data and multidimensional meteorological data, wherein the multidimensional meteorological data includes environmental parameters that affect cable temperature changes; The raw data is cleaned to construct a preprocessed time-series dataset. Step 2: Identification and Marking of Missing Data Missing data detection is performed on the preprocessed time series dataset to generate a missing data marker matrix; Step 3: Feature Sample Construction The sliding window technique is used to extract features from the time series dataset and construct a fused feature vector; Based on data integrity, the extracted samples are divided into a complete sample set for model training and a sample set to be imputed for missing data imputation. Step 4: Transformer Model Training Construct a Transformer time series prediction model; The model is iteratively trained using the training sample set to minimize the prediction error until the model meets the preset convergence condition. Step 5: Imputing missing data: Input the sample set to be interpolated into the trained Transformer time series prediction model, output the temperature interpolation values ​​for the missing time periods, and combine them with the original data to obtain the complete Lasso temperature time series dataset.

2. The method according to claim 1, characterized in that, In step one, the environmental parameters include atmospheric temperature, atmospheric humidity, wind speed, wind direction, and radiation.

3. The method according to claim 1, characterized in that, In step one, the preprocessing includes: The Z-score normalization method was used to process the multidimensional meteorological feature data and the Lasso temperature data to eliminate dimensional differences.

4. The method according to claim 1, characterized in that, In step three, the fused feature vector includes time dimension features, meteorological environment features, and historical temperature features.

5. The method according to claim 1, characterized in that, In step three, the window size of the sliding window technology is dynamically adjusted according to the missing duration, wherein the window size is positively correlated with the missing duration, and the window size is greater than the missing duration.

6. The method according to claim 1, characterized in that, In step three, the sample partitioning includes: for data segments containing missing positions, extracting a sample set to be imputed centered on the missing time period; for complete data segments without missing positions, extracting a sample set to be used for model training.

7. The method according to claim 1, characterized in that, In step four, the step of using the model training sample set to iteratively train the model with the goal of minimizing the prediction error until the model meets the preset convergence condition includes: The model training sample set is divided into a training set and a validation set; The model is iteratively trained using an adaptive learning rate scheduling strategy. The iterative training uses the Huber loss function as the optimization objective and takes the prediction error on the validation set being less than a preset threshold as the convergence condition.

8. A method for long-term prediction of cable temperature based on multidimensional meteorological characteristics, characterized in that: The method includes the steps of any one of claims 1 to 7; it also includes the following steps: Step Six: Long-Term Temperature Prediction The complete time-series dataset of cable temperature is used as the historical input sequence and input into the trained Transformer time-series prediction model. The output is the predicted result of the temperature change trend of the cable structure in the future target period.

9. The method according to claim 8, characterized in that, The target period for the future is between 720 and 1440 hours.

10. A cable temperature monitoring and data processing system, characterized in that, include: The data acquisition layer includes a cable temperature sensor deployed on the cable under test, and a meteorological parameter sensor for collecting multi-dimensional meteorological parameters that affect the temperature change of the cable. The data transmission and storage layer includes a data acquisition module that samples at a preset period, as well as a data transmission module and a data storage module for transmitting the acquired data to a remote location. The data analysis layer includes a computer analysis module with embedded computer programs; The computer analysis module is configured to perform the cable temperature missing data interpolation method based on multidimensional meteorological characteristics as described in any one of claims 1 to 7, or to perform the cable temperature long-term prediction method based on multidimensional meteorological characteristics as described in claim 8 or 9.

Images