High-precision laser ranging-based dynamic detection method for highway construction quality

Abstract

The application relates to the technical field of laser ranging, and more particularly to a highway construction quality dynamic detection method based on high-precision laser ranging, which comprises the following steps: step 1, nodes of a plurality of high-precision laser ranging terminals are arranged at equal intervals on both sides of a center line of a to-be-built highway, and a unified reference three-dimensional coordinate system is established; step 2, a receiving end of each node receives original echo signals returned through pavement scattering; multi-scale wavelet transformation is performed on the original echo signals to form a construction quality multi-dimensional parameter matrix, so that the current construction quality state can be represented in real time in a high dimension; and step 3, principal component analysis is performed on the construction quality multi-dimensional parameter matrix to obtain a highway construction quality detection result. The application has the advantages of high ranging precision, strong space-time resolution, high automation degree, strong real-time performance, outstanding abnormality identification capability and the like.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of laser ranging, and particularly relates to a highway construction quality dynamic detection method based on high-precision laser ranging. BACKGROUND

[0002] In the process of modern infrastructure construction, the detection and monitoring of highway construction quality has become the core link to ensure structural safety and service life. The traditional highway construction quality detection methods mainly include manual section measurement, level gauge height measurement, total station control, static load settlement test, etc. Although these methods meet the basic needs of construction quality evaluation in a certain period, with the large-scale construction of modern highways, intelligent transportation, heavy-load roads and special structure road sections, they gradually expose technical bottlenecks and precision blind spots in dynamic response capture, high-frequency disturbance identification and millimeter-level displacement measurement.

[0003] Traditional section measurement usually relies on manual line pulling, flat scale and target device, and cooperates with theodolite to read the height difference of discrete measurement points. The accuracy of this method is limited by the instrument resolution and human operation error, and it cannot realize millimeter-level high-precision dynamic tracking under continuous spatial distribution. At the same time, this method does not have real-time performance and can only be used for single detection during construction intervals, making it difficult to use for immediate feedback of the quality of key processes (such as rolling, compaction, filling, etc.) in the construction process.

[0004] The total station control method introduces automatic angle and distance measurement capability and can perform multi-point synchronous observation, but the measurement frequency is generally several times per second, and complex manual reflection mark layout is required. For non-rigid structure continuous pavement materials (such as water stable layer, asphalt layer), the dependence of total station on echo stability makes its stability significantly decrease in strong sunlight, high dust or wet environment, and the change of reflection surface will cause distance jump, which seriously affects the displacement continuity analysis. SUMMARY

[0005] The main purpose of the application is to provide a highway construction quality dynamic detection method based on high-precision laser ranging. By arranging laser ranging terminal nodes on both sides of the highway center line, using a multi-frequency modulation coding sequence to emit a ranging light beam, combining wavelet transform, tensor decomposition, manifold clustering and deep time series network, millimeter-level displacement settlement monitoring and prediction are realized. Finally, a multi-dimensional parameter matrix of construction quality is constructed and a comprehensive quality index is extracted to evaluate the construction state in real time. This method has the advantages of high ranging accuracy, strong spatio-temporal resolution, high automation, strong real-time performance and outstanding abnormality identification capability, and can be widely applied to intelligent highway construction quality control scenes.

[0006] To solve the above problems, the technical scheme of the application is as follows:

[0007] The method comprises the following steps:

[0008] Step 1: A plurality of high-precision laser ranging terminal nodes are arranged at equal intervals on both sides of the center line of the to-be-built highway, the interval between the nodes is a pre-set fixed design interval, the transmitting end of each node transmits a ranging beam processed by a light beam multi-frequency modulation coding sequence at a frequency of one thousand times per second, and the absolute three-dimensional coordinates of the nodes are recorded synchronously to establish a unified reference three-dimensional coordinate system;

[0009] Step 2: The receiving end of each node receives the original echo signal returned by the pavement scattering; the original echo signal is subjected to multi-scale wavelet transform to generate a multi-scale power spectrum tensor; the multi-scale power spectrum tensor is separated into a signal tensor and a noise tensor by using a tensor sparse kernel decomposition method, and a three-dimensional similarity spectrum graph is constructed; a variable weight Lorentz manifold measure clustering is applied on the three-dimensional similarity spectrum graph to obtain a characteristic cluster with maximum spectral coherence; the characteristic cluster is input into a deep time sequence gated recurrent network to obtain a predicted displacement settlement combined deviation of the next period; the current displacement settlement combined deviation and the predicted displacement settlement combined deviation are aggregated in time sequence to form a construction quality multi-dimensional parameter matrix to realize real-time high-dimensional representation of the current construction quality state;

[0010] Step 3: The construction quality multi-dimensional parameter matrix is subjected to principal component analysis to generate a highway construction quality detection result.

[0011] Further, in step 1, the process of light beam multi-frequency modulation coding sequence processing includes: in a unified frame time slot, according to a preset multi-layer frequency domain stack layer rule, dividing a target transmission frequency band into a plurality of sub-frequency bands; inserting a dedicated start-stop synchronization identifier into each sub-frequency band, so that the sub-frequency bands are aligned in the time-frequency plane, forming a basic light beam multi-frequency modulation coding sequence frame; implementing inter-layer orthogonal code spreading on the basic light beam multi-frequency modulation coding sequence frame, and mapping each sub-frequency band to a pseudo-random spreading code orthogonal to each other; through inter-code cyclic shift, realizing self-synchronous interference suppression between different sub-frequency bands, and ensuring that there is no cross talk during multi-frequency parallel transmission; performing time-frequency interleaving rearrangement on the light beam multi-frequency modulation coding sequence frame subjected to orthogonal code spreading by using a preset interleaving matrix; in the time-frequency interleaving rearrangement process, alternately exchanging positions of adjacent time slots and adjacent sub-frequency bands to obtain an interleaved sequence, bipolar phase randomization is performed on the interleaved sequence; introducing a pseudo-random phase flip mode, embedding a phase randomization mask into the light beam multi-frequency modulation coding sequence frame header; inserting a multi-level self-correcting error correction embedded code segment into the light beam multi-frequency modulation coding sequence frame in which the phase randomization mask has been embedded, including a frame-level block check code and a cross-frame-level cyclic check code; the embedded code segment adopts a combination of flip parity check and redundancy segmentation; performing spectrum shape robust suppression shaping on the light beam multi-frequency modulation coding sequence frame in which the embedded code is completed; using a mixed shaping strategy of symmetric Hann window and raised cosine window, bidirectional decreasing shaping is performed on the sideband energy of each sub-frequency band to reduce the emission spectrum leakage; and an adaptive suppression condensation zone is inserted in the spectrum center to suppress sidelobe noise caused by multipath echoes.

[0012] Further, the process of inserting an adaptive suppression condensation zone in the spectrum center includes: extracting the overall spectrum envelope of the light beam multi-frequency modulation coding sequence frame in which the embedded code is completed in real time, determining the spectrum center position and the symmetric energy threshold surface; according to the symmetric energy threshold surface, opening the adaptive suppression condensation zone start-stop boundary in the spectrum center, and setting a gradual connection band on both sides of the boundary; performing amplitude decreasing processing on the inside of the condensation zone according to the adaptive suppression strategy, and performing smooth transition processing on the connection band to make the spectrum shape continuous and consistent; in the subsequent transmission period of the light beam multi-frequency modulation coding sequence frame, the width and the decreasing depth of the condensation zone are dynamically corrected using the echo monitoring result to keep the sidelobe noise continuously low; whenever a sudden change in the characteristics of the multipath echo is detected, the adaptive reconstruction of the condensation zone is triggered immediately to ensure that the emission spectrum shape and the suppression effect are updated synchronously.

[0013] Further, the process of implementing inter-layer orthogonal code spreading on the base light beam multi-frequency modulation coding sequence frame includes: loading a unique and pairwise orthogonal pseudo-random spreading code index field in each sub-band of the light beam multi-frequency modulation coding sequence frame; performing inter-code cyclic shift with the frame sequence number as the seed, and shifting the pseudo-random spreading code of adjacent sub-bands clockwise by a fixed code relative to the previous frame; setting an adaptive trigger gate at the transmitting end of the node, and immediately latching the bit sequence when the inter-code cyclic shift is detected to be completed, thereby realizing self-synchronization of the sub-band boundary; at the receiving end of the node, the pseudo-random spreading code is recovered in reverse according to the same cyclic shift sequence, thereby automatically canceling the cross-sub-band crosstalk; if the inter-code alignment deviation is detected to be over the threshold, the next cyclic shift period is immediately triggered, thereby dynamically maintaining the mutual non-crosstalk of multi-frequency parallel transmission.

[0014] Further, the process of separating the multi-scale power spectrum tensor into a signal tensor and a noise tensor by using the tensor sparse kernel decomposition method in step 2, and constructing a three-dimensional similarity spectrum map includes: performing multi-directional resampling on the multi-scale power spectrum tensor to establish an initial index table of the tensor sparse kernel; in the initial index table of the tensor sparse kernel, weakly associated kernel blocks are iteratively removed according to a sparse density threshold, and the remaining kernel blocks are aggregated to generate a signal tensor; the removed kernel blocks are simultaneously captured to form a noise tensor; the signal tensor is expanded along the frequency domain, the scale domain and the time domain layer by layer, the three-way similarity weight is calculated, and a primary association matrix is generated; the three-dimensional similarity spectrum map is obtained by applying segmented progressive mapping and cross-stretch processing on the primary association matrix.

[0015] Further, the process of applying variable weight Lorentz manifold measurement clustering on the three-dimensional similarity spectrum map in step 2 to obtain a feature cluster with the maximum spectral coherence includes: constructing an initial grid of the Lorentz manifold based on the spectral density in the entire domain of the three-dimensional similarity spectrum map, and assigning a variable weight vector to each grid node; performing double-threshold adaptive update on the variable weight vector according to the local spectral gradient of the node to form a multi-scale weight field; performing layer-by-layer equidistant expansion search on the grid nodes under the variable weight Lorentz manifold measurement, and recording the manifold measurement distance between nodes in real time; iteratively merging node sets according to the principle of minimum manifold path and maximum weight gain to generate a candidate feature cluster set; performing spectral coherence increment screening on the candidate feature cluster set, and outputting the feature cluster with the maximum spectral coherence.

[0016] Further, the process of constructing the initial grid of the Lorentz manifold based on spectral density in the whole domain of the three-dimensional similarity spectrum includes: decomposing the three-dimensional similarity spectrum into continuous density layers according to the spectral density grading rule; performing bidirectional Manhattan scanning on each density layer to generate equipotential surfaces, and mapping the equipotential surfaces into curved surface grid pieces by using the Lorentz coordinate embedding method; directionally weaving along the spectral gradient tangent to seamlessly splice all the curved surface grid pieces to form the initial grid of the Lorentz manifold covering the whole domain; synchronously collecting the neighborhood spectral density, local curvature and time similarity at each grid node to splice into the original feature vector; converting the original feature vector into a variable weight vector through a normalization and reweighting cycle; and writing the variable weight vector into the grid node metadata to complete the weight initialization of the grid node.

[0017] Further, the process of performing bidirectional Manhattan scanning on each density layer to generate equipotential surfaces, and mapping the equipotential surfaces into curved surface grid pieces includes: selecting the boundary of the density layer as the starting reference line, and setting bidirectional Manhattan scanning paths in the horizontal and vertical directions; collecting density threshold surface intersection points along the horizontal scanning path row by row, and marking equipotential surface candidate nodes in real time; switching to the vertical scanning path after completing the horizontal scanning, and supplementing the equipotential surface gap nodes column by column to form a bidirectional closed node network; performing neighborhood connectivity detection on the bidirectional closed node network to remove isolated nodes to obtain a continuous and smooth equipotential surface; establishing a local Lorentz tangent plane on the equipotential surface, and converting the equipotential surface segment into a curved surface parameter domain according to the Lorentz coordinate embedding method; sequentially splicing the curved surface segments in the curved surface parameter domain to generate a completely conformal curved surface grid piece; and writing the density layer identifier and spatial index for each curved surface grid piece to complete the global registration of the curved surface grid piece.

[0018] Further, step 3 specifically includes: performing dimension-wise normalization on the multi-dimensional parameter matrix of construction quality to generate a standardized matrix and reserve an inverse normalization index; constructing a sliding covariance grid, updating the covariance weight network according to the time sequence correlation, and inputting the obtained network into a principal component analysis core engine; the principal component analysis core engine extracts a principal axis vector set according to a progressive variance absorption rate criterion, and implements orthogonal rotation to obtain stable principal axes; performing cascade coordinate transformation on the standardized matrix under the stable principal axes to obtain a compressed feature score sequence; performing adaptive contribution aggregation on the compressed feature score sequence according to the time scale to output a comprehensive quality index column; and performing hierarchical comparison between the comprehensive quality index column and a design tolerance library to mark the qualified state and the abnormal state, and form a highway construction quality detection result.

[0019] The highway construction quality dynamic detection method based on high-precision laser ranging has the following beneficial effects: the highway construction quality dynamic detection method based on high-precision laser ranging realizes continuous, real-time and high-resolution quality state capture and evaluation of key structural layers in the construction process on the basis of the prior art. By arranging high-precision laser ranging terminal nodes at equal intervals on both sides of the highway center line and performing multi-frequency modulation coding sequence laser scanning on the ground at a transmission frequency of 1000 times per second, the application can obtain the absolute three-dimensional coordinate changes of the ground scattering points in real time with millimeter-level precision, thereby effectively identifying the construction states such as compaction, filling, settlement and uplift. Compared with the traditional measurement method mainly based on low-frequency static sampling, the application realizes continuous acquisition of high-temporal and spatial resolution data, and overcomes the problems of long detection period and lack of dynamic change information. The application can extract signal components with physical representation from a large amount of unstable echo data by constructing a unified reference three-dimensional coordinate system and performing multi-scale wavelet transform and tensor sparse kernel decomposition on the original echo signal, thereby significantly enhancing the anti-interference ability. At the same time, by using the three-dimensional similarity spectrum diagram and the variable weight Lorentz manifold degree measurement clustering method, the spatial coherent feature clusters in the construction process are effectively identified, thereby providing highly relevant input feature sequences for subsequent prediction based on the deep time sequence gated recurrent network. Further, by constructing a multi-dimensional parameter matrix of construction quality and performing principal component analysis processing, the application can realize multi-dimensional comprehensive evaluation and qualification determination of construction quality, and form a digital quality index. The index has the characteristics of quantification, traceability and partition labeling, and is suitable for whole-process quality supervision, abnormal early warning and construction decision feedback. Therefore, the application has significant engineering application value and technical popularization prospect in aspects of guaranteeing construction precision, improving engineering quality control efficiency and supporting intelligent construction site construction. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The method flowchart of the highway construction quality dynamic detection method based on high-precision laser ranging provided by the embodiment of the application is shown in the figure.

[0021] Figure 2 The characteristic distribution and processing effect diagram of the light beam multi-frequency modulation coding sequence in the frequency domain in the embodiment of the application is shown in the figure.

[0022] Figure 3 The process diagram of the multi-scale power spectrum tensor sparse kernel decomposition in the embodiment of the application is shown in the figure.

[0023] Figure 4 The process diagram of the variable weight Lorentz manifold degree measurement clustering applied on the basis of the three-dimensional similarity spectrum diagram in the embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0024] In order to make the person skilled in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the scope of protection of the present application.

[0025] Reference Figure 1 The highway construction quality dynamic detection method based on high-precision laser ranging comprises:

[0026] Step 1: A plurality of high-precision laser ranging terminal nodes are arranged at equal intervals on both sides of the center line of the to-be-built highway, the interval between each node is a pre-set fixed design interval; the transmitting end of each node transmits a ranging beam processed by beam multi-frequency modulation coding sequence at a frequency of one thousand times per second, and synchronously records the absolute three-dimensional coordinates of the node, and establishes a unified reference three-dimensional coordinate system;

[0027] Step 2: The receiving end of each node receives the original echo signal returned by the road surface scattering; the original echo signal is subjected to multi-scale wavelet transform to generate a multi-scale power spectrum tensor; the tensor sparse kernel decomposition method is used to separate the multi-scale power spectrum tensor into a signal tensor and a noise tensor, and a three-dimensional similarity spectrum graph is constructed; variable weight Lorentz manifold measure clustering is applied on the three-dimensional similarity spectrum graph to obtain a characteristic cluster with maximum spectral coherence; the characteristic cluster is input into a deep time sequence gated recurrent network to obtain a predicted displacement settlement joint deviation of the next period; the current displacement settlement joint deviation and the predicted displacement settlement joint deviation are aggregated in time sequence to form a construction quality multi-dimensional parameter matrix to represent the current construction quality state in real time;

[0028] Step 3: The construction quality multi-dimensional parameter matrix is subjected to principal component analysis to generate a highway construction quality detection result.

[0029] The laser ranging body relies on phase accumulation and time-of-flight dual determination mechanism. When the node transmitting end outputs the ranging beam at a frequency of 1000 times per second, the phase accumulation can lock the sub-wavelength level change, and the time-of-flight gives an absolute distance reference, which complementarily solves the integer wavelength ambiguity, thereby maintaining the ranging consistency on the order of centimeters or even millimeters. In order to avoid the ranging jitter caused by phase jump, the beam is processed into a beam multi-frequency modulation coding sequence, which simultaneously occupies multiple channel resources in the time domain, frequency domain and code domain, and through orthogonal distribution, the channels in the same node are not interfered with each other, and through cyclic shift, the interference of adjacent nodes is avoided. The synchronization identification and error correction embedded code are embedded at the transmitting end, which ensures that the echo identification and error correction can be completed at the receiving end. The spectral shape robust suppression shaping and the design of inserting an adaptive suppression condensation zone in the center of the spectrum are derived from the understanding of the mechanism of multipath echo: multipath causes sidelobe noise to concentrate in the frequency band adjacent to the main lobe, and dynamic weakening of the energy at the edge of the main lobe in the center condensation zone can reduce the relative height of the sidelobe and improve the feasibility of subsequent signal separation.

[0030] When the node transmitting end outputs the ranging beam at a frequency of 1000 times per second, the ranging result of the first time is:

[0031] ;

[0032] Wherein, is the number of subcarriers in the beam multi-frequency modulation coding sequence; is the speed of light; is the adaptive amplitude weight corresponding to the subcarrier signal-to-noise ratio; is the echo phase of the th subcarrier at the receiving end in the current frame; is the carrier initial phase of the th subcarrier at the transmitting end in the current frame; is the round-trip time of flight of the th subcarrier in the current frame; is the center frequency of the th subcarrier.

[0033] The signal tensor carries the three-way similarity information in the frequency domain, scale domain and time domain. The traditional Euclidean metric cannot describe the curvature difference of the three similarities at the same time. This technology introduces the Lorentz manifold metric, whose time axis has a negative weight feature. It can shorten the node distance in the area with high similarity and fast evolution speed, and lengthen the node distance in the area with low similarity and slow evolution speed, avoiding the dilution of high-density clusters by time dragging. In actual implementation, it is necessary to first construct the initial mesh of the manifold on the three-dimensional similarity spectrum graph. In order to make the mesh fit the high-density area, the system performs bidirectional Manhattan scanning on each density layer, alternately collects isopotential nodes in the horizontal and vertical directions, and then uses the Lorentz coordinate embedding method to map the discrete isopotential surface into a curved mesh sheet. Finally, the sheet-shaped mesh is stitched along the spectral gradient tangent. Each mesh node collects spectral density, local curvature and time similarity, and after normalization, a variable weight vector is formed. In the clustering stage, the variable weight vector is searched in the manifold metric space according to the equal distance expansion, and the clustering front is expanded only in the high similarity area by the minimum manifold path constraint, and the class fusion is driven by the maximum weight gain, so as to obtain the feature cluster with the maximum spectral coherence.

[0034] The feature cluster contains the core mode of the pavement micro-deformation. In order to predict the pavement settlement and displacement deviation in the future period of time, a network structure capable of modeling long-range dependence and selectively remembering local mutations is needed. The deep time sequence gated recurrent network uses the gating unit to determine the degree of information retention or forgetting, and stacking multiple layers can capture trends and oscillations at different time scales. When the network input is the current time sequence of feature clusters, its output is the predicted displacement settlement joint deviation in the next period. The system splices the current displacement settlement joint deviation and the predicted displacement settlement joint deviation in time sequence to obtain the construction quality multi-dimensional parameter matrix. The rows and columns of the matrix correspond to the three dimensions of spatial position, time window and state index respectively, and the matrix values depict various state quantities such as displacement, settlement, amplitude stability, coherence diffusion degree, etc. The comprehensive dimension makes the matrix can simultaneously present the horizontal uniformity, vertical continuity and time stability of the current construction state.

[0035] However, the multi-dimensional parameter matrix data is huge, and directly used for decision-making will bring redundancy and overfitting risk. In this method, principal component analysis is used to reduce the dimension of the matrix. The analysis process first unifies the dimension in all dimensions to ensure that the variance measure is comparable, and then a sliding covariance grid is established to adjust the influence degree of near and far data with time sequence weight. Principal component extraction is performed along the increasing principle of variance absorption rate, and the proportion of extracted amount to matrix variance information is dynamically adjusted. In the high volatility stage, the number of principal components is reduced to highlight significant events, and in the stable stage, the number of principal components is increased to maintain fine-grained monitoring. The stable principal axis vector and the matrix product obtain the compressed feature score sequence, and the system generates a comprehensive quality index column by adaptively aggregating the contributions of the sequence. The comprehensive quality index column is compared with the construction design tolerance library in layers, and if the index is within the threshold of each layer, the corresponding construction state is marked as qualified, otherwise it is marked as abnormal and the deviation level is output. The continuous fluctuation trend of the index column can also indicate the gradual deterioration or improvement of quality, providing guidance for on-site decision-making.

[0036] Further, the system cuts the complete frequency band into several sub-frequency bands in a unified frame time slot, and each sub-frequency band has a clear and constant bandwidth interface. This hard division ensures that different nodes do not have cross-band interference even if there is slight local oscillator drift in the case of high-density node parallel layout. Then, a special start and end synchronization identifier is inserted for each sub-frequency band. The start synchronization identifier provides a reference for the transmission start point, and the end synchronization identifier determines the frame tail alignment. The two identifiers together form an alignment reference that can be observed in both time and frequency domains, so that all sub-frequency bands remain aligned in the time-frequency plane and thus form a basic light beam multi-frequency modulation coding sequence frame. Inter-layer orthogonal code spreading is implemented within the basic light beam multi-frequency modulation coding sequence frame. Each sub-frequency band is mapped to a unique and pairwise orthogonal pseudo-random spreading code to achieve uniform energy spreading through the cyclic characteristics of the pseudo-random sequence, thereby reducing the possibility of single-frequency peak interference; the orthogonal characteristic also ensures that the cross-correlation between different layers in the same frame is approximately zero.

[0037] In order to eliminate the low-frequency beat effect between the pseudo-random spread code and the modulated sub-carrier in long-term operation, the system introduces inter-code cyclic shift, which performs a circular shift on the pseudo-random spread code between consecutive frames, so that the spread code of adjacent sub-frequencies is shifted by a fixed number of bits clockwise or counterclockwise relative to the previous frame. This shift creates a self-synchronous interference suppression mechanism: when the echo of a certain sub-frequency leaks into the adjacent frequency, the receiving end of the adjacent frequency can easily determine the source of the interference due to the correlation shift caused by the circular shift, and can eliminate it in the signal reconstruction stage, thereby ensuring that the multiple frequencies are not interfered with each other when transmitting in parallel. After completing the inter-layer orthogonal code spreading, the system reorders the coded frames according to the preset interleaving matrix, which is designed according to the principle of "row and column coprime, diagonal crossing", so that adjacent time slots and adjacent sub-frequencies are alternately exchanged. After interleaving, the sequence becomes a chessboard distribution in the time-frequency two-dimensional domain, which not only disperses the possible concentrated periodic phase noise, but also averages the influence of channel instantaneous fading on subsequent decoding. Then, the interleaved sequence is subjected to bipolar phase randomization. The specific method is to introduce a pseudo-random phase flipping pattern and write a phase randomization mask into the beam multi-frequency modulation coding sequence frame header. The mask is updated once per frame. By randomly switching the transmission phase between 0 degrees and 180 degrees, it is difficult for any potential interceptor to reconstruct the effective information even if they synchronize to the carrier. The receiving end performs inverse flipping according to the frame header mask to achieve lossless recovery.

[0038] In order to further improve the error code flexibility, a multi-level self-correcting error correction code segment is embedded in the beam multi-frequency modulation coding sequence frame which has embedded a phase randomization mask. The frame-level block check code is responsible for locating short-time burst errors, and the cross-frame level cyclic check code provides boundary protection for randomly distributed sparse errors. The combination of the two takes into account the local correction speed and global redundancy efficiency. The combination of flipping parity check and redundant segment within the embedded code segment can maintain stable error correction ability under the scattering fluctuation caused by changes in the route environment. After completing the embedded code, the system performs spectrum shape robust suppression shaping on the beam multi-frequency modulation coding sequence frame. The symmetric Hanning window and the lifting cosine window are mixed and shaped simultaneously to the sideband energy of each sub-frequency. The bidirectional decreasing shaping makes the spectrum shape like a clock-shaped landslide, with fast edge energy decay and high central energy. This not only reduces the emission spectrum leakage, but also reduces the adjacent frequency interference.

[0039] Considering that actual highway construction site is often accompanied by metal reflecting surface generated by concrete mixer, road roller and other large mechanical equipment, which causes multi-path echoes to concentrate in the frequency band near the main lobe to form sidelobe noise peak, the application inserts adaptive suppression condensation zone in the center of the frequency spectrum. The system first extracts the envelope of the transmission spectrum shape in real time and determines the center energy threshold surface, then dynamically reduces the main lobe width in the center threshold surface, further gathers energy to the center and at the same time makes smooth transition in the transition band. Since the width and the depth of the condensation zone are adaptively adjusted according to the echo monitoring results, the sidelobe noise suppression effect can be maintained stable for a long time. When the characteristics of multi-path echoes are monitored to change suddenly, the condensation zone is immediately reconstructed to ensure that the transmission spectrum shape and the suppression strategy are updated synchronously. The above continuous multi-link combination enables the beam multi-frequency modulation coding sequence processing to still provide high signal-to-noise ratio, high error correction margin and low crosstalk performance in the harsh highway construction site environment, providing a highly distinguishable original signal for the subsequent multi-scale wavelet transform and tensor sparse kernel decomposition links. With the help of this signal, the system can compare and measure the distance change between nodes in real time, capture the micron to millimeter level of road surface settlement and displacement information, so that the highway construction quality dynamic detection method has both real-time and accuracy, and provides reliable quality trend data support for project managers.

[0040] Further, in order to suppress the sidelobe noise introduced by multi-path echoes in the frequency spectrum and ensure that the ranging beam maintains high signal-to-noise ratio for a long time, the system performs a special processing of inserting an adaptive suppression condensation zone in the center of the frequency spectrum on the beam multi-frequency modulation coding sequence frame completed by the embedded code in the last stage of the entire transmission link. The core principle is to capture the overall spectrum envelope of the transmission signal in the frequency domain in real time, first accurately locate the center point with the highest energy distribution and the most complete symmetry, then determine the symmetric energy threshold surface on both sides of the point, use the threshold surface as the basis to draw the start and end boundaries in the center of the frequency spectrum, and form an adaptive suppression condensation zone. Inside the adaptive suppression condensation zone, the main lobe edge energy is forced to be suppressed according to the amplitude decreasing rule from the center to the outside, and a gradual transition band is constructed outside the boundary, and smooth transition is made in the transition band area, so that the suppression curve is continuous and consistent on the overall spectrum shape, and no new sharp corner or abrupt depression is produced.

[0041] To ensure that this suppression process does not fail over time due to drift, the system uses the results of the node receiver end echo monitoring to dynamically correct the width and decreasing depth of the adaptive suppression condensation zone at each transmission cycle. The correction logic is driven by the real-time ratio of the sidelobe noise peak value to the main lobe peak value in the echo. When the ratio rises, indicating that the sidelobe is rising, the system immediately expands the condensation zone width or increases the decreasing gradient. When the ratio decreases, it is appropriately relaxed to avoid excessive suppression leading to loss of effective power of the main lobe. This adaptive process is fully automated and does not require on-site human intervention, so it can maintain a low level of sidelobe noise in the construction environment where large equipment such as road rollers, watering trucks, and reflective surfaces change instantaneously. Further, to deal with the sudden change in multipath echo characteristics caused by the removal of metal templates or the exposure of steel reinforcement cages during the concrete pouring stage, the system sets up an event detection unit inside the echo monitoring module. The unit uses a change rate threshold to determine whether a mutation has occurred. Once it is determined to be true, it sends a reconstruction instruction to the transmitter. The transmitter immediately triggers the adaptive reconstruction of the adaptive suppression condensation zone. The reconstruction process first clears the existing condensation zone parameters, then recalculates the center, threshold surface and boundary using the new spectrum envelope after the mutation, and then updates the real-time control table of the shaper according to the latest amplitude decreasing strategy, so as to ensure that the transmission spectrum shape and suppression effect are always synchronized. In this way, no matter what changes in the reflective environment occur during the road construction stage, the laser ranging link can suppress the sidelobe noise introduced by multipath to a constant and controllable low level, providing clean and stable raw data sources for subsequent high-order signal processing links such as multi-scale wavelet transform, tensor sparse kernel decomposition, and variable weight Lorentz manifold metric clustering, making the calculation of the predicted displacement settlement combined deviation and the multi-dimensional parameter matrix of the construction quality more accurate, and making the road construction quality detection results output by the principal component analysis more reliable.

[0042] Further, the inter-layer orthogonal code spreading of the base beam multi-frequency modulation code sequence frame is a key process that determines the signal purity and node mutual interference ability. The system first writes a unique and pairwise orthogonal pseudo-random spreading code index field in each sub-band divided by the beam multi-frequency modulation code sequence frame. The orthogonal characteristics isolate the different sub-bands transmitted by the same node into independent channels with extremely low cross-correlation, ensuring that the receiving end can still locate the exact sub-band attribution by the correlation peak in a high scattering noise environment. Then the system performs inter-code cyclic shift using the frame sequence number as the seed, and shifts the pseudo-random spreading code of adjacent sub-bands clockwise by a fixed code bit relative to the previous frame. This bit shift operation forms a ring arrangement of the spreading code rolling with the frame number on the time axis, which is equivalent to continuously refreshing the relative misalignment angle across the sub-bands at a transmission rhythm of one thousand times per second, so that any interference from a fixed position cannot maintain stable coupling in consecutive frames. In order to align this dynamic misalignment with the local reference of the node, the system specially sets an adaptive trigger gate at the transmitting end of the node. The adaptive trigger gate monitors the displacement progress in real time, and immediately latches the bit sequence when it detects that the inter-code cyclic shift is complete. The sub-band boundary at the current time is written into the local register and broadcast to the receiving front end of the same node through the synchronization bus, forming a physical layer self-synchronous closed loop.

[0043] After the receiving end obtains the latched bit sequence, it reversely recovers the pseudo-random spreading code according to the same cyclic shift sequence, reconstructs the original orthogonal relationship through the correlator array, thereby automatically canceling the cross-sub-band interference, while still preserving the maximum channel capacity using the orthogonality when there is frequency offset or amplitude imbalance. When large machinery moves or steel formwork is dismantled on the construction site, causing the number of multipath reflection paths to suddenly increase, the correlation peak group will show a sudden rise in the statistical quantity, which is the alignment deviation of the spreading code. When the system detects that the inter-code alignment deviation exceeds the threshold, it determines that the current cyclic shift strategy cannot meet the isolation requirements, and immediately triggers the next cyclic shift period. The spreading code index field is iterated forward by a fixed code bit, and the adaptive trigger gate resets the latched counter, ensuring that the mutual interference performance of multi-frequency parallel transmission is updated in real time with the environment. Since the entire process is completely driven by the frame sequence number and relies on the hardware clock, the high-precision laser ranging signal can still maintain a stable pseudo-random orthogonal pattern under complex electromagnetic and optical conditions on the highway site, providing high signal-to-noise, high orthogonality, and low mutual interference for the subsequent multi-scale wavelet transform, tensor sparse kernel decomposition, and variable weight Lorentz manifold metric clustering calculation stages. This ensures that the estimation accuracy of the predicted displacement settlement joint deviation and the multi-dimensional parameter matrix of the construction quality remains at the millimeter or even micrometer level, further improving the reliability and real-time performance of the dynamic detection results of highway construction quality.

[0044] Further, the node receiving end first performs continuous-time sampling on the road surface scattering echo and completes multi-scale wavelet transform, projects the energy bands and phase shift at different scales to the two-dimensional grid formed by the intersection of the frequency domain and the scale domain, and then stacks them to form a three-dimensional multi-scale power spectrum tensor. This tensor is distributed in the spatial dimension as a block, and contains long-range evolution traces in the time dimension. In order to enable the subsequent algorithm to grasp the core structure that truly reflects the micro-deformation of the road surface, the system first performs multi-directional resampling on the multi-scale power spectrum tensor. The multi-directional resampling adjusts the sampling step along the frequency domain, scale domain and time domain at the same time, compresses the original tensor with high resolution and containing redundant noise to the minimum information loss range, thereby reducing the computational burden and eliminating high-frequency false textures. Then the system establishes an initial index table for the sparse kernel of the tensor, which divides the tensor into equal-volume kernel blocks and records the energy density, adjacent correlation and time persistence of each kernel block.

[0045] The core idea of sparse kernel decomposition is to assume that the useful information inside the signal tensor presents strong block correlation within a three-dimensional neighborhood, while the environmental noise is scattered in the form of discrete points or fine pieces. Therefore, the system sets a sparse density threshold in the initial index table of the tensor sparse kernel, and uses an iterative exclusion strategy to gradually exclude weakly correlated kernel blocks. In each iteration, the correlation weight of all kernel blocks and their three-dimensional neighborhood is calculated first, and then the weakly correlated kernel blocks below the sparse density threshold are marked and deleted, while the remaining kernel blocks are marked as retained kernel blocks. The deletion action synchronously captures the weakly correlated kernel blocks that are excluded, and aggregates them in real time to form a noise tensor; the retained kernel blocks are re-aggregated to generate a signal tensor at the end of each round. This iterative process continues until the change rate of the number of retained kernel blocks is below a predetermined convergence threshold, ensuring that the final signal tensor only contains three-dimensional energy blocks with high correlation, high sparsity and time sequence continuity, while the noise tensor collects low-energy and low-correlation information fragments. Subsequently, the system expands the signal tensor along the frequency domain, scale domain and time domain, calculates the three-way similarity weight between kernel blocks through a sliding window, and forms a primary correlation matrix covering the entire tensor.

[0046] The primary correlation matrix still retains the multi-resolution characteristics of the original tensor, but the weight distribution in it may present discrete stripes visually, which is difficult to be directly used for manifold clustering. To make the correlation information more coherent and available for subsequent variable weight Lorentz manifold metric clustering, the system applies piecewise progressive mapping and cross-stretch processing on the primary correlation matrix. The piecewise progressive mapping divides the weight interval into multiple levels, and performs dynamic gain or compression on each level according to a nonlinear mapping function, so that the low weight region is moderately lifted and the high weight region is softened, avoiding the dominance of extreme values on the overall structure. The cross-stretch processing alternately selects a reference line between the frequency-scale domain plane and the time axis, simultaneously stretching the local high weight ridge along two orthogonal paths, and filling the gaps in the primary correlation matrix caused by sparse sampling. After these two steps of remodeling, the primary correlation matrix is transformed into a three-dimensional similarity spectrum, which visually presents continuous and smooth energy mountains and valleys, not only retaining the frequency-scale fingerprints of pavement deformation, but also strengthening the time evolution context with stretched structure. At this time, the three-dimensional similarity spectrum has all the conditions to provide high-resolution input for variable weight Lorentz manifold metric clustering, making the subsequent feature cluster extraction obtain clear and noise-free spectral coherence peaks. Through this whole tensor sparse kernel decomposition process, the method purifies the signal tensor representing the real pavement settlement and displacement pattern from the complex background noise on the construction site, while constructing a structured and clusterable three-dimensional similarity spectrum, laying a solid data foundation for dynamic prediction and quality evaluation. This separation strategy purely relying on the internal sparse characteristics and similarity topology of the tensor does not need to introduce prior environmental models, and is suitable for different seasons, different materials and different combinations of mechanical equipment in construction conditions, so it significantly improves the robustness and universality of the highway construction quality dynamic detection method to changes in site conditions.

[0047] Spectral coherence evaluation index of three-dimensional similarity spectrum is:

[0048] ;

[0049] wherein, is the power spectral density of the signal tensor at frequency band index , scale index , and time index ; is the power spectral density of the noise tensor at the corresponding position; , and are the discrete division numbers of the frequency domain, scale domain, and time domain, respectively. is the power spectral density of the noise tensor at the corresponding position.

[0050] Further, the three-dimensional similarity spectrum assumes the task of explicitly geometrizing the high-order correlation of the signal tensor, while the variable weight Lorentz manifold metric clustering is responsible for extracting the feature cluster that best represents the micro-deformation evolution law of the road surface from the geometric image. The system first constructs the initial grid of the Lorentz manifold based on the spectral density in the global domain of the three-dimensional similarity spectrum. The grid construction process regards the spectral density as the source of spatial curvature, and the grid nodes in the area with high curvature are more dense, and the grid nodes in the area with low curvature are appropriately sparse, so that the grid naturally fits the energy distribution. Then, a variable weight vector is assigned to each grid node, which is composed of the spectral density, local curvature and time similarity at the node location after normalization, and the vector dimension is one-to-one corresponding to the three axes of the three-dimensional similarity spectrum, so as to unify the importance of different scales into the same metric framework. In order to let the weight automatically adjust with the local structure, the system performs double-threshold adaptive update on the variable weight vector according to the local spectral gradient of the node: when the spectral gradient is lower than the first threshold, it is considered that the node is in the plateau area, and the weight slowly decreases to avoid overemphasis on the flat background; when the spectral gradient is higher than the second threshold, it is considered that the node is located in the ridge or crack, and the weight rapidly rises to highlight the significant features; the region between the two thresholds adopts linear interpolation for smooth transition. After updating, the variable weight vector presents a multi-scale weight field in the three-dimensional space, and the continuity of the weight field ensures that the weight change on any curved surface is homologous to the spectral gradient change.

[0051] Under the variable weight Lorentz manifold metric, the system implements layer-by-layer equidistant expansion search on the grid nodes. The Lorentz manifold metric adopts a negative weight form in the time axis, which physically means that the nodes with high similarity and fast time evolution are pulled closer, and the nodes with low similarity and slow time evolution are pulled farther, thereby maintaining the spatial aggregation degree of rapidly changing features. In the process of layer-by-layer equidistant expansion search, the system starts from the node with the highest spectral density peak value, expands the search layer outward according to the manifold equidistance principle, and records the manifold metric distance between nodes in real time at each expansion step. The manifold metric distance depends on both the geometric path length and the variable weight difference, so it can bypass the valleys with steep weight drops in space, and quickly move along the ridges with similar weights. When the expansion boundary touches the manifold metric distance threshold, the current search layer stops, and the system starts to merge the node set according to the minimum manifold path and maximum weight gain principle among the visited nodes, generating a candidate feature cluster set. The merging adopts a greedy strategy: preferentially connecting the node pair with the shortest manifold path and the maximum weight gain at both ends, avoiding low-value nodes from occupying the cluster center, and preventing nodes with too far distance from destroying internal consistency.

[0052] After obtaining the candidate feature cluster set, the system calculates the spectral coherence of each cluster, which is an index of the consistency of the cumulative energy of all nodes within the cluster on the three-dimensional similarity spectrum. To ensure that the final output feature cluster has both local density and global representativeness, the system performs spectral coherence incremental screening on the candidate feature cluster set: first, sort the spectral coherence in descending order, then evaluate the redundancy coverage between adjacent clusters one by one. If two clusters are highly overlapped in space or time and the spectral coherence difference is less than the set step threshold, the system retains the cluster with higher spectral coherence and discards the cluster with lower spectral coherence; if the overlap is less than the threshold, both clusters are retained to avoid excessive compression leading to loss of key patterns. After multiple rounds of incremental screening, the feature cluster with the highest spectral coherence is determined as the final output. This feature cluster not only contains the most active microscopic settlement patterns of the current stage of the construction pavement, but also ensures high similarity within the cluster in both time and space dimensions with variable weight Lorentz manifold measure, so that it can maximize the activation of the gating unit to capture long-term and short-term dependencies in the subsequent deep time sequence gating recurrent network, thereby improving the accuracy of predicting the combined deviation of displacement settlement. With this clustering process, the highway construction quality dynamic detection method avoids the cluster boundary distortion problem in high curvature spectral space that occurs in traditional Euclidean clustering, and introduces the endogenous adjustment ability of differentiated features in frequency domain, scale domain and time domain through weight self-adaptation, ensuring that the feature cluster with the maximum spectral coherence can be stably extracted in multiple scenarios, multiple materials and all-weather construction sites, thereby injecting high confidence and high resolution basic input into the dynamic detection framework.

[0053] Further, the system first divides the three-dimensional similarity spectrum into continuous density layers according to the spectral density grading rule. The spectral density grading rule follows the principle of equal energy decrease, so that the energy distribution within each density layer is as smooth as possible, and the layers are monotonically decreasing. This hierarchical structure provides a clear topological order for subsequent surface extraction. After obtaining the density layers, the system generates equipotential surfaces by performing bidirectional Manhattan scanning on each density layer. The bidirectional Manhattan scanning alternately moves along the horizontal and vertical directions of the Cartesian coordinate plane, captures the equal-energy contour points through staggered scanning, and fills in the missing nodes in real time, thereby drawing closed and non-intersecting equipotential surfaces for each density layer. To convert the discrete equipotential surfaces into calculable surface entities, the system uses the Lorentz coordinate embedding method to embed the equipotential surfaces into surface mesh sheets. The Lorentz coordinate embedding method uses the negative weight property of the time axis of the three-dimensional similarity spectrum to naturally include the time curvature information of the surface mesh sheet while smoothly embedding it, ensuring that the subsequent manifold measure can perceive the speed difference of dynamic evolution.

[0054] After the generation of the curved mesh patches, the system performs directional weaving along the spectral gradient tangent to seamlessly stitch all the curved mesh patches into a global Lorenz manifold initial mesh. The core idea of directional weaving is to spread the mesh edges along the maximum spectral gradient direction and then cross-stitch in the tangent direction. This avoids the creases caused by manual stitching and preserves the continuous trend of the spectral density main lines. When all the curved mesh patches are stitched, a Lorenz manifold initial mesh is formed, which is topologically identical to the three-dimensional similarity spectrum energy map. The nodes are dense in the high-density ridge peak area and sparse in the low-density valley bottom area, which meets the differentiated needs of subsequent clustering in terms of neighborhood resolution.

[0055] After the establishment of the Lorenz manifold initial mesh, the system synchronously collects the neighborhood spectral density, local curvature, and time similarity for each mesh node. The neighborhood spectral density describes the energy concentration around the node, the local curvature depicts the bending rate of the mesh surface at the point, and the time similarity measures the coherence degree of the signal corresponding to the point over time. The three indicators are orthogonal in space and complementary in information, so they can be combined into the original feature vector. In order to make features of different dimensions comparable in the same dimension space, the system performs a normalization and re-weighting cycle on the original feature vector. In the normalization stage, each dimension of data is mapped to a unified range of zero to one. In the re-weighting stage, the weight coefficients are dynamically adjusted according to the current construction scene, for example, the time similarity weight is appropriately increased during the high-frequency vibration phase of the road roller, and the spectral density weight is increased during the concrete pouring phase. The normalization and re-weighting cycle is iteratively executed multiple rounds on each node until the contribution of all dimensions converges to the designed threshold. This way, the original feature vector is converted into a variable weight vector. The variable weight vector not only encodes the local geometric and dynamic information of the node, but also reflects the adjustment of the external environmental intensity on the intrinsic feature significance during construction, thus greatly improving the adaptability of subsequent metric calculations to different construction conditions.

[0056] In the final step, the system writes the variable weight vector into the mesh node metadata, completing the initialization of the mesh node weights. From this moment on, the Lorenz manifold initial mesh becomes a weighted metric mesh carrying dynamic information, which can not only calculate the distance between nodes in the variable weight Lorenz manifold metric, but also respond in real time to the weight changes caused by subsequent double-threshold adaptive updates. Since the mesh has previously encoded the spectral density, curvature, and time information uniformly, its topological structure and weight distribution are highly consistent with the microscopic settlement pattern of the highway construction site, providing dual sensitivity in terms of geometric shape and time evolution speed for feature cluster extraction. In the entire highway construction quality dynamic detection process, this approach creates a structured and interpretable data foundation for the input of deep temporal models, significantly improving the credibility of the predicted displacement settlement joint deviation and the accuracy of the subsequent comprehensive quality index in reflecting the actual construction quality.

[0057] Further, the algorithm first selects the four peripheral boundaries as the starting reference line for a certain density layer, and the boundary selection follows the principle of energy decreasing step by step, so that the transverse and longitudinal sweeping paths have clear threshold coordinates at the beginning. Then the system moves row by row on the transverse sweeping path, and marks the potential isosurface nodes in real time according to the positions of the multiple intersection points with the density threshold surface on the current row. This way can capture the continuous trend of the energy contour in the transverse direction, while reducing the transverse jumping points caused by discrete sampling. After the transverse sweeping is completed, the longitudinal sweeping path is switched immediately, and the isosurface gap nodes that are not touched in the transverse process are supplemented in the longitudinal way. The two sets of staggered sampling compensate for each other, and finally a bidirectional closed node net is formed in the density layer, which not only retains the high-resolution curved details, but also avoids the omission of isolated islands. In order to ensure that the generated node net does not appear cracks and hanging in subsequent geometric calculations, the system performs neighborhood connectivity detection on the bidirectional closed node net, and removes all isolated nodes and small clusters irrelevant to the main connected branch, only retaining the main structure with a connectivity threshold. In this way, it can ensure that the surface converges normally during the subsequent interpolation fitting, and a continuous and smooth isosurface is obtained.

[0058] After obtaining the isosurface, the system establishes a local Lorenz tangent plane on the isosurface block by block. The tangent plane not only provides the first-order approximation of the surface, but also embeds the negative weight information of the time axis, so that the surface can perceive the difference in the speed of time evolution in a small range. Inside each Lorenz tangent plane, the algorithm uses the Lorenz coordinate embedding method to map the surface segment to a two-dimensional surface parameter domain. The mapping process keeps the angles unchanged, so it can still maintain geometric preservation in areas with dramatic changes in curvature, such as ridges or valleys. All surface parameter domains are sequentially spliced according to the generation order, and the system uses the common control points of the overlapping edges for flexible fusion to avoid scale misplacement due to seams, thereby generating a completely conformal surface mesh piece. Finally, the density layer identifier and spatial index are written for each surface mesh piece, so that it has a globally unique identity. This registration step enables subsequent network construction, weight allocation, manifold measurement, and even depth timing analysis to quickly find the target segment in a unified coordinate system, and corresponds to the precise mapping of the energy terrain and the real road mileage in the construction site. The entire process from sampling, connectivity detection to embedding and registration is completed by pure geometric means, and does not rely on any statistical priori, so it can stably operate on construction surfaces of different materials, thicknesses, and compaction degrees. It provides high-resolution, topologically correct, and time-sensitive surface mesh piece basic data for the dynamic detection framework, greatly improving the credibility of subsequent feature clustering and settlement prediction.

[0059] Further, the system will write the dynamic feature flow returned by each node in real time and purified through multiple stages into a construction quality multi-dimensional parameter matrix, the dimensions of which cover multiple indicators such as spatial position, ranging displacement, settlement trend, amplitude stability, and coherent dispersion. Once the data volume rapidly expands with the progress of construction, dimensional redundancy and inconsistent scales will hinder engineers' quick perception of the overall quality evolution. Therefore, the first step is to perform dimension reduction normalization on the construction quality multi-dimensional parameter matrix. By applying linear mapping and nonlinear compression to columns of different dimensions respectively, the data in each column is scaled to a comparable interval of zero to one, and an inverse normalization index is generated at the same time to facilitate the subsequent back calculation of the dimension reduction result to the original engineering dimension. The standardized matrix obtained after normalization retains all the information but eliminates the dimensional differences, laying a foundation for calculating the synergy between related features. Subsequently, the system constructs a sliding covariance grid along the construction time axis, calculates the sample covariance for each time window and writes it into the grid cell, and then dynamically updates the covariance weight network according to the time correlation: when the correlation between consecutive windows is high, the weight network thickens the connection in space, emphasizing stable patterns; when the correlation drops sharply, the weight network weakens the connection in the corresponding channel, highlighting the sudden events.

[0060] The covariance weight network generated in this way is input into the principal component analysis core engine, which performs eigenvalue decomposition on the sliding covariance grid and selects the principal axis vector set from high to low according to the progressive variance absorption rate criterion. When the cumulative variance coverage rate reaches the preset threshold, it stops extracting, ensuring that only the feature directions that best explain the overall variation are retained. Then the core engine performs orthogonal rotation on the principal axis vector set, making each principal axis independent and interpretable. The stable principal axes obtained after rotation are not disturbed by slight disturbances from single abnormal values during continuous monitoring, so they can represent the baseline trend of construction quality for a long time. The system then performs cascading coordinate transformation on the standardized matrix under the stable principal axes, projecting the high-dimensional state into the principal component space and generating a compressed feature score sequence. This sequence is compressed to tens or even single digits in dimension, but still maintains the millisecond-level refresh frequency of the original data on the time axis, thus balancing real-time performance and visualization convenience. In order to further convert the compressed feature score sequence into a single dimension that is easy to make decisions, the system performs adaptive contribution aggregation on the compressed feature score sequence according to the minute, hour, team, or day time scale defined by the construction schedule: the weight scheduler will refer to the quality risk focus of the current construction phase to dynamically increase or decrease the contribution of each principal component within the same time window, generating a comprehensive quality index column. The comprehensive quality index column is presented in the form of percentage or thousandth, which is a health curve that can be easily understood by project managers.

[0061] Finally, the system performs a hierarchical comparison between the comprehensive quality index and the design tolerance library. The design tolerance library encodes the allowable fluctuation ranges of different processes, materials, and equipment combinations into multi-level threshold intervals. The residence time of the comprehensive quality index within each interval becomes the quantitative basis for quality compliance. If the index exceeds the limit, the system not only marks the abnormal state but also traces the cause of the deviation to the specific original dimension through inverse normalization indexing, indicating which mileage, material, or equipment caused the quality risk. At this point, the highway construction quality inspection results are output in a closed loop. Engineers can observe the index curve and spatial heat map in real time on the visualization panel, and can also export abnormal segments for immediate rectification by on-site construction personnel. This achieves data-driven, refined quality control, maintaining millimeter-level ranging accuracy and second-level quality update cycle throughout the entire process, providing reliable protection for the construction safety and lifespan of high-grade highways.

[0062] Comprehensive Quality Index for:

[0063] ;

[0064] in, For a moment The overall quality index; The number of principal components contributing to the aggregation; For the first Principal components at time Adaptive weights; For the first Principal components at time Feature scoring; This is the reference value for the k-th principal component under the design baseline condition; The tolerance library is designed to provide the permissible deviation of the k-th principal component.

[0065] The following examples use a length of The straight test roadbed was used as the object, and the monitoring time was... Lay out on both the left and right sides of the center line. A high-precision laser ranging terminal node, laterally away from the centerline Vertical axis Left node coordinates Right node coordinates The node has its own three-axis inertial navigation system, initially aligned to a unified reference coordinate system. .

[0066] The multi-level frequency domain stack parameters are set to three levels, with each level containing a certain number of subcarriers. The example only lists the four lowest-level subcarriers. Start and end synchronization identifier time slot width Pseudo-random diffusion code length Frame sequence number from Counting, inter-symbol cyclic shift .

[0067] Single-frame ranging calculation (based on the left node) , No. (Taking a frame as an example) Echo phase difference measurement Round-trip flight times were locked based on relevant peak times. Signal-to-noise ratio Mapped to magnitude weights numerical value .

[0068] Fusion ranging formula ; Substitute have to ; and initialization Comparison, instantaneous rise .

[0069] In continuous The frame window performs a three-level discrete wavelet packet operation on each frame ranging sequence. Frequency band division Scale division The multiscale power spectral tensor is obtained. Multi-directional resampling step size form One core block .

[0070] Define the sparse density of the kernel block Threshold Iteration: 1st Round rejection Block, Reserve ;No. Round rejection Block, Reserve ;No. Wheel retention rate of change Stop. Retain the sum of core block power. Noise power .

[0071] Three-dimensional similarity spectrum and coherence, calculating weights ;make Spectral coherence .

[0072] Lorentzian manifold initial grid, five energy levels, nodes obtained by bidirectional Manhattan sweep of equipotential surfaces. Each node's original feature vector. ;in For neighborhood spectral density, For local curvature, is time similarity. Normalization ; re-weighting . get variable weight vector .

[0073] variable weight lorentz manifold metric clustering, negative weight time axis parameter . manifold distance ;

[0074] layer-by-layer equidistant threshold . search termination gets candidate feature cluster shortest manifold path with maximum weight gain after normalization, greedy merge screening retains the highest coherence cluster, node number .

[0075] deep time series gated recurrent network prediction, gated unit number , history window . input feature sequence ; output next frame prediction .

[0076] construction quality multi-dimensional parameter matrix, matrix column number (node, frequency band, scale, displacement, settlement, coherent diffusion, etc.). write current and predicted deviation, time index .

[0077] sliding covariance and principal component analysis, sliding window length . covariance weight ; variance absorption rate select three principal components .

[0078] comprehensive quality index, adaptive weight . tolerance .

[0079] comprehensive quality index .

[0080] quality determination, threshold setting: ; current , belongs to the anomaly. Reverse normalization positioning finds that the longitudinal displacement residual error corresponding to the principal component is out of limit, position interval there is insufficient compaction. The system issues a rework instruction and records an abnormal event.

[0081] Figure 2The characteristics distribution and processing effect of the light beam multi-frequency modulation coding sequence in the frequency domain are shown in detail. The abscissa represents the frequency range, and the ordinate represents the power spectral density. In the unified frame time slot, according to the preset multi-layer frequency domain stack layer rule, the target transmission frequency band is divided into four main sub-frequency bands, which are marked as sub-frequency band one, sub-frequency band two, sub-frequency band three and sub-frequency band four. Each sub-frequency band presents a specific spectral shape, among which the sub-frequency band one is located in the lower frequency region, and the power spectral density curve presents a clock-shaped distribution of rising first and then falling, with the peak value at about 150Hz. The sub-frequency band two is located in the low-middle frequency region, and its spectral envelope is relatively wide, and the power distribution is more uniform. The sub-frequency band three and the sub-frequency band four are located in the middle-high frequency and high frequency regions respectively, and show the increasing spectral characteristics. At the starting and ending positions of each sub-frequency band, special start and end synchronization marks are inserted, represented by black rectangular blocks, to ensure that each sub-frequency band is accurately aligned in the time-frequency plane. The setting of these synchronization marks enables the basic light beam multi-frequency modulation coding sequence frame to realize strict timing synchronization. The interlayer orthogonal code diffusion effect shown in the figure is reflected by the mutual independence between different sub-frequency bands, and each sub-frequency band is mapped to an orthogonal pseudo-random diffusion code, effectively avoiding the code interference phenomenon. Figure 2 The core feature is the adaptive suppression condensation zone located in the center of the spectrum, marked by a dashed rectangular box. The setting of this condensation zone is to suppress the sidelobe noise caused by multipath echoes. The boundary position of the condensation zone is determined according to the symmetric energy threshold surface, and the inside adopts an amplitude decreasing processing strategy, while the gradual transition bands are set on both sides of the boundary to ensure spectral continuity. The application effect of the mixed shaping strategy of symmetric Hann window and raised cosine window is also shown in the figure, and in the two end regions of the spectrum, the sub-band energy of each sub-band presents a bidirectional decreasing shaping feature, effectively reducing the emission spectrum leakage phenomenon. The whole spectrum analysis diagram clearly reflects the final frequency domain characteristics of the light beam multi-frequency modulation coding sequence after the complete processing process, providing an optimized signal basis for subsequent high-precision laser ranging.

[0082] Figure 3The complete process and effect of multi-scale power spectrum tensor sparse kernel decomposition are presented in three-dimensional form. The three coordinate axes represent the time domain, frequency domain and scale domain, respectively, forming the basic coordinate system of tensor analysis. After multi-directional resampling processing of the original multi-scale power spectrum tensor, the initial index table of tensor sparse kernel is established, providing data basis for subsequent decomposition operation. The cube structure on the left side of the figure represents the signal tensor extracted after sparse kernel decomposition. The internal signal tensor contains multiple black blocks, representing the strong correlation kernel blocks retained after iterative removal of weak correlation kernel blocks according to the sparse density threshold. These kernel blocks show regular distribution in three-dimensional space, reflecting the spatial aggregation characteristics of effective signal components. The cube boundary of the signal tensor is clear, indicating that the signal components after aggregation processing have good spatial continuity and integrity. The cube structure on the right side of the figure represents the noise tensor captured synchronously. Unlike the signal tensor, the noise tensor shows a sparse point distribution inside, represented by small round dots. These noise components are relatively random in space, with low density, fully embodying the incoherence characteristics of noise signals. The three-dimensional similarity spectrum based on the signal tensor is shown on the far right of the figure. The spectrum shows the similarity distribution of different regions in the form of contour lines, and the density and shape of the contour lines reflect the correlation strength of the signal in three-dimensional space. After the segmented progressive mapping and cross-stretching processing of the primary correlation matrix, a three-dimensional similarity distribution pattern with obvious hierarchical structure is formed. The central arrow marks the "tensor sparse kernel decomposition" process, clearly indicating the conversion direction from the original tensor to the signal tensor and the noise tensor. The entire decomposition process embodies the advantages of multi-scale wavelet transform, which can effectively identify and separate signal features at different scale levels, providing a reliable data basis for subsequent feature extraction and quality assessment.

[0083] Figure 4The detailed process and results of applying variable weight Lorentz manifold metric clustering on the basis of three-dimensional similarity spectrum map are presented. The curved grid structure in the figure represents the initial grid of the Lorentz manifold constructed based on the spectral density. These grid lines exhibit non-linear surface characteristics, reflecting the spatial geometric properties in the Lorentz coordinate system. The construction of the grid follows the spectral density grading rule, with each equipotential curve corresponding to a specific density layer. The grid is generated by bidirectional Manhattan sweeping and mapped to the curved grid sheet using the Lorentz coordinate embedding method. Three main feature clusters are identified in the figure: feature cluster A, feature cluster B, and feature cluster C, represented by different size ellipse bounding boxes. Feature cluster A is located in the relatively low spectral coherence region, containing 5 nodes, with a relatively compact ellipse boundary, indicating a high similarity among the nodes in this cluster. Feature cluster B is located in the medium spectral coherence region, containing 6 nodes, with a relatively dispersed distribution, reflecting the diversity characteristics of the data points in this region. Feature cluster C is located in the high spectral coherence region, containing 7 nodes, with the most elongated ellipse boundary, showing the strong correlation of the cluster in a specific direction. The black dots in each feature cluster represent the grid nodes, which are assigned variable weight vectors, represented by short straight line segments emanating from the nodes. The direction and length of the variable weight vectors reflect the characteristics of the local spectral gradient, and these vectors are dynamically adjusted according to the double-threshold adaptive update strategy, forming a multi-scale weight field. The dashed lines in the figure connect the representative nodes between different feature clusters, labeling the "minimum manifold path" and "maximum weight gain", embodying the optimization criteria for node merging in the clustering process. The entire clustering process uses a layer-by-layer equidistant expansion search strategy, recording the manifold metric distance between nodes in real time. Through iterative merging of node sets, a candidate feature cluster set is generated, and finally the feature cluster with the maximum spectral coherence is output through spectral coherence increment screening. The final clustering result shown in the figure fully embodies the superior performance of the Lorentz manifold metric clustering method in handling complex high-dimensional data, providing high-quality feature input for subsequent deep time series gated recurrent network prediction.

[0084] The above-described embodiments are merely intended to illustrate the technical solutions of the present application, not to limit the same; even though the foregoing embodiments of the present application have been described in detail, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A dynamic detection method for highway construction quality based on high-precision laser ranging, characterized in that, The method includes: Step 1: Install several high-precision laser ranging terminal nodes at equal intervals on both sides of the centerline of the highway to be built. The distance between each node is a pre-set fixed design distance. The transmitting end of each node emits a ranging beam processed by a beam multi-frequency modulation coding sequence at a frequency of one thousand times per second, and simultaneously records the absolute three-dimensional coordinates of the node to establish a unified reference three-dimensional coordinate system. Step 2: The receiver at each node receives the original echo signal returned by road surface scattering; multi-scale wavelet transform is performed on the original echo signal to generate a multi-scale power spectrum tensor; tensor sparse kernel decomposition method is used to separate the multi-scale power spectrum tensor into signal tensor and noise tensor, and a three-dimensional similarity spectrum is constructed; variable weight Lorentz manifold metric clustering is applied to the three-dimensional similarity spectrum to obtain feature clusters with maximum spectral coherence; the feature clusters are input into a deep time-series gated recurrent network to obtain the predicted displacement-settlement joint deviation for the next time period; the current displacement-settlement joint deviation and the predicted displacement-settlement joint deviation are aggregated according to the time series to form a multi-dimensional parameter matrix of construction quality, which represents the current construction quality status in real time with high dimension; Step 3: Perform principal component analysis on the multidimensional parameter matrix of construction quality to generate highway construction quality inspection results. Specifically, this includes: performing fractal normalization on the multidimensional parameter matrix of construction quality to generate a standardized matrix and retaining the inverse normalization index; constructing a sliding covariance grid, updating the covariance weight network based on temporal correlation, and inputting the resulting network into the principal component analysis core engine; the principal component analysis core engine extracting the principal axis vector set according to the progressive variance absorption rate criterion and performing orthogonal rotation to obtain stable principal axes; performing cascaded coordinate transformation on the standardized matrix under the stable principal axes to obtain a compressed feature scoring sequence; performing adaptive contribution aggregation on the compressed feature scoring sequence according to the time scale to output a comprehensive quality index column; and performing a hierarchical comparison between the comprehensive quality index column and the design tolerance library to indicate qualified and abnormal states, thus forming the highway construction quality inspection results.

2. The method for dynamic detection of highway construction quality based on high-precision laser ranging as described in claim 1, characterized in that, In step 1, the process of processing the beam multi-frequency modulation coding sequence includes: within a unified frame time slot, dividing the target transmission frequency band into several sub-frequency bands according to a preset multi-layer frequency domain stacking rule; inserting a dedicated start and end synchronization identifier into each sub-frequency band to keep each sub-frequency band aligned on the time-frequency plane, forming a basic beam multi-frequency modulation coding sequence frame; performing inter-layer orthogonal code diffusion on the basic beam multi-frequency modulation coding sequence frame, mapping each sub-frequency band to mutually orthogonal pseudo-random diffusion codes; achieving self-synchronization interference suppression between different sub-frequency bands through inter-code cyclic shifting to ensure no crosstalk during multi-frequency parallel transmission; performing time-frequency interleaving rearrangement on the beam multi-frequency modulation coding sequence frame after orthogonal code diffusion using a preset interleaving matrix; during the time-frequency interleaving rearrangement process, adjusting the time slots and adjacent sub-frequency bands... The frequency bands are alternately swapped to obtain an interleaved sequence. This interleaved sequence is then subjected to bipolar phase randomization: a pseudo-random phase flipping mode is introduced, and the phase randomization mask is embedded in the beam multi-frequency modulation coding sequence frame header. Multi-level self-correcting error correction embedded code segments, including frame-level block check codes and cross-frame-level cyclic check codes, are inserted into the beam multi-frequency modulation coding sequence frames with embedded phase randomization masks. The embedded code segments employ a combination of parity-flipping and redundancy segmentation. The beam multi-frequency modulation coding sequence frames with completed embedding are then subjected to spectral robust suppression and shaping: a hybrid shaping strategy using symmetrical Hanning windows and raised cosine windows is employed to bidirectionally reduce the sideband energy of each sub-frequency band, thereby reducing emission spectrum leakage. Simultaneously, an adaptive suppression and cohesion region is inserted at the center of the spectrum to suppress sidelobe noise caused by multipath echoes.

3. The method for dynamic detection of highway construction quality based on high-precision laser ranging as described in claim 2, characterized in that, The process of inserting an adaptive suppression cohesion region at the center of the spectrum includes: extracting the overall spectral envelope of the completed beam multi-frequency modulation coded sequence frame in real time to determine the position of the spectral center and the symmetric energy threshold surface; opening the start and end boundaries of the adaptive suppression cohesion region at the center of the spectrum based on the symmetric energy threshold surface, and setting progressive transition bands on both sides of the boundaries; performing amplitude reduction processing inside the cohesion region according to the adaptive suppression strategy, and performing smooth transition processing on the transition bands to make the spectral shape continuous and consistent; dynamically correcting the width and reduction depth of the cohesion region using echo monitoring results in subsequent transmission cycles of the beam multi-frequency modulation coded sequence frame to maintain continuous low suppression of sidelobe noise; and immediately triggering adaptive reconstruction of the cohesion region whenever a sudden change in multipath echo characteristics is detected to ensure that the transmission spectral shape and suppression effect are updated synchronously.

4. The method for dynamic detection of highway construction quality based on high-precision laser ranging as described in claim 3, characterized in that, The process of implementing inter-layer orthogonal code diffusion for the basic beam multi-frequency modulation and coding sequence frame includes: loading a unique and pairwise orthogonal pseudo-random diffusion code index field in each sub-band of the beam multi-frequency modulation and coding sequence frame; performing inter-code cyclic shift using the frame sequence number as a seed to fix the pseudo-random diffusion codes of adjacent sub-bands by clockwise displacement relative to the previous frame; setting an adaptive trigger gate at the transmitting end of the node to latch the bit sequence immediately after the inter-code cyclic shift is detected, thereby achieving self-synchronization of the sub-band boundaries; recovering the pseudo-random diffusion code in reverse according to the same cyclic shift sequence at the receiving end of the node to automatically cancel cross-sub-band crosstalk; and immediately triggering the next cyclic shift cycle if the inter-code alignment deviation exceeds the threshold to dynamically maintain the non-crosstalk of multi-frequency parallel transmission.

5. The method for dynamic detection of highway construction quality based on high-precision laser ranging as described in claim 4, characterized in that, Step 2, which uses tensor sparse kernel decomposition to separate the multi-scale power spectrum tensor into signal and noise tensors and construct a three-dimensional similarity spectrum, includes: performing multi-directional resampling on the multi-scale power spectrum tensor to establish an initial index table of tensor sparse kernels; iteratively eliminating weakly correlated kernel blocks in the initial index table according to the sparse density threshold, and aggregating the remaining kernel blocks to generate a signal tensor; simultaneously capturing and eliminating kernel blocks to form a noise tensor; expanding the signal tensor layer by layer along the frequency, scale, and time domains, calculating the three-dimensional similarity weights, and generating a primary correlation matrix; and applying piecewise progressive mapping and cross-stretching processing to the primary correlation matrix to obtain a three-dimensional similarity spectrum.

6. The method for dynamic detection of highway construction quality based on high-precision laser ranging as described in claim 5, characterized in that, Step 2, which involves applying variable-weight Lorentz manifold metric clustering to the 3D similarity spectrum to obtain the feature cluster with the maximum spectral coherence, includes: constructing an initial Lorentz manifold grid based on spectral density across the entire 3D similarity spectrum and assigning a variable weight vector to each grid node; performing a dual-threshold adaptive update on the variable weight vector based on the local spectral gradient of the nodes to form a multi-scale weight field; performing a layer-by-layer equidistant expansion search on the grid nodes under the variable-weight Lorentz manifold metric, and recording the manifold metric distance between nodes in real time; iteratively merging the node set according to the principle of minimum manifold path and maximum weight gain to generate a candidate feature cluster set; and performing incremental spectral coherence screening on the candidate feature cluster set to output the feature cluster with the maximum spectral coherence.

7. The method for dynamic detection of highway construction quality based on high-precision laser ranging as described in claim 6, characterized in that, The process of constructing an initial Lorentz manifold mesh based on spectral density across the entire 3D similarity spectrum and assigning a variable weight vector to each mesh node includes: decomposing the 3D similarity spectrum into continuous density layers according to the spectral density grading rules; performing a bidirectional Manhattan sweep on each density layer to generate equipotential surfaces and mapping them to curved mesh patches using the Lorentz coordinate interpolation method; weaving along the spectral gradient tangent to seamlessly stitch all the curved mesh patches together to form an initial Lorentz manifold mesh covering the entire domain; synchronously collecting neighborhood spectral density, local curvature, and temporal similarity at each mesh node and combining them into an original feature vector; converting the original feature vector into a variable weight vector through a normalized reweighting loop; and writing the variable weight vector into the mesh node metadata to complete the initialization of the mesh node weights.

8. The method for dynamic detection of highway construction quality based on high-precision laser ranging as described in claim 7, characterized in that, The process of generating equipotential surfaces by performing bidirectional Manhattan sweep for each density layer and mapping them to curved mesh patches using the Lorentz coordinate interpolation method includes: selecting the four boundaries of the density layer as the starting reference line and setting bidirectional Manhattan sweep paths in the horizontal and vertical directions; collecting the intersection points of density threshold surfaces row by row along the horizontal sweep path and marking candidate nodes of equipotential surfaces in real time; after completing the horizontal sweep, switching to the vertical sweep path and supplementing the missing nodes of equipotential surfaces column by column to form a bidirectional closed node network; performing neighborhood connectivity detection on the bidirectional closed node network to remove isolated nodes and obtain continuous and smooth equipotential surfaces; establishing local Lorentz tangent planes on the equipotential surfaces and converting equipotential surface segments into curved surface parameter domains according to the Lorentz coordinate interpolation method; splicing the curved surface segments in the order of the curved surface parameter domains to generate a fully conformal curved mesh patch; and writing the density layer identifier and spatial index to each curved mesh patch to complete the global registration of the curved mesh patch.

Images