Early warning method based on the causal model of optical links in broadcast backbone networks

Abstract

This invention discloses an early warning method based on a causal model of optical links in a broadcast backbone network, comprising the following steps: Step 1, constructing the causal relationships and topology of optical links in a provincial broadcast backbone network based on a causal relationship model; Step 2, establishing a dynamic reference benchmark and variation law representation system in the causal relationships and topology of optical links based on statistical analysis methods and dynamic programming; Step 3, integrating short-time memory networks and moving average autoregressive models to linearly predict the future trend of optical power, comparing the prediction results with the dynamic reference benchmark and variation law representation system to achieve high-quality intelligent early warning of trend-based optical link losses. This invention can achieve high-quality, highly generalized intelligent early warning of trend-based optical link losses, with the advantage of accurate results.

Description

Technical Field

[0001] This invention relates to the field of broadcast communication technology, and in particular to an early warning method based on a causal model of optical links in broadcast backbone networks. Background Technology

[0002] The Zhejiang Provincial Broadcasting Backbone Network's optical links cover all 11 prefecture-level cities in Zhejiang Province, consisting of 95 network element sites and 220 optical links. Spanning from the hilly and mountainous southwest to the Hangzhou-Jiaxing-Huzhou Plain in the northeast, and then to the Zhoushan Islands in the east, the network faces a complex geographical environment with dynamic variations in construction year, network structure, optical path distance, and transmission quality. The entire network topology can be roughly divided into a northern ring, a southern ring, and a partial ring around Quzhou in the southwest. Sites are generally equipped with bidirectional, four-way (primary and backup) optical cables, presenting a relatively complex dynamic network topology. To achieve early prediction of fiber optic transmission system quality using artificial intelligence technology, it is first necessary to quantify and model various factors in the backbone, clarifying the correlation and causality between each factor and optical transmission quality. This information will serve as prior knowledge for deep learning training, enhancing model performance. Traditional Granger causality only considers the impact of noise in factor prediction, neglecting other factors, which is insufficient in real-world optical link scenarios and makes accurate early warning difficult. Summary of the Invention

[0003] The purpose of this invention is to provide an early warning method based on a causal model of optical links in a broadcast backbone network. This invention can achieve high-quality, highly generalized intelligent early warning of trend-based optical link losses, and has the advantage of accurate results.

[0004] The technical solution of this invention: an early warning method based on a causal model of optical links in a broadcast backbone network, comprising the following steps:

[0005] Step 1: Based on the causal relationship model, construct the causal relationship and topology of the provincial broadcasting backbone optical links;

[0006] Step 2: Based on statistical analysis methods and dynamic programming, establish a characterization system for the causal relationships and dynamic reference benchmarks and change patterns in the topology of optical links;

[0007] Step 3: Integrate the short-time memory network and the moving average autoregressive model to make a linear prediction of the future trend of optical power. Compare the prediction results with the dynamic reference benchmark and the change law characterization system to achieve intelligent early warning of trend loss of high-quality optical links.

[0008] The aforementioned early warning method based on the causal model of the optical link in the broadcasting backbone network has the following formula for the causal relationship model:

[0009]

[0010] In the formula: n represents the time series Xi For time series X k The causal influence value is given by: X represents the optical power value that changes over time; D represents the transmission and reception direction of the optical cable line; N represents the observation length of the time series; t represents the current time point; m is the lag length of the time series; a is the model coefficient; tj represents the time points before j; h is the sample number; and η represents the model error.

[0011] The aforementioned early warning method based on the causal model of optical links in the broadcast backbone network uses network element sites as graph nodes and optical links between sites as graph boundaries. It uses a weighted directed efficient optical network to measure the efficient transmission in the optical link topology and a binary undirected optical network to analyze the optical link topology. The causal correlation, optical power, and optical signal-to-noise ratio quantification indicators are used as weight information in the weighted directed optical network.

[0012] The aforementioned early warning method based on the causal model of the optical link in the broadcast backbone network establishes a dynamic reference benchmark and a system for representing the changing patterns. This process is based on statistical analysis and dynamic programming. The time series data is solved one by one using a sliding window to find the local optimum. The sequence is analyzed to determine whether it is in a stable period or a loss period. The benchmark with the largest window size in the stable period is taken as the global optimum and the benchmark is updated dynamically. When the optical link is in a loss period, the benchmark is no longer updated until a special event occurs. At this time, the previous training results are discarded, and the next cycle of iterative search is started, recursively updating the benchmark.

[0013] The aforementioned early warning method based on the causal model of optical links in the broadcast backbone network involves linearly predicting the future trend of optical power by using historical data of the optical links as samples to input into the training model of the long short-term memory network, and then using a moving average autoregressive model to linearly predict the future trend of optical power.

[0014] Compared with existing technologies, this invention constructs the causal relationship and topology of provincial broadcast backbone optical links based on a causal relationship model. It explores in depth the topological connectivity and causal synergy mechanism of provincial optical link backbones under time-varying conditions. Then, based on statistical analysis methods and dynamic programming, it establishes a dynamic reference benchmark and change law representation system in the causal relationship and topology of optical links, thereby obtaining the corresponding benchmark. Finally, it integrates short-time memory networks and moving average autoregressive models to linearly predict the future trend of optical power. The prediction results are compared with the benchmark of the dynamic reference benchmark and change law representation system to achieve intelligent early warning of trend loss of high-quality optical links. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the causal relationship and topology of the provincial broadcasting backbone network optical links of the present invention;

[0016] Figure 2Analysis of common periods and events in optical links;

[0017] Figure 3 A schematic diagram of three optical link references;

[0018] Figure 4 This represents the error between real data and the prediction model at the link loss time point.

[0019] Figure 5 It represents the error between real data and the prediction model at the link maintenance time point;

[0020] Figure 6 This is a schematic diagram illustrating the dynamic programming algorithm for searching a dynamic optical link reference benchmark.

[0021] Figure 7 It is the prediction curve of the linear model;

[0022] Figure 8 It is the prediction curve of the nonlinear model;

[0023] Figure 9 It is a comparison of optical power prediction simulation. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments, but this should not be construed as limiting the present invention.

[0025] Example: An early warning method based on a causal model of optical links in a broadcast backbone network, comprising the following steps:

[0026] Step 1: Based on the causal relationship model, construct the causal relationship and topology of the provincial broadcasting backbone optical links; specifically, firstly, define the autoregressive model among multiple variables (common...

[0027] Equation 1) and the joint regression model (Equation 2):

[0028]

[0029]

[0030] In the formula: X represents the optical power value changing over time; N represents the observation length of the time series; t represents the current time point; m is the lag length of the time series; a is the model coefficient; tj represents the time points before the j-th time point; X 1，t X 2，t It is a variable, ∈ 1，t ,∈ 2，t η 1，t η 2，t These are model errors, which are caused by random noise. This noise may be introduced during optical link transmission due to system quality degradation or construction.

[0031] In the above formulas, formula (1) is a set of autoregressive models, where X1 and X2 are independent of each other and do not affect each other. Formula (2) establishes a set of multivariate regression models, where X1 and X2 have mutual causal influence, that is, the historical data of X1 affects the current optical power of X2, and the historical data of X2 also affects the current optical power of X1.

[0032] In traditional Granger causality, if the addition of a factor B leads to a decrease in the autoregressive prediction model error of another factor A, it indicates that factor B has a Granger causal effect on A, as shown in Equation 3:

[0033]

[0034] In the formula: It is variance.

[0035] Formula (3) calculates the causal value F between the two factors, representing the degree of influence of the X2 link optical power on X1. ln is calculated using log(10), where the numerator is the noise variance of formula (1) and the denominator is the noise variance of formula (2). If the numerator is greater than the denominator, it indicates that the noise variance decreases after the introduction of X2, meaning that X2 has an impact on X1. If the numerator is less than the denominator, it indicates that the prediction success rate of X1 decreases after the introduction of X2.

[0036] However, the traditional Granger causality method only considers the influence of noise terms in factor prediction. This invention proposes a causal relationship based on factor proportions, and the causal relationship model is shown in Equation 4:

[0037]

[0038] In the formula: n represents the time series X i For time series X k The causal influence value is given by: X represents the optical power value that changes over time; D represents the transmission and reception direction of the optical cable line; N represents the observation length of the time series; t represents the current time point; m is the lag length of the time series; a is the model coefficient; tj represents the time points before j; h is the sample number; and η represents the model error.

[0039] Based on this causal relationship model, for the optical links of the Zhejiang Provincial Broadcasting Backbone Network, network element sites are used as graph nodes, and the optical links between sites are used as graph boundaries. Weighted directed optical networks are used to measure the efficient transmission in the optical link topology, and binary undirected optical networks are used to analyze the optical link topology structure. Figure 1 As shown in the figure. Among them, the quantitative indicators such as optical power, optical signal-to-noise ratio, and causal correlation derived from the causal relationship model can be used as weight information in the weighted directed optical network.

[0040] Step 2: Based on statistical analysis methods and dynamic programming, establish a characterization system for the causal relationships and dynamic reference benchmarks and change patterns in the topology of optical links;

[0041] like Figure 2 As shown, Figure 2 These are several common stages and special events in optical links. First, unforeseen events, whether caused by natural factors (such as ground subsidence) or human factors (such as construction and maintenance, vehicle damage, etc.), can significantly impact optical power. This impact manifests as a sudden drop in optical power followed by a rapid recovery, failing to reflect the overall trend of the optical link and being unpredictable; therefore, it is not considered within the scope of this embodiment. In subsequent data preprocessing, these events can be removed as anomalies using statistical analysis methods such as moving averages and box plots.

[0042] The phase of optical power can be broadly divided into two types: the plateau phase and the loss phase. Plateau phase...

[0043] During the initial stage, the overall data is significantly affected by noise disturbances, but remains relatively stable and does not cause a significant decrease in the optical signal-to-noise ratio. Disturbances below a certain threshold are acceptable to engineering maintenance units. During the attenuation stage, optical power will significantly decay over time. This process often lasts for several months, is relatively insidious, and poses a significant threat to system transmission quality. Engineering maintenance and repairs are only initiated when indicators fall below warning values, making it relatively reactive. Meanwhile, the reference values ​​for most equipment are static and often initially set. With natural system aging and frequent engineering maintenance, the optical link quality may not be able to recover to the level of the initial construction, or it may exceed the factory-set reference value due to the presence of gain equipment. At this point, the static initial reference becomes meaningless and has no reference value. Therefore, studying the dynamic reference values ​​and their changing patterns of the optical link is crucial for system quality monitoring.

[0044] This invention, based on statistical analysis and dynamic programming methods, studies the dynamic reference benchmark and its changing patterns in optical links. For example... Figure 3 As shown, a dynamic reference baseline can reflect the system's true reference level in real time and can be divided into two cases. First, for optical power in the loss phase, a dynamic historical data baseline should be referenced. This baseline value can be dynamically obtained through training an artificial intelligence model and usually represents the system's best performance at a certain stage, essentially solving an optimization problem. Furthermore, for optical links in the loss phase, artificial intelligence technology can be used to determine their development trend and provide early warnings of potential future degradation indicators. Second, when an optical link undergoes engineering maintenance, the baseline needs to be updated in real time. This baseline should not be affected by historical data and can be automatically updated based on the on-site maintenance level.

[0045] Model training and statistical differencing methods can intuitively obtain specific event time nodes in the optical link. Figure 4 and Figure 5These are two time series data points from a typical optical link. Curve 1 represents the predicted optical power after the AI ​​model was trained, and curve 2 represents the actual optical power. It can be observed that the model's predictions have a certain lag, which is caused by uncertainties in the optical link. By calculating the error between the actual and predicted values, the time corresponding to the peak value is observed. Figure 4 In the data, times t1, t2, and t3 correspond to three local peaks, and according to the actual optical power data and maintenance work orders, significant power loss occurred at each of these times. Figure 5 The appearance of a global peak indicates a sharp increase in the error between the model prediction and the actual value, corresponding to the time point of engineering maintenance. These two time points can serve as starting points for dynamic benchmark updates.

[0046] Based on the above analysis, the dynamic programming algorithm used in this invention searches for a dynamic optical link reference. From left to right, a sliding window is used to solve for local optima one by one on the time series data, analyzing whether the sequence is in a stationary or lossy period. For example... Figure 6 As shown, the maximum benchmark during the stationary period is taken as the global optimal solution. Figure 6 (Section 3 in the middle), the reference is dynamically updated. When the optical link is in a loss period, the reference is no longer updated. Until a special event occurs, such as engineering maintenance ( Figure 5 When the prediction error peaks (in the middle), discard the previous training results, start the next iteration search, and recursively update the baseline ( Figure 6 (4 line segments)

[0047] Step 3: Integrate the short-time memory network and the moving average autoregressive model to make a linear prediction of the future trend of optical power. Compare the prediction results with the dynamic reference benchmark and the change law characterization system to achieve intelligent early warning of trend loss of high-quality optical links.

[0048] This embodiment integrates both linear and nonlinear models for optical link prediction. Figure 7 This illustrates the prediction curve of the linear model. Figure 8 The diagram illustrates the prediction curves of two nonlinear models. The linear model is relatively simple and can generally reflect the overall development trend of the optical link, unaffected by local disturbances. However, its drawbacks are also obvious: when sudden events such as link maintenance, updates, or losses occur, the model underfits and cannot match data trends in real time. The nonlinear model can realistically reflect the local characteristics of the optical link, especially during link updates. Figure 7(Right end) The model fits well. However, nonlinear models are particularly complex in terms of mathematical definition, parameter selection, and sample training, and are prone to overfitting and are sensitive to noise. Therefore, this invention combines the advantages of two prediction models. On the one hand, it leverages the advantages of long short-term memory networks in time series pattern recognition, using historical data of optical links as sample input to train the model. On the other hand, it employs a moving average autoregressive model to linearly predict the future trend of optical power.

[0049] The nonlinear local characteristics of historical data are particularly important, mainly used for searching for optical link variation patterns and dynamic benchmarks. The linear model ARIMA is an important tool in time series forecasting analysis; with confidence intervals, it can characterize the future data trends and probabilities of optical links. This embodiment compares optical power prediction simulations. Figure 9 As shown, line 1 represents the actual optical link, line 2 is the LSTM fitted data, line 3 is the ARIMA linear fit, interval 4 represents the confidence level of this linear prediction, line 5 represents the dynamic reference baseline, and region 6 represents the time window in which the baseline is located. Figure 9 The invention can continuously track the historical optical power benchmark level and automatically discover maintenance nodes to update the benchmark; at the same time, the invention can make trend predictions and, given a confidence interval, achieve intelligent early warning.

Claims

1. A warning method based on a broadcast backbone network optical link causal model, characterized in that: Includes the following steps: Step 1: Based on the causal relationship model, construct the causal relationship and topology of the provincial broadcasting backbone optical links; Step 2: Based on statistical analysis methods and dynamic programming, establish a characterization system for the causal relationships and dynamic reference benchmarks and change patterns in the topology of optical links; Step 3: Integrate the Long Short-Term Memory Network and the Moving Average Autoregressive Model to make a linear prediction of the future trend of optical power. Compare the prediction results with the dynamic reference benchmark and the change law characterization system to achieve intelligent early warning of trend loss of high-quality optical links. The formula for the causal relationship model is as follows: ； In the formula: Representing time series Time series The causal impact value, This represents the optical power value that changes over time; This refers to the transmission and reception direction of the optical cable line; Represents the observation length of the time series; Represents the current point in time; It is the lag length of the time series; These are model coefficients; express Before a certain point in time; The sample number; Indicates model error; The optical link causal relationship and topology are defined with network element sites as graph nodes and optical links between sites as graph boundaries. The efficient transmission in the optical link topology is measured by a weighted directed efficient optical network, and the optical link topology is analyzed by a binary undirected optical network. The causal correlation, optical power, and optical signal-to-noise ratio quantification indicators are used as weight information in the weighted directed optical network.

2. The early warning method based on the causal model of the optical link in the broadcasting backbone network according to claim 1, characterized in that: The establishment process of the dynamic reference benchmark and change law representation system is based on statistical analysis and dynamic programming methods. The time series data is solved one by one by using a sliding window to find the local optimum. The sequence is analyzed to determine whether it is in a stationary period or a loss period. The maximum benchmark in the stationary period is taken as the global optimum and the benchmark is updated dynamically. When the optical link is in a loss period, the benchmark is no longer updated until a special event occurs. At this time, the previous training results are discarded and the next cycle of iterative search is started, and the benchmark is updated recursively.

3. The early warning method based on the causal model of the optical link in the broadcasting backbone network according to claim 1, characterized in that: The linear prediction of future optical power trends involves using historical data from optical links as samples to train a long short-term memory network model, and then employing a moving average autoregressive model to linearly predict future optical power trends.

Images