A method for dynamic correlation analysis of roadbed compaction trajectory and compaction quality

Abstract

This invention relates to the field of big data analysis and processing technology, and discloses a method for dynamic correlation analysis of roadbed compaction trajectory and compaction quality. The method includes: Step 1: collecting the position information, attitude information, and vibration response information of the road roller during the roadbed compaction process to obtain a set of hidden state samples and a set of observation sequence samples of corresponding partial trajectory analysis units; Step 2: executing a joint dynamic correlation analysis algorithm based on Hidden Markov Model (HMM) and manifold learning; Step 3: using the joint dynamic correlation analysis model to process the observation sequence of real-time trajectory analysis units constructed from the real-time collected data, and combining it with the compaction manifold embedded coordinates to generate a dynamic compaction quality grade distribution map covering the roadbed construction area. This invention not only improves the continuity and stability of compaction quality assessment, but also enhances the responsiveness to changes in construction conditions.

Description

Technical Field

[0001] This invention relates to the field of data processing and analysis technology, and in particular to a method for dynamic correlation analysis of roadbed compaction trajectory and compaction quality. Background Technology

[0002] In road construction, the compaction quality of the subgrade directly affects the bearing capacity, deformation performance, and long-term service condition of the pavement layer. Therefore, accurate assessment and dynamic monitoring of the compaction process during construction has always been a key focus in the engineering field. Traditional methods for obtaining subgrade compaction quality data rely primarily on on-site sampling using compaction quality testing devices. This involves measuring dry density or degree of compaction at several sampling points after construction is completed to determine the compaction quality grade at each point. While this method is widely used in engineering practice and its results can serve as the final quality acceptance criterion, the limited number of testing points often fails to reflect the overall quality status of the entire construction area. During compaction, variations in roller speed, vibration amplitude, vibration frequency, and the coverage path of each pass all contribute to a spatially uneven distribution of compaction quality, making it difficult to accurately capture this variation using only a small number of testing points.

[0003] To overcome the shortcomings of traditional methods, the industry has begun to introduce sensor data from road rollers to digitally record the construction process. For example, positioning devices installed on the machine acquire the travel trajectory, inertial measurement units acquire attitude changes, and vibration response sensors record vibration acceleration or energy. This data reflects the operating status of the road roller at each trajectory point and can be used to infer the compaction process at different locations. Existing technologies attempt to statistically correlate these operating parameters with compaction quality test data, for example, by setting thresholds or establishing empirical formulas to map combinations of speed, vibration amplitude, and vibration frequency to compaction quality grades. These methods can achieve visualization of compaction quality to some extent, but because they rely on manually set rules or simple linear relationships, they often cannot adapt to complex and dynamically changing construction conditions and are less sensitive to abnormal situations and complex working conditions. When the operating state of the construction machinery changes, such as speed fluctuations, changes in the energy output of the vibration system, or changes in the moisture content of the subgrade, models based on fixed thresholds are prone to misjudgment. Summary of the Invention

[0004] The purpose of this invention is to provide a method for dynamic correlation analysis of roadbed compaction trajectory and compaction quality. By constructing a trajectory analysis unit observation sequence, a hidden state sample set, and a compaction feature vector set, the method integrates roller operating parameters, compaction quality detection data, and compaction behavior characteristics into the analysis process. This involves sequentially completing hidden Markov model training, compaction manifold embedding space construction, and joint iterative correction of both, thereby obtaining a joint dynamic correlation analysis model that simultaneously reflects the temporal evolution of the compaction process and the spatial similarity structure of compaction behavior. In real-time applications, this invention uses real-time observation symbol sequences to infer the final hidden state number and forms a compaction quality region based on the compaction manifold embedding coordinates, further generating a dynamic compaction quality level distribution map covering the construction area. Therefore, this invention not only improves the continuity and stability of compaction quality assessment but also enhances the responsiveness to changes in construction conditions, achieving refined, real-time, and full-area dynamic monitoring of the compaction process, providing a more reliable basis for work path planning and compaction pass control.

[0005] To solve the above-mentioned technical problems, this invention provides a method for dynamic correlation analysis of roadbed compaction trajectory and compaction quality, the method comprising:

[0006] Step 1: Collect the position, attitude, and vibration response information of the road roller during the roadbed compaction process, divide the roadbed construction area into trajectory analysis units, and construct a trajectory analysis unit observation sequence containing the road roller's operating parameters; obtain compaction quality test data, and establish the correspondence between compaction quality grade and latent state to obtain the latent state sample set and the corresponding partial trajectory analysis unit observation sequence sample set;

[0007] Step 2: Execute the joint dynamic correlation analysis algorithm based on Hidden Markov Model (HMM) and manifold learning, including:

[0008] Step 2.1: Initially train the Hidden Markov Model (HMM) using the hidden state sample set and the partial trajectory analysis unit observation sequence sample set to obtain the initial HMM model.

[0009] Step 2.2: Construct the compaction manifold embedding space based on the compaction feature vectors of all trajectory analysis units, and obtain the compaction manifold embedding coordinates of each trajectory analysis unit;

[0010] Step 2.3: Perform joint iterative correction on the initial model of the Hidden Markov Model (HMM) and the embedded space of the compacted manifold to obtain a joint dynamic correlation analysis model containing the final hidden state number of the convergence.

[0011] Step 3: Using the joint dynamic correlation analysis model, the observation sequence of the real-time trajectory analysis unit constructed from the real-time collected data is processed, and combined with the compaction manifold embedded coordinates, a dynamic compaction quality grade distribution map covering the roadbed construction area is generated.

[0012] Furthermore, the process of constructing the trajectory analysis unit observation sequence in step 1 specifically includes: collecting position information using a positioning device, collecting attitude information using an attitude detection device, and collecting vibration response information using a vibration response acquisition device; dividing the roadbed construction area into several trajectory analysis units with a side length of one meter on the plane, and assigning a unique number to each trajectory analysis unit; assigning the position information at each acquisition time to the corresponding trajectory analysis unit according to the plane coordinates; assigning a compaction pass number to each complete round-trip compaction process, and considering sampling points belonging to the same trajectory analysis unit and having the same compaction pass number as observation segments of the same compaction pass on the corresponding trajectory analysis unit; for each trajectory analysis unit, arranging the observation segments corresponding to each compaction pass on the corresponding trajectory analysis unit in sequence according to the acquisition time, forming a trajectory analysis unit observation sequence with a length of not less than five; each observation segment includes the average driving speed, average vibration amplitude, average vibration frequency, and marking information of the corresponding compaction pass on the corresponding trajectory analysis unit.

[0013] Furthermore, in step 1, in the trajectory analysis unit with compaction quality test data, the compaction quality corresponding to the trajectory analysis unit is divided into three compaction quality levels: insufficient compaction level, normal compaction level, and high compaction level, based on the numerical range of the compaction quality test data. A one-to-one correspondence is established between the three compaction quality levels and the latent states of insufficient compaction, normal compaction, and high compaction, respectively, to obtain the latent state sample set and the partial trajectory analysis unit observation sequence sample set corresponding to the latent state sample set.

[0014] Furthermore, the initial training process of the Hidden Markov Model in step 2.1 specifically includes: selecting no fewer than fifty trajectory analysis units with compaction quality detection data as training samples; for each training sample, dividing each observation segment in its trajectory analysis unit observation sequence into one of three speed interval codes according to the average driving speed, one of three vibration amplitude interval codes according to the average vibration amplitude, and one of three frequency interval codes according to the average vibration frequency; concatenating the three interval codes in a fixed order to form discrete observation symbols; obtaining the observation symbol sequence sequentially for the observation segments in the same trajectory analysis unit observation sequence; in each training sample, within the same trajectory analysis unit observation sequence, counting the number of transitions between adjacent hidden states in the order of their hidden states, and for each... The combination of the initial hidden state and the target hidden state is used to divide the number of transitions of the combination by the total number of transitions of the corresponding initial hidden state, resulting in a hidden state transition probability table with three rows and three columns. The sum of the three values ​​in each row of the table is rounded to three decimal places and equals one. In all training samples, the first hidden state of the observation sequence of each trajectory analysis unit is taken as the initial hidden state. The occurrence frequency of each of the three hidden states is counted, and the occurrence frequency of each hidden state is divided by the total number of training samples, resulting in a hidden state initial probability table with three rows and one column. In each training sample, for each hidden state, the occurrence frequency of each discrete observation symbol in the corresponding hidden state is counted, and the occurrence frequency of each discrete observation symbol is divided by the sum of the occurrence frequency of all observation symbols in the corresponding hidden state, resulting in three observation output probability tables with several columns.

[0015] Furthermore, the initial training process for the Hidden Markov Model (HMM) includes: using the initial probability table of hidden states, the hidden state transition probability table, and the observation output probability table as the initial parameters of the HMM, and using the observation symbol sequence of all training samples as input; in each iteration, performing forward and backward computations on each observation symbol sequence in chronological order, with the forward computation calculating the forward probabilities of the three hidden states sequentially at each time step, and the backward computation calculating the backward probabilities of the three hidden states sequentially from the last observation time step backward; then updating the probability values ​​in the initial probability table of hidden states, the hidden state transition probability table, and the observation output probability table based on the forward and backward probabilities, retaining all probability values ​​to three decimal places after each update; stopping the iteration when the change in any value in the hidden state transition probability table is less than 0.001 in two consecutive iterations or when the number of iterations reaches fifty, thus obtaining the initial model of the Hidden Markov Model (HMM).

[0016] Furthermore, the process of constructing the compaction manifold embedding space in step 2.2 specifically includes: for each trajectory analysis unit, calculating the average driving speed, average vibration amplitude, average vibration frequency, and cumulative dwell time of all observed segments in its trajectory analysis unit observation sequence, and forming a compaction feature vector in the compaction feature vector set in a fixed order; for each feature in the compaction feature vector set, calculating the average value and standard deviation of each feature in all trajectory analysis units, and then using the calculated average value and standard deviation to linearly scale each feature of each trajectory analysis unit so that the average value of each feature of all trajectory analysis units is zero and the standard deviation is one.

[0017] Furthermore, the process of constructing the compacted manifold embedding space also includes: for any two trajectory analysis units, calculating the Euclidean distance according to the scaled values ​​of the four features, and storing the Euclidean distances between all trajectory analysis units in the distance matrix; for each trajectory analysis unit, selecting the five different trajectory analysis units with the smallest distance in the distance matrix as neighboring trajectory analysis units, and constructing an adjacency relationship set containing all trajectory analysis units; on the adjacency relationship set, using the shortest path algorithm, calculating the shortest path distance through the adjacency relationship set for each pair of trajectory analysis units, and storing all shortest path distances in the shortest path distance matrix; performing dimensionality reduction processing on the shortest path distance matrix, specifically, correcting the shortest path distance matrix by row mean and column mean to obtain a symmetric matrix, then performing eigenvalue decomposition on the symmetric matrix, selecting the two eigenvectors corresponding to the two largest eigenvalues, and using the components of each trajectory analysis unit in the two eigenvectors as the first coordinate and second coordinate of the corresponding trajectory analysis unit in the compacted manifold embedding space, respectively, to obtain the compacted manifold embedding coordinates of each trajectory analysis unit.

[0018] Furthermore, in step 2.3, the joint iterative correction process of the Hidden Markov Model (HMM) and the compacted manifold embedding space specifically includes: for all obtained trajectory analysis units, the initial model of the Hidden Markov Model (HMM) is adopted, and Viterbi decoding is performed on the observation symbol sequence of each trajectory analysis unit. Specifically, at the first observation time, the initial scores of the three hidden states are calculated respectively, and at the second and subsequent observation times, the scores of the three hidden states at the current time are calculated sequentially. At each observation time, the path number of the hidden state with the largest score from the previous time is recorded. At the last observation time, the hidden state number with the largest score is selected as the termination hidden state number of the observation sequence. Then, the hidden state number at each time is obtained by backtracking along the recorded hidden state path number. The hidden state number at the last observation time is used as the final hidden state number of the corresponding trajectory analysis unit.

[0019] Furthermore, the joint iterative correction process also includes: initializing the iteration count to one; in each iteration, for each trajectory analysis unit, calculating the Euclidean distance between the trajectory analysis unit and other trajectory analysis units in the compacted manifold embedding space; selecting the five trajectory analysis units with the smallest Euclidean distance as the neighboring trajectory analysis units in the compacted manifold embedding space; counting the occurrences of the three hidden state numbers in the final hidden state numbers of the five neighboring trajectory analysis units; if the occurrence count of a certain hidden state number is not less than three and greater than the current final hidden state number of the corresponding trajectory analysis unit... The number of times the hidden state number appears in the neighborhood trajectory analysis unit is calculated, and the final hidden state number of the corresponding trajectory analysis unit is replaced with the hidden state number that appears most frequently. After one iteration, the ratio of the number of trajectory analysis units where the final hidden state number is replaced to the total number of trajectory analysis units is calculated. If the calculated ratio is less than 0.01, or the iteration count reaches 10, the iteration is stopped, and the converged final hidden state number of each trajectory analysis unit is obtained. The initial model of the Hidden Markov Model (HMM), the converged final hidden state number, and the compacted manifold embedded coordinates are used together as the joint dynamic correlation analysis model.

[0020] Further, step 3 specifically includes: during the roadbed compaction construction process, real-time acquisition of new roadbed compaction trajectory data, attitude data, and vibration response data, and construction of a real-time trajectory analysis unit observation sequence; for the real-time observation symbol sequence of each trajectory analysis unit, Viterbi decoding is performed using the Hidden Markov Model (HMM) in the joint dynamic correlation analysis model to obtain the real-time final hidden state number; and using the compaction manifold embedding space in the joint dynamic correlation analysis model, several trajectory analysis units with a distance less than a set distance threshold between the compaction manifold embedding coordinates and the same final hidden state number are divided into the same compaction quality region; the insufficiently compacted hidden state, the normal compacted hidden state, and the high-compaction hidden state are respectively corresponding to the unqualified compaction quality grade, the qualified compaction quality grade, and the excellent compaction quality grade, and the compaction quality grade is displayed in color on the plane according to the spatial position of the trajectory analysis unit to generate a dynamic compaction quality grade distribution map covering the entire roadbed construction area; and the dynamic compaction quality grade distribution map is output as the result of the dynamic correlation analysis between the roadbed compaction trajectory and the compaction quality, which is used to guide the adjustment of the roller operation path and the control of the compaction passes.

[0021] The present invention provides a dynamic correlation analysis method for roadbed compaction trajectory and compaction quality, which has the following beneficial effects: The present invention constructs a hidden state sample set, a trajectory analysis unit observation sequence sample set, and a compaction feature vector set, simultaneously incorporating roller operating parameters, compaction quality detection data, and compaction behavior characteristics into a unified analysis system. Based on this, a joint dynamic correlation analysis model of a hidden Markov model and a compaction manifold embedding space is established, enabling the judgment of compaction quality to simultaneously depend on the temporal evolution of the compaction process and the spatial similarity of compaction characteristics. The present invention utilizes discrete observation symbols to reflect the combined changes in velocity, vibration amplitude, and vibration frequency during each compaction pass, allowing the hidden Markov model to accurately capture the state transition trends of the compaction stage. By standardizing compaction feature vectors, constructing an adjacency relationship set, and calculating the shortest path distance, the compaction manifold embedding coordinates generated by the present invention can preserve the structure of compaction behavior in a high-dimensional feature space, allowing similar compaction behaviors to naturally cluster in this space. The joint iterative correction process utilizes the compaction manifold embedded coordinates to spatially consistently adjust the latent state results, automatically correcting local anomalies and significantly improving the continuity and stability of compaction state determination. Furthermore, in the real-time application phase, this invention can rapidly generate real-time trajectory analysis unit observation sequences based on real-time collected roller operating parameters. It then infers the final real-time latent state number through a joint dynamic correlation analysis model, combining the compaction manifold embedded coordinates with a set distance threshold to form a compaction quality region. This allows the dynamic compaction quality level distribution map to reflect the latest state of the construction site. Compared to traditional methods that rely solely on detection point data or infer compaction quality based on a single model, this invention simultaneously achieves temporal analysis, spatial aggregation, and dynamic updating of the compaction process, improving the comprehensiveness, accuracy, and real-time nature of compaction quality evaluation. This provides a more reliable reference for roller operation path adjustment, compaction pass control, and regional quality management. Attached Figure Description

[0022] Figure 1 This is a visualization diagram of the hidden state transition probability matrix of a hidden Markov model provided in an embodiment of the present invention.

[0023] Figure 2 This is a schematic diagram of the optimal path for Viterbi decoding of a Hidden Markov Model provided in an embodiment of the present invention. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] A method for dynamic correlation analysis of roadbed compaction trajectory and compaction quality, the method comprising:

[0026] Step 1: Collect the position, attitude, and vibration response information of the road roller during the roadbed compaction process, divide the roadbed construction area into trajectory analysis units, and construct a trajectory analysis unit observation sequence containing the road roller's operating parameters; obtain compaction quality detection data, and establish the correspondence between compaction quality level and latent state to obtain the latent state sample set and the corresponding partial trajectory analysis unit observation sequence sample set.

[0027] In one implementation, step 1 involves installing a positioning device, an attitude detection device, and a vibration response acquisition device on the road roller. The positioning device can be a receiver supporting a Global Positioning System (GPS), with a sampling period of 0.1 seconds. This ensures that when the road roller travels at approximately 1 meter per second, the displacement between adjacent sampling moments is approximately 0.1 meters. This allows for subsequent division into trajectory analysis units with 1-meter sides, where each unit typically contains multiple sampling points, facilitating the statistical analysis of road roller operating parameters such as average travel speed. The attitude detection device can be an inertial measurement unit (IMU) used to acquire attitude information such as the road roller's pitch and roll angles within the same sampling period. The vibration response acquisition device can be installed near the vibrating wheel to collect vibration acceleration data or vibration energy data along the vertical direction. Its sampling period can be consistent with that of the positioning device for alignment on the time axis.

[0028] Before subgrade construction, the subgrade construction area is divided into several trajectory analysis units in a plane coordinate system according to the designed subgrade width and length. Preferably, the trajectory analysis unit can be set as a square unit with a side length of 1 meter. Each trajectory analysis unit corresponds to a unique row number and column number in the coordinate plane, and is further assigned a unique number. After converting the roller position information into plane coordinates, for each acquisition time, the plane coordinates of the acquisition time are compared with the boundary range of the trajectory analysis unit to determine which trajectory analysis unit the acquisition point falls into, and the roller position information, attitude information, and vibration response information at that acquisition time are assigned to the corresponding trajectory analysis unit. By dividing the data according to trajectory analysis units, the continuous compaction trajectory can be discretized into units with clear spatial boundaries, which facilitates statistical analysis within each trajectory analysis unit and also facilitates the direct alignment of subsequent compaction quality test data to the corresponding trajectory analysis unit.

[0029] In terms of construction organization, the round trip of the road roller along the same route can be considered as a complete compaction round trip. Each complete compaction round trip is assigned a compaction pass number; for example, the compaction pass number for the first round trip is 1, the compaction pass number for the second round trip is 2, and so on. For the same trajectory analysis unit, all sampling points located within that unit and with the same compaction pass number are selected. These sampling points correspond to the compaction process of the same compaction pass within that trajectory analysis unit. By statistically analyzing the position, attitude, and vibration response information of these sampling points, the average travel speed, average vibration amplitude, and average vibration frequency of that compaction pass within the trajectory analysis unit can be calculated. For example, if 20 sampling points are collected within a trajectory analysis unit at compaction pass number 3, the travel distance within the corresponding time interval can be calculated using the temporal order of these 20 sampling points, thus obtaining the average travel speed within that trajectory analysis unit. Simultaneously, the vibration response information of these 20 sampling points is averaged to obtain the average vibration amplitude and average vibration frequency. To align with subsequent compaction quality testing data, marker information can be added to each observation segment to indicate whether compaction quality testing data exists in that compaction pass and that trajectory analysis unit.

[0030] For each trajectory analysis unit, the observation segments of each compaction pass belonging to that trajectory analysis unit are arranged in chronological order of acquisition time to form a trajectory analysis unit observation sequence. Preferably, to ensure that the trajectory analysis unit observation sequence can reflect the entire process of the gradual evolution of compaction quality with each compaction pass, the length of the trajectory analysis unit observation sequence can be specified to be no less than 5, that is, at least 5 compaction passes must be completed on the same trajectory analysis unit and the corresponding observation segments must be acquired. The advantage of this length setting is that it can cover multiple stages from the initial uncompacted state, through gradual compaction, to a near-stable compaction state, so that subsequent state identification based on the trajectory analysis unit observation sequence can more completely reflect the dynamic correlation between changes in roller operating parameters and changes in compaction quality.

[0031] For compaction quality testing, testing points can be arranged at predetermined intervals within the roadbed construction area. For example, one testing point can be arranged every 10 meters along the direction of travel, and a row of testing points can be arranged every 2 meters along the lateral direction. At each testing point, compaction quality testing equipment is used to acquire compaction quality test data, such as measuring dry density or degree of compaction. After converting the spatial position of each testing point into planar coordinates, these planar coordinates are matched to the corresponding trajectory analysis unit, so that each set of compaction quality test data is associated with a unique trajectory analysis unit. In some implementations, if a testing point is located near the boundary of two adjacent trajectory analysis units, the compaction quality test data can be assigned to the closer trajectory analysis unit based on the distance between the testing point and the geometric center of the adjacent trajectory analysis unit to avoid the same test data being reused.

[0032] To establish the correspondence between compaction quality grades and latent states, compaction quality can be classified according to the numerical range of compaction quality test data. For example, in one implementation, when the compaction quality test data is less than 90% of the design target value, the compaction quality of the corresponding trajectory analysis unit is classified as insufficiently compacted; when the compaction quality test data is between 90% and 110% of the design target value, the compaction quality of the corresponding trajectory analysis unit is classified as normal compaction; and when the compaction quality test data is greater than 110% of the design target value, the compaction quality of the corresponding trajectory analysis unit is classified as high compaction. This classification method converts continuous test values ​​into discrete compaction quality grades, facilitating the establishment of a one-to-one correspondence with discrete latent states. Subsequently, a fixed correspondence is established between the insufficiently compacted, normal, and high compaction grades and the insufficiently compacted, normal, and high compaction latent states, respectively, thereby obtaining a latent state sample set. The trajectory analysis unit observation sequence of the trajectory analysis unit with compaction quality test data is then used as the corresponding partial trajectory analysis unit observation sequence sample set. The establishment of this correspondence enables the inference of the hidden state of the trajectory analysis unit even when relying solely on the observation sequence of the trajectory analysis unit, thereby indirectly estimating the compaction quality level.

[0033] In another alternative implementation, the side length of the trajectory analysis unit can be set to 0.5 meters to improve spatial resolution. In this case, the sampling period for the roller position information can be shortened to 0.05 seconds to ensure that each trajectory analysis unit still has a sufficient number of sampling points for statistical analysis of average travel speed and vibration response information. For acquiring compaction quality test data, a random sampling method can also be used. At the end of each work shift, no fewer than 30 trajectory analysis units are randomly selected within the construction area of ​​that shift for compaction quality testing, and the test point locations are matched to the corresponding trajectory analysis units in the same manner. Regardless of whether a regular layout or random sampling is used, as long as each trajectory analysis unit with compaction quality test data can establish a corresponding compaction quality level and latent state, a latent state sample set and a partial trajectory analysis unit observation sequence sample set can be constructed, providing a training basis for subsequent steps.

[0034] Step 2: Execute the joint dynamic correlation analysis algorithm based on Hidden Markov Model (HMM) and manifold learning, including:

[0035] Step 2.1: Initially train the Hidden Markov Model (HMM) using the hidden state sample set and the partial trajectory analysis unit observation sequence sample set to obtain the initial HMM model.

[0036] Step 2.2: Construct the compaction manifold embedding space based on the compaction feature vectors of all trajectory analysis units, and obtain the compaction manifold embedding coordinates of each trajectory analysis unit;

[0037] Step 2.3: Perform joint iterative correction on the initial model of the Hidden Markov Model (HMM) and the embedded space of the compacted manifold to obtain a joint dynamic correlation analysis model containing the convergent final hidden state number.

[0038] In one implementation, after constructing the hidden state sample set and the partial trajectory analysis unit observation sequence sample set in step 1, a joint dynamic correlation analysis algorithm based on Hidden Markov Model (HMM) and manifold learning is executed. The overall idea of ​​this algorithm is as follows: First, an initial HMM model capable of characterizing the temporal evolution of the compaction process is obtained by training the hidden state sample set and the partial trajectory analysis unit observation sequence sample set. Then, a compaction manifold embedding space that maintains the geometric structure relationship of compaction behavior is constructed based on the compaction feature vectors of all trajectory analysis units. Finally, feedback iteration between the initial HMM model and the compaction manifold embedding space is used to make the hidden state evolution on the same time series consistent with the compaction behavior between adjacent trajectory analysis units, thereby obtaining a joint dynamic correlation analysis model containing the convergent final hidden state number.

[0039] In the initial training of the Hidden Markov Model (HMM), at least 50 trajectory analysis units with compaction quality detection data are selected as training samples. The sample size is set to at least 50 because the hidden state sample set includes three types of hidden states: insufficient compaction, normal compaction, and high compaction. If the sample size is too small, the occurrence frequency of each hidden state will be insufficient, leading to a large deviation between the statistically derived transition relationship and the observed output relationship, which is not conducive to the stable characterization of the dynamic changes in compaction quality. For each training sample, the continuous observation segments in its trajectory analysis unit observation sequence need to be converted into discrete observation symbols. Specifically, based on engineering experience and common ranges in historical construction data, the average driving speed can be divided into three speed ranges (low speed, medium speed, and high speed), the average vibration amplitude into three vibration amplitude ranges (low vibration amplitude, medium vibration amplitude, and high vibration amplitude), and the average vibration frequency into three frequency ranges (low frequency, medium frequency, and high frequency). For example, an average driving speed of 0.8 m / s or less can be classified as a low-speed range, a speed greater than 0.8 m / s but less than or equal to 1.2 m / s as a medium-speed range, and a speed greater than 1.2 m / s as a high-speed range. Similarly, three ranges for vibration amplitude and vibration frequency can be set based on equipment specifications and test data. The reason for dividing into three ranges is that the sensitivity of compaction effect to speed and vibration parameters can generally be categorized into three levels: low, moderate, and high. Using a three-level discrete encoding preserves the coarse-grained variation trend related to compaction effect while avoiding excessively sparse distribution of training samples across ranges due to too many ranges. The speed range code, vibration amplitude range code, and vibration frequency range code for each observation segment are concatenated in a fixed order to form a discrete observation symbol, thus converting each trajectory analysis unit observation sequence into a sequence of discrete observation symbols. For example, in a trajectory analysis unit observation sequence of length 8, 8 discrete observation symbols can be obtained.

[0040] Based on the constructed discrete observation symbol sequence, and combined with the correspondence between compaction quality grades and latent states established in step 1, the latent state at the corresponding time can be directly determined at the observation location with compaction quality detection data. Within the observation sequence of the same trajectory analysis unit, the combinations of adjacent latent states are statistically analyzed in chronological order, and the number of transitions from each initial latent state to the three target latent states is accumulated. For example, in a certain training sample, if the transition from the insufficiently compacted latent state to the normal compacted latent state occurs 5 times, while the transition from the insufficiently compacted latent state to the high-compaction latent state occurs only once, it can be seen that in this sample, the insufficiently compacted latent state is more likely to evolve into the normal compacted latent state. The number of transitions in all training samples is merged, and the number of transitions in each row indexed by the initial latent state is normalized, that is, the number of transitions from each initial latent state to each target latent state is divided by the sum of the number of transitions from that initial latent state, resulting in a latent state transition probability table containing 3 rows and 3 columns. The sum of the three values ​​in each row of the table is rounded to three decimal places and equals 1. This ensures that for any current hidden state, the change of the hidden state at the next time step can always be explained according to three possible transition directions.

[0041] To determine the initial probabilities of hidden states in a Hidden Markov Model (HMM), it is necessary to statistically analyze the distribution of the first hidden state in the observation sequence of each trajectory analysis unit. In the training samples, the occurrence frequency of each of the three hidden states as the first hidden state is counted. Then, the occurrence frequency of each hidden state is divided by the total number of training samples, resulting in a hidden state initial probability table with 3 rows and 1 column. This setup reflects the overall distribution tendency of each hidden state at the beginning of roadbed compaction; for example, most road sections begin in an insufficiently compacted hidden state, while reworked sections may begin directly in a normally compacted hidden state. Further, in each training sample, for each hidden state, the occurrence frequency of each discrete observation symbol in that hidden state is counted. Each hidden state is then normalized, and the occurrence frequency of each discrete observation symbol in that hidden state is divided by the sum of the occurrence frequencies of all observation symbols in that hidden state, resulting in three observation output probability tables with several columns. The observed output probability table describes the probability distribution of combinations of average travel speed, average vibration amplitude, and average vibration frequency of the road roller under different hidden states, thereby establishing the correspondence between the combination of road roller operating parameters and the hidden state of compaction quality.

[0042] After constructing the initial probability table of hidden states, the probability table of hidden state transitions, and the probability table of observations and outputs, these are used as the initial parameters of the Hidden Markov Model (HMM). The observed symbol sequences of all training samples are used as input to perform an iterative training process. In each iteration, for each observed symbol sequence, forward computation is performed in chronological order, starting from the first observation time and calculating the forward probability of being in one of the three hidden states at each time step. Intuitively, the forward probability represents the likelihood of a certain hidden state being the current state, given the prefix of the current observed symbol sequence. Correspondingly, backward computation is performed starting from the last observation time step, calculating the backward probability of being in one of the three hidden states at each time step and subsequently generating the remaining observed symbol sequences. After both forward and backward probabilities are calculated, the expected number of occurrences of each hidden state at each time step and the expected number of occurrences of each hidden state transition combination at adjacent time steps can be estimated based on the probability product relationship. These expected values ​​are accumulated and normalized over all training samples to update the values ​​in the initial hidden state probability table, the hidden state transition probability table, and the observation output probability table. After each update, all probability values ​​are retained to three decimal places to avoid excessive accumulation of tail differences. Through repeated forward and backward calculations, the Hidden Markov Model (HMM) gradually converges to a parameter combination that better reflects the statistical regularity of the training data. Iteration stops when the change in any value in the hidden state transition probability table is less than 0.001 in two consecutive iterations, or when the number of iterations reaches 50. The initial hidden state probability table, the hidden state transition probability table, and the observation output probability table at this point are used as the initial model of the HMM. In another optional implementation, the allowable change threshold can be set to 0.0005 or the maximum number of iterations can be set to 100 to accommodate construction scenarios with larger data volumes or higher convergence requirements.

[0043] In one implementation, Figure 1 This section presents the hidden state transition probability matrix A in the initial Hidden Markov Model (HMM) obtained through step 2.1 training. This 3x3 square matrix characterizes the probability distribution of the evolution of the compaction mass hidden state from the current time step to the next time step during roadbed compaction. Specifically, the row index of the hidden state transition probability matrix A represents the initial hidden state at the current time step, and the column index represents the target hidden state at the next time step. Each element aij in the matrix represents the probability that when the compaction mass is in the i-th hidden state in the current compaction pass, it will transition to the j-th hidden state in the next compaction pass. In this embodiment, the three hidden states are, in order, the insufficiently compacted hidden state, the normally compacted hidden state, and the high-compaction hidden state, corresponding to the 1st row / column, 2nd row / column, and 3rd row / column of the matrix. Figure 1In the illustrated embodiment, the element a11 = 0.152 in the first row and first column of the matrix indicates that when a trajectory analysis unit is in an insufficiently compacted hidden state in the current compaction pass, the probability of it remaining in the insufficiently compacted hidden state in the next compaction pass is 15.2%. This value is relatively small, indicating that the insufficiently compacted state has strong instability and tends to transition to other states. The element a12 = 0.765 in the first row and second column of the matrix indicates that the probability of transitioning from the insufficiently compacted hidden state to the normal compacted hidden state is 76.5%. This is the largest value in this row, indicating that in the insufficiently compacted state, the most likely evolution direction after additional compaction is to reach the normal compacted state. The element a13 = 0.083 in the first row and third column of the matrix indicates that the probability of directly transitioning from the insufficiently compacted hidden state to the high-compacted hidden state is only 8.3%. This low probability is reasonable because the improvement of compaction quality is usually a gradual process, and it is rare to directly reach high compaction without sufficient compaction. The second row of the matrix represents the transition pattern of the implicit state of normal compaction. Element a21 = 0.042 indicates that the probability of degenerating from the normal compaction state to the insufficiently compacted state is only 4.2%. This extremely low probability means that once the normal compaction state is reached, due to the compaction memory effect of the subgrade material itself, even if subsequent rolling parameters fluctuate slightly, it is difficult for the compaction quality to decrease significantly. Element a22 = 0.687 indicates that the probability of the normal compaction state maintaining itself is 68.7%. This value indicates that the normal compaction state has strong stability and is likely to remain in the current state during subsequent rolling. Element a23 = 0.271 indicates that the probability of improving from the normal compaction state to the high-compaction state is 27.1%, indicating that on the basis of normal compaction, by adding rolling passes or optimizing rolling parameters, there is a high probability of further improving the compaction quality to an excellent level. The third row of the matrix shows the transition characteristics of the implicit state of high-compaction. Element a31 = 0.018 indicates that the probability of degrading from a high-compaction state to an insufficiently compacted state is only 1.8%. This extremely low probability demonstrates the strong stability of the high-compaction state, making a significant quality decrease during subsequent compaction virtually impossible. Element a32 = 0.235 indicates that the probability of regressing from a high-compaction state to a normal compaction state is 23.5%. This probability is relatively moderate and may correspond to slight loosening caused by continuing to apply high-intensity vibration to an already over-compacted area. Element a33 = 0.747 indicates that the probability of the high-compaction state maintaining itself is as high as 74.7%, the largest value among all diagonal elements. This means that once the roadbed reaches a high-compaction state, its internal structure has formed a stable, dense skeleton, which can maintain excellent compaction quality for a long time during subsequent compaction. To visually illustrate the differences in the magnitude of the transition probabilities, Figure 1The matrix elements are visualized using a grayscale filling method, where darker grayscale values ​​represent higher probability values. It is clearly observed that the darkest grayscale values ​​are in the 1st row, 2nd column; the 2nd row, 2nd column; and the 3rd row, 3rd column, corresponding to probability values ​​of 0.765, 0.687, and 0.747, respectively. These three elements represent the three main evolutionary paths: transition from insufficient compaction to normal compaction, self-sustaining normal compaction, and self-sustaining high compaction. In contrast, the elements representing quality degradation (such as a21=0.042 and a31=0.018) have extremely light grayscale values, indicating that the probability of a significant decrease in compaction quality under normal construction conditions is extremely low. Figure 1 Below, the four main transition directions and their corresponding transition probabilities are further marked. The probability of the "Insufficient → Normal" transition is 0.765, the highest among all transitions, indicating that increasing the number of compaction passes during construction can effectively improve compaction quality. The probability of the "Normal → Normal" transition is 0.687, indicating that normal compaction can maintain stability. The probability of the "High-Pressure Compaction → High-Pressure Compaction" transition is 0.747, showing that the high-pressure compaction state has high stability. The probability of the "Normal → High-Pressure Compaction" transition is 0.271, indicating that there is still considerable room for quality improvement on the basis of normal compaction. The construction of this hidden state transition probability matrix is ​​based on the training process described in step 2.1, and is obtained by statistical analysis of no less than 50 trajectory analysis units with compaction quality detection data. The sum of the elements in each row of the matrix is ​​equal to 1 (after rounding to 3 decimal places), satisfying the normalization condition of the probability distribution. This matrix, together with the observation output probability table and the hidden state initial probability table, will form the core parameters of the initial model of the Hidden Markov Model (HMM) in subsequent steps. It will be used to perform Viterbi decoding on the observation sequence of the real-time trajectory analysis unit, thereby inferring the hidden state of compaction quality of each trajectory analysis unit in the current rolling stage.

[0044] When constructing the compaction manifold embedding space, the first step is to construct a compaction feature vector based on all trajectory analysis units. For each trajectory analysis unit, the average travel speed of all observed segments in its trajectory analysis unit observation sequence is averaged to obtain the average average travel speed; the average vibration amplitude of all observed segments is averaged to obtain the average vibration amplitude; and the average vibration frequency of all observed segments is averaged to obtain the average vibration frequency. Simultaneously, the cumulative residence time of the roller within the trajectory analysis unit is recorded. These four statistics are combined into a compaction feature vector in a fixed order. These four statistics are chosen because the average average travel speed, average average vibration amplitude, and average average vibration frequency reflect the typical operating conditions experienced by the trajectory analysis unit during long-term compaction, while the cumulative residence time reflects the cumulative intensity of the action on the trajectory analysis unit. Combining these four can comprehensively summarize the compaction behavior characteristics of the trajectory analysis unit.

[0045] To eliminate differences in the dimensions and numerical ranges of different features and prevent any single feature from dominating subsequent distance calculations due to an excessively large numerical range, the compacted feature vector set needs to be standardized. Specifically, for each feature, the mean and standard deviation of that feature across all trajectory analysis units are calculated. The standard deviation can be obtained by averaging and then taking the square root of the squared difference between the feature and the mean for each trajectory analysis unit. Then, using this mean and standard deviation, the corresponding feature for each trajectory analysis unit is linearly scaled, so that the scaled feature has a mean close to 0 and a standard deviation close to 1 across all trajectory analysis units. After standardization, the four features are on the same order of magnitude in terms of numerical scale, which is beneficial for fairly measuring the contribution of each feature when calculating the Euclidean distance.

[0046] After obtaining the standardized set of compaction feature vectors, for any two trajectory analysis units, the Euclidean distance is calculated based on the scaled values ​​of the four features. The Euclidean distances between all trajectory analysis units are stored in a distance matrix. The distance matrix reflects the similarity relationships between trajectory analysis units in the compaction feature space. To capture local structure, for each trajectory analysis unit, the five distinct trajectory analysis units with the smallest distance from it are selected from the distance matrix. These are designated as neighboring trajectory analysis units, and the Euclidean distances between the current trajectory analysis unit and its five neighbors are recorded, constructing an adjacency set containing all trajectory analysis units. Selecting five neighboring trajectory analysis units is a compromise; too few neighbors would result in an unconnected adjacency set, while too many would obscure the local structure, causing slight differences to be considered strong similarities. On the adjacency set, a shortest path algorithm is used. For each pair of trajectory analysis units, the shortest path distance through the adjacency set is calculated by finding a path in the adjacency set and accumulating the Euclidean distances along that path. All shortest path distances are stored in a shortest path distance matrix. Compared with using Euclidean distance directly, the shortest path distance not only considers the direct differences between two trajectory analysis units, but also the indirect differences that are gradually transitioned through multiple intermediate trajectory analysis units, thus better reflecting the overall geometric structure of the entire compaction feature space.

[0047] After obtaining the shortest path distance matrix, dimensionality reduction is performed on it. Specifically, the mean of the shortest path distance matrix can be calculated for each row and column, and the mean of the corresponding row and column can be subtracted from each element of the matrix, and then the overall mean can be added to obtain a symmetric matrix. Eigenvalue decomposition is performed on the symmetric matrix to obtain a set of eigenvalues ​​and corresponding eigenvectors. The two eigenvectors corresponding to the two largest eigenvalues ​​are selected, and the components of each trajectory analysis unit in these two eigenvectors are used as the first and second coordinates of the trajectory analysis unit in the compaction manifold embedding space, thus obtaining the compaction manifold embedding coordinates of each trajectory analysis unit. The compaction manifold embedding space constructed in this way can preserve the relative positional relationships between trajectory analysis units in the original high-dimensional compaction feature space as much as possible in two-dimensional space. Intuitively, trajectory analysis units with similar compaction behaviors will be close to each other in the compaction manifold embedding space, while trajectory analysis units with significantly different compaction behaviors will be far apart, which is beneficial for subsequent neighborhood statistics and latent state correction in this space. In another alternative implementation, the three feature vectors corresponding to the three largest feature values ​​can be selected to construct a three-dimensional compaction manifold embedding space. When more detailed visualization is required, the compaction behavior distribution can be displayed through a three-dimensional interactive method.

[0048] After obtaining the initial Hidden Markov Model (HMM) model and the compacted manifold embedding space, a joint iterative correction is needed to obtain a joint dynamic correlation analysis model containing the convergent final hidden state numbers. The joint iterative correction process first uses the initial HMM model to perform Viterbi decoding on the observation symbol sequences of all trajectory analysis units obtained in step 1. Specifically, for each trajectory analysis unit's observation symbol sequence, at the first observation time, the initial scores of the three hidden states are calculated. The initial score can be understood as the degree of matching between the hidden state and the current observation symbol after considering the initial probability and the observation output probability. At the second and subsequent observation times, the scores of the three hidden states at the current time are calculated sequentially. This score can be obtained by traversing the scores of the three hidden states at the previous time, the corresponding hidden state transition probabilities, and the observation output probability corresponding to the current observation symbol. During the calculation process, the hidden state number corresponding to each current hidden state score at the previous time is recorded, allowing the complete hidden state sequence to be gradually recovered from back to front along the path with the highest score. At the last observation time, the hidden state number with the highest score is selected as the terminating hidden state number of that observation symbol sequence. Then, the hidden state number is obtained by backtracking along the recorded hidden state path numbers. The hidden state number at the last observation time is used as the final hidden state number of the corresponding trajectory analysis unit, thus obtaining the initial and final hidden state number set for all trajectory analysis units.

[0049] After assigning initial final hidden state numbers to all trajectory analysis units, the joint iterative correction process begins. The initial iteration count is 1. In each iteration, for each trajectory analysis unit, the Euclidean distance between that trajectory analysis unit and other trajectory analysis units is calculated in the compacted manifold embedding space. The five trajectory analysis units with the smallest Euclidean distances are selected as the neighboring trajectory analysis units in the compacted manifold embedding space. Then, the occurrence counts of the three hidden state numbers (insufficiently compacted, normally compacted, and heavily compacted) in the final hidden state numbers of these five neighboring trajectory analysis units are counted. If the occurrence count of a certain hidden state number in the neighboring trajectory analysis units is not less than 3 and is greater than the occurrence count of the current final hidden state number of that trajectory analysis unit in the neighboring trajectory analysis units, then the final hidden state number of that trajectory analysis unit is replaced with the hidden state number that occurs most frequently in the neighborhood. The reason for this setup is that when a trajectory analysis unit is surrounded by multiple neighboring trajectory analysis units with the same hidden state number in the compaction manifold embedding space, it indicates that the trajectory analysis unit is very similar to these neighboring trajectory analysis units in terms of compaction characteristics. Therefore, adjusting the final hidden state number of the trajectory analysis unit to be consistent with the mainstream hidden state of the neighborhood helps to eliminate abnormal hidden state judgments caused by local observation noise or individual abnormal iterations, making the spatial distribution of the final hidden state number more continuous and smooth. After each iteration, the number of trajectory analysis units whose final hidden state numbers were replaced in this iteration is counted, and the ratio of this number to the total number of trajectory analysis units is calculated. If this ratio is less than 0.01, it means that there are very few trajectory analysis units that need to be corrected in the current iteration, and the overall hidden state distribution tends to be stable, so the iteration can be terminated; if the iteration count has reached 10, the iteration is also terminated to limit computational overhead. Otherwise, the iteration count is incremented by 1, and the next round of iterations continues. After the above iterations, when the termination condition is met, the convergent final hidden state number of each trajectory analysis unit is obtained. At this point, the initial Hidden Markov Model (HMM), the compacted manifold embedding space, and the convergent final hidden state number consistent with the compacted manifold embedding coordinates together constitute the joint dynamic correlation analysis model. In another optional implementation, the number of neighborhood trajectory analysis units can be adjusted to 7, the condition of occurrence frequency not less than 3 can be adjusted to not less than 4, or the proportion threshold can be adjusted from 0.01 to 0.02, in order to enhance spatial consistency constraints in scenarios with higher road grades and stricter compaction quality requirements.

[0050] Step 3: Using the joint dynamic correlation analysis model, the observation sequence of the real-time trajectory analysis unit constructed from the real-time collected data is processed, and combined with the compaction manifold embedded coordinates, a dynamic compaction quality grade distribution map covering the roadbed construction area is generated.

[0051] In one implementation, after training the joint dynamic correlation analysis model, it is deployed on the onboard computing device of the road roller or on-site edge computing device to generate a dynamic compaction quality grade distribution map covering the roadbed construction area in real time during the roadbed compaction construction process. To this end, it is necessary to continuously collect new roadbed compaction trajectory data, attitude data, and vibration response data during construction, and construct a real-time trajectory analysis unit observation sequence in the same manner as described above. Then, the joint dynamic correlation analysis model is used to process the real-time trajectory analysis unit observation sequence, and the compaction quality area is divided and graphically displayed by combining the compaction manifold embedded coordinates.

[0052] During construction, the positioning device, attitude detection device, and vibration response acquisition device continue to operate at a preset sampling period, for example, 0.1 seconds. For each sampling, the position information, attitude information, and vibration response information of the roller are recorded, and the position information is converted into planar coordinates and assigned to the corresponding trajectory analysis unit. To maintain consistency with the statistical characteristics in the joint dynamic correlation analysis model, for each trajectory analysis unit, sampling points are aggregated according to the compaction pass number within the current construction time window (e.g., within the last 5 minutes). The average travel speed, average vibration amplitude, and average vibration frequency of each compaction pass on the trajectory analysis unit are calculated, thus obtaining a real-time trajectory analysis unit observation sequence containing multiple observation segments. To ensure that the real-time trajectory analysis unit observation sequence sufficiently reflects the current compaction process, preferably, when the length of the real-time trajectory analysis unit observation sequence of a certain trajectory analysis unit is less than 3, the hidden state inference of that trajectory analysis unit can be temporarily suspended. After the trajectory analysis unit accumulates 3 or more observation segments, the joint dynamic correlation analysis model is then applied. This avoids instability in the hidden state judgment due to too few observation segments.

[0053] After constructing the observation sequence of the real-time trajectory analysis unit, each observation segment needs to be converted into discrete observation symbols in the same way as in the offline training phase. Specifically, for each observation segment, it is mapped to the corresponding speed interval code, vibration amplitude interval code, and frequency interval code according to the interval to which the average driving speed, average vibration amplitude, and average vibration frequency belong. For example, an average driving speed of no more than 0.8 m / s is mapped to a low-speed interval code, between 0.8 m / s and 1.2 m / s is mapped to a medium-speed interval code, and greater than 1.2 m / s is mapped to a high-speed interval code. Then, the three interval codes are concatenated in a fixed order to form discrete observation symbols, thus forming a real-time observation symbol sequence. Since the initial model of the Hidden Markov Model (HMM) in the joint dynamic correlation analysis model is trained under the same encoding method, Viterbi decoding can be directly performed on the real-time observation symbol sequence.

[0054] For each trajectory analysis unit's real-time observed symbol sequence, Viterbi decoding is performed using a Hidden Markov Model (HMM) in the joint dynamic correlation analysis model. Specifically, at the first observation time, initial scores are calculated for the insufficiently compacted, normally compacted, and highly compacted hidden states, respectively. These initial scores comprehensively consider the initial probability of the hidden state and the observation output probability, measuring the degree of matching between the hidden state and the current observed symbol combination. At the second and subsequent observation times, the current score is calculated for each hidden state sequentially. The current score is derived from the superposition of the three hidden state scores and the hidden state transition probability from the previous time, multiplied by the observation output probability corresponding to the current observed symbol. To recover the complete hidden state path, the hidden state number from the previous time that resulted in the maximum current score is recorded at each observation time. At the last observation time, the hidden state number with the largest score is selected as the terminating hidden state number of the real-time observed symbol sequence. Then, the hidden state path numbers are traced backwards along the recorded paths to obtain the hidden state numbers for each observation time. Since the dynamic compaction quality grade distribution map focuses on the comprehensive compaction state of each trajectory analysis unit at the current moment, the hidden state number of the last observation moment can be used as the real-time final hidden state number of that trajectory analysis unit.

[0055] In one implementation, Figure 2 This paper demonstrates the complete process of using a Hidden Markov Model (HMM) to perform Viterbi decoding on the observation symbol sequence of a trajectory analysis unit, thereby determining the optimal hidden state path for that unit. In this example, the observation symbol sequence is 6 in length, corresponding to the observation segments collected by the trajectory analysis unit in 6 consecutive rolling passes, denoted as O1, O2, O3, O4, O5, and O6. The goal of Viterbi decoding is to find, among all possible hidden state sequences, the path that best matches the observation symbol sequence, maximizing the joint probability.

[0056] Specifically, Figure 2 A grid structure is used to illustrate the Viterbi decoding calculation process. The horizontal axis represents the time dimension, from left to right: six observation times from t=1 to t=6, corresponding to six compaction passes. The vertical axis represents the hidden state dimension, from top to bottom: three possible states: insufficiently compacted hidden state, normally compacted hidden state, and high-compaction hidden state. At the intersection of each time t and each hidden state i, a circular node is drawn, with the name of the hidden state labeled inside the node. All nodes form a grid of 18 nodes in 3 rows and 6 columns.

[0057] At time t=1, the Viterbi algorithm first uses the initial probability of the hidden state and the observation output probability corresponding to the first observation symbol O1 to calculate the initial scores of the three hidden states. Figure 2In the analysis, the initial score for the insufficiently compacted latent state at t=1 is δ1(1)=0.32, the initial score for the normally compacted latent state is δ1(2)=0.58, and the initial score for the high-compaction latent state is δ1(3)=0.10. These three scores indicate that when the operating parameter combination O1 of the first compaction pass is observed, the trajectory analysis unit is most likely in the normally compacted state (score 0.58 is the highest), followed by the insufficiently compacted state (score 0.32), while the probability of being in the high-compaction state is low (score only 0.10). This score distribution reflects that the subgrade is usually in the normally compacted or insufficiently compacted state in the initial compaction stage, and rarely reaches the high-compaction state after the first compaction pass.

[0058] At time t=2, the algorithm needs to calculate the scores of the three hidden states at time t=2 based on the scores at time t=1, the hidden state transition probabilities, and the observation output probability of the current observation symbol O2. For any hidden state j at time t=2, the algorithm iterates through all three hidden states i at time t=1, calculating the path score for transitioning from state i to state j. This path score is equal to δ1(i) multiplied by the transition probability aij and then multiplied by the observation output probability bj(O2). Among the three possible transition paths, the path with the highest score is selected as the optimal predecessor path to state j, and the predecessor state number corresponding to this path is recorded. Figure 2 In the diagram, since all possible state transition paths are represented by thin gray lines, while the optimal path is marked by a thick black line, it is clear that in the transition from t=1 to t=2, the optimal path transitions from the insufficiently compacted state at t=1 (score 0.32) to the normally compacted state at t=2 (score 0.72). This transition conforms to... Figure 2 The transition probability matrix shown has the largest element in the row, with a transition probability of 0.765 from the uncompacted state to the normal compacted state. At time t=2, the scores for the three states are δ2(1)=0.15, δ2(2)=0.72, and δ2(3)=0.13, respectively. The score for the normal compacted state is significantly higher than that for the other two states, indicating that after two passes of compaction, the compaction quality of this trajectory analysis unit has been improved to a normal level.

[0059] At time t=3, the algorithm continues to perform the same calculation process. From Figure 2As can be seen, the optimal path continues from the normal compaction state at t=2 to the normal compaction state at t=3, with a corresponding score of δ3(2)=0.81. This score further increases, indicating that as the number of compaction passes increases, the model's confidence in judging that the trajectory analysis unit is in a normal compaction state continuously strengthens. At this moment, the score for the insufficiently compacted state has dropped to δ3(1)=0.08, and the score for the high-compacted state is δ3(3)=0.11, both of which are much lower than the score for the normal compaction state, indicating that the trajectory analysis unit stably maintains a normal compaction state during the third compaction pass.

[0060] At time t=4, the observation symbol O4 indicates a change in the combination of operating parameters of the roller. According to the calculation results, the score of the normal compaction state is δ4(2)=0.68, which is still the highest among the three states, but the value has decreased compared to 0.81 at time t=3. At the same time, the score of the high compaction state rises to δ4(3)=0.27, and the score of the insufficiently compacted state further decreases to δ4(1)=0.05. This trend indicates that the trajectory analysis unit is still in the normal compaction state during the fourth compaction, but the observed combination of operating parameters has begun to approach the parameter characteristics corresponding to the high compaction state, laying the groundwork for the transition to the high compaction state at the next time step. Figure 2 The optimal path from t=3 to t=4 remains in a normal compaction state.

[0061] At time t=5, the compaction quality underwent a significant evolution. The score for the high-compaction state jumped to δ5(3)=0.55, exceeding the score for the normal compaction state δ5(2)=0.43 for the first time, becoming the highest-scoring latent state at that time. The score for the insufficiently compacted state dropped to an extremely low δ5(1)=0.02, indicating that this state could be largely ruled out. Figure 2 In this process, the optimal path shifts from the normal compaction state at t=4 to the high compaction state at t=5. This shift corresponds to... Figure 2 The transition probability a23 = 0.271 indicates that, based on normal compaction, optimizing rolling parameters or increasing the number of rolling passes has a high probability of further improving the compaction quality to an excellent level. This transition path reflects the typical evolutionary law of gradual improvement in compaction quality during roadbed compaction.

[0062] At time t=6, the last observation time, the score for the high-compaction state further increased to δ6(3)=0.74, far higher than the score for the normal-compaction state δ6(2)=0.25 and the score for the insufficiently compacted state δ6(1)=0.01. The optimal path continued from the high-compaction state at t=5 to the high-compaction state at t=6, corresponding to Figure 2The transition probability a33 = 0.747 indicates that the high-voltage real state has a strong self-sustaining ability. In the final step of Viterbi decoding, the algorithm selects the hidden state with the highest score at time t=6, i.e., the high-voltage real state, as the terminating hidden state.

[0063] After determining the terminating hidden state, the Viterbi algorithm backtracks backward along the recorded optimal predecessor path numbers. Starting from the high compaction state at t=6, it searches for the optimal predecessor of this state at t=5, obtaining the high compaction state at t=5; continuing backtracking, it obtains the normal compaction state at t=4, the normal compaction state at t=3, the normal compaction state at t=2, and the insufficiently compacted state at t=1. Arranging these six hidden states in chronological order yields the complete optimal hidden state sequence: insufficiently compacted → normal compacted → normal compacted → normal compacted → high compaction → high compaction. This sequence... Figure 2 The lines connected by thick black lines clearly show the complete process of the compaction quality of the trajectory analysis unit evolving from an initial insufficiently compacted state to a high-compacted state.

[0064] Figure 2 In the legend at the bottom, the thick black solid line represents the Viterbi optimal path, i.e., the path with the highest probability among all possible hidden state sequences; the thin gray lines represent other possible but non-optimal transition paths. At each time step, theoretically there are 3 × 3 = 9 possible transition paths (transitioning from the 3 states of the previous time step to the 3 states of the current time step respectively). However, the Viterbi algorithm, through a dynamic programming strategy, retains only one optimal predecessor path at each state node, thus reducing the exponential path search complexity to linear complexity. In this example, there are 5 state transitions from t=1 to t=6, each with 9 possibilities, for a total of 9^5 = 59049 possible combinations of hidden state sequences. The Viterbi algorithm, through time-by-time local optimal selection and dynamic programming recording, can efficiently find the globally optimal hidden state sequence without enumerating all possible combinations.

[0065] After obtaining the real-time final hidden state number, it is necessary to perform spatial clustering of the trajectory analysis units, dividing them into multiple compaction quality regions, using the compaction manifold embedding space in the joint dynamic correlation analysis model. The joint dynamic correlation analysis model has pre-calculated compaction manifold embedding coordinates for each trajectory analysis unit. These coordinates have the following property: trajectory analysis units with similar compaction behaviors are close to each other in the compaction manifold embedding space, while those with significantly different compaction behaviors are far apart. Therefore, in real-time applications, these compaction manifold embedding coordinates can be directly used to divide the compaction quality regions. Specifically, in one update cycle (e.g., every 10 seconds or every 30 seconds), for all trajectory analysis units that have obtained real-time final hidden state numbers, Euclidean distances are calculated pairwise according to the compaction manifold embedding coordinates. For each trajectory analysis unit, other trajectory analysis units with Euclidean distances less than a set distance threshold are found. The distance threshold can be preset based on the coordinate scale of the compaction manifold embedding space. For example, when the values ​​of the first and second coordinates of the compaction manifold embedding coordinates are approximately between -3 and 3, the distance threshold can be set to 0.8 or 1.0. A smaller distance threshold will result in more and finer division of compaction quality areas, but adjacent trajectory analysis units may be assigned to different areas; a larger distance threshold will merge more trajectory analysis units into the same compaction quality area, but may mask some local differences. By comparing the compaction quality area division results under different distance thresholds during the trial operation phase, a set distance threshold that reflects the differences in construction quality without excessive fragmentation can be selected.

[0066] When specifically dividing compaction quality zones, it is required that not only is the distance between the compaction manifold embedded coordinates less than a set distance threshold, but also that the final hidden state numbers of the trajectory analysis units are the same. This is because the compaction manifold embedded coordinates primarily reflect the overall similarity of long-term compaction behavior, while the real-time final hidden state number reflects the state of the current compaction stage. If clustering is performed solely based on the compaction manifold embedded coordinates, trajectory analysis units at different compaction stages may be grouped into the same region; if division is performed solely based on the real-time final hidden state number, the cumulative differences in compaction behavior may be ignored. By requiring that the distance between the compaction manifold embedded coordinates be less than a set distance threshold and that the final hidden state numbers be the same, a high degree of consistency in historical behavior and current state is ensured among trajectory analysis units within the same compaction quality zone, making them more suitable as basic units for construction control.

[0067] After forming compaction quality zones, the latent state information needs to be mapped to easily understood compaction quality levels. In the joint dynamic correlation analysis model, the latent states of insufficient compaction, normal compaction, and high compaction are respectively corresponding to unqualified, qualified, and excellent compaction quality levels. Therefore, for each compaction quality zone, the compaction quality level of that zone can be directly determined based on the final latent state number of the trajectory analysis unit within that zone. For example, when the final latent state numbers of all trajectory analysis units within a certain compaction quality zone correspond to the number corresponding to the insufficient compaction latent state, the compaction quality zone is marked as unqualified compaction quality level; when they all correspond to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number corresponding to the number of ...

[0068] When generating a dynamic compaction quality grade distribution map, it is necessary to map the compaction quality grade information to the planar location of the subgrade construction area. Since each trajectory analysis unit has a unique spatial location in the subgrade plane coordinate system, a grid consistent with the trajectory analysis unit division can be constructed on the display terminal, representing each trajectory analysis unit as a small rectangle or polygon on the screen. Based on the compaction quality area and compaction quality grade to which the trajectory analysis unit belongs, the small rectangle or polygon is filled with the corresponding color; for example, unqualified compaction quality grade is set to red, qualified compaction quality grade to yellow, and excellent compaction quality grade to green. Trajectory analysis units that have not yet completed compaction or lack real-time trajectory analysis unit observation sequences can be represented in gray. In this way, a dynamic compaction quality grade distribution map covering the entire subgrade construction area can be generated on the display terminal, allowing construction personnel to intuitively see which areas have reached the excellent compaction quality grade, which areas are only at the qualified compaction quality grade, and which areas are still at the unqualified compaction quality grade.

[0069] To reflect dynamic characteristics, the dynamic compaction quality level distribution map can be updated at preset time intervals. For example, every 10 seconds or every 30 seconds, the currently collected real-time trajectory analysis unit observation sequence is batch-processed, the real-time final hidden state number of each trajectory analysis unit is recalculated, the compaction quality area is re-divided, and the color distribution on the display terminal is updated. Since the displacement of the road roller within 10 or 30 seconds is usually within the range of 10 to 30 meters, each update introduces a batch of new observation segments, enabling the dynamic compaction quality level distribution map to better reflect the current construction progress. In some implementations, a certain number of historical layers can be retained for different time slices, such as retaining 10 dynamic compaction quality level distribution maps generated in the last 5 minutes, so that construction managers can compare the compaction quality change trends at different times and identify quality fluctuations caused by factors such as changes in moisture content and equipment status.

[0070] When outputting the dynamic compaction quality grade distribution map as the result of the dynamic correlation analysis between subgrade compaction trajectory and compaction quality, related statistical information can be output simultaneously, such as the total area of ​​unqualified compaction quality grade areas, the total area of ​​qualified compaction quality grade areas, the total area of ​​excellent compaction quality grade areas, and the proportion of each compaction quality grade area to the entire construction area. Construction managers can use this statistical information and the graphical distribution to guide the adjustment of roller operation paths and control of compaction passes. For example, for unqualified compaction quality grade areas shown in red on the dynamic compaction quality grade distribution map, rollers can be prioritized for additional compaction along these areas in the next round of construction; for qualified compaction quality grade areas shown in yellow, if the construction specifications require an excellent compaction quality grade, one or two additional compaction passes can be reasonably arranged based on the area of ​​the area and equipment capacity; for excellent compaction quality grade areas shown in green, unnecessary repeated compaction can be reduced, thereby reducing fuel consumption and equipment wear. In another optional implementation, the real-time location and travel trajectory of the road roller can be superimposed on the dynamic compaction quality grade distribution map, so that the driver can directly observe the compaction quality grade distribution of the road section ahead through the display terminal in the cab, and make spontaneous path fine-tuning based on the current location of the road roller, thereby improving work efficiency.

[0071] The present invention has been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the invention. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and core ideas of the present invention. It should be noted that those skilled in the art can make various improvements and modifications to the present invention without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

Claims

1. A method for dynamically correlating subgrade rolling track and compaction quality, characterized in that, The method includes: Step 1: Collect the position, attitude, and vibration response information of the road roller during the roadbed compaction process, divide the roadbed construction area into trajectory analysis units, and construct a trajectory analysis unit observation sequence containing the road roller's operating parameters; obtain compaction quality test data, and establish the correspondence between compaction quality grade and latent state to obtain the latent state sample set and the corresponding partial trajectory analysis unit observation sequence sample set; Step 2: Execute the joint dynamic correlation analysis algorithm based on Hidden Markov Model (HMM) and manifold learning, including: Step 2.1: Initially train the Hidden Markov Model (HMM) using the hidden state sample set and the partial trajectory analysis unit observation sequence sample set to obtain the initial HMM model. Step 2.2: Construct the compaction manifold embedding space based on the compaction feature vectors of all trajectory analysis units, and obtain the compaction manifold embedding coordinates of each trajectory analysis unit; Step 2.3: Perform joint iterative correction on the initial model of the Hidden Markov Model (HMM) and the embedded space of the compacted manifold to obtain a joint dynamic correlation analysis model containing the final hidden state number of the convergence. Step 3: Using the joint dynamic correlation analysis model, the observation sequence of the real-time trajectory analysis unit constructed from the real-time collected data is processed, and combined with the compaction manifold embedded coordinates, a dynamic compaction quality grade distribution map covering the roadbed construction area is generated; Step 1, the process of constructing the trajectory analysis unit observation sequence, specifically includes: collecting position information using a positioning device, collecting attitude information using an attitude detection device, and collecting vibration response information using a vibration response acquisition device; dividing the roadbed construction area into several trajectory analysis units with a side length of one meter on the plane, and assigning a unique number to each trajectory analysis unit; assigning the position information at each acquisition time to the corresponding trajectory analysis unit according to the plane coordinates; assigning a compaction pass number to each complete round-trip compaction process, and considering sampling points belonging to the same trajectory analysis unit and having the same compaction pass number as observation segments of the same compaction pass on the corresponding trajectory analysis unit; for each trajectory analysis unit, arranging the observation segments corresponding to each compaction pass on the corresponding trajectory analysis unit in chronological order according to the acquisition time, forming a trajectory analysis unit observation sequence with a length of no less than five; each observation segment includes the average driving speed, average vibration amplitude, average vibration frequency, and marking information of the corresponding compaction pass on the corresponding trajectory analysis unit. In step 1, within the trajectory analysis unit with compaction quality detection data, the compaction quality corresponding to the trajectory analysis unit is divided into three compaction quality levels: insufficient compaction level, normal compaction level, and high compaction level, based on the numerical range of the compaction quality detection data. A one-to-one correspondence is then established between the three compaction quality levels and the latent states of insufficient compaction, normal compaction, and high compaction, respectively, to obtain the latent state sample set and the partial trajectory analysis unit observation sequence sample set corresponding to the latent state sample set.

2. The method of claim 1, wherein, Step 2.1, the initial training process of the Hidden Markov Model, specifically includes: selecting no fewer than fifty trajectory analysis units with compaction quality detection data as training samples; for each training sample, dividing each observation segment in its trajectory analysis unit observation sequence into one of three speed interval codes according to the average driving speed, one of three vibration amplitude interval codes according to the average vibration amplitude, and one of three frequency interval codes according to the average vibration frequency; concatenating the three interval codes in a fixed order to form discrete observation symbols; obtaining the observation symbol sequence sequentially for the observation segments in the same trajectory analysis unit observation sequence; in each training sample, within the same trajectory analysis unit observation sequence, counting the number of transitions between adjacent hidden states in the order of their hidden states, and for each initial... The combination of the hidden state and the target hidden state is used to divide the number of transitions of the combination by the total number of transitions of the corresponding initial hidden state, resulting in a hidden state transition probability table with three rows and three columns. The sum of the three values ​​in each row of the table is rounded to three decimal places and equals one. In all training samples, the first hidden state of the observation sequence of each trajectory analysis unit is taken as the initial hidden state. The occurrence frequency of each of the three hidden states as the initial hidden state is counted. The occurrence frequency of each hidden state is divided by the total number of training samples, resulting in a hidden state initial probability table with three rows and one column. In each training sample, for each hidden state, the occurrence frequency of each discrete observation symbol in the corresponding hidden state is counted. The occurrence frequency of each discrete observation symbol is divided by the sum of the occurrence frequencies of all observation symbols in the corresponding hidden state, resulting in three observation output probability tables with several columns.

3. The method according to claim 2, characterized in that, The initial training process for the Hidden Markov Model (HMM) includes: using the initial probability table of hidden states, the hidden state transition probability table, and the observation output probability table as the initial parameters of the HMM, and using the observation symbol sequence of all training samples as input; in each iteration, performing forward and backward computations on each observation symbol sequence in chronological order, with the forward computation calculating the forward probabilities of the three hidden states sequentially at each time step, and the backward computation calculating the backward probabilities of the three hidden states sequentially from the last observation time step backward; then updating the probability values ​​in the initial probability table of hidden states, the hidden state transition probability table, and the observation output probability table based on the forward and backward probabilities, retaining all probability values ​​to three decimal places after each update; stopping the iteration when the change in any value in the hidden state transition probability table is less than 0.001 in two consecutive iterations or when the number of iterations reaches fifty, thus obtaining the initial model of the HMM.

4. The method according to claim 1, characterized in that, The process of constructing the compaction manifold embedding space in step 2.2 specifically includes: for each trajectory analysis unit, calculating the average driving speed, average vibration amplitude, average vibration frequency, and cumulative dwell time of all observed segments in its trajectory analysis unit observation sequence, and forming a compaction feature vector in the compaction feature vector set in a fixed order; for each feature in the compaction feature vector set, calculating the average value and standard deviation of each feature in all trajectory analysis units, and then using the calculated average value and standard deviation to linearly scale each feature of each trajectory analysis unit so that the average value of each feature of all trajectory analysis units is zero and the standard deviation is one.

5. The method according to claim 4, characterized in that, The process of constructing the compacted manifold embedding space also includes: for any two trajectory analysis units, calculating the Euclidean distance according to the scaled values ​​of the four features, and storing the Euclidean distances between all trajectory analysis units in the distance matrix; for each trajectory analysis unit, selecting the five different trajectory analysis units with the smallest distance in the distance matrix as neighboring trajectory analysis units, and constructing an adjacency relationship set containing all trajectory analysis units; on the adjacency relationship set, using the shortest path algorithm, calculating the shortest path distance through the adjacency relationship set for each pair of trajectory analysis units, and storing all shortest path distances in the shortest path distance matrix; performing dimensionality reduction processing on the shortest path distance matrix, specifically, correcting the shortest path distance matrix by row mean and column mean to obtain a symmetric matrix, then performing eigenvalue decomposition on the symmetric matrix, selecting the two eigenvectors corresponding to the two largest eigenvalues, and using the components of each trajectory analysis unit in the two eigenvectors as the first coordinate and second coordinate of the corresponding trajectory analysis unit in the compacted manifold embedding space, respectively, to obtain the compacted manifold embedding coordinates of each trajectory analysis unit.

6. The method according to claim 1, characterized in that, In step 2.3, the joint iterative correction process of the Hidden Markov Model (HMM) and the compacted manifold embedding space specifically includes: for all obtained trajectory analysis units, the initial model of the Hidden Markov Model (HMM) is adopted, and Viterbi decoding is performed on the observation symbol sequence of each trajectory analysis unit. Specifically, at the first observation time, the initial scores of the three hidden states are calculated respectively, and at the second and subsequent observation times, the scores of the three hidden states at the current time are calculated sequentially. At each observation time, the path number of the hidden state with the largest score from the previous time is recorded. At the last observation time, the hidden state number with the largest score is selected as the termination hidden state number of the observation sequence. Then, the hidden state number at each time is obtained by backtracking along the recorded hidden state path number. The hidden state number at the last observation time is taken as the final hidden state number of the corresponding trajectory analysis unit.

7. The method according to claim 6, characterized in that, The joint iterative correction process also includes: initializing the iteration count to one; in each iteration, for each trajectory analysis unit, calculating the Euclidean distance between the trajectory analysis unit and other trajectory analysis units in the compacted manifold embedding space; selecting the five trajectory analysis units with the smallest Euclidean distance as the neighboring trajectory analysis units in the compacted manifold embedding space; counting the occurrence counts of the three hidden state numbers in the final hidden state numbers of the five neighboring trajectory analysis units; if the occurrence count of a certain hidden state number is not less than three and greater than the hidden state number corresponding to the current final hidden state number of the corresponding trajectory analysis unit... The number of times a state number appears in a neighborhood trajectory analysis unit is used to replace the final hidden state number of the corresponding trajectory analysis unit with the hidden state number that appears most frequently. After one iteration, the ratio of the number of trajectory analysis units where the final hidden state number is replaced to the total number of trajectory analysis units is calculated. If the calculated ratio is less than 0.01, or the iteration count reaches 10, the iteration is stopped, and the converged final hidden state number of each trajectory analysis unit is obtained. The initial model of the Hidden Markov Model (HMM), the converged final hidden state number, and the compacted manifold embedded coordinates are used together as the joint dynamic correlation analysis model.

8. The method according to claim 1, characterized in that, Step 3 specifically includes: during the subgrade compaction construction process, real-time acquisition of new subgrade compaction trajectory data, attitude data, and vibration response data, and construction of a real-time trajectory analysis unit observation sequence; for each trajectory analysis unit's real-time observation symbol sequence, Viterbi decoding is performed using the Hidden Markov Model (HMM) in the Joint Dynamic Association Analysis Model to obtain the real-time final hidden state number; and using the compaction manifold embedding space in the Joint Dynamic Association Analysis Model, several trajectory analysis units with a compaction manifold embedding coordinate distance less than a set distance threshold and the same final hidden state number are divided into the same compaction quality region; the insufficiently compacted hidden state, the normal compacted hidden state, and the high-compaction hidden state are respectively corresponding to unqualified compaction quality grade, qualified compaction quality grade, and excellent compaction quality grade, and each compaction quality grade is displayed in color on the plane according to the spatial position of the trajectory analysis unit, generating a dynamic compaction quality grade distribution map covering the entire subgrade construction area; and the dynamic compaction quality grade distribution map is output as the result of the dynamic association analysis between the subgrade compaction trajectory and compaction quality, used to guide the adjustment of the roller's operating path and the control of the compaction passes.

Images