Link quality evaluation method based on test data fusion statistics

Abstract

The invention relates to the field of network link quality evaluation, and discloses a link quality evaluation method based on test data fusion statistics, which comprises the following steps: S1, collecting time delay, flow and topological parameters of a network link, and carrying out sliding window standardization; s2, constructing a four-dimensional non-negative tensor comprising a link number, a parameter dimension, a network slice and time; s3, features are extracted through non-negative Tucker decomposition with sparse constraints; s4, solving a dynamic weight based on a three-party game model of time delay sensitivity, bandwidth competition and reliability guarantee; s5, calculating a link quality score in combination with the weight and the tensor core; and S6, generating a thermodynamic diagram. Through the technical scheme of sliding window standardization and four-dimensional non-negative tensor construction, the technical effects of eliminating parameter dimension differences and adapting to topology dynamic changes are achieved by utilizing window dynamic calculation of mean values and standard deviations and sparse constraint modeling, and the input data quality of a subsequent game model and weight calculation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of network link quality assessment, and in particular to a link quality assessment method based on test data fusion statistics. Background Art

[0002] In the current field of network link quality assessment, the mainstream link quality assessment solutions usually adopt a multi-dimensional parameter acquisition and weighted scoring mechanism to realize link status monitoring. For example, by calculating the statistics of parameters such as delay and traffic in a fixed time window, and setting static weight coefficients based on expert experience or historical data, and finally triggering bandwidth adjustment instructions by combining linear weighted scoring. Some improved methods introduce machine learning models for feature extraction, store real-time status information using two-dimensional data tables (link × parameter), and supplement with a threshold warning mechanism to achieve resource scheduling. At the visualization level, common technologies map link quality levels through red-green dual-color heat maps to provide basic decision-making references for operation and maintenance personnel.

[0003] However, with the popularization of technologies such as network slicing and dynamic topology, the above solutions gradually expose adaptability bottlenecks in engineering practice. The standardized processing of the fixed time window is difficult to adapt to bursty traffic scenarios, and the lag of parameter statistics leads to input deviation of the model; although the two-dimensional data structure simplifies storage, it discards key dimensions such as network slice identification and time series, resulting in limited extraction of associated features in multi-service scenarios. In addition, the static weight allocation lacks dynamic game considerations for delay-sensitive services and bandwidth-competitive traffic, and the threshold-driven strategy is prone to cause bandwidth allocation oscillations due to ignoring historical state backtracking. The heat map visualization has a single dimension and is difficult to support cross-period link state comparison analysis, indirectly reducing the resource scheduling efficiency. Summary of the Invention

[0004] The purpose of the present invention is to provide a link quality assessment method based on test data fusion statistics, which solves the problems of insufficient adaptability and accuracy in the existing technology in complex network environments.

[0005] To achieve the above purpose, the present invention is realized through the following technical solutions: A link quality assessment method based on test data fusion statistics, comprising the following steps: S1. Collect the delay, traffic and topology parameters of the network link and perform sliding window standardization; S2. Construct a four-dimensional non-negative tensor including the number of links, parameter dimensions, network slices and time; S3. Extract features through non-negative Tucker decomposition with sparse constraints; S4. Solve the dynamic weights based on a three-party game model of delay sensitivity, bandwidth competition and reliability guarantee; S5. Calculate the link quality score by combining the weights and the tensor core; S6. Dynamically allocate bandwidth according to the score and generate a heat map.

[0006] Preferably, in the S1 step, the specific steps of sliding window normalization are as follows: Calculate the mean value of each parameter according to a 5 - second time window and the standard deviation , and perform normalization: ; where, is the normalized value of the th parameter of the th link; is the value of the th parameter collected originally; is the mean value of the th parameter of the time window; is the standard deviation of the th parameter of the time window; is the current timestamp; And truncate the data points that satisfy .

[0007] Preferably, in the S2 step, the construction of the four - dimensional non - negative tensor satisfies: Sparsity constraint: ; where, is a four - dimensional tensor with dimensions ; is the number of dynamically changing links; is the parameter dimension; is the network slice type identifier; is a time series with a sampling interval of 1 millisecond; is the L1 norm of the tensor; is the L2 norm of the tensor; Dimension definition: The first dimension is the number of dynamic links , which automatically expands with the change of topology; The third dimension is the network slice identifier , which is mapped to 32 discrete values according to the policy.

[0008] Preferably, in the S3 step, the optimization objective function of non - negative Tucker decomposition is: ; where G is the core tensor, and its dimensions are determined by the decomposition rank; is the factor matrix of the th mode, Corresponding link, parameter, slice, time dimension; is the Frobenius norm of the matrix; And during the decomposition process, for each factor matrix Retain the singular values with an energy occupancy ratio ≥ 95%.

[0009] Preferably, in the step S4, the utility function of the three-party game model is defined as: Delay-sensitive type: ; Bandwidth competition type: ; Reliability guarantee type: ; Wherein, is the link delay variance; is the indicator function; is the threshold of the delay-sensitive service; is the penalty coefficient for exceeding the delay; is the packet sending rate of the link, which is collected in real time through the SNMP protocol; is the packet receiving rate of the link, and is collected synchronously; is the smoothing constant to prevent the denominator from being zero; is the packet loss rate of the link; is the indicator function, which takes the value of 1 when the packet loss rate exceeds 5%, otherwise 0; is the penalty coefficient; is the packet loss rate threshold, and exceeding this value triggers additional penalties.

[0010] Preferably, in the step S5, the update formula of the weight is: ; Wherein, is the dynamic weight of the th parameter at the moment; is the comprehensive quality function partial derivative with respect to the th parameter; is the learning rate adjustment coefficient; is the exponential function; are the 8 corresponding collected parameters, among which, there are 3 delay features, 3 traffic features, and 2 topology features; is the sum of the gradient response values of all parameters.

[0011] Preferably, in the step S5, the calculation formula of the link quality score includes: Linear term: ; Nonlinear interaction term: ; Among them, is the dynamic weight of the th parameter; is the normalized value of the th item of the th parameter of the th link; is the corresponding 8 collected parameters, including 3 delay features, 3 traffic features, and 2 topology features; G(p,q): core tensor at position in is the factor matrix element of the th link under the th core mode; is the factor matrix element of the th parameter dimension for the th link; is the decomposition dimension of the core tensor

[0012] Preferably, in the step S6, the calculation method of the RGB channel values of the heat map is: ; Among them, is the red channel value of the th link in the heat map, and red indicates lower link quality; is the blue channel value of the th link in the heat map, and blue indicates higher link quality; is the real-time quality score of the th link; is the lowest quality score among all links in the current network; is the highest quality score among all links in the current network; is the maximum value of the RGB color channels.

[0013] Preferably, in the learning rate adjustment coefficient , the update method of the learning rate parameter is: ; Among them, is the L2 error between the predicted score and the actual score; is the sign function; is the learning rate adjustment step size; is the gradient truncation threshold.

[0014] Preferably, in the step S6, the adjustable bandwidth in the allocated bandwidth is dynamically calculated according to the following formula: ; wherein, is the dynamic bandwidth increment that can be allocated at the current moment; is the total available network bandwidth resource pool; is the time interval of the previous bandwidth scheduling; is the time decay factor, and the coefficient is the experimental optimization value; is the sum of all link quality scores in the previous 1 second, reflecting the real-time network state; is the sum of all link quality scores in the previous 5 seconds, used to smooth historical fluctuations; is the th link at the moment quality score; is the current number of active links, which is dynamically determined by the network topology.

[0015] In summary, the present invention includes at least one of the following beneficial technical effects: 1. Through the technical solutions of sliding window normalization and four-dimensional non-negative tensor construction, the present invention uses the window to dynamically calculate the mean, standard deviation and sparse constraint modeling, achieving the technical effects of eliminating the dimensional difference of parameters and adapting to the dynamic changes of the topology. Compared with the fixed time window or single-dimensional data table storage scheme in the prior art, it solves the problem of feature alignment distortion caused by its inability to be compatible with multi-source heterogeneous parameters, and improves the input data quality of subsequent game models and weight calculations.

[0016] 2. By extracting multi-dimensional correlation features through non-negative Tucker decomposition and dynamically solving weights in combination with a three-party game model, the present invention realizes the adaptive balance between network resource allocation and service quality requirements. Compared with traditional static weight allocation methods, it effectively overcomes the resource mismatch problems such as latency-sensitive service jamming and bandwidth competition-based traffic overload caused by rigid weights in the prior art, making the scoring results more suitable for the actual business scenario.

[0017] 3. Through the adaptive learning rate parameter update rule and the calculation of the core tensor non-linear term, the present invention solves the problems of weight oscillation or slow convergence caused by a fixed learning rate in gradient descent algorithms. Compared with the manual parameter adjustment or fixed step size strategy in the prior art, this scheme automatically optimizes the iteration speed through error feedback, improving the response sensitivity of the model to abnormal events such as burst traffic and topology changes while ensuring the scoring stability.

[0018] 4. The present invention transforms the quality score into an intuitive graphical output and executable resource instructions through the mapping of the RGB channels of the heat map and the calculation of the dynamic bandwidth increment. Compared with the traditional single-threshold warning or static bandwidth allocation scheme, it uses the tensor slicing in the time dimension to realize the historical state backtracking, solves the problem of resource allocation lag caused by the lack of spatio-temporal correlation analysis in the prior art, and improves the network utilization rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flowchart of the method steps of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] The following will further describe the present invention in detail Figure 1 in conjunction with the accompanying drawings.

[0021] The present invention provides a link quality evaluation method based on the fusion and statistics of test data, as Figure 1 shown, this link quality evaluation method based on the fusion and statistics of test data may include the following steps: S1. Collect the delay, traffic, and topology parameters of the network link and perform sliding window normalization; In step S1, the specific steps of sliding window normalization are as follows: Calculate the mean value and standard deviation of each parameter according to a 5-second time window, and perform normalization: ; wherein, is the normalized value of the th link and the th parameter; is the original collected value of the th parameter; is the mean value of the th parameter in the time window; is the standard deviation of the th parameter in the time window; is the current timestamp; and truncate the data points that satisfy .

[0022] Specifically, in this embodiment, step S1 includes the following technical content: Real-time collect the delay characteristics (such as delay mean, variance, jitter), traffic characteristics (such as packet sending rate, packet receiving rate, packet loss rate), and topology characteristics (such as number of hops, number of neighbor nodes) of the network link, a total of 8 parameters.

[0023] Dynamically standardize parameters according to a fixed time window to adapt to network state fluctuations. Generally, the window length needs to balance real-time performance and smoothing effect. If it is too short, it is vulnerable to noise interference, and if it is too long, it will cause response delay.

[0024] Specifically, the normalization calculation of the j-th parameter of the i-th link is as follows: ; Among them, is the normalized value of the j-th parameter of the -th link; is the original collected value of the j-th parameter; is the mean value of the j-th parameter in the time window; is the standard deviation of the j-th parameter in the time window; is the current timestamp.

[0025]

[0026] In a possible implementation, the standardized parameters need to be truncated: when is satisfied, force the assignment to suppress the interference of extreme outliers on subsequent model training.

[0026] And the normalized parameter value is used as the input of the linear term in step S5, which directly affects the quality score calculation; Dynamic statistics , : Participate in the construction of the four-dimensional tensor in step S2 (the dimension includes the time series), and are used for feature extraction of non-negative Tucker decomposition.

[0027] For example, in the case of traffic bursts, the standard deviation of the traffic parameters in the window increases significantly, and the standardized will reflect the relative deviation degree of the parameter, and then adjust the weight allocation strategy through the game model in step S4.

[0028] As an option, the 5-second interval of the sliding window has been experimentally verified to balance the following requirements: The ability to capture rapidly changing parameters such as delay jitter; Avoid the system overhead caused by frequent recalculation; Form a multiple relationship with the 1-second update frequency of bandwidth allocation in step S6 to ensure data synchronization consistency.

[0029] In addition, a smoothing constant is introduced into the denominator of the standard deviation in the normalization formula, which can maintain numerical stability even when the parameter fluctuation is extremely low (such as when the topological hop count is fixed), and avoid the problem of ill-conditioned matrices in subsequent tensor decomposition.

[0030] S2. Construct a four-dimensional non-negative tensor containing the number of links, parameter dimension, network slice, and time; In step S2, the four-dimensional non-negative tensor is constructed to satisfy: Sparsity constraint: ; where is a four-dimensional tensor with dimensions ; is the dynamically changing number of links; is the parameter dimension; is the network slice type identifier; is a time series with a sampling interval of 1 millisecond; is the L1 norm of the tensor; is the L2 norm of the tensor; Dimension definition: The first dimension is the dynamic number of links , which automatically expands with the change of topology; The third dimension is the network slice identifier , which is mapped to 32 discrete values according to the policy.

[0031] Specifically, in this embodiment, the technical implementation of step S2 includes the following: The four-dimensional non-negative tensor is constructed to satisfy: Sparsity constraint: ; where is a four-dimensional tensor with dimensions ; is the dynamically changing number of links; is the parameter dimension; is the network slice type identifier; is a time series with a sampling interval of 1 millisecond; is the L1 norm of the tensor; is the L2 norm of the tensor; represents the slice of the tensor at a specific index and timestamp ; Dimension definition: The first dimension is the dynamic number of links , which automatically expands with the change of topology; The third dimension is the network slice identifier , which is mapped to 32 discrete values according to the policy.

[0032] In a possible implementation, the dynamic expansion mechanism of the first - dimension link number is completed through memory pre - allocation and real - time indexing. When a new link is detected to be accessed, the tensor expands along the first dimension and initializes the parameter values. For example, when a large number of Internet of Things devices are accessed in batches, resulting in a sudden increase in the link number, the system automatically expands the storage space of the tensor and fills the new link data into the corresponding positions. The third - dimension network slice identifier is implemented by hash coding, converting the policy identifier into a 32 - bit integer value to ensure the orthogonal distribution of different slice types in the tensor space.

[0033] Specifically, the time dimension synchronizes with the 5 - second normalization window in step S1 through a sliding window mechanism, generating a timestamp slice every 1 millisecond to record the parameter snapshots of all current links. At milliseconds, the corresponding slice in the tensor contains the normalized parameter values of all active links at that moment and forms a continuous time - series relationship with the millisecond slice. The sparsity constraint is achieved through the ratio of L1 and L2 norms, restricting the proportion of non - zero elements in the tensor to not exceed 10% to avoid the curse of dimensionality in subsequent decomposition processes.

[0034] As an option, the sparse constraint coefficient of 0.1 has been experimentally verified to balance data fidelity and computational complexity. When the network load is high, the parameter values of a large number of links approach zero (such as the packet - sending rate of low - traffic links). At this time, the constraint condition automatically triggers zero - value compression to reduce invalid calculations. For example, during the low - load period in the early morning, the traffic characteristics of 80% of the links are close to zero, and the sparsification process can reduce the tensor storage overhead to 35% of the original data.

[0035] The tensor construction process forms a closed - loop with the data pre - processing in step S1. The truncated and normalized values output by step S1 are written into the corresponding positions of the tensor, and the specific coordinates are , where the value is determined by the network slice type to which the link belongs. For example, for a link carrying video services , its time - delay parameter at a certain moment is written into , the traffic parameter is written into , and the topology parameter is written into . The high - precision sampling (1 millisecond) in the time dimension supports real - time decision - making for bandwidth allocation in step S6. For example, predicting the traffic mutation trend based on the tensor slices in the recent 100 milliseconds.

[0036] S3. Extract features through non - negative Tucker decomposition with sparse constraints; In step S3, the optimization objective function of non - negative Tucker decomposition is as follows: ; where \(G\) is the core tensor, and its dimension is determined by the decomposition rank; is the factor matrix of the th mode, corresponding to the link, parameter, slice, and time dimensions; is the Frobenius norm of the matrix; and during the decomposition process, for each factor matrix the singular values with an energy proportion ≥ 95% are retained; Specifically, in this embodiment, the specific implementation of step S3 includes the following technical content: The optimization objective function of non - negative Tucker decomposition is defined as: ; where \(G\) is the core tensor, and its dimension is determined by the decomposition rank; is the factor matrix of the th mode, corresponding to the link, parameter, slice, and time dimensions; is the Frobenius norm of the matrix.

[0037] In a possible implementation, the method for determining the decomposition rank is as follows: perform singular value decomposition on the mode - unfolded matrix of the original tensor, retain the first singular values with an accumulated energy proportion ≥ 95%, and take . For example, after calculating the singular values of the parameter - dimension matrix expansion, if the cumulative proportion of the first 3 singular values reaches 96%, then set to achieve a compressed expression of parameter correlation.

[0038] Specifically, the non - negative constraint of the factor matrix is realized by the projected gradient descent algorithm, and all elements are forced to be ≥ 0 during the iteration process. For example, each row of the link - dimension factor matrix represents the eigenvector of a specific link in the latent space, and the non - negativity ensures that the features can be superimposed and interpreted, avoiding ambiguity in weight calculation caused by positive - negative cancellation. Each column of the time - dimension factor matrix represents the time - evolution pattern, and the sparse constraint makes only key time points (such as the moment of traffic peak) have significant non - zero values, reducing interference from redundant features.

[0039] As an option, the core tensor is optimized using the alternating least - squares method, and in each iteration, other variables are fixed to update a single - mode factor matrix. For example, when updating the parameter dimension, fix and solve the following sub - problem: ; Among them, is the matrix expansion form of the tensor along the parameter dimension; represents the Kronecker product. This optimization process forms a nested iteration with the weight update in step S5, sharing gradient calculation resources; represents parameter matrices of different dimensions; represents the matrix (or tensor) to be optimized, corresponding to the parameters of the second dimension; represents the weight matrix.

[0040] The decomposition result outputs two types of key data: the core tensor is used for the calculation of the non - linear interaction term for the quality score in step S5, and the factor matrix is input into the three - party game model in step S4 to drive the dynamic weight update. For example, for the parameter dimension factor matrix the th row represents the projection of the th parameter (such as delay variance) in the latent space. Its sparsity reflects the contribution intensity of the parameter to the quality assessment, directly affecting the calculation of the partial derivative of the utility function in step S4.

[0041] Generally, during the decomposition process, singular values with an energy occupancy ratio ≥ 95% are retained for each factor matrix to ensure the completeness of feature extraction. For example, when the number of links suddenly increases, resulting in dimension expansion, the singular value recomputation is automatically triggered, and the value is dynamically adjusted to adapt to the topological changes, avoiding manual intervention. The decomposition rank of the time - dimension factor matrix is usually low because the network state shows stationarity within a short - time window and can be characterized by a few time basis functions.

[0042] S4. Solve the dynamic weight using a three - party game model based on delay sensitivity, bandwidth competition, and reliability guarantee; In step S4, the utility function of the three - party game model is defined as: Delay - sensitive type: ; Bandwidth - competition type: ; Reliability - guarantee type: ; Among them, is the link delay variance; is the indicator function; is the threshold of the delay - sensitive service; is the penalty coefficient for exceeding the delay; is the packet - sending rate of the link, which is collected in real - time through the SNMP protocol; is the packet - receiving rate of the link, and Synchronous acquisition; Smoothing constant to prevent the denominator from being zero; Packet loss rate of the link; Indicator function, taking the value of 1 when the packet loss rate exceeds 5%, otherwise 0; Penalty coefficient; Packet loss rate threshold, exceeding which triggers additional penalties; Specifically, in the embodiment, the technical implementation of step S4 includes the following: The utility function of the three-party game model is defined as follows: Delay-sensitive type: ; Bandwidth competition type: ; Reliability guarantee type: ; Among them, Link delay variance; Indicator function; Threshold of delay-sensitive service; Penalty coefficient for excessive delay; Packet sending rate of the link, collected in real time through the SNMP protocol; Packet receiving rate of the link, and Synchronous acquisition; Smoothing constant to prevent the denominator from being zero; Packet loss rate of the link; Indicator function, taking the value of 1 when the packet loss rate exceeds 5%, otherwise 0; Penalty coefficient; Packet loss rate threshold, exceeding which triggers additional penalties.

[0043] In a possible implementation, the calculation of the partial derivative of the utility function depends on the decomposition result of step S3. For example, the partial derivative of the delay-sensitive utility function with respect to the parameter weight is calculated through the column vector covariance of the link dimension factor matrix , reflecting the contribution intensity of different parameters to the delay fluctuation. Specifically, if the parameter corresponds to the delay mean, the absolute value of its partial derivative increases with increasing, triggering a weight reduction to suppress the resource occupancy of delay-sensitive services.

[0044] Specifically, the Nash equilibrium solution of the three-party game adopts the gradient projection method, and the weight vector is iteratively updated until the following convergence condition is satisfied: ; Among them, Comprehensive quality function, is the dynamic adjustment coefficient (the initial values are 0.4, 0.4, and 0.2 respectively); is the weight parameter vector in the model; is the L2 norm, which calculates the square root of the sum of the squares of the vectors and reflects the overall magnitude of the partial derivative.

[0045] As an option, the penalty coefficient and are adaptively adjusted according to the energy distribution of the core tensor in step S3. For example, when the network slice type is a video stream, the element values of the slice dimension in the core tensor are relatively large, and is automatically increased to 0.6 to strengthen the delay constraint; when the slice is an IoT device, is reduced to 0.4 to prioritize reliability. is reduced to 0.4 to prioritize reliability.

[0046] The game model outputs the dynamic weight vector , which is directly input into the linear scoring term in step S5. For example, when the packet loss rate of a certain link suddenly increases to 7%, the penalty term of the reliability utility function is activated, resulting in a 20% decrease in the corresponding weight , thereby weakening the influence of this parameter in the quality score and avoiding low-reliability links from occupying too much bandwidth resources.

[0047] Generally, the interaction relationship of the three-party game is modeled through the non-linear interaction term in step S3. For example, the element in the core tensor represents the coupling effect between the parameter dimension and the slice dimension. When there is a conflict between the delay parameter and the traffic parameter of a high-priority slice, the value of this element increases, forcing the game model to balance resource allocation through weight adjustment.

[0048] S5. Combine the weight and the tensor core to calculate the link quality score; In step S5, the update formula for the weight is: ; Among them, is the dynamic weight of the th parameter at time; is the partial derivative of the comprehensive quality function with respect to the th parameter; is the learning rate adjustment coefficient; is the exponential function; are the 8 corresponding collected parameters, among which, there are 3 delay features, 3 traffic features, and 2 topological features; is the sum of the gradient response values for all parameters; In step S5, the calculation formula of the link quality score includes: Linear term: ; Non-linear interaction term: ; Among them, is the dynamic weight of the th parameter; is the normalized value of the th item parameter of the th link; is the corresponding 8 collected parameters, including 3 delay features, 3 traffic features, and 2 topology features; G(p,q): core tensor The element at position in is the factor matrix element of the th link under the rd core mode; is the factor matrix element of the th parameter dimension of the th link; is the decomposition dimension of the core tensor ; Learning rate adjustment coefficient In, the update method of the learning rate parameter is: ; Among them, is the L2 error between the predicted score and the actual score; is the sign function; is the learning rate adjustment step size; is the gradient truncation threshold; is the learning rate parameter of the th iteration.

[0049] Specifically, in this embodiment, the technical implementation of step S5 includes the following: Weight update module: The iteration formula of the dynamic weight is: ; Among them, is the dynamic weight of the th parameter at time; is the partial derivative of the comprehensive quality function with respect to the th parameter; is the learning rate adjustment coefficient; is the exponential function; For the eight parameters collected correspondingly, among which, there are three latency features, three traffic features, and two topology features; It is to sum up the gradient response values of all parameters.

[0050] In a possible implementation manner, the update of the learning rate parameter β(t)β(t) is achieved through the following formula: ; Wherein, is the L2 error between the predicted score and the actual score; is the sign function; is the learning rate adjustment step size; is the gradient truncation threshold.

[0051] Quality score calculation module: The link quality score QiQi is the weighted sum of the linear term and the non - linear interaction term: ; Wherein, is the element at position in the core tensor, representing the interaction strength between the parameter dimension and the link dimension; is the element in the th row and th column of the link dimension factor matrix , reflecting the latent space feature of the th link; is the element in the j - th row and q - th column of the parameter dimension factor matrix A(2)A(2), representing the latent space projection of the j - th parameter.

[0052] Specifically, the non - linear interaction term captures the cross - dimensional correlation effect through the tensor core . For example, when in the core tensor, it means that there is a strong positive correlation between the third item of the parameter dimension (such as the traffic packet sending rate) and the second mode of the link dimension, and the eigenvalue of this link in this mode will amplify the contribution of the packet sending rate to the quality score.

[0053] As an option, the dimension rank of the non - linear term is dynamically determined by the decomposition result of step S3. For example, if the parameter dimension decomposition rank , then the non - linear term only calculates the first three - order interaction effects to avoid overfitting. When the network load is low, the rank value is automatically reduced to reduce the calculation overhead.

[0054] The scoring result is input to the heat map generation and bandwidth allocation module in step S6. For example, the real - time score of a certain link, and the blue channel value of its heat mapwill be mapped to high brightness, indicating the priority allocation of bandwidth resources. When the score is lower than 0.3, the red channel dominates, triggering a link status alarm and starting the redundant path switching; wherein, is the red channel value of the th link in the heat map, and red indicates lower link quality; is the blue channel value of the th link in the heat map, and blue indicates higher link quality; is the real-time quality score of the th link; is the lowest quality score among all links in the current network; is the highest quality score among all links in the current network; is the maximum value of the RGB color channels.

[0055] Generally, the learning rate adjustment step size and the gradient clipping threshold are optimized and determined through offline experiments. For example, in the training stage, too large a step size will lead to frequent oscillations, and too small a step size will result in slow convergence; the clipping threshold can limit the change range of in a single iteration to not exceed 10%, avoiding score distortion caused by parameter mutations.

[0056] S6. Dynamically allocate bandwidth according to the score and generate a heat map; In step S6, the calculation method of the RGB channel values of the heat map is: ; wherein, is the red channel value of the th link in the heat map, and red indicates lower link quality; is the blue channel value of the th link in the heat map, and blue indicates higher link quality; is the real-time quality score of the th link; is the lowest quality score among all links in the current network; is the highest quality score among all links in the current network; is the maximum value of the RGB color channels; In step S6, the adjustable bandwidth in the allocated bandwidth is dynamically calculated according to the following formula: ; wherein, is the dynamically adjustable dynamic bandwidth increment at the current moment; is the total available bandwidth resource pool for the network; is the time interval of the previous bandwidth scheduling; is the time decay factor, coefficient is the experimental optimization value; is the sum of all link quality scores in the previous 1 second, reflecting the real-time network status; is the sum of all link quality scores in the previous 5 seconds, used to smooth historical fluctuations; is the th link's quality score at time ; is the current number of active links, dynamically determined by the network topology.

[0057] Specifically, in the embodiment, the technical implementation of step S6 includes the following: Heat map generation module: The RGB channel values of each link are calculated as: ; where is the red channel value of the th link in the heat map, and red indicates lower link quality; is the blue channel value of the th link in the heat map, and blue indicates higher link quality; is the real-time quality score of the th link; is the lowest quality score among all links in the current network; is the highest quality score among all links in the current network; is the maximum value of the RGB color channels.

[0058] In a possible implementation, the heat map rendering engine periodically extracts historical score data from the time dimension slices of step S2 to generate a time series animation. For example, it traces back the score changes in the most recent 60 seconds at 1-second intervals to dynamically display the evolution trend of link quality, assisting operation and maintenance personnel in identifying sudden traffic congestion.

[0059] Dynamic bandwidth allocation module: The calculation formula for the adjustable bandwidth increment ΔB(t)ΔB(t) is: ; where is the adjustable dynamic bandwidth increment at the current moment; is the total available bandwidth resource pool for the network; is the time interval of the previous bandwidth scheduling; is the time decay factor, coefficient is the experimental optimization value; is the sum of all link quality scores in the previous 1 second, reflecting the real-time network status; is the sum of all link quality scores in the first 5 seconds, used to smooth historical fluctuations; is the th link's quality score at time ; is the current number of active links, dynamically determined by the network topology.

[0060] Specifically, the bandwidth allocation strategy adjusts the weights through the non - linear interaction term in step S5. For example, when in the core tensor, it means that the second item of the parameter dimension (such as the packet reception rate) has a strong positive impact on bandwidth allocation. At this time, links with high packet reception rates will preferentially obtain a larger share.

[0061] As an option, the optimized value of the time - decay factor is dynamically fine - tuned through the decomposition result of step S3. For example, when the decomposition rank of the factor matrix in the time dimension is , it indicates that the network state is highly stationary in the time dimension, and the time - decay factor is automatically increased to 0.8 to strengthen the real - time scoring weight; when

[0062] The dynamic bandwidth increment ΔB(t)ΔB(t) is proportionally allocated to each link, and the specific formula is: ; where is the allocated bandwidth of the th link at time. For example, if the real - time score of a link is , and the sum is 5.0, it will obtain 14% of .

[0063] Generally, the red - channel value of the heat map is complementary to the blue - channel value . When is close to , approaches , and the link is displayed as dark blue, indicating a high - quality state; when is less than 0.3, exceeds 200, triggering an alarm log and starting the reliability guarantee penalty mechanism of step S4.

[0064] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A link quality assessment method based on test data fusion statistics, characterized in that It includes the following steps: S1. Collect the delay, traffic, and topology parameters of the network link and perform sliding window normalization; S2. Construct a four-dimensional non-negative tensor including the number of links, parameter dimension, network slice, and time; S3. Extract features through non-negative Tucker decomposition with sparse constraints; S4. Solve the dynamic weight based on a three-party game model of delay sensitivity, bandwidth competition, and reliability guarantee; S5. Calculate the link quality score by combining the weight and the tensor core; S6. Dynamically allocate bandwidth according to the score and generate a heat map.

2. The link quality evaluation method based on test data fusion statistics according to claim 1, characterized in that, In the S1 step, the specific steps of sliding window normalization are: Calculate the mean value of each parameter within a 5 - second time window and the standard deviation , and perform normalization: ； Among them, is the normalized value of the th parameter of the th link; is the value of the th parameter collected originally; is the mean value of the th parameter in the time window; is the standard deviation of the th parameter in the time window; is the current timestamp; And truncate the data points that meet .

3. A link quality assessment method based on test data fusion statistics according to claim 1, characterized in that In the step S2, the construction of the four-dimensional non-negative tensor satisfies: Sparse constraint: ； Among them, is a four-dimensional tensor with dimensions of ; is the number of dynamically changing links; is the parameter dimension; is the network slice type identifier; is a time series with a sampling interval of 1 millisecond; is the L1 norm of the tensor; is the L2 norm of the tensor; Dimension definition: The first dimension is the number of dynamic links , which automatically expands with the change of topology; The third dimension is the network slice identifier , which is mapped to 32 discrete values according to the policy 4. A link quality assessment method based on test data fusion statistics according to claim 1, characterized in that In the S3 step, the optimization objective function of non-negative Tucker decomposition is: ； Among them, G is the core tensor, and its dimension is determined by the decomposition rank; is the factor matrix of the th modality, corresponding to the link, parameter, slice, and time dimensions; is the Frobenius norm of the matrix; And during the decomposition process, for each factor matrix retain the singular values with an energy proportion ≥ 95%.

5. A link quality assessment method based on test data fusion statistics according to claim 1, characterized in that In the S4 step, the utility function of the three-party game model is defined as: Delay-sensitive type: ; Bandwidth competition type: ; Reliability Assurance Type: ; Among them, is the link delay variance; is the indicator function; is the threshold of delay-sensitive services; is the penalty coefficient for exceeding the delay; is the packet sending rate of the link, which is collected in real time through the SNMP protocol; is the packet receiving rate of the link, and is collected synchronously with; is the smoothing constant to prevent the denominator from being zero; is the packet loss rate of the link; is an indicator function, which takes the value of 1 when the packet loss rate exceeds 5%, otherwise 0; is the penalty coefficient; is the packet loss rate threshold, and when the packet loss rate exceeds this value, additional penalties are triggered.

6. The link quality evaluation method based on test data fusion statistics according to claim 1, wherein, In the S5 step, the update formula of the weight is: ； Among them, is the dynamic weight of the th parameter at moment; is the comprehensive quality function partial derivative of the th parameter; is the learning rate adjustment coefficient; is the exponential function; are the 8 parameters corresponding to the collected data, among which, there are 3 delay features, 3 traffic features, and 2 topology features; is the sum of the gradient response values of all parameters.

7. A link quality assessment method based on test data fusion statistics according to claim 1, characterized in that, In the S5 step, the calculation formula of the link quality score includes: Linear term: ; Nonlinear interaction term: ; Among them, is the dynamic weight of the th parameter; is the normalization value of the th item of the th parameter of the th link; is the element at the position in the core tensor is the factor matrix element of the th link under the th core mode; is the factor matrix element of the th parameter dimension of the th link; is the decomposition dimension of the core tensor 8. A link quality assessment method based on test data fusion statistics according to claim 1, characterized in that, In the S6 step, the calculation method of the RGB channel values of the heat map is: ； Among them, is the red channel value of the th link in the heat map, and red indicates lower link quality; is the blue channel value of the th link in the heat map, and blue indicates higher link quality; is the real-time quality score of the th link; is the lowest quality score among all links in the current network; is the highest quality score among all links in the current network; is the maximum value of the RGB color channels.

9. A link quality assessment method based on test data fusion statistics according to claim 6, characterized in that The learning rate adjustment coefficient In, the update method of the learning rate parameter is as follows: ； Among them, is the L2 error between the predicted score and the actual score; is the sign function; is the learning rate adjustment step size; is the gradient truncation threshold.

10. A link quality assessment method based on test data fusion statistics according to claim 1, characterized in that In the step S6, the adjustable bandwidth in the allocated bandwidth is dynamically calculated according to the following formula: ； Among them, is the dynamic bandwidth increment that can be allocated at the current moment; is the total available network bandwidth resource pool; is the time interval of the previous bandwidth scheduling; is the time decay factor, coefficient is the experimental optimization value; is the sum of all link quality scores in the previous 1 second, reflecting the real-time network state; is the sum of all link quality scores in the previous 5 seconds, used to smooth historical fluctuations; is the th link at the moment quality score; is the current number of active links, dynamically determined by the network topology.

Images