Asymmetric Delay Compensation Method for Mine Wireless PTP Timing Based on Bidirectional Link Feature Fusion

By integrating bidirectional link characteristics into PTP timing technology and employing a dual-branch multilayer perceptron model and a Ridge linear basis model, the synchronization error problem caused by uplink and downlink asymmetry in underground coal mines was solved, achieving high-precision delay compensation and improved synchronization accuracy.

CN122908862BActive Publication Date: 2026-11-17XIAN UNIV OF SCI & TECH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611394660.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-09-09
Publication Date
2026-11-17
Estimated Expiration
2046-09-09

AI Technical Summary

Technical Problem

In the complex wireless propagation environment of underground coal mines, existing PTP timing technology suffers from increased synchronization errors due to uplink and downlink asymmetry. Existing methods struggle to explicitly distinguish and estimate propagation delay differences, leading to decreased synchronization accuracy.

Method used

A mine wireless PTP timing asymmetric delay compensation method based on bidirectional link feature fusion is adopted. By collecting PTP bidirectional timing data, bidirectional link quality parameters, multipath statistical parameters and mine structure scene tags are constructed. Asymmetric delay error estimation and compensation are performed using a two-branch multilayer perceptron model and a Ridge linear basis model.

Benefits of technology

It enables accurate estimation and compensation of errors caused by inconsistencies in uplink and downlink propagation delays in complex mining environments, improves synchronization accuracy and model generalization stability in different mining scenarios, and supports online deployment and outlier suppression.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122908862B_ABST
    Figure CN122908862B_ABST
Patent Text Reader

Abstract

This invention discloses a method for asymmetric delay compensation in mine wireless PTP timing based on bidirectional link feature fusion, belonging to the field of mine wireless communication and time synchronization technology. It aims to solve the problem of increased synchronization error caused by uplink and downlink asymmetry in the complex wireless propagation environment of mines using existing PTP timing technology. The method includes: collecting timestamps, collecting downlink quality parameters, uplink quality parameters, and multipath statistical parameters, generating contextualized tags, constructing a complete feature system, estimating the asymmetric delay error, and performing fusion compensation. This method can correct the synchronization deviation of the original PTP, compensating for the uplink and downlink asymmetry in mines at its root.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the field of mine wireless communication and time synchronization technology, specifically involving a method for asymmetric delay compensation of mine wireless PTP timing based on bidirectional link feature fusion. Background Technology

[0002] The underground environment of coal mine roadways has significant unique characteristics. The roadways are long and confined, with numerous turns and branches, significant equipment obstruction, and a large number of reflectors and scattering sources. This complex spatial structure leads to problems such as multipath effects, non-line-of-sight propagation, shadow fading, and local dynamic obstruction when wireless signals propagate underground. Under these propagation conditions, the downlink and uplink of the same pair of timing nodes often experience completely different propagation paths, with different signal-to-noise ratio characteristics, different multipath spreads, and different signal arrival times, resulting in inconsistencies in uplink and downlink propagation delays.

[0003] Precision Time Protocol (PTP), a widely used time synchronization protocol in computer networks, calculates time deviation through bidirectional message exchange between master and slave nodes. Traditional PTP-based bidirectional time synchronization methods typically assume that downlink and uplink propagation delays are approximately equal, assuming perfectly symmetrical propagation conditions in the bidirectional link. However, in the complex wireless propagation environment of underground coal mines, this symmetry assumption is often untenable. When there are differences in uplink and downlink propagation delays, these differences become asymmetric errors mixed into the original time synchronization deviation, leading to a decrease in synchronization accuracy. Existing technologies lack effective solutions to this problem, mainly in the following aspects: (1) The traditional PTP method calculates time deviation based on the symmetry assumption, which cannot explicitly distinguish and estimate the propagation delay difference between bidirectional links. Therefore, it is difficult to fundamentally eliminate the impact of asymmetric error on the synchronization result.

[0004] (2) Although methods such as mean filtering, sliding filtering or Kalman filtering can reduce some random fluctuations, these methods mainly target noise smoothing and short-term disturbance suppression, and cannot explain and compensate for systematic asymmetric errors caused by factors such as line-of-sight and non-line-of-sight states, roadway turning structure, branch direction, multipath delay spread and additional delay.

[0005] (3) Fixed statistical models usually rely on preset channel distribution patterns or empirical parameters. When the geometry, obstruction conditions, multipath intensity or link quality of the mine roadway change, the model parameters are prone to mismatch with the actual propagation environment, resulting in a significant decrease in compensation effect.

[0006] (4) In addition, if a single machine learning model directly splices together the features of the uplink and downlink and inputs them, it is easy to learn the statistical regularity of the overall average, and it cannot fully retain the differences between the uplink and downlink in terms of signal propagation direction, reflection path features, signal-to-noise ratio distribution and multipath spread characteristics, thereby weakening the model's ability to express the mechanism of asymmetric time delay error.

[0007] Based on this, the present invention proposes a method for asymmetric delay compensation of mine wireless PTP timing based on bidirectional link feature fusion, in order to solve the problems existing in the prior art. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method for asymmetric delay compensation of mine wireless PTP timing based on bidirectional link feature fusion, in order to solve the problem of increased synchronization error caused by uplink and downlink asymmetry in the complex wireless propagation environment of mines in existing PTP timing technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: A method for asymmetric delay compensation in mine wireless PTP timing based on bidirectional link feature fusion includes the following steps: Step 1: Collect PTP two-way time synchronization data and form basic timestamp samples; Step 2: Within the same synchronization period or adjacent aligned window as the data acquisition process in Step 1, record the downlink quality parameters, uplink quality parameters, and multipath statistical parameters respectively; Step 3: Generate contextualized tags for the mine structure; Step 4: Combine the timestamps, link quality, multipath statistics, and contextual labels obtained in Steps 1 to 3 according to a unified sample number to construct a complete feature system for bidirectional links; Step 5: Estimate the asymmetric time delay error of the feature system based on the DualBranchMLP model; Step 6: The asymmetric time delay error estimation is fused using the Ridge linear basis model and the DualBranchMLP model to obtain the asymmetric error estimate. Then, the original PTP time deviation is compensated based on the asymmetric error estimate, and the compensated synchronization deviation is output.

[0010] In a preferred embodiment, step 1, the process of acquiring PTP two-way time synchronization data, includes: Step 1.1: Collect PTP two-way time synchronization data and obtain timestamp data; Step 1.2: After obtaining the timestamp data, generate a derived value of the timestamp data; Among them, the derived quantities include the original PTP timing deviation estimate, the two-way round-trip time observation, the synchronization cycle number, the message sequence number, the transmit / receive direction identifier, and the abnormal message flag; Step 1.3: Finally, preprocess the collected timestamp data samples to obtain basic timestamp samples with direction, period, and quality markers.

[0011] In a preferred embodiment, in step 2, the downlink quality parameters include downlink received signal strength indication, downlink signal-to-noise ratio, downlink channel total gain, and downlink noise floor; the uplink quality parameters include uplink received signal strength indication, uplink signal-to-noise ratio, uplink channel total gain, and uplink noise floor; and the multipath statistical parameters include the number of multipaths, the dominant path energy percentage, root mean square delay spread, maximum excess delay, energy dispersion, power entropy, and delay kurtosis.

[0012] All of the above observations are marked with an upward or downward direction and a sample number.

[0013] In a preferred embodiment, step 3, the process of generating scene-based labels for the mine structure, includes: Based on the tunnel map, node locations, tunnel width and height, turning angle, branch angle, node height difference, equipment obstruction, support obstruction, vehicle obstruction, and manual inspection records, generate sight distance / non-sight distance signs, straight tunnel signs, turning tunnel signs, branch tunnel signs, obstruction type signs, and complex multipath level signs.

[0014] In a preferred embodiment, in step 4, the feature system includes directional channel features, uplink-downlink differential features, uplink-downlink ratio features, joint product interaction features, absolute difference features, and mine scenario enhancement features.

[0015] In a preferred embodiment, the directional channel features include uplink received signal strength, uplink signal-to-noise ratio, uplink total channel gain, downlink received signal strength, downlink signal-to-noise ratio, and downlink total channel gain, which are used to characterize the directional channel states of the uplink and downlink, respectively. The uplink and downlink differential characteristics include differential received signal strength, differential signal-to-noise ratio, differential channel total gain, differential root mean square delay spread, and differential maximum excess delay, which are used to directly characterize the asymmetric propagation characteristics of the uplink and downlink wireless links. The uplink / downlink ratio characteristics include signal strength ratio, signal-to-noise ratio ratio, channel gain ratio, delay spread ratio, and tail delay ratio, which are used to characterize the relative variation and correlation characteristics of the uplink / downlink bidirectional link channel features. The joint product interaction features include the joint product of delay spread, the joint product of tail delay, and the joint product of power entropy, which are used to characterize the nonlinear coupling combination effect between the basic channel features. The absolute difference features include channel state absolute deviation, power entropy absolute deviation, and delay spread absolute deviation, which are used to characterize the degree of absolute difference between the corresponding feature parameters of uplink and downlink. The enhanced features of the mine scene include main roadway length, sending and receiving distance, turning angle, branching angle, line-of-sight and non-line-of-sight signs, and scene complexity score.

[0016] In a preferred embodiment, in step 4, after the feature system is constructed, the feature set of the complete feature system is split into three types of neural network inputs: uplink branch input, downlink branch input, and differential auxiliary branch input, as well as a table model input containing the complete feature set; The uplink branch input includes the received signal strength indication, signal-to-noise ratio, channel gain, and multipath statistics in the uplink direction; the downlink branch input includes the corresponding features in the downlink direction; and the differential auxiliary branch input includes uplink and downlink differential features, proportional features, joint product features, and absolute difference features.

[0017] In a preferred embodiment, in step 5, the DualBranchMLP model uses the up-row branch and the down-row branch as two main branches, and includes differential auxiliary branches and a fusion layer. The relationship between the layers is as follows: h down = f down (X down ) h up = f up (X up ) h diff = f diff (X diff ) = g([h down , h up , h diff ]) Among them, f up For the upward branch, f down For the downward branch, f diff For differential auxiliary branches, g is the fusion layer. X represents the error estimate output by the DualBranchMLP model. up X is the uplink feature vector. down X is the downlink feature vector. diff h is the difference auxiliary feature vector. up h is the latent vector of the upward branch. down h is the latent vector of the downlink branch. diff The latent vector for the difference auxiliary branch.

[0018] In a preferred embodiment, the uplink branch, the downlink branch, and the differential auxiliary branch all include two fully connected layers with a hidden dimension of 128. Each fully connected layer is followed by a batch normalization layer, a ReLU activation function, and a Dropout layer in sequence, with a Dropout ratio of 0.02. The three branches output 128-dimensional implicit vectors: the uplink implicit representation, the downlink implicit representation, and the difference-aided implicit representation, respectively. The input to the fusion layer is the concatenation of three latent vectors. It first passes through a 384-dimensional to 128-dimensional fully connected layer, a ReLU activation function, a batch normalization layer, and a Dropout layer. Then it passes through a 128-dimensional to 64-dimensional fully connected layer and a ReLU activation function. Finally, it passes through a 64-dimensional to 1-dimensional linear output layer to obtain the asymmetric error estimate.

[0019] In a preferred embodiment, in step 6, when performing fusion compensation, the fusion formula is: and in, This represents the asymmetric time delay error estimate output by the fusion model. This represents the error estimate output by the Ridge linear basis model. The coefficients are the error estimates output by the DualBranchMLP model. and These are the corresponding fusion weights.

[0020] Compared with existing technologies, this invention provides an asymmetric delay compensation method for mine wireless PTP timing based on bidirectional link feature fusion, which has the following beneficial effects: This invention breaks the assumption of uplink and downlink symmetry in traditional PTP timing by explicitly estimating the PTP residual error caused by the inconsistency between uplink and downlink propagation delays. It avoids simply incorporating the asymmetry of the mine wireless link into the original PTP deviation, thus providing a compensation scheme from the root cause of the error.

[0021] This invention utilizes RSSI, SNR, and G up / G down τ rms τ max,exc The system analyzes and explains the sources of error in mine multipath statistics and scene characteristics, including line-of-sight and non-line-of-sight states, turning angles, branch angles, and occlusion types. This enables interpretable modeling of the error mechanism and gives the compensation process a clear physical meaning.

[0022] This invention constructs a complete feature system for bidirectional links, which can accurately distinguish between uplink and downlink and quantify the differences between bidirectional links, effectively ensuring the accuracy of asymmetric error prediction results.

[0023] The uplink and downlink dual-branch neural network designed in this invention learns the bidirectional propagation law through independent branches, explicitly learns the asymmetric relationship through differential auxiliary branches, and performs error estimation by integrating the three-way information through a fusion layer. Compared with single-branch networks and simple feature splicing methods, it significantly improves the error estimation accuracy.

[0024] This invention improves the generalization stability and resistance to extreme errors of the model under different mining scenarios by fusing and compensating for the Ridge linear model and the DualBranchMLP model, taking into account both linear interpretable error terms and nonlinear error terms in complex scenarios.

[0025] This invention supports online deployment, outlier suppression, weight updates or model recalibration under changing scenarios, and has good engineering applicability, making it easy to promote and apply in actual mining environments. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the asymmetric delay error modeling and compensation process for wireless time synchronization in mines, as described in this invention.

[0027] Figure 2 This is a representative RMSE comparison chart of the Raw PTP and fusion compensation methods of this invention.

[0028] Figure 3 This is a comparison chart of RMSE values ​​for the various models in this invention on the test set.

[0029] Figure 4 This is a comparison chart of the MAE test sets for each model in this invention.

[0030] Figure 5 This is a distribution diagram of the prediction error of the Ridge and DualBranchMLP fusion model of this invention.

[0031] Figure 6 This is a comparison chart of the predicted values ​​and reference values ​​of the Ridge and DualBranchMLP fusion model of this invention.

[0032] Figure 7 This is a distribution diagram of the prediction error of the DualBranchMLP model for mine scenarios in this invention.

[0033] Figure 8 This is a comparison chart of the predicted values ​​and reference values ​​of the DualBranchMLP model for mine scenarios in this invention.

[0034] Figure 9 This is a comparison chart of the RMSE of ablation experiments for the features and model structure of this invention.

[0035] Figure 10 This is a ranking diagram of the Top 20 key features of this invention.

[0036] Figure 11 This is a comparison chart of LOS / NLOS grouped RMSE of the present invention.

[0037] Figure 12 This is a comparison chart of RMSE for different mining scenarios in this invention.

[0038] Figure 13 This is a comparison chart of the group RMSE for different delay extension intervals of the present invention.

[0039] Figure 14 This is a comparison chart of the mean and standard deviation of RMSE in 30 independent paired repeated experiments of the method of this invention. Detailed Implementation

[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] As per the instruction manual Figure 1 - Appendix Figure 14 As shown, to address the problem of increased synchronization errors caused by uplink and downlink asymmetry in the complex wireless propagation environment of mines using existing PTP timing technology, this invention proposes an asymmetric delay compensation method for mine wireless PTP timing based on bidirectional link feature fusion. This method adds an error estimation and compensation layer outside the PTP bidirectional message exchange process. Without changing the basic transmission and reception logic of PTP messages, it synchronously collects or constructs uplink and downlink quality, multipath statistics, and mine scene characteristics to estimate the asymmetric error caused by the inconsistency in uplink and downlink propagation delays. This estimate is then used to correct the original time deviation, thereby achieving asymmetric delay compensation for mine wireless PTP timing. Specific steps include: Step 1: Collect PTP two-way time synchronization data and use the PTP two-way time synchronization data acquisition module to form basic timestamp samples.

[0042] Step 1.1: In the PTP bidirectional timing data acquisition module, both the master node and the slave node are configured with an IEEE 1588v2-compatible PTP protocol stack, and forward messages through a wireless access point (AP). The master node sends a synchronization message (Sync) under local clock control and records the sending time t1. This synchronization message is propagated to the slave node through the mine wireless channel. After receiving the message, the slave node records the reception time t2. The slave node then sends a delay request (Delay_Req) ​​and records the sending time t3. This return message propagates along the uplink to the master node. After receiving the return message, the master node records the reception time t4.

[0043] The four PTP timestamps t1, t2, t3, and t4 constitute the core time base for PTP two-way time synchronization and can be used to calculate the original PTP time synchronization deviation. Here, t1 is the time when the synchronization message is sent, t2 is the time when the synchronization message is received, t3 is the time when the delay request message is sent, and t4 is the time when the delay request message is received.

[0044] Step 1.2: After acquiring the timestamp data, the PTP two-way time synchronization data acquisition module is used to further generate the original PTP time synchronization deviation estimate within the same synchronization period. Two-way round-trip time observation Derived variables such as synchronization cycle number, message sequence number, transmit / receive direction identifier, and abnormal message flag; Among them, the original PTP timing deviation estimate The calculation formula is: = ((t2- t1) - (t4- t3)) / 2 Two-way round trip time observation The calculation formula is: = (t4- t1) - (t3- t2) Step 1.3: To ensure the data quality used in subsequent modeling, the PTP two-way time synchronization data acquisition module is used to remove samples from the acquired data that show obvious packet loss, reversed timestamps, or exceed the physical propagation range. Specifically, this includes: Step 1.3.1: Perform an integrity check on the timestamp sequence. When a message missing for more than a set threshold is detected within a consecutive period, mark the data of that period as abnormal. Step 1.3.2: When a reversed timestamp phenomenon such as t3 being less than t2 or t4 being less than t1 is detected within the same clock domain, the sample is removed; Step 1.3.3: When the calculated round-trip time exceeds the pre-calibrated reasonable time range, the sample is removed.

[0045] After the above processing, the output is a basic timestamp sample with direction, period and quality markings, which provides a unified time reference for subsequent asymmetric error modeling.

[0046] It should be noted that in actual mine deployments, PTP timestamp acquisition needs to consider the unique characteristics of mine wireless channels. While the propagation speed of wireless signals underground in coal mines is close to the speed of light, for tunnels hundreds of meters long, the propagation delay is only on the order of hundreds of nanoseconds. When multipath effects and non-line-of-sight propagation exist, the actual arrival time of the signal will experience additional delays due to the superposition of multipath paths. This additional delay may be inconsistent between uplink and downlink, leading to asymmetric errors. The PTP two-way timing data acquisition module uses the synchronization period number as the primary key to associate all observation data within the same period, forming time-aligned data samples.

[0047] Step 2: Perform the data acquisition in the same synchronization period or adjacent aligned window as in Step 1, and use the wireless link observation module to record the downlink quality parameters, uplink quality parameters, and multipath statistics parameters respectively.

[0048] The downlink quality parameters include downlink received signal strength indication. Downlink signal-to-noise ratio Total downlink channel gain and downlink noise level Characteristic parameters, etc.

[0049] The uplink quality parameters include the uplink received signal strength indicator. Uplink signal-to-noise ratio Total uplink channel gain and uplink noise level Characteristic parameters, etc.

[0050] The multipath statistical parameters include the number of multipaths N. paths η, the dominant path energy ratio dom Root mean square delay spread τ rms Maximum excess delay τ max,exc Energy dispersion E disp Power entropy H pow Timely extension of kurtosis K τ Characteristic parameters, etc.

[0051] Wherein, the root mean square delay spread τ rms Used to describe the dispersion of signal energy along the time axis; the maximum excess delay τ max,exc Used to describe the maximum time delay difference between the latest significant multipath component and the first-arrival path; the dominant path energy percentage η dom Used to describe the proportion of dominant path energy to total energy; the energy dispersion E dispUsed to describe the statistical distribution characteristics of multipath energy; the power entropy H pow Used to describe the uniformity of multipath power distribution; the time delay kurtosis K τ These parameters are used to describe the sharpness of multipath delay distribution. Together, they characterize the temporal dispersion properties of the mine wireless channel and are an important basis for estimating asymmetric delay errors.

[0052] All the above observations are labeled with either uplink or downlink direction and sample number to avoid confusing uplink and downlink data. The downlink propagation delay is denoted as d. down The uplink propagation delay is denoted as d. up The symmetry assumption in traditional PTP calculations can be expressed as d down Approximately equal to d up When this assumption is not true, asymmetric errors will be introduced into the original time offset. Specifically, the theoretical asymmetric error term introduced by the difference in uplink and downlink propagation delays is: = (d down -d up ) / 2 in, This represents the asymmetric time delay error estimate output by the fusion model, which can be obtained through simulation or experimental calibration during the model training phase. As a supervisory label; This indicates the time synchronization correction result after compensation; This represents the original PTP timing deviation estimate.

[0053] During the online deployment phase, the system uses the timestamp of the current synchronization period, link quality, multipath parameters, and the asymmetric delay error estimate output by the fusion model. and according to Calculate the compensated time synchronization correction result Then As the corrected synchronization deviation output.

[0054] It should be noted that wireless signal propagation exhibits significant directional differences in the underground coal mine environment. Due to the influence of tunnel geometry, in straight tunnels, both uplink and downlink may maintain line-of-sight propagation, resulting in good and symmetrical channel quality. However, in curved tunnels, signals from the master node to the slave node may be reflected by the curve, while signals from the slave node to the master node may propagate via different paths, leading to significant differences in path loss and multipath structure between the two links. In branch tunnels, signals may diffract and scatter at the branch points, causing differences in the angle of arrival, delay spread, and signal strength of the uplink and downlink. In areas with dense equipment, obstruction from electromechanical equipment and electromagnetic interference further exacerbate the asymmetry of the uplink and downlink.

[0055] Step 3: Using the mine scene recognition module, generate scene-based labels for the mine structure based on the roadway map, node locations, roadway width and height, turning angles, branch angles, node height differences, equipment obstruction, support obstruction, vehicle obstruction, and quantifiable obstruction event flags (such as a Boolean value for "temporary obstruction exists" or a 1-5 level number representing the severity of obstruction) confirmed and entered manually through the terminal.

[0056] Among them, such as Figure 1 As shown, the generated scenario-based label types include line-of-sight / non-line-of-sight (LOS / NLOS) signs, straight roadway signs, turning roadway signs, branch roadway signs, obstruction type signs (including one or more combinations of equipment obstruction, support obstruction, vehicle obstruction, and manual inspection obstruction), and complex multipath level signs, etc.

[0057] The specific judgment criteria are as follows: Step 3.1: LOS / NLOS flag determination: Determine whether there is a directly visible propagation path based on the tunnel map and the location of the transmitting and receiving nodes, and determine the LOS / NLOS flag by combining RSSI, SNR and root mean square delay spread; Step 3.2: Determination of straight roadway markings: When the turning angle is less than 15° and there are no branches, it is marked as a straight roadway; Step 3.3: Determination of turning roadway markings: When the turning angle is greater than or equal to 15°, it is marked as a turning roadway. The turning angle is the angle between the main axis of the roadway and the connecting line. Step 3.4: Branch roadway marking determination: When a branch angle greater than 30° is detected, it is marked as a branch roadway; Step 3.5: Obstruction type identification: Based on the type of obstruction, it is divided into equipment obstruction, support obstruction, vehicle obstruction, and no obstruction; Step 3.6: Complex Multipath Classification: Based on the root mean square delay extension τ rms Classified into low level (τ) rms <50ns), medium level (50ns≤τ) rms <150ns) and high-level (τ) rms ≥150ns).

[0058] It should be noted that for scenario items that cannot be directly obtained, this embodiment employs an indirect inference strategy. Specifically: A sudden change in link quality may indicate a change in line-of-sight status. An increase in multipath tail energy may indicate non-line-of-sight propagation; When the root mean square time delay spreads τ rms A significant increase may indicate a complex multipath environment; When a node is located near a bend or branch point in a lane, the scene type can be inferred through location segment mapping.

[0059] The mine scene recognition module further constructs a tunnel scene complexity score S. mine The model is used to quantitatively assess the combined complexity of the geometry, occlusion status, and multipath propagation of underground roadways. The S... mine The calculation formula is: S mine =w1·ΔW+w2·θ corner +w3·θ branch +w4·D device + w5·τ rms,norm ; Where: ΔW is the variation range of the tunnel width, θ corner Let θ be the turning angle. branch D is the branch angle. device For equipment layout density, τ rms,norm For the normalized root mean square time delay spread, w1 to w5 are weighting coefficients, which are obtained by calibration using actual mine environment data.

[0060] In this embodiment, w1, w2, w3, w4 and w5 can be 0.15, 0.20, 0.20, 0.15 and 0.30 respectively; in other embodiments, each weight coefficient can be adjusted according to the calibration data.

[0061] The scoring model integrates characteristic parameters such as the variation range of roadway width, roadway turning angle, branch angle, equipment deployment density, and normalized root mean square time delay spread to quantitatively characterize the complexity of underground propagation scenarios.

[0062] After processing in step 3 above, the output result is a mine scene-based label table corresponding to each synchronization cycle. This label table contains information such as the line-of-sight or non-line-of-sight status identifier, roadway type identifier, and scene complexity score corresponding to each sample, providing scene-related prior knowledge for subsequent feature construction.

[0063] Step 4: Using the feature construction module, merge the timestamps, link quality, multipath statistics, and contextual labels obtained in Steps 1 to 3 according to a unified sample number to construct a complete bidirectional link feature system F. bi .

[0064] The above feature system F bi The construction process is not simply a matter of stacking RSSI or SNR, but rather a unified modeling of PTP temporal relationships, uplink and downlink directional link observations, multipath propagation, contextualized labels, and interaction terms. This allows the model to simultaneously see both the "current error magnitude" and the "propagation mechanism leading to the error." Specifically, this includes: Step 4.1: First, perform data alignment and save tags; Align the PTP timestamps, uplink and downlink observations, multipath statistics, and mine scenario labels within the same synchronization period according to sample numbers, and use the PTP residual error values ​​obtained from simulation calculations or calibration tests. Stored independently as a supervisory label.

[0065] In the simulation environment, the PTP residual error value corresponds to the residual error between the original PTP estimate and the actual clock deviation; in the actual test environment, a high-precision reference clock or fiber optic loopback link is used to establish the reference time deviation between the master and slave nodes, and the reference time deviation is used as the reference for the actual clock deviation; the round-trip flight time measurement is obtained through bidirectional probe messages or a known delay reference link, and the sending time and receiving time of the probe message are associated with the corresponding link direction and recorded.

[0066] Step 4.2: Then perform feature standardization; For continuous input features, the StandardScaler standardization method is used to reduce the impact of differences in the dimensions and numerical scales of different features. This transforms each continuous feature into a standardized form with a mean close to 0 and a standard deviation close to 1 according to the training set statistics, thereby achieving feature dimensionality balance. This standardization process can simultaneously adapt to the model input requirements of Ridge linear basis models and multilayer perceptron neural networks.

[0067] The normalizer calculates the mean and standard deviation of each continuous feature using only the training set and saves the corresponding parameters; the validation set, test set, and online real-time data are all transformed using the same set of parameters saved during the training phase to avoid data leakage.

[0068] Step 4.3: Next, generate directional channel features, uplink / downlink differential features, uplink / downlink ratio features, joint product interaction features, absolute difference features, and mine scenario enhancement features, respectively. Specifically, the above features include: Directional channel characteristics are selected from the aforementioned uplink and downlink quality parameters, including: uplink received signal strength indication. Uplink signal-to-noise ratio Total uplink channel gain and downlink received signal strength indication Downlink signal-to-noise ratio and total downlink channel gain , used to characterize the directional channel states of the uplink and downlink respectively.

[0069] Uplink and downlink differential characteristics include: differential received signal strength Differential signal-to-noise ratio Total gain of differential channel Differential root mean square delay spread Differential maximum excess delay These parameters can directly characterize the asymmetric propagation characteristics of uplink and downlink wireless links.

[0070] Uplink / downlink ratio characteristics include: signal strength ratio r RSSI Signal-to-noise ratio r ξ Channel gain ratio r gain Latency spread ratio r spread Tail delay ratio r tail These parameters can effectively characterize the relative variation patterns and correlation characteristics of uplink and downlink bidirectional link channel features.

[0071] Joint product interaction features include: delay-spread joint product P js The joint product of tail delay P jt The joint product of power and entropy P je Interactive features are used to characterize the nonlinear coupling and combination effects between various basic channel features.

[0072] Absolute difference characteristics include: absolute channel state deviation |Δ regime | Absolute deviation of power entropy |ΔH | Absolute deviation of time delay spread |Δτ rms The parameters are used to characterize the absolute difference in the corresponding feature parameters of the uplink and downlink.

[0073] Enhanced features for mine scenes include: main roadway length L tunnel Transmit / receive distance L tx-rx Turning angle θ corner Branch angle θ branch Visual distance and non-visual distance markers δ NLOS Scene complexity score S scene Time-varying complexity S time Deployment asymmetry score S asym Material support complexity S mate and dynamic environment complexity S dyn wait.

[0074] Wherein, the true time offset τ offset PTP residual error ε ptp Downlink delay deviation τ d,down Uplink delay deviation τ d,up It is only used in the simulation and calibration stages, serving as a supervision label, error analysis indicator, and performance evaluation indicator. It does not participate in the feature input and prediction calculation process during the model inference stage. In the simulation embodiment, the link asymmetry parameter S... link,asymThis value is used solely for generating simulation scenarios and conducting result grouping analysis and performance evaluation; it is not used as input features during model training, validation, testing, or online inference. In actual deployments, this value is not directly used as a manually predicted input, but can be obtained through uplink and downlink received signal strength RSSI, signal-to-noise ratio SNR, and root mean square delay spread τ. rms The differences in packet loss rate and retransmission rate are combined with scene complexity scoring to achieve an approximate representation.

[0075] Step 4.4: After the feature system is constructed, the feature set of the complete feature system is split into three types of neural network inputs: uplink branch input, downlink branch input, and differential auxiliary branch input, and one type of tabular model input. Specifically: The complete feature set is split into an uplink branch input X. up Downward branch input X down Differential auxiliary branch input X diff and table model input X tab ; Among them, the uplink branch input X up Features include received signal strength index (RSSI), signal-to-noise ratio (SNR), channel gain, and multipath statistics in the uplink direction; downlink branch input X down Including corresponding features in the downlink direction; differential auxiliary branch input X diff Includes uplink / downlink difference features, proportional features, joint product features, and absolute difference features; tabular model input includes complete F... bi Feature set, used by tabular models such as Ridge regression.

[0076] Step 5: Use the asymmetric error estimation module to perform asymmetric time delay error estimation on the feature system based on the DualBranchMLP model.

[0077] like Figure 1 As shown, the design of the DualBranchMLP model is based on the fact that wireless timing errors in mines mainly stem from inconsistencies in uplink and downlink propagation conditions. If uplink and downlink features are simply concatenated and input into a single network, the model may treat bidirectional links as having identical statistical laws, failing to fully preserve the differences between uplink and downlink in terms of signal propagation direction, reflection path characteristics, signal-to-noise ratio distribution, and multipath spread characteristics. The dual-branch structure inputs uplink and downlink features into two separate branches, first independently learning directional propagation laws, and then learning the asymmetric relationship between the two in the fusion layer.

[0078] The DualBranchMLP model includes an uplink main branch. Downward main branch Differential auxiliary branch and fusion layer The upward main branch and the downward main branch constitute two directional main branches. The relationship between each layer is as follows: h down = f down (X down ) h up = f up (X up ) h diff = f diff (X diff ) = g([h down , h up , h diff ]) The aforementioned downward branch f down and the upward branch f up Using a StableMLPBlock structure with the same number of layers but no shared parameters, differential auxiliary branches f diff The same StableMLPBlock is used. The reason why the two directional branches do not share parameters is that the main-to-slave and slave-to-main directions in the mine may exhibit different statistical patterns due to different antenna orientations, node heights, roadway turns, obstruction positions, and multipath reflection paths. Independent training of branches can learn patterns of large downlink propagation delay, large uplink propagation delay, or alternating patterns of both.

[0079] In one specific implementation, the uplink branch f up Downward branch f down Sum of auxiliary branches f diff Each layer consists of two fully connected layers, with a hidden dimension D. hid,dual The value is 128; each fully connected layer is followed by a batch normalization layer, a ReLU activation function, and a Dropout layer in sequence, with λ... drop,dual Set the value to 0.02. The three branches each output a 128-dimensional latent vector h. up h down and h diff The fusion layer takes the concatenation of three latent vectors as input, first passing through a 384-dimensional to 128-dimensional fully connected layer, ReLU, batch normalization, and Dropout, then through a 128-dimensional to 64-dimensional fully connected layer and ReLU, and finally through a 64-dimensional to 1-dimensional linear output layer to obtain the asymmetric error estimate. .

[0080] The fusion layer g is used to integrate the uplink branch f up Downward branch f down Sum of auxiliary branches f diffThe vectors are concatenated to form a fused vector Z. The difference auxiliary branch f... diff Effective difference information can be automatically extracted from difference-class features (diff), ratio-class features (ratio), joint interaction-class features (joint), and absolute deviation-class features (|Δ|), that is, the differential representation of uplink and downlink is explicitly introduced into the network structure. On this basis, the fused feature vector Z is used to complete feature mapping through a fully connected layer, and finally the prediction result is output. .

[0081] During the model training phase, the residual error ε of PTP is used. PTP The label serves as the supervisory label for supervised learning training, corresponding to the residual error between the original PTP estimate and the actual clock skew in the simulation. The loss function is SmoothL1Loss, the optimizer is AdamW, and the learning rate is η. lr,dual for The weight decay coefficient is λ wd for The ReduceLROnPlateau learning rate adjustment mechanism and early stopping mechanism were adopted, and the optimal model was selected based on the root mean square error (RMSE) of the validation set. The maximum number of training rounds in the conventional mine scenario experiment was set to 180 rounds, and the maximum number of training rounds in the formal 30 independent paired repeated experiments was set to 80 rounds, with a batch size of 128 for both. Training was stopped when the RMSE of the validation set did not improve for 24 consecutive rounds, and the model parameters when the RMSE of the validation set was optimal were retained.

[0082] During the deployment phase, the uplink feature vector constructed in the current cycle is... Downlink feature vector Sum of auxiliary feature vectors Enter the up branch respectively Downward branch Sum of auxiliary branches The outputs of the three branches are processed by the fusion layer to obtain the asymmetric error estimate for the current cycle. .

[0083] In the aforementioned DualBranchMLP model, the downlink branch inputs include downlink RSSI, downlink SNR, downlink channel gain, downlink multipath statistics, downlink estimated propagation delay, and downlink scenario-related features, used to characterize the wireless propagation state from the master node to the slave node. The uplink branch inputs include uplink RSSI, uplink SNR, uplink channel gain, uplink multipath statistics, uplink estimated propagation delay, and uplink scenario-related features, used to characterize the wireless propagation state from the slave node to the master node. After the two directional branches extract independent features for the uplink and downlink respectively, they are then combined with the asymmetric features extracted by the differential auxiliary branch and fed into the fusion layer.

[0084] The differential auxiliary inputs include uplink / downlink differential features, proportional features, LOS / NLOS status, turn / branch labels, complex scenario scoring, and interaction features, used to further characterize the asymmetric relationship between bidirectional links. The model ultimately outputs an estimate of the asymmetric delay error within the current synchronization cycle. This estimate can be directly used as a PTP bias correction term to compensate for systematic delay errors in the original PTP synchronization results.

[0085] Step 6: In the fusion compensation module, the asymmetric time delay error estimation is fused using the Ridge linear basis model and the DualBranchMLP model to obtain the asymmetric error estimate. Then, the original PTP time deviation is compensated based on the asymmetric error estimate, and the compensated synchronization deviation is output.

[0086] The Ridge linear basis model is used to learn relatively stable and interpretable linear error terms such as uplink / downlink delay differences, channel quality differences, and multipath spread differences; the DualBranchMLP model is used to learn nonlinear error terms caused by turns, branches, NLOS, and complex multipath effects. Fusing the two models reduces the risk of overfitting or underfitting in a single model and ensures that the final compensation value possesses both stability and nonlinear expressive power. The fusion formula is: and ; in, This represents the asymmetric time delay error estimate output by the fusion model. This represents the error estimate output by the Ridge linear basis model; The error estimate is the output of the DualBranchMLP model; coefficients and These are the corresponding fusion weights, determined by minimizing the RMSE on the validation set. The specific steps are as follows: Step 6.1: First, train the Ridge linear basis model and the DualBranchMLP model on the training set respectively; Step 6.2: Calculate the prediction sequences of the two models on the validation set. and ; Step 6.3: Then command Perform a grid search with a step size of 0.01 between 0 and 1, and calculate: And select α that minimizes the RMSE of the validation set. ridge As a Ridge weight, and at the same time As weights in DualBranchMLP.

[0087] Let α ridge Perform a grid search between 0 and 1 with a step size of 0.01, and let... = The weight combination that minimizes the RMSE of the validation set is selected as the fusion weight for the current experiment; the weight search process is repeated in each independent repeated experiment.

[0088] To avoid overcorrection introduced by fusion compensation, in an optional engineering implementation, a limiting and rollback mechanism can be set, wherein the rollback mechanism control unit executes the following logic: Continuously monitor the degree of divergence D=| - |, when D exceeds a preset fixed threshold Th for K1 consecutive periods div At that time, a rollback is triggered. After triggering, the weights are merged. The value decreases linearly from the optimum to 0 over M periods, during which only the output of the Ridge linear basis model is used as the final estimate; if D is still greater than Th during the rollback process... div Then it further reverts to the Raw PTP strategy. When D is below another smaller safety threshold Th for K2 consecutive periods... safe At that time, weight It linearly recovers to the optimal value of the validation set within M periods. The Th div ,Th safe K1, K2, and M are all pre-set configuration parameters based on the mine site environment. This adaptive mechanism can effectively prevent the model from generating excessive compensation jumps under abnormal conditions.

[0089] Asymmetric error estimate obtained based on fusion compensation Calculate the time deviation after compensation: equal minus and will As the corrected synchronization deviation output.

[0090] Step 7: Training, Validation, and Deployability Process; Step 7.1: Training and Validation Process; 1. Generate or collect a dataset containing PTP timestamps, uplink and downlink status, multipath parameters, and mine-specific labels.

[0091] 2. The theoretical asymmetric time delay error e obtained by simulation calculation or calibration. asym Or, equivalently defined, the PTP residual error ε PTPAs a supervisory label, the uncompensated Raw PTP results are retained as the baseline. In the current simulation embodiment, ε res,PTP The residual error between the original PTP estimate and the actual clock deviation is represented, which can be equivalently characterized by the compensation target introduced by asymmetric propagation.

[0092] 3. Divide the data into training, validation, and test sets; train Linear, Ridge, SVR, RandomForest, HistGBR, SingleBranchMLP, DualBranchMLP, and fusion models.

[0093] 4. Calculate MAE, RMSE, and R. 2 It also outputs error distribution plots, prediction scatter plots, model comparison plots, and scene grouping error plots.

[0094] 5. Perform F bi Feature ablation, comparison of basic features / multipath features / scene features, Top20 / Top50 key feature experiments, and multiple random seed repeated experiments.

[0095] Step 7.2: Online deployment process; 1. The nodes exchange synchronization messages according to the PTP process and record four timestamps.

[0096] 2. The link observation module records uplink and downlink received signal strength indication (RSSI), signal-to-noise ratio (SNR), channel gain, and available multipath statistics within the same synchronization period.

[0097] 3. The scene recognition unit generates LOS / NLOS, turn, branch and occlusion labels based on the alley location, link quality or manual configuration.

[0098] 4. Feature building blocks generate F bi The features are then broken down into uplink input, downlink input, and differential auxiliary input.

[0099] 5. Model outputs asymmetric time delay error estimate The compensation module calculates the corrected synchronization deviation. It can also perform moving average, outlier suppression, or scene-adaptive weight adjustment on continuous periodic outputs according to deployment needs.

[0100] Step 8: Experimental Results and Technical Effects; To verify the effectiveness, stability, and potential engineering applicability of the method of the present invention under the simulated conditions of a mine scenario, in a feasible application scenario, the simulation process improves the fidelity in the following ways: during the simulation channel modeling stage, pulse noise disturbances that may be generated by the start-up and shutdown of electromechanical equipment are introduced, and time drift and jitter effects based on typical crystal oscillator parameters are simulated in the PTP timestamp generation stage.

[0101] The experiments covered general stochastic multipath simulation, mine-scenario parameter adjustment, basic model comparison, dual-branch structure verification, fusion of Ridge linear basis model and dual-branch multilayer perceptron (DualBranchMLP) model, complex scene correction, feature ablation, TopK key feature selection, and repeated experiments with multiple random seeds. The experimental results are shown in Table 1 and the figures are shown below. Figures 2-14 As shown, the model prediction error on the general stochastic multipath simulation dataset cannot be directly compared numerically with the Raw PTP baseline of the mine scenario simulation dataset.

[0102] Table 1: Experimental Results This embodiment introduces real mine UWB CIR records to constrain and calibrate multipath statistical parameters under basic straight-tunnel propagation conditions; PTP residual error, uplink and downlink propagation asymmetry, and complex mine scenarios are generated through mine scenario simulation. Therefore, this embodiment belongs to a mine scenario simulation experiment under real mine CIR statistical constraints.

[0103] The results shown in Table 1 illustrate that, in the multi-random seed repeated experiments, the results of 30 independent paired repeated experiments were used as the statistical basis. The RMSE of the uncompensated baseline Raw PTP was 60.7187 ± 0.9426 ns, and the RMSE of the fusion compensation model of the Ridge linear basis model and the DualBranchMLP model was 18.0690 ± 0.7221 ns; the fusion compensation model reduced the RMSE by an average of 70.24% compared to the uncompensated baseline Raw PTP. These results indicate that, under the conditions of mine-based simulation data and consistent evaluation set in this embodiment, the fusion compensation model has good repeatability.

[0104] like Figures 2 to 4 As shown in the performance comparison chart of the various models of this invention on different test sets, the fusion model Fusion_Ridge_DualBranchMLP proposed in this invention has achieved the best performance on different test sets, which is significantly better than traditional machine learning models Linear, Ridge, HistGBR and single-branch MLP models.

[0105] like Figure 5 and Figure 6The error distribution plot and prediction scatter plot of the fusion model shown can be seen that the prediction error is mainly concentrated around 0 ns, the overall error distribution is approximately symmetrical, most errors are in the range of -100 ns to 100 ns, and the long-tail phenomenon is relatively weak; the scatter points are closely distributed around the diagonal of y equal to x, covering a range from -1000 ns to 1000 ns, indicating that the model can still maintain high linearity prediction ability under a large range of errors.

[0106] like Figure 7 and Figure 8 The error distribution plot and prediction scatter plot of the DualBranchMLP model in the mine scenario shown can be seen that the error distribution is more concentrated, mainly between -50 and 50 ns, with extremely high peak values. The data points range from -200 to 200 ns, closely following the diagonal, indicating that the model's predicted values ​​have good consistency with the reference values.

[0107] like Figure 9 The summary diagram of ablation experiments for features and model structures shown illustrates the progression from using only basic link features to using only scene geometry features, then to multipath statistics plus downlink differences, multipath statistics plus scene features, and finally to the complete F-axis model. bi As features are progressively added, model performance continuously improves. The dual-branch MLP structure has significant advantages over the single-branch MLP structure. The Ridge linear basis model and the dual-branch fusion model (a fusion and compensation model of the Ridge linear basis model and the dual-branch multilayer perceptron DualBranchMLP model) achieved the best performance.

[0108] like Figure 10 As shown in the Top 20 key feature ranking chart, differential features rank relatively high in the ranking, including power entropy-related differential features Δ. H spread Channel state correlation differential features Δ R spread Root mean square delay spread correlation difference feature Δ τ rms,spread etc.; among which, Δ H spread Δ R spread and Δ τ rms,spread These are symbolic representations of actual characteristic fields such as diffusion entropy spread (diff_entropy_spread), diffusion state spread (diff_regime_spread), and diffusion root mean square delay spread (diff_rms_delay_spread_ns). The above ranking results indicate that the asymmetric differences between uplink and downlink are important factors affecting PTP error estimation.

[0109] like Figure 11 The performance comparison chart of LOS and NLOS groups shows that the root mean square error of the LOS line-of-sight group is about 11 ns, while the root mean square error of the NLOS non-line-of-sight group is about 28 ns, which verifies the robustness of the model in working effectively in non-line-of-sight environments.

[0110] like Figure 12 The comparison charts of RMSE for different mine scenarios show that the RMSE for the straight roadway scenario is approximately 6 ns, for the branch roadway scenario it is approximately 11 ns, for the turning roadway scenario it is approximately 13 ns, for the mixed and complex scenario it is approximately 24 ns, and for the non-line-of-sight scenario it is approximately 28 ns. These results indicate that as the mine propagation scenario changes from straight roadways to branch roadways, turning roadways, mixed and complex scenarios, and non-line-of-sight scenarios, the effects of multipath propagation and occlusion gradually increase, making it more difficult to estimate the residual error of PTP, and the overall model error shows an upward trend.

[0111] like Figure 13 The comparison chart of RMSE for different delay spread intervals shows that the RMSE for the low delay spread interval is approximately 3 ns, the RMSE for the medium delay spread interval is approximately 12 ns, and the RMSE for the high delay spread interval is approximately 30 ns. These results indicate that as the root mean square delay spread increases, the dispersion of multipath components on the time axis intensifies, significantly increasing the difficulty of estimating PTP residual error. This invention improves the model's ability to compensate for asymmetric delay errors under complex multipath conditions by introducing multipath statistical features such as root mean square delay spread, maximum excess delay, power entropy, and delay kurtosis.

[0112] like Figure 14 The results show that this embodiment conducted 30 independent paired repeated experiments on the original uncompensated PTP baseline, the Ridge linear basis model, the DualBranchMLP model, and the fusion compensation model of the Ridge linear basis model and the DualBranchMLP model. The fusion compensation model of the Ridge linear basis model and the DualBranchMLP model is the core embodiment model of this invention; the complex scene correction method is only used for supplementary comparison. Formal statistical results show that the RMSE of the uncompensated PTP baseline is 60.7187 ± 0.9426 ns, and the RMSE of the fusion compensation model of the Ridge linear basis model and the DualBranchMLP model is 18.0690 ± 0.7221 ns. The fusion compensation model of the Ridge linear basis model and the DualBranchMLP model shows an average 70.24% decrease in RMSE compared to the uncompensated PTP baseline.

[0113] The above results show that, under the conditions of mine scenario simulation data and consistent evaluation set in this embodiment, the fusion compensation method can reduce the original PTP timing error, and the core effect results have good repeatability.

[0114] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for asymmetric delay compensation of mine wireless PTP timing based on bidirectional link feature fusion, characterized in that, Includes the following steps: Step 1: Collect PTP two-way time synchronization data and form basic timestamp samples; Step 2: Within the same synchronization period or adjacent aligned window as the data acquisition process in Step 1, record the downlink quality parameters, uplink quality parameters, and multipath statistical parameters respectively; Step 3: Generate contextualized tags for the mine structure; Step 4: Combine the timestamps, link quality, multipath statistics, and contextual labels obtained in Steps 1 to 3 according to a unified sample number to construct a complete feature system for bidirectional links; In step 4, the feature system includes directional channel features, uplink-downlink differential features, uplink-downlink ratio features, joint product interaction features, absolute difference features, and mine scenario enhancement features; Directional channel features include uplink received signal strength, uplink signal-to-noise ratio, uplink total channel gain, downlink received signal strength, downlink signal-to-noise ratio, and downlink total channel gain, which are used to characterize the directional channel states of the uplink and downlink, respectively. The uplink and downlink differential characteristics include differential received signal strength, differential signal-to-noise ratio, differential channel total gain, differential root mean square delay spread, and differential maximum excess delay, which are used to directly characterize the asymmetric propagation characteristics of the uplink and downlink wireless links. The uplink / downlink ratio characteristics include signal strength ratio, signal-to-noise ratio ratio, channel gain ratio, delay spread ratio, and tail delay ratio, which are used to characterize the relative variation and correlation characteristics of the uplink / downlink bidirectional link channel features. The joint product interaction features include the joint product of delay spread, the joint product of tail delay, and the joint product of power entropy, which are used to characterize the nonlinear coupling combination effect between the basic channel features. The absolute difference features include channel state absolute deviation, power entropy absolute deviation, and delay spread absolute deviation, which are used to characterize the degree of absolute difference between the corresponding feature parameters of uplink and downlink. The enhanced features of the mine scene include main roadway length, receiving and sending distance, turning angle, branch angle, line-of-sight and non-line-of-sight signs, and scene complexity score; In step 4, after the feature system is constructed, the feature set of the complete feature system is split into three types of neural network inputs: uplink branch input, downlink branch input, and differential auxiliary branch input, as well as a table model input containing the complete feature set; The uplink branch input includes the received signal strength indication, signal-to-noise ratio, channel gain, and multipath statistics in the uplink direction; the downlink branch input includes the corresponding features in the downlink direction; and the differential auxiliary branch input includes uplink and downlink differential features, uplink and downlink ratio features, joint product interaction features, and absolute difference features. Step 5: Estimate the asymmetric time delay error of the feature system based on the DualBranchMLP model; In step 5, in the DualBranchMLP model, the up-row branch and the down-row branch are the two main branches, and differential auxiliary branches and fusion layers are set up. The relationship between the layers is as follows: h down = f down (X down ); h up = f up (X up ); h diff = f diff (X diff ); = g([h down , h up , h diff ]); Among them, f up For the upward branch, f down For the downward branch, f diff For differential auxiliary branches, g is the fusion layer. X represents the error estimate output by the DualBranchMLP model. up X is the uplink feature vector. down X is the downlink feature vector. diff h is the difference auxiliary feature vector. up h is the latent vector of the upward branch. down h is the latent vector of the downlink branch. diff The latent vector for the difference auxiliary branch; Step 6: The asymmetric time delay error estimation is fused using the Ridge linear basis model and the DualBranchMLP model to obtain the asymmetric error estimate. Then, the original PTP time deviation is compensated based on the asymmetric error estimate, and the compensated synchronization deviation is output.

2. The asymmetric delay compensation method for mine wireless PTP timing based on bidirectional link feature fusion as described in claim 1, characterized in that, Step 1, the PTP two-way time synchronization data acquisition process includes: Step 1.1: Collect PTP two-way time synchronization data and obtain timestamp data; Step 1.2: After obtaining the timestamp data, generate a derived value of the timestamp data; Among them, the derived quantities include the original PTP timing deviation estimate, the two-way round-trip time observation, the synchronization cycle number, the message sequence number, the transmit / receive direction identifier, and the abnormal message flag; Step 1.3: Finally, preprocess the collected timestamp data samples to obtain basic timestamp samples with direction, period, and quality markers.

3. The asymmetric delay compensation method for mine wireless PTP timing based on bidirectional link feature fusion as described in claim 1, characterized in that, In step 2, the downlink quality parameters include downlink received signal strength indication, downlink signal-to-noise ratio, downlink channel total gain, and downlink noise floor; the uplink quality parameters include uplink received signal strength indication, uplink signal-to-noise ratio, uplink channel total gain, and uplink noise floor; the multipath statistical parameters include the number of multipaths, the energy share of the dominant path, the root mean square delay spread, the maximum excess delay, the energy dispersion, the power entropy, and the delay kurtosis. All of the above observations are marked with an upward or downward direction and a sample number.

4. The asymmetric delay compensation method for mine wireless PTP timing based on bidirectional link feature fusion as described in claim 1, characterized in that, Step 3, the process of generating contextualized labels for the mine structure, includes: Based on the tunnel map, node locations, tunnel width and height, turning angle, branch angle, node height difference, equipment obstruction, support obstruction, vehicle obstruction, and manual inspection records, generate sight distance / non-sight distance signs, straight tunnel signs, turning tunnel signs, branch tunnel signs, obstruction type signs, and complex multipath level signs.

5. The asymmetric delay compensation method for mine wireless PTP timing based on bidirectional link feature fusion as described in claim 1, characterized in that, The uplink branch, downlink branch, and differential auxiliary branch all consist of two fully connected layers with a hidden dimension of 128. Each fully connected layer is followed by a batch normalization layer, a ReLU activation function, and a Dropout layer in sequence, with a Dropout ratio of 0.

02. The three branches output 128-dimensional implicit vectors: the uplink implicit representation, the downlink implicit representation, and the difference-aided implicit representation, respectively. The input to the fusion layer is the concatenation of three latent vectors. It first passes through a 384-dimensional to 128-dimensional fully connected layer, a ReLU activation function, a batch normalization layer, and a Dropout layer. Then it passes through a 128-dimensional to 64-dimensional fully connected layer and a ReLU activation function. Finally, it passes through a 64-dimensional to 1-dimensional linear output layer to obtain the asymmetric error estimate.

6. The asymmetric delay compensation method for mine wireless PTP timing based on bidirectional link feature fusion as described in claim 1, characterized in that, In step 6, when performing fusion compensation, the fusion formula is: ; and ; in, This represents the asymmetric time delay error estimate output by the fusion model. This represents the error estimate output by the Ridge linear basis model. The coefficients are the error estimates output by the DualBranchMLP model. and These are the corresponding fusion weights.

Citation Information

Patent Citations

  • High-precision time synchronization method and model based on multi-channel fusion perception

    CN114650114A

  • Method and device for implementing automatic compensation for asymmetric delay of 1588 link

    WO2013020499A1