Strain key feature missing repair method and system for traffic infrastructure monitoring

Abstract

The application provides a strain key feature missing repair method and system for traffic infrastructure monitoring, belongs to the technical field of data missing repair, and comprises the following steps: acquiring strain time series for traffic infrastructure monitoring and performing feature extraction, acquiring strain data features, and constructing a structured feature data set; establishing a basic regression model based on a multilayer perception machine; introducing a local residual error estimation mechanism based on K neighbors to perform local deviation correction and a hyperparameter optimization strategy to find optimal hyperparameters; acquiring strain time series for traffic infrastructure monitoring to be repaired, inputting the strain time series into the basic regression model established based on the multilayer perception machine, and outputting repaired strain key features.

Description

Technical Field

[0001] This invention belongs to the field of data missing repair technology, and in particular relates to a method and system for repairing missing key strain features for traffic infrastructure monitoring. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] The long-term performance scientific observation network for transportation infrastructure provides crucial data support for infrastructure lifecycle performance assessment and maintenance decisions by continuously collecting multi-source monitoring data such as temperature, humidity, load, stress, and strain. Among these, strain, as a core monitoring indicator that directly reflects the local deformation and stress state of a structure, is fundamental data for stress analysis, fatigue life assessment, and structural safety evaluation. Its completeness and accuracy are of irreplaceable importance.

[0004] However, under long-term service environments and complex operating conditions, strain monitoring systems generally face problems such as sensor aging, physical damage, and environmental interference, leading to the loss or distortion of key physical characteristics in the monitoring data, such as peak-to-valley differences. This loss of characteristics is not a simple random occurrence, but often happens at critical moments when the load is significant or the structure's response is intense, directly obscuring the true stress and deformation patterns of the structure, thus causing systematic biases in fatigue life assessment and safety status determination. Several infrastructure safety accidents in recent years have also demonstrated that the long-term loss of key characteristics in monitoring data significantly weakens the structural health monitoring system's ability to identify and warn of potential risks.

[0005] The inventors discovered in their research that traditional interpolation or statistical repair methods are difficult to characterize the highly nonlinear mapping relationship between external environmental excitation and structural strain response. They can often only restore the overall trend of the data, making it difficult to guarantee the physical consistency and engineering credibility of the repair results. Summary of the Invention

[0006] To overcome the shortcomings of the existing technologies, this invention provides a method and system for repairing missing key strain features in traffic infrastructure monitoring. It introduces data-driven technology with nonlinear modeling capabilities to deeply mine the inherent coupling mechanism between environmental factors, traffic load and strain response from multi-source monitoring information, thereby achieving high-precision repair of missing key features.

[0007] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:

[0008] Firstly, a method for repairing missing key strain characteristics in traffic infrastructure monitoring is disclosed, including:

[0009] Obtain strain time series data for traffic infrastructure monitoring and extract features to obtain strain data features and construct a structured feature dataset;

[0010] A basic regression model is established based on a multilayer perceptron: a local residual estimation mechanism based on K nearest neighbors is introduced to perform local bias correction and hyperparameter optimization strategy to find the optimal hyperparameters;

[0011] Then, the strain time series of the traffic infrastructure to be repaired is obtained, input into the basic regression model based on multilayer perceptron, and the key strain features after repair are output.

[0012] As a further technical solution, the acquisition of strain data characteristics specifically includes:

[0013] The acquired strain time series data for traffic infrastructure monitoring are grouped into master sensor sequences and slave sensor sequences.

[0014] Identify significant peaks in the main sensor sequence and their corresponding timestamps, and establish them as the global synchronization benchmark for this loading loop;

[0015] The time difference between the first significant peaks in the master and slave channels is calculated as the global phase offset. This offset is then used to construct a dynamic search window centered on the reference timestamp in each slave sensor channel. This window adaptively searches for local peaks that physically correspond to the master peak. After locking the peak, the window preceding the peak is used to search for the nearest significant valley in reverse order, thus constructing a pair of "peak-valley" feature pairs.

[0016] As a further technical solution, it also includes: using the minimum peak spacing and relative peak value as constraints, and only when the amplitude difference between a peak and its adjacent trough exceeds a set multiple of the signal standard deviation, it is considered a valid feature point.

[0017] As a further technical solution, adaptive anomaly cleaning is also included for the feature sequences in the structured feature dataset:

[0018] Set a sliding time window with a set length;

[0019] For the data sequence within the window, the lower quartile and upper quartile of the data within the window are calculated in real time, and a dynamic interquartile range is defined. Based on this, an adaptive anomaly detection boundary that changes dynamically over time is constructed.

[0020] Data points falling outside this dynamic boundary are all considered outliers;

[0021] For identified outliers, local linear interpolation or median of normal data within the window is used for replacement.

[0022] As a further technical solution, the multilayer perceptron network structure in the basic regression model based on the multilayer perceptron includes:

[0023] Input layer: Receives preprocessed feature vectors;

[0024] The first hidden layer maps low-dimensional input features to a high-dimensional space to increase the linear separability of features and capture the initial interaction relationships between input variables.

[0025] The second hidden layer: compresses and refines high-dimensional features, removes redundant information, and extracts the most critical abstract expression for predicting strain features.

[0026] Output layer: Contains 1 neuron, which directly outputs the predicted strain feature value.

[0027] As a further technical solution, the establishment of a local residual estimation mechanism includes:

[0028] Define the observation domain and missing domain of the monitoring dataset;

[0029] Based on the multilayer perceptron trained above, a basic regression model is established, and back-substituted prediction is performed for all time i within the observation domain to calculate the deviation between the actual observed value and the basic predicted value of the model.

[0030] If the residuals exhibit a random distribution of alternating positive and negative values ​​within a local time window, it is considered measurement noise; however, if the residuals exhibit a continuous positive or negative bias trend locally, it indicates the presence of local systematic errors.

[0031] For any missing point to be repaired, assuming that its potential prediction error has a high degree of similarity and transitivity to the residuals of known points in its neighborhood, a dynamic search window is constructed to find the k nearest valid observation points in the observation domain, forming a neighborhood set:

[0032] Based on the known residual information within the neighborhood set, an inverse distance weighting algorithm is used for the target point. The weight of each reference point in the neighborhood is inversely proportional to its distance. Thus, the estimated residual of the target point can be expressed as a weighted average of the neighborhood residuals. A correction intensity coefficient is introduced to adaptively adjust the correction magnitude.

[0033] The estimated residual compensation is superimposed on the basic predicted value of the multilayer perceptron-based regression model to obtain the final reconstructed value after local correction.

[0034] Secondly, a system for repairing missing key strain characteristics for traffic infrastructure monitoring was disclosed, including:

[0035] The structured feature dataset construction module is configured to: acquire strain time series data for traffic infrastructure monitoring and extract features, obtain strain data features, and construct a structured feature dataset.

[0036] The basic regression model construction module is configured to: build a basic regression model based on a multilayer perceptron; introduce a local residual estimation mechanism based on K nearest neighbors for local bias correction and a hyperparameter optimization strategy to find the optimal hyperparameters;

[0037] The repair module is configured to: acquire the strain time series of the traffic infrastructure to be repaired, input it into the basic regression model based on the multilayer perceptron, and output the key features of the strain after repair.

[0038] As a further technical solution, when acquiring strain data features in the structured feature dataset construction module, the specific steps include:

[0039] The acquired strain time series data for traffic infrastructure monitoring are grouped into master sensor sequences and slave sensor sequences.

[0040] Identify significant peaks in the main sensor sequence and their corresponding timestamps, and establish them as the global synchronization benchmark for this loading loop;

[0041] The time difference between the first significant peaks in the master and slave channels is calculated as the global phase offset. This offset is then used to construct a dynamic search window centered on the reference timestamp in each slave sensor channel. This window adaptively searches for local peaks that physically correspond to the master peak. After locking the peak, the window preceding the peak is used to search for the nearest significant valley in reverse order, thus constructing a pair of "peak-valley" feature pairs.

[0042] The above one or more technical solutions have the following beneficial effects:

[0043] To address the issues of data redundancy and ambiguous load-response mapping when directly using raw strain time series, this invention provides a sub-technical solution to acquire strain time series for monitoring transportation infrastructure and extract features to construct a structured feature dataset. This dataset can extract key information reflecting the structural limit state and fatigue damage from the raw data, strip away redundant intermediate processes, and establish a clearer load-response relationship, thereby providing high-quality feature input for subsequent analysis.

[0044] The sub-technical solution of this invention utilizes the powerful function approximation capability of multilayer perceptrons to establish a global regression model. To address potential local fitting biases in the global model, a residual correction mechanism based on neighborhood information is introduced for refined compensation, achieving high-fidelity recovery of key features. This provides a solid and reliable data foundation for subsequent fatigue damage analysis and structural safety status assessment.

[0045] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0046] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0047] Figure 1 This is a flowchart of the master-slave synchronization mechanism according to an embodiment of the present invention;

[0048] Figure 2 This is a graph showing the peak and valley detection results;

[0049] Figure 3 A heatmap showing the correlation of feature data;

[0050] Figure 4 A global result diagram of anomaly detection for strain characteristics;

[0051] Figure 5 This is a local result diagram of anomaly detection for strain characteristics;

[0052] Figure 6 Here is a diagram of the MLP network structure;

[0053] Figure 7 The comparison diagrams of the repair results before and after correction under the continuous missing mode are shown. (a) is a schematic diagram of the overall repair effect of the first 3000 data, and (b) is a schematic diagram of the corresponding local repair results.

[0054] Figure 8 The comparison diagrams of the repair results before and after correction under the discontinuous missing mode are shown. (a) is a schematic diagram of the overall repair effect of the first 3000 data, and (b) is a schematic diagram of the corresponding local repair results.

[0055] Figure 9 The figures show the model repair results under different missing rates in the discontinuous missing pattern. (a), (b), (c), (d), and (e) respectively show the repair effect of the model under missing rates of 10%, 30%, 50%, 70%, and 90% in the discontinuous missing pattern.

[0056] Figure 10The figures show the model repair results under various missing rates in the continuous missing pattern. (a), (b), (c), (d), and (e) respectively demonstrate the repair effect of the model under continuous missing rates of 10%, 30%, 50%, 70%, and 90%. Detailed Implementation

[0057] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0058] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0059] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0060] Example 1

[0061] This embodiment discloses a method for repairing missing key strain features in traffic infrastructure monitoring, including:

[0062] Step 1: By extracting and cleaning physically meaningful features from the raw data, a high-quality model input-output pair is constructed. The raw data for this step is the collected raw time series data.

[0063] Step 2: Utilize the powerful function approximation capability of the multilayer perceptron to establish a global regression model;

[0064] Step 3: To address potential local fitting biases in the global model, a residual correction mechanism based on neighborhood information is introduced to refine the output of the global regression model.

[0065] In one implementation example, specifically in step one, regarding the extraction of strain data features, directly using the original strain time series in structural health monitoring suffers from data redundancy and ambiguity in the load-response mapping relationship. Therefore, this embodiment's sub-technical solution extracts strain features with clear physical meaning, such as peak / valley differences, as an effective preprocessing strategy. This method can extract key information reflecting the structural limit state and fatigue damage—namely, the peak / valley difference—from the original data. This peak / valley difference can reflect this structural limit state and fatigue damage, stripping away redundant intermediate processes and establishing a clearer response relationship between load and strain features, thereby providing high-quality feature input for subsequent analysis.

[0066] In practice, a master-slave synchronization-based feature extraction strategy is adopted. In response to the micro-hour sequence deviation problem that is common in multi-channel strain monitoring data, the traditional single-channel independent feature extraction method is prone to cause the physical events between different measuring points to fail to accurately align with the responses under the same load, which in turn affects the training accuracy of the subsequent load and strain feature mapping model.

[0067] To address this, the sub-solution of this embodiment proposes an adaptive feature extraction strategy based on a master-slave synchronization mechanism. By locking the peak moment of the master sensor to establish a global time reference, it achieves accurate matching and difference calculation of multi-channel strain peak and valley features. The specific process is as follows: Figure 1 As shown.

[0068] In multi-channel monitoring systems, due to signal transmission delay, differences in sensor response speed, and structural damping effects, the peak responses of sensors at different locations to the same external excitation often exhibit a slight phase difference on the time axis. To address this issue, firstly, based on signal-to-noise ratio and load correlation analysis, the sensor channel with the best data quality and highest sensitivity to load response is selected as the master sensor, while the remaining sensors with strong correlation are defined as slave sensors. Using an improved peak detection algorithm, the master sensor feature pairs and their corresponding slave sensor feature pairs are identified, and significant peaks in the master sensor sequence and their corresponding timestamps, which also serve as reference timestamps, are determined. This is established as the global synchronization benchmark for this loading loop.

[0069] Based on this, the time difference between the first significant peaks of the master and slave channels is first identified as the global phase offset. That is, identifying the first peak of the main sensor and the first peak of the slave sensor; the time difference between these two peaks is... This offset is then combined with a reference timestamp built from each sensor channel. Centered dynamic search window The adaptive search corresponds to the local peaks that physically correspond to the significant peaks in the main sensor sequence. The specific method is as follows: after obtaining the peak of the main sensor, within the time range corresponding to this peak... In the process, the peak corresponding to the sensor is found, and then the trough is found based on the peak. After locking the peak, the nearest significant trough is searched backward within the time window preceding the peak. This constructs peak-valley feature pairs, which are the extracted strain data features and subsequently serve as model input-output pairs.

[0070] Subsequently, for the constructed paired peak-valley feature pairs, minimum peak spacing and relative peak value were introduced as constraints to eliminate spurious peak interference caused by high-frequency noise. Specifically, a peak is considered a valid feature point only when the amplitude difference between it and its adjacent valley exceeds α times the signal standard deviation; in this example, α = 0.3.

[0071] Based on the matching results, calculate the value of each sensor channel j at the 1st... i Strain peak-valley difference under each fluctuation segment The peak-valley feature is a pair of peaks and valleys, calculated using the following formula:

[0072] (1)

[0073] in, and Indicates that each sensor channel j is at the 1st... i Peak and trough values ​​under each fluctuation segment. The peak-trough difference features are extracted, and then the relationship between them and the load is established in the model to output the final result.

[0074] Figure 2 The detection results are for the main channel and a peak-valley segment randomly selected from one of the channels. Figure 2 The proposed master-slave synchronization feature extraction strategy can effectively identify the peaks and valleys of strain data in the master and slave channels.

[0075] By traversing all loading events, i.e., the original data, a structured feature dataset is ultimately constructed, including all peak-valley features. This data structure based on physical event alignment effectively eliminates temporal redundancy caused by asynchronous sampling, transforming high-dimensional time-series data into a compact feature vector matrix, providing a high-quality data foundation for subsequent model training.

[0076] Subsequently, in order to select the most representative target feature data from the multi-channel strain data for interpolation research, the sub-method of this embodiment performed Pearson correlation analysis on all calculated peak-valley features and plotted a correlation heatmap of the multi-feature data, such as... Figure 3 As shown in the figure, the analysis results show that strain 3 exhibits a significant and strong correlation with other feature data, and its average correlation coefficient is significantly better than that of strain 1 and strain 2. This statistical characteristic indicates that the feature data of strain 3 contains the richest structural response information. Therefore, this study ultimately selects the feature data of strain 3 as the core interpolation and reconstruction object to verify the effectiveness of the proposed algorithm. Stress feature data and other feature data show a low correlation; therefore, stress data is not included in the calculation of the sub-technical solution in this embodiment.

[0077] Regarding anomaly detection in feature data: In the actual engineering monitoring feature extraction process, although a master-slave synchronization strategy is adopted, due to transient noise interference or local signal distortion, some significant outliers are still inevitably mixed in among all the extracted peak-valley features, such as abnormal values ​​caused by mismatched peak-valley matching. If these abnormal feature values ​​are not removed, they will seriously interfere with the training convergence and generalization ability of the subsequent regression model as dirty data.

[0078] Given that strain feature sequences also exhibit non-stationary time-varying characteristics, traditional fixed threshold methods are ill-suited to their dynamic fluctuations. Therefore, this embodiment introduces a rolling interquartile range (IQR) algorithm to adaptively clean up anomalies in all peak-valley features.

[0079] This algorithm captures local statistical characteristics of the data by setting a sliding time window of length L=30. For the data sequence within the window, the lower quartiles of the data within the window are calculated in real time. and the upper quartile And define dynamic interquartile range Based on this, an adaptive anomaly detection boundary that dynamically changes over time is constructed: .

[0080] in, The sensitivity coefficient is set to 1.2 in this embodiment. Any data point falling outside this dynamic boundary is considered an outlier. For identified outliers, local linear interpolation or the median of the normal data within the window is used for replacement. This effectively eliminates non-physical impulse interference while preserving the true fluctuation trend of the signal, ensuring the quality of subsequent modeling data.

[0081] Figure 4 and Figure 5 This demonstrates the global and local magnification effects of the rolling window interquartile range algorithm for anomaly detection in strain feature sequences. From Figure 4 As seen in the global view, for strain characteristic sequences with drastic data changes, the dynamic statistical boundaries generated by the scrolling window can determine reasonable data ranges within different local windows, effectively avoiding interference from cross-scale statistics, and thus accurately marking outliers caused by extraction errors within each window. Further observation... Figure 5 The magnified view shows that this method can sensitively determine the normal fluctuation range of the data within the detection window, thus accurately removing individual outliers without mistakenly deleting normal load response extrema. Results indicate that this strategy effectively cleanses the feature dataset, providing high-quality data for subsequent model training.

[0082] It is worth noting that, given the superiority and adaptability of the rolling window IQR algorithm in capturing local abnormal fluctuations, this method is not only applied to the data cleaning stage after feature extraction to remove erroneous feature samples, but also constitutes a key link in the quality control of subsequent model prediction results.

[0083] In step two, the key to data restoration using the MLP data restoration method based on adaptive residual correction is establishing a high-precision mapping model from environmental loads to structural response characteristics after feature extraction of structural health monitoring data. Although Multilayer Perceptrons (MLPs) possess powerful global nonlinear fitting capabilities, single models often exhibit fitting residuals when facing localized drastic fluctuations or complex nonlinear boundaries. Therefore, this patented technical solution proposes an MLP data restoration method based on adaptive residual correction. This method focuses on establishing a basic regression model based on MLP, innovatively introducing a local residual estimation mechanism based on K-Nearest Neighbors (KNN) for local bias correction and an Optuna algorithm hyperparameter optimization strategy to find the optimal hyperparameters. These methods further compensate for the prediction bias of the original MLP model, thereby achieving high-precision feature restoration.

[0084] Construction and Implementation of the Basic MLP Regression Model: As the most fundamental and core feedforward neural network architecture in deep learning, the Multilayer Perceptron (MLP) plays a crucial role in solving complex regression problems due to its powerful ability to approximate nonlinear functions. In the strain feature repair task of this example, it is necessary to establish a mapping relationship from multidimensional environmental variables such as temperature and load characteristics to single structural response characteristics such as strain peak-valley differences. This is essentially a nonlinear regression problem from high-dimensional input to low-dimensional output. MLP, through hierarchical neuron connections and nonlinear activation, can effectively capture this complex physical mapping law.

[0085] Regarding the topology of neural networks: A typical MLP network consists of an input layer, several hidden layers, and an output layer. Signal transmission between layers is achieved through fully connected connections, meaning that every neuron in one layer is connected to all neurons in the next layer. The MLP network structure design in this patented implementation is as follows: Figure 6 As shown.

[0086] Input layer: Responsible for receiving preprocessed feature vectors xi represents the peak-to-valley difference. In the implementation example, the input dimension n corresponds to the number of selected, highly relevant auxiliary variables, such as sensor data from different locations, ambient temperature, etc.

[0087] Hidden layers: These are the core layers for extracting deep features from the data. To balance the model's fitting ability and computational complexity, this example design uses a network structure with two hidden layers:

[0088] The first hidden layer has 64 neurons. Its purpose is to map low-dimensional input features to a high-dimensional space, increasing the linear separability of features and capturing the initial interactions between input variables.

[0089] The second hidden layer consists of 32 neurons. This layer is responsible for compressing and refining high-dimensional features, removing redundant information, and extracting the most critical abstract representations for strain feature prediction.

[0090] Output layer: Contains one neuron, directly outputting the predicted strain feature value. .

[0091] Its mathematical expression can be formalized as a series of matrix operations. Assume the... l The output vector of the layer is Then the first l The output of layer +1 can be represented as:

[0092] (2)

[0093] (3)

[0094] in, No. l +1 layer of inactive linear combination input, specifically For the first l layer to the first l +1 weight matrix, For bias vectors, It is a non-linear activation function. For the first l The output vector of the activation of layer +1. For the two hidden layer network designed in this study, This is the initial output of the model; the final result is the result after further correction.

[0095] (4)

[0096]

[0097] in, For the input feature vector, This is the weight matrix of the first hidden layer. This is the weight matrix of the second hidden layer. The weight matrix of the output layer. This is the bias vector for the first hidden layer. This is the bias vector for the second hidden layer. This is the bias vector for the output layer. It is a non-linear activation function.

[0098] Regarding the selection and characteristic analysis of activation functions: In neural networks, activation functions introduce nonlinear factors and are key to enabling the model to approximate any nonlinear function. Traditional Sigmoid or Tanh functions are prone to the vanishing gradient problem in deep networks, meaning that when the input value is large or small, the function derivative approaches 0, causing the gradient to fail to propagate effectively during backpropagation, making the network difficult to train.

[0099] To address this issue, this study uses ReLU (Rectified Linear Unit) as the activation function in all hidden layers. The mathematical definition of ReLU is:

[0100] (5)

[0101] The main basis for ReLU includes:

[0102] (1) Solving gradient vanishing: When the input x>0, the derivative of ReLU is always 1, which ensures that the gradient can propagate backwards without loss in the positive interval, which significantly accelerates the convergence speed of the model.

[0103] (2) Sparse activation: When the input x < 0, the neuron outputs 0. This means that at any given time, only some neurons in the network are in an activated state. This sparsity not only reduces the interdependence of parameters, but also effectively reduces the risk of model overfitting and enhances the generalization ability of the model.

[0104] (3) Computationally efficient: Compared with the exponential operation of Sigmoid / Tanh, ReLU only involves threshold judgment, with extremely low computational overhead, making it very suitable for processing massive amounts of monitoring data.

[0105] Regarding data standardization strategies: Due to the multi-source and heterogeneous nature of structural health monitoring data, the dimensions and numerical ranges of different physical quantities (such as temperature, strain, and load) vary greatly. For example, strain values ​​are typically in the range of 10... -6 Temperature data ranges from 0 to 40 degrees Celsius. If the raw data is directly input into the MLP, features with larger values ​​will dominate the direction of weight updates, making the model insensitive to features with smaller values. Furthermore, the contour lines of the loss function will be flattened ellipsoids, causing the gradient descent process to oscillate and converge slowly.

[0106] To eliminate the influence of dimensions and accelerate model convergence, this study preprocesses all input features using Z-score standardization before inputting the data into the network. The calculation formula is as follows:

[0107] (6)

[0108] Where x represents the original data, u represents the mean of the feature on the training set, σ represents the standard deviation, and x* represents the standardized input value. After processing, all input features follow a standard normal distribution with a mean of 0 and a standard deviation of 1, making the solution space of the loss function closer to a sphere, thus ensuring that the gradient descent algorithm can quickly converge to the global minimum along the optimal path.

[0109] Model Training and Optimization Algorithm: The model training process is essentially a parameter optimization problem, that is, finding a set of optimal weights W and biases b that minimizes the difference between the predicted and true values. This study uses Mean Squared Error (MSE) as the loss function to measure the model's predictive performance.

[0110] (7)

[0111] in, Represents all parameters of the model. For the true value, These are the model's predicted values. This represents the sample batch size.

[0112] To minimize the loss function, the Adam (Adaptive Moment Estimation) optimization algorithm is used. Adam is an optimization algorithm with an adaptive learning rate that combines the advantages of the momentum method and the RMSProp algorithm.

[0113] Specifically, the Adam algorithm adjusts the current update step size in each iteration based on the moving average of historical gradients. This allows the model to descent rapidly in the early stages of training, and automatically reduces the step size to avoid oscillations as it approaches the optimal solution. Compared to traditional stochastic gradient descent (SGD), the Adam algorithm is more robust when dealing with sparse gradients and non-stationary objective functions, significantly improving the training efficiency and final accuracy of the nonlinear regression task in this study.

[0114] A Residual Estimation Mechanism Based on Local Neighborhood: Although Multilayer Perceptrons (MLPs), with their powerful global nonlinear mapping capabilities, can effectively fit the overall trend of monitoring data as load and environment change, in practical engineering applications, due to complex factors such as local microenvironmental fluctuations, sensor temperature drift effects, and structural transient response hysteresis, there is often a non-random systematic deviation between model predictions and actual observations. These deviations (i.e., prediction residuals) typically exhibit significant local correlations in their temporal or spatial distribution, rather than being pure white noise. To further mine this effective information not captured by MLPs and improve the precision of data reconstruction, this embodiment proposes a residual estimation mechanism based on local neighborhood.

[0115] 1. Construction and analysis of the global residual field

[0116] First, the observation domain of the feature dataset is defined as follows: The missing field is Based on the MLP model trained above, for all objects located in the observation domain... Back-substitution prediction is performed at time i within the range to calculate the actual observed value. Compared with the model's baseline predictions Deviation between:

[0117] (8)

[0118] The residual sequence This reflects the inadequacy of the current MLP model in fitting specific time periods or operating conditions. If the residuals exhibit a random distribution of alternating positive and negative values ​​within a local time window, it is usually considered as measurement noise; however, if the residuals show a continuous positive or negative bias locally, it indicates the existence of local systematic errors, which is precisely what this mechanism aims to capture and correct.

[0119] 2. Dynamic Neighborhood Search Strategy: For any missing point to be repaired... Assuming that its potential prediction error has a high degree of similarity and transitivity to the residuals of known points in its neighborhood, a dynamic search window is constructed within the observation domain. Find the k nearest valid observation points to the target point j, and form a neighborhood set. :

[0120] (9)

[0121] in, This is a distance metric between sample points i and j. In time series interpolation scenarios, this distance is typically defined as the absolute value of the difference between time indices. In a multidimensional feature space, this can also be expanded to the Euclidean distance between feature vectors. This can be achieved by setting a distance threshold. Further eliminate reference points that are too far away or have weak correlation to ensure the effectiveness of neighborhood information.

[0122] 3. Inverse Distance Weighted Residual Correction Model: Based on Neighborhood Set Given the known residual information, the Inverse Distance Weighting (IDW) algorithm is used to assign weights to each reference point i in the neighborhood of the target point j. It is inversely proportional to its distance:

[0123] (10)

[0124] in, To prevent the use of tiny constants with a denominator of zero, the estimated residual for target point j is further... It can be expressed as a weighted average of the neighborhood residuals, with a correction intensity coefficient introduced. Adaptive adjustment of the correction magnitude:

[0125] (11)

[0126] Finally, the estimated residual compensation is superimposed on the MLP's base prediction to obtain the final reconstructed value after local correction.

[0127] (12)

[0128] This mechanism, by introducing prior information from neighborhood residuals, is equivalent to superimposing a local correction filter on top of the global regression model. When local systematic biases occur in MLP predictions, such as overall shifts due to temperature inertia, the known residuals within the neighborhood guide the predicted values ​​back to the true trend, thereby significantly improving the robustness and accuracy of the method under non-stationary conditions.

[0129] Optuna-based hyperparameter optimization strategy: In the aforementioned residual estimation and correction mechanism, the correction effect is highly sensitive to the values ​​of hyperparameters. Core parameters include the number of neighborhood samples k and the effective distance threshold. and correction strength coefficient If the value of k is too small, the estimation results are easily affected by individual noise points and fluctuate; if the value of k is too large, it may introduce weakly correlated samples from afar, diluting local features; and The value directly determines the magnitude of residual compensation; improper setting can lead to undercompensation or overcompensation. Since the data distribution and noise characteristics of different monitoring channels vary, relying on trial and error based on manual experience is not only inefficient but also makes it difficult to obtain the globally optimal solution.

[0130] To address this, the patented technical solution introduces the Optuna automated hyperparameter optimization framework, constructing an adaptive parameter optimization strategy. Optuna employs a Tree-structured Parzen Estimator (TPE) as its core sampling algorithm. By probabilistically modeling historical experimental results, it can intelligently focus the search on parameter regions more likely to improve model performance, significantly outperforming traditional grid search.

[0131] The optimization process aims to minimize the prediction error on the validation set. The objective function is defined as follows:

[0132] (13)

[0133] in, For the parameter combination to be optimized, To verify the true value of the set, Represents the residual correction function. These are the predictions of the original MLP model on the validation set. This represents the residual of the training set. In this implementation example, the following search space is constructed for each parameter: , , The optimization trial was set to 100 rounds. The algorithm automatically performed iterative sampling and evaluation within the search space, and finally output the optimal parameter combination that minimizes the MAE. This strategy enables the model to adapt to different data features, ensuring the robustness and optimal performance of the residual correction mechanism under various operating conditions.

[0134] Accuracy Evaluation and Result Analysis of Strain Feature Repair: This section analyzes the performance of the proposed MLP model based on adaptive residual correction in the strain feature repair task. Through quantitative error statistics and qualitative visualization comparison, the system evaluates the repair accuracy of the model under different data integrity levels to verify the improvement of model performance by the adaptive residual correction mechanism.

[0135] Validation of the effectiveness of the adaptive residual correction mechanism: The adaptive residual correction module aims to compensate for the basic prediction bias of the MLP using local neighborhood information. To verify the universality and effectiveness of this mechanism under different data distribution characteristics, this example selects strain data features with a 10% missing rate and compares the repair effects under discontinuous and continuous missing modes. Figure 7 and Figure 8 The repair results before and after correction are shown for two different missing modes. Figure 7 (a) and Figure 8 (a) shows the overall repair effect of the first 3000 data points. Figure 7(b) and Figure 8 (b) shows the corresponding local repair results, so as to more intuitively compare and analyze the repair performance.

[0136] Figure 7 This visually demonstrates the comparison between the basic MLP prediction and the repair results after adaptive residual correction under a 10% continuous missing data pattern. For example... Figure 7 As shown in (a), in regions with continuous data loss, the predicted curves before and after correction closely follow the distribution of the original data, indicating that the MLP model can capture the overall trend of change quite well. However, in extreme value regions where the values ​​are too large or too small, the repair effect is still insufficient. This is mainly because, under long periods of data loss, the model loses the local constraints of the neighborhood data, and it is difficult to accurately capture local abrupt changes by relying solely on global rules, resulting in underfitting at extreme points. Nevertheless, this stable grasp of the global trend avoids excessive outlier error, verifying that the model still has good robustness and core mapping ability when lacking local guidance. Figure 7 In the magnified region of (b) in the diagram, the repaired curve (green) after adaptive residual correction is significantly closer to the original data (blue) than the basic MLP prediction curve (red). Especially at local peaks and inflection points, the correction mechanism effectively corrects the bias of the basic model. This indicates that introducing neighborhood residual information can significantly improve the model's ability to capture local details, making the repaired result closer to the real physical response.

[0137] Figure 8 This demonstrates the repair effect under 10% discontinuous deletion. For example... Figure 8 As shown in (a), in stable data segments, the overall trends of both the base MLP and the corrected prediction curves closely match the actual data, reflecting the model's excellent prediction performance for missing segments with small data fluctuations. However, at occasional missing maxima or minima, both models exhibit some overestimation or underestimation of the predicted values. This indicates that at locations with drastic fluctuations, the MLP model's sensitivity to extreme features amplifies the prediction bias, making it difficult to fully reproduce the true extreme value fluctuations. Meanwhile... Figure 8 In the magnified region of (b) in the model, the result after adaptive residual correction is significantly better than the original MLP, and the repair curve is closer to the real data trajectory. Especially at the edges of continuous missing segments and at drastic fluctuations, the corrected curve shows a stronger correction capability, effectively corrects the deviation of the basic model, and significantly improves the physical consistency and continuity of the data reconstruction.

[0138] contrast Figure 7 and Figure 8It can be seen that in the discontinuous missing mode, because a lot of effective neighborhood information is preserved around the missing point, the residual correction mechanism can make full use of these local clues to make fine corrections to small prediction biases, exhibiting the characteristics of overall stability and local accuracy in repair. In the continuous missing mode, due to the loss of large segments of data, neighborhood reference information is scarce, and the model relies more on the MLP's ability to extrapolate the global trend. Although it is susceptible to overshoot at extreme points, the residual correction mechanism still plays a key role at the edges of missing segments and at fluctuation inflection points. Through smooth transitions and local corrections, it ensures the overall continuity and physical rationality of the repaired data, demonstrating the adaptability and robustness of this hybrid strategy under different missing scenarios.

[0139] Table 1. Comparison of results before and after correction of different deletion patterns at a 10% deletion rate.

[0140]

[0141] Table 1 shows a comparison of the quantization accuracy of two different missing patterns before and after residual correction at a 10% missing rate. Data analysis reveals significant differences.

[0142] For discontinuous missing data scenarios, the adaptive residual correction mechanism demonstrated superior correction capabilities: MAE decreased significantly from 7.8579 to 6.1500, Wmape decreased from 20.7672% to 16.2534%, and RMSE decreased from 13.0060 to 10.2665, representing relative improvements of 21.7%, 21.7%, and 21.1%, respectively. This indicates that in discontinuous missing data scenarios, due to the rich real-observation neighbors around the missing points, the algorithm can efficiently capture and utilize residual information within the neighborhood to powerfully correct the basic predicted values.

[0143] For the continuous missing data pattern, although the MAE slightly decreased from 2.5099 to 2.5072 after introducing the correction mechanism, the overall accuracy improvement was relatively limited. This is mainly because the loss of large, continuous data segments leads to a severe lack of effective neighborhood information. In this scenario, many of the time-series neighboring points around the point to be repaired are missing, making it difficult for the algorithm to capture the true observation residuals as a basis for correction. This prevents the residual correction module based on neighborhood similarity from fully functioning. Therefore, the final repair result in this scenario relies more on the MLP model's ability to fit the global nonlinear trend of the data, rather than on local residual compensation.

[0144] Furthermore, comparing the absolute error values ​​of the two modes reveals that the prediction error in the continuous missing data mode is significantly lower than that in the discontinuous missing data mode. This indicates that the MLP base model exhibits strong adaptability when handling continuous missing data segments. This is because the data changes in continuous missing data segments typically follow the overall trend of structural stress evolution over time. This trend is strongly correlated with the input load and temperature variables, making it easily captured by the MLP model. In contrast, discontinuous missing data points are often scattered in areas of drastic signal fluctuations or containing random noise. The base regression model struggles to fully fit these instantaneous high-frequency changes, resulting in larger initial prediction errors and thus requiring a greater reliance on the compensation effect of the residual correction mechanism.

[0145] Evaluation of model repair performance under different missing patterns and missing rates: In order to systematically evaluate the adaptability and performance of the proposed "adaptive residual correction MLP model" under different missing degree and missing pattern data, this example designed a comparative experiment covering five missing rate gradients of 10%, 30%, 50%, 70%, and 90%, and tested them under two missing modes: non-continuous missing and continuous missing. Figure 9 and Figure 10 The model repair performance under different missing rates in both discontinuous and continuous missing modes is shown. Figure 9 (a), (b), (c), (d), and (e) in the figure show the repair effect of the model under non-continuous deletions of 10%, 30%, 50%, 70%, and 90% deletion rates, respectively. Figure 10 Tables (a), (b), (c), (d), and (e) in the table show the model's repair performance at missing rates of 10%, 30%, 50%, 70%, and 90% under continuous deletion conditions, respectively. Table 2 shows the MAE, RMSE, and Wmape quantitative indicators for different missing rates under both continuous and non-continuous deletion modes.

[0146] Table 2. Evaluation of Model Repair Effectiveness at Different Missing Rates under Different Missing Patterns

[0147]

[0148] from Figure 9 and Figure 10 The comparison of time-domain waveforms shows that, regardless of whether it's a discontinuous missing data pattern with random data point loss or a continuous missing data pattern with long data segments, the model's predicted curve consistently and closely follows the fluctuation trend of the true value curve. Even with an extreme missing rate of 90%, the model still accurately reproduces the main peak and trough characteristics and periodic patterns of the strain response, without significant trend deviation or oscillation divergence. This indicates that the method has strong adaptability and robustness to different types and degrees of data missingness, and can meet the high requirements for data integrity in complex engineering environments.

[0149] Subsequent analysis of the specific evaluation metrics in Table 2 reveals that, regardless of whether the model faces discrete, discontinuous missing data or large segments of continuous missing data, it achieves low error metrics at all missing rate levels. This further demonstrates that the proposed adaptive residual correction MLP model possesses stable and reliable repair performance across various missing data scenarios. Furthermore, Table 2 shows that, especially in the discontinuous missing data mode, the evaluation metrics do not change significantly with increasing missing rate. This is because there is a strong physical mapping relationship between the input temperature and load auxiliary variables and the target strain, enabling the model to accurately infer strain changes based on complete driving variables even when the target data is extremely sparse.

[0150] As shown in the above figures, the proposed adaptive residual correction MLP model exhibits good repair performance in most missing scenarios. However, compared to discontinuous missing modes, the reconstruction accuracy of the model still decreases to some extent under continuous long missing segments. This is mainly because the loss of continuous large segments of data severs local spatiotemporal correlations, making it difficult for the neighborhood-based residual correction mechanism to obtain effective reference information, thus reducing the model's prediction accuracy in these regions.

[0151] It should be noted that, in order to solve the problem of physical consistency repair of key features, the sub-technical solution in this embodiment adopts the implementation method of constructing high-quality features, global nonlinear regression, and local adaptive correction in sequence.

[0152] This embodiment introduces a data-driven method with nonlinear modeling capabilities to deeply mine the inherent coupling mechanism between environmental factors, traffic loads and strain responses from multi-source monitoring information, thereby achieving high-precision repair of missing key features.

[0153] A feature-driven repair model based on a multi-layer perceptron (MLP) is designed to learn the complex nonlinear mapping between multi-source driving variables (such as temperature and strain at other measuring points) and the target strain response, thereby achieving high-fidelity recovery of key features and providing a solid and reliable data foundation for subsequent fatigue damage analysis and structural safety status assessment.

[0154] Example 2

[0155] The purpose of this embodiment is to provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.

[0156] Example 3

[0157] The purpose of this embodiment is to provide a computer-readable storage medium.

[0158] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the steps of the above method.

[0159] Example 4

[0160] The purpose of this embodiment is to provide a system for repairing missing critical strain features in traffic infrastructure monitoring, including:

[0161] The structured feature dataset construction module is configured to: acquire strain time series data for traffic infrastructure monitoring and extract features, obtain strain data features, and construct a structured feature dataset.

[0162] The basic regression model construction module is configured to: build a basic regression model based on a multilayer perceptron; introduce a local residual estimation mechanism based on K nearest neighbors for local bias correction and a hyperparameter optimization strategy to find the optimal hyperparameters;

[0163] The repair module is configured to: acquire the strain time series of the traffic infrastructure to be repaired, input it into the basic regression model based on the multilayer perceptron, and output the key features of the strain after repair.

[0164] Example 5

[0165] The purpose of this embodiment is to provide a computer program product containing instructions that, when run on a computer, cause the computer to perform the methods and functions involved in any of the above embodiments.

[0166] The steps and methods involved in the apparatus of the above embodiments correspond to those in Embodiment 1. For specific implementation details, please refer to the relevant description section of Embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood as including any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and enabling the processor to perform any of the methods in this invention.

[0167] Those skilled in the art will understand that the modules or steps of the present invention described above can be implemented using general-purpose computer devices. Optionally, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computer device, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. The present invention is not limited to any particular combination of hardware and software.

[0168] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for repairing missing key strain features in traffic infrastructure monitoring, characterized by: include: Obtain strain time series data for traffic infrastructure monitoring and extract features to obtain strain data features and construct a structured feature dataset; A basic regression model is established based on a multilayer perceptron. A local residual estimation mechanism based on K-nearest neighbors is introduced to perform local bias correction and hyperparameter optimization strategies to find the optimal hyperparameters. When establishing a local residual estimation mechanism, the following should be included: Define the observation domain and missing domain of the monitoring dataset; A basic regression model is established based on the trained multilayer perceptron. Back-substitution prediction is performed on all time i within the observation domain to calculate the deviation between the actual observed value and the basic predicted value of the model. If the residuals exhibit a random distribution of alternating positive and negative values ​​within a local time window, it is considered measurement noise; however, if the residuals exhibit a continuous positive or negative bias trend locally, it indicates the presence of local systematic errors. For any missing point to be repaired, assuming that its potential prediction error has a high degree of similarity and transitivity to the residuals of known points in its neighborhood, a dynamic search window is constructed to find the k nearest valid observation points in the observation domain, forming a neighborhood set: Based on the known residual information within the neighborhood set, an inverse distance weighting algorithm is used for the target point. The weight of each reference point in the neighborhood is inversely proportional to its distance. Thus, the estimated residual of the target point can be expressed as a weighted average of the neighborhood residuals. A correction intensity coefficient is introduced to adaptively adjust the correction magnitude. The estimated residual compensation is superimposed on the basic predicted value of the basic regression model established by the multilayer perceptron to obtain the final reconstructed value after local correction. Then, the strain time series of the traffic infrastructure to be repaired is obtained, input into the basic regression model based on multilayer perceptron, and the key strain features after repair are output.

2. The method for repairing missing key strain features in transportation infrastructure monitoring as described in claim 1, characterized in that, When acquiring strain data characteristics, the specific steps include: The acquired strain time series data for traffic infrastructure monitoring are grouped into master sensor sequences and slave sensor sequences. Identify significant peaks in the main sensor sequence and their corresponding timestamps, and establish them as the global synchronization benchmark for this loading loop; The time difference between the first significant peaks in the master and slave channels is calculated as the global phase offset. This offset is then used to construct a dynamic search window centered on the reference timestamp in each slave sensor channel. This window adaptively searches for local peaks that physically correspond to the master peak. After locking the peak, the window preceding the peak is used to search for the nearest significant valley in reverse order, thus constructing a pair of peak-valley feature pairs.

3. The method for repairing missing key strain features in traffic infrastructure monitoring as described in claim 1, characterized in that it also... include: Using minimum peak spacing and relative peak value as constraints, a peak is considered a valid feature point only when the amplitude difference between it and its adjacent trough exceeds a set multiple of the signal standard deviation.

4. The method for repairing missing key strain features in transportation infrastructure monitoring as described in claim 1, characterized in that, It also includes adaptive anomaly cleaning of feature sequences in structured feature datasets: Set a sliding time window with a set length; For the data sequence within the window, the lower quartile and upper quartile of the data within the window are calculated in real time, and a dynamic interquartile range is defined. Based on this, an adaptive anomaly detection boundary that changes dynamically over time is constructed. Data points falling outside this dynamic boundary are all considered outliers; For identified outliers, local linear interpolation or median of normal data within the window is used for replacement.

5. A strain key feature missing repair system for traffic infrastructure monitoring, characterized by: include: The structured feature dataset construction module is configured to: acquire strain time series data for traffic infrastructure monitoring and extract features, obtain strain data features, and construct a structured feature dataset. The basic regression model construction module is configured to: build a basic regression model based on a multilayer perceptron, introduce a local residual estimation mechanism based on K nearest neighbors for local bias correction and a hyperparameter optimization strategy to find the optimal hyperparameters; When establishing a local residual estimation mechanism, the following should be included: Define the observation domain and missing domain of the monitoring dataset; A basic regression model is established based on the trained multilayer perceptron. Back-substitution prediction is performed on all time i within the observation domain to calculate the deviation between the actual observed value and the basic predicted value of the model. If the residuals exhibit a random distribution of alternating positive and negative values ​​within a local time window, it is considered measurement noise; however, if the residuals exhibit a continuous positive or negative bias trend locally, it indicates the presence of local systematic errors. For any missing point to be repaired, assuming that its potential prediction error has a high degree of similarity and transitivity to the residuals of known points in its neighborhood, a dynamic search window is constructed to find the k nearest valid observation points in the observation domain, forming a neighborhood set: Based on the known residual information within the neighborhood set, an inverse distance weighting algorithm is used for the target point. The weight of each reference point in the neighborhood is inversely proportional to its distance. Thus, the estimated residual of the target point can be expressed as a weighted average of the neighborhood residuals. A correction intensity coefficient is introduced to adaptively adjust the correction magnitude. The estimated residual compensation is superimposed on the basic predicted value of the basic regression model established by the multilayer perceptron to obtain the final reconstructed value after local correction. The repair module is configured to: acquire the strain time series of the traffic infrastructure to be repaired, input it into the basic regression model based on the multilayer perceptron, and output the key features of the strain after repair.

6. The strain critical feature missing repair system for traffic infrastructure monitoring as described in claim 5, characterized in that, In the structured feature dataset construction module, when acquiring strain data features, the specific steps include: The acquired strain time series data for traffic infrastructure monitoring are grouped into master sensor sequences and slave sensor sequences. Identify significant peaks in the main sensor sequence and their corresponding timestamps, and establish them as the global synchronization benchmark for this loading loop; The time difference between the first significant peaks in the master and slave channels is calculated as the global phase offset. This offset is then used to construct a dynamic search window centered on the reference timestamp in each slave sensor channel. This window adaptively searches for local peaks that physically correspond to the master peak. After locking the peak, the window preceding the peak is used to search for the nearest significant valley in reverse order, thus constructing a pair of peak-valley feature pairs.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 4.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method described in any one of claims 1-4.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it performs the steps of the method described in any one of claims 1-4 above.

Images