A Road Defect Detection Method Based on Dynamic Weight Allocation and Confidence Optimization

Abstract

This invention discloses a road defect detection method based on dynamic weight allocation and confidence optimization, belonging to the field of road defect detection technology. The method includes: acquiring raw electromagnetic wave signals through ground-penetrating radar to construct radar spectral features and time-frequency domain features; training a YOLO-based spectral defect detection model and an XGBoost-based time-frequency domain feature recognition model respectively; fusing the outputs of the two models using a dynamic weight allocation mechanism, with weights dynamically adjusted through Bayesian updates; optimizing the prediction confidence using information entropy; and filtering the final result using a confidence threshold. This invention effectively improves the accuracy and robustness of multimodal data fusion and is suitable for the automated detection of road defects such as cracks and voids.

Description

Technical Field

[0001] This invention belongs to the field of road inspection technology, specifically relating to a road defect detection method based on dynamic weight allocation and confidence optimization. Background Technology

[0002] With the rapid development of transportation infrastructure, the requirements for the accuracy and real-time performance of road defect detection (such as cracks, cavities, and delamination) are increasing. Current mainstream detection methods mainly rely on ground-penetrating radar (GPR) to acquire radar image data of the road surface structure, followed by manual identification and interpretation of image features. This type of method has significant drawbacks:

[0003] Highly dependent on human experience: Test results are easily affected by the operator's technical level and subjective judgment, and there are significant differences in the interpretation results between different operators, resulting in insufficient reliability and consistency of the test.

[0004] Limitations of single-modal data: Traditional methods are mostly based on single radar spectrum data, which has limited information dimensions and cannot fully reflect the complex physical characteristics of road structural defects (such as energy distribution at material interfaces, frequency shift, etc.), which can easily lead to missed or false detections.

[0005] To improve detection accuracy, some studies have attempted to introduce multimodal data fusion techniques (e.g., combining spectral images with time-frequency domain features). However, existing fusion methods mostly employ fixed weight allocation or simple averaging strategies, lacking a dynamic adjustment mechanism for the importance of different modal data. This results in unstable fusion effects and difficulty in adapting to complex and changing road surface environments (such as different materials, humidity, or noise interference). Furthermore, current methods generally lack confidence assessment and optimization mechanisms for the recognition results, failing to effectively quantify prediction uncertainty, further limiting their widespread application in engineering practice. Summary of the Invention

[0006] The purpose of this invention is to provide a road defect detection method based on dynamic weight allocation and confidence optimization, which effectively improves the accuracy and robustness of multimodal data fusion and is applicable to the automated detection of road defects such as cracks and voids, thereby solving at least one of the technical problems mentioned in the background art.

[0007] To solve the above-mentioned technical problems, the present invention is implemented as follows:

[0008] This invention provides a road defect detection method based on dynamic weight allocation and confidence optimization, comprising the following steps:

[0009] Step S1, Multimodal Data Acquisition and Preprocessing:

[0010] The raw electromagnetic wave signals of the road are collected by ground-penetrating radar;

[0011] The original electromagnetic wave signal was reconstructed into a radar spectral image, and time-frequency domain features were extracted.

[0012] Preprocessing of radar spectrum image data includes zero bias removal, zero-point adjustment, gain adjustment, digital filtering, background removal, moving average, and gain compensation;

[0013] Based on the disease detection frame, the channel range and depth sampling range in the original electromagnetic wave signal data are inferred, and multiple A-scan single-channel waveform signals in the disease area are extracted to form a local signal matrix, which is then subjected to time-frequency transformation and normalization.

[0014] Step S2, Model Training:

[0015] Construct a YOLO-based map-based disease detection model, inputting a radar map image and outputting disease category, location bounding box, and confidence score;

[0016] Construct a time-frequency domain feature-based disease identification model based on machine learning, input the time-frequency domain features of the diseased area, and output the disease category and confidence level;

[0017] Step S3, Decision-level Fusion:

[0018] Optimize the confidence scores of the two models separately: adjust the original confidence scores based on information entropy, using the following formula:

[0019] ;

[0020] in For adjustment coefficients, As the initial confidence level, The optimized confidence level; where:

[0021] ;

[0022] ;

[0023] in For the first The sample model can provide the first... j The probability of each category, ; Total number of categories; Information entropy; The normalized information entropy;

[0024] Dynamic weight allocation: Model weights are calculated based on feature correlation and prediction error feedback, and a Bayesian update mechanism is used to iteratively optimize the weights. The formula is as follows:

[0025] ;

[0026] in It is the likelihood function and ， For the first The total error of each model within the current time window. To adjust the parameters; For prior confidence and , For the first The weights of each model in the previous round; For observation results; For the first One model; For the first One model;

[0027] The outputs of the two models are weighted and fused. Regions with confidence scores below a threshold after fusion are marked as uncertain regions, and conflicting results are decided by voting or eliminated.

[0028] Optionally, in step S1, the time-frequency domain feature extraction includes:

[0029] The local signal matrix is ​​aggregated, and wavelet transform or short-time Fourier transform is applied to obtain spectral features;

[0030] Principal component analysis was used to reduce the dimensionality of time-frequency features, and key features were screened through feature importance analysis.

[0031] Optionally, in step S2:

[0032] The atlas disease detection model uses the YOLOv8m network, and data augmentation and nonmaximum suppression are introduced during training.

[0033] The time-frequency domain feature disease identification model uses the XGBoost classifier, and combines Bayesian parameter tuning and cross-validation during training.

[0034] Optionally, in step S3, the dynamic weight allocation includes:

[0035] Initialize weights based on the correlation coefficient between features and target labels: ,in The correlation coefficient between the feature and the target;

[0036] Based on the model's prediction error within the sliding time window Update weights: ,in For the first The total error of each model; For the first The total error of the model.

[0037] Optionally, in step S3, the spatial positioning information of the fusion result is output in the following manner:

[0038] Vehicle radar data is combined with GPS positioning information to convert road markers and lane numbers;

[0039] The ground-based fixed radar outputs the two-dimensional plane coordinates of the disease relative to the monitoring point.

[0040] Optionally, the final output includes disease category, location information, and overall confidence level, which is calculated from the following factors:

[0041] The confidence scores of each sub-model after information entropy optimization;

[0042] Dynamic weight allocation results;

[0043] The maximum confidence score in the fusion probability distribution.

[0044] Compared with the prior art, the advantages of this invention are as follows:

[0045] 1. This invention adaptively adjusts the fusion weights of the YOLO graph model and the XGBoost time-frequency model through a dynamic weight allocation mechanism (Bayesian update) and confidence optimization (information entropy adjustment), effectively suppressing high-uncertainty prediction results (such as misjudgments caused by noise interference) and overcoming the limitations of traditional fixed-weight fusion strategies.

[0046] 2. The dual-model decision-level fusion of this invention fully utilizes the complementarity of spatial spectral features (image structures such as cracks and potholes) and time-frequency physical features (spectral characteristics such as interlayer peeling and loose structure) to comprehensively cover multiple types of road defects and reduce missed detections and false detections.

[0047] 3. The dynamic weighting mechanism of this invention automatically adjusts the weight allocation based on the real-time performance of the model (prediction error feedback), adapts to different road conditions (such as material differences and environmental noise) and data fluctuations, and ensures the stability of the fusion results in various scenarios.

[0048] 4. This invention automates the entire process (data acquisition → preprocessing → model recognition → fusion decision → result output), significantly reducing reliance on human experience and solving the problems of strong subjectivity and poor consistency in traditional methods.

[0049] 5. The output results of this invention include a confidence score (0–1 range) and support tiered early warning (e.g., areas marked as "uncertain" require manual review), improving the usability of the project.

[0050] 6. This invention supports high-speed detection by vehicle-mounted mobile radar (≥80km / h), and automatically associates the location of defects with road markers by combining GPS positioning, making it suitable for large-scale road inspection.

[0051] 7. The present invention provides a structured output of the detection report (disease type, spatial coordinates, confidence level), which directly serves maintenance decision-making and shortens the response cycle. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments 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 these drawings without creative effort, wherein:

[0053] Figure 1 A flowchart of a road defect detection method based on dynamic weight allocation and confidence optimization provided in an embodiment of the present invention;

[0054] Figure 2 This is a structural block diagram of a road defect detection system based on dynamic weight allocation and confidence optimization provided in an embodiment of the present invention. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] The terms "first," "second," etc., used in the specification and claims of this invention are used to distinguish similar objects and not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class and the number of objects is not limited; for example, a first object can be one or more. Furthermore, in the specification and claims, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0057] Please see Figure 1 As shown, this embodiment of the invention provides a road defect detection method based on dynamic weight allocation and confidence optimization, including the following steps:

[0058] Step S1, Multimodal Data Acquisition and Preprocessing:

[0059] The raw electromagnetic wave signals of the road are collected by ground-penetrating radar;

[0060] The original electromagnetic wave signal was reconstructed into a radar spectral image, and time-frequency domain features were extracted.

[0061] Preprocessing of radar spectrum image data includes zero bias removal, zero-point adjustment, gain adjustment, digital filtering, background removal, moving average, and gain compensation;

[0062] Based on the disease detection frame, the channel range and depth sampling range in the original electromagnetic wave signal data are inferred, and multiple A-scan single-channel waveform signals in the disease area are extracted to form a local signal matrix, which is then subjected to time-frequency transformation and normalization.

[0063] Step S2, Model Training:

[0064] Construct a YOLO-based map-based disease detection model, inputting a radar map image and outputting disease category, location bounding box, and confidence score;

[0065] Construct a time-frequency domain feature-based disease identification model based on machine learning, input the time-frequency domain features of the diseased area, and output the disease category and confidence level;

[0066] Step S3, Decision-level Fusion:

[0067] Optimize the confidence scores of the two models separately: adjust the original confidence scores based on information entropy, using the following formula:

[0068] ;

[0069] in To adjust the coefficient and control the strength of the entropy's correction to the confidence level, a suitable coefficient is determined experimentally. The value can effectively suppress prediction results with high uncertainty and improve the credibility of the overall identification results; As the initial confidence level, The optimized confidence level; where:

[0070] ;

[0071] ;

[0072] ;

[0073] in For the first The sample model can provide the first... j The probability of each category, ; Total number of categories; Information entropy; The normalized information entropy;

[0074] Dynamic weight allocation: Model weights are calculated based on feature correlation and prediction error feedback, and a Bayesian update mechanism is used to iteratively optimize the weights. The formula is as follows:

[0075] ;

[0076] in It is the likelihood function and ， For the first The total error of each model within the current time window. To adjust the parameters; For prior confidence and , For the first The weights of each model in the previous round; For observation results; For the first One model; For the first One model;

[0077] It should be noted that this formula implements round-by-round adaptive updates of the weights, so that models that have performed well recently receive higher fusion weights, while models that have performed poorly are dynamically suppressed.

[0078] The outputs of the two models are weighted and fused. Regions with confidence scores below a threshold after fusion are marked as uncertain regions, and conflicting results are decided by voting or eliminated.

[0079] In step S1, the ground-penetrating radar uses a 3D GPR radar device from Zhongdian Zhongyi Company, suitable for non-destructive fault detection on highways and urban roads. The device employs a "one-transmit, two-receive" mode, possessing excellent penetration capability and lateral resolution, and supports continuous data acquisition at high speeds. To avoid missed scans, the scanning frequency needs to be dynamically adjusted with vehicle speed. For example, at a vehicle speed of 72 km / h, the corresponding scanning frequency and scanning width (1.9 m) ensure gapless coverage. The device achieves high-precision positioning through the coordinated use of Doppler radar speed measurement, differential GPS positioning, and satellite timing systems. The acquired electromagnetic signals (A-scan) are stored in a standard format as the basis for subsequent feature extraction.

[0080] The original electromagnetic wave signal is reconstructed into a radar image. Specifically, the original electromagnetic wave signal is reconstructed according to the spatial dimension to form a B-scan or C-scan radar image, which is used for spatial structure analysis and disease image recognition.

[0081] Time-frequency domain feature extraction includes:

[0082] The local signal matrix is ​​aggregated, and wavelet transform or short-time Fourier transform is applied to obtain spectral features. The spectral features reflect the physical characteristics of the interface and diseased area, such as energy distribution and dominant frequency shift.

[0083] Principal component analysis (PCA) is used to reduce the dimensionality of time-frequency features, and key features are selected through feature importance analysis to improve model training efficiency and prediction performance.

[0084] Zero bias removal specifically includes: to avoid some zero bias caused during data acquisition, the zero bias component in the data is removed by correcting the zero bias. This can be solved by an averaging algorithm, which removes the zero bias by subtracting the average value of the data from each data point.

[0085] Zero-point adjustment specifically includes: zero-point adjustment eliminates the slight differences in the starting sampling points between channels by aligning the ground of each channel, ensuring that the starting points of all channels are consistent, and improving the imaging effect.

[0086] Gain adjustment specifically includes: enhancing the anomalies in radar data and performing gain processing using an energy attenuation method based on an exponential gain function.

[0087] Digital filtering specifically includes improving the signal-to-noise ratio by filtering out signal components in the data whose frequencies exceed the main operating bandwidth of the ground-penetrating radar system.

[0088] Background removal specifically includes: using a moving average method to remove background signals. This highlights abnormal targets by subtracting background signals from the original data.

[0089] The moving average specifically involves processing several adjacent data points and selecting a time window for the moving average.

[0090] The compensation gain specifically includes: after the moving average, readjusting the gain to make the disease characteristics more obvious.

[0091] The time-frequency transformation specifically includes: performing aggregation processing on the local signal matrix (such as channel mean, representative waveform extraction, etc.), and then applying algorithms such as wavelet transform and short-time Fourier transform (STFT) to obtain the spectral characteristics of the diseased area, such as the dominant frequency, energy distribution, and frequency centroid.

[0092] The normalization process specifically includes: considering the amplitude differences between different data sources (simulation / experimental), using the maximum-minimum normalization method to standardize the extracted features, ensuring that the scale of the model input features is uniform and comparable.

[0093] In step S2:

[0094] The atlas disease detection model uses the YOLOv8m network. During training, data augmentation (such as cropping, flipping, scaling, etc.) is introduced to improve the model's generalization ability. In the post-processing stage, non-maximum suppression (NMS) is used to remove redundant boxes, so as to achieve rapid localization and early warning of disease areas in the atlas.

[0095] The time-frequency domain feature disease identification model uses the XGBoost classifier, and combines Bayesian parameter tuning and cross-validation during training to provide physical feature support and secondary judgment capability for the subsequent fusion module.

[0096] In step S3, the dynamic weight allocation includes:

[0097] Initialize weights based on the correlation coefficient between features and target labels: ,in The correlation coefficient between the feature and the target;

[0098] The weights are updated based on the model's prediction error within the sliding time window: ;

[0099] in, For the first The total error of each model; , This represents the prediction error of the model within the most recent sliding time window. For the first The total error of each model; This represents the number of samples within the time window.

[0100] This weighting formula means that the smaller the error of a model, the greater its weight; conversely, the larger the error of a model, the weaker its contribution will be.

[0101] It should be noted that the initial weights reflect the degree of linear correlation between each model or feature channel and the label, serving as an initial reference for the fusion decision.

[0102] In step S3, the outputs of the two models are weighted and fused. Regions with confidence levels below a threshold after fusion are marked as uncertain regions. Conflicting class results are decided by voting or eliminated. Specifically, this includes:

[0103] The fusion strategy consists of two phases: intermediate decision fusion and confidence-driven final decision selection.

[0104] (a) Intermediate integration decision

[0105] First, the posterior weights of each model in the current round are obtained based on the aforementioned Bayesian update mechanism. , representing the model Relative credibility in this round of integration.

[0106] Let the discrimination result of each model be... This can be represented as: In classification tasks, This is the distribution vector of the model's predicted categories for the samples; in the detection task, This refers to the disease categories and bounding box set predicted by the model.

[0107] The fusion result Expressed as a weighted combination of the outputs of each model:

[0108] ;

[0109] in, The number of models participating in the fusion (2 in this invention). and The fusion results at this stage form an intermediate prediction set, retaining the important judgment information of each model and fusing their weight distributions.

[0110] (b) Confidence-driven outcome selection and conflict decision-making

[0111] For intermediate fusion results A confidence optimization mechanism (see the aforementioned information entropy adjustment method) is introduced to calculate the final confidence of the fused result. Based on the confidence threshold Perform a second screening:

[0112] like If so, the fusion result will be output directly;

[0113] like If so, the area is marked as an uncertain area and submitted for manual review or secondary judgment.

[0114] If the model results have class conflicts (e.g., YOLO classifies it as "cracked" while XGBoost classifies it as "loose"), then the results are ranked according to confidence score or confidence score difference threshold. Determine whether to vote to decide or eliminate inconsistent results.

[0115] In step S3, the spatial positioning information of the fusion result is output in the following manner:

[0116] Vehicle radar data is combined with GPS positioning information to convert road markers and lane numbers;

[0117] The ground-based fixed radar outputs the two-dimensional plane coordinates of the disease relative to the monitoring point.

[0118] The final output includes disease type, location information, and overall confidence level, which is calculated from the following factors:

[0119] The confidence scores of each sub-model after information entropy optimization;

[0120] Dynamic weight allocation results;

[0121] The maximum confidence score in the fusion probability distribution.

[0122] Following step S3, step S4 is also included, which involves visualizing the results and generating a report, including:

[0123] ① Automatic image annotation: Different colored borders are used to annotate various diseases on the original B-scan image, with the disease category name and confidence level value attached to the border;

[0124] ②Structured inspection report: Outputs the category, spatial location (station number or relative coordinates), and confidence value for each defect;

[0125] ③ Multi-mode positioning output: Vehicle data combined with GPS coordinates generates real road segment markers; Ground-based radar data provides two-dimensional planar positioning;

[0126] ④ Confidence score: Each result is accompanied by a confidence score in the range [0,1], which is used to indicate the reliability of the identification result.

[0127] Combined Figure 2 As shown, the present invention also provides a road defect detection system based on dynamic weight allocation and confidence optimization, for implementing the method, including a data acquisition module 1, a preprocessing module 2, a storage module 3, and a fusion decision module 4.

[0128] The data acquisition module 1 is used to acquire raw electromagnetic wave signals through ground-penetrating radar.

[0129] The preprocessing module 2 is used to generate radar spectral images and time-frequency domain features.

[0130] The storage module 3 is used to store the trained YOLO map disease detection model and XGBoost time-frequency domain feature recognition model.

[0131] The fusion decision module 4 is used to perform confidence optimization, dynamic weight allocation, and Bayesian weighted fusion.

[0132] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0133] Furthermore, it should be noted that the scope of the methods and systems in the embodiments of the present invention is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. In addition, features described with reference to certain examples may be combined in other examples.

[0134] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A road defect detection method based on dynamic weight allocation and confidence optimization, characterized in that, Includes the following steps: Step S1, Multimodal Data Acquisition and Preprocessing: The raw electromagnetic wave signals of the road are collected by ground-penetrating radar; The original electromagnetic wave signal was reconstructed into a radar spectral image, and time-frequency domain features were extracted. Preprocessing of radar spectrum image data includes zero bias removal, zero-point adjustment, gain adjustment, digital filtering, background removal, moving average, and gain compensation; Based on the disease detection frame, the channel range and depth sampling range in the original electromagnetic wave signal data are inferred, and multiple A-scan single-channel waveform signals in the disease area are extracted to form a local signal matrix, which is then subjected to time-frequency transformation and normalization. Step S2, Model Training: Construct a YOLO-based map-based disease detection model, inputting a radar map image and outputting disease category, location bounding box, and confidence score; Construct a time-frequency domain feature-based disease identification model based on machine learning, input the time-frequency domain features of the diseased area, and output the disease category and confidence level; Step S3, Decision-level Fusion: Optimize the confidence scores of the two models separately: adjust the original confidence scores based on information entropy, using the following formula: ； in For adjustment coefficients, As the initial confidence level, The optimized confidence level; where: ； ； in For the first The sample model can provide the first j The probability of each category, ; Total number of categories; Information entropy; The normalized information entropy; Dynamic weight allocation: Model weights are calculated based on feature correlation and prediction error feedback, and a Bayesian update mechanism is used to iteratively optimize the weights. The formula is as follows: ； in It is the likelihood function and ， For the first The total error of each model within the current time window. To adjust the parameters; For prior confidence and , For the first The weights of each model in the previous round; For observation results; For the first One model; For the first One model; The outputs of the two models are weighted and fused. Regions with confidence scores below a threshold after fusion are marked as uncertain regions, and conflicting results are decided by voting or eliminated.

2. The method according to claim 1, characterized in that, In step S1, the time-frequency domain feature extraction includes: The local signal matrix is ​​aggregated, and wavelet transform or short-time Fourier transform is applied to obtain spectral features; Principal component analysis was used to reduce the dimensionality of time-frequency features, and key features were screened through feature importance analysis.

3. The method according to claim 1, characterized in that, In step S2: The atlas disease detection model uses the YOLOv8m network, and data augmentation and nonmaximum suppression are introduced during training. The time-frequency domain feature disease identification model uses the XGBoost classifier, and combines Bayesian parameter tuning and cross-validation during training.

4. The method according to claim 1, characterized in that, In step S3, the dynamic weight allocation includes: Initialize weights based on the correlation coefficient between features and target labels: ,in The correlation coefficient between the feature and the target; Based on the model's prediction error within the sliding time window Update weights: ,in For the first The total error of the model; For the first The total error of the model.

5. The method according to claim 1, characterized in that, In step S3, the spatial positioning information of the fusion result is output in the following manner: Vehicle radar data is combined with GPS positioning information to convert road markers and lane numbers; The ground-based fixed radar outputs the two-dimensional plane coordinates of the disease relative to the monitoring point.

6. The method according to claim 1, characterized in that, The final output includes disease type, location information, and overall confidence level, which is calculated from the following factors: The confidence scores of each sub-model after information entropy optimization; Dynamic weight allocation results; The maximum confidence score in the fusion probability distribution.

Images