A method for compensating for clock synchronization delay in a distributed computer network

Abstract

This invention relates to the field of clock synchronization technology for audio / video stream scheduling networks, specifically a method for clock synchronization delay compensation in distributed computer networks. The method includes: acquiring multiple sets of original timestamp data through bidirectional message exchange between master and slave clock nodes in the audio / video stream scheduling data network; filtering and removing abnormal data sets with timestamp intervals exceeding a threshold to obtain valid timestamp data; calculating the path delay estimate corresponding to the valid timestamp data; modeling the delay distribution probability density function using a kernel density estimation algorithm; extracting the path delay value corresponding to the density peak as the network transmission reference delay; correcting the initial value of the clock deviation using the reference delay; and obtaining the optimized clock deviation compensation amount through iterative convergence using a recursive Kalman filter. This method can accurately determine the reference delay, improve the accuracy of clock deviation compensation, and adapt to the clock synchronization requirements of audio / video stream scheduling data networks.

Description

Technical Field

[0001] This invention relates to the field of network clock synchronization technology for audio and video stream scheduling, and in particular to a method for compensating for clock synchronization delay in distributed computer networks. Background Technology

[0002] Audio and video stream scheduling data networks have stringent requirements for clock synchronization accuracy. Existing distributed computer network clock synchronization delay compensation technologies mostly rely on the bidirectional message exchange mechanism between master and slave clock nodes to obtain the original timestamp data. They directly calculate the path delay value through simple statistical methods such as mean and median, and then use conventional Kalman filtering to complete the clock deviation compensation calculation based on the initial value of clock deviation directly calculated by the slave clock node. This is the mainstream implementation method for clock synchronization delay compensation in audio and video stream scheduling scenarios.

[0003] The raw timestamp data obtained by existing technologies is susceptible to interference from fluctuations in the operating conditions of the audio and video stream scheduling data network, resulting in a large amount of invalid data with abnormal timestamp intervals. Simple statistical methods cannot match the actual distribution pattern of path delays, leading to significant path delay estimation biases. Directly using uncorrected initial clock offset values ​​for filtering calculations means that the baseline delay error is directly transmitted to the clock offset calculation stage, and the recursive calculation process is prone to convergence deviations, failing to meet the high precision and high stability requirements of clock synchronization in audio and video stream scheduling data networks.

[0004] The original timestamp data needs to be filtered and removed for abnormal data. The distribution of path delay estimates is modeled based on the kernel density estimation algorithm. The value corresponding to the probability density peak is extracted to determine the reference delay of network transmission. At the same time, the initial value of clock deviation is corrected with this reference delay. Then, the corrected clock deviation is iteratively converged by recursive Kalman filtering to improve the problems of inaccurate delay estimation and insufficient accuracy of clock deviation compensation. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for compensating for clock synchronization delay in distributed computer networks.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for compensating clock synchronization delay in a distributed computer network, comprising: Between the master clock node and any slave clock node of the audio and video stream scheduling data network, multiple sets of original timestamp data for the clock synchronization process are obtained through a two-way message exchange mechanism. The multiple sets of original timestamp data are filtered and processed to remove abnormal data sets whose timestamp intervals exceed the preset network operating condition fluctuation threshold, thereby obtaining multiple sets of valid timestamp data. For each set of valid timestamp data, the corresponding path delay estimate is calculated. The path delay estimate is derived from the transmission time difference between the synchronization message and the delay request message on the bidirectional path, combined with the symmetry assumption of the bidirectional path. The path delay estimates from multiple sets are aggregated, and the distribution of the path delay estimates is modeled using a kernel density estimation algorithm to obtain the delay distribution probability density function. Extract the path delay value corresponding to the probability density peak from the delay distribution probability density function, and determine the path delay value as the final network transmission reference delay; Based on the network transmission reference delay and the original timestamp data, the initial value of the clock deviation calculated from the clock node is corrected, and the corrected clock deviation is iteratively converged using a recursive Kalman filter to obtain the optimized clock deviation compensation amount.

[0007] As a further aspect of the present invention, the acquired sets of original timestamp data are filtered, including: The original timestamp data includes the departure time of the synchronization message sent by the master clock node, the arrival time of the synchronization message received by the slave clock node, the departure time of the delay request message sent by the slave clock node, and the arrival time of the delay request message received by the master clock node. Calculate the complete round-trip time from the departure time of the synchronization message to the arrival time of the delay request message in each set of original timestamp data; Calculate the absolute value of the difference between the complete round-trip time in two adjacent exchanges; The absolute value of the difference is compared with the preset network operating condition fluctuation threshold; If the absolute value of the difference is greater than the network operating condition fluctuation threshold, the corresponding original timestamp data is determined to be the abnormal data group and is removed. If the absolute value of the difference is not greater than the network operating condition fluctuation threshold, then the original timestamp data is retained as the valid timestamp data.

[0008] As a further aspect of the present invention, the step of calculating the corresponding path delay estimate for each set of valid timestamp data includes: Four time values ​​are extracted from a set of valid timestamp data, namely the departure time T1 of the synchronization message, the arrival time T2 of the synchronization message, the departure time T3 of the delay request message, and the arrival time T4 of the delay request message; Assume that the downlink path delay from the master clock node to the slave clock node is equal to the uplink path delay from the slave clock node to the master clock node; Based on the time relationship of the bidirectional message exchange mechanism, the total bidirectional transmission time is calculated as the arrival time T4 of the delay request message minus the departure time T1 of the synchronization message; The dwell time from the clock node is calculated as the departure time T3 of the delay request message minus the arrival time T2 of the synchronization message; Subtract the dwell time of the slave clock node from the total bidirectional transmission time, and then divide by the number two. The result is used as the path delay estimate corresponding to the valid timestamp data.

[0009] As a further aspect of the present invention, the step of modeling the distribution of the path delay estimate using a kernel density estimation algorithm includes: The Gaussian kernel function is set as the kernel function of the kernel density estimation algorithm; For all the calculated path delay estimates, a smoothing bandwidth parameter is determined. This smoothing bandwidth parameter is used to control the smoothness of the probability density curve output by the kernel density estimation algorithm. A kernel defined by the Gaussian kernel function is placed centered on each path delay estimate sample; The Gaussian kernel functions corresponding to all path delay estimates are summed and then normalized by dividing by the total number of samples. A continuous curve describing the probability of the path delay estimate occurring on the entire real number axis is obtained, and the continuous curve is the probability density function of the delay distribution.

[0010] As a further aspect of the present invention, extracting the path delay value corresponding to the probability density peak from the delay distribution probability density function includes: A global optimization search is performed on the time-delay probability density function to find the point with the largest function value; Record the x-coordinate value corresponding to the point with the largest function value; The horizontal axis value is used as the final network transmission reference delay.

[0011] As a further aspect of the present invention, based on the network transmission reference delay and the original timestamp data, the initial value of the clock deviation calculated from the clock node is corrected, including: Read the arrival time T2 of the most recent message exchange from the original timestamp data buffer of the clock node; Read the departure time T1 of the synchronization message corresponding to the arrival time T2 of the synchronization message from the original timestamp data cache of the master clock node; Calculate the difference between the arrival time T2 and the departure time T1 of the synchronization message; Subtract the final network transmission reference delay from the difference; The result of the subtraction is used as the initial value of the clock deviation calculated from the clock node.

[0012] As a further aspect of the present invention, the iterative convergence calculation of the corrected clock skew using a recursive Kalman filter includes: Establish a state-space model describing the clock bias and frequency offset of the slave clock node, where the state variables include the corrected clock bias and the relative frequency offset of the slave clock node's local oscillator relative to the master clock node. Define the state transition matrix and process noise covariance matrix of the state-space model; The observation matrix and the observation noise variance are defined using the initial value of the new clock offset calculated after each message exchange as the observation value. Initialize the state estimation vector and error covariance matrix of the recursive Kalman filter; Each time a new observation is obtained, the standard time update step and measurement update step of the recursive Kalman filter are performed; In the measurement update step, the state component defined as clock bias in the output state estimation vector is the optimized clock bias compensation amount at the current moment.

[0013] As a further aspect of the present invention, after obtaining the optimized clock skew compensation amount, the method further includes: applying the optimized clock skew compensation amount to the local clock drive circuit of the slave clock node to achieve frequency and phase adjustment of the slave clock node, specifically including: The optimized clock skew compensation output from the recursive Kalman filter is converted from a digital quantity into a high-precision analog voltage control signal. The analog voltage control signal is applied to the control voltage input terminal of the local voltage-controlled temperature-compensated crystal oscillator from the clock node; The output frequency of the local voltage-controlled temperature-compensated crystal oscillator is finely adjusted under the action of the analog voltage control signal; The local voltage-controlled temperature-compensated crystal oscillator outputs an adjusted clock signal to drive all timing and sampling circuits of the slave clock node, thereby achieving frequency and phase synchronization with the master clock node.

[0014] As a further aspect of the present invention, it also includes a step of periodically recalibrating the network transmission reference delay: After the clock synchronization system enters a stable synchronization state, the bidirectional message exchange mechanism is triggered again according to a preset long period to collect a new set of original timestamp data. Perform the filtering process and the path delay estimation calculation steps on the new original timestamp data to obtain a new path delay estimate; The new path delay estimate, together with multiple historically stored path delay estimates, forms an expanded path delay estimate sample set. For the extended path delay estimate sample set, re-execute all steps from modeling with the kernel density estimation algorithm to extracting the final network transmission reference delay; The recalculated final network transmission reference delay is used to update the delay reference upon which subsequent clock offset initial value corrections depend.

[0015] As a further aspect of the present invention, it also includes a step of adaptively adjusting the noise covariance of the recursive Kalman filtering process: The innovation sequence output by the recursive Kalman filter is continuously monitored, and the innovation sequence is the difference sequence between the observed value and the predicted observed value. Calculate the statistical variance of the innovation sequence within the sliding time window; The calculated statistical variance is compared with the preset nominal observation noise variance; If the statistical variance is consistently greater than the nominal observed noise variance, it is determined that the network environment disturbance has intensified, and the corresponding element value in the process noise covariance matrix of the recursive Kalman filter is increased according to a preset strategy. If the statistical variance is consistently less than the nominal observed noise variance, the network environment is determined to be stable, and the corresponding element value in the process noise covariance matrix of the recursive Kalman filter is reduced according to a preset strategy. The adjusted process noise covariance matrix is ​​used for the next round of iterative calculation of the recursive Kalman filter.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are as follows: Multiple sets of raw timestamp data obtained from the bidirectional message exchange between the master and slave clock nodes of the audio / video stream scheduling data network are filtered. Abnormal data sets with timestamp intervals exceeding a preset network condition fluctuation threshold are removed, as are invalid timestamp data caused by network condition fluctuations. Valid timestamp data reflecting the actual transmission status is retained. For each set of valid timestamp data, the corresponding path delay estimate is calculated. A kernel density estimation algorithm is used to model the distribution of multiple sets of path delay estimates, generating a delay distribution probability density function. The path delay value corresponding to the peak probability density value is extracted from this function as the network transmission baseline delay, accurately matching the actual distribution characteristics of path delay and obtaining a baseline delay value that matches the network transmission status.

[0017] The initial value of the clock deviation calculated from the clock node is corrected based on a defined network transmission reference delay to eliminate the initial value error of the clock deviation caused by the reference delay deviation. A recursive Kalman filter is used to perform iterative convergence calculations on the corrected clock deviation, gradually optimizing the calculation results to form a clock deviation compensation amount adapted to the audio / video stream scheduling data network. This weakens the interference of network fluctuations and initial value errors on the clock synchronization compensation results, improving the iterative convergence effect of clock deviation compensation. Attached Figure Description

[0018] Figure 1 This is a flowchart of a distributed computer network clock synchronization delay compensation method according to the present invention; Figure 2 A flowchart for calculating the corresponding path delay estimate for each set of valid timestamp data; Figure 3 Comparison of the effects of correcting the initial value of clock deviation; Figure 4 The kernel density estimation distribution of the path delay estimate; Figure 5 This is a comparison chart of the distribution of historical delay samples and the characteristics of newly added delay samples under periodic recalibration. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0021] See Figure 1 This invention provides a method for compensating clock synchronization delay in a distributed computer network. The overall implementation scheme includes: In the audio / video stream scheduling data network, multiple sets of raw timestamp data for the clock synchronization process are obtained between the master clock node and any slave clock node through a bidirectional message exchange mechanism. These raw timestamp data sets are then filtered to remove abnormal data sets whose timestamp intervals exceed a preset network operating condition fluctuation threshold, resulting in multiple sets of valid timestamp data. For each set of valid timestamp data, a corresponding path delay estimate needs to be calculated. This path delay estimate is derived based on the transmission time difference between the synchronization message and the delay request message on the bidirectional path, and is based on the symmetry assumption of the bidirectional path. The multiple path delay estimates are then aggregated, and a kernel density estimation algorithm is used to model the distribution of these estimates to obtain a delay distribution probability density function. From the established delay distribution probability density function, the path delay value corresponding to the peak probability density is extracted and determined as the final network transmission reference delay. Finally, based on the network transmission reference delay and the original timestamp data, the initial value of the clock deviation calculated from the clock node is corrected, and the corrected clock deviation is iteratively converged using a recursive Kalman filter to obtain the optimized clock deviation compensation amount.

[0022] In one embodiment of the present invention, the process of filtering and processing the original timestamp data and calculating the path delay estimate is specifically implemented. See also... Figure 2 The original timestamp data includes the departure time T1 of the synchronization message sent by the master clock node, the arrival time T2 of the synchronization message received from the clock node, the departure time T3 of the delay request message sent by the clock node, and the arrival time T4 of the delay request message received by the master clock node. During processing, the complete round-trip time from the departure time T1 of the synchronization message to the arrival time T4 of the delay request message is calculated sequentially for each set of original timestamp data. Then, the absolute value of the difference between the complete round-trip times in two adjacent message exchanges is calculated. The calculated absolute value of the difference is compared with a preset network condition fluctuation threshold. If the absolute value of the difference is greater than the network condition fluctuation threshold, the corresponding original timestamp data is determined to be an abnormal data group and is removed. If the absolute value of the difference is not greater than the network condition fluctuation threshold, the corresponding original timestamp data is retained as valid timestamp data. For each set of valid timestamp data obtained after filtering, four time values ​​T1, T2, T3, and T4 are extracted. Assuming the downlink path delay from the master clock node to the slave clock node is equal to the uplink path delay from the slave clock node to the master clock node, the total bidirectional transmission time is calculated as T4 minus T1. The dwell time of the slave clock node is calculated as T3 minus T2. Subtracting the dwell time of the slave clock node from the total bidirectional transmission time and then dividing the result by two, the result is used as the path delay estimate corresponding to this set of valid timestamp data.

[0023] In specific implementation, the original timestamp data includes the departure time T1 of the synchronization message sent by the master clock node, the arrival time T2 of the synchronization message received from the clock node, the departure time T3 of the delay request message sent by the clock node, and the arrival time T4 of the delay request message received by the master clock node. For multiple sets of original timestamp data, the complete round-trip time from the departure time T1 of the synchronization message to the arrival time T4 of the delay request message in each set of original timestamp data is calculated in turn. The complete round-trip time is obtained by subtracting T1 from T4. The absolute value of the difference between the complete round-trip times in two adjacent exchanges is calculated. The absolute value of the difference is compared with the preset network operating condition fluctuation threshold. If the absolute value of the difference is greater than the network operating condition fluctuation threshold, the corresponding original timestamp data is determined to be an abnormal data group and is removed. If the absolute value of the difference is not greater than the network operating condition fluctuation threshold, the original timestamp data is retained as valid timestamp data. In some embodiments, the network condition fluctuation threshold is set to 80 microseconds. Considering two sets of consecutive raw timestamp data, the complete round-trip time of the first set of raw timestamp data is 600 microseconds, and the complete round-trip time of the second set of raw timestamp data is 500 microseconds. The absolute value of the difference is 100 microseconds. Since 100 microseconds is greater than the threshold of 80 microseconds, the second set of raw timestamp data is determined to be an abnormal data set and is removed. The complete round-trip time of the third set of raw timestamp data is 580 microseconds. The absolute value of the difference between the third set of raw timestamp data and the first set of raw timestamp data (600 microseconds) is 20 microseconds. Since 20 microseconds is not greater than the threshold of 80 microseconds, the third set of raw timestamp data is retained as valid timestamp data. In practical implementation, the path delay estimate is calculated for each set of valid timestamp data. Four time values, T1, T2, T3, and T4, are extracted from each set of valid timestamp data. Assuming the downlink path delay from the master clock node to the slave clock node is equal to the uplink path delay from the slave clock node to the master clock node, the total bidirectional transmission time is calculated based on the time relationship of the bidirectional message exchange mechanism: the arrival time T4 of the delay request message minus the departure time T1 of the synchronization message. The dwell time of the slave clock node is calculated as the departure time T3 of the delay request message minus the arrival time T2 of the synchronization message. The dwell time of the slave clock node is subtracted from the total bidirectional transmission time and then divided by two. The result is used as the path delay estimate corresponding to the valid timestamp data. The formula for calculating the path delay estimate is as follows: Where: d represents the path delay estimate, T1 represents the departure time of the synchronization message, T2 represents the arrival time of the synchronization message, T3 represents the departure time of the delay request message, and T4 represents the arrival time of the delay request message. In some embodiments, for a set of valid timestamp data, assuming T1 is 1000 microseconds, T2 is 1050 microseconds, T3 is 1100 microseconds, and T4 is 1200 microseconds, then the total bidirectional transmission time T4 minus T1 is 200 microseconds, the clock node dwell time T3 minus T2 is 50 microseconds, and the path delay estimate d is (200 microseconds minus 50 microseconds) divided by 2 equals 75 microseconds. Optionally, before calculating the path delay estimate, the time values ​​in the valid timestamp data can be standardized, for example, all time values ​​can be converted to microsecond precision. It can be understood that the calculation of the path delay estimate relies on the bidirectional path symmetry assumption, which holds true in scenarios where the path symmetry of the audio / video stream scheduling data network is good. In practice, the filtering and path delay estimation calculations are repeated cyclically until a predetermined number of valid timestamp data sets are accumulated, such as 100 sets. Optionally, the network condition fluctuation threshold can be dynamically adjusted according to network load; for example, the threshold can be increased during peak periods to tolerate greater network fluctuations. It is understood that removing abnormal data sets helps reduce the impact of sudden network congestion or packet loss on delay estimation.

[0024] In one embodiment of the present invention, a kernel density estimation algorithm is used to model the distribution of path delay estimates and extract a baseline delay value. This process sets a Gaussian kernel function as the kernel function of the kernel density estimation algorithm. For all calculated path delay estimate samples, a smoothing bandwidth parameter is determined, which controls the smoothness of the probability density curve output by the kernel density estimation algorithm. During modeling, a kernel defined by the Gaussian kernel function is placed centered on each path delay estimate sample. The Gaussian kernel functions corresponding to all path delay estimate samples are superimposed and summed, then normalized by dividing by the total number of samples, resulting in a continuous curve describing the probability of the path delay estimate occurring on the entire real number axis; this curve is the delay distribution probability density function. Extracting the path delay value corresponding to the peak probability density value from the delay distribution probability density function requires a global optimization search of the delay distribution probability density function to find the point with the largest function value. The x-coordinate value corresponding to the point with the largest function value is recorded, and this x-coordinate value is used as the final network transmission baseline delay.

[0025] In specific implementations, a Gaussian kernel function is set as the kernel function for the kernel density estimation algorithm. The mathematical form of the Gaussian kernel function is a symmetric probability density function centered at a sample point. For all calculated path delay estimates, a smoothing bandwidth parameter is determined. This parameter controls the smoothness of the probability density curve output by the kernel density estimation algorithm, and its selection directly affects the bias and variance characteristics of the probability density function estimation. In some embodiments, a set of path delay estimates is collected, with sample data in the range [75.2, 76.1, 74.8, 150.3, 75.9, 76.5, 73.9, 75.0] microseconds, including an outlier of 150.3 microseconds that significantly deviates from the main distribution. The smoothing bandwidth parameter is set to 1.5. In specific implementations, a kernel defined by the Gaussian kernel function is placed centered on each path delay estimate sample. Each kernel function contributes a probability density value at any point x to be estimated. The Gaussian kernel functions corresponding to all path delay estimates are summed and then normalized by dividing by the total number of samples n. This yields a continuous curve describing the probability of a path delay estimate occurring on the entire real number axis. This continuous curve is the probability density function of the delay distribution, and the formula for calculating the probability density function of the delay distribution is as follows: in: This represents the estimated probability density at point x. This represents the total number of samples for the path delay estimate. Indicates the smoothing bandwidth parameter. Represents the standard Gaussian kernel function. The points representing the independent variables whose probability density is to be estimated. This represents the sample of the i-th path delay estimate. Optionally, the smoothing bandwidth parameter h can be determined by empirical rules or optimization methods, such as calculating it by multiplying the sample standard deviation by a coefficient. In some embodiments, the probability density is calculated based on the above sample data. In the range of x from 70 microseconds to 160 microseconds, the probability density function exhibits a significant peak at approximately 75.5 microseconds, while the probability density value is very low near the outlier of 150.3 microseconds. This indicates that the kernel density estimation algorithm can effectively suppress the interference of isolated outlier samples on the overall distribution pattern judgment. To extract the path delay value corresponding to the probability density peak from the delay distribution probability density function, a global optimization search needs to be performed on the delay distribution probability density function to find the point with the largest function value. The global optimization search can traverse and calculate the probability density function value within a reasonable delay value range with a preset step size. The x-coordinate value corresponding to the point with the largest function value is recorded, and the x-coordinate value is used as the final network transmission reference delay. It can be understood that the x-coordinate corresponding to the probability density peak represents the most likely and representative path delay value in the sample data. In practice, the probability density function described above is searched, and it reaches its maximum value of 0.25 at x = 75.5 microseconds. Therefore, the baseline network transmission delay is determined to be 75.5 microseconds. Optionally, the global optimization search can be performed using either gradient ascent or direct search. It can be understood that, compared to directly calculating the sample arithmetic mean, the method of extracting the baseline network transmission delay based on kernel density estimation of peak values ​​has stronger robustness to outliers in the sample.

[0026] In one embodiment of the present invention, the initial value of clock deviation is corrected based on network transmission reference delay and original timestamp data, and the noise covariance of the recursive Kalman filter process is adaptively adjusted. When correcting the initial value of clock deviation, the arrival time T2 of the most recent message exchange synchronization message is read from the original timestamp data buffer of the slave clock node. The departure time T1 of the synchronization message corresponding to T2 is read from the original timestamp data buffer of the master clock node. The difference between T2 and T1 is calculated, and the final network transmission reference delay is subtracted from this difference. The result of the subtraction is used as the initial value of clock deviation calculated by the slave clock node. Adaptive adjustment of the noise covariance of the recursive Kalman filter process requires continuous monitoring of the innovation sequence output by the recursive Kalman filter, where the innovation sequence is the difference sequence between the observed value and the predicted observed value. The statistical variance of this innovation sequence within a sliding time window is calculated. The calculated statistical variance is compared with a preset nominal observation noise variance. If the statistical variance consistently exceeds the nominal observed noise variance, the network environment is considered to be experiencing increased disturbance, and the corresponding element values ​​in the process noise covariance matrix of the recursive Kalman filter are increased according to a preset strategy. If the statistical variance consistently falls below the nominal observed noise variance, the network environment is considered to be stabilizing, and the corresponding element values ​​in the process noise covariance matrix of the recursive Kalman filter are decreased according to a preset strategy. Finally, the adjusted process noise covariance matrix is ​​used for the next round of iterative calculations of the recursive Kalman filter.

[0027] In practice, the correction process reads the arrival time T2 of the most recent message exchange from the original timestamp data buffer of the clock node. Assuming the read T2 value is 100525.3 microseconds, it reads the departure time T1 of the synchronization message corresponding to the arrival time T2 from the original timestamp data buffer of the master clock node. Assuming the read T1 value is 100000.0 microseconds, it calculates the difference between the arrival time T2 and the departure time T1 of the synchronization message, which is 525.3 microseconds. It then subtracts the final network transmission reference delay from the difference. Assuming the final network transmission reference delay is 75.5 microseconds, subtracting 75.5 microseconds from 525.3 microseconds yields 449.8 microseconds. This 449.8 microseconds is used as the initial value of the clock deviation calculated from the clock node. In some embodiments, if a subsequent message exchange obtains a new T1' value of 101000.0 microseconds and a T2' value of 101524.7 microseconds, the difference is calculated to be 524.7 microseconds. This difference is corrected using the same network transmission reference delay of 75.5 microseconds, resulting in a new initial clock skew value of 449.2 microseconds. Adaptive adjustment of the noise covariance of the recursive Kalman filter process is performed according to claim 10. This requires continuous monitoring of the innovation sequence output by the recursive Kalman filter, where the innovation sequence is the difference sequence between the observed values ​​and the predicted observed values. In a specific implementation, the observed values ​​of the recursive Kalman filter are the initial clock skew values ​​calculated in the above steps. For example, the observed values ​​at multiple consecutive times are 449.8 microseconds, 449.2 microseconds, 450.1 microseconds, and 448.9 microseconds. The recursive Kalman filter will provide corresponding predicted observed values ​​based on the state prediction, and the statistical variance of the innovation sequence within the sliding time window is calculated. The formula for calculating the statistical variance of the innovation sequence is expressed as: in: This represents the statistical variance of the innovation sequence calculated within the sliding time window. Indicates the size of the sliding window. This represents the innovation value at time k. This represents the average value of the information within the sliding window. This indicates the starting time index of the sliding window. The calculated statistical variance is compared with the preset nominal observation noise variance, which is a baseline value set based on prior knowledge of the system's observation errors. See Table 1 for an example of news sequence monitoring data.

[0028] Table 1: Example Table of New Information Sequence Monitoring Data Optionally, the sliding window size N can be set to 5. The statistical variance of the innovation sequence is calculated based on the data in Table 1. The average value of the five innovation values ​​in the window in Table 1 is calculated to be 0.66 microseconds, and the statistical variance is... The variance is approximately 0.086 square microseconds. Assuming a preset nominal observation noise variance of 0.01 square microseconds, the calculated statistical variance of 0.086 square microseconds is greater than the nominal observation noise variance of 0.01 square microseconds. In some embodiments, if the statistical variance is consistently greater than the nominal observation noise variance, it is determined that the network environment disturbance has intensified. The corresponding element value in the process noise covariance matrix of the recursive Kalman filter is increased according to a preset strategy. For example, the matrix element value describing clock skew noise in the process noise covariance matrix is ​​adjusted from 0.001 to 0.01. It can be understood that increasing the process noise covariance means that the filtering model considers the uncertainty of the system state to increase over time, thus making the filter more confident in new observations. In specific implementations, if the statistical variance is consistently less than the nominal observation noise variance, it is determined that the network environment is becoming stable. The corresponding element value in the process noise covariance matrix of the recursive Kalman filter is decreased according to a preset strategy. For example, the matrix element value describing clock skew noise is adjusted from 0.001 to 0.0001. Optionally, the adjustment strategy can be linear, for example, setting the element values ​​of the process noise covariance matrix to be proportional to the ratio of the statistical variance to the nominal observation noise variance. The adjusted process noise covariance matrix is ​​then used for the next round of iterative calculation of the recursive Kalman filter. It can be understood that through this adaptive adjustment, the recursive Kalman filter can automatically adjust the confidence weights of its state estimates as the network environment changes, thereby maintaining the robustness of the estimates.

[0029] See Figure 3In the clock synchronization delay compensation method for distributed computer networks, the effect of correcting the initial value of the clock deviation is intuitively presented by comparing the clock deviation sequences before and after correction. Specifically, the clock deviation before correction is calculated directly from the clock node based on the original timestamp data. The calculation logic is the difference between the arrival time T2 and departure time T1 of the synchronization message. Because the influence of network transmission delay is not eliminated, this result remains in the high range of approximately 520-530 μs and exhibits slight oscillations due to network operating condition fluctuations. The corrected clock deviation is obtained by subtracting the network transmission reference delay extracted based on kernel density estimation (approximately 75.5 μs in this embodiment) from the above difference. The corrected clock deviation is stably compressed to the low range of approximately 450 μs, and the sequence fluctuation amplitude remains consistent with that before correction. This verifies that the correction method only eliminates the systematic deviation of the fixed network transmission delay without introducing additional noise. The comparison results show that the clock deviation initial value correction method based on network transmission reference delay can effectively eliminate the interference of network transmission delay on clock deviation calculation, provide high-precision initial input for the iterative convergence of subsequent recursive Kalman filtering, and ensure the overall synchronization accuracy of the clock synchronization system.

[0030] In one embodiment of the present invention, a recursive Kalman filter is used to iteratively converge to calculate the corrected clock skew and to apply the optimized clock skew compensation to the local clock. A state-space model describing the clock skew and frequency offset of the slave clock node is established, with state variables including the corrected clock skew and the relative frequency offset of the slave clock node's local oscillator relative to the master clock node. The state transition matrix and process noise covariance matrix of the state-space model are defined. Using the new initial value of the clock skew calculated after each message exchange as the observation value, the observation matrix and observation noise variance are defined. The state estimation vector and error covariance matrix of the recursive Kalman filter are initialized. Each time a new observation value is obtained, the standard time update step and measurement update step of the recursive Kalman filter are executed. In the measurement update step, the state component defined as the clock skew in the output state estimation vector is the optimized clock skew compensation at the current moment. Applying the optimized clock skew compensation to the local clock drive circuit of the slave clock node requires converting the optimized clock skew compensation output by the recursive Kalman filter from a digital quantity into a high-precision analog voltage control signal. The analog voltage control signal is applied to the control voltage input of the local temperature-controlled crystal oscillator (TCC-O) at the slave clock node. The output frequency of the local TCC-O is finely adjusted under the action of the analog voltage control signal. The adjusted clock signal output by the local TCC-O drives all timing and sampling circuits at the slave clock node, thereby achieving frequency and phase synchronization with the master clock node.

[0031] In specific implementation, a recursive Kalman filter is used to iteratively converge the corrected clock deviation according to claim 7. A state-space model describing the clock deviation and frequency offset of the slave clock node is established. The state variables include the corrected clock deviation and the relative frequency offset of the slave clock node's local oscillator relative to the master clock node. The state transition matrix and process noise covariance matrix of the state-space model are defined. The state transition formula of the state-space model is expressed as: in: This represents the state vector at time k, containing the clock offset state component and the relative frequency offset state component. Represents the state transition matrix. This represents the state vector at time k-1. Let k represent the process noise vector at time k. Using the initial value of the new clock offset calculated after each message exchange as the observation value, the observation matrix and observation noise variance are defined. The observation equation can be expressed as: ,in This represents the observation value at time k (i.e., the corrected initial value of the clock offset). Represents the observation matrix. This represents the observation noise at time k. The state estimation vector and error covariance matrix of the recursive Kalman filter are initialized. The initial value of the state estimation vector can be set to zero, and the initial value of the error covariance matrix can be set to a diagonal matrix. Each time a new observation is obtained, the standard time update step and measurement update step of the recursive Kalman filter are executed. The time update step is responsible for predicting the state and error covariance at the next time step, and the measurement update step uses the new observation to correct the prediction. In the measurement update step, the state component defined as the clock skew in the output state estimation vector is the optimized clock skew compensation amount at the current time. Refer to Table 2, which shows a simplified data segment of the recursive Kalman filter iteration.

[0032] Table 2: Example of iterative data for recursive Kalman filter In some embodiments, the state transition matrix It can be set to ,in This represents the time interval between two message exchanges. Optionally, the process noise covariance matrix and the observation noise variance can be set based on prior knowledge of the oscillator noise characteristics and network delay measurement noise. It is understood that the recursive Kalman filter, by fusing historical state information and the latest observations, can effectively smooth out random fluctuations in the observation noise, thereby obtaining a more stable and accurate clock skew compensation. In a specific implementation, the optimized clock skew compensation is applied to the local clock drive circuit of the clock node according to claim 8, converting the optimized clock skew compensation output by the recursive Kalman filter from a digital quantity into a high-precision analog voltage control signal. The conversion process is completed by a high-resolution digital-to-analog converter. In some embodiments, the digital-to-analog converter converts the digital clock skew compensation into an analog voltage signal ranging from 0 to 3.3 volts. The analog voltage control signal is applied to the control voltage input of the local voltage-controlled temperature-compensated crystal oscillator of the clock node. The output frequency of the local voltage-controlled temperature-compensated crystal oscillator is finely adjusted under the action of the analog voltage control signal. The frequency-voltage control characteristics of the voltage-controlled temperature-compensated crystal oscillator determine the correspondence between the output voltage change and the output frequency change. Optionally, a low-pass filter can be used to smooth the voltage signal before applying the analog voltage control signal to reduce high-frequency jitter. The local voltage-controlled temperature-compensated crystal oscillator outputs an adjusted clock signal, driving all timing and sampling circuits from the clock node. The frequency and phase of the adjusted clock signal gradually align with the clock signal of the master clock node. This can be understood as a closed-loop control process. The recursive Kalman filter continuously outputs the optimized clock deviation compensation amount, which is converted into a voltage signal to adjust the local voltage-controlled temperature-compensated crystal oscillator, thereby dynamically achieving frequency and phase synchronization with the master clock node.

[0033] See Figure 4In the distributed computer network clock synchronization delay compensation method, the determination of the network transmission reference delay relies on the kernel density estimation algorithm to model the distribution of path delay estimates. Specifically, multiple sets of raw timestamp data are obtained through the bidirectional message exchange mechanism between the master and slave clock nodes of the audio / video stream scheduling data network. After filtering out abnormal data, valid timestamp data sets are obtained. For each set of valid timestamp data, path delay estimate samples are calculated based on the assumption of bidirectional path symmetry. All path delay estimate samples are input into the kernel density estimation algorithm. Based on the Gaussian kernel function, a smoothing bandwidth parameter is set to control the curve smoothness. The Gaussian kernels corresponding to each sample are superimposed, summed, and normalized, ultimately generating the delay distribution probability density curve (solid line) shown in the figure. The filled area under the curve represents the probability density integral interval. A global optimization search is performed on this probability density curve to extract the horizontal axis value corresponding to the probability density peak, which is determined as the final network transmission reference delay. The dashed line in the figure marks the specific value of this reference delay as 449.5 ns. This value provides an accurate delay reference for subsequent correction from the initial value of the clock node clock deviation. During parameter configuration, a Gaussian kernel function is used, and the smoothing bandwidth parameter is adaptively determined based on the statistical characteristics of the path delay estimation sample to balance the smoothness of the probability density curve with the degree of distribution detail reproduction.

[0034] In one embodiment of the present invention, the step of periodically recalibrating the network transmission reference delay is specifically implemented. After the clock synchronization system enters a stable synchronization state, a bidirectional message exchange mechanism is triggered again according to a preset long period to collect a new set of original timestamp data. The new original timestamp data is filtered and path delay estimation is calculated to obtain a new path delay estimate. The new path delay estimate, together with multiple historically stored path delay estimates, constitutes an expanded path delay estimate sample set. For the expanded path delay estimate sample set, all steps from kernel density estimation algorithm modeling to extraction of the final network transmission reference delay are re-executed. The recalculated final network transmission reference delay is used to update the delay reference upon which subsequent clock offset initial value corrections depend.

[0035] In specific implementation, the step of periodically recalibrating the network transmission reference delay is performed according to the content of claim 9. After the clock synchronization system enters a stable synchronization state, the bidirectional message exchange mechanism is retried according to a preset long period to collect a new set of original timestamp data. The preset long period can be set according to the actual network environment and accuracy requirements, for example, 24 hours or 1000 seconds. In some embodiments, after the clock synchronization system has been running continuously for 24 hours, it determines that it has entered a stable synchronization state. The system automatically triggers a new bidirectional message exchange. The master clock node sends a synchronization message, the slave clock node records the arrival time of the synchronization message T2 as 86400075 microseconds and sends a delay request message to the master clock node, the master clock node records the departure time of the synchronization message T1 as 86400000 microseconds and the arrival time of the delay request message T4 as 86400180 microseconds, and the slave clock node records the departure time of the delay request message T3 as 86400100 microseconds, thereby collecting a new set of original timestamp data [T1=86400000,T2=86400075,T3=86400100,T4=86400180]=. In practice, the new raw timestamp data undergoes filtering and path delay estimation. The complete round-trip time is calculated as T4 minus T1, equaling 180 microseconds. The dwell time from the clock node is calculated as T3 minus T2, equaling 25 microseconds. Based on the path delay estimation formula, a new path delay estimate is calculated as (180-25) / 2 = 77.5 microseconds. Optionally, the filtering process checks whether the absolute value of the difference between the new and previous complete round-trip times exceeds a network operating condition fluctuation threshold. If it does not exceed the threshold, the new data set is considered valid. The new path delay estimate, along with multiple historically stored path delay estimates, forms an expanded path delay estimate sample set. The historically stored path delay estimates can be samples calculated and retained during the initial system synchronization phase or previous recalibration cycles. In some embodiments, the historically stored path delay estimate sample set contains 100 values, such as [75.2, 76.1, 74.8, 75.9, 76.5, 75.0, ...]. After adding the new path delay estimate of 77.5 microseconds, the expanded path delay estimate sample set contains 101 sample values. In a specific implementation, all steps from kernel density estimation algorithm modeling to extracting the final network transmission baseline delay are re-executed on the expanded path delay estimate sample set. The delay distribution probability density function is recalculated using a Gaussian kernel function and all expanded samples. The recalculation process of kernel density estimation is expressed as follows: in: This represents the probability density at point y estimated based on an expanded sample set. This represents the total number of samples for the expanded path delay estimate. This indicates the smoothing bandwidth parameter used in the recalibration step. Represents the standard Gaussian kernel function. The independent variable represents the probability density to be estimated. This represents the j-th path delay estimate sample in the expanded set. Optionally, the smoothing bandwidth parameter b can be the same as the parameter used in the initial modeling, or it can be recalculated based on the characteristics of the expanded samples. A global optimization search is performed on the obtained delay distribution probability density function to find the point with the largest function value, and the x-coordinate value corresponding to the point with the largest function value is recorded. It is understood that due to the addition of new samples, the shape and peak position of the probability density function may change slightly. In specific implementation, assuming that the search finds that the probability density function reaches its maximum value at y=76.2 microseconds, then 76.2 microseconds is used as the final network transmission reference delay recalculated. The recalculated final network transmission reference delay is used to update the delay reference on which the subsequent clock skew initial value correction depends. It is understood that in the next round or the next time the clock skew initial value needs to be calculated, 76.2 microseconds will be used instead of the old reference value for correction calculation. Optionally, the historically stored sample set can be managed using a first-in-first-out queue, discarding the oldest sample while adding new samples to maintain a fixed-size sample set.

[0036] See Figure 5 This study presents the distribution characteristics of historical path delay estimates and the numerical location of newly added delay samples. Specifically, after statistical analysis, the distribution of path delay estimates in the historical delay sample set exhibits clear discrete characteristics: the box range covers the interval from 75.0 μs to 75.9 μs, the median of the box corresponds to a typical delay level of approximately 75.4 μs, and the whisker line extends to define the sample fluctuation range from 74.0 μs to 77.0 μs. The outliers below the box plot reflect short-term extreme delay values ​​present in the historical samples. The newly added delay sample, marked with a dot, has a value of 77.5 μs, which is significantly higher than the upper whisker boundary and upper limit of the box in the historical samples. This represents a delay measurement result exceeding the historical normal fluctuation range collected through the new message exchange mechanism during the periodic recalibration phase. At the data integration level, after the new sample of 77.5μs is incorporated into the original 100 sets of historical samples to form an extended sample set of 101 sets, the kernel density estimation algorithm needs to be re-executed based on the extended set to model. A new delay distribution probability density function is generated by superimposing and normalizing the Gaussian kernel function. Then, the network transmission reference delay corresponding to the peak probability density is obtained through global optimization, thereby completing the dynamic update of the clock deviation correction reference.

[0037] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for compensating for clock synchronization delays in a distributed computer network, characterized by, include: Between the master clock node and any slave clock node of the audio and video stream scheduling data network, multiple sets of original timestamp data for the clock synchronization process are obtained through a two-way message exchange mechanism. The multiple sets of original timestamp data are filtered and processed to remove abnormal data sets whose timestamp intervals exceed the preset network operating condition fluctuation threshold, thereby obtaining multiple sets of valid timestamp data. For each set of valid timestamp data, the corresponding path delay estimate is calculated. The path delay estimate is derived from the transmission time difference between the synchronization message and the delay request message on the bidirectional path, combined with the symmetry assumption of the bidirectional path. The path delay estimates from multiple sets are aggregated, and the distribution of the path delay estimates is modeled using a kernel density estimation algorithm to obtain the delay distribution probability density function. Extract the path delay value corresponding to the probability density peak from the delay distribution probability density function, and determine the path delay value as the final network transmission reference delay; Based on the network transmission reference delay and the original timestamp data, the initial value of the clock deviation calculated from the clock node is corrected, and the corrected clock deviation is iteratively converged using a recursive Kalman filter to obtain the optimized clock deviation compensation amount.

2. The method for compensating clock synchronization delay in a distributed computer network according to claim 1, characterized in that, The acquired sets of raw timestamp data are filtered, including: The original timestamp data includes the departure time of the synchronization message sent by the master clock node, the arrival time of the synchronization message received by the slave clock node, the departure time of the delay request message sent by the slave clock node, and the arrival time of the delay request message received by the master clock node. Calculate the complete round-trip time from the departure time of the synchronization message to the arrival time of the delay request message in each set of original timestamp data; Calculate the absolute value of the difference between the complete round-trip time in two adjacent exchanges; The absolute value of the difference is compared with the preset network operating condition fluctuation threshold; If the absolute value of the difference is greater than the network operating condition fluctuation threshold, the corresponding original timestamp data is determined to be the abnormal data group and is removed. If the absolute value of the difference is not greater than the network operating condition fluctuation threshold, then the original timestamp data is retained as the valid timestamp data.

3. The distributed computer network clock synchronization delay compensation method according to claim 2, characterized in that, The step of calculating the corresponding path delay estimate for each set of valid timestamp data includes: Four time values ​​are extracted from a set of valid timestamp data, namely the departure time T1 of the synchronization message, the arrival time T2 of the synchronization message, the departure time T3 of the delay request message, and the arrival time T4 of the delay request message; Assume that the downlink path delay from the master clock node to the slave clock node is equal to the uplink path delay from the slave clock node to the master clock node; Based on the time relationship of the bidirectional message exchange mechanism, the total bidirectional transmission time is calculated as the arrival time T4 of the delay request message minus the departure time T1 of the synchronization message; The dwell time from the clock node is calculated as the departure time T3 of the delay request message minus the arrival time T2 of the synchronization message; Subtract the dwell time of the slave clock node from the total bidirectional transmission time, and then divide by the number two. The result is used as the path delay estimate corresponding to the valid timestamp data.

4. The distributed computer network clock synchronization delay compensation method according to claim 3, characterized in that, The step of modeling the distribution of the path delay estimates using a kernel density estimation algorithm includes: The Gaussian kernel function is set as the kernel function of the kernel density estimation algorithm; For all the calculated path delay estimates, a smoothing bandwidth parameter is determined. This smoothing bandwidth parameter is used to control the smoothness of the probability density curve output by the kernel density estimation algorithm. A kernel defined by the Gaussian kernel function is placed centered on each path delay estimate sample; The Gaussian kernel functions corresponding to all path delay estimates are summed and then normalized by dividing by the total number of samples. A continuous curve describing the probability of the path delay estimate occurring on the entire real number axis is obtained, and the continuous curve is the probability density function of the delay distribution.

5. A method for compensating clock synchronization delay in a distributed computer network according to claim 4, characterized in that, Extracting the path delay value corresponding to the probability density peak from the delay distribution probability density function includes: A global optimization search is performed on the time-delay probability density function to find the point with the largest function value; Record the x-coordinate value corresponding to the point with the largest function value; The horizontal axis value is used as the final network transmission reference delay.

6. A method for compensating clock synchronization delay in a distributed computer network according to claim 5, characterized in that, Based on the network transmission reference delay and the original timestamp data, the initial value of the clock offset calculated from the clock node is corrected, including: Read the arrival time T2 of the most recent message exchange from the original timestamp data buffer of the clock node; Read the departure time T1 of the synchronization message corresponding to the arrival time T2 of the synchronization message from the original timestamp data cache of the master clock node; Calculate the difference between the arrival time T2 and the departure time T1 of the synchronization message; Subtract the final network transmission reference delay from the difference; The result of the subtraction is used as the initial value of the clock deviation calculated from the clock node.

7. A method for compensating clock synchronization delay in a distributed computer network according to claim 6, characterized in that, The iterative convergence calculation of the corrected clock skew using a recursive Kalman filter includes: Establish a state-space model describing the clock bias and frequency offset of the slave clock node, where the state variables include the corrected clock bias and the relative frequency offset of the slave clock node's local oscillator relative to the master clock node. Define the state transition matrix and process noise covariance matrix of the state-space model; The observation matrix and the observation noise variance are defined using the initial value of the new clock offset calculated after each message exchange as the observation value. Initialize the state estimation vector and error covariance matrix of the recursive Kalman filter; Each time a new observation is obtained, the standard time update step and measurement update step of the recursive Kalman filter are performed; In the measurement update step, the state component defined as clock bias in the output state estimation vector is the optimized clock bias compensation amount at the current moment.

8. A method for compensating clock synchronization delay in a distributed computer network according to claim 7, characterized in that, After obtaining the optimized clock skew compensation amount, the method further includes: applying the optimized clock skew compensation amount to the local clock drive circuit of the slave clock node to adjust the frequency and phase of the slave clock node, specifically including: The optimized clock skew compensation output from the recursive Kalman filter is converted from a digital quantity into a high-precision analog voltage control signal. The analog voltage control signal is applied to the control voltage input terminal of the local voltage-controlled temperature-compensated crystal oscillator from the clock node; The output frequency of the local voltage-controlled temperature-compensated crystal oscillator is finely adjusted under the action of the analog voltage control signal; The local voltage-controlled temperature-compensated crystal oscillator outputs an adjusted clock signal to drive all timing and sampling circuits of the slave clock node, thereby achieving frequency and phase synchronization with the master clock node.

9. A method for compensating clock synchronization delay in a distributed computer network according to claim 8, characterized in that, It also includes the step of periodically recalibrating the network transmission reference delay: After the clock synchronization system enters a stable synchronization state, the bidirectional message exchange mechanism is triggered again according to a preset long period to collect a new set of original timestamp data. Perform the filtering process and the path delay estimation calculation steps on the new original timestamp data to obtain a new path delay estimate; The new path delay estimate, together with multiple historically stored path delay estimates, forms an expanded path delay estimate sample set. For the extended path delay estimate sample set, re-execute all steps from modeling with the kernel density estimation algorithm to extracting the final network transmission reference delay; The recalculated final network transmission reference delay is used to update the delay reference upon which subsequent clock offset initial value corrections depend.

10. A method for compensating clock synchronization delay in a distributed computer network according to claim 9, characterized in that, It also includes a step of adaptively adjusting the noise covariance of the recursive Kalman filtering process: Continuously monitor the innovation sequence output by the recursive Kalman filter, where the innovation sequence is the difference sequence between the observed value and the predicted observed value; Calculate the statistical variance of the innovation sequence within the sliding time window; The calculated statistical variance is compared with the preset nominal observation noise variance; If the statistical variance is consistently greater than the nominal observed noise variance, it is determined that the network environment disturbance has intensified, and the corresponding element value in the process noise covariance matrix of the recursive Kalman filter is increased according to the preset strategy. If the statistical variance is consistently less than the nominal observed noise variance, the network environment is determined to be stable, and the corresponding element value in the process noise covariance matrix of the recursive Kalman filter is reduced according to a preset strategy. The adjusted process noise covariance matrix is ​​used for the next round of iterative calculation of the recursive Kalman filter.

Images