A method and system for intelligently improving the quality of long-term performance monitoring data of asphalt pavement

Abstract

This invention discloses a method and system for intelligently improving the quality of long-term performance monitoring data for asphalt pavements. The method includes: extracting multi-dimensional features from raw sensor data and automatically identifying sensor layers; performing differentiated preprocessing based on layer labels; calculating a quality score Q and adaptively adjusting the anomaly detection threshold accordingly to intelligently distinguish and repair measurement anomalies; and performing hierarchical physical constraint verification on the data and establishing a feedback loop. This invention solves problems such as unclear sensor layers, distorted preprocessing, accidental deletion of anomalies, and lack of physical constraints for repair, achieving intelligent governance of the entire data process, significantly improving the reliability of monitoring data, and providing a high-quality data foundation for accurate pavement health assessment.

Description

Technical Field

[0001] This invention relates to the field of road engineering monitoring and data processing technology, and in particular to a method and system for intelligently improving the quality of long-term performance monitoring data of asphalt pavement. Background Technology

[0002] As transportation infrastructure operation and maintenance develops towards intelligence and refinement, long-term online monitoring of asphalt pavement structure has become an important means of assessing its health status and achieving preventive maintenance. However, the quality of monitoring data is the fundamental bottleneck restricting its effective analysis and application. The raw data generally has the following problems: (1) The actual installation layer of the sensor is unknown due to construction deviation or recording errors, resulting in unclear data source objects; (2) Signals from different structural layers (viscoelastic surface layer, elastoplastic base layer, loose subbase layer) have different physical characteristics, but traditional methods use a unified preprocessing strategy, resulting in loss of effective signals or noise residue; (3) It is difficult to distinguish whether the outliers in the data are "sensor measurement errors" or "actual pavement damage responses", and simple removal will lead to the loss of key information; (4) The processed data may be mathematically smooth but violate basic engineering physical laws, resulting in distortion of subsequent analysis.

[0003] In existing technologies, many studies focus on using monitoring data for health assessment, but lack systematic solutions to the source quality issues of the data itself. For example, existing literature CN118246245A discloses a monitoring method for solid waste asphalt pavements, which assesses performance degradation by establishing an initial state regression function. However, this method highly relies on the accuracy and reliability of the input strain, stress, and other monitoring data, and assumes that the layer information of the sensors is correct, without addressing how to ensure and improve the quality of these raw data. When the basic data contains layer misalignment or noise interference, the reliability of its assessment results will be significantly reduced.

[0004] Existing literature CN119848467A discloses an integrated asphalt pavement health status monitoring system, which includes a data processing module and adopts a generalized data cleaning, quality assessment, and anomaly handling process, and utilizes a deep learning model for feature extraction. The system's shortcomings lie in its generalized and black-box data processing strategy, failing to closely integrate the inherent physical and mechanical properties of each structural layer of the asphalt pavement to formulate differentiated governance rules with clear physical meaning. It also struggles to automatically address core issues such as sensor layer identification, signal fidelity preprocessing based on layer knowledge, and physical verification to ensure the rationality of data engineering.

[0005] Therefore, how to construct an intelligent data governance system that does not rely on manual recording, can automatically identify sensor layers, perform differentiated signal processing based on layer characteristics, intelligently diagnose and repair anomalies, and whose final output data conforms to engineering physical laws has become an urgent technical problem to be solved in order to achieve accurate perception and analysis of the long-term performance of asphalt pavement. Summary of the Invention

[0006] The main objective of this invention is to provide a method and system for intelligently improving the quality of long-term performance monitoring data for asphalt pavements, aiming to solve fundamental quality problems existing in the current asphalt pavement monitoring data sources, such as unclear layer location, mixed features, intertwined noise, and difficulty in distinguishing between genuine and fake data.

[0007] To achieve the above objectives, this invention provides a method for intelligently improving the quality of long-term performance monitoring data for asphalt pavements, the method comprising the following steps:

[0008] Step 1: Receive the raw time-series data collected by the sensor and extract multi-dimensional features based on the raw time-series data;

[0009] Step 2: Match the multidimensional features with the pre-built feature library of different road structure layers, automatically determine the road structure layer to which the sensor belongs, and output the corresponding layer label and corresponding confidence score;

[0010] Step 3: Based on the labels of each layer, call the preset data processing strategies of each structural layer to preprocess the raw time series data. The data processing strategies include differentiated noise reduction, detrending and interpolation strategies.

[0011] Step 4: Perform a quality assessment on the preprocessed data from Step 3, and calculate the quality score Q from four dimensions: data integrity, smoothness, signal-to-noise ratio, and physical rationality.

[0012] Step 5: Perform anomaly detection and intelligent diagnosis on the preprocessed data from Step 3 to distinguish between measurement anomalies and actual physical responses, and implement corresponding processing strategies for measurement anomaly and actual physical response data.

[0013] Step 6: Perform hierarchical physical constraint verification on the data processed in Step 5. If the verification fails, the error information will be fed back to the previous step for iterative processing. If the verification is successful, the processing result will be output.

[0014] Optionally, in step 1, the original time-series data includes strain time series and temperature time series. The multidimensional features include frequency domain features, time domain statistical features, temperature correlation features, and dynamic response features extracted from the strain and temperature time-series data. The frequency domain features are obtained by performing a fast Fourier transform on the strain time series. The time domain statistical features are obtained by performing statistical calculations on the strain time series. The temperature correlation features are obtained by performing linear regression analysis on the strain time series and temperature time series. The dynamic response features are obtained through time domain analysis.

[0015] Optionally, in step 2, the feature library contains typical feature ranges of surface layer, base layer and subbase layer established based on pavement mechanics theory, simulation or historical calibration data; the matching is achieved by calculating weighted Euclidean distance or using a machine learning classifier, and outputs a normalized confidence score; when the confidence score is lower than a first threshold, it is marked as requiring manual review.

[0016] Optionally, in step 3, the noise reduction strategy is as follows: for the bottom data of the surface layer, a bandpass filter with a passband range of 20~60Hz is used; for the bottom data of the base layer, a bandpass filter with a passband range of 3~40Hz is used; and for the bottom data of the bottom layer, a low-pass filter with a cutoff frequency of 8Hz is used.

[0017] Optionally, in step 4, the quality score Q is calculated as follows: scores are given for the four dimensions of data integrity, smoothness, signal-to-noise ratio and physical rationality, and then the scores of the four dimensions are weighted and summed, wherein the weight of integrity is 0.3, the weight of smoothness is 0.2, the weight of signal-to-noise ratio is 0.3 and the weight of physical rationality is 0.2.

[0018] Optionally, in step 5, anomaly detection employs a statistical unsupervised learning method, which identifies anomalies by calculating the degree of deviation of data points relative to a local neighborhood, wherein the local neighborhood consists of 50 data points before and after the data points.

[0019] Optionally, in step 5, the sensitivity threshold for anomaly detection is adaptively adjusted based on the quality score Q, and the formula for adaptively adjusting the sensitivity threshold is:

[0020] T = T0 × (1 / Q), where T is the adjusted threshold and T0 is the baseline threshold, set to 3 times the standard deviation.

[0021] Optionally, in step 5, the intelligent diagnosis performs five levels of judgment on the detected anomalies sequentially:

[0022] The first step is to determine whether the absolute value of the strain exceeds the sensor's range.

[0023] The second stage involves determining whether there are signal mutations exceeding a preset mutation threshold.

[0024] The third step is to determine whether the strain temperature coefficient deviates from the typical range of the layer by a preset ratio.

[0025] The fourth level is to determine whether the proportion of high-frequency energy is abnormally high.

[0026] Level 5 involves preliminary physical constraint testing;

[0027] If any one of the first four levels of judgment is true, it is determined to be a measurement anomaly, and the corresponding method is used to repair it according to the anomaly type; if none of the first four levels are true and the fifth level of judgment is passed, it is determined to be a real physical response and is retained and marked.

[0028] Optionally, in step 6, the hierarchical physical constraint verification includes three levels:

[0029] First-level strong constraint verification verifies whether the data meets the sensor's physical range and the limiting strain range of materials in each layer.

[0030] Second-level constraint verification verifies whether the elastic modulus calculated from the back-calculated data is within the reasonable range of the corresponding layer material, and whether the strain temperature coefficient conforms to the typical value.

[0031] The three-level weak constraint verification includes: verifying whether the spatial correlation between multiple sensor data in the same layer is higher than a preset threshold; and verifying whether the strain values ​​at the bottom of different layers satisfy the attenuation relationship from top to bottom under the same load.

[0032] Furthermore, to achieve the above objectives, the present invention also provides a system for improving the quality of long-term performance monitoring data of asphalt pavement, for implementing the method described in any of the above claims, the system comprising:

[0033] The data access and feature extraction module is used to receive raw data and extract multidimensional features;

[0034] The layer intelligent identification module is connected to the data access and feature extraction module and is used to output the layer label and confidence score of the sensor based on feature matching.

[0035] A differentiated preprocessing module, connected to the layer intelligent identification module, is used to preprocess the data by calling the corresponding strategy according to the layer label;

[0036] The data quality comprehensive assessment module is connected to the differential preprocessing module and is used to evaluate the preprocessed data and calculate the quality score Q.

[0037] An abnormal intelligent diagnosis and repair module is connected to the differential preprocessing module and the data quality comprehensive evaluation module, and is used to adaptively adjust the detection threshold according to the quality score Q to perform abnormal detection, diagnosis and repair.

[0038] The physical constraint verification module is connected to the anomaly intelligent diagnosis and repair module and is used to perform hierarchical physical constraint verification on the repaired data. When the verification of the physical constraint verification module fails, the control will feed back the error information to the layer intelligent identification module or the anomaly intelligent diagnosis and repair module to trigger reprocessing.

[0039] Beneficial effects:

[0040] (1) This invention pioneered a sensor layer automatic identification method based on the characteristics of the data itself rather than construction records, which fundamentally solves the problem of data identity confusion caused by construction deviations and provides a correct premise for all subsequent analyses.

[0041] (2) This invention proposes a differentiated preprocessing concept of “layer label driven”, which customizes the signal processing flow according to the completely different material mechanical properties of the surface layer, base layer and subbase layer, and retains the key information reflecting the real behavior of each layer to the greatest extent, thus achieving signal fidelity.

[0042] (3) The anomaly diagnosis module of the present invention can effectively distinguish between “bad data” and “bad road conditions” by integrating multi-level logic and physical common sense, avoiding the accidental deletion of the actual road damage response and protecting the key structural state information.

[0043] (4) The present invention establishes a three-level physical constraint verification system to ensure that the final output data is not only mathematically clean, but also reasonable and credible in engineering physics, thus eliminating the generation of data that is "mathematically correct but physically incorrect".

[0044] (5) This invention constructs an adaptive feedback closed loop with data quality score Q as the core, enabling the system to have self-optimization and error correction capabilities, and greatly improving the robustness of the entire data governance process and the reliability of the output results. This system transforms profound knowledge of road engineering into computable and executable rules, forming a complete, intelligent, and closed-loop solution from data source governance to high-quality data output. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0046] Figure 1 This is a flowchart illustrating the intelligent improvement method for long-term performance monitoring data quality of asphalt pavement according to the present invention.

[0047] Figure 2 This is a schematic diagram of the feedback mechanism process disclosed in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the five-level decision tree disclosed in an embodiment of the present invention.

[0049] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0050] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0051] See Figure 1-3 This invention provides a flowchart illustrating a method for intelligently improving the quality of long-term performance monitoring data for asphalt pavements, wherein the method includes the following steps:

[0052] Step 1, Data Feature Extraction: Based on the raw time-series data collected by the sensors, extract multi-dimensional features that can characterize the dynamic mechanical properties of the asphalt pavement structural layers.

[0053] This step receives raw time-series data uploaded by strain sensors and temperature sensors embedded in the road surface. This raw time-series data includes strain time series and temperature time series. The strain time series is the dynamic strain response data of the road structure under load, collected by the strain sensors, while the temperature time series is the temperature data of the road structure, collected synchronously by the temperature sensors. The asphalt pavement structure typically consists of a surface layer, a base layer, and a subbase layer from top to bottom. Preferably, the sensor data sampling frequency is not less than 100Hz, based on the Nyquist sampling theorem. To accurately capture the high-frequency response components (up to approximately 50Hz) at the bottom of the asphalt pavement surface layer, the sampling frequency needs to be at least twice that. This sampling rate also fully meets the requirements for the base layer and subbase layer, which have even lower response frequencies.

[0054] Preferably, the surface layer is made of asphalt mixture, which has viscoelastic characteristics, is temperature sensitive, and generates a rapid high-frequency response under vehicle load, with a dominant frequency band of 30~50Hz, a maximum tensile strain of 800με, a maximum compressive strain of 2000με, and a strain temperature coefficient of 15~25με per degree Celsius. Preferably, the sensors embedded in the surface layer are located at the bottom of the asphalt mixture. The base layer is made of semi-rigid material, which has a high modulus, obvious elastic characteristics, a dominant frequency band of load response of 5~30Hz, a peak strain range of -400~400με, and a strain temperature coefficient of 5~10με per degree Celsius. Preferably, the sensors embedded in the surface layer are located at the bottom of the semi-rigid material. The subbase layer uses graded crushed stone. While this material experiences relatively low stress after attenuation by the upper layers, its low modulus results in significant actual strain and a slow response. The dominant frequency band is 0.5–5 Hz, the peak strain range is -1500–300 με, and the strain temperature coefficient is 1–3 με per degree Celsius. Preferably, the sensor embedded in the subbase layer is located at the bottom of the graded crushed stone. The strain temperature coefficient, which reflects the response of strain to temperature changes, comprehensively reflects the thermal expansion effect of the material and the influence of temperature on the material modulus. It is determined based on empirical statistical values ​​from a large amount of engineering monitoring data. Furthermore, based on these physical differences, layer identification can be achieved through data characteristics, and differentiated preprocessing and verification strategies can be adopted accordingly.

[0055] Furthermore, the multidimensional features include frequency domain features, time domain statistical features, temperature-related features, and dynamic response features.

[0056] The frequency domain characteristics are obtained by performing a Fast Fourier Transform on the strain time series using a 10-second time window. This transformation from the time domain to the frequency domain primarily includes the dominant frequency (the frequency component with the highest energy in the spectrum), energy percentage (the percentage of the signal's total energy within a specific frequency band), and bandwidth (the width of the frequency range where the signal's main energy is distributed). The energy percentage is defined as the high-frequency energy percentage in the 30-50Hz band, the mid-frequency energy percentage in the 5-30Hz band, and the low-frequency energy percentage in the 0.5-5Hz band. Generally, the surface layer has the highest response frequency, followed by the base layer, and the subbase layer has the lowest.

[0057] The time-domain statistical features are obtained by statistically calculating the strain time series using a 10-second time window. The time-domain statistical features include peak strain (the maximum absolute value of strain within the time window), strain standard deviation (the standard deviation of strain values ​​within the time window, reflecting the dispersion of the data), and coefficient of variation (the ratio of standard deviation to mean, used to compare the dispersion of data at different levels).

[0058] The temperature-related characteristics were obtained by performing linear regression analysis on the synchronously acquired strain time series and temperature time series using a 10-second time window. These characteristics include the correlation coefficient between strain and temperature (a statistic that measures the degree of linear correlation between strain changes and temperature changes within the time period) and the strain-temperature coefficient (the slope obtained from the linear regression, which physically represents the amount of strain change caused by each 1-degree Celsius temperature change; this parameter is key to distinguishing stratigraphic layers).

[0059] The dynamic response characteristics are obtained by analyzing the strain pulse waveform caused by a single load event (such as a vehicle passing through) in the time domain, and the extracted features include response time (the time required for the strain to rise from a reference value (such as a static load state) to a peak value) and decay rate (the time or decay rate required for the strain to decay from the peak value to a certain proportion (such as 50%)).

[0060] Step 2, Automatic Layer Identification: The multi-dimensional features are matched with a pre-built feature library of different pavement structure layers to automatically determine the layer to which the sensor belongs and output a layer label. The feature library contains typical feature ranges of the surface layer, base layer, and subbase layer based on pavement mechanics theory, simulation, or historical calibration data. Specifically: Surface layer bottom: high dominant frequency (30-50Hz), high-frequency energy percentage >70%, large strain temperature coefficient (15-25με / ℃). Base layer bottom: medium dominant frequency (5-30Hz), medium-frequency energy percentage >60%, medium strain temperature coefficient (5-10με / ℃). Subbase layer bottom: low dominant frequency (0.5-5Hz), low-frequency energy percentage >80%, small strain temperature coefficient (1-3με / ℃).

[0061] The matching algorithm employs either weighted Euclidean distance (assigning weights to different features reflecting their discriminative power) or a machine learning classifier (such as a Support Vector Machine, SVM). The system uses the layer with the highest similarity or classification probability as the discrimination result, outputting the layer label and corresponding credibility score. The credibility score is calculated using normalized similarity, with the formula being the maximum similarity divided by the sum of the similarities of the three layers. A credibility score higher than 0.8 is considered high credibility; a credibility score between 0.6 and 0.8 is considered medium credibility, and the system accepts the recognition result but enhances monitoring in subsequent processing. If physical constraint verification fails, layer recognition errors are suspected first, and re-recognition is requested; a credibility score lower than 0.6 is considered low credibility and marked for manual review.

[0062] Step 3, Differentiated preprocessing step: Based on the layer label, invoke the preset data processing strategy for that layer to preprocess the original time series data.

[0063] The data processing strategies include differentiated noise reduction, drift correction, and interpolation strategies. Specifically,

[0064] (1) Select a noise reduction strategy based on the layer label. For the bottom data of the surface layer, use a 4th order Butterworth bandpass filter with a passband range of 20~60Hz to retain the high frequency response and filter out the low frequency drift. For the bottom data of the base layer, use a 4th order Butterworth bandpass filter with a passband range of 3~40Hz to retain the mid-to-high frequency components and filter out the extremely low and extremely high frequency noise. For the bottom data of the sub-base layer, use a 4th order Butterworth low-pass filter with a cutoff frequency of 8Hz to retain only the low frequency components and remove the high frequency noise.

[0065] (2) Select a detrending strategy based on the layer label. For the bottom data of the surface layer, use third-order polynomial fitting to remove the high-order nonlinear trend components caused by temperature. For the bottom data of the base layer, use linear detrending to remove the linear trend components in the data. For the bottom data of the bottom base layer, since the temperature effect is weak and the trend components in the data are not significant, the original trend is retained and no active detrending is performed.

[0066] (3) Select the interpolation method according to the layer label. Use cubic spline interpolation for the bottom data of the surface layer to maintain data smoothness, use linear interpolation for the bottom data of the base layer, and use forward filling for the bottom data of the sub-base layer.

[0067] Step 4: Perform a quality assessment on the preprocessed data and calculate the quality score Q from four dimensions: data integrity, smoothness, signal-to-noise ratio, and physical rationality.

[0068] The quality assessment module is used to score the quality of the preprocessed data. The module evaluates the data from four dimensions: data integrity, smoothness, signal-to-noise ratio (SNR), and physical reasonableness. Data integrity is calculated using the missing rate: a missing rate below 5% is excellent, 5%–10% is good, and above 10% is poor. Smoothness is assessed by calculating the first-order difference standard deviation of the data; a smaller standard deviation indicates smoother data. The SNR is calculated as the ratio of signal power to noise power: an SNR above 20dB is excellent, 10–20dB is good, and below 10dB is poor. The SNR evaluation standard is determined with reference to the "Highway Engineering Quality Inspection and Evaluation Standard." Physical reasonableness is assessed by whether the test data falls within the typical strain range of the stratum. The weighted sum of the scores from the four dimensions yields a comprehensive quality score Q, ranging from 0 to 1. The completeness score has a weight of 0.3, the smoothness score has a weight of 0.2, the signal-to-noise ratio score has a weight of 0.3, and the physical plausibility score has a weight of 0.2. A higher score indicates better data quality. This quality score is used for adaptive threshold adjustment in the subsequent anomaly detection module.

[0069] Step 5, Intelligent Anomaly Diagnosis and Repair: Perform anomaly detection and intelligent diagnosis on the preprocessed data to distinguish between measurement anomalies and actual physical responses, and implement corresponding processing strategies for measurement anomalies and actual physical response data.

[0070] Anomaly detection employs a statistically based unsupervised learning method. Anomalies are identified by calculating the deviation of each data point from its local neighborhood. The local neighborhood consists of 50 data points before and after the data point, corresponding to a time window of approximately 1 second at a sampling frequency of 100Hz. The anomaly detection threshold is adaptively adjusted based on the quality score Q. The adjustment formula is: threshold T equals the baseline threshold T0 multiplied by the reciprocal of the quality score Q. The baseline threshold T0 is set to three standard deviations, determined based on the 3-sigma principle in statistics. Higher quality scores result in lower thresholds and more stringent detection; lower quality scores result in higher thresholds and more lenient detection, avoiding misclassification of constant fluctuations as anomalies.

[0071] Furthermore, the anomaly detection step sets differentiated detection weights based on layer labels: high-frequency anomalies are prioritized for the bottom data of the surface layer, medium-frequency anomalies for the bottom data of the base layer, and low-frequency anomalies for the bottom data of the sub-base layer. After anomaly detection, the location of the anomaly point and the anomaly feature F are output and passed to the subsequent intelligent repair processing.

[0072] Furthermore, intelligent repair processing involves diagnosing the causes of detected anomalies and intelligently addressing them. For example... Figure 3 As shown, the intelligent repair module employs a five-level judgment process to distinguish between measurement anomalies and true responses. The first level checks whether the absolute strain value exceeds the sensor's range of 3000 με. This range limit is a typical technical parameter for commonly used strain gauge sensors. If it exceeds this limit, it is judged as a measurement anomaly and is deleted or interpolated. The second level checks for abrupt changes. If the absolute change in strain value relative to the previous moment exceeds 500 με and the data is smooth, it is judged as a measurement anomaly and median filtering is used. The abrupt change threshold of 500 με is determined based on the maximum strain change rate of the pavement material under normal load. The third level checks for temperature compensation failure. If strain and temperature are linearly correlated but the strain temperature coefficient deviates from the typical range of this layer by more than 30%, it is judged as temperature drift and re-compensation is performed, specifically by subtracting the abnormal temperature response component. The fourth level checks for high-frequency noise. If power spectrum analysis reveals an abnormally high proportion of high-frequency energy (above 85% for surface layer data, above 75% for base layer data, and above 30% for subbase layer data), it is judged as measurement noise and is subjected to low-pass filtering again. The fifth level of judgment checks whether the data meets the preliminary physical constraints. If none of the above judgments are met and the data passes the preliminary physical test, it is determined to be a possible true response, which is then retained and proceeds to the subsequent physical constraint verification process.

[0073] Preferably, different processing strategies are adopted for measurement anomalies and true responses. For measurement anomalies, corresponding repair methods are used according to the anomaly type, including deletion, interpolation, filtering, and temperature compensation. The repaired data replaces the original anomaly values. For true responses, the original data is retained without modification, but marked as a suspected true response, and engineers are alerted to this in the output report. The above processing strategy ensures that true information is not lost due to over-repair, nor is data quality affected by retaining measurement anomalies. At the same time, the judgment result and processing method of each anomaly point are recorded to form a repair log for easy traceability and review.

[0074] Furthermore, a conservative strategy is adopted for outliers that cannot be clearly identified. When an outlier meets multiple criteria or when the results are contradictory, it is prioritized as a true response rather than a measurement anomaly, avoiding accidental deletion or alteration of potentially accurate information. Such outliers are marked as highly suspected and requiring manual review in the output report, along with detailed characteristic information including strain values, temperature, spectral characteristics, and compliance with physical constraints, to assist engineers in their judgment. This conservative strategy reflects the system's respect for data authenticity and its responsible attitude towards engineering practice.

[0075] Step 6, Physical constraint verification step: Perform hierarchical physical constraint verification on the data processed in step 5.

[0076] The physical constraint verification module is configured with three levels of constraints: Level 1 strong constraints, Level 2 medium constraints, and Level 3 weak constraints. Level 1 strong constraints are mandatory and include strain range constraints and sensor range constraints. The strain range constraint stipulates that the absolute value of tensile strain at the bottom of the surface layer should be less than 1000 με, the absolute value of compressive strain should be less than 2000 με, the absolute value of strain at the bottom of the base layer should be less than 500 με, and the absolute value of strain at the bottom of the subbase layer should be less than 1500 με. These strain limits are determined based on the material's ultimate strain in the "Specifications for Design of Asphalt Pavement on Highways" (JTGD50-2017). Exceeding these limits indicates that the material may have cracked or undergone plastic deformation. The sensor range constraint stipulates that the absolute value of strain should be less than 3000 με. Violation of Level 1 constraints is considered a serious problem and is returned to step 5 for reprocessing.

[0077] Level II constraints are those that should be satisfied, including stress-strain relationship constraints and temperature-strain relationship constraints. Stress-strain relationship constraints are verified by checking whether the elastic modulus is within a reasonable range. The elastic modulus at the bottom of the surface layer should be 1500~4000 MPa, varying with temperature; approximately 1500 MPa at high temperatures and approximately 4000 MPa at low temperatures. The elastic modulus at the bottom of the base course should be 8000~15000 MPa, and the elastic modulus at the bottom of the subbase course should be 200~400 MPa. These modulus ranges are determined based on the "Highway Subgrade Design Specification" (JTGD30-2015) and engineering experience. Temperature-strain relationship constraints verify whether the strain-temperature coefficient is within the typical range for that layer. Violation of Level II constraints is marked as a warning and manual review is recommended.

[0078] Level 3 weak constraints serve as reference verification constraints, including spatial correlation constraints and longitudinal consistency constraints. The spatial correlation constraint verifies the consistency of data from multiple sensors within the same stratum. For multiple sensors identified as being at the bottom of the same stratum, the Pearson correlation coefficient is calculated. The correlation coefficient should be greater than a preset threshold, determined based on sensor spacing and engineering experience, typically ranging from 0.6 to 0.8. The smaller the sensor spacing, the higher the threshold: 0.8 for sensors with a spacing less than 10 meters, 0.7 for sensors with a spacing of 10-30 meters, and 0.6 for sensors with a spacing greater than 30 meters. A correlation coefficient below the threshold indicates insufficient correlation between sensor data within the same stratum. Possible causes include incorrect stratum identification, sensor damage, or special local conditions. In this case, the system systematically checks the sensor data, comparing the similarity of each sensor to the stratum feature library. Sensors with significantly low similarity (defined as below 80% of the average similarity) are identified and fed back to the stratum identification module for re-identification. The longitudinal consistency constraint examines the decreasing relationship of strain at the bottom of different layers. Under the same load, the strain at the bottom of the surface layer should be greater than that at the bottom of the base layer, which in turn should be greater than that at the bottom of the subbase layer, with an allowable deviation range of 20%. This deviation range is determined based on the material modulus ratio and load transfer law. If the longitudinal consistency constraint is violated, the process is also fed back to step 2 for re-identification. This three-level constraint tiered verification strategy balances the rigor and flexibility of the verification process.

[0079] The three-level physical constraint verification (Level 1 strong constraint, Level 2 medium constraint, and Level 3 weak constraint) transforms engineering experience into computable verification rules, ensuring that all output data conforms to the physical limits of sensors, material constitutive relations, and fundamental mechanical laws of load transfer (such as longitudinal attenuation consistency and spatial correlation). This eliminates the possibility of generating mathematically correct but physically incorrect data, fundamentally improving the credibility and reliability of data used in advanced evaluation models such as fatigue analysis and life prediction. It also constructs a clear hierarchical feedback loop. When physical verification fails, the system automatically diagnoses the root cause of the problem (e.g., improper repair of Level 1 constraint violations, or incorrect layer identification of Level 3 constraint violations), and accurately feeds back the error information to the corresponding preceding processing step (repair processing or layer identification processing) to trigger reprocessing. Combined with an upper limit on the number of iterations and a manual review trigger mechanism, the system achieves an intelligent closed loop of processing-verification-feedback-optimization, significantly improving the automation level of the data governance process and the consistency of the final results, while reducing the workload of engineers' manual verification and intervention.

[0080] If the verification is successful, the processing results will be output, including high-quality monitoring data, layer labels, quality scores, and data quality reports. The high-quality monitoring data are time-series data that have passed physical constraint verification. Each data point is associated with a sensor number, layer label, and timestamp. The layer label indicates whether the sensor data belongs to the bottom of the surface layer, the bottom of the base layer, or the bottom of the subbase layer. The quality score is a value from 0 to 1, reflecting the overall quality level of the data. The data quality report includes the layer identification results and reliability of each sensor, outlier marking and handling methods, physical constraint verification results, and suggestions for manual review.

[0081] The intelligent improvement method for long-term performance monitoring data quality of asphalt pavement of this invention systematically solves the quality problems across the entire chain of long-term performance monitoring of asphalt pavement, from data identification, fidelity preprocessing, accurate anomaly detection, to physical reliability verification, through a series of innovative and organically combined technical features. It produces highly reliable structured monitoring data with accurate layer labels, high-quality cleaning, conformity to engineering physical laws, and potentially containing early damage warning information. This lays an irreplaceable data foundation for the digital and intelligent management of pavement assets, and significantly improves upon existing technologies in terms of data accuracy, reliability, intelligent processing efficiency, and deep information mining capabilities.

[0082] Furthermore, the present invention also provides a system for improving the quality of long-term performance monitoring data of asphalt pavement, for implementing the method described in any one of the above, the system comprising:

[0083] The data access and feature extraction module is used to receive raw data and extract multidimensional features;

[0084] The layer intelligent recognition module is connected to the feature extraction module and has a built-in feature library for matching features and outputting sensor layer labels;

[0085] The differentiated preprocessing module is connected to the layer intelligent identification module and is used to call the corresponding strategy to perform data preprocessing based on the layer label;

[0086] The quality assessment module is used to score the quality of preprocessed data from four dimensions: data integrity, smoothness, signal-to-noise ratio, and physical reasonableness.

[0087] An anomaly detection and repair module connects the preprocessing module and the quality assessment module. The anomaly detection and repair module includes an anomaly detection unit and an intelligent repair unit. The anomaly detection unit is used to detect and diagnose anomalies, and the intelligent repair unit performs corresponding repair strategies on the data processed by the anomaly detection unit.

[0088] The physical constraint verification module is connected to the anomaly diagnosis and repair module and is used to perform hierarchical physical constraint verification on the repaired data.

[0089] Furthermore, the feature library in the layer intelligent identification module stores frequency domain, time domain, and temperature-sensitive response features predefined based on the material properties of the surface layer, base layer, and subbase layer.

[0090] Furthermore, the physical constraint verification module is also connected to the adaptive control module. That is, when the verification result from the physical constraint verification module fails, the adaptive control module sends the verification failure information and suggestions to one or more preceding modules to trigger reprocessing. Specifically, for example... Figure 2 As shown, the system has a two-level feedback mechanism. The first level is automatic feedback. When the first-level constraint verification of the physical constraint verification module fails, it is automatically fed back to the intelligent repair module to re-determine the anomaly type and repair it. A single piece of data is allowed a maximum of 3 intelligent repair feedbacks. If it still fails verification after more than 3 times, it is marked as a suspected sensor failure. When the third-level constraint verification of the physical constraint verification module fails, it is fed back to the layer identification module to re-perform layer identification. Layer identification feedback is allowed once. After re-identification, the complete process from differential preprocessing to physical constraint verification is executed in sequence, and the number of intelligent repair feedbacks is reset to zero and counted again. The second level is manual intervention. When the credibility score output by the layer identification module is lower than 0.6, or when the number of feedback repairs exceeds the upper limit, the system marks the data as pending manual review.

[0091] To better illustrate the method and system of the present invention, specific examples are provided below. Details are as follows:

[0092] This embodiment is based on a pavement monitoring project of the reconstruction and expansion of a provincial highway G207, to verify the practical application effect of the intelligent improvement method and system for long-term performance monitoring data of asphalt pavement of the present invention.

[0093] This section of road is a major arterial road with heavy traffic, averaging 8,500 vehicles per day (AADT), with heavy vehicles accounting for 28%. The road structure, from top to bottom, consists of: a 5cm AC-13 asphalt concrete surface layer, a 20cm cement-stabilized crushed stone base course, and a 20cm graded crushed stone subbase course. The construction unit deployed six strain gauge sensors, numbered S1 to S6, at the bottom of different layers of the road structure, with a sampling frequency of 100Hz. Temperature sensors were also installed to monitor road surface temperature changes. According to construction records, the sensor layer placement was as follows: S1 and S2 at the bottom of the surface layer, S3 and S4 at the bottom of the base course, and S5 and S6 at the bottom of the subbase course. However, during actual operation and monitoring, it was found that some sensor data characteristics did not match their nominal layer positions, and there were issues with missing construction records leading to uncertain layer information. Furthermore, various anomalies were found in the monitoring data, making it difficult to effectively improve data quality using traditional unified preprocessing methods.

[0094] The system continuously collected monitoring data for 7 days, spanning from June 15th to June 21st, 2025. This period covered different temperature conditions (daytime high of 37°C and nighttime low of 19°C) and different traffic load conditions. The raw data includes strain time series and corresponding temperature time series for each sensor. Preliminary inspection revealed that all six sensors could collect data normally, but the data contained quality issues such as noise interference, missing values, and suspected anomalies. Furthermore, the response characteristics of some sensors differed significantly from the layer markings in the construction records.

[0095] The system first extracts features from the raw data of six sensors. A 10-second time window is used to perform a Fast Fourier Transform on the time-series data to obtain the spectrum, and frequency domain features, time domain statistical features, temperature correlation features, and dynamic response features are extracted. Table 1 lists the key feature parameters extracted from each sensor.

[0096]

[0097] The layer identification module matches the extracted feature vectors with a pre-established layer feature library and calculates a weighted similarity. Table 2 lists the similarity between each sensor and the three layer feature libraries, as well as the final identification results.

[0098]

[0099] As shown in Table 2, the identification results of S1 and S2 are consistent with the construction records, both being the bottom of the surface layer, with confidence scores of 0.92 and 0.94 respectively, indicating high confidence. Although the construction record for S3 indicates the bottom of the base layer, its similarity to the surface layer bottom feature library is as high as 0.92, far exceeding the similarity of 0.45 with the base layer bottom feature library. The system identifies it as the bottom of the surface layer, with a confidence score of 0.89. Similarly, although the construction record for S4 indicates the bottom of the base layer, its similarity to the subbase layer bottom feature library is 0.88, significantly higher than the similarity of 0.52 with the base layer bottom feature library. The system identifies it as the bottom of the subbase layer, with a confidence score of 0.85. The identification results of S5 and S6 are consistent with the construction records, both being the bottom of the subbase layer. A global rationality check was performed, revealing that the identification results included 3 surface layer bottom sensors and 3 subbase layer bottom sensors, but no base layer bottom sensors were identified. Although this differs from the construction design, the feature matching results are clear. The system retains the identification results and marks them as requiring manual verification. After on-site verification by engineers and tracing of construction records, it was confirmed that No. S3 was actually installed at the bottom of the surface layer instead of the designed bottom of the base layer during construction, and No. S4 was actually installed at the bottom of the subbase layer instead of the base layer. There were errors in the construction records, and the automatic identification results of the layer identification module were completely correct.

[0100] Based on the layer identification results, a targeted preprocessing strategy was applied to the data from the six sensors using a differentiated preprocessing module. For S1, S2, and S3, identified as the bottom of the surface layer, a fourth-order Butterworth bandpass filter with a passband range of 20–60 Hz was used for noise reduction. Third-order polynomial fitting was used to remove nonlinear drift caused by temperature. Cubic spline interpolation was used for missing values, processing a total of 12 missing values. For S4, S5, and S6, identified as the bottom of the sub-base layer, a fourth-order Butterworth low-pass filter with a cutoff frequency of 8 Hz was used for noise reduction. No drift removal was performed. For missing values, forward padding was used, processing a total of 8 missing values.

[0101] Meanwhile, to verify the effectiveness of differentiated preprocessing, the method of this invention was compared with the traditional uniform preprocessing method. The traditional method uniformly applies a bandpass filter with a passband range of 3-40Hz and linear detrending processing to all sensor data. Table 3 lists the comparison of the preprocessing effects of the two methods.

[0102]

[0103] As shown in Table 3, the differentiated preprocessing improved the signal-to-noise ratio (SNR) by an average of 8.9 dB compared to the unified preprocessing, representing an improvement of 52.7%. In terms of smoothness, the first-order difference standard deviation decreased by an average of 21.3, a reduction of 49.8%. For the surface layer bottom sensors S1, S2, and S3, the unified preprocessing used a 3–40 Hz bandpass filter with an excessively low cutoff frequency, filtering out the effective high-frequency response signal in the 30–60 Hz band, resulting in limited SNR improvement. In contrast, the differentiated preprocessing used a 20–60 Hz bandpass filter, which fully preserved the high-frequency characteristics of the surface layer, significantly improving the SNR. For the bottom layer bottom sensors S4, S5, and S6, the unified preprocessing retained the mid-to-high frequency components of 8–40 Hz, which are mainly noise, resulting in noise residue. The differentiated preprocessing used a low-pass filter with a cutoff frequency of 8 Hz, effectively removing high-frequency noise and significantly improving data quality.

[0104] The preprocessed data was evaluated using a quality assessment module across four dimensions: data integrity, smoothness, signal-to-noise ratio, and physical plausibility, resulting in a comprehensive quality score Q. The evaluation results showed that the quality scores for the six sensors were: S1=0.88, S2=0.91, S3=0.85, S4=0.87, S5=0.83, and S6=0.89, with an average quality score of 0.87, indicating good preprocessed data quality.

[0105] The anomaly detection unit employs a statistically based unsupervised learning method to detect anomalies in the preprocessed data, with a local neighborhood size of 50 data points before and after the anomaly. The anomaly detection threshold is adaptively adjusted based on the quality score Q. For example, for S2, with a quality score Q=0.91 and a baseline threshold T0=3σ, the adjusted threshold T=3σ / 0.91=3.30σ. In 7 days of monitoring data, the system detected a total of 23 anomalies: 5 in S1, 3 in S2, 4 in S3, 6 in S4, 3 in S5, and 2 in S6. The system extracted features from each anomaly, including strain value, strain change, temperature, and spectral characteristics, and passed these to the intelligent repair module for further analysis.

[0106] The intelligent repair unit sequentially executes a five-level judgment process on the 23 detected anomalies to distinguish between measurement anomalies and actual responses. Three typical cases are listed below to illustrate the judgment process.

[0107] Case 1: Measurement Anomaly - Signal Sudden Change. Sensor S2 detected an anomaly at 14:23:15 on the third day (June 17th). The strain value suddenly changed from 205 με to 1850 με, a change of 1645 με. The five-level judgment process is as follows: First level judgment: The absolute strain value of 1850 με does not exceed the sensor's range of 3000 με, so it passes. Second level judgment: The absolute strain change of 1645 με is much greater than the sudden change threshold of 500 με, and the data before and after the sudden change are smooth and stable for 10 seconds, so it is judged to be a signal sudden change type of measurement anomaly. The system uses median filtering to repair the anomaly. After repair, the strain value is 198 με, which is continuous with the data before and after.

[0108] Case 2: Real-world Response - Heavy-duty Vehicle. Sensor S1 detected an anomaly at 10:36:42 on day 5 (June 19th), with a strain value of 1680 με, significantly higher than the average peak strain of 850 με over the past 7 days. The five-level judgment process is as follows: Level 1 Judgment: The absolute strain value of 1680 με does not exceed the sensor's range of 3000 με, passing. Level 2 Judgment: Checking the data before and after, it was found that the strain gradually increased from the baseline of 120 με to the peak value of 1680 με over 0.35 seconds, and then gradually decreased back to the baseline. The entire process was smooth and continuous, without any abrupt changes, passing. Level 3 Judgment: Correlation analysis of strain and temperature during this period showed that the strain temperature coefficient was 19.5 με / ℃, within the typical range of 15~25 με / ℃ at the bottom of the surface layer, passing. Level 4 Judgment: Power spectrum analysis showed that the dominant frequency was 39Hz, with a high-frequency energy ratio of 77.2%, and the spectrum characteristics were normal, passing. Level 5 Judgment: The strain value of 1680 με is within the peak tensile strain range of -800 to 2000 με at the bottom of the surface layer, meeting the preliminary physical constraints, and is deemed acceptable. The system determines this anomaly to be a genuine response under heavy vehicle load, retains the original data and marks it as a suspected genuine response, alerting engineers to its importance in the data quality report. Manual verification confirmed that a heavy truck did indeed pass through during this period, validating the system's judgment.

[0109] Case 3: Temperature Compensation Failure - Temperature Drift. Sensor S5 detected four anomalies in the data collected throughout the day on June 16th. Linear regression analysis of strain and temperature showed a strain temperature coefficient of 8.5 με / ℃, while the typical range for the bottom of the substrate is 1~3 με / ℃, a deviation of 183%, far exceeding the 30% threshold. The five-level judgment process is as follows: Levels 1 and 2 were passed. Level 3 judgment: The strain temperature coefficient was abnormally high, indicating temperature drift caused by temperature compensation failure. The system underwent temperature compensation again, deducting the abnormal temperature response component. After repair, the strain temperature coefficient decreased to 2.3 με / ℃, returning to the normal range.

[0110] After a five-level assessment, 18 of the 23 anomalies were identified as measurement anomalies and repaired, including 8 signal abrupt changes, 6 high-frequency noise, and 4 temperature drifts; 5 were identified as possible true responses and retained, including 3 heavy-load condition responses and 2 low-temperature high-strain responses. The repair methods included 8 rounds of median filtering, 6 rounds of low-pass filtering, 4 rounds of temperature compensation, and 5 rounds of data retention of the original data.

[0111] The physical constraint verification module performs three levels of physical constraint verification on the repaired data. Level 1 strong constraint verification results: All repaired data meet strain range constraints and sensor range constraints. The absolute values ​​of tensile strain at the bottom of the surface layer (S1, S2, S3) are all less than 1000 με, and the absolute values ​​of compressive strain are all less than 2000 με; the absolute values ​​of strain at the bottom of the subbase layer (S4, S5, S6) are all less than 1500 με; and the absolute values ​​of strain at all sensors are less than 3000 με. Level 2 medium constraint verification results: The stress-strain relationship constraint test shows that the elastic modulus calculated by the sensors at the bottom of the surface layer is between 2100 and 3600 MPa, which is consistent with the modulus range of asphalt mixture at different temperatures; the elastic modulus calculated by the sensors at the bottom of the subbase layer is between 250 and 380 MPa, which is consistent with the modulus range of graded crushed stone; the temperature-strain relationship constraint test shows that the strain temperature coefficient of each sensor is within the typical range of its respective layer. Only S4 showed an elastic modulus of 255 MPa during verification, slightly close to the lower limit of the modulus range of 200 MPa at the bottom of the subbase layer. The system marked this as a warning and recommended manual review. After review, the value was deemed to be within a reasonable range and was accepted. Level 3 weak constraint verification results: Spatial correlation constraint test: The Pearson correlation coefficients between the bottom layers S1, S2, and S3 were 0.89, 0.87, and 0.91, respectively, all greater than the threshold of 0.8, indicating good consistency of sensor data within the same layer. The correlation coefficients between the bottom layers S4, S5, and S6 were 0.82, 0.79, and 0.85, respectively, all meeting the threshold requirements. Longitudinal consistency constraint test: Peak strains at the bottom of different layers at the same time were compared. The results showed that the average peak strain at the bottom of the surface layer (1550 με) was greater than the average peak strain at the bottom of the subbase layer (1160 με, absolute value), consistent with the load transfer attenuation law and within the allowable deviation range.

[0112] During the verification process, two corrected anomalies failed the Level 1 strong constraint verification. The system automatically fed back these anomalies to the anomaly detection and repair module for reassessment and repair. After one feedback and repair, the verification was passed. All sensor data ultimately passed the Level 3 physical constraint verification without triggering the layer identification feedback.

[0113] To comprehensively evaluate the effectiveness of the method of this invention, it is systematically compared with traditional data processing methods. The traditional method employs the following process: determining sensor layer locations based on construction records, uniformly applying a 3-40Hz bandpass filter and linear detrending to all sensors, using a fixed threshold based on 3 times the standard deviation for anomaly detection, and uniformly repairing detected anomalies using linear interpolation, without physical constraint verification. Table 4 lists the comparison results of the two methods on several key indicators.

[0114]

[0115] In this embodiment, as shown in Table 4, the method of the present invention has advantages over traditional methods in several aspects. First, in terms of layer identification, traditional methods rely on construction records. In this case, the layer labeling errors of S3 and S4 were not detected, while the method of the present invention automatically identifies layers through data features and successfully corrects the two errors. Second, in terms of data preprocessing, the uniform preprocessing of traditional methods leads to the loss of high-frequency information in the surface layer and residual noise in the subbase layer, with an average signal-to-noise ratio of only 18.1 dB. The differentiated preprocessing of the method of the present invention adopts the optimal strategy for different layers, improving the average signal-to-noise ratio to 26.5 dB, an improvement of 46.4%, and improving the data smoothness index by 51.9%. Third, in terms of anomaly detection, traditional methods use a fixed threshold to detect 31 anomalies, including some over-detection. The method of the present invention uses an adaptive threshold for quality scoring, detecting 23 anomalies, making the detection more accurate and avoiding the problems of over-sensitivity or under-sensitivity. Fourth, regarding data quality, traditional methods lack a systematic quality assessment, with data integrity at 98.7% and a comprehensive quality score (Q) of 0.67. The method of this invention, through a four-dimensional quality assessment system, improves data integrity to 99.8% and achieves a comprehensive quality score (Q) of 0.87, an improvement of 29.9%. Fifth, regarding physical constraint verification, traditional methods do not perform physical law verification on the repaired data, potentially leading to results that are "mathematically correct but physically incorrect." The method of this invention, through three levels of physical constraint verification, ensures that all output data conforms to material mechanical properties, sensor physical limitations, and engineering common sense. All data passes verification through first-level strong constraints, second-level medium constraints, and third-level weak constraints.

[0116] Furthermore, this embodiment verifies the effectiveness and practicality of the intelligent improvement method and system for long-term performance monitoring data quality of asphalt pavement. The system successfully identified and corrected layer labeling errors from two sensors, improved the average signal-to-noise ratio of the data by 46.4% through differentiated preprocessing, and targeted the processing of 23 anomalies through a five-level judgment mechanism, including the repair of 18 measurement anomalies and the retention of 5 suspected true responses. Three-level physical constraint verification ensured that all repaired data conformed to engineering physical laws, and the final high-quality data comprehensive score Q reached 0.87, an improvement of 29.9% compared to traditional methods. The processed data provides a reliable data foundation for pavement structural mechanics analysis, fatigue life assessment, and maintenance decisions for this road section. In particular, the retained suspected true response data provided engineers with key clues for focus. This embodiment demonstrates that the present invention can significantly improve the quality of monitoring data in practical engineering applications, solve the layer confusion problem caused by construction deviations, achieve intelligent repair driven by differentiated preprocessing and physical constraints, and improve data quality and reliability.

[0117] The above are merely preferred embodiments of the present invention and do not limit the patent scope of the present invention. All equivalent structural transformations made using the contents of the present invention's specification and drawings under the inventive concept of the present invention, or direct / indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A method for intelligently improving the quality of long-term performance monitoring data of asphalt pavement, characterized in that, The method includes the following steps: Step 1: Receive the raw time-series data collected by the sensor and extract multi-dimensional features based on the raw time-series data; Step 2: Match the multidimensional features with the pre-built feature library of different road structure layers, automatically determine the road structure layer to which the sensor belongs, and output the corresponding layer label and corresponding confidence score; Step 3: Based on the labels of each layer, call the preset data processing strategies of each structural layer to preprocess the raw time series data. The data processing strategies include differentiated noise reduction, detrending and interpolation strategies. Step 4: Perform a quality assessment on the preprocessed data from Step 3, and calculate the quality score Q from four dimensions: data integrity, smoothness, signal-to-noise ratio, and physical rationality. Step 5: Perform anomaly detection and intelligent diagnosis on the preprocessed data from Step 3 to distinguish between measurement anomalies and actual physical responses, and implement corresponding processing strategies for measurement anomaly and actual physical response data. Step 6: Perform hierarchical physical constraint verification on the data processed in Step 5. If the verification fails, the error information will be fed back to the previous step for iterative processing. If the verification is successful, the processing result will be output.

2. The method according to claim 1, characterized in that, In step 1, the original time-series data includes strain time series and temperature time series. The multidimensional features include frequency domain features, time domain statistical features, temperature correlation features, and dynamic response features extracted from the strain and temperature time-series data. The frequency domain features are obtained by performing a fast Fourier transform on the strain time series. The time domain statistical features are obtained by performing statistical calculations on the strain time series. The temperature correlation features are obtained by performing linear regression analysis on the strain and temperature time series. The dynamic response features are obtained through time domain analysis.

3. The method according to claim 2, characterized in that, In step 2, the feature library contains typical feature ranges of surface layer, base layer and subbase layer established based on pavement mechanics theory, simulation or historical calibration data; the matching is achieved by calculating weighted Euclidean distance or using a machine learning classifier, and outputs a normalized confidence score; when the confidence score is lower than a first threshold, it is marked as requiring manual review.

4. The method according to claim 1, characterized in that, In step 3, the noise reduction strategy is as follows: for the bottom data of the surface layer, a bandpass filter with a passband range of 20~60Hz is used; for the bottom data of the base layer, a bandpass filter with a passband range of 3~40Hz is used; and for the bottom data of the bottom layer, a low-pass filter with a cutoff frequency of 8Hz is used.

5. The method according to claim 1, characterized in that, In step 4, the quality score Q is calculated as follows: scores are given for four dimensions: data integrity, smoothness, signal-to-noise ratio, and physical rationality. Then, the scores for the four dimensions are weighted and summed, with integrity having a weight of 0.3, smoothness having a weight of 0.2, signal-to-noise ratio having a weight of 0.3, and physical rationality having a weight of 0.

2.

6. The method according to claim 1, characterized in that, In step 5, anomaly detection employs a statistical unsupervised learning method, which identifies anomalies by calculating the degree of deviation of data points relative to their local neighborhood, where the local neighborhood consists of 50 data points before and after the data points.

7. The method according to claim 6, characterized in that, In step 5, the sensitivity threshold for anomaly detection is adaptively adjusted based on the quality score Q. The formula for adaptively adjusting the sensitivity threshold is: T = T0 × (1 / Q), where T is the adjusted threshold and T0 is the baseline threshold, set to 3 times the standard deviation.

8. The method according to claim 6, characterized in that, In step 5, intelligent diagnosis involves performing five levels of judgment on the detected anomalies: The first step is to determine whether the absolute value of the strain exceeds the sensor's range. The second stage involves determining whether there are signal mutations exceeding a preset mutation threshold. The third step is to determine whether the strain temperature coefficient deviates from the typical range of the layer by a preset ratio. The fourth level is to determine whether the proportion of high-frequency energy is abnormally high. Level 5 involves preliminary physical constraint testing; If any one of the first four levels of judgment is true, it is determined to be a measurement anomaly, and the corresponding method is used to repair it according to the anomaly type; if none of the first four levels are true and the fifth level of judgment is passed, it is determined to be a real physical response and is retained and marked.

9. The method according to claim 1, characterized in that, In step 6, the hierarchical physical constraint verification includes three levels: First-level strong constraint verification verifies whether the data meets the sensor's physical range and the limiting strain range of materials in each layer. Second-level constraint verification verifies whether the elastic modulus calculated from the back-calculated data is within the reasonable range of the corresponding layer material, and whether the strain temperature coefficient conforms to the typical value. Level 3 weak constraint verification includes: verifying whether the spatial correlation between multiple sensor data at the same level is higher than a preset threshold. Verify whether the strain values ​​at the bottom of different layers satisfy the attenuation relationship from top to bottom under the same load.

10. A system for improving the quality of long-term performance monitoring data for asphalt pavements, used to implement the method according to any one of claims 1-9, characterized in that, The system includes: The data access and feature extraction module is used to receive raw data and extract multidimensional features; The layer intelligent identification module is connected to the data access and feature extraction module and is used to output the layer label and confidence score of the sensor based on feature matching. A differentiated preprocessing module, connected to the layer intelligent identification module, is used to preprocess the data by calling the corresponding strategy according to the layer label; The data quality comprehensive assessment module is connected to the differential preprocessing module and is used to evaluate the preprocessed data and calculate the quality score Q. An abnormal intelligent diagnosis and repair module is connected to the differential preprocessing module and the data quality comprehensive evaluation module, and is used to adaptively adjust the detection threshold according to the quality score Q to perform abnormal detection, diagnosis and repair. The physical constraint verification module is connected to the anomaly intelligent diagnosis and repair module and is used to perform hierarchical physical constraint verification on the repaired data. When the verification of the physical constraint verification module fails, the control will feed back the error information to the layer intelligent identification module or the anomaly intelligent diagnosis and repair module to trigger reprocessing.

Images