Paving control method and system based on artificial intelligence ai architecture

Abstract

The application provides a paving control method and system based on an artificial intelligence architecture, which comprises the following steps: determining domain construction attributes based on a construction reference model and a positioning uncertainty covariance matrix generated from satellite raw observation data, and then determining a positioning failure risk field in the cloud; weighting the spatial scope of the positioning uncertainty covariance matrix according to the reliability to obtain a reliable simulation state in the cloud; inputting the reliable simulation state, the domain construction attributes and the positioning failure risk field into an artificial intelligence optimization model to generate a pre-correction parameter set, and then downloading the pre-correction parameter set and the positioning failure risk field to an edge terminal. The edge terminal determines satellite positioning quality parameters, and pre-starts a laser three-dimensional imaging device and a millimeter wave radar to obtain pose estimation data before entering the positioning failure risk area; when the positioning quality meets the judgment rule, the pre-correction parameter set is injected according to the global optimization mode, and when the positioning quality does not meet the judgment rule, the local material sensing mode is switched to and a local correction instruction bit is generated based on the relative pose information.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction, and in particular to a paving control method and system based on an artificial intelligence (AI) architecture. Background Technology

[0002] Currently, asphalt pavement paving control largely relies on satellite positioning, conventional sensing, and classical control algorithms to adjust construction parameters. This makes it difficult to accurately characterize the non-equilibrium evolution of asphalt mixtures, non-Gaussian migration of aggregates, and viscoelastic-plastic deformation of the structure throughout the paving process. Existing modeling methods typically use fixed spatial domains without dynamic constraints based on satellite positioning reliability. This easily introduces simulation biases when satellite signals deteriorate, affecting control reliability. Furthermore, current paving control largely depends on passive feedback correction after deviations occur, failing to address differences in positioning reliability across different paving sub-regions and lacking proactive detection of positioning failure zones. This results in delayed control mode switching when satellite positioning quality fluctuates. In addition, paving control lacks a relative pose perception and adaptive switching mechanism independent of satellite signals in satellite positioning failure scenarios, failing to guarantee the continuity and control accuracy of paving operations under complex conditions and failing to meet the needs for refined and intelligent control of the paving process. Summary of the Invention

[0003] The first aspect of this application provides a paving control method based on an artificial intelligence architecture, applied to a paving control system. The paving control system includes a cloud platform, an edge terminal of the paver, a vehicle-mounted satellite receiver, a laser 3D imaging device, a millimeter-wave radar, and a paver motion controller. The method includes: the cloud platform determining the domain construction attributes corresponding to different paving sub-regions in the paving area based on a construction benchmark model of the paving area and a positioning uncertainty covariance matrix generated from raw satellite observation data from the vehicle-mounted satellite receiver; the cloud platform determining a positioning failure risk field based on the positioning uncertainty covariance matrix and the domain construction attributes; and the cloud platform... A fractional-order spatiotemporal coupled constitutive model is established on the construction benchmark model, and the spatial domain of the fractional-order spatiotemporal coupled constitutive model is weighted with credibility according to the positioning uncertainty covariance matrix to obtain the credible simulation state of each paving sub-region; the cloud inputs the credible simulation state, the sub-domain construction attributes, and the positioning failure risk field into the artificial intelligence optimization model to jointly optimize the paving speed curve, tamping frequency, and / or auger speed gradient, generate a pre-correction parameter set, and sends the positioning failure risk field and the pre-correction parameter set to the edge terminal; the edge terminal acquires the real-time satellite original output by the vehicle-mounted satellite receiver during the paving operation. The system begins by observing data and determining satellite positioning quality parameters based on the real-time satellite raw observation data. It then determines whether the paver will enter the positioning failure risk zone based on the positioning failure risk field and the paver's real-time positioning information. When it is determined that the paver will enter the positioning failure risk zone, the edge device pre-activates the laser three-dimensional imaging device and the millimeter-wave radar before entering the risk zone to obtain pose pre-estimation data between the paver and the paved surface. When the satellite positioning quality parameters meet the positioning quality discrimination rules, the edge device injects the pre-correction parameter set into the paver motion controller according to a global optimization mode to control... The paving speed curve and / or tamping frequency; when the satellite positioning quality parameters do not meet the positioning quality discrimination rules, the edge end switches from the global optimization mode to the local material sensing mode, stops using satellite positioning coordinates as paving position control input, and generates a local correction command based on the pose pre-estimation data and the relative pose information obtained by the laser three-dimensional imaging device and the millimeter-wave radar in the local material sensing mode; the edge end injects the local correction command into the paver motion controller to adjust the paver's working parameters; the working parameters include at least one of the following: auger speed gradient, screed inclination angle, and overlap zone vibration phase.

[0004] The second aspect provides a paving control system based on an artificial intelligence architecture, including a cloud platform, an edge device of the paver, an onboard satellite receiver, a laser 3D imaging device, a millimeter-wave radar, and a paver motion controller. The cloud platform is used to determine the domain-specific construction attributes corresponding to different paving sub-regions within the paving area based on a construction benchmark model of the paving area and a positioning uncertainty covariance matrix generated from raw satellite observation data from the onboard satellite receiver. Based on the positioning uncertainty covariance matrix and the domain-specific construction attributes, a positioning failure risk field is determined. A fractional-order spatiotemporal coupled constitutive model is established based on the construction benchmark model. The spatial domain of the fractional-order spatiotemporal coupled constitutive model is weighted with a confidence level according to the positioning uncertainty covariance matrix to obtain the credible simulation state of each paving sub-region. The credible simulation state, the sub-domain construction attributes, and the positioning failure risk field are input into an artificial intelligence optimization model to jointly optimize the paving speed curve, tamper frequency, and / or auger speed gradient, generating a pre-correction parameter set. The positioning failure risk field and the pre-correction parameter set are then sent to the edge terminal. The edge terminal is used to acquire real-time raw satellite observation data output by the vehicle-mounted satellite receiver during the paving operation, based on the... Real-time satellite raw observation data determines satellite positioning quality parameters, and based on the positioning failure risk field and the paver's real-time positioning information, it is determined whether the paver will enter the positioning failure risk zone. When it is determined that the paver will enter the positioning failure risk zone, the laser three-dimensional imaging device and the millimeter-wave radar are pre-activated before entering the positioning failure risk zone to obtain pose pre-estimation data between the paver and the paved surface. When the satellite positioning quality parameters meet the positioning quality discrimination rules, the pre-correction parameter set is injected into the paver motion controller according to the global optimization mode to control the paving speed curve. Line and / or tamping hammer frequency; when the satellite positioning quality parameters do not meet the positioning quality discrimination rules, switch from the global optimization mode to the local material sensing mode, stop using satellite positioning coordinates as paving position control input, and generate a local correction command based on the pose pre-estimation data and the relative pose information obtained by the laser three-dimensional imaging device and the millimeter-wave radar in the local material sensing mode; inject the local correction command into the paver motion controller to adjust the paver's working parameters; the working parameters include at least one of the following: auger speed gradient, screed inclination angle, and overlap zone vibration phase.

[0005] A third aspect provides a computer-readable storage medium storing computer-executable instructions; the computer-executable instructions, when executed by a processor, can implement the method of the first aspect described above.

[0006] The beneficial effects of this invention are as follows: The solution in this application uses a cloud-based construction benchmark model and a positioning uncertainty covariance matrix generated from raw satellite observation data to determine the domain construction attributes. This allows the paving area to be characterized by a combination of geometric, thermal, material properties, and positioning reliability. This breaks through the limitations of traditional paving modeling, which only focuses on spatial geometry and design parameters, and lays the foundation for precise domain control.

[0007] The solution in this application establishes a fractional-order spatiotemporal coupled constitutive model based on the construction benchmark model, and performs credibility weighting on the spatial domain of the constitutive model according to the positioning uncertainty covariance matrix to obtain the credible simulation state of each paving sub-region. This can suppress the pollution of the global simulation by low-credibility regions when the positioning uncertainty increases, solve the problem that the traditional fixed modeling domain is easily affected by signal degradation, and improve the robustness of the model in complex environments.

[0008] The solution in this application integrates the credible simulation state, the domain construction attributes, and the location failure risk field into the artificial intelligence optimization model for joint optimization to generate a pre-correction parameter set. This enables the control strategy to simultaneously satisfy physical constraints and multi-objective optimization criteria, overcoming the limitation of traditional single control algorithms in their insufficient optimization capabilities under complex constraints.

[0009] The solution in this application embodiment determines the positioning failure risk field based on the positioning uncertainty covariance matrix and the domain construction attributes in the cloud. Based on this, the edge device pre-starts the laser three-dimensional imaging device and millimeter-wave radar to obtain the pose pre-estimation data before the paver enters the positioning failure risk area. This realizes the forward identification of positioning failure risk and the early readiness of the relative pose perception channel, avoiding the time delay caused by the cold start of the sensing channel.

[0010] The solution in this application embodiment involves the edge terminal to self-correct and switch between global optimization mode and local material sensing mode based on satellite positioning quality parameters. When the satellite positioning quality is good, the pre-correction parameter set is injected according to the global optimization mode to control the paving speed curve and tamping frequency. When the satellite positioning quality deteriorates, the solution switches to local material sensing mode and generates local correction commands based on relative pose information, which effectively solves the problem of control interruption or accuracy reduction when the satellite positioning is unstable.

[0011] The solution in this application embodiment stops using satellite positioning coordinates as the paving position control input in the local material sensing mode. Instead, it adjusts the auger speed gradient, screed inclination angle and vibration phase of the overlap area based on the pose pre-estimation data and relative pose information. Even when satellite positioning fails, the paving operation can be maintained, which greatly improves the adaptability and continuous operation capability of the paving control system under complex working conditions. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below. The accompanying 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 content of the embodiments of the present invention and these drawings.

[0013] Figure 1 This is a field schematic diagram of a paving control method based on an artificial intelligence (AI) architecture provided in an embodiment of this disclosure; Figure 2 This is a flowchart illustrating another paving control method based on an artificial intelligence (AI) architecture provided in this disclosure embodiment; Figure 3 This is a schematic diagram of another paving control system based on an artificial intelligence (AI) architecture provided in this disclosure embodiment; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0014] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the application and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only partial structures relevant to the present application, not the complete structure.

[0015] This application provides a paving control method based on an artificial intelligence architecture, which is applied to a paving control system. The paving control system includes a cloud platform, an edge device located on the paver, a vehicle-mounted satellite receiver, a laser three-dimensional imaging device, a millimeter-wave radar, and a paver motion controller.

[0016] The cloud platform can be a remote server cluster with computing, storage, and networking capabilities, a local server deployed in the construction command center, or a distributed computing platform composed of several server nodes. The cloud platform undertakes global, computationally intensive tasks with relatively low real-time requirements, such as global physical simulation, multi-objective optimization, and model updates.

[0017] The edge device is a computing unit deployed on the paver, close to the data source. The edge device can be at least one of an industrial control computer, an embedded computing module, or a vehicle-mounted domain controller. The edge device is responsible for real-time data acquisition, preliminary processing, and issuing low-latency control commands. The cloud and the edge device interact via a wireless communication network, which can be at least one of a fifth-generation mobile communication network, a fourth-generation mobile communication network, or a wireless local area network.

[0018] The vehicle-mounted satellite receiver is installed on the paver and is used to receive satellite signals and output raw satellite observation data. The raw satellite observation data includes at least one of carrier phase, pseudorange, signal-to-noise ratio, and position dilution of precision (PDOP). The satellite can be a satellite of the BeiDou Navigation Satellite System, or a satellite of the Global Positioning System, GLONASS, or Galileo system, or a combination of satellites from multiple systems.

[0019] The laser 3D imaging device is installed at the rear or side of the paver to collect point cloud data of the paved surface. The laser 3D imaging device can be at least one of a laser structured light imaging device, a line laser scanning device, or a surface-array lidar. The millimeter-wave radar is installed on the paver to transmit millimeter waves to the paved surface and overlap areas and receive echo data. The millimeter-wave radar can operate in at least one of the 24 GHz, 77 GHz, or 79 GHz frequency bands. The laser 3D imaging device and the millimeter-wave radar together constitute a relative pose sensing channel independent of satellite signals.

[0020] The paver motion controller is connected to the paver's travel drive mechanism, tamping mechanism, auger spreading mechanism, and screed mechanism, and is used to receive control commands and drive the aforementioned mechanisms to operate. The control commands received by the paver motion controller include at least one corresponding control quantity among the following: paving speed curve, tamping frequency, auger speed gradient, screed inclination angle, and vibration phase of the overlap area.

[0021] The signal flow of the above-mentioned components can be understood as follows: the output data of the vehicle-mounted satellite receiver, the laser three-dimensional imaging device, and the millimeter-wave radar converge at the edge terminal; the edge terminal and the cloud terminal perform bidirectional data interaction; and the edge terminal sends control commands to the paver motion controller. This application embodiment does not limit the specific installation location of the above-mentioned components; any arrangement that can achieve the above-mentioned data acquisition, data interaction, and command issuance functions can be adopted.

[0022] For ease of understanding, several terms involved in the embodiments of this application will be explained below. The explanation of the terms does not constitute a limitation on the scope of protection of the claims.

[0023] A paving sub-area refers to a local area obtained by dividing the paving area spatially. In some embodiments, the entire paving surface is divided into multiple grid-like spatial sub-areas, with each grid corresponding to a paving sub-area. The grid size can be one meter by one meter, or 0.5 meters by 0.5 meters, or two meters by two meters. This application does not limit the grid size.

[0024] The positioning uncertainty covariance matrix is ​​used to quantify the error distribution of satellite positioning results and the correlation between error components. In some embodiments, the positioning uncertainty covariance matrix is ​​a third-order square matrix, whose diagonal elements represent the position variances in the east, north, and elevation directions, respectively, and whose off-diagonal elements represent the error correlations between these directions. The positioning uncertainty covariance matrix can be obtained from the raw satellite observation data of the vehicle-mounted satellite receiver through least-squares estimation or Kalman filtering, or it can be obtained by combining inertial navigation data through combined filtering.

[0025] Domain construction attributes refer to the category labels that different paving sub-areas possess in terms of geometry, thermal response, material migration, and positioning reliability. In some embodiments, the domain construction attributes include at least two of the following: conventional paving domains, variable slope curve domains, overlapping joint domains, and structural boundary domains. Conventional paving domains refer to areas with gentle slope changes and relatively stable construction conditions; variable slope curve domains refer to areas with significant changes in slope or direction, such as uphill / downhill sections and sharp curves; overlapping joint domains refer to the connection areas between different paving widths or different construction sections; and structural boundary domains refer to areas adjacent to fixed structures such as bridges and tunnel entrances.

[0026] A positioning failure risk field refers to a field quantity continuously distributed along the paver's planned travel path, characterizing the probability of satellite positioning failure at each location. In some embodiments, the positioning failure risk field can be represented as a geographic information system layer, where each location in the layer corresponds to a risk value or risk level.

[0027] A credible simulation state refers to the simulation result obtained by weighting the spatial domain of a fractional-order spatiotemporally coupled constitutive model according to its location credibility. In some embodiments, the credible simulation state includes at least one of the predicted temperature, predicted segregation index, and predicted deformation of each paving sub-region over a future period, and the corresponding credibility weights.

[0028] The pre-calibration parameter set refers to a set of control parameters pre-calculated by the cloud before the paving operation starts, used for controlling the paver in global optimization mode. In some embodiments, the pre-calibration parameter set includes at least one of paving speed curves, tamping hammer frequency sequences, and auger speed gradient sequences corresponding to different domain construction attributes.

[0029] Position and pose pre-estimation data refers to data characterizing the paver's position and attitude relative to the paved surface, pre-collected and fused by the laser 3D imaging device and the millimeter-wave radar before the satellite positioning signal deteriorates to the point of triggering a mode switch. Local correction commands refer to control commands generated based on relative pose information in local material sensing mode, used to adjust the paver's operating parameters.

[0030] A Spatio-Temporal Graph Neural Network (STGNN) is a network that adds temporal dimension modeling capabilities to a graph neural network. A Deep Temporal Optimization Network (DTONet) is an end-to-end deep network for temporal optimization of control parameters. The STGNN and the DTONet together constitute the artificial intelligence optimization model in this embodiment.

[0031] Combination Figure 2 The paving control method in this application adopts a cloud-edge collaborative execution approach. Before the paving operation starts, the cloud loads a construction baseline model and a fractional-order spatiotemporally coupled constitutive model, performs global physical simulation and multi-objective optimization, and generates a pre-correction parameter set, a location failure risk field, and an error source map. During the paving operation, the edge device collects real-time data and performs low-latency control, while simultaneously transmitting paving operation data back to the cloud; the cloud updates the model accordingly and transmits the updated constraints and instructions back. This task allocation method allows complex, computationally intensive tasks to be completed offline in the cloud, while control tasks with high real-time requirements are completed online at the edge, thus achieving a balance between computational efficiency and real-time response.

[0032] It should be noted that the embodiments of this application use a paving control system as the execution environment for the method, and each step of the method is executed by the cloud and the edge, respectively. In the following embodiments, when describing the method steps, it will be specified whether the execution entity for that step is the cloud or the edge. The embodiments of this application do not limit the execution order of steps that are not explicitly defined in sequence.

[0033] This application provides a paving control method based on an artificial intelligence architecture, which is applied to a paving control system. The paving control system includes a cloud platform, an edge device located on the paver, a vehicle-mounted satellite receiver, a laser three-dimensional imaging device, a millimeter-wave radar, and a paver motion controller.

[0034] The cloud can be a remote server cluster with computing, storage, and networking capabilities, a local server deployed in the construction command center, or a distributed computing platform composed of several server nodes. The cloud handles global, computationally intensive tasks with relatively low real-time requirements, such as global physical simulation, multi-objective optimization, and model updates. The edge is a computing unit deployed on the paver, close to the data source. The edge can be at least one of an industrial control computer, an embedded computing module, or a vehicle-mounted domain controller. The edge is responsible for real-time data acquisition, preliminary processing, and issuing low-latency control commands. The cloud and the edge interact via a wireless communication network, which can be at least one of a fifth-generation mobile communication network, a fourth-generation mobile communication network, or a wireless local area network.

[0035] The vehicle-mounted satellite receiver is installed on the paver to receive satellite signals and output raw satellite observation data. The raw satellite observation data includes at least one of carrier phase, pseudorange, signal-to-noise ratio, and position accuracy attenuation factor. The satellite can be a satellite of the BeiDou Navigation Satellite System, or a satellite of the Global Positioning System, GLONASS, or Galileo system, or a combination of satellites from multiple systems. The laser 3D imaging device is installed at the rear or side of the paver to collect point cloud data of the paved surface. The laser 3D imaging device can be at least one of a laser structured light imaging device, a line laser scanning device, or a surface-array lidar. The millimeter-wave radar is installed on the paver to transmit millimeter waves to the paved surface and overlapping areas and receive echo data. The millimeter-wave radar can operate in at least one of the 24 GHz, 77 GHz, or 79 GHz bands. The laser 3D imaging device and the millimeter-wave radar together constitute a relative pose sensing channel independent of satellite signals. The paver motion controller is connected to the paver's travel drive mechanism, tamping mechanism, auger spreading mechanism, and screed mechanism, and is used to receive control commands and drive the aforementioned mechanisms to operate. The control commands received by the paver motion controller include at least one corresponding control quantity among the following: paving speed curve, tamping frequency, auger speed gradient, screed inclination angle, and vibration phase of the overlap area.

[0036] like Figure 2 As shown, the paving control method of this application embodiment includes the following steps.

[0037] S101, the cloud-based system determines the domain construction attributes corresponding to different paving sub-regions in the paving area based on the construction benchmark model of the paving area and the positioning uncertainty covariance matrix generated by the original satellite observation data from the vehicle-mounted satellite receiver.

[0038] Specifically, before the paving operation begins, the cloud-based system determines the domain-specific construction attributes corresponding to different paving sub-areas within the paving area based on a construction benchmark model of the paving area and a positioning uncertainty covariance matrix generated from raw satellite observation data from the vehicle-mounted satellite receiver. The construction benchmark model is a digital model describing the comprehensive characteristics of the paving area.

[0039] In some embodiments, the cloud platform divides the paving area into multiple paving sub-regions and extracts geometric, thermal, and material features from the construction baseline model for each sub-region. The geometric features include at least one of elevation, slope, and curvature; the thermal features include surface temperature and its spatial dispersion; and the material features include the base layer moisture content and its anomaly degree. The positioning uncertainty covariance matrix reflects the reliability of satellite positioning at each paving sub-region.

[0040] In some embodiments, the cloud platform substitutes the raw satellite observation data collected by the vehicle-mounted satellite receiver at various locations in the paving area into the positioning calculation, estimates the corresponding positioning uncertainty covariance matrix for each paving sub-region, and takes the trace or the largest eigenvalue of the matrix as the positioning uncertainty of the paving sub-region. Based on this, the cloud platform can determine the domain construction attributes as follows: for each paving sub-region, its slope curvature characteristics, temperature dispersion characteristics, moisture content anomaly characteristics, and positioning uncertainty are compared with the corresponding discrimination thresholds, and the paving sub-region is labeled as one of the following according to the comparison results: a conventional paving domain, a variable slope curve domain, an overlap joint domain, and a structural boundary domain.

[0041] As an optional implementation, the cloud platform can also input the aforementioned features into a pre-trained classification model, which will then output the domain-specific construction attributes of the paving area. As yet another optional implementation, the cloud platform can combine the road alignment information from the building information model design parameters to directly mark curves and slope-changing sections in the design alignment as slope-changing curve domains, and mark the locations of the lane joints as overlapping joint domains. These three implementation methods can be used individually or in combination.

[0042] Furthermore, in an optional embodiment of this application, the cloud platform also constructs a dynamic migration map of zonal construction attributes based on historical paving data. This dynamic migration map uses paving sub-regions as nodes and the probability of construction attribute transfer between adjacent regions as edges. The cloud platform uses the dynamic migration map to perform forward correction on the zonal construction attributes of the current paving sub-region. Specifically, when a paving sub-region is initially determined to be a regular paving region but its adjacent region is determined to be a slope-curve region, the cloud platform corrects the zonal construction attributes of the region based on the cumulative probability of the region transferring to the slope-curve region in the dynamic migration map. If the cumulative probability exceeds a preset transfer threshold, the region is re-marked as a slope-curve region. This avoids attribute misjudgments caused by solely relying on static feature thresholds for classification and improves the continuity and rationality of zonal construction attributes in the spatial transition zone.

[0043] S102, the cloud determines the positioning failure risk field based on the positioning uncertainty covariance matrix and the domain construction attributes.

[0044] Specifically, the cloud platform determines the positioning failure risk field based on the positioning uncertainty covariance matrix and the domain construction attributes.

[0045] In some embodiments, the cloud platform organizes the positioning uncertainty of each paving sub-area along the planned travel direction of the paver, forming a risk value sequence distributed along the planned travel path. The cloud platform further adjusts the risk value sequence based on the sub-domain construction attributes, assigning greater weight to locations corresponding to sub-domain construction attributes with poor positioning reliability. After the above processing, the cloud platform obtains a positioning failure risk field continuously distributed along the paver's planned travel path. This positioning failure risk field can be represented as a continuous risk value curve or as a risk interval sequence graded by risk level.

[0046] Furthermore, in an optional embodiment of this application, the cloud platform also incorporates historical satellite signal attenuation maps to correct the positioning failure risk field. These historical satellite signal attenuation maps record the statistical characteristics of signal attenuation in the paving area at different times and satellite elevation angles. The cloud platform performs spatiotemporal convolution fusion of the positioning failure risk field and the historical satellite signal attenuation maps to obtain the corrected positioning failure risk field. Specifically, the cloud platform extracts historical signal attenuation statistics corresponding to each location on the paver's planned path within a preset spatiotemporal window. These statistics are then multiplied by a preset attenuation-risk mapping coefficient and superimposed onto the corresponding location in the positioning failure risk field. Thus, the positioning failure risk field not only reflects the theoretical risk based on positioning uncertainty but also incorporates the satellite signal attenuation patterns observed at that location in the actual construction history, making the risk characterization more closely aligned with the actual site conditions.

[0047] S103, the cloud platform establishes a fractional-order spatiotemporal coupled constitutive model based on the construction benchmark model, and performs credibility weighting on the spatial domain of the fractional-order spatiotemporal coupled constitutive model according to the positioning uncertainty covariance matrix to obtain the credible simulation state of each paving sub-region.

[0048] Specifically, the cloud platform establishes a fractional-order spatiotemporal coupled constitutive model based on the construction benchmark model, and weights the spatial domain of the fractional-order spatiotemporal coupled constitutive model with credibility according to the positioning uncertainty covariance matrix to obtain the credible simulation state of each paving sub-region. The fractional-order spatiotemporal coupled constitutive model is used to describe the non-equilibrium evolution behavior of asphalt mixture during the paving process.

[0049] In some embodiments, the fractional-order spatiotemporal coupled constitutive model includes at least one of the following: a Caputo-type fractional-order time derivative model for characterizing the history-dependent effect during heat conduction; a Lévy flight stochastic differential equation model for capturing the non-Gaussian migration characteristics of paving aggregate in paving sections with varying slopes or curves; and a viscoelastic-plastic coupled model for jointly simulating the instantaneous elastic recovery and cumulative plastic deformation of the base-surface layer system under rolling loads. The credibility weighting of the spatial domain can be understood as introducing location-varying weighting factors into the spatial differential operators in the constitutive model.

[0050] As an optional implementation, the cloud calculates a confidence weight w for each paving sub-region. w is negatively correlated with the positioning uncertainty of that paving sub-region; for example, w can be set to one minus the normalized positioning uncertainty, with values ​​ranging from zero to one. When performing spatial discretization of the constitutive equations for a given paving sub-region, the cloud multiplies the state variables of adjacent paving sub-regions participating in the spatial differential operator calculation of that sub-region by their corresponding confidence weights before aggregation. When the positioning uncertainty of an adjacent paving sub-region is large, its confidence weight is small, thereby reducing the impact of this low-confidence region on the simulation results of the current paving sub-region.

[0051] As another optional implementation, the cloud can set a shrinkage rule: when the PDOP of a certain sub-region is greater than a first threshold or the signal-to-noise ratio is less than a second threshold, the neighborhood radius of that sub-region participating in the spatial differential operator calculation is shrunk, retaining only the high-confidence neighboring sub-regions for aggregation. The above two weighting methods can be used individually or in combination. After the above confidence-weighted solution, the cloud obtains the confidence simulation state of each sub-region. The confidence simulation state includes the predicted state quantity of each sub-region and its confidence weight, enabling subsequent optimization to differentiate between simulation results of high-confidence and low-confidence regions.

[0052] Furthermore, in an optional embodiment of this application, when establishing the fractional-order spatiotemporal coupled constitutive model, the cloud platform configures differentiated fractional-order order parameters for different sub-domain construction attributes. Specifically, for the variable-slope curve domain, the cloud platform configures fractional-order parameters lower than the first-order threshold to enhance the model's ability to capture rapid transient responses; for the lap joint domain, the cloud platform configures fractional-order parameters higher than the second-order threshold to enhance the model's memory strength of heat conduction history dependence effects; for the conventional paving domain, the cloud platform configures default order parameters. The cloud platform stores the fractional-order order parameters in association with the corresponding sub-domain construction attributes, and calls the corresponding fractional-order order parameters to solve the constitutive equations according to the sub-domain construction attributes of each paving sub-region during reliable simulation. Thus, the spatial differentiation modeling capability of the fractional-order constitutive model is further enhanced, and regions with different construction attributes can be physically simulated with the most suitable order parameters.

[0053] S104, the cloud inputs the trusted simulation state, the domain construction attributes and the positioning failure risk field into the artificial intelligence optimization model, performs joint optimization on the paving speed curve, tamping frequency and / or auger speed gradient, generates a pre-correction parameter set, and sends the positioning failure risk field and the pre-correction parameter set to the edge terminal.

[0054] Specifically, the cloud inputs the trusted simulation state, the domain construction attributes, and the positioning failure risk field into the artificial intelligence optimization model to jointly optimize the paving speed curve, tamping frequency, and / or auger rotation speed gradient, generate a pre-correction parameter set, and send the positioning failure risk field and the pre-correction parameter set to the edge terminal.

[0055] In some embodiments, after receiving the above input, the artificial intelligence optimization model uses at least one of the following as optimization objectives: smoothness, density, uniformity, energy consumption, and carbon emission intensity. Under the premise of meeting construction specification limits, it searches for at least one of the following: paving speed curve, tamping frequency, and auger speed gradient, which optimizes the objectives. The searched control parameters are then organized into a pre-correction parameter set corresponding to different sub-domain construction attributes. The cloud sends the positioning failure risk field and the pre-correction parameter set to the edge terminal via the wireless communication network. The edge terminal receives and caches these parameters locally for real-time control during the paving operation.

[0056] Furthermore, in an optional embodiment of this application, after the cloud generates the pre-calibration parameter set, it also performs virtual simulation verification of the pre-calibration parameter set in a digital twin environment. Specifically, the cloud constructs a digital twin model of the paving process, inputs the pre-calibration parameter set into the digital twin model for virtual paving simulation, and obtains virtual simulation results; the virtual simulation results include at least one of virtual smoothness, virtual density, and virtual temperature uniformity. The cloud compares the virtual simulation results with the corresponding quality threshold. When the virtual simulation results do not meet the quality threshold, the cloud re-executes the artificial intelligence optimization model to generate an updated pre-calibration parameter set, and performs virtual simulation verification again in the digital twin environment until the virtual simulation results meet the quality threshold. When the virtual simulation results meet the quality threshold, the cloud sends the verified pre-calibration parameter set to the edge terminal. Thus, through virtual verification in the digital twin environment, the quality risks and material waste caused by directly applying unqualified control parameters to on-site construction are avoided.

[0057] S105, the edge end acquires real-time raw satellite observation data output by the vehicle-mounted satellite receiver during the paving operation, determines satellite positioning quality parameters based on the real-time raw satellite observation data, and determines whether the paver will enter the positioning failure risk area based on the positioning failure risk field and the real-time positioning information of the paver.

[0058] Specifically, during the paving operation, the edge terminal acquires real-time raw satellite observation data output by the vehicle-mounted satellite receiver, determines satellite positioning quality parameters based on the real-time raw satellite observation data, and determines whether the paver will enter the positioning failure risk area based on the positioning failure risk field and the paver's real-time positioning information. The satellite positioning quality parameters are used to characterize the reliability of the current satellite positioning signal. In some embodiments, the edge terminal extracts at least one of PDOP and signal-to-noise ratio from the real-time raw satellite observation data as the satellite positioning quality parameters.

[0059] As an optional implementation, the edge terminal can also combine PDOP, signal-to-noise ratio, and number of visible satellites according to preset weights to obtain a comprehensive positioning quality index as the satellite positioning quality parameter.

[0060] Regarding risk zone determination, in some embodiments, the edge device locates the paver's real-time positioning information within the positioning failure risk field, and reads the risk value corresponding to the paver's current position and a segment of its forward travel path. When there is a location on the forward travel path where the risk value reaches a risk threshold, the edge device determines that the paver will enter the positioning failure risk zone. The real-time positioning information can come from the real-time positioning results of the vehicle-mounted satellite receiver, or from the combined positioning results of the vehicle-mounted satellite receiver and the inertial navigation device.

[0061] Furthermore, in an optional embodiment of this application, the edge terminal introduces a speed adaptive prediction time domain adjustment mechanism when determining whether the paver will enter a positioning failure risk zone. Specifically, the edge terminal acquires the paver's current travel speed and acceleration, and dynamically adjusts the value of the preset prediction time domain based on the current travel speed and acceleration. When the current travel speed is greater than a speed threshold, the preset prediction time domain is shortened; when the current travel speed is less than the speed threshold, the preset prediction time domain is extended. As an optional implementation, the value of the preset prediction time domain is the product of the basic prediction time domain and a speed adjustment factor, where the speed adjustment factor is negatively correlated with the current travel speed. Thus, at high speeds, misjudgment of risk zones caused by excessively long prediction time domains can be avoided; at low speeds, omissions in risk identification caused by excessively short prediction time domains can be avoided, improving the accuracy of determining entry into positioning failure risk zones.

[0062] S106, when it is determined that the paver will enter the positioning failure risk area, the edge end pre-activates the laser three-dimensional imaging device and the millimeter-wave radar before entering the positioning failure risk area to obtain the pose pre-estimation data between the paver and the paved surface.

[0063] Specifically, when it is determined that the paver will enter the positioning failure risk area, the edge end pre-activates the laser three-dimensional imaging device and the millimeter-wave radar before entering the positioning failure risk area to obtain the pose pre-estimation data between the paver and the paved surface.

[0064] In some embodiments, the edge device sends a start command to the laser 3D imaging device and the millimeter-wave radar at a distance or time before the paver reaches the positioning failure risk area. The laser 3D imaging device begins to acquire point cloud data of the paved surface, and the millimeter-wave radar begins to acquire echo data of the overlap area. The edge device processes the acquired data to obtain pose pre-estimation data. Since the laser 3D imaging device and the millimeter-wave radar have already entered the working state and completed preliminary data acquisition before the paver actually enters the risk area, the relative pose perception channel is already in a ready state when the satellite positioning signal actually deteriorates, which can avoid the time delay caused by the cold start of the sensing channel.

[0065] Furthermore, in an optional embodiment of this application, after the edge terminal pre-starts the laser 3D imaging device and the millimeter-wave radar, a readiness self-check of the relative pose perception channel is performed before the paver actually enters the positioning failure risk area. Specifically, the edge terminal acquires the startup status, data acquisition quality, and data synchronization status of the laser 3D imaging device and the millimeter-wave radar. When the startup status, data acquisition quality, and data synchronization status all meet preset readiness conditions, the edge terminal marks the relative pose perception channel as ready. When any one of the startup status, data acquisition quality, and data synchronization status fails to meet the readiness conditions, the edge terminal generates a channel anomaly alarm and sends it back to the cloud. Thus, by performing a readiness self-check before entering the risk area, the availability of the relative pose perception channel is ensured when the satellite positioning signal actually fails, avoiding the problem that the channel cannot provide effective relative pose information after pre-start due to incomplete initialization or abnormal data acquisition.

[0066] S107, when the satellite positioning quality parameters meet the positioning quality discrimination rules, the edge end injects the pre-correction parameter set into the paver motion controller according to the global optimization mode to control the paving speed curve and / or the tamping frequency.

[0067] Specifically, when the satellite positioning quality parameters meet the positioning quality discrimination rules, the edge terminal injects the pre-calibration parameter set into the paver motion controller according to the global optimization mode to control the paving speed curve and / or tamping frequency. In some embodiments, the positioning quality discrimination rules include: PDOP is less than or equal to a third threshold, and the signal-to-noise ratio is greater than or equal to a fourth threshold. When the satellite positioning quality parameters meet the above discrimination rules, the edge terminal determines that the current satellite positioning signal is reliable and enters the global optimization mode. In the global optimization mode, the edge terminal selects the corresponding paving speed curve and tamping frequency from the pre-calibration parameter set according to the domain construction attributes of the paving sub-area where the paver is currently located, and injects them into the paver motion controller, which then drives the paver's travel drive mechanism and tamping mechanism accordingly.

[0068] Furthermore, in an optional embodiment of this application, the edge terminal also performs control parameter pre-smoothing processing in global optimization mode. Specifically, when selecting the first control parameter corresponding to the current paving sub-region from the pre-correction parameter set, the edge terminal also obtains the second control parameter corresponding to the previously passed paving sub-region, performs weighted smoothing on the first and second control parameters to obtain transition control parameters, and injects the transition control parameters into the paver motion controller. The weighting coefficients of the weighted smoothing are determined based on the spatial distance between the current paving sub-region and the previously passed paving sub-region and the differences in regional construction attributes. This avoids paving quality jumps caused by abrupt changes in control parameters due to switching of regional construction attributes, ensuring construction continuity at the boundaries of different paving sub-regions.

[0069] S108, when the satellite positioning quality parameters do not meet the positioning quality discrimination rules, the edge end switches from the global optimization mode to the local material sensing mode, stops using the satellite positioning coordinates as the paving position control input, and generates a local correction command based on the pose pre-estimation data and the relative pose information obtained by the laser three-dimensional imaging device and the millimeter-wave radar in the local material sensing mode.

[0070] Specifically, when the satellite positioning quality parameters do not meet the positioning quality discrimination rules, the edge terminal switches from the global optimization mode to the local material sensing mode, stops using satellite positioning coordinates as paving position control input, and generates a local correction command based on the pose pre-estimation data and the relative pose information acquired by the laser 3D imaging device and the millimeter-wave radar in the local material sensing mode. In some embodiments, when PDOP is greater than a third threshold or the signal-to-noise ratio is less than a fourth threshold, the edge terminal determines that the satellite positioning signal is unreliable and switches from the global optimization mode to the local material sensing mode. After switching, the edge terminal no longer uses the satellite positioning coordinates output by the vehicle-mounted satellite receiver as the input for paving position control, thereby avoiding unreliable positioning coordinates from contaminating the control process. In the local material sensing mode, the laser 3D imaging device and the millimeter-wave radar continuously collect data. Based on the initial relative pose provided by the pose pre-estimation data, and combined with the relative pose information acquired in real time by the two sensors in this mode, the edge terminal determines the real-time relative pose of the paver relative to the paved surface. The edge terminal generates a local correction command based on the deviation between the real-time relative pose and the target paving state.

[0071] As an optional implementation, the edge device uses the pose pre-estimation data as the initial value for relative pose tracking, uses the continuous elevation boundary provided by the laser 3D imaging device and the overlap area boundary provided by the millimeter-wave radar as the observation, obtains the real-time relative pose through filtering estimation, and then calculates the local correction command based on this.

[0072] As another optional implementation, the edge terminal can input real-time relative pose and local material state into an online learning control strategy, which will then output local correction instructions.

[0073] Furthermore, in an optional embodiment of this application, when the edge terminal switches from the global optimization mode to the local material sensing mode, it also performs a smooth switching of control parameters. Specifically, the edge terminal records the global optimization control parameters injected into the paver motion controller in the last frame before the switch as a switching reference value, obtains the local adjustment amount in the local correction command, superimposes the local adjustment amount on the switching reference value to form a smooth switching control amount, and injects the smooth switching control amount into the paver motion controller. The initial value of the local adjustment amount at the time of switching is zero, and it gradually increases to the target value according to a preset gradual change rate after the switch. Thus, the impact on the paver actuator caused by the jump in control amount at the moment of mode switching is avoided, and a smooth transition from the global optimization mode to the local material sensing mode is achieved.

[0074] S109, the edge end injects the local correction command into the paver motion controller to adjust the paver's operating parameters; the operating parameters include at least one of the following: auger speed gradient, screed inclination angle, and overlap zone vibration phase.

[0075] Specifically, the edge end injects the local correction command into the paver motion controller to adjust the paver's operating parameters. These operating parameters include at least one of the following: auger speed gradient, screed inclination angle, and overlap zone vibration phase. In some embodiments, the local correction command includes at least one of the following: adjustment amount for the auger speed gradient, adjustment amount for the screed inclination angle, and adjustment amount for the overlap zone vibration phase. After receiving the local correction command, the paver motion controller drives the auger spreading mechanism to adjust the auger speed gradient, drives the screed mechanism to adjust the screed inclination angle, and drives the vibration mechanism to adjust the overlap zone vibration phase, respectively. Thus, even during periods of satellite positioning signal failure, the paver can still maintain paving operations based on the results of local material sensing.

[0076] In one embodiment of this application, the method further includes: the edge device transmitting paving operation data back to the cloud; the cloud updating the positioning uncertainty covariance matrix based on the paving operation data; and simultaneously calculating a predicted value for satellite positioning quality recovery while updating the positioning uncertainty covariance matrix. The paving operation data may include at least one of the following: raw satellite observation data, multispectral sensor data, laser three-dimensional imaging data, and millimeter-wave radar data collected during the paving operation. The predicted value for satellite positioning quality recovery is used to characterize the improvement trend of satellite positioning signal quality over a future period.

[0077] As an optional implementation, the cloud uses historical PDOP and signal-to-noise ratio to construct a time series model, extrapolates the PDOP and signal-to-noise ratio at future moments, and calculates the recovery prediction value accordingly; for example, the recovery prediction value can be the remaining time required for the predicted signal quality to reach a usable level, or it can be the probability of signal quality recovery within a preset time period in the future.

[0078] As another optional implementation, the cloud-based system combines satellite ephemeris data with the occlusion distribution on both sides of the paver's planned path to infer the time when the paver leaves the occluded area, and thus determines the recovery prediction value. The above implementation methods can be used individually or in combination. The cloud-based system sends the satellite positioning quality recovery prediction value to the edge device. The edge device, in the local material sensing mode, adjusts the sampling frequency of the laser 3D imaging device and the millimeter-wave radar based on the satellite positioning quality recovery prediction value.

[0079] As an optional implementation, when the satellite positioning quality recovery prediction signal is expected to recover within a short time, the edge device reduces the sampling frequency of the laser 3D imaging device and the millimeter-wave radar to save computing resources and energy. When the satellite positioning quality recovery prediction signal is difficult to recover within a longer time, the edge device maintains or increases the sampling frequency to maintain sufficient local sensing accuracy in the absence of satellite positioning. When the cloud determines that the satellite positioning quality meets the recovery conditions, the edge device switches back to the global optimization mode according to the global optimization mode switchback command sent by the cloud.

[0080] In some embodiments, the recovery condition can be set as follows: PDOP is continuously lower than a fifth threshold for a first preset duration, and the signal-to-noise ratio is continuously higher than a sixth threshold for the first preset duration. When the actual satellite positioning quality continuously monitored by the cloud meets the recovery condition, the cloud sends a global optimization mode switchback command to the edge device, and the edge device switches back from the local material sensing mode to the global optimization mode accordingly.

[0081] Furthermore, in an optional embodiment of this application, when the edge device switches back to the global optimization mode, it also performs transfer learning from local sensing data to the global optimization model. Specifically, the edge device packages the relative pose data and local material state data collected in the local material sensing mode into a transfer data package and sends the transfer data package back to the cloud. The cloud uses the transfer data package as incremental training samples to fine-tune and update the artificial intelligence optimization model, enabling the artificial intelligence optimization model to learn the actual pose deviation and material state characteristics obtained in the local material sensing mode, generate an updated pre-correction parameter set, and send the updated pre-correction parameter set to the edge device. Thus, the local sensing data of each satellite positioning failure event is fed back to the cloud for model iterative optimization, enabling the artificial intelligence optimization model to more accurately predict the construction deviation of the positioning failure area in subsequent global optimization, and realizing the ability to continuously learn from failure events.

[0082] In one embodiment of this application, the construction benchmark model is constructed in the following manner: The system acquires terrain elevation point clouds collected by a lidar-equipped UAV, surface thermal radiation distribution and subgrade moisture content anomaly zones collected by a multispectral sensor on a UAV, building information model (BIM) design parameters, and raw satellite observation data. The lidar generates terrain elevation point clouds characterizing surface undulations by emitting laser beams and receiving reflected signals to measure distance. The multispectral sensor collects reflection or radiation information of ground features within multiple specific wavelength ranges, thereby obtaining the surface thermal radiation distribution and subgrade moisture content anomaly zones. The BIM design parameters include at least one of the following: road cross slope, longitudinal slope, thickness, and material composition.

[0083] Subsequently, the topographic elevation point cloud, the surface thermal radiation distribution, the base layer moisture content anomaly zone, the building information model design parameters, and the raw satellite observation data are spatiotemporally registered to form a geometric layer, a thermal layer, a material layer, and a positioning uncertainty layer under the same construction coordinate system. The spatiotemporal registration refers to the process of unifying data from different sensors and coordinate systems into the same construction coordinate system and aligning them temporally. As an optional implementation, the cloud uses the building information model's design coordinate system as the construction coordinate system. The topographic elevation point cloud is registered to this coordinate system through control point matching, and multispectral data is associated with the corresponding locations using timestamp alignment. The geometric layer is constructed from the topographic elevation point cloud, representing the geometric shape of the paving area; the thermal layer is constructed from the surface thermal radiation distribution, representing the temperature distribution; the material layer is constructed from the base layer moisture content anomaly zone, representing material properties; and the positioning uncertainty layer is constructed from the positioning uncertainty covariance matrix generated from the raw satellite observation data, representing the positioning reliability at various locations. Finally, the construction reference model is constructed based on the geometric layer, the thermal layer, the material layer, and the positioning uncertainty layer.

[0084] Furthermore, in an optional embodiment of this application, after the cloud-based system constructs the construction baseline model, it performs cross-layer cross-validation based on the spatial correlation between the geometric layer, thermal layer, and material layer. Specifically, the cloud-based system extracts abrupt slope changes in the geometric layer and anomalous temperature gradients in the thermal layer. When the spatial overlap between the abrupt slope changes and the anomalous temperature gradients exceeds a preset overlap threshold, the cloud-based system marks the overlapping area as a thermo-mechanical coupling sensitive region and configures incremental simulation accuracy parameters for the thermo-mechanical coupling sensitive region. When solving the fractional-order spatiotemporal coupling constitutive model, the cloud-based system uses a finer solution step size smaller than the default step size for the thermo-mechanical coupling sensitive region. Thus, by automatically identifying regions with geometric and thermo-mechanical anomalies through cross-layer cross-validation, the simulation accuracy of key areas is improved.

[0085] In one embodiment of this application, the domain construction attributes include at least two of the following: conventional paving domain, variable slope curve domain, overlapping joint domain, and structural boundary domain. The cloud platform determines the domain construction attributes of the corresponding paving sub-region based on the slope curvature characteristics of different paving sub-regions in the geometric layer, the temperature dispersion characteristics in the thermal layer, the moisture content anomaly characteristics in the material layer, and the positioning uncertainty in the positioning uncertainty layer. As an optional implementation, the cloud platform sets the following criteria: when the slope curvature characteristics exceed a curvature threshold, it is determined to be a variable slope curve domain; when the temperature dispersion characteristics exceed a temperature dispersion threshold and are located at a width joint, it is determined to be an overlapping joint domain; when the paving sub-region is adjacent to a fixed structure, it is determined to be a structural boundary domain; the rest are determined to be conventional paving domains. When the same paving sub-region simultaneously meets multiple determination conditions, one of the following can be selected based on the priority of structural boundary domain, overlapping joint domain, variable slope curve domain, and conventional paving domain: As an optional implementation, for variable-slope curves and structural boundary areas with poor positioning reliability, the cloud-based system is configured with a larger local sensing pre-start distance and a longer local sensing pre-start time; for conventional paving areas, the cloud-based system is configured with a smaller local sensing pre-start distance or no pre-start. The local sensing pre-start distance can be between ten and fifty meters, and the local sensing pre-start time can be between three and fifteen seconds; this application does not limit the specific values. The edge device determines the pre-start time of the laser three-dimensional imaging device and the millimeter-wave radar based on the local sensing pre-start distance, the local sensing pre-start time, and the current travel speed of the paver. As an optional implementation, the edge device divides the local sensing pre-start distance by the current travel speed of the paver to obtain a distance-based pre-start advance; the larger of this advance and the local sensing pre-start time is taken as the final pre-start advance; accordingly, the laser three-dimensional imaging device and the millimeter-wave radar are pre-started at the time corresponding to this advance before the paver reaches the positioning failure risk area.

[0086] In one embodiment of this application, the positioning failure risk field is determined as follows: the position uncertainty corresponding to the positioning uncertainty covariance matrix is ​​projected along the planned travel direction of the paver to obtain the travel direction uncertainty. As an optional implementation, the cloud platform performs a quadratic projection of the positioning uncertainty covariance matrix along the unit vector of the planned travel direction, i.e., multiplying the unit vector of the travel direction on the left by the covariance matrix and then on the right by the unit vector, and the square root of the resulting scalar is taken as the travel direction uncertainty at that position. The travel direction uncertainty obtained in this way reflects the positioning error level in the paver's forward direction.

[0087] Subsequently, based on the domain construction attributes, regional weights are assigned to the uncertainty of the travel direction, and a positioning failure risk field continuously distributed along the paver's planned travel path is formed based on the weighted travel direction uncertainty. As an optional implementation, the cloud-based system assigns regional weights greater than one to the variable slope curve domain, overlap joint domain, and structural boundary domain, and assigns a regional weight equal to one to the conventional paving domain. The travel direction uncertainty is multiplied by the corresponding regional weight to obtain the risk value at each location. The risk values ​​at each location are arranged along the planned travel path to constitute the positioning failure risk field. The edge terminal extrapolates the paver's travel coverage area within the positioning failure risk field based on the paver's real-time positioning information and a preset prediction time domain. When the travel coverage area intersects with the positioning failure risk area, it is determined that the paver will enter the positioning failure risk area. As an optional implementation, the edge end is taken as the starting point of the paver's current position, and the distance obtained by multiplying the current travel speed by a preset prediction time domain is used as the length. This distance is then pushed forward along the planned travel path to obtain the travel coverage area. The travel coverage area is compared with the area in the positioning failure risk field where the risk value reaches the risk threshold. If the two intersect, it is determined that the paver will enter the positioning failure risk area. The preset prediction time domain can be a value between three and ten seconds.

[0088] In one embodiment of this application, the pose pre-estimation data includes at least two of the following: lateral offset of the paver relative to the paved surface, elevation deviation, and heading angle deviation. The lateral offset characterizes the lateral deviation of the paver relative to the boundary of the paved surface; the elevation deviation characterizes the elevation deviation of the paver's screed bottom surface relative to the paved surface; and the heading angle deviation characterizes the angular deviation of the paver's travel direction relative to the extension direction of the paved surface boundary. After the edge end pre-activates the laser 3D imaging device and the millimeter-wave radar, it extracts the continuous elevation boundary of the paved surface based on the point cloud of the paved surface collected by the laser 3D imaging device, and extracts the overlap zone boundary based on the echo data collected by the millimeter-wave radar. As an optional implementation, the edge end scans the point cloud of the paved surface laterally line by line, detects the locations where the elevation changes abruptly, and connects these locations along the travel direction to obtain the continuous elevation boundary; the edge end processes the millimeter-wave radar echo according to distance-angle, detects the change in echo intensity at the material density difference on both sides of the overlap zone, and extracts the overlap zone boundary accordingly. The edge device performs time-synchronized fusion of the continuous elevation boundary and the overlap area boundary to obtain the pose pre-estimation data. As an optional implementation, the edge device first aligns the data from the laser 3D imaging device and the millimeter-wave radar according to their acquisition timestamps. Then, the continuous elevation boundary is used to calculate the elevation deviation and heading angle deviation, and the overlap area boundary is used to calculate the lateral offset. Finally, the above calculation results are combined into the pose pre-estimation data. Because the data from the two sensors are fused in time synchronization, the pose pre-estimation data can comprehensively reflect the relative pose of the paver with respect to the paved surface in the lateral, elevation, and heading directions at the same time.

[0089] In one embodiment of this application, the edge end generates a local correction command, including: determining the local deviation of the paver relative to the target paving trajectory based on the pose pre-estimation data and the relative pose information. The target paving trajectory refers to the trajectory that the paver should travel and operate according to the construction design. As an optional implementation, the edge end uses the pose pre-estimation data as the initial relative pose, uses the relative pose information continuously acquired in the local material sensing mode as the update quantity, obtains the current relative pose through filtering estimation, and then compares the current relative pose with the target paving trajectory to obtain a local deviation including at least two of lateral deviation, elevation deviation, and heading deviation. The edge end determines the local material state based on the paving surface smoothness information acquired by the laser three-dimensional imaging device and the overlap zone boundary information acquired by the millimeter-wave radar. The local material state is used to characterize the material distribution and surface quality of the current paving area. As an optional implementation, the edge end calculates the smoothness index from the paving surface points and calculates the degree of lateral material distribution unevenness from the material accumulation difference on both sides of the overlap zone boundary; the two together constitute the local material state. The edge end generates the local correction command based on the local deviation and the local material state. The local correction command adjusts the spiral speed gradient according to the difference in lateral material accumulation in the overlap area, adjusts the ironing plate angle according to the elevation deviation, and adjusts the vibration phase of the overlap area according to the position change of the overlap area boundary. As an optional implementation, when there is more material accumulation on one side of the overlap area, the local correction command increases the spiral speed on that side and decreases the spiral speed on the other side, thereby adjusting the spiral speed gradient to balance the lateral fabric distribution. When the elevation deviation is positive (i.e., the ironing plate is too high), the local correction command decreases the ironing plate angle, and vice versa. When the overlap area boundary position changes, the local correction command adjusts the vibration phase of the overlap area accordingly, aligning the vibration with the overlap area position.

[0090] Furthermore, in an optional embodiment of this application, when the edge end generates the local correction command, a collaborative constraint is introduced between the helical speed gradient, the ironing plate tilt angle, and the vibration phase of the overlap area. Specifically, the edge end establishes a collaborative constraint matrix between the helical speed gradient adjustment, the ironing plate tilt angle adjustment, and the vibration phase adjustment of the overlap area. The off-diagonal elements of the collaborative constraint matrix represent the mutual influence coefficients between different operating parameters. When generating the local correction command, the edge end uses the collaborative constraint matrix to couple and constrain the helical speed gradient adjustment, the ironing plate tilt angle adjustment, and the vibration phase adjustment of the overlap area, ensuring that the adjustment of any operating parameter does not exceed the collaborative feasible region determined by the collaborative constraint matrix. This avoids control conflicts caused by mutual coupling when multiple operating parameters are adjusted independently, and ensures the collaborative consistency among multiple actuators in the local material sensing mode.

[0091] In one embodiment of this application, the artificial intelligence optimization model includes a spatiotemporal graph neural network and a deep temporal optimization network. The spatiotemporal graph neural network uses paving sub-regions as graph nodes and spatial adjacency relationships, temperature propagation relationships, and positioning residual correlation relationships between adjacent paving sub-regions as graph edges to determine the state influence relationships between the paving sub-regions. As an optional implementation, the node state of each graph node includes the geometric features, material state features, and positioning quality features of that paving sub-region. The geometric features include at least one of elevation, slope, and curvature; the material state features include at least one of predicted temperature and segregation index; and the positioning quality features include positioning residuals or uncertainty weights obtained by PDOP mapping. The spatiotemporal graph neural network determines the state influence relationships through two stages: spatial aggregation and temporal evolution. In the spatial aggregation stage, each node collects states from its neighboring nodes and aggregates them using a weighted attention mechanism. That is, for node i, its aggregation result is obtained by weighted summation of the transformed states of neighboring node j according to attention weights, where the attention weights reflect the influence strength between node i and node j. In the temporal evolution stage, a gated recurrent unit is used, taking the current node state and spatial aggregation result as input and outputting the node state at the next moment, thereby characterizing the dynamic evolution of the construction state over time. The deep temporal optimization network generates the pre-correction parameter set based on the state influence relationship and the reliable simulation state. As an optional implementation, the deep temporal optimization network sequentially includes a temporal embedding layer, a bidirectional gated recurrent unit layer, a spatiotemporal attention fusion layer, a multi-objective optimization head, and a constraint pruning layer. The temporal embedding layer maps the paving speed sequence, tamper frequency sequence, auger speed gradient sequence, temperature sequence, and high-order sequence into a feature vector of a unified dimension; the bidirectional gated recurrent unit layer captures the temporal dependencies of the paving process; the spatiotemporal attention fusion layer weightedly fuses the temporal features output by the bidirectional gated recurrent unit layer with the spatial features output by the spatiotemporal graph neural network; the multi-objective optimization head synchronously outputs the control parameter sequences corresponding to the paving speed, tamper frequency, and auger speed gradient; the constraint pruning layer imposes limit constraints on the output parameters according to construction specifications, ensuring that the output parameters are within the engineering feasible range. The pre-correction parameter set includes paving speed curves, tamping hammer frequency sequences, and auger speed gradient sequences corresponding to different domain construction attributes.

[0092] In one embodiment of this application, when generating the pre-correction parameter set, the loss function of the artificial intelligence optimization model includes the residuals of the fractional-order constitutive equation system, the smoothness prediction error, the carbon emission intensity, and a penalty term for consistency between satellite signals and inertial navigation trajectories. The residuals of the fractional-order constitutive equation system characterize the degree of consistency between the simulation state corresponding to the control parameters during the optimization process and the fractional-order spatiotemporally coupled constitutive model; the smoothness prediction error characterizes the deviation between the predicted paving smoothness and the target smoothness; the carbon emission intensity characterizes the carbon emission per unit paving area; and the penalty term for consistency between satellite signals and inertial navigation trajectories characterizes the degree of deviation between the trajectory determined by the satellite signals and the trajectory determined by inertial navigation. The loss function can be constructed by weighted summation of the above items according to their corresponding weights. When the position uncertainty corresponding to the positioning uncertainty covariance matrix increases, the weight of the penalty term for consistency between satellite signals and inertial navigation trajectories in the loss function is increased, and the node aggregation weight of the low-confidence paving sub-region in the spatiotemporal graph neural network is decreased. As an optional implementation, the cloud-based system sets the weight of the consistency penalty term between satellite signals and inertial navigation trajectories to monotonically increase with position uncertainty. Simultaneously, in the spatial aggregation stage of the spatiotemporal graph neural network, the attention weights of nodes corresponding to low-confidence localization sub-regions are multiplied by a suppression factor less than one, reducing their contribution to the aggregation of neighboring node states. Therefore, when the localization confidence of a certain region decreases, on the one hand, increasing the weight of the consistency penalty term makes the optimization result more dependent on the inertial navigation trajectory; on the other hand, reducing the aggregation weight of nodes in that region reduces its contamination of the global state estimation.

[0093] Furthermore, in an optional embodiment of this application, the loss function further includes a time-domain covariance penalty term between the paving speed and the tamping frequency. This time-domain covariance penalty term is used to constrain the coordination of the changes in paving speed and tamping frequency in the time domain. The cloud calculates the time-domain covariance between the paving speed sequence and the tamping frequency sequence within a preset time window. When the absolute value of the time-domain covariance exceeds a preset covariance threshold, the weight of the time-domain covariance penalty term in the loss function is increased to constrain the optimization process to tend to generate a control parameter sequence where the changes in paving speed and tamping frequency are coordinated. This avoids density fluctuations caused by opposite directions of change in paving speed and tamping frequency in the optimization results, improving the physical consistency of the joint optimization.

[0094] In one embodiment of this application, the method further includes: the cloud platform generating an error source map based on the paving operation simulation results. The error source map includes the expected error value of each paving sub-region, the symbolic causal rules corresponding to the expected error values, and a positioning sensitivity index. The expected error value characterizes the expected construction error in each paving sub-region; the symbolic causal rules describe the causal relationship between construction parameters, environmental factors, and paving quality in the form of logical expressions, such as a rule describing how exceeding a threshold for a certain construction parameter will lead to a deviation in a certain quality indicator; the positioning sensitivity index is used to indicate the vulnerability of the expected error value to satellite signal degradation. As an optional implementation, the cloud platform sorts the symbolic causal rules by causal contribution and extracts the causal rules whose causal contribution reaches the contribution threshold as causal rules affecting paving quality. The causal contribution can be obtained through at least one of Granger causality test, mutual information, structural equation modeling, or machine learning-based interpretability analysis. The cloud platform sends the symbolic causal rules and the positioning sensitivity index to the edge device. The edge endpoint, based on the symbolic causal rules and the positioning sensitivity index, increases the data processing frequency of a first state variable and decreases the data processing frequency of a second state variable. The first state variable includes at least one of raw satellite observation data, PDOP, elevation deviation, and positioning uncertainty covariance matrix. In some embodiments, the edge endpoint identifies state variables indicated by the symbolic causal rules as having a significant impact on paving quality or being highly correlated with positioning sensitivity as the first state variable, and identifies the remaining state variables as the second state variables. As an optional implementation, the edge endpoint increases the sampling rate, filtering frequency, or model update frequency for the first state variable, for example, increasing the solution frequency of raw satellite observation data from once per second to five or ten times per second; and decreases the corresponding processing frequency for the second state variable to save computational resources and energy consumption at the edge endpoint.

[0095] Furthermore, in an optional embodiment of this application, after the cloud generates the error source map, it also constructs a self-healing control strategy library based on the error source map. Specifically, the cloud associates a pre-configured self-healing action sequence with each expected error value according to the causal chain relationship in the symbolic causal rules. The self-healing action sequence includes at least one control parameter adjustment action. The cloud stores the association between the expected error value and the self-healing action sequence as a self-healing control strategy library and sends the self-healing control strategy library to the edge terminal. When the edge terminal detects that the deviation between the current construction deviation and the expected error value exceeds a preset deviation threshold, it queries the self-healing control strategy library according to the current construction deviation, calls the corresponding self-healing action sequence to generate a self-healing correction instruction, and superimposes the self-healing correction instruction onto the local correction instruction. Thus, through the pre-constructed self-healing control strategy library, the edge terminal can quickly call the contingency plan for proactive correction when it detects a construction error of the expected type, improving the timeliness and effectiveness of error response.

[0096] like Figure 3 As shown in the illustration, this application also provides a paving control system based on an artificial intelligence architecture, including a cloud platform, an edge terminal of the paver, a vehicle-mounted satellite receiver, a laser three-dimensional imaging device, a millimeter-wave radar, and a paver motion controller. The cloud platform is used to execute the steps performed by the cloud platform in the above method, and the edge terminal is used to execute the steps performed by the edge terminal in the above method. For the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple; relevant details can be found in the description of the method embodiment.

[0097] like Figure 4 As shown, this application embodiment provides an electronic device, which can be a component of a satellite-based photovoltaic piling guidance system, including a processor 10 and a memory 11. Optionally, the device may further include a communication interface 12 and a bus 9. The processor 10, communication interface 12, and memory 11 can communicate with each other via the bus 9. The communication interface 12 can be used for information transmission. The processor 10 can call logical instructions in the memory 11 to execute the paving control method based on the AI ​​architecture provided in any of the above embodiments.

[0098] Furthermore, the logical instructions in the aforementioned memory 11 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0099] The memory 11, as a computer-readable storage medium, can be used to store software programs and computer-executable programs, such as program instructions / modules corresponding to the methods in the embodiments of this application. The processor 10 executes functional applications and data processing by running the program instructions / modules stored in the memory 11, thereby implementing the paving control method in the artificial intelligence architecture described above.

[0100] This application also provides one or more computer-readable media storing instructions that, when executed by one or more processors, cause the processors to perform the paving control method described in any of the above embodiments. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state drive).

[0101] The solution in the above embodiments of this application integrates geometric, thermal, material properties and positioning uncertainty covariance matrix into a construction benchmark model, breaking through the limitation of traditional paving modeling that only focuses on spatial geometry and design parameters, and making the construction scene representation more complete.

[0102] The above-described embodiments of this application employ a fractional-order spatiotemporal coupled constitutive model to model the non-equilibrium evolution behavior of asphalt mixtures. This model can characterize the history-dependent effect of heat conduction, the non-Gaussian migration characteristics of aggregates, and the viscoelastic-plastic deformation of the base and surface layers. Compared with conventional empirical models, this model better reflects complex construction mechanics and improves the physical consistency of the simulation.

[0103] The above-described embodiments of this application apply a confidence weighting to the spatial domain of the constitutive model based on the covariance matrix of the positioning uncertainty. This can suppress the influence of low-confidence regions on the global simulation when the positioning uncertainty increases, thereby improving the robustness of the model in complex environments.

[0104] The solution of the above embodiments of this application achieves forward-looking determination of entering the positioning failure risk area by extrapolating the positioning failure risk field and the travel coverage area. Based on this, the relative pose perception channel is pre-activated before entering the risk area, avoiding the cold start delay of the sensing channel. This enables seamless switching to the local material perception mode when the satellite positioning quality deteriorates, ensuring the continuity of paving operations.

[0105] The solution in the above embodiments of this application enables the edge end to proactively adjust the sensing sampling frequency by simultaneously calculating the recovery prediction value when updating the positioning uncertainty covariance matrix, and promptly switch back to the global optimization mode when the recovery conditions are met, thereby avoiding mode stagnation and resource waste.

[0106] The solution in the above embodiments of this application links the loss function weights and node aggregation weights with the positioning credibility, enabling the artificial intelligence optimization model to adaptively adjust its dependence on different information sources and regions when the positioning credibility changes, thereby improving the reliability of the optimization results in complex positioning environments.

[0107] The solution in the above embodiments of this application, by ranking the causal contribution of error source maps and symbolic causal rules, enables the edge end to concentrate computing resources on the first state variable, thereby improving real-time response capability and control accuracy while reducing the processing overhead of secondary variables.

[0108] The above preferred embodiments are only used to illustrate the technical solutions of this application and are not restrictive. Although this application has been described in detail, those skilled in the art should understand that changes in form and detail can be made without departing from the scope defined by the claims of this application. The dimensions in the drawings are not related to the specific physical object, and the physical object dimensions can be arbitrarily changed.

Claims

1. A paving control method based on an artificial intelligence architecture, applied to a paving control system, the paving control system comprising a cloud platform, an edge device mounted on the paver, a vehicle-mounted satellite receiver, a laser three-dimensional imaging device, a millimeter-wave radar, and a paver motion controller, characterized in that, The method includes: The cloud-based system determines the domain construction attributes corresponding to different paving sub-regions in the paving area based on the construction benchmark model of the paving area and the positioning uncertainty covariance matrix generated by the original satellite observation data from the vehicle-mounted satellite receiver. The cloud platform determines the positioning failure risk field based on the positioning uncertainty covariance matrix and the domain construction attributes. The cloud platform establishes a fractional-order spatiotemporal coupled constitutive model based on the construction benchmark model, and performs a credibility weighting on the spatial domain of the fractional-order spatiotemporal coupled constitutive model according to the positioning uncertainty covariance matrix to obtain the credible simulation state of each paving sub-area. The cloud inputs the trusted simulation state, the domain construction attributes, and the positioning failure risk field into the artificial intelligence optimization model, performs joint optimization on the paving speed curve, tamping frequency, and / or auger speed gradient, generates a pre-correction parameter set, and sends the positioning failure risk field and the pre-correction parameter set to the edge terminal. During the paving operation, the edge end acquires real-time raw satellite observation data output by the vehicle-mounted satellite receiver, determines satellite positioning quality parameters based on the real-time raw satellite observation data, and determines whether the paver will enter the positioning failure risk area based on the positioning failure risk field and the real-time positioning information of the paver. When it is determined that the paver will enter the positioning failure risk area, the edge end pre-activates the laser three-dimensional imaging device and the millimeter-wave radar before entering the positioning failure risk area to obtain pose pre-estimation data between the paver and the paved surface. When the satellite positioning quality parameters meet the positioning quality discrimination rules, the edge end injects the pre-correction parameter set into the paver motion controller according to the global optimization mode to control the paving speed curve and / or the tamping frequency. When the satellite positioning quality parameters do not meet the positioning quality discrimination rules, the edge end switches from the global optimization mode to the local material sensing mode, stops using the satellite positioning coordinates as the paving position control input, and generates a local correction command based on the pose pre-estimation data and the relative pose information obtained by the laser three-dimensional imaging device and the millimeter-wave radar in the local material sensing mode. The edge end injects the local correction command into the paver motion controller to adjust the paver's operating parameters; the operating parameters include at least one of the following: auger speed gradient, screed inclination angle, and overlap zone vibration phase.

2. The method according to claim 1, characterized in that, The method further includes: The edge device transmits paving operation data back to the cloud. The cloud updates the positioning uncertainty covariance matrix based on the paving operation data and simultaneously calculates the satellite positioning quality recovery prediction value when updating the positioning uncertainty covariance matrix. The cloud sends the predicted value of satellite positioning quality recovery to the edge device. The edge device adjusts the sampling frequency of the laser three-dimensional imaging device and the millimeter-wave radar according to the predicted value of satellite positioning quality recovery in the local material sensing mode. When the cloud determines that the satellite positioning quality meets the recovery conditions, the edge device switches back to the global optimization mode according to the global optimization mode switching command sent by the cloud.

3. The method according to claim 1 or 2, characterized in that, The construction benchmark model is constructed in the following manner: The system acquires terrain elevation point clouds collected by lidar on a drone, surface thermal radiation distribution and ground moisture content anomaly areas collected by multispectral sensors on a drone, building information model design parameters, and the raw satellite observation data. The topographic elevation point cloud, the surface thermal radiation distribution, the base layer moisture content anomaly zone, the building information model design parameters, and the original satellite observation data are spatiotemporally registered to form a geometric layer, thermal layer, material layer, and positioning uncertainty layer under the same construction coordinate system; The construction reference model is constructed based on the geometric layer, the thermal layer, the material layer, and the positioning uncertainty layer.

4. The method according to claim 3, characterized in that, The domain construction attributes include at least two of the following: conventional paving domain, variable slope curve domain, lap joint domain, and structural boundary domain. The cloud platform determines the domain construction attributes of the corresponding paving area based on the slope curvature characteristics of the geometric layer, the temperature dispersion characteristics of the thermal layer, the moisture content anomaly characteristics of the material layer, and the positioning uncertainty of the positioning uncertainty layer for different paving sub-areas. The cloud-based system configures local sensing pre-start distance and local sensing pre-start time for the corresponding paving sub-area according to the domain construction attributes. The edge device determines the pre-start time of the laser three-dimensional imaging device and the millimeter-wave radar based on the local sensing pre-start distance, the local sensing pre-start time, and the current travel speed of the paver.

5. The method according to claim 1, characterized in that, The location failure risk field is determined in the following way: Projecting the position uncertainty corresponding to the positioning uncertainty covariance matrix along the planned travel direction of the paver yields the travel direction uncertainty. Based on the domain construction attributes, regional weights are assigned to the uncertainty of the travel direction, and a positioning failure risk field continuously distributed along the paver's planned travel path is formed based on the weighted uncertainty of the travel direction. The edge end extrapolates the paver's travel coverage area in the positioning failure risk field based on the paver's real-time positioning information and the preset prediction time domain, and determines that the paver will enter the positioning failure risk area when the travel coverage area intersects with the positioning failure risk area.

6. The method according to claim 1, characterized in that, The pose prediction data includes at least two of the following: the lateral offset of the paver relative to the paved surface, the elevation deviation, and the heading angle deviation; After the laser three-dimensional imaging device and the millimeter-wave radar are pre-activated at the edge end, the elevation continuous boundary of the paved surface is extracted based on the point cloud of the paved surface collected by the laser three-dimensional imaging device, and the overlap area boundary is extracted based on the echo data collected by the millimeter-wave radar. The edge end performs time-synchronized fusion of the elevation continuous boundary and the overlapping area boundary to obtain the pose pre-estimation data.

7. The method according to claim 6, characterized in that, The edge end generates local correction instructions, including: The local deviation of the paver relative to the target paving trajectory is determined based on the pose pre-estimation data and the relative pose information; The local material condition is determined based on the surface flatness information of the paving obtained by the laser three-dimensional imaging device and the overlap zone boundary information obtained by the millimeter-wave radar. The local correction command is generated based on the local deviation and the local material state, wherein the local correction command adjusts the spiral speed gradient according to the difference in lateral material accumulation in the overlap area, adjusts the ironing plate inclination angle according to the elevation deviation, and adjusts the vibration phase of the overlap area according to the position change of the overlap area boundary.

8. The method according to claim 1, characterized in that, The artificial intelligence optimization model includes a spatiotemporal graph neural network and a deep temporal optimization network; The spatiotemporal graph neural network uses the paving sub-regions as graph nodes and the spatial adjacency relationship, temperature propagation relationship, and positioning residual correlation relationship between adjacent paving sub-regions as graph edges to determine the state influence relationship between the paving sub-regions. The deep temporal optimization network generates the pre-correction parameter set based on the state influence relationship and the credible simulation state; The pre-correction parameter set includes paving speed curves, tamping hammer frequency sequences, and auger speed gradient sequences corresponding to different domain construction attributes.

9. The method according to claim 8, characterized in that, When generating the pre-correction parameter set, the loss function of the artificial intelligence optimization model includes the residuals of the fractional constitutive equation system, the smoothness prediction error, the carbon emission intensity, and the consistency penalty term between the satellite signal and the inertial navigation trajectory. When the position uncertainty corresponding to the positioning uncertainty covariance matrix increases, the weight of the consistency penalty term between the satellite signal and the inertial navigation trajectory in the loss function is increased, and the node aggregation weight of the low-confidence positioning paving sub-region in the spatiotemporal graph neural network is reduced.

10. The method according to claim 1 or 2, characterized in that, The method further includes: The cloud platform generates an error source map based on the paving operation simulation results. The error source map includes the expected error value of each paving sub-region, the symbolic causal rules corresponding to the expected error value, and the positioning sensitivity index. The cloud sends the symbolic causal rules and the location sensitivity index to the edge device; The edge terminal increases the data processing frequency of the first state variable and decreases the data processing frequency of the second state variable according to the symbolic causal rule and the positioning sensitivity index. The first state variable includes at least one of the following: raw satellite observation data, position accuracy attenuation factor PDOP, elevation deviation, and positioning uncertainty covariance matrix.

Images