Microwave network topology optimization method and system based on graph neural network

Abstract

The invention provides a microwave network topology optimization method and system based on a graph neural network, and the method comprises the steps: firstly obtaining a real-time operation data set of a plurality of microwave nodes in a microwave communication network, covering a link quality index, a node load state parameter and an environment interference coefficient, and then carrying out the dynamic feature extraction of the real-time operation data set, the method comprises the steps of obtaining a dynamic propagation feature set containing inter-node signal attenuation features and the like, calling a preset graph neural network model to carry out topological relation modeling on the dynamic propagation feature set, generating an association weight matrix of microwave nodes, and carrying out iterative matching calculation based on the association weight matrix and a preset topological optimization constraint condition, generating a microwave network topology optimization strategy; and finally, the microwave network topology optimization strategy is sent to a microwave network management system to trigger link reconstruction operation, and the microwave network topology can be effectively optimized.

Description

Technical Field

[0001] The present invention relates to the field of microwave communication technology, and in particular to a microwave network topology optimization method and system based on graph neural network. Background Art

[0002] In the field of microwave communication network operation and maintenance, with the continuous growth of communication demand and the expansion of network scale, the optimization of microwave network topology has become increasingly important and complex. Traditional microwave network topology optimization methods mainly rely on static network parameter analysis and empirical rule-making. These methods can achieve certain results when dealing with small-scale, simple networks, but their limitations are clearly revealed when dealing with large-scale, dynamically changing microwave communication networks.

[0003] Specifically, existing technologies often limit the collection of microwave node operational data to historical or periodically sampled data, making it difficult to reflect dynamic network changes in real time. This lag in data acquisition prevents optimization strategies from keeping pace with actual changes in network status, thus impacting optimization effectiveness. Furthermore, traditional feature extraction methods typically employ fixed feature templates or simple statistical methods, making it difficult to fully and accurately capture the complex propagation characteristics and interference relationships between microwave nodes. This incomplete feature extraction prevents optimization models from fully understanding the inherent laws of the network, further limiting the scientific nature and effectiveness of optimization strategies.

[0004] Existing technologies for modeling topological relationships often rely on traditional graph-theory-based algorithms, such as the shortest path algorithm and the minimum spanning tree algorithm. While these algorithms are effective for simple network structures, they often struggle with the complex topologies and dynamically changing node states found in microwave communication networks. These algorithms struggle to automatically learn and adapt to the nonlinear relationships and complex patterns within the network, resulting in topologies that fail to fully reflect the network's actual operational requirements and performance bottlenecks.

[0005] Finally, when it comes to developing and executing optimization strategies, existing technologies often rely on pre-set rules or heuristic algorithms. These methods may be effective in specific scenarios or with simple constraints, but they often struggle to find a global optimal solution when faced with the diverse optimization objectives and complex constraints in microwave communication networks. Furthermore, existing technologies often lack effective integration with microwave network management systems when executing optimization strategies. This prevents the optimization strategies from being translated into actual link reconstruction operations in a timely and accurate manner, thus hindering the effectiveness of optimization. Summary of the Invention

[0006] In view of the above-mentioned problems, in combination with the first aspect of the present invention, an embodiment of the present invention provides a microwave network topology optimization method based on a graph neural network, the method comprising: Acquire a real-time operation data set of a plurality of microwave nodes in a microwave communication network, wherein the real-time operation data set includes a link quality indicator, a node load state parameter, and an environmental interference coefficient between each microwave node; Performing dynamic feature extraction processing on the real-time operation data set to generate a dynamic propagation feature set for each microwave node, the dynamic propagation feature set including inter-node signal attenuation features, channel occupancy fluctuation features, and interference superposition features; Calling a preset graph neural network model to perform topological relationship modeling on the dynamic propagation feature set to generate an association weight matrix of microwave nodes; Performing iterative matching calculation based on the association weight matrix and preset topology optimization constraints to generate a microwave network topology optimization strategy; The microwave network topology optimization strategy is sent to a microwave network management system to trigger a link reconstruction operation.

[0007] On the other hand, an embodiment of the present invention also provides a microwave network topology optimization system based on graph neural network, including a processor and a machine-readable storage medium, wherein the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0008] Based on the above aspects, the embodiments of the present invention achieve dynamic adaptive optimization of microwave network topology by integrating multi-source real-time operation data, deep feature mining, intelligent modeling and optimization calculations, significantly improving the overall performance and operational efficiency of microwave communication networks. Specifically, by comprehensively collecting key information such as link quality indicators, node load status parameters and environmental interference coefficients of microwave nodes, a comprehensive data set reflecting the real-time operation status of the network is constructed. On this basis, through dynamic feature extraction and processing, the raw data is converted into a dynamic propagation feature set with clear physical meaning. It not only reveals subtle changes in signal transmission between nodes, but also captures the complex dynamics of channel occupancy and environmental interference. Furthermore, the preset graph neural network model is effectively called to perform deep topological relationship modeling on the dynamic propagation features, generating an association weight matrix that can accurately reflect the strength of the association between microwave nodes. This not only overcomes the limitations of traditional methods in processing complex network relationships, but also uses the powerful representation learning capabilities of graph neural networks to mine the deep topological structure information hidden behind the data. Based on the association weight matrix and the preset topology optimization constraints, iterative matching calculations are performed, and ultimately a scientific and reasonable microwave network topology optimization strategy is generated. This microwave network topology optimization strategy not only considers the current network status but also future development trends and potential risks, thereby achieving global optimization of the network topology. By sending the optimization strategy to the microwave network management system and triggering link reconstruction operations, the network structure can be adjusted in real time to effectively respond to dynamic changes in the network, such as node failures and link congestion, significantly improving network reliability, stability, and transmission efficiency. This achieves intelligent optimization of the microwave network topology, helping to ensure the efficient operation of the microwave communication network. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 It is a schematic diagram of the execution flow of the microwave network topology optimization method based on graph neural network provided in an embodiment of the present invention.

[0010] Figure 2 This is a schematic diagram of exemplary hardware and software components of a microwave network topology optimization system based on a graph neural network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0011] The present invention will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of a microwave network topology optimization method based on a graph neural network provided by an embodiment of the present invention. The microwave network topology optimization method based on a graph neural network is introduced in detail below.

[0012] Step S110: obtaining a real-time operation data set of a plurality of microwave nodes in a microwave communication network, wherein the real-time operation data set includes link quality indicators, node load status parameters, and environmental interference coefficients between the microwave nodes.

[0013] In this example, to implement microwave network topology optimization based on a graph neural network, we first obtain a set of real-time operational data for multiple microwave nodes in the microwave communication network. For example, we consider a medium-sized microwave communication network consisting of 20 microwave nodes, distributed across different geographic locations and communicating with each other via microwave links.

[0014] To obtain link quality indicators, specialized signal monitoring equipment is installed at each microwave node. This equipment measures parameters such as signal strength, bit error rate, and signal-to-noise ratio between nodes in real time. For example, the signal strength between nodes A and B may fluctuate at different times. Through continuous monitoring, a series of signal strength values can be obtained. Similarly, the bit error rate reflects the proportion of erroneous code elements during transmission. The signal-to-noise ratio is the ratio of signal power to noise power, which directly affects signal transmission quality. By comprehensively collecting these parameters, link quality indicators between each microwave node can be obtained.

[0015] Obtaining node load status parameters primarily focuses on the node's computing and storage resource usage. System monitoring software can be deployed on each microwave node to collect real-time information such as CPU usage, memory usage, and disk I / O. For example, when a node is processing a large number of data transfer tasks, its CPU usage may rise rapidly, and its memory usage will also increase accordingly. By monitoring these parameters, we can understand the node's load status and determine whether it is overloaded or idle.

[0016] Obtaining the environmental interference coefficient relies on environmental monitoring sensors, which can detect the impact of factors such as electromagnetic interference and weather conditions on microwave signal transmission. For example, during a thunderstorm, the ionosphere in the atmosphere changes, disrupting microwave signal propagation and increasing the environmental interference coefficient. By monitoring and analyzing these environmental factors, the interference coefficient of each node's environment can be determined.

[0017] The link quality indicators, node load status parameters, and environmental interference coefficients obtained above are integrated to form a real-time operation data set of multiple microwave nodes in the microwave communication network. The data in this real-time operation data set is dynamically changing and needs to be continuously updated and maintained to ensure the accuracy and effectiveness of subsequent topology optimization.

[0018] Step S120: performing dynamic feature extraction processing on the real-time operation data set to generate a dynamic propagation feature set of each microwave node, wherein the dynamic propagation feature set includes inter-node signal attenuation features, channel occupancy fluctuation features, and interference superposition features.

[0019] After obtaining the real-time operation data set, it is necessary to perform dynamic feature extraction processing on it to generate the dynamic propagation feature set of each microwave node. The specific extraction process is as follows: Step S121: performing time domain segmentation processing on the link quality indicator to obtain a link signal-to-noise ratio sequence in a plurality of continuous time windows.

[0020] First, the signal-to-noise ratio data in the link quality indicator needs to be segmented in the time domain. The time window length is set to an integer multiple of the communication frame period to ensure that each time window contains complete communication frame information. Assuming the communication frame period is 5ms, the time window length is set to 50ms, which is 10 times the communication frame period.

[0021] Step S1211: Divide the original link quality indicator into equal intervals according to the time window length.

[0022] Using the previously described network of 20 microwave nodes as an example, the signal-to-noise ratio data for the link quality indicators between each microwave node is divided into equally spaced time windows. Starting from the start time of data collection, a time window is created every 50ms. For example, the first time window is 0-50ms, the second is 50-100ms, and so on. Each time window contains multiple signal-to-noise ratio measurements, which reflect the signal quality of the link during that time period.

[0023] Step S1212: In each time window, the signal-to-noise ratio measurement values of the uplink and downlink between all microwave nodes are collected.

[0024] Within each defined time window, the signal-to-noise ratio (SNR) values for the uplink and downlink between all microwave nodes are collected using signal monitoring equipment installed on the nodes. For example, within the first 50ms window, the uplink SNR values between nodes A and B are 25dB, 26dB, 27dB, and so on, respectively, while the downlink SNR values are 24dB, 25dB, 26dB, and so on. For a 20-node network, a large number of link SNR measurements must be collected.

[0025] Step S1213: performing outlier elimination processing on the signal-to-noise ratio measurement value.

[0026] The collected signal-to-noise ratio measurements may contain some outliers due to equipment failure, external interference, etc. These outliers will affect subsequent data analysis and feature extraction, so they need to be eliminated. The specific steps are as follows: Step S12131: Calculate the standard deviation of the signal-to-noise ratio measurement values within the time window, and mark the measurement values that exceed the range of N times the standard deviation as abnormal data points.

[0027] Taking the uplink signal-to-noise ratio measurements between nodes A and B in the first time window as an example, first calculate the average of these measurements. Assuming the measured values are 25dB, 26dB, 27dB, 28dB, and 29dB, the average is (25+26+27+28+29) / 5 = 27dB. Next, square the difference between each measurement and the average, average these squared values, and finally take the square root of the average to obtain the standard deviation. Assuming a standard deviation of 1dB and setting N to 3, any measurement outside the range of 24dB-30dB will be marked as an outlier. If there is a measurement of 32dB, it will be marked as an outlier.

[0028] Step S12132: Use linear interpolation to replace the abnormal data points to generate smoothed link signal-noise.

[0029] Determine the valid data points preceding and following the outlier data point within the time window. Assume that the previous valid data point before the outlier data point 32dB is 29dB, and the next valid data point after it is 28dB. Calculate the SNR difference between the two valid data points: 28dB - 29dB = -1dB. Assume the time interval difference is 5ms (determined by the measurement time). The SNR change rate per unit time is -1dB / 5ms = -0.2dB / ms. If the time offset of the outlier data point relative to the previous valid data point is 2ms, the interpolated SNR estimate is 29dB + (-0.2dB / ms) × 2ms = 28.6dB. Replace the original outlier data point 32dB with 28.6dB and update the link SNR sequence.

[0030] Step S12132-1: Determine the valid data points before and after the abnormal data point within the time window.

[0031] By sorting the signal-to-noise ratio measurements within the time window, we can find the valid data points before and after the abnormal data point. These two valid data points are the basis for linear interpolation, and their accuracy directly affects the reliability of the interpolation result.

[0032] Step S12132-2: Calculate the signal-to-noise ratio difference and time interval difference between adjacent valid data points.

[0033] Based on the SNR values of the two adjacent valid data points, calculate the difference between them. Simultaneously, based on the measurement time records, determine the time interval difference between the two data points. These two differences are used to calculate the rate of change of the SNR per unit time.

[0034] Step S12132-3: Determine the signal-to-noise ratio change rate per unit time according to the ratio of the signal-to-noise ratio difference to the time interval difference.

[0035] Divide the SNR difference by the time interval difference to get the rate of change of the SNR per unit time. This rate of change reflects the trend of the SNR over time and is used for subsequent interpolation calculations.

[0036] Step S12132-4: Calculate an interpolated signal-to-noise ratio estimate based on the signal-to-noise ratio change rate per unit time and the time offset of the abnormal data point.

[0037] The time offset of an abnormal data point is the time difference between that data point and the previous valid data point. Multiplying the rate of change of the SNR per unit time by the time offset and adding the SNR value of the previous valid data point gives the interpolated SNR estimate.

[0038] Step S12132-5: Replace the original abnormal data point with the estimated signal-to-noise ratio value, and update the link signal-to-noise ratio sequence to obtain a replaced link signal-to-noise ratio sequence.

[0039] The original abnormal data point is replaced with the calculated SNR estimate, and then the entire link SNR sequence is updated. The resulting replaced link SNR sequence no longer contains abnormal data points and is more consistent with actual signal transmission conditions.

[0040] Step S12133: performing a differential operation on the smoothed link SNR sequence to extract a SNR variation trend between adjacent measurement points.

[0041] Perform a differential operation on the replaced link SNR sequence, calculating the SNR differences between adjacent measurement points. For example, for the updated sequence of 25dB, 26dB, 27dB, 28dB, and 28.6dB, the differences between adjacent measurement points are 26dB - 25dB = 1dB, 27dB - 26dB = 1dB, 28dB - 27dB = 1dB, and 28.6dB - 28dB = 0.6dB, respectively. These differences reflect the SNR trend between adjacent measurement points and help understand the dynamic changes in signal quality.

[0042] Step S1214: constructing a time domain feature vector according to the signal-to-noise ratio variation trend as input data of the link signal-to-noise ratio sequence.

[0043] The calculated SNR differences between adjacent measurement points are combined into a vector, which is the time-domain feature vector. For example, the time-domain feature vector in the above example is [1dB, 1dB, 1dB, 0.6dB]. This time-domain feature vector serves as the input data for the link SNR sequence and is used for subsequent feature extraction and analysis.

[0044] Step S1215: verify whether the smoothed signal-to-noise ratio sequence meets a preset fluctuation tolerance. If it exceeds the fluctuation tolerance, reduce the time window and re-execute the time domain segmentation process.

[0045] Set a preset fluctuation tolerance, such as ±2dB. Check the smoothed SNR sequence to determine whether its fluctuation range is within this tolerance. If the maximum difference in the SNR sequence exceeds 2dB, it is considered to have exceeded the fluctuation tolerance. In this case, it is necessary to reduce the time window length, for example, from 50ms to 25ms, and then re-execute the time domain segmentation process, including data acquisition, outlier removal, and difference calculation, until the smoothed SNR sequence meets the preset fluctuation tolerance.

[0046] Step S122: for the link signal-to-noise ratio sequence in each time window, perform the following operations.

[0047] After obtaining the link signal-to-noise ratio sequence in each time window, it needs to be further processed to extract more valuable features.

[0048] Step S1221: Calculate the mean offset of the link signal-to-noise ratio sequence, and extract the signal-to-noise ratio change gradient between adjacent time windows.

[0049] For the link SNR sequence within each time window, first calculate its mean. Taking the uplink SNR sequence between node A and node B in the first time window as an example, assuming the sequence is 25dB, 26dB, 27dB, 28dB, and 28.6dB, the mean is (25 + 26 + 27 + 28 + 28.6) / 5 = 26.92dB. Then, based on historical data or a preset reference value, calculate the offset of this mean relative to the reference value. Assuming the reference value is 27dB, the mean offset is 26.92dB - 27dB = -0.08dB.

[0050] The SNR gradient between adjacent time windows is also extracted. The mean values of the link SNR sequences within two adjacent time windows are compared and their difference is calculated. For example, if the mean value of the first time window is 26.92 dB and the mean value of the second time window is 27.1 dB, the SNR gradient is 27.1 dB - 26.92 dB = 0.18 dB. This gradient reflects the rate of change of the SNR between adjacent time windows and helps understand the dynamic trend of signal quality.

[0051] Step S1222: constructing a first dynamic propagation feature subset according to the mean offset and the signal-to-noise ratio change gradient.

[0052] The calculated mean shift and signal-to-noise ratio gradient are combined to form a first dynamic propagation feature subset. For example, in the above example, the first dynamic propagation feature subset is [-0.08dB, 0.18dB]. This first dynamic propagation feature subset reflects the variation characteristics of the link signal-to-noise ratio over time.

[0053] Step S123: performing frequency domain decomposition processing on the node load state parameter to obtain load fluctuation amplitudes and load occupancy periods in different frequency bands.

[0054] Perform frequency domain decomposition on node load parameters such as CPU usage and memory usage. Use methods like Fourier transform to convert time-domain load data into the frequency domain, obtaining load fluctuation amplitudes and load occupancy periods in different frequency bands.

[0055] For example, consider the CPU usage data for a node, assuming the measured values over a period of time are 50%, 55%, 60%, 58%, 56%, and so on. Using a Fourier transform, this time-domain data is converted to frequency-domain data, obtaining the amplitude and phase information for different frequency components. The load fluctuation amplitudes in different frequency bands reflect the degree of load variation at different frequencies, while the load duty cycle indicates how the load varies over a complete cycle. For example, in a low-frequency band, the load fluctuation amplitude is 5%, and the load duty cycle is 100ms. This means that the node's load fluctuates by 5% every 100ms in this low-frequency band.

[0056] Step S124: performing a nonlinear combination of the load fluctuation amplitude and the load occupancy period to generate a second dynamic propagation feature subset.

[0057] Perform a nonlinear combination on the load fluctuation amplitudes and load occupancy periods obtained for different frequency bands. Nonlinear functions, such as exponential functions and logarithmic functions, can be used to combine the two. For example, if the load fluctuation amplitude of a certain frequency band is 5% and the load occupancy period is 100ms, assuming an exponential function is used for the combination and a coefficient a is set to 0.5, the combined characteristic value is exp(0.5×5%×100ms). This combination is performed on the load fluctuation amplitudes and load occupancy periods for all frequency bands, and the results are combined to form a second dynamic propagation feature subset. This second dynamic propagation feature subset reflects the comprehensive characteristics of the node load in the frequency domain.

[0058] Step S125: performing spatial cluster analysis on the environmental interference coefficient to identify the distribution density of interference sources and the diffusion range of interference intensity.

[0059] Spatial cluster analysis of the environmental interference coefficient is performed to identify the distribution density of interference sources and the diffusion range of interference intensity. First, the nodes in the microwave communication network are divided into different regions based on their geographical location. Then, these regions are clustered based on the environmental interference coefficient of each node.

[0060] For example, using the K-means clustering algorithm, 20 nodes are divided into five regions. For each region, the mean and standard deviation of the environmental interference coefficient are calculated. If the mean environmental interference coefficient for nodes in a particular region is high and the standard deviation is low, this indicates that the density of interference sources in that region is high and the interference intensity is relatively stable. By analyzing each region, the distribution density of interference sources and the range of interference intensity can be determined. For example, it may be found that the interference sources in a particular region are primarily concentrated in a small area with high interference intensity, while the interference in other regions is relatively weak.

[0061] Step S126: constructing a third dynamic propagation feature subset according to the interference source distribution density and the interference intensity diffusion range.

[0062] The identified interference source distribution density and interference intensity diffusion range are combined to form a third dynamic propagation feature subset. For example, for a certain area, the interference source distribution density is 10 interference sources per square kilometer, and the interference intensity diffusion range is a radius of 5 kilometers. Then the third dynamic propagation feature subset of this area is [10 / square kilometer, 5 kilometers]. This processing is performed on the interference source distribution density and interference intensity diffusion range of all areas, and the results are combined to form a complete third dynamic propagation feature subset. This third dynamic propagation feature subset reflects the characteristics of environmental interference in the spatial dimension.

[0063] Step S127: Map the first dynamic propagation feature subset, the second dynamic propagation feature subset, and the third dynamic propagation feature subset into inter-node signal attenuation features, channel occupancy fluctuation features, and interference superposition features, respectively, and merge them into the dynamic propagation feature set.

[0064] Among them, the first dynamic propagation feature subset reflects the change characteristics of the link signal-to-noise ratio in the time dimension and can be mapped to the signal attenuation characteristics between nodes. Specifically, the mean offset in this first dynamic propagation feature subset reflects the deviation of the link signal-to-noise ratio relative to the reference value in the current time window. If the mean offset is negative, it means that the current link signal-to-noise ratio is lower than the reference level, which means that the signal may be attenuated during transmission; the signal-to-noise ratio change gradient reflects the rate of change of the signal-to-noise ratio between adjacent time windows. If the change gradient is negative and the absolute value is large, it indicates that the signal is attenuating rapidly. For example, the first dynamic propagation feature subset is [-0.08dB, 0.18dB], where -0.08dB as the mean offset reflects that the current link signal-to-noise ratio is lower than the reference value and there is a certain degree of signal attenuation; 0.18dB as the signal-to-noise ratio change gradient reflects the dynamic change trend of signal quality, which is related to the signal attenuation between nodes.

[0065] The second dynamic propagation feature subset reflects the comprehensive characteristics of node load in the frequency domain and can be mapped into channel occupancy fluctuation characteristics. This second dynamic propagation feature subset is derived through a nonlinear combination of the load fluctuation amplitude and load occupancy period in different frequency bands. The load fluctuation amplitude reflects the degree of load variation at different frequencies, and the load occupancy period represents the load variation within a complete cycle. Together, these two reflect the fluctuating impact of node load on channel occupancy. For example, the second dynamic propagation feature subset is [exp(0.5×5%×100ms), exp(0.5×3%×200ms)]. These combined feature values reflect the comprehensive situation of node load in different frequency bands, thereby reflecting the fluctuating characteristics of channel occupancy.

[0066] The third dynamic propagation feature subset reflects the spatial characteristics of environmental interference and can be mapped into interference superposition features. The interference source distribution density in this third dynamic propagation feature subset describes the spatial distribution density of interference sources, while the interference intensity diffusion range indicates the range of interference impact in space. These two factors together determine the spatial superposition effect of environmental interference. For example, the third dynamic propagation feature subset is [10 / km², 5km], indicating that there are 10 interference sources per square kilometer and the interference intensity diffusion range is a 5km radius, reflecting the superposition of environmental interference within the area.

[0067] The mapped inter-node signal attenuation characteristics, channel occupancy fluctuation characteristics, and interference superposition characteristics are merged. This can be done by concatenating the elements of the three subsets together to form a new set. For example, if the first dynamic propagation feature subset is [-0.08dB, 0.18dB], the second dynamic propagation feature subset is [exp(0.5×5%×100ms), exp(0.5×3%×200ms)], and the third dynamic propagation feature subset is [10 / km2, 5km], then the merged dynamic propagation feature set is [-0.08dB, 0.18dB, exp(0.5×5%×100ms), exp(0.5×3%×200ms), 10 / km2, 5km]. This dynamic propagation feature set contains information on inter-node signal attenuation characteristics, channel occupancy fluctuation characteristics, and interference superposition characteristics.

[0068] Step S130: calling a preset graph neural network model to perform topological relationship modeling on the dynamic propagation feature set to generate an association weight matrix of microwave nodes.

[0069] After obtaining the dynamic propagation feature set of each microwave node, the pre-set graph neural network model is used to perform topological relationship modeling to generate the microwave node's association weight matrix. The specific process is as follows: Step S131: each microwave node is regarded as a vertex in a graph structure, and the physical connection relationship between microwave nodes is regarded as an edge in the graph structure.

[0070] Each microwave node in a microwave communication network is abstracted as a vertex in a graph structure, and the physical connections between nodes are represented as edges in the graph. For example, in the 20-node network described above, if there is a physical connection between node A and node B, then in the graph structure, there is an edge connecting vertex A and vertex B. This constructs a graph that reflects the topology of the microwave network.

[0071] Step S132: Calculate the initial edge weight between any two adjacent vertices based on the inter-node signal attenuation feature in the dynamic propagation feature set.

[0072] Internode signal attenuation features are extracted from the dynamic propagation feature set. Based on these internode signal attenuation features, the initial edge weights between any two adjacent vertices are calculated. Internode signal attenuation features can be represented by signal strength attenuation values, changes in the signal-to-noise ratio, and other indicators. For example, for adjacent vertices A and B, a calculation formula is established to calculate the initial edge weights based on the signal strength attenuation value and the change in the signal-to-noise ratio between them. Assuming a signal strength attenuation value of 10dB and a change in the signal-to-noise ratio of 5dB, and setting coefficients a and b to 0.6 and 0.4, the initial edge weight is 0.6 × 10dB + 0.4 × 5dB = 8dB. This calculation is repeated for all adjacent vertices in the graph to obtain the initial edge weights for all edges.

[0073] Step S133: performing sliding window variance analysis on the channel occupancy fluctuation characteristics to determine a channel occupancy stability index.

[0074] After obtaining the dynamic propagation feature set, in order to further rationally adjust the initial edge weights, it is necessary to conduct a sliding window variance analysis on the channel occupancy fluctuation characteristics to determine the channel occupancy stability index. Specifically, the sliding window size must first be set, taking into account factors such as the characteristics of the microwave communication network and the temporal granularity of the data. Assume that, based on actual conditions, the sliding window length is set to 10 time units. For example, a microwave communication network with 20 time units of channel occupancy fluctuation data is used. These data represent the channel occupancy rates for each time unit, which are 30%, 32%, 35%, 33%, 36%, 34%, 37%, 35%, 38%, 36%, 39%, 37%, 40%, 38%, 41%, 39%, 42%, 40%, 43%, and 41%, respectively.

[0075] Starting from the beginning of the data sequence, a sliding window is applied across the channel occupancy fluctuation characteristic sequence, with each time unit as the step size. The first window covers the channel occupancy data for the first 10 time units, namely 30%, 32%, 35%, 33%, 36%, 34%, 37%, 35%, 38%, and 36%. Then, the window slides back one time unit, and the data in the second window becomes 32%, 35%, 33%, 36%, 34%, 37%, 35%, 38%, 36%, and 39%, and so on, until the window reaches the end of the sequence.

[0076] For the data within each window, calculate its average. Taking the first window as an example, add the 10 data points: 30% + 32% + 35% + 33% + 36% + 34% + 37% + 35% + 38% + 36% = 346%. Then divide this sum by the number of data points, 10, for an average of 34.6%. Use the same method to calculate the average for each window. For example, the sum of the data in the second window is 32% + 35% + 33% + 36% + 34% + 37% + 35% + 38% + 36% + 39% = 355%, resulting in an average of 35.5%.

[0077] After calculating the mean value for each window, we further calculate the variance of the data within each window. Using the first window as an example, we first calculate the difference between each data point and the mean. The difference between the first data point (30%) and the mean value (34.6%) is -4.6%. Square this difference to get (-4.6%)² = 0.2116. The difference between the second data point (32%) and the mean value is -2.6%. Square this difference to get (-2.6%)² = 0.0676. Similarly, we calculate the square of each data point's difference from the mean. We then add these squared values. Assuming the result is 1.2, we divide this sum by the number of data points (10). This gives us the variance of the channel occupancy data within that window: 1.2 / 10 = 0.12. Repeat this calculation for each window to obtain a series of variance values.

[0078] After obtaining a series of variance values, they need to be further processed to determine the channel occupancy stability index. The average of these variance values can be taken as a baseline value for the channel occupancy stability index. Assuming the calculated variance values are 0.12, 0.11, 0.13, 0.1, 0.12, 0.11, 0.13, 0.1, 0.12, and 0.11, the sum of these values is 1.15. Dividing this by the number of variance values (10) yields an average of 0.115.

[0079] To make the index more consistent with practical application requirements, the base value must be normalized. A maximum possible variance is set as the upper limit. For example, assuming the maximum possible variance is 0.2, the calculated average value of 0.115 is divided by the upper limit of 0.2, resulting in a normalized channel occupancy stability index of 0.575. The closer the index is to 0, the more stable the channel occupancy; the closer it is to 1, the more unstable the channel occupancy.

[0080] Step S134: adjusting the initial edge weights according to the channel occupancy stability index to generate an intermediate edge weight set.

[0081] Step S1341: Establish a mapping relationship table between the channel occupancy stability index and the edge weight adjustment coefficient.

[0082] Since a lower channel occupancy stability index indicates more stable channel occupancy and a higher link reliability in the network, the corresponding edge weight adjustment coefficient should be larger. Assume the following mapping relationship: when the channel occupancy stability index is between 0 and 0.2, the edge weight adjustment coefficient is 1.5; when it is between 0.2 and 0.4, the edge weight adjustment coefficient is 1.2; when it is between 0.4 and 0.6, the edge weight adjustment coefficient is 0.8; when it is between 0.6 and 0.8, the edge weight adjustment coefficient is 0.5; and when it is between 0.8 and 1, the edge weight adjustment coefficient is 0.2.

[0083] Step S1342: query the mapping relationship table to obtain the adjustment coefficient that matches the channel occupancy stability index corresponding to each edge.

[0084] For each edge in the microwave network, the corresponding adjustment coefficient is searched in the mapping table based on its corresponding channel occupancy stability index. For example, if the channel occupancy stability index of an edge is 0.575, the corresponding edge weight adjustment coefficient is 0.8.

[0085] Step S1343: performing a product operation on the initial edge weight and the adjustment coefficient to obtain a preliminarily adjusted edge weight.

[0086] Assume that the initial edge weight of an edge is 8dB and its corresponding adjustment coefficient is 0.8. Multiply the two together, that is, 8dB×0.8=6.4dB, and the initial adjusted edge weight of this edge is 6.4dB.

[0087] Step S1344: Detect whether the edge weight after preliminary adjustment exceeds a preset edge weight upper limit threshold; if so, set the edge weight to the upper limit threshold.

[0088] To ensure that edge weights remain within a reasonable range and prevent excessive edge weights from negatively impacting subsequent topology modeling and optimization, a preset upper threshold for edge weights is required. Assuming the preset upper threshold for edge weights is 7dB, no special processing is required for the initially adjusted edge weight of 6.4dB, as it does not exceed 7dB. If the initially adjusted edge weight of an edge is greater than 7dB, for example, 7.5dB, the edge weight for that edge is set to 7dB.

[0089] Step S1345: Normalize the edge weights that do not exceed the upper threshold so that they are distributed in the interval [0, 1].

[0090] Linear normalization is used to process edge weights that do not exceed the upper threshold. Assuming that the edge weight range is 0-10dB, for an edge weight of 6.4dB, the normalization calculation process is (6.4dB-0dB) / (10dB-0dB)=0.64, that is, the edge weight of this edge is normalized to 0.64, making it distributed in the interval [0, 1].

[0091] Step S1346: performing weighted summation on the normalized edge weights and the remaining bandwidth capacity of the microwave node to generate the intermediate edge weight set.

[0092] Obtain the remaining bandwidth capacity of the microwave node and normalize it. Assume that a node's maximum bandwidth capacity is 100 Mbps and its current remaining bandwidth capacity is 50 Mbps. The normalized remaining bandwidth capacity is 50 Mbps / 100 Mbps = 0.5. Set weighting coefficients, for example, 0.6 for the normalized edge weight and 0.4 for the remaining bandwidth capacity. For the edge with a normalized edge weight of 0.64, its final intermediate edge weight is 0.6 × 0.64 + 0.4 × 0.5 = 0.584. Perform this calculation for all edges in the microwave network and combine the results to generate a set of intermediate edge weights.

[0093] Step S135: extracting the cross-link interference coupling coefficient from the interference superposition feature, and dynamically modifying the intermediate edge weight set based on the cross-link interference coupling coefficient.

[0094] The cross-link interference coupling coefficient, part of the interference superposition feature, is extracted from the dynamic propagation feature set. This coefficient reflects the degree of mutual interference between different links. Its calculation requires comprehensive consideration of multiple factors, including the distribution of interference sources and the spread of interference intensity. For example, through a detailed analysis of factors such as the density of interference sources and the range of interference intensity, the cross-link interference coupling coefficient between two adjacent links is determined to be 0.3.

[0095] For each edge in the intermediate edge weight set, dynamic correction is performed based on the cross-link interference coupling coefficient between its link and other links. For example, suppose the intermediate edge weight of an edge is 0.584, and the cross-link interference coupling coefficient between its link and its adjacent links is 0.3. A correction coefficient of 0.2 is set. The corrected edge weight is then 0.584 - 0.2 × 0.3 = 0.524. This correction is performed on all edges in the intermediate edge weight set to obtain the corrected intermediate edge weight set.

[0096] Step S136: Input the corrected intermediate edge weight set into the graph neural network model, and generate implicit association features between vertices through multi-layer feature aggregation operations.

[0097] The modified set of intermediate edge weights is input into a pre-set graph neural network model. A graph neural network model typically contains multiple layers, each of which performs feature aggregation operations.

[0098] In the first layer, for each vertex, edge weight information for its adjacent vertices is collected. For example, in a network containing multiple microwave nodes, vertex A has adjacent vertices B and C. The corrected median edge weight of edge AB is 0.524, and the corrected median edge weight of edge AC is 0.6. This edge weight information is concatenated with the features of vertex A to generate a new feature vector. Suppose vertex A's feature vector is [0.2, 0.3], and the concatenated edge weight information of edges AB and AC is [0.524, 0.6]. The concatenated feature vector is [0.2, 0.3, 0.524, 0.6].

[0099] Then, a nonlinear transformation is performed on the feature vector, such as using the ReLU activation function. The ReLU function outputs the value if the input value is greater than 0, and outputs 0 if the input value is less than or equal to 0. Assume that after the concatenated feature vector is processed by the ReLU activation function, the new feature vector obtained is [0.2, 0.3, 0.524, 0.6] (because these values are all greater than 0).

[0100] This new feature vector is passed to the next layer. In the second layer, the process of collecting adjacent vertex information, concatenating, and performing nonlinear transformations is repeated. After multiple layers of this operation, each vertex receives a new feature vector, which reflects the implicit association characteristics between vertices. For example, after three layers of this operation, the implicit association feature vector for vertex A is [0.4, 0.5, 0.6, 0.7].

[0101] Step S137: performing weighted fusion based on the implicit association features and the physical distance parameters between vertices to generate the association weight matrix.

[0102] In this embodiment, the physical distance parameters between vertices are first obtained. The physical distance between each vertex (microwave node) can be obtained through measurement or a geographic information system. For example, the physical distance between vertex A and vertex B is 10 kilometers. After normalization, assuming the maximum possible physical distance is 100 kilometers, the normalized physical distance is 10 kilometers / 100 kilometers = 0.1.

[0103] Furthermore, weighting coefficients are set, for example, the weighting coefficient of the implicit association feature is 0.7, and the weighting coefficient of the physical distance parameter is 0.3. For any two vertices, their implicit association feature vectors are subjected to a certain metric calculation, such as calculating their cosine similarity. Assume that the implicit association feature vector of vertex A is [0.4, 0.5, 0.6, 0.7], and the implicit association feature vector of vertex B is [0.3, 0.4, 0.5, 0.6], and calculate their cosine similarity. First, calculate the dot product of the two vectors, that is, 0.4×0.3+0.5×0.4+0.6×0.5+0.7×0.6=0.92. Then calculate the modulus of the two vectors. The modulus of vector A is calculated to be approximately 1.02, and the modulus of vector B is approximately 0.91. The cosine similarity is 0.92 / (1.02×0.91)≈0.99.

[0104] The cosine similarity and the normalized physical distance parameter are weighted and combined to obtain the association weight between the two vertices: 0.7 × 0.99 + 0.3 × 0.1 = 0.723. This calculation is repeated for all vertex pairs, and the resulting association weights are arranged in vertex order to form an association weight matrix.

[0105] Step S140: performing iterative matching calculation based on the association weight matrix and preset topology optimization constraints to generate a microwave network topology optimization strategy.

[0106] Step S141: extracting the out-degree weight and in-degree weight of each vertex from the association weight matrix, and calculating the comprehensive load balancing degree of the vertex.

[0107] The association weight matrix records the association relationships and corresponding weights between each vertex (i.e., microwave node) in the microwave network. Taking a microwave network containing five microwave nodes (labeled as node 1, node 2, node 3, node 4, and node 5) as an example, the association weight matrix is as follows (the actual values are calculated according to the previous steps): Node 1 Node 2 Node 3 Node 4 Node 5 Node 1 0 0.3 0.2 0.1 0.4 Node 2 0.2 0 0.3 0.2 0.1 Node 3 0.1 0.2 0 0.3 0.4 Node 4 0.3 0.1 0.2 0 0.3 Node 5 0.4 0.2 0.1 0.3 0 For node 1, its out-degree weight is the sum of the weights of the edges pointing from node 1 to other nodes, that is, 0.3+0.2+0.1+0.4=1; the in-degree weight is the sum of the weights of the edges pointing from other nodes to node 1, that is, 0.2+0.1+0.3+0.4=1.

[0108] The comprehensive load balance can be measured by the absolute value of the difference between the out-degree weight and the in-degree weight. For node 1, its comprehensive load balance is |1-1|=0. Using the same method, the out-degree weight of node 2 is calculated to be 0.2+0.3+0.2+0.1=0.8, the in-degree weight is 0.3+0.2+0.1+0.2=0.8, and the comprehensive load balance is |0.8-0.8|=0; the out-degree weight of node 3 is 0.1+0.2+0.3+0.4=1, the in-degree weight is 0.2+0.3+0.2+0.1=0.8, and the comprehensive load balance is |1- 0.8|=0.2; the out-degree weight of node 4 is 0.3+0.1+0.2+0.3=0.9, the in-degree weight is 0.1+0.2+0.3+0.3=0.9, and the overall load balancing degree is |0.9-0.9|=0; the out-degree weight of node 5 is 0.4+0.2+0.1+0.3=1, the in-degree weight is 0.4+0.2+0.1+0.3=1, and the overall load balancing degree is |1-1|=0.

[0109] Step S142: determining a set of potential congested nodes in the microwave network according to the comprehensive load balancing degree.

[0110] In this embodiment, a comprehensive load balancing threshold, such as 1.5, can be set. If the comprehensive load balancing degree of a vertex exceeds this threshold, the vertex is marked as a potential congested node. In the above example, the comprehensive load balancing degree of all nodes does not exceed 1.5, so there are currently no potential congested nodes in the network. If the comprehensive load balancing degree of a node is 1.6, the node belongs to the set of potential congested nodes.

[0111] Step S143: Based on the maximum hop limit in the topology optimization constraint, a set of candidate paths from the source node to the target node is generated.

[0112] Assume that the maximum hop count is 3, the source node is S, and the destination node is T. Starting from the source node S, perform a breadth-first search to find all paths with a hop count of no more than 3. For example, possible paths include SABT, SCDT, etc. These paths constitute the candidate path set.

[0113] Step S144: performing the following operations on each path in the candidate path set.

[0114] Step S1441: Count the sum of the weights of all edges in the path and calculate the interference superposition accumulation on the path.

[0115] Taking the path SABT as an example, assuming that the weight of edge SA is 0.8, the weight of edge AB is 0.7, and the weight of edge BT is 0.6, then the total weight of this path is 0.8+0.7+0.6=2.1.

[0116] Calculate the cumulative interference along the path. Based on the nodes and edges the path passes through and the previously obtained interference superposition characteristics, calculate the cumulative interference. Assume that the interference intensities in the areas traversed by the SABT path are 0.2, 0.3, and 0.4, respectively. Adding these intensities yields the cumulative interference superposition of 0.2 + 0.3 + 0.4 = 0.9.

[0117] Step S1442: constructing a path optimization score according to the weight sum and the interference superposition accumulation.

[0118] Step S14421: assign a dynamic priority to each edge in the path, where the dynamic priority is determined based on the historical number of failures of the edge and the real-time load rate.

[0119] When assigning dynamic priorities to each edge in a path, two key factors must be considered: the edge's historical failure count and real-time load ratio. The historical failure count reflects the edge's reliability during past operations, while the real-time load ratio reflects the edge's current workload.

[0120] Taking path SABT as an example, assume that edge SA has a historical fault count of 2 and a real-time load rate of 30%; edge AB has a historical fault count of 1 and a real-time load rate of 20%; and edge BT has a historical fault count of 0 and a real-time load rate of 10%.

[0121] To determine dynamic priority, a reasonable calculation method is required. A linear combination approach can be used to assign different weights to the number of historical failures and the real-time load rate. For example, assume the weight of the number of historical failures is 0.5, and the weight of the real-time load rate is 0.5.

[0122] For the SA edge, first calculate its corresponding priority component based on the number of historical failures. This number can be normalized. Assuming the maximum number of historical failures is 5, the normalized value for the SA edge's historical failure count is 2 / 5 = 0.4. The real-time load rate is inherently in the [0, 1] range and does not require normalization, resulting in a value of 0.3. Based on the set weights, the SA edge's dynamic priority is 0.5 × 0.4 + 0.5 × 0.3 = 0.35.

[0123] For edge AB, the normalized value of its historical fault count is 1 / 5=0.2, and the real-time load rate is 0.2. Therefore, the dynamic priority of edge AB is 0.5×0.2+0.5×0.2=0.2.

[0124] For edge BT, the normalized value of its historical fault count is 0 / 5=0, and the real-time load rate is 0.1. Then the dynamic priority of edge BT is 0.5×0+0.5×0.1=0.05.

[0125] Step S14422: performing weighted correction on the sum of the path weights based on the dynamic priority to obtain a corrected path weight.

[0126] After obtaining the dynamic priority of each edge, we can perform a weighted correction on the path weight sum. The weight sum of the path SABT has been calculated to be 2.1.

[0127] For edge SA, its weight is 0.8 and its dynamic priority is 0.35; the weight of edge AB is 0.7 and its dynamic priority is 0.2; the weight of edge BT is 0.6 and its dynamic priority is 0.05.

[0128] The weighted correction process is to multiply the weight of each edge by its dynamic priority and then add these products. That is, the corrected path weight is 0.8×0.35+0.7×0.2+0.6×0.05=0.28+0.14+0.03=0.45.

[0129] Step S14423: Calculate the average value of the comprehensive load balancing degree of all nodes on the path and use it as the load balancing score.

[0130] First, we need to obtain the comprehensive load balance of each node on the path. Assume that the comprehensive load balance of node S is 0.5, the comprehensive load balance of node A is 0.3, the comprehensive load balance of node B is 0.2, and the comprehensive load balance of node T is 0.4.

[0131] Calculate the average of these comprehensive load balancing degrees, which is (0.5 + 0.3 + 0.2 + 0.4) / 4 = 1.4 / 4 = 0.35. This average is used as the load balancing score of the path SABT.

[0132] Step S14424: input the corrected path weight, the interference superposition accumulation and the load balancing score into a scoring function to generate a path optimization score.

[0133] The scoring function is defined as: path optimization score = α × modified path weight + β × interference superposition accumulation + γ × load balancing score, where α, β, and γ are preset weight coefficients and satisfy α+β+γ=1.

[0134] Assuming α = 0.6, β = 0.3, and γ = 0.1, the modified path weight of path SABT is 0.45, the interference superposition accumulation is 0.9, and the load balancing score is 0.35.

[0135] Substituting these values into the scoring function, the path optimization score = 0.6 × 0.45 + 0.3 × 0.9 + 0.1 × 0.35 = 0.27 + 0.27 + 0.035 = 0.575.

[0136] Step S1443: Select the candidate path with the highest path optimization score as the target optimization path.

[0137] Calculate the path optimization score for each path in the candidate path set using the above steps. Assume that in addition to the SABT path, there is also the SCDT path, and its path optimization score is calculated to be 0.6. Comparing the scores of the two paths, 0.6 > 0.575, so the SCDT path is selected as the target optimization path.

[0138] Step S1444: updating the edge weight distribution in the association weight matrix based on the target optimization path, and recalculating the comprehensive load balancing degree.

[0139] For the target optimized path SCDT, the weights of each edge on the path are adjusted based on the path's usage and the network's real-time status. For example, if the original weight of edge SC is 0.7 and the selected path may increase its load, its weight is adjusted to 0.8; the original weight of edge CD is 0.6 and is adjusted to 0.7; and the original weight of edge DT is 0.5 and is adjusted to 0.6.

[0140] The weights of other edges in the association weight matrix that are not related to the target optimization path remain unchanged for the time being. The updated association weight matrix reflects the dynamic changes of the network topology.

[0141] Based on the updated association weight matrix, the out-degree weight and in-degree weight of each vertex are recalculated, and the overall load balance is recalculated. For example, for node C related to the target optimization path, its out-degree weight and in-degree weight will change due to changes in the weights of edges SC and CD, and its overall load balance will be recalculated.

[0142] Step S1445: When the comprehensive load balancing degree meets a preset convergence threshold, a microwave network topology optimization strategy including a target optimization path is generated.

[0143] Set a preset convergence threshold, such as 0.2. After each update of the associated weight matrix and recalculation of the overall load balance, check the change in the overall load balance of all vertices. If the change in the overall load balance of all vertices is within 0.2, it indicates that the network load state has stabilized and the convergence condition has been met.

[0144] When convergence conditions are met, the generated microwave network topology optimization strategy includes the target optimization path SCDT and related edge weight adjustment information. This information will be used to guide subsequent link reconstruction operations to optimize the microwave network topology and improve network performance and reliability.

[0145] Step S150: Send the microwave network topology optimization strategy to a microwave network management system to trigger a link reconstruction operation.

[0146] Step S151: parsing the microwave network topology optimization target optimization path to determine microwave node port configuration parameters that need to be adjusted.

[0147] For the target optimized path SCDT, it is necessary to analyze the port connection status of each microwave node on the path in detail to determine the port configuration parameters that need to be adjusted.

[0148] Taking node S as an example, through analysis of network topology and signal transmission requirements, it was found that in order to better adapt to the target optimization path, its port frequency needed to be adjusted from 2.4 GHz to 2.5 GHz, the transmit power needed to be calibrated from 10 dBm to 12 dBm, and the antenna pointing angle needed to be corrected from 30° to 35°.

[0149] For node C, based on path planning and network performance requirements, it may be necessary to adjust its port frequency from 2.5 GHz to 2.6 GHz, calibrate the transmit power from 12 dBm to 13 dBm, and correct the antenna pointing angle from 40° to 45°.

[0150] Similarly, perform similar analysis and parameter determination for nodes D and T. Assume that node D needs to adjust its port frequency from 2.6 GHz to 2.7 GHz, calibrate its transmit power from 13 dBm to 14 dBm, and correct its antenna pointing angle from 45° to 50°. Node T needs to adjust its port frequency from 2.7 GHz to 2.8 GHz, calibrate its transmit power from 14 dBm to 15 dBm, and correct its antenna pointing angle from 50° to 55°.

[0151] Step S152: Generate a link switching instruction set according to the port configuration parameters, wherein the link switching instruction set includes a port frequency adjustment value, a transmit power calibration value, and an antenna pointing angle correction value.

[0152] According to the determined port configuration parameters, a corresponding link switching instruction is generated for each node.

[0153] For node S, the port frequency adjustment value is +0.1 GHz (from 2.4 GHz to 2.5 GHz), the transmit power calibration value is +2 dBm (from 10 dBm to 12 dBm), and the antenna pointing angle correction value is +5° (from 30° to 35°). The generated link switching instruction is: "Increase the port frequency by 0.1 GHz, the transmit power by 2 dBm, and the antenna pointing angle by 5°."

[0154] For node C, the port frequency adjustment value is +0.1 GHz (from 2.5 GHz to 2.6 GHz), the transmit power calibration value is +1 dBm (from 12 dBm to 13 dBm), and the antenna pointing angle correction value is +5° (from 40° to 45°). The generated link switching instruction is: "Increase the port frequency by 0.1 GHz, the transmit power by 1 dBm, and the antenna pointing angle by 5°."

[0155] For nodes D and T, link switching instructions are generated in the same way. The link switching instruction for node D is: "Increase the port frequency by 0.1 GHz, increase the transmit power by 1 dBm, and increase the antenna pointing angle by 5°." The link switching instruction for node T is: "Increase the port frequency by 0.1 GHz, increase the transmit power by 1 dBm, and increase the antenna pointing angle by 5°."

[0156] These instructions are combined together to form a link switching instruction set.

[0157] Step S153: performing conflict detection on the link switching instruction set to identify conflicting node pairs that simultaneously trigger configuration changes.

[0158] Step S1531: construct a geographical coordinate topology map of microwave nodes and calculate the Euclidean distance between any two microwave nodes.

[0159] Obtain the geographic coordinates of each microwave node using a geographic information system or surveying equipment. Assume that the coordinates of node S are (x1, y1), the coordinates of node C are (x2, y2), the coordinates of node D are (x3, y3), and the coordinates of node T are (x4, y4).

[0160] According to the Euclidean distance formula, the Euclidean distance between nodes S and C is calculated as follows: first calculate the sum of the squared difference in the horizontal coordinates and the squared difference in the vertical coordinates, that is, (x² - x¹)² + (y² - y¹)², and then take the square root of this sum to get the Euclidean distance. Suppose the calculated Euclidean distance between nodes S and C is 5 kilometers.

[0161] Using the same method, the Euclidean distance between any two nodes is calculated to construct the geographic coordinate topology of the microwave nodes.

[0162] Step S1532: If two microwave nodes receive the frequency adjustment instruction at the same time and the Euclidean distance is less than a preset interference radius threshold, they are marked as a frequency conflict node pair.

[0163] Assume that the preset interference radius threshold is 3 kilometers. In the link switching instruction set, if nodes S and C receive frequency adjustment instructions at the same time and the Euclidean distance between them is 5 kilometers, which is greater than the interference radius threshold, then they are not a frequency conflict node pair.

[0164] If node M and node N receive frequency adjustment instructions at the same time, and the Euclidean distance between them is 2 kilometers, which is less than the interference radius threshold, node M and node N are marked as a frequency conflict node pair.

[0165] Step S1533: If two microwave nodes receive the transmit power adjustment instruction at the same time and belong to the same communication frequency band, they are marked as a power conflict node pair.

[0166] Check the link switching instruction set. If node P and node Q receive the transmit power adjustment instruction at the same time and they belong to the same communication frequency band, for example, both are in the 2.4GHz-2.5GHz frequency band, then node P and node Q are marked as a power conflict node pair.

[0167] Step S1534: If the overlapping area of the main lobe directions of the two microwave nodes exceeds a preset ratio after the antenna pointing angles of the two microwave nodes are adjusted, they are marked as a beam conflict node pair.

[0168] By simulating and calculating the main lobe direction of the beams after adjusting the antenna pointing angle, the overlap area of the main lobe directions of the two microwave nodes is determined. Assume that the preset ratio is 30%. If the main lobe overlap area of nodes R and S exceeds the preset ratio after adjusting the antenna pointing angle, nodes R and S are marked as a beam conflicting node pair.

[0169] Step S1535: Merge the frequency conflict node pairs, power conflict node pairs, and beam conflict node pairs into a global conflict node set.

[0170] The marked frequency conflict node pairs, power conflict node pairs, and beam conflict node pairs are aggregated to form a global conflict node set. For example, if the frequency conflict node pairs are (M, N), the power conflict node pairs are (P, Q), and the beam conflict node pairs are (R, S), then the global conflict node set is {(M, N), (P, Q), (R, S)}.

[0171] Step S1536: Assign a conflict resolution priority to each conflicting node pair according to the conflict type, with frequency conflict having the highest priority, power conflict having the second highest priority, and beam conflict having the lowest priority.

[0172] For each conflicting node pair in the global conflicting node set, a priority is assigned based on the conflict type. The frequency conflicting node pair (M, N) is assigned the highest priority; the power conflicting node pair (P, Q) is assigned the second highest priority; and the beam conflicting node pair (R, S) is assigned the lowest priority.

[0173] Step S154: reallocating the link switching timing based on the physical location coordinates of the conflicting node pairs, and generating a phased execution instruction queue.

[0174] Replan the link switching sequence based on the physical location coordinates and conflict priority of the conflicting node pairs. The frequency-conflicting node pair, which has the highest priority, should be handled first. You can first perform a link switch on one node in the frequency-conflicting node pair, wait for a period of time to ensure signal stability, and then switch to the other node.

[0175] For power conflicting node pairs and beam conflicting node pairs, the switching order is arranged in order of priority. For example, the switching of frequency conflicting node pairs is processed first, followed by the switching of power conflicting node pairs, and finally the switching of beam conflicting node pairs.

[0176] Arrange the link switching instructions for each node according to the reallocated time sequence to generate a phased execution queue. For example, execute the link switching instruction for node M first, wait 5 minutes, then execute the link switching instruction for node N; then execute the link switching instruction for node P, wait 3 minutes, then execute the link switching instruction for node Q; and finally execute the link switching instructions for nodes R and S.

[0177] Step S155: Send the phased execution instruction queue to the corresponding microwave node controller, and perform port frequency adjustment, transmit power calibration and antenna pointing angle correction operations in sequence to achieve link reconstruction operation.

[0178] The generated phased execution command queue is sent to the controller corresponding to each microwave node. The controller performs port frequency adjustment, transmit power calibration, and antenna pointing angle correction operations in sequence according to the command queue.

[0179] For example, the controller of node M first adjusts the port frequency to the specified value according to the instructions, waits for a period of time to ensure the frequency is stable, then calibrates the transmit power and finally corrects the antenna pointing angle. In this way, the corresponding operations are performed on each node in the queue in sequence to achieve link reconstruction.

[0180] Step S156: monitoring the network performance index after link reconstruction in real time. If the network performance index does not reach the expected threshold, re-triggering the dynamic feature extraction process and modeling process.

[0181] After the link reconstruction operation is completed, the network performance indicators are monitored in real time, such as link quality indicators (signal-to-noise ratio, bit error rate, etc.), node load status parameters (CPU usage, memory usage, etc.), environmental interference coefficient, etc.

[0182] Assume the expected signal-to-noise ratio threshold is 30dB and the bit error rate threshold is 1%. Real-time monitoring reveals that the signal-to-noise ratio of a link in the reconfigured network is 28dB, lower than the expected threshold, or the bit error rate is 1.2%, higher than the expected threshold, indicating that network performance does not meet expectations.

[0183] At this point, the dynamic feature extraction process is retriggered, recollecting and processing the real-time operational data set of multiple microwave nodes in the microwave communication network to generate a new set of dynamic propagation features. A pre-configured graph neural network model is then used to perform topological modeling on this new set of dynamic propagation features, regenerating the microwave node association weight matrix. Iterative matching calculations are then performed again to generate a new microwave network topology optimization strategy to further optimize the network topology and improve network performance.

[0184] Figure 2 A schematic diagram illustrating exemplary hardware and software components of a graph neural network-based microwave network topology optimization system 100 that can implement the concepts of the present application, as provided in some embodiments of the present application. For example, the processor 120 can be used in the graph neural network-based microwave network topology optimization system 100 to perform the functions described in the present application.

[0185] The microwave network topology optimization system 100 based on a graph neural network can be a general-purpose server or a special-purpose server, both of which can be used to implement the microwave network topology optimization method based on a graph neural network of the present application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0186] For example, the microwave network topology optimization system 100 based on graph neural network may include a network port 110 connected to the network, one or more processors 120 for executing program instructions, a communication bus 130, and storage media 140 in different forms, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the microwave network topology optimization system 100 based on graph neural network may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The method of the present application may be implemented according to these program instructions. The microwave network topology optimization system 100 based on graph neural network also includes an I / O interface 150 between the computer and other input and output devices.

[0187] For ease of explanation, only one processor is described in the microwave network topology optimization system 100 based on graph neural networks. However, it should be noted that the microwave network topology optimization system 100 based on graph neural networks in this application may also include multiple processors, so the steps performed by one processor described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the microwave network topology optimization system 100 based on graph neural networks performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually in one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor perform steps A and B together.

[0188] In addition, an embodiment of the present invention also provides a readable storage medium, in which computer-executable instructions are preset. When a processor executes the computer-executable instructions, the microwave network topology optimization method based on graph neural network as described above is implemented.

[0189] It should be noted that in order to simplify the description of the present invention and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present invention, multiple features are sometimes combined into one embodiment, figure or description thereof.

Claims

1. A microwave network topology optimization method based on graph neural network, characterized in that: The method comprises: Acquire a real-time operation data set of a plurality of microwave nodes in a microwave communication network, wherein the real-time operation data set includes a link quality indicator, a node load state parameter, and an environmental interference coefficient between each microwave node; Performing dynamic feature extraction processing on the real-time operation data set to generate a dynamic propagation feature set for each microwave node, the dynamic propagation feature set including inter-node signal attenuation features, channel occupancy fluctuation features, and interference superposition features; Calling a preset graph neural network model to perform topological relationship modeling on the dynamic propagation feature set to generate an association weight matrix of microwave nodes; Performing iterative matching calculation based on the association weight matrix and preset topology optimization constraints to generate a microwave network topology optimization strategy; The microwave network topology optimization strategy is sent to a microwave network management system to trigger a link reconstruction operation.

2. The microwave network topology optimization method based on graph neural network according to claim 1, characterized in that: The performing dynamic feature extraction processing on the real-time operation data set to generate a dynamic propagation feature set for each microwave node includes: Performing time domain segmentation processing on the link quality indicator to obtain a link signal-to-noise ratio sequence in a plurality of continuous time windows; For the link SNR sequence within each time window, perform the following operations: Calculating the mean offset of the link signal-to-noise ratio sequence and extracting the signal-to-noise ratio change gradient between adjacent time windows; Constructing a first dynamic propagation feature subset according to the mean offset and the signal-to-noise ratio change gradient; Performing frequency domain decomposition processing on the node load state parameters to obtain load fluctuation amplitudes and load occupancy periods in different frequency bands; performing a nonlinear combination of the load fluctuation amplitude and the load occupancy period to generate a second dynamic propagation feature subset; Performing spatial cluster analysis on the environmental interference coefficient to identify the distribution density of interference sources and the diffusion range of interference intensity; Constructing a third dynamic propagation feature subset according to the interference source distribution density and the interference intensity diffusion range; The first dynamic propagation feature subset, the second dynamic propagation feature subset and the third dynamic propagation feature subset are respectively mapped to inter-node signal attenuation features, channel occupancy fluctuation features and interference superposition features, and merged into the dynamic propagation feature set.

3. The microwave network topology optimization method based on graph neural network according to claim 2, characterized in that: The performing time domain segmentation processing on the link quality indicator to obtain a link signal-to-noise ratio sequence in a plurality of continuous time windows includes: Setting the time window length to an integer multiple of the communication frame period, and dividing the original link quality indicator into equal intervals according to the time window length; In each time window, the uplink and downlink signal-to-noise ratio measurements between all microwave nodes are collected; Performing outlier elimination processing on the signal-to-noise ratio measurement value specifically includes: Calculate the standard deviation of the signal-to-noise ratio measurements within the time window and mark the measurements that exceed N times the standard deviation as abnormal data points; Using linear interpolation to replace the abnormal data points to generate smoothed link signal-noise; performing a differential operation on the smoothed link signal-to-noise ratio sequence to extract a signal-to-noise ratio variation trend between adjacent measurement points; Constructing a time domain feature vector according to the signal-to-noise ratio variation trend as input data of the link signal-to-noise ratio sequence; The step of replacing the abnormal data points with a linear interpolation method to generate a smoothed link signal-noise comprises: Determine the valid data points before and after the abnormal data point within the time window; Calculate the signal-to-noise ratio difference and time interval difference of adjacent valid data points before and after; Determine the signal-to-noise ratio change rate per unit time according to the ratio of the signal-to-noise ratio difference to the time interval difference; Calculating an interpolated signal-to-noise ratio estimate based on the signal-to-noise ratio change rate per unit time and the time offset of the abnormal data point; Replacing the original abnormal data point with the estimated signal-to-noise ratio value, and updating the link signal-to-noise ratio sequence to obtain a replaced link signal-to-noise ratio sequence; The replaced link SNR sequence is processed with cubic spline interpolation to obtain the smoothed SNR sequence: Verify whether the smoothed signal-to-noise ratio sequence meets the preset fluctuation tolerance. If it exceeds the fluctuation tolerance, reduce the time window and re-execute the time domain segmentation process.

4. The microwave network topology optimization method based on graph neural network according to claim 1, characterized in that: The calling of a preset graph neural network model to perform topological relationship modeling processing on the dynamic propagation feature set to generate an association weight matrix of microwave nodes includes: Each microwave node is regarded as a vertex in the graph structure, and the physical connection relationship between microwave nodes is regarded as an edge in the graph structure; Calculating the initial edge weight between any two adjacent vertices based on the inter-node signal attenuation feature in the dynamic propagation feature set; Performing a sliding window variance analysis on the channel occupancy fluctuation characteristics to determine a channel occupancy stability index; Adjusting the initial edge weights according to the channel occupancy stability index to generate an intermediate edge weight set; Extracting a cross-link interference coupling coefficient from the interference superposition feature, and dynamically modifying the intermediate edge weight set based on the cross-link interference coupling coefficient; Inputting the modified set of intermediate edge weights into the graph neural network model, and generating implicit correlation features between vertices through multi-layer feature aggregation operations; The association weight matrix is generated by performing weighted fusion based on the implicit association features and the physical distance parameters between vertices.

5. The microwave network topology optimization method based on graph neural network according to claim 4 is characterized in that: The step of adjusting the initial edge weights according to the channel occupancy stability index to generate an intermediate edge weight set includes: Establish a mapping relationship table between the channel occupancy stability index and the edge weight adjustment coefficient, where the lower the channel occupancy stability index, the larger the corresponding edge weight adjustment coefficient; Query the mapping relationship table to obtain the adjustment coefficient matched by the channel occupancy stability index corresponding to each edge; Performing a product operation on the initial edge weight and the adjustment coefficient to obtain a preliminarily adjusted edge weight; Detecting whether the edge weight after the initial adjustment exceeds a preset edge weight upper threshold, and if so, setting the edge weight to the upper threshold; Normalize the edge weights that do not exceed the upper threshold so that they are distributed in the interval [0, 1]; The normalized edge weights and the remaining bandwidth capacity of the microwave node are weighted and summed to generate the intermediate edge weight set.

6. The microwave network topology optimization method based on graph neural network according to claim 1, characterized in that: The iterative matching calculation based on the association weight matrix and the preset topology optimization constraint conditions is performed to generate a microwave network topology optimization strategy, including: Extracting the out-degree weight and in-degree weight of each vertex from the association weight matrix and calculating the comprehensive load balancing degree of the vertex; determining a set of potential congested nodes in the microwave network according to the comprehensive load balancing degree; generating a set of candidate paths from a source node to a target node based on a maximum hop count limit in the topology optimization constraint; For each path in the candidate path set, perform the following operations: Count the sum of the weights of all edges in the path and calculate the interference superposition accumulation on the path; Constructing a path optimization score according to the weight sum and the interference superposition accumulation; Select the candidate path with the highest path optimization score as the target optimization path; updating the edge weight distribution in the association weight matrix based on the target optimization path, and recalculating the comprehensive load balancing degree; When the comprehensive load balancing degree meets a preset convergence threshold, a microwave network topology optimization strategy including a target optimization path is generated.

7. The microwave network topology optimization method based on graph neural network according to claim 6, characterized in that: The constructing a path optimization score according to the weight sum and the interference superposition accumulation includes: Assign a dynamic priority to each edge in the path, where the dynamic priority is determined based on the edge's historical failure count and real-time load rate. Performing weighted correction on the sum of the path weights based on the dynamic priority to obtain a corrected path weight; Calculate the average of the comprehensive load balancing degrees of all nodes on the path and use it as the load balancing score; Inputting the corrected path weight, the interference superposition accumulation and the load balancing score into a scoring function to generate a path optimization score; The scoring function is defined as: Path optimization score = α × modified path weight + β × interference superposition accumulation + γ × load balancing score Among them, α, β, and γ are preset weight coefficients, and they satisfy α+β+γ=1.

8. The microwave network topology optimization method based on graph neural network according to claim 7, characterized in that: The sending of the microwave network topology optimization strategy to a microwave network management system to trigger a link reconstruction operation includes: Analyzing the target optimization path of the microwave network topology optimization and determining the microwave node port configuration parameters that need to be adjusted; Generate a link switching instruction set according to the port configuration parameters, wherein the link switching instruction set includes a port frequency adjustment value, a transmit power calibration value, and an antenna pointing angle correction value; Performing conflict detection on the link switching instruction set to identify conflicting node pairs that simultaneously trigger configuration changes; reallocating the link switching timing based on the physical location coordinates of the conflicting node pairs, and generating a phased execution instruction queue; Sending the phased execution instruction queue to the corresponding microwave node controller to sequentially perform port frequency adjustment, transmit power calibration, and antenna pointing angle correction operations to achieve link reconstruction operations; The network performance index after the link reconstruction is monitored in real time. If the network performance index does not reach the expected threshold, the dynamic feature extraction process and modeling process are re-triggered.

9. The microwave network topology optimization method based on graph neural network according to claim 8, characterized in that: The performing conflict detection on the link switching instruction set to identify conflicting node pairs that simultaneously trigger configuration changes includes: Construct a geographical coordinate topology map of microwave nodes and calculate the Euclidean distance between any two microwave nodes; If two microwave nodes receive the frequency adjustment instruction at the same time and the Euclidean distance is less than the preset interference radius threshold, they are marked as a frequency conflict node pair; If two microwave nodes receive a transmit power adjustment instruction at the same time and belong to the same communication frequency band, they are marked as a power conflict node pair; If the overlap area of the main lobe direction of the two microwave nodes exceeds the preset ratio after the antenna pointing angles of the two microwave nodes are adjusted, they are marked as a beam conflict node pair; Merging the frequency conflict node pairs, power conflict node pairs and beam conflict node pairs into a global conflict node set; A conflict resolution priority is assigned to each conflicting node pair based on the conflict type, with frequency conflict having the highest priority, power conflict the second highest, and beam conflict the lowest.

10. A microwave network topology optimization system based on graph neural network, characterized in that: It includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the microwave network topology optimization method based on graph neural network as described in any one of claims 1 to 9.

Images