Road surface disease identification method based on arrayed intelligent particles and multi-feature fusion

Abstract

The invention relates to the technical field of road engineering and structural health monitoring, and discloses a pavement disease identification method based on arrayed intelligent particles and multi-feature fusion. The method comprises the following steps: acquiring and constructing a five-dimensional data cube of'node-horizon-lateral offset-time-channel 'through intelligent particle sensing arrays arranged at multiple horizon and lateral positions in a pavement structure; on the basis of the data cube, multi-dimensional spatial and temporal characteristics such as cross-layer mechanical difference, transverse pressure gradient and humidity-stress time-delay correlation are extracted and input into a machine learning model for fusion analysis, and identification results of disease types such as ruts, cracks and water seepage, spatial positions and risk levels are output. Furthermore, the five-dimensional data cube is directly input into a time-space unified model based on a Transform architecture, disease sensitive features are automatically mined through a self-attention mechanism, high-precision intelligent diagnosis of early and hidden diseases is realized, and limitation of traditional single-point monitoring is overcome.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of road engineering and structural health monitoring, and particularly relates to a pavement internal disease precise identification and positioning method based on arrayed intelligent particle sensor network and multi-dimensional space-time feature fusion. BACKGROUND

[0002] As the main form of highway pavement, the internal structure of asphalt pavement is prone to typical diseases such as rutting, cracking, loosening and water damage under the long-term and repeated coupling action of driving load and natural environment. These diseases are usually hidden in the internal structure of the pavement at the initial stage of their occurrence and development, and are difficult to be effectively discovered by traditional manual inspection, appearance photography or surface ground penetrating radar detection technology. However, when these internal diseases develop to the surface of the pavement, they have caused irreversible damage to the structure layer, not only greatly increasing the difficulty and cost of repair work, but also posing a potential threat to driving safety.

[0003] Currently, embedded sensing technology provides an effective means for directly obtaining internal state information of the pavement. For example, some technical solutions use a single or a few stress, strain or humidity sensors embedded in the pavement to monitor the changes of physical parameters at local positions. However, such existing technologies mainly rely on monitoring data of single points or sparse points, and their monitoring angle is limited, which is essentially a "point-to-area" sensing mode. This mode has obvious limitations: first, it is difficult to capture the evolution process and distribution pattern of diseases in the spatial dimension; second, it cannot effectively distinguish between local material abnormalities and distributed diseases with structure; third, it is difficult to reveal the correlation between mechanical responses of different structure layers and the causal time sequence relationship between environmental factors and mechanical responses. This leads to low recognition sensitivity, poor spatial positioning accuracy and inaccurate disease type judgment for early and hidden diseases. Further, even if an embedded sensor array is used, most current methods still rely on manually designed features (such as stress gradient, interlayer difference, etc.) and traditional machine learning models for diagnosis. Although this method has certain effect, its diagnosis ability is severely dependent on expert prior knowledge, and the designed features are often subjective and have weak generalization ability, and it is difficult to fully exploit the complex nonlinear coupling relationships in high-dimensional space-time data, such as the dynamic time lag correlation between multiple physical fields (stress, temperature, humidity) and other deep patterns. This limits the sensitive perception and accurate identification ability of early and hidden diseases.

[0004] Therefore, there is an urgent need in the art for a pavement internal disease diagnosis method that can perform collaborative and comprehensive perception in spatial and temporal dimensions, and can deeply fuse multi-source information, in order to achieve early warning, precise identification and intelligent evaluation of diseases. SUMMARY

[0005] The present application aims at the problems that existing internal diseases of asphalt pavement are difficult to be identified early, the spatial positioning precision is insufficient, the traditional single-point sensing method is difficult to capture the coupling relationship between structural layers, and the existing method based on feature engineering relies on prior knowledge and has limited feature expression capability, and provides a pavement disease identification method based on arrayed intelligent particles and spatiotemporal feature intelligent learning. The method constructs a five-dimensional data cube which can reflect the spatiotemporal evolution law of the internal multi-layer, multi-lateral position and multi-channel physical quantity of the pavement structure, and based on the data cube, through feature engineering and machine learning fusion, or through an end-to-end deep learning method, the type identification, spatial positioning and risk level assessment of multiple types of pavement diseases such as rut, crack and water seepage are realized. The present application overcomes the limitations of traditional point monitoring such as narrow view, single data dimension and insensitivity to hidden diseases, and significantly improves the intelligent diagnosis capability of early diseases of pavement structure.

[0006] The specific scheme is as follows:

[0007] In the first aspect, the present application provides a pavement disease identification method based on arrayed intelligent particles and multi-feature fusion, comprising the following steps:

[0008] S1, collecting time series monitoring data through an intelligent particle sensor array arranged at multiple layers and lateral positions in the pavement structure, wherein the time series monitoring data at least includes triaxial stress, triaxial attitude, temperature and humidity;

[0009] S2, fusing the time series monitoring data and the spatial position information of the intelligent particle sensor array to construct a five-dimensional data cube of node-layer-lateral offset-time-channel; the five-dimensional structure comprehensively describes the spatial distribution and time evolution of the internal mechanical response of the pavement, and provides a high-dimensional data basis for disease identification;

[0010] S3, obtaining spatiotemporal features for disease identification based on the five-dimensional data cube;

[0011] S4, inputting the extracted spatiotemporal features into a pre-trained machine learning model to output comprehensive identification results including disease type, spatial position and risk level.

[0012] Further, in step S1, the intelligent particle sensor array is composed of multiple intelligent particle sensor nodes in an array, which is arranged in a grid-like manner under different structural layers in the pavement wheel track, and at least two sensor nodes are arranged along the lateral direction in the same layer. To form a multi-dimensional array layout with longitudinal, lateral and layer resolution.

[0013] Further, in step S3, the spatiotemporal features at least include cross-layer difference features, lateral gradient features and humidity-stress time lag correlation features.

[0014] The cross-layer difference feature is a difference in physical quantity of nodes at the same lateral position and different structural layer positions at the same time, and is used to identify diseases such as rutting with interlayer mechanical imbalance characteristics; the lateral gradient feature is a variation rate of physical quantity measured by sensor nodes at the same structural layer position and different lateral offset positions at the same time along the lateral direction, and is used to identify spatial non-uniformity diseases such as transverse cracks and looseness; and the humidity-stress time lag correlation feature is a maximum cross-correlation coefficient and a corresponding time lag value between a humidity data sequence and a stress data sequence within a specific time window, and is used to reflect the coupling relationship between'seepage-in-lag-structure response' in water damage.

[0015] The above multi-dimensional spatio-temporal features jointly constitute key inputs for disease identification, and realize deep coupling representation of structural mechanics, environmental changes and time sequence correlation.

[0016] Further, in step S4, the machine learning model includes an ensemble learning model based on gradient boosting decision trees or a convolutional neural network model, which can identify typical feature combination patterns of different diseases through multi-dimensional feature fusion learning, and realize high-precision identification and positioning of early and hidden diseases in the pavement.

[0017] Further, after obtaining the five-dimensional data cube in step S2, the five-dimensional data cube can be directly input into an end-to-end deep learning model, and the model automatically learns spatio-temporal features; wherein the end-to-end deep learning model is a sequence model based on a Transformer architecture. The model automatically mines deep features such as cross-layer, lateral and multi-physical field time sequence coupling in the data through its internal self-attention mechanism and four-dimensional position encoding, and directly generates identification results of disease types, spatial positions and risk levels through a multi-task output head. This method avoids the subjectivity and information loss of manual feature design, can learn more complex disease patterns, and further improves the identification accuracy and generalization ability of early and hidden diseases.

[0018] Further, in step S4, the disease types include rutting, cracking, looseness and water seepage; and the risk level is used to represent the severity and development urgency of the disease.

[0019] In a second aspect, the present application provides a pavement disease identification system based on an arrayed intelligent particle and multi-feature fusion, which is used to realize the above method, and includes:

[0020] An intelligent particle sensor array is arranged in the pavement structure to collect time sequence monitoring data;

[0021] A data processing unit is used to construct a five-dimensional data cube and obtain feature information for disease identification;

[0022] The intelligent diagnosis unit is internally provided with a pre-trained machine learning model, and is used for generating a disease identification result according to the feature information.

[0023] Compared with the prior art, the present application has the following remarkable advantages:

[0024] (1) Strong multi-dimensional monitoring capability: by constructing a five-dimensional data cube, time series data and spatial distribution information are fused, and the internal structure state of the pavement can be presented in all directions;

[0025] (2) High disease identification accuracy: multi-feature fusion can reflect the layer characteristics, lateral non-uniformity and environmental coupling relationship at the same time, improve the model discrimination ability, and especially the end-to-end learning mode can automatically explore deep sensitive features beyond artificial design, and the accuracy is higher;

[0026] (3) Early diagnosis can be realized: the humidity-stress time lag feature can sensitively capture the initiation process of water damage and give early warning;

[0027] (4) Strong spatial positioning capability: relying on the arrayed arrangement mode, the present application can accurately locate the disease position in the layer and the lateral direction;

[0028] (5) Flexible technical path and high degree of intelligence: the method is not only suitable for typical pavement diseases such as rut, crack, loose, water seepage, etc., but also provides a complete technical spectrum from traditional feature fusion to advanced end-to-end learning, which can adapt to different application scenarios and data conditions, and the degree of intelligence is significantly improved.

[0029] The method provided by the present application has comprehensive spatial coverage capability, rich data dimension expression capability and excellent intelligent identification performance, and can be used as an important technical path for monitoring the health of the whole life cycle of the pavement structure, and also provides a stable data basis and decision basis for the road asset management and intelligent maintenance system. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 The flowchart of the method of the present application. DETAILED DESCRIPTION

[0031] In the following description, specific values of the number of RIS units are proposed for explanation rather than for limitation, so as to thoroughly understand the present application. However, it should be clear to those skilled in the art that the present application can also be implemented in other embodiments without these specific details.

[0032] This embodiment: end-to-end identification method based on space-time unified Transformer

[0033] In the laboratory phase, intelligent particle sensors were first embedded within the experimental road surface structure according to an array-based deployment scheme identical to that of actual roads. Monitoring sections of the sensors were set up along the driving direction at intervals of 2-5 meters, with each section deployed across at least two structural layers. Within the same layer, sensors maintained a lateral spacing of 0.5-1.5 meters, forming a standardized three-dimensional monitoring network.

[0034] By simulating normal driving loads and typical defect formation processes using loading equipment, intelligent particle sensors simultaneously collect time-series data such as triaxial stress, triaxial attitude, temperature, and humidity. The data processing unit applies Kalman filtering to denoise the raw data and constructs a five-dimensional data cube consisting of nodes, layers, lateral offsets, time, and channels. .

[0035] Based on this five-dimensional data cube, this embodiment employs an end-to-end identification model based on a spatiotemporal unified Transformer to directly identify diseases without the need for manual design or feature extraction. The specific steps are as follows:

[0036] 1. Input Representation and Embedding Layer

[0037] The five-dimensional data cube consisting of nodes, layers, lateral offsets, time, and channels is denoted as...

[0038]

[0039] in, To monitor the number of nodes; Number of structural layers; This represents the number of horizontal positions. The length of the time series; The number of channels (three-axis stress, three-axis attitude, temperature, humidity).

[0040] Flatten the first four dimensions to a length of...

[0041]

[0042] The spatiotemporal sequence is obtained

[0043]

[0044] The original physical quantities are mapped to a dimension of through linear projection (fully connected layer). The high-dimensional feature space is superimposed with spatiotemporal location encoding to obtain the initial input representation:

[0045]

[0046] in, , , It is a four-dimensional spatiotemporal location encoding.

[0047] 2. Four-dimensional spatiotemporal position encoding mechanism

[0048] To explicitly depict the geometric and temporal position information in the four dimensions of "node—horizon—lateral—time", the embodiment constructs four types of position encodings of node, horizon, lateral and time, and adds them to obtain the total position encoding:

[0049]

[0050] wherein, , , , are the position encodings of node, horizon, lateral and time respectively;

[0051] Taking the time dimension position encoding as an example, let the integer index of a certain sequence position in the time dimension be , then its encoding in the feature dimension , is respectively:

[0052]

[0053]

[0054] The node position encoding, horizon position encoding and lateral position encoding are respectively constructed in the same form according to the node index , horizon index and lateral index . When the index corresponding to a spatiotemporal point is , the total position encoding is the element-wise sum of the four types of encodings, thereby giving each "node—horizon—lateral—time" data point a unique spatiotemporal identity.

[0055] 3. Hierarchical spatiotemporal attention working mechanism

[0056] The Transformer main body is stacked by several layers, and each layer includes a multi-head self-attention sublayer and a feedforward network sublayer. For the hth attention head of the lth layer, the input is the output of the last layer , and the query, key and value are obtained through linear mapping:

[0057]

[0058] wherein . The basic form of self-attention is:

[0059]

[0060] To simultaneously depict local spatial correlation and temporal causality, the embodiment introduces a spatial local mask matrix M_local and a temporal causal mask matrix M_causal in the above attention calculation to obtain spatio-temporal constraint attention weights:

[0061]

[0062]

[0063] wherein:

[0064] For limiting the attention strength between non-physically adjacent nodes, a sufficiently large negative value (approximately regarded as -∞) is given at the corresponding position for the spatio-temporal points not belonging to the same monitoring section or exceeding the preset threshold, so as to only allow strong connection between physically adjacent sensor nodes, highlighting the local damage pattern;

[0065] For ensuring the causality in the time dimension, a negative infinity is also given to the element of "paying attention to the future time at the current time", so as to only allow paying attention to the historical time at the current time, and accurately capture the time-dependent relationship such as "humidity penetration→stress response" which has a lag feature.

[0066] The output of the multi-head attention is:

[0067]

[0068] wherein is the number of attention heads.

[0069] Then, the nonlinear features are further refined by a feedforward network with residual connection and layer normalization to form the output of the first layer . After the layer Transformer stack, the final spatio-temporal feature representation is obtained.

[0070] 4. Intrinsic learning mechanism of cross-layer, lateral and multi-physical field features

[0071] Under the above end-to-end network structure, the embodiment does not need to explicitly construct "cross-layer difference", "lateral gradient", "humidity-stress time lag correlation" and other manual features, but automatically learns the corresponding physical patterns by the model through attention weights:

[0072] Cross-layer mechanical transmission features: for nodes at the same lateral position and different structural layers, their mutual attention weights can reflect the mechanical response difference between the upper and lower layers, which is equivalent to automatically learning the "difference response" between different layers;

[0073] Transverse damage distribution feature: In the same layer, the attention connection between nodes along the transverse direction can depict the transverse variation gradient of the physical quantity, so as to capture the spatial non-uniformity mode of rut, transverse crack, local looseness and other diseases;

[0074] Multi-physical field time sequence coupling feature: based on the time causal mask, the attention mechanism can automatically mine the high correlation response mode under different time lags from the multi-channel sequence of humidity, temperature and stress, and posture, implicitly learn the "humidity-stress time lag correlation" and the corresponding lag scale.

[0075] Therefore, the Transformer model automatically completes the extraction and fusion of multi-dimensional features in a unified spatio-temporal representation space, avoiding the information loss and subjective bias that may be introduced by traditional manual feature engineering.

[0076] 5. Global feature aggregation and multi-task output head design

[0077] In order to output disease type, spatial position and risk level at the same time, a multi-task output layer is set after the Transformer main body. Firstly, the final layer output is globally aggregated to obtain the global feature vector of the pavement state. Preferably, the global aggregation can adopt the average pooling method:

[0078]

[0079] On this basis, three types of output heads are constructed:

[0080] (1) Disease type identification head

[0081]

[0082] Among them, is the probability distribution of rut, crack, looseness, seepage and other disease types, , are trainable parameters.

[0083] (2) Spatial position regression head

[0084]

[0085] Among them, , , respectively represent the estimated position of the disease center in the layer, transverse offset and depth direction, , are trainable parameters.

[0086] (3) Risk level evaluation head

[0087]

[0088] wherein may be a continuous risk score, or quantified into discrete risk levels in the post-processing stage, , are trainable parameters.

[0089] 6. End-to-end multi-task training and loss function

[0090] This embodiment adopts an end-to-end method to train the Transformer model. Assuming that the true disease type label is (encoded by one-hot), the true spatial position is , and the true risk score or level is , the overall loss function is defined as the weighted sum of multiple tasks:

[0091]

[0092] wherein:

[0093] Classification loss (cross-entropy)

[0094]

[0095] Position regression loss (preferably Smooth L1 loss)

[0096]

[0097] Risk assessment loss (preferably mean square error)

[0098]

[0099] , , are the weight coefficients of each sub-task loss, which can be optimized through experiments to balance the disease type recognition accuracy, spatial positioning accuracy, and risk assessment reliability.

[0100] In the training process, a laboratory calibration dataset containing multiple typical disease working conditions is used to iteratively update the above parameters and the corresponding bias until the multi-task loss function converges and meets the preset recognition accuracy requirement.

[0101] After completing the laboratory training, the verified complete system is directly deployed to the actual road. In the actual road surface track area, a standardized monitoring network is established using the same array layout scheme as in the laboratory. This ensures that the spatial structure of the input data when the model is applied in the field is consistent with the training stage.

[0102] During system operation, the field sensor array continuously collects data, constructs a five-dimensional data cube through the same preprocessing process as the laboratory stage, and directly inputs the trained Transformer model for real-time end-to-end identification. The final output of disease type, location and risk level information is transmitted to the management platform through wireless transmission to provide accurate decision support for road maintenance.

[0103] The above merely illustrates the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical range disclosed by the present application, which shall be covered within the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the protection scope of the claims.

Claims

1. A method for identifying pavement defects based on arrayed intelligent particles and multi-feature fusion, characterized in that, Includes the following steps: S1. Collect time-series monitoring data by deploying an array of intelligent particle sensors at multiple layers and lateral positions inside the road structure. The time-series monitoring data includes at least triaxial stress, triaxial attitude, temperature and humidity. S2. Integrate the time-series monitoring data with the spatial location information of the intelligent particle sensor array to construct a five-dimensional data cube of node-layer-lateral offset-time-channel; S3. Based on the five-dimensional data cube, obtain the spatiotemporal features for disease identification; S4. Input the extracted spatiotemporal features into the pre-trained machine learning model and output a comprehensive identification result including disease type, spatial location and risk level.

2. The method according to claim 1, characterized in that, In step S1, the intelligent particle sensor array is composed of multiple intelligent particle sensor nodes arranged in an array, which are arranged in a grid pattern in different structural layers below the road wheel track, and at least two sensor nodes are arranged laterally in the same layer.

3. The method according to claim 1, characterized in that, In step S3, the spatiotemporal features include at least cross-layer differential features, lateral gradient features, and humidity-stress time delay correlation features.

4. The method according to claim 3, characterized in that, The cross-layer differential feature reflects the difference in physical quantities at nodes in the same lateral position but different structural layers at the same time, and is used to identify defects such as ruts that have inter-layer mechanical imbalance characteristics; the lateral gradient feature refers to the rate of change of physical quantities measured by sensor nodes in the same structural layer but at different lateral offset positions at the same time along the lateral direction, and is used to identify spatial non-uniform defects such as lateral cracks and loosening; the humidity-stress time lag correlation feature refers to the maximum cross-correlation coefficient and the corresponding time lag value between the humidity data sequence and the stress data sequence within a specific time window, and is used to reflect the coupling relationship between "infiltration-hysteresis-structural response" in water damage.

5. The method according to claim 1, characterized in that, In step S4, the machine learning model includes an ensemble learning model based on gradient boosting decision trees or a convolutional neural network model.

6. The method according to claim 1, characterized in that, After obtaining the five-dimensional data cube from step S2, the five-dimensional data cube is directly input into the end-to-end deep learning model. The model automatically learns the spatiotemporal features and directly generates the final identification results of disease type, spatial location, and risk level through the multi-task output head. The end-to-end deep learning model is a sequence model based on the Transformer architecture.

7. The method according to claim 1, characterized in that, In step S4, the types of defects include ruts, cracks, loosening, and seepage; the risk level is used to characterize the severity and urgency of the defects.

8. A pavement defect identification system based on arrayed intelligent particles and multi-feature fusion, used to implement the method described in any one of claims 1-7, characterized in that, include: A smart particle sensor array is deployed inside the road surface structure to collect time-series monitoring data; The data processing unit is used to construct a five-dimensional data cube and obtain feature information for disease identification. The intelligent diagnostic unit has a built-in pre-trained machine learning model, which is used to generate disease identification results based on the feature information.

Images