A multi-sensor coupled compaction energy two-dimensional map generation system and method

Abstract

The application discloses a compaction energy two-dimensional atlas generation system and method coupled with multiple sensors, and relates to the technical field of atlas generation.The system comprises the following steps: constructing a cross-scale prior distribution based on a sparse physical quantity field dataset, generating a partial differential constraint, embedding a neural network model, defining a composite loss function, and outputting an untrained neural network model; uniformly sampling in a two-dimensional plane of a region to be compacted to generate a two-dimensional coordinate point set; combining the sparse physical quantity field dataset to train the untrained neural network model, minimizing the composite loss function, and obtaining a trained neural network model; and dividing a regular grid in the two-dimensional plane of the region to be compacted to generate dense grid point coordinates.The application solves the problem that a static prior distribution ignores the heterogeneity of sparse data in a traditional method by constructing a cross-scale prior distribution based on a sparse physical quantity field dataset.

Description

Technical Field

[0001] This invention relates to the field of spectrum generation technology, and in particular to a multi-sensor coupled compaction energy two-dimensional spectrum generation system and method. Background Technology

[0002] Multi-sensor coupling technology and compaction energy distribution analysis have wide applications in civil engineering, road construction, and geological exploration. In recent years, with advancements in sensor technology, devices such as mechanical sensors, lidar, and inertial measurement units (IMUs) have been widely used to capture physical quantity data during the compaction process, such as force, displacement, and attitude information. Traditional methods generate compaction energy distribution through single sensors or simple data fusion, typically based on physical field reconstruction techniques using regular grids, combined with finite element analysis or interpolation algorithms, to generate energy distribution maps.

[0003] However, existing map generation methods still have room for improvement. First, traditional interpolation algorithms (such as spline interpolation) are prone to producing smooth transitions or artifacts in unsampled areas, reducing the realism of energy distribution. The construction of cross-scale prior distributions is mostly aimed at uniform data, ignoring the heterogeneity of local gradient distributions in sparse data, making it difficult to accurately capture the multi-scale characteristics of compaction energy. Second, in terms of anomaly detection, traditional methods mostly rely on manual experience or threshold segmentation, resulting in low automation and difficulty in efficiently identifying unevenly compacted areas. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method for generating two-dimensional compaction energy spectra using multi-sensor coupling to solve the problem of accurately capturing the multi-scale characteristics of compaction energy.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for generating a two-dimensional compaction energy spectrum using multi-sensor coupling, comprising,

[0008] Within the area to be compacted, the raw data from the mechanical sensors, lidar, and inertial measurement center are aligned and converted into a sparse physical quantity field dataset through time synchronization and attitude correction of the inertial measurement center.

[0009] Based on the sparse physical quantity field dataset, a cross-scale prior distribution is constructed, partial differential constraints are generated, a neural network model is embedded, a composite loss function is defined, and an untrained neural network model is output.

[0010] Uniform sampling is performed in the two-dimensional plane of the area to be compacted to generate a set of two-dimensional coordinate points. Combined with the sparse physical quantity field dataset, the untrained neural network model is trained, and the composite loss function is minimized to obtain the trained neural network model.

[0011] A regular grid is divided in the two-dimensional plane of the area to be compacted to generate dense grid point coordinates. The grid point coordinates are input into the trained neural network model to generate an energy matrix, which is mapped into a two-dimensional heat map. Abnormal areas are identified through gradient field analysis, and a two-dimensional compaction energy map is output.

[0012] As a preferred embodiment of the multi-sensor coupled compaction energy two-dimensional spectrum generation method of the present invention, the step of aligning the original data of the mechanical sensor, lidar, and inertial measurement unit through time synchronization and attitude correction of the inertial measurement unit, and converting them into a sparse physical quantity field dataset, specifically involves:

[0013] Acquire pressure force signals collected by mechanical sensors, point cloud data generated by lidar, and attitude data measured by inertial measurement unit, and align the timestamps.

[0014] Using the pitch and roll angles of the attitude data, geometric correction is performed on the point cloud data to extract the spatial coordinates of the contact point between the roller and the ground;

[0015] The displacement increment is calculated based on the compressive force signal and spatial coordinates, and the dot product operation of force and displacement is performed to generate the instantaneous compaction energy value.

[0016] By associating spatial coordinates with instantaneous compaction energy values, a sparse physical quantity field dataset is generated.

[0017] As a preferred embodiment of the multi-sensor coupled compaction energy two-dimensional spectrum generation method of the present invention, wherein: the construction of a cross-scale prior distribution based on the sparse physical quantity field dataset specifically includes:

[0018] Extract instantaneous compaction energy values ​​and corresponding spatial coordinates from a sparse physical field dataset;

[0019] Calculate the local energy gradient between adjacent coordinate points based on the instantaneous compaction energy value, and generate a gradient distribution map;

[0020] Based on the gradient distribution map, regions are divided into coarse-scale and fine-scale regions. The gradient change characteristics of each scale region are extracted to construct a cross-scale prior distribution.

[0021] As a preferred embodiment of the multi-sensor coupled compaction energy two-dimensional spectrum generation method of the present invention, wherein: the generation of partial differential constraints, embedded in a neural network model, specifically includes:

[0022] By utilizing cross-scale prior distributions, smoothness constraint parameters at both coarse and fine scales are determined, and partial differential constraint equations reflecting the multi-scale characteristics of compaction energy are constructed.

[0023] Initialize the neural network model, embed the partial differential constraint equations into the weight update rules of the neural network model, and complete the embedding.

[0024] As a preferred embodiment of the multi-sensor coupled compaction energy two-dimensional spectrum generation method of the present invention, wherein: the definition of the composite loss function and the output of the untrained neural network model are specifically as follows:

[0025] Extract instantaneous compaction energy values ​​and corresponding spatial coordinates from a sparse physical field dataset;

[0026] Spatial coordinates are input into the neural network model for forward propagation to generate the corresponding predicted compaction energy value. The data mean square error loss is then calculated by combining the instantaneous compaction energy value.

[0027] Based on the partial differential constraint equations, the second derivative of the predicted compaction energy value output by the neural network model is calculated in spatial coordinates to generate the physical law loss.

[0028] By fusing the mean squared error loss of the data and the loss due to physical laws, a composite loss function is generated, resulting in an untrained neural network model.

[0029] As a preferred embodiment of the multi-sensor coupled compaction energy two-dimensional spectrum generation method of the present invention, wherein: the step of uniformly sampling within the two-dimensional plane of the region to be compacted to generate a two-dimensional coordinate point set specifically comprises:

[0030] Determine the two-dimensional planar boundary range of the area to be compacted, divide it into uniform grid cells according to the boundary range, and set the sampling interval of the grid cells;

[0031] Sampling points are randomly selected within each grid cell to generate a two-dimensional coordinate point set.

[0032] As a preferred embodiment of the multi-sensor coupled compaction energy two-dimensional spectrum generation method of the present invention, the step of training the untrained neural network model by combining the sparse physical quantity field dataset and minimizing the composite loss function to obtain the trained neural network model specifically involves:

[0033] Input the set of two-dimensional coordinate points into an untrained neural network model to generate iteratively predicted compaction energy values;

[0034] Using the instantaneous compaction energy values ​​from the sparse physical quantity field dataset, the data error between the iteratively predicted compaction energy values ​​and the actual energy values ​​is calculated;

[0035] Calculate the weighted loss value of data error and partial differential constraint based on the composite loss function;

[0036] The gradient descent algorithm is used to iteratively adjust the parameters of the neural network model and minimize the weighted loss value. When the weighted loss value converges, the trained neural network model is generated.

[0037] As a preferred embodiment of the multi-sensor coupled compaction energy two-dimensional spectrum generation method of the present invention, the step of dividing the two-dimensional plane of the region to be compacted into a regular grid to generate dense grid point coordinates, and inputting the grid point coordinates into a trained neural network model to generate an energy matrix, specifically:

[0038] Based on the two-dimensional planar boundary range of the area to be compacted, high-resolution grid cells are divided, and dense grid point coordinates are generated;

[0039] The coordinates of the dense grid points are input into the trained neural network model to calculate the predicted compaction energy value for each coordinate point;

[0040] Based on the predicted compaction energy value, high-resolution grid cells are filled to generate an energy matrix.

[0041] As a preferred embodiment of the multi-sensor coupled compaction energy two-dimensional spectrum generation method of the present invention, wherein: the mapping is a two-dimensional thermogram, and abnormal regions are identified through gradient field analysis to output a two-dimensional compaction energy spectrum, specifically:

[0042] Based on the compaction energy value of the energy matrix, the color mapping rule is determined, and the compaction energy value of the energy matrix is ​​converted into pixel color value to generate a two-dimensional heat map;

[0043] Calculate the spatial gradient of the energy matrix to generate a gradient field distribution;

[0044] Based on the gradient field distribution, regions where the absolute value of the gradient exceeds a preset gradient threshold are extracted and identified as areas of uneven compaction.

[0045] The contour markers of the unevenly compacted areas are superimposed on the two-dimensional thermogram to generate a two-dimensional compaction energy spectrum.

[0046] Secondly, this invention provides a multi-sensor coupled compaction energy two-dimensional spectrum generation system, comprising,

[0047] The acquisition module, within the area to be compacted, aligns and converts the raw data from the mechanical sensors, lidar, and inertial measurement center into a sparse physical field dataset through time synchronization and attitude correction of the inertial measurement center.

[0048] The construction module constructs a cross-scale prior distribution based on the sparse physical quantity field dataset, generates partial differential constraints, embeds a neural network model, defines a composite loss function, and outputs an untrained neural network model.

[0049] The training module performs uniform sampling within the two-dimensional plane of the area to be compacted to generate a set of two-dimensional coordinate points. Combined with the sparse physical quantity field dataset, it trains the untrained neural network model, minimizes the composite loss function, and obtains the trained neural network model.

[0050] The generation module divides the two-dimensional plane of the area to be compacted into a regular grid, generates the coordinates of dense grid points, inputs the grid point coordinates into the trained neural network model, generates an energy matrix, maps it into a two-dimensional heat map, identifies abnormal areas through gradient field analysis, and outputs a two-dimensional compaction energy map.

[0051] The beneficial effects of this invention are as follows: By constructing a cross-scale prior distribution based on a sparse physical quantity field dataset and combining it with the local energy gradient distribution to divide coarse-scale and fine-scale regions, it overcomes the shortcomings of traditional interpolation algorithms in producing smooth transitions or artifacts in unsampled areas, significantly improving the realism of compaction energy distribution; by capturing multi-scale characteristics through the cross-scale prior distribution, it solves the problem of static prior distribution ignoring the heterogeneity of sparse data in traditional methods, enhancing the model's adaptability to complex terrain compaction scenarios; in addition, by generating a high-resolution energy matrix and calculating the gradient field distribution, it automatically extracts compaction non-uniform regions where the absolute value of the gradient exceeds the threshold, and superimposes contour markers onto a two-dimensional heat map, replacing the inefficient detection method that relies on manual experience or threshold segmentation, significantly improving the automation and efficiency of abnormal area identification, and providing an accurate quality assessment tool for road construction and foundation compaction. Attached Figure Description

[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0053] Figure 1 This is a flowchart of a method for generating two-dimensional compaction energy spectra using multi-sensor coupling.

[0054] Figure 2 This is a schematic diagram of a multi-sensor coupled compaction energy two-dimensional spectrum generation system.

[0055] Figure 3 A flowchart for constructing a cross-scale prior distribution.

[0056] Figure 4 This describes the neural network training and graph generation process. Detailed Implementation

[0057] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0058] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0059] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0060] Reference Figures 1-4 As one embodiment of the present invention, this embodiment provides a method for generating a two-dimensional compaction energy spectrum using multi-sensor coupling, comprising the following steps:

[0061] S1: Within the area to be compacted, the raw data from the mechanical sensors, lidar, and inertial measurement center are aligned and converted into a sparse physical quantity field dataset through time synchronization and attitude correction of the inertial measurement center.

[0062] The system acquires pressure signals from mechanical sensors to record the magnitude of the vertical force exerted by the rollers on the ground; it acquires point cloud data from lidar, which includes the three-dimensional coordinates of the contact points between the rollers and the ground within the compaction area; and it acquires attitude data from the inertial measurement center, which includes the pitch and roll angles of the roller equipment.

[0063] Perform timestamp alignment on pressure force signals, point cloud data, and attitude data. Compare the timestamps of each data point, adjust the sampling time deviation, and ensure that the three correspond at the same point in time.

[0064] Using the pitch and roll angles from the attitude data, a rotation matrix is ​​constructed to generate a three-dimensional rotation matrix. The transformation formulas for rotating the pitch angle around the horizontal axis and the roll angle around the vertical axis are applied sequentially and then merged into a single rotation matrix. Each coordinate point in the point cloud data generated by the lidar is multiplied with the single rotation matrix to adjust the coordinate direction and eliminate the deviation caused by the roller tilt, generating corrected point cloud coordinates. In the corrected point cloud coordinates, the ground reference frame is set to a horizontal plane, and points with vertical coordinates close to the ground height are selected to filter out the spatial coordinates of the roller-ground contact point.

[0065] Based on the spatial coordinates of consecutive timestamps, the coordinate changes between adjacent time points are determined, and the displacement increment is generated, expressed as:

[0066] ;

[0067] in, Indicates the displacement increment. Indicates at a point in time The x-coordinate value of the spatial coordinates, Indicates at a point in time The x-coordinate value of the spatial coordinates, Indicates at a point in time The ordinate value of the spatial coordinates, Indicates at a point in time The ordinate value of the spatial coordinates.

[0068] By combining the compressive strength signal and the displacement increment, a dot product operation is performed on the compressive strength signal and the displacement increment to generate the instantaneous compaction energy value.

[0069] The spatial coordinates at each time point are paired with the corresponding instantaneous compaction energy value to establish a one-to-one correspondence, and the data is organized into a sparse physical field dataset containing spatial coordinates and instantaneous compaction energy values.

[0070] S2: Construct a cross-scale prior distribution based on a sparse physical quantity field dataset, generate partial differential constraints, embed a neural network model, define a composite loss function, and output an untrained neural network model.

[0071] S2.1: Construct a cross-scale prior distribution based on a sparse physical quantity field dataset. The specific steps are as follows:

[0072] Extract instantaneous compaction energy values ​​and corresponding spatial coordinates from sparse physical quantity field datasets, and record the energy value and horizontal and vertical coordinates on a two-dimensional plane at each time point.

[0073] The local energy gradient between adjacent coordinate points is calculated based on the instantaneous compaction energy value, and the expression is as follows:

[0074] ;

[0075] in, Represents spatial coordinates The rate of change of instantaneous compaction energy, i.e., the local energy gradient. Represents spatial coordinates The instantaneous compaction energy value at the point, Represents spatial coordinates The instantaneous compaction energy value at the point, Represents spatial coordinates The x-coordinate value, Represents spatial coordinates The x-coordinate value, Represents spatial coordinates The ordinate value, Represents spatial coordinates The ordinate value.

[0076] Organize all local energy gradient values ​​to form a gradient distribution map, and record the gradient magnitude at each spatial coordinate point; determine the maximum and minimum values ​​of the gradient values ​​in the gradient distribution map, and obtain the gradient value range based on the maximum and minimum values, and divide the gradient value range into two intervals: the gradient value region below the median of the range is marked as the low gradient region, and the gradient value region above or equal to the median is marked as the high gradient region.

[0077] The low gradient region is divided into a coarse-scale region, representing a region where the compaction energy changes gradually; the high gradient region is divided into a fine-scale region, representing a region where the compaction energy changes drastically.

[0078] Within the coarse-scale region, the gradient values ​​at each spatial coordinate point are traversed, and the arithmetic mean of all gradient values ​​is calculated to generate the gradient mean of the coarse-scale region. The sum of squares of the deviations of each gradient value from the mean is calculated and divided by the number of points to generate the gradient variance of the coarse-scale region. The gradient mean and variance of the fine-scale region are calculated in the same way as those of the coarse-scale region.

[0079] For coarse-scale and fine-scale regions, extract their respective gradient value sets, iterate through the gradient values ​​of each region one by one, construct a statistical interval by the maximum and minimum gradient values, count the frequency of gradient values ​​within the statistical interval, construct a gradient value histogram, and calculate the probability of each interval by dividing the number of times the gradient value appears in the interval by the total number of times.

[0080] The gradient mean and gradient variance of the coarse-scale region are combined with those of the fine-scale region to form gradient change characteristics, which describe the energy change patterns of both coarse-scale and fine-scale regions. Based on these gradient change characteristics, a probability density function is constructed, where the gradient mean and variance of the coarse-scale region define the probability of a smooth energy distribution, and the gradient mean and variance of the fine-scale region define the probability of a drastic energy distribution, thus generating a cross-scale prior distribution that reflects the energy distribution at both coarse-scale and fine-scale.

[0081] S2.2: Generate partial differential constraints and embed them into the neural network model. The specific process is as follows:

[0082] Based on the statistical characteristics that coarse-scale regions belong to low-gradient regions, with gradient mean lower than the overall gradient median and small gradient variance, a larger smoothness constraint parameter is set, for example, setting the smoothness constraint parameter value to twice the mean, to ensure that the energy distribution in coarse-scale regions maintains a smooth transition. Conversely, based on the characteristics that fine-scale regions belong to high-gradient regions, with gradient mean higher than or equal to the overall gradient median and large gradient variance, a smaller smoothness constraint parameter is set, for example, setting the smoothness constraint parameter value to half the mean, to allow the energy distribution in fine-scale regions to capture local drastic changes.

[0083] By utilizing the larger smoothness constraint parameters in the coarse-scale region and the smaller smoothness constraint parameters in the fine-scale region, a partial differential constraint equation is constructed. The smoothness parameters in the coarse-scale region are applied to the low-frequency components of the partial differential constraint equation to ensure gradual changes; the smoothness parameters in the fine-scale region are applied to the high-frequency components of the equation to ensure dramatic changes, thereby generating a partial differential constraint equation that reflects the multi-scale variation characteristics of compaction energy.

[0084] The structure of the fully connected neural network model is set up as follows: the input layer receives two-dimensional spatial coordinates and contains two neurons; three hidden layers are configured, each containing 128 neurons; the output layer generates the predicted compaction energy value and contains one neuron; the weights of the hidden layers are randomly initialized using a normal distribution, with a mean of 0 and a standard deviation of 0.01; according to the smoothness requirements of the partial differential equation, the hidden layer weight update rule is adjusted to limit the magnitude of weight changes, for example, by multiplying the weight update step size by the reciprocal of the smoothness constraint parameter to ensure that the energy distribution conforms to the multi-scale variation characteristics; according to the magnitude of the smoothness constraint parameter, the initialization range of the hidden layer weights is narrowed, for example, by halving the standard deviation, to generate a weight distribution that conforms to the partial differential equation, thus completing the embedding of the partial differential constraint equation.

[0085] S2.3: Define the composite loss function to output the untrained neural network model. The specific process is as follows:

[0086] The instantaneous compaction energy value and its corresponding spatial coordinates are extracted again from the sparse physical quantity field dataset. The spatial coordinates are then input into the neural network model for forward propagation to generate the predicted compaction energy value. The predicted compaction energy value is compared with the instantaneous compaction energy value, and the sum of squares of their differences is calculated to generate the mean square error loss. Based on the partial differential constraint equations, the second-order partial derivative operation is performed on the predicted compaction energy value output by the neural network model in spatial coordinates to generate the physical law loss, expressed as:

[0087] ;

[0088] in, Indicates the loss of physical laws, Represents the neural network model at spatial coordinate points The output is the predicted compaction energy value, in joules. The predicted compaction energy value is represented on the horizontal axis. The second partial derivative on, in joules per square meter. The predicted compaction energy value is represented on the horizontal axis. The second partial derivative on, in joules per square meter. Represents the smoothness constraint parameter. This represents the total number of spatial coordinate points.

[0089] The mean squared error loss and the physical law loss are summed with preset weights to generate a composite loss function. The preset weights are set based on the proportion of gradient change characteristics in coarse-scale and fine-scale regions of the gradient distribution map. The weight of the mean squared error loss is 0.4 and the weight of the physical law loss is 0.6, reflecting the balance between data fitting and physical consistency. If the weight of the mean squared error loss is higher than 0.4 and the weight of the physical law loss is lower than 0.6, the neural network model will overfit the instantaneous compaction energy value of the sparse physical quantity field dataset, which will cause the predicted compaction energy value in the unsampled area to deviate from the physical law, reducing the smoothness and consistency of the energy distribution across scales. If the weight of the mean squared error loss is lower than 0.4 and the weight of the physical law loss is higher than 0.6, the neural network model will rely too much on the partial differential equation constraint, which may lead to an increase in the deviation between the predicted compaction energy value and the actual data, reducing the model's fitting accuracy for sparse data, especially in fine-scale regions.

[0090] By attaching a composite loss function to the training objective configuration of the neural network model, the initialization structure, embedded partial differential constraint equations, and composite loss function are integrated to output an untrained neural network model.

[0091] Preferably, this method constructs a cross-scale prior distribution, combines the gradient change characteristics of coarse and fine scales, generates partial differential constraint equations, embeds them into a neural network model, defines a composite loss function, and outputs an untrained neural network model. Compared with traditional methods, this method can more accurately reflect the multi-scale distribution characteristics of compaction energy, balance the constraints of data fitting and physical laws, improve the model's generalization ability in sparse data scenarios, and enhance the smoothness and consistency of predictions in unsampled areas, especially performing better in complex terrain compaction scenarios.

[0092] S3: Uniformly sample within the two-dimensional plane of the area to be compacted to generate a set of two-dimensional coordinate points. Combine this with a sparse physical quantity field dataset to train an untrained neural network model. Minimize the composite loss function to obtain a trained neural network model.

[0093] S3.1: Uniform sampling is performed within the two-dimensional plane of the area to be compacted to generate a set of two-dimensional coordinate points. The specific process is as follows:

[0094] The two-dimensional planar boundary range of the area to be compacted is determined by the spatial coordinates of the lidar point cloud, and the minimum and maximum values ​​of the abscissa and ordinate of the area are recorded.

[0095] Based on the range of the horizontal and vertical coordinates, divide the area into uniform grid cells. For example, divide the horizontal and vertical coordinate ranges into a fixed number of intervals to generate square grid cells. Set the side length of each grid cell as the sampling interval. The fixed number is determined by the engineering precision, for example, 50 to 500.

[0096] Within each grid cell, a point is randomly selected as a sampling point. The x and y coordinates of the selected point are recorded to ensure that the sampling points are evenly distributed across the entire two-dimensional plane. The coordinates of all sampling points are then organized to generate a two-dimensional coordinate point set.

[0097] S3.2: Using a sparse physical quantity field dataset, train the untrained neural network model, minimize the composite loss function, and obtain the trained neural network model. The specific process is as follows:

[0098] Traverse the set of two-dimensional coordinate points, extract the x-coordinate and y-coordinate values ​​of each spatial coordinate, and form a two-dimensional vector. Input each two-dimensional vector sequentially into the input layer of an untrained neural network model, where the input layer contains two neurons, receiving the x-coordinate and y-coordinate values ​​respectively. In the neural network model, perform a forward propagation operation: the two-dimensional vectors of the input layer undergo a linear transformation through the weight matrix and bias vector of the first hidden layer to generate the output of the first hidden layer. Apply the ReLU activation function to the output of the first hidden layer to generate nonlinear features. Pass the nonlinear features of the first hidden layer to the second hidden layer, repeating the linear transformation and ReLU activation function processing. Continue passing to the third hidden layer, performing the same operation. Transform the output of the third hidden layer through the weights and biases of the output layer to generate a single predicted value, i.e., the iteratively predicted compaction energy value corresponding to each spatial coordinate. Record the x-coordinate, y-coordinate, and corresponding iteratively predicted compaction energy value of each spatial coordinate, establish a one-to-one correspondence, and store it as a predicted energy value dataset.

[0099] Instantaneous compaction energy values ​​and their corresponding spatial coordinates are extracted from a sparse physical quantity field dataset. Points with the same spatial coordinates in a two-dimensional coordinate point set are matched, and the iteratively predicted compaction energy values ​​are compared with the instantaneous compaction energy values. The sum of the squares of the differences between the two is calculated to generate the data error. Based on the composite loss function, the data error and the physical law loss of the partial differential constraint equation are summed with preset weights of 0.4 and 0.6 to generate a weighted loss value.

[0100] It should also be noted that the preset weights are set based on the balance principle in sparse data scenarios. Since the number of sampling points in a sparse physical quantity field dataset is less than that in traditional dense measurements, if the data loss weight is too high (greater than 0.5), the neural network model will overfit the sparse sampling points, resulting in non-physical oscillations in the unsampled areas. If the physical loss weight is too low (less than 0.5), the neural network model will satisfy the partial differential constraint, but the deviation from the actual measurement value will increase significantly.

[0101] The gradient descent algorithm is used to iteratively adjust the weights and biases of the neural network model based on the weighted loss value. In each iteration, the weights are updated to reduce the weighted loss value. The iteration is repeated until the change in the weighted loss value is less than a preset threshold, thus generating a trained neural network model.

[0102] It should also be noted that the preset threshold of 0.001 is set according to the convergence criterion of the gradient descent algorithm in the training of physical information neural networks, to ensure that the weighted loss value changes sufficiently in continuous iterations, the neural network model parameters are stable, and the computational efficiency and prediction accuracy are balanced.

[0103] S4: Divide the two-dimensional plane of the area to be compacted into a regular grid, generate the coordinates of dense grid points, input the grid point coordinates into the trained neural network model, generate an energy matrix, map it into a two-dimensional heat map, and identify abnormal areas through gradient field analysis, and output a two-dimensional compaction energy map.

[0104] S4.1: Generate dense grid point coordinates in the two-dimensional plane of the area to be compacted, and input the grid point coordinates into the trained neural network model to generate an energy matrix. The specific process is as follows:

[0105] Extract the two-dimensional planar boundary range of the area to be compacted, and record the minimum and maximum values ​​of the horizontal coordinate and the vertical coordinate.

[0106] Based on the range of the x and y coordinates, high-resolution grid cells are divided. For example, the x and y coordinate ranges are divided into smaller intervals to generate square grid cells with smaller side lengths. Ensure that the side length of the grid cells is smaller than the sampling interval during uniform sampling, for example, set to one-fifth of the sampling interval.

[0107] At the center point of each high-resolution grid cell, record the x-coordinate and y-coordinate values, organize the coordinates of all center points, and generate a set of dense grid point coordinates.

[0108] Traverse the set of coordinates of dense grid points, extract the x-coordinate and y-coordinate values ​​of each spatial coordinate, and form a two-dimensional vector.

[0109] Each two-dimensional vector is sequentially input into the input layer of the trained neural network model, and a forward propagation operation is performed. Based on the initialization structure and the embedded partial differential constraint equations, the predicted compaction energy value corresponding to each spatial coordinate is generated, and the pairing relationship between these predicted energy values ​​and the corresponding spatial coordinates is recorded.

[0110] Based on the division order of high-resolution grid cells, the row and column indices of the spatial coordinates of the center point of each grid cell in the two-dimensional coordinate system are determined, where the row index corresponds to the grid division of the vertical axis and the column index corresponds to the grid division of the horizontal axis. The predicted compaction energy value of the center point of each grid cell is assigned to the corresponding position in the two-dimensional matrix. The number of rows in the matrix is ​​equal to the number of grid cells in the vertical axis and the number of columns is equal to the number of grid cells in the horizontal axis, ensuring that each matrix element reflects the predicted compaction energy value of the corresponding grid cell, thus generating an energy matrix.

[0111] S4.2: Mapping to a two-dimensional thermogram, identifying abnormal areas through gradient field analysis, and outputting a two-dimensional compaction energy spectrum. The specific process is as follows:

[0112] Extract all compaction energy values ​​from the energy matrix, determine the maximum and minimum values, and define color mapping rules. For example, map the minimum compaction energy value to blue (RGB value 0,0,255), map the maximum compaction energy value to red (RGB value 255,0,0), and generate a green to yellow gradient for intermediate values ​​through linear interpolation.

[0113] Iterate through each element of the energy matrix, converting each compaction energy value into its corresponding RGB pixel color value according to color mapping rules. This generates a color value matrix of the same size as the energy matrix, which is then plotted as a two-dimensional heatmap. The horizontal axis of the heatmap corresponds to the matrix columns, and the vertical axis corresponds to the matrix rows. For each element of the energy matrix, calculate the gradient in both the horizontal and vertical axes. The gradient in the horizontal direction is the difference between the compaction energy value of the current element and its right-hand neighbor, divided by the grid cell edge length. The gradient in the vertical direction is the difference between the compaction energy value of the current element and its bottom neighbor, divided by the grid cell edge length. This generates a gradient field distribution containing both horizontal and vertical gradient components.

[0114] The gradient field distribution is traversed, and the absolute value of the gradient for each element is calculated. The absolute value of the gradient is compared with a preset gradient threshold, which is twice the mean of the absolute values ​​of the gradients. This threshold is set based on the statistical characteristics of abnormal areas in the gradient field distribution. It can effectively distinguish between normal and smooth areas and abnormal areas with uneven compaction, ensuring the capture of significant gradient changes while avoiding too many false positives, thus balancing detection sensitivity and accuracy. The spatial coordinates corresponding to elements whose absolute gradient values ​​exceed the preset gradient threshold are extracted and marked as areas with uneven compaction.

[0115] On the two-dimensional heat map, for each spatial coordinate of the unevenly compacted area, an outline marker is drawn, for example, the boundary of the area is outlined with a black line (RGB value 0,0,0), ensuring that the marker is clearly visible, and a two-dimensional compaction energy map containing the heat map and the outline marker is generated.

[0116] Preferably, existing methods rely on manual judgment for anomaly detection, which has limited efficiency and accuracy. This method generates high-resolution dense grid point coordinates, uses a trained neural network model to generate an energy matrix, maps it to a two-dimensional heat map, and automatically identifies uneven compaction areas through gradient field analysis, outputting a two-dimensional compaction energy spectrum. Compared to traditional inspection methods, this method combines multi-sensor sparse data and high-resolution grids to improve the precision of energy distribution; gradient field analysis automates the detection of anomalies, significantly improving the efficiency of compaction quality assessment; and the combination of heat map and contour marking visually presents compaction unevenness, enhancing engineering practicality, and is particularly suitable for compaction scenarios in complex terrain.

[0117] This embodiment also provides a multi-sensor coupled two-dimensional compaction energy spectrum generation system, including:

[0118] The acquisition module, within the area to be compacted, aligns and converts the raw data from the mechanical sensors, lidar, and inertial measurement center into a sparse physical field dataset through time synchronization and attitude correction of the inertial measurement center.

[0119] The construction module constructs a cross-scale prior distribution based on the sparse physical quantity field dataset, generates partial differential constraints, embeds a neural network model, defines a composite loss function, and outputs an untrained neural network model.

[0120] The training module performs uniform sampling within the two-dimensional plane of the area to be compacted to generate a set of two-dimensional coordinate points. Combined with the sparse physical quantity field dataset, it trains the untrained neural network model, minimizes the composite loss function, and obtains the trained neural network model.

[0121] The generation module divides the two-dimensional plane of the area to be compacted into a regular grid, generates the coordinates of dense grid points, inputs the grid point coordinates into the trained neural network model, generates an energy matrix, maps it into a two-dimensional heat map, identifies abnormal areas through gradient field analysis, and outputs a two-dimensional compaction energy spectrum.

[0122] This embodiment also provides a computer device applicable to the multi-sensor coupled compaction energy two-dimensional spectrum generation method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to realize the multi-sensor coupled compaction energy two-dimensional spectrum generation method as proposed in the above embodiment.

[0123] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0124] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the method for generating a two-dimensional compaction energy spectrum with multi-sensor coupling as proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0125] In summary, this invention overcomes the shortcomings of traditional interpolation algorithms in producing smooth transitions or artifacts in unsampled areas by constructing a cross-scale prior distribution based on a sparse physical quantity field dataset and combining it with local energy gradient distribution to divide large-scale and fine-scale regions, significantly improving the realism of compaction energy distribution. Furthermore, by capturing multi-scale characteristics through the cross-scale prior distribution, it solves the problem of static prior distribution ignoring the heterogeneity of sparse data in traditional methods, enhancing the model's adaptability to complex terrain compaction scenarios. In addition, by generating a high-resolution energy matrix and calculating the gradient field distribution, it automatically extracts compaction inhomogeneous regions where the absolute gradient value exceeds a threshold, and superimposes contour markers onto a two-dimensional heatmap, replacing the inefficient detection methods that rely on manual experience or threshold segmentation. This significantly improves the automation and efficiency of abnormal area identification, providing a precise quality assessment tool for road construction and foundation compaction.

[0126] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for generating a two-dimensional compaction energy spectrum using multi-sensor coupling, characterized in that: include, Within the area to be compacted, the raw data from the mechanical sensors, lidar, and inertial measurement center are aligned and transformed into a sparse physical field dataset through time synchronization and attitude correction at the inertial measurement center. Specifically: The system acquires pressure signals collected by mechanical sensors, point cloud data generated by lidar, and attitude data measured by inertial measurement unit. After timestamp alignment, it uses the pitch and roll angles of the attitude data to perform geometric correction on the point cloud data and extract the spatial coordinates of the contact point between the roller and the ground. The displacement increment is calculated based on the compressive force signal and spatial coordinates. The dot product operation of force and displacement is performed to generate the instantaneous compaction energy value, which is then correlated with the spatial coordinates to obtain a sparse physical quantity field dataset. Based on the sparse physical quantity field dataset, a cross-scale prior distribution is constructed, partial differential constraints are generated, a neural network model is embedded, a composite loss function is defined, and an untrained neural network model is output. Uniform sampling is performed in the two-dimensional plane of the area to be compacted to generate a set of two-dimensional coordinate points. Combined with the sparse physical quantity field dataset, the untrained neural network model is trained, and the composite loss function is minimized to obtain the trained neural network model. The construction of the cross-scale prior distribution based on the sparse physical quantity field dataset specifically involves: Extract instantaneous compaction energy values ​​and corresponding spatial coordinates from a sparse physical field dataset; Calculate the local energy gradient between adjacent coordinate points based on the instantaneous compaction energy value, generate a gradient distribution map, divide the region into coarse-scale and fine-scale regions, extract the gradient change characteristics of each scale region, and construct a cross-scale prior distribution. The generation of partial differential constraints, embedded in a neural network model, specifically involves: By utilizing cross-scale prior distributions, smoothness constraint parameters at both coarse and fine scales are determined, and partial differential constraint equations reflecting the multi-scale characteristics of compaction energy are constructed. Initialize the neural network model, embed the partial differential constraint equations into the weight update rules of the neural network model, and complete the embedding; A regular grid is divided in the two-dimensional plane of the area to be compacted to generate dense grid point coordinates. The grid point coordinates are input into the trained neural network model to generate an energy matrix, which is mapped to a two-dimensional heat map. Abnormal areas are identified through gradient field analysis, and a two-dimensional compaction energy map is output.

2. The method for generating a two-dimensional compaction energy spectrum using multi-sensor coupling as described in claim 1, characterized in that: The definition of the composite loss function, which outputs the untrained neural network model, is as follows: Extract instantaneous compaction energy values ​​and corresponding spatial coordinates from a sparse physical field dataset; Spatial coordinates are input into the neural network model for forward propagation to generate the corresponding predicted compaction energy value. The data mean square error loss is then calculated by combining the instantaneous compaction energy value. Based on the partial differential constraint equations, the second derivative of the predicted compaction energy value output by the neural network model is calculated in spatial coordinates. After generating the physical law loss, it is fused with the data mean square error loss to generate a composite loss function, thus obtaining the untrained neural network model.

3. The method for generating a two-dimensional compaction energy spectrum using multi-sensor coupling as described in claim 1, characterized in that: The process of uniformly sampling within the two-dimensional plane of the area to be compacted to generate a set of two-dimensional coordinate points refers to determining the boundary range of the two-dimensional plane of the area to be compacted, dividing it into uniform grid cells according to the boundary range, setting the sampling interval of the grid cells, and randomly selecting sampling points within each grid cell.

4. The method for generating a two-dimensional compaction energy spectrum using multi-sensor coupling as described in claim 1, characterized in that: The process of training the untrained neural network model by combining the sparse physical quantity field dataset and minimizing the composite loss function to obtain the trained neural network model is as follows: The two-dimensional coordinate point set is input into an untrained neural network model to generate iteratively predicted compaction energy values. The instantaneous compaction energy values ​​of the sparse physical quantity field dataset are then used to calculate the data error between the iteratively predicted compaction energy values ​​and the actual energy values. Calculate the weighted loss value of data error and partial differential constraint based on the composite loss function; The gradient descent algorithm is used to iteratively adjust the parameters of the neural network model and minimize the weighted loss value. When the weighted loss value converges, the trained neural network model is generated.

5. The method for generating a two-dimensional compaction energy spectrum using multi-sensor coupling as described in claim 1, characterized in that: The process involves dividing the two-dimensional plane of the area to be compacted into a regular grid, generating dense grid point coordinates, and inputting these grid point coordinates into a trained neural network model to generate an energy matrix. Specifically: Based on the two-dimensional planar boundary range of the area to be compacted, high-resolution grid cells are divided, dense grid point coordinates are generated and input into the trained neural network model to calculate the predicted compaction energy value of each coordinate point. Based on the predicted compaction energy value, high-resolution grid cells are filled to generate an energy matrix.

6. The method for generating a two-dimensional compaction energy spectrum using multi-sensor coupling as described in claim 1, characterized in that: The mapping is a two-dimensional thermogram, and abnormal areas are identified through gradient field analysis, outputting a two-dimensional compaction energy spectrum, specifically: Based on the compaction energy value of the energy matrix, the color mapping rule is determined, and the compaction energy value of the energy matrix is ​​converted into pixel color value to generate a two-dimensional heat map; The spatial gradient of the energy matrix is ​​calculated to generate a gradient field distribution, and regions whose absolute gradient values ​​exceed a preset gradient threshold are extracted and identified as areas of uneven compaction. The contour markers of the unevenly compacted areas are superimposed on the two-dimensional thermogram to generate a two-dimensional compaction energy spectrum.

7. A multi-sensor coupled two-dimensional compaction energy spectrum generation system, based on the multi-sensor coupled two-dimensional compaction energy spectrum generation method according to any one of claims 1 to 6, characterized in that: include, The acquisition module, within the area to be compacted, aligns and converts the raw data from the mechanical sensors, lidar, and inertial measurement center into a sparse physical field dataset through time synchronization and attitude correction of the inertial measurement center. The module constructs a cross-scale prior distribution based on the sparse physical quantity field dataset, generates partial differential constraints, embeds a neural network model, defines a composite loss function, and outputs an untrained neural network model. The training module performs uniform sampling within the two-dimensional plane of the area to be compacted to generate a set of two-dimensional coordinate points. Combined with the sparse physical quantity field dataset, it trains the untrained neural network model, minimizes the composite loss function, and obtains the trained neural network model. The generation module divides the two-dimensional plane of the area to be compacted into a regular grid, generates the coordinates of dense grid points, inputs the grid point coordinates into the trained neural network model, generates an energy matrix, maps it into a two-dimensional heat map, identifies abnormal areas through gradient field analysis, and outputs a two-dimensional compaction energy spectrum.

Images