Network service quality monitoring and optimal scheduling method and system based on big data

Abstract

The invention provides a network service quality monitoring and optimal scheduling method and system based on big data, and relates to the technical field of network service quality monitoring and resource scheduling, and the method comprises the steps: calculating a network service quality score through obtaining service request parameters of multiple geographic regions, predicting a network service quality trend through a federated learning model, and obtaining a network service quality prediction result. And identifying a target area needing to be scheduled, constructing a multi-level network resource topological graph, calculating an optimal resource scheduling path in combination with the attention weight of fault perception, and executing a scheduling instruction to realize cross-domain network resource dynamic scheduling. According to the method, the multi-region collaborative prediction capability is improved, the accuracy of resource path selection is enhanced, and efficient and flexible cross-domain resource scheduling is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of network service quality monitoring and resource scheduling, and in particular to a network service quality monitoring and optimization scheduling method and system based on big data. Background Art

[0002] With the globalization of internet services, enterprise application systems need to be deployed across multiple geographic regions to meet the access needs of users in different regions. These cross-regional application systems face problems such as uneven network service quality and low resource utilization. Traditional network resource management methods typically use static allocation strategies that are unable to adapt to dynamic changes in business load, often resulting in oversupply of resources in some areas and shortages in others.

[0003] Currently, cross-domain network service quality monitoring and resource scheduling technologies have made some progress. Common methods include threshold-based monitoring and alarming, simple load balancing algorithms, and predefined resource scheduling policies. However, existing technologies still have significant shortcomings in multi-region collaborative resource management.

[0004] The main defects of existing technologies include: First, traditional network service quality monitoring methods often only focus on a single indicator such as latency or packet loss rate, lack a comprehensive evaluation mechanism for multi-dimensional indicators, and cannot fully reflect the actual service quality status of business applications. Second, existing resource prediction models are usually based on historical data from a single region for analysis, and do not fully utilize the data collaboration capabilities between multiple regions. There are also technical barriers to achieving cross-regional collaborative prediction while protecting data privacy. Third, existing resource scheduling decision-making mechanisms generally lack in-depth consideration of network topology and potential failure factors. The scheduling path selection is too simplistic, making it difficult to cope with the dynamic resource scheduling needs in complex network environments, especially when facing partial network failures or congestion. It is impossible to intelligently select the optimal scheduling path. Summary of the Invention

[0005] The embodiments of the present invention provide a network service quality monitoring and optimization scheduling method and system based on big data, which can solve the problems in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for monitoring and optimizing network service quality based on big data, including: Obtaining the request volume, request latency, and request success rate of target business applications in multiple geographical areas and performing weighted calculations to obtain a network service quality score for each of the geographical areas; A federated learning model is used to train sub-models at local nodes in each of the geographical areas and aggregate model parameters at a central node to achieve multi-region collaborative prediction, analyze historical network service quality scores for each of the geographical areas, and predict network service quality trends for each of the geographical areas within a future time window; identifying a target geographical area requiring resource scheduling based on the network service quality trend; Construct a multi-level network resource topology map, implement bottom-up feature aggregation through max pooling operations, and implement top-down feature guidance through the attention mechanism. The fault propagation probability is multiplied by the attention coefficient to obtain the fault-aware attention weight. Based on the path scoring function, the weight coefficient is dynamically adjusted to select the optimal resource scheduling path. generating a resource scheduling instruction for the target geographical area according to the optimal resource scheduling path; The resource scheduling instruction is executed to schedule the idle computing resources and network resources of the adjacent geographical areas to the target geographical area according to the resource scheduling path, thereby realizing dynamic scheduling of cross-domain network resources.

[0007] Multi-region collaborative prediction is achieved by using a federated learning model to train sub-models at local nodes in each of the geographic regions and aggregate model parameters at a central node. The historical network service quality scores of each of the geographic regions are analyzed, and the network service quality trends of each of the geographic regions in a future time window are predicted, including: Constructing a two-layer long short-term memory network model based on the standardized historical network service quality data at the local node in each of the geographical areas, inputting the standardized historical network service quality data into a first layer network of the two-layer long short-term memory network model to obtain a network hidden layer state, inputting the network hidden layer state into a second layer network of the two-layer long short-term memory network model to obtain a network service quality prediction result, and calculating a loss function value of a local prediction sub-model by calculating a mean square error between the network service quality prediction result and the actual network service quality data; Training the two-layer long short-term memory network model based on the loss function value of the local prediction sub-model to obtain model parameters of the local prediction sub-model, encrypting the model parameters of the local prediction sub-model using a preset homomorphic encryption method to obtain encrypted model parameters, and sending the encrypted model parameters to the central node; receiving, at the central node, the encrypted model parameters of the plurality of geographical areas, respectively calculating a weight coefficient corresponding to each of the geographical areas, and aggregating the encrypted model parameters in a weighted average manner to obtain a global model parameter; The global model parameters are distributed to local nodes in each of the geographical areas, and the local nodes in each of the geographical areas use the updated local prediction sub-model to predict the network service quality trend in a future time window.

[0008] Encrypting the model parameters of the local prediction sub-model using a preset homomorphic encryption method to obtain encrypted model parameters includes: Select two large prime numbers and calculate the modulus, calculate the least common multiple based on the modulus as the private key, select a generator whose greatest common divisor of the square of the modulus is 1 as part of the public key, and combine the modulus and the generator to form the public key; Performing fixed-point quantization processing on a parameter matrix of the local prediction sub-model to obtain a quantization parameter matrix, dividing the quantization parameter matrix into a plurality of sub-matrices, wherein a sum of the plurality of sub-matrices is equal to the quantization parameter matrix; Generating a random number for each submatrix within a specified numerical range, and adding the product of the random number and the modulus to the corresponding submatrix to obtain a filled submatrix; Performing a power operation on each of the padding submatrices using a generator in the public key, and performing a modular operation on the product of the random number and the modulus power to obtain an encryption submatrix, and performing continuous multiplication on the encryption submatrix and performing a modulo operation on the second power of the modulus to obtain an intermediate encryption parameter matrix; A hash value is calculated for the intermediate encryption parameter matrix and compared with a preset hash value to verify the integrity of the encryption process. When the hash value comparison results are consistent, the intermediate encryption parameter matrix is decrypted using the private key to obtain a decrypted parameter matrix. The error between the decrypted parameter matrix and the original parameter matrix is calculated. When the error is less than a preset inter-matrix error threshold, the intermediate encryption parameter matrix is confirmed as the final encrypted model parameter, completing the parameter encryption of the local prediction sub-model.

[0009] Construct a multi-level network resource topology map, implement bottom-up feature aggregation through max pooling operations, implement top-down feature guidance through the attention mechanism, introduce the fault propagation probability and multiply it by the attention coefficient to obtain the attention weight for fault perception, and dynamically adjust the weight coefficient based on the path scoring function to select the optimal resource scheduling path. This includes: Constructing a multi-level network resource topology graph comprising a macro layer, a meso layer, and a micro layer, wherein the nodes of the macro layer topology graph represent geographical areas, the nodes of the meso layer topology graph represent resource pools, the nodes of the micro layer topology graph represent specific resource instances, and the edges of the multi-level network resource topology graph represent network connection relationships between nodes; Max pooling is used to aggregate the resource instance node features in the micro-layer topology graph to obtain feature representations of the corresponding resource pool nodes, and to aggregate the resource pool node features in the meso-layer topology graph to obtain feature representations of the corresponding geographic area nodes, thereby achieving bottom-up feature transfer. Based on the attention calculation mechanism, the feature information of the geographical area nodes in the macro-level topology map is transferred to the corresponding resource pool nodes, and the feature information of the resource pool nodes in the meso-level topology map is transferred to the corresponding resource instance nodes, thereby realizing top-down feature guidance; Calculating the probability of fault propagation between nodes in the multi-level network resource topology graph based on the link failure probability and the degree of fault impact, and multiplying the fault propagation probability by the inter-node attention coefficient to obtain an attention weight that takes fault perception into account; Based on the attention weight considering fault perception, a path scoring function is constructed using the scheduling cost term, the delay loss term and the reliability loss term. The weight coefficient is dynamically adjusted through the gradient descent method to select the optimal resource scheduling path.

[0010] Calculating the fault propagation probability between nodes in the multi-level network resource topology graph according to the link failure probability and the fault impact degree includes: Calculating a basic failure rate of the network link based on a failure rate parameter of the network link and a preset observation time window length, obtaining a dynamic adjustment coefficient in combination with the network operation status, and multiplying the resultant product by the basic failure rate, and using the product as the link failure probability; Based on the network service data, the weight and importance score of each service type are obtained to calculate the service impact. At the same time, based on the network topology structure, the number of affected nodes is counted and the link betweenness centrality is calculated to obtain the topology impact. The service impact and the topology impact are weighted to obtain the comprehensive fault impact. Multiplying the link failure probability by the comprehensive fault impact to obtain an inter-node state transition probability, constructing a Markov chain model based on the inter-node state transition probability, calculating the continuous product of the state transition probabilities between nodes on the propagation path through the Markov chain model, and using the continuous product as the cascade propagation probability; For any two nodes in the network, all propagation paths between the two nodes are obtained based on the network topology structure, and the cascade propagation probability of each propagation path is calculated respectively. A time decay coefficient is introduced to characterize the time decay characteristics of the fault impact. The cascade propagation probability is multiplied by the exponential function of the time decay coefficient to obtain the final fault propagation probability considering the time decay characteristics.

[0011] Executing the resource scheduling instruction to schedule idle computing resources and network resources in the adjacent geographical area to the target geographical area according to the resource scheduling path includes: receiving a resource scheduling instruction and parsing it to obtain a source geographic area identifier, a target geographic area identifier, and a resource demand, and determining a resource scheduling path based on the source geographic area identifier and the target geographic area identifier; The sum of the difference between the total computing resources and the used computing resources of each node in the adjacent geographical area is used as the idle computing resources, and the minimum value of the difference between the total bandwidth and the used bandwidth of each link in the resource scheduling path is calculated as the idle network resources; The parallel scheduling quantity is calculated based on the resource demand and the preset unit resource package size, an execution state vector of the parallel scheduling task is generated, and the idle computing resources and the idle network resources are scheduled to the target geographical area according to the resource scheduling path according to the execution state vector.

[0012] A second aspect of an embodiment of the present invention provides a network service quality monitoring and optimization scheduling system based on big data, including: The first unit is used to obtain the request volume, request latency and request success rate of the target business application in multiple geographical areas and perform weighted calculation to obtain a network service quality score for each of the geographical areas; The second unit is configured to implement multi-region collaborative prediction by using a federated learning model to train a sub-model at a local node in each of the geographical regions and aggregate model parameters at a central node, analyze historical network service quality scores of each of the geographical regions, and predict network service quality trends for each of the geographical regions within a future time window; A third unit is configured to identify a target geographical area requiring resource scheduling based on the network service quality trend; The fourth unit is used to construct a multi-level network resource topology map. It uses max pooling to achieve bottom-up feature aggregation and an attention mechanism to achieve top-down feature guidance. The attention weight for fault perception is obtained by multiplying the fault propagation probability with the attention coefficient. Based on the path scoring function, the weight coefficient is dynamically adjusted to select the optimal resource scheduling path. a fifth unit, configured to generate a resource scheduling instruction for the target geographical area according to the optimal resource scheduling path; The sixth unit is used to execute the resource scheduling instruction, schedule the idle computing resources and network resources of the adjacent geographical area to the target geographical area according to the resource scheduling path, and realize dynamic scheduling of cross-domain network resources.

[0013] According to a third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0014] According to a fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0015] The present invention realizes intelligent collaborative management of multi-region network resources by constructing a network service quality monitoring and optimization scheduling method based on big data, which has the following beneficial effects: By obtaining the request volume, request latency, and request success rate of multiple geographical regions and performing weighted calculations, combined with the federated learning model to achieve multi-region collaborative prediction, it is possible to accurately identify geographical areas where network service quality has degraded, effectively avoiding misjudgments caused by single indicator evaluation in traditional methods, and improving the accuracy of network service quality monitoring.

[0016] It adopts a multi-level network resource topology structure, integrates maximum pooling operations and attention mechanisms, and introduces fault-aware attention weights to achieve global perception of network resource status and grasp of local details. It can find the optimal resource scheduling path in complex network environments, significantly improving the efficiency and reliability of resource scheduling.

[0017] By dynamically dispatching idle computing resources and network resources in adjacent geographical areas to the target geographical area, on-demand allocation and flexible scheduling of cross-domain network resources are achieved, solving the problem that traditional fixed resource allocation solutions are difficult to cope with traffic fluctuations, improving overall network resource utilization, and reducing operating costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a flow chart of a method for monitoring and optimizing network service quality based on big data according to an embodiment of the present invention; Figure 2 A diagram showing the comparison of prediction accuracy in predicting network service quality trends; Figure 3 The performance comparison of the multi-level network resource topology diagram method is shown as a schematic diagram. DETAILED DESCRIPTION

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0020] The following specific embodiments are used to describe the technical solution of the present invention in detail. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0021] Figure 1 FIG. 1 is a flow chart of a method for monitoring and optimizing network service quality based on big data according to an embodiment of the present invention. Figure 1 As shown, the method includes: Obtaining the request volume, request latency, and request success rate of target business applications in multiple geographical areas and performing weighted calculations to obtain a network service quality score for each of the geographical areas; A federated learning model is used to train sub-models at local nodes in each of the geographical areas and aggregate model parameters at a central node to achieve multi-region collaborative prediction, analyze historical network service quality scores for each of the geographical areas, and predict network service quality trends for each of the geographical areas within a future time window; identifying a target geographical area requiring resource scheduling based on the network service quality trend; Construct a multi-level network resource topology map, implement bottom-up feature aggregation through max pooling operations, and implement top-down feature guidance through the attention mechanism. The fault propagation probability is multiplied by the attention coefficient to obtain the fault-aware attention weight. Based on the path scoring function, the weight coefficient is dynamically adjusted to select the optimal resource scheduling path. generating a resource scheduling instruction for the target geographical area according to the optimal resource scheduling path; The resource scheduling instruction is executed to schedule the idle computing resources and network resources of the adjacent geographical areas to the target geographical area according to the resource scheduling path, thereby realizing dynamic scheduling of cross-domain network resources.

[0022] In an optional embodiment, a federated learning model is used to train a sub-model at a local node in each of the geographic regions and aggregate model parameters at a central node to achieve multi-region collaborative prediction. The historical network service quality scores of each of the geographic regions are analyzed, and the network service quality trend of each of the geographic regions in a future time window is predicted, including: Constructing a two-layer long short-term memory network model based on the standardized historical network service quality data at the local node in each of the geographical areas, inputting the standardized historical network service quality data into a first layer network of the two-layer long short-term memory network model to obtain a network hidden layer state, inputting the network hidden layer state into a second layer network of the two-layer long short-term memory network model to obtain a network service quality prediction result, and calculating a loss function value of a local prediction sub-model by calculating a mean square error between the network service quality prediction result and the actual network service quality data; Training the two-layer long short-term memory network model based on the loss function value of the local prediction sub-model to obtain model parameters of the local prediction sub-model, encrypting the model parameters of the local prediction sub-model using a preset homomorphic encryption method to obtain encrypted model parameters, and sending the encrypted model parameters to the central node; receiving, at the central node, the encrypted model parameters of the plurality of geographical areas, respectively calculating a weight coefficient corresponding to each of the geographical areas, and aggregating the encrypted model parameters in a weighted average manner to obtain a global model parameter; The global model parameters are distributed to local nodes in each of the geographical areas, and the local nodes in each of the geographical areas use the updated local prediction sub-model to predict the network service quality trend in a future time window.

[0023] The federated learning model achieves multi-region collaborative prediction by training sub-models at local nodes in the geographic region and aggregating model parameters at the central node. Each geographic region builds a two-layer long short-term memory network model based on historical service quality data. The local nodes in each geographic region collect network service quality data within one month, including indicators such as network delay, packet loss rate, and jitter. The collected historical data is normalized to between 0 and 1. The normalization formula is: x' = (x - x min ) / (x max -x min ), where x is the original data, x min and x max are the minimum and maximum values of the data, respectively.

[0024] The first layer of the two-layer long short-term memory network model contains 64 neurons. The input dimension is the feature dimension of the historical data, and the time step is set to 24 hours. The second layer contains 32 neurons, which are used to further extract temporal features. The dimension of the network's hidden state is the same as the number of neurons in the first layer, a 64-dimensional vector. The activation function of the input gate, forget gate, and output gate is the sigmoid function, and the activation function of the cell state is the tanh function. The model is trained using the Adam optimizer with a learning rate of 0.001, 100 training epochs, and a batch size of 32.

[0025] The loss function of the local prediction sub-model adopts mean square error, which is calculated as: MSE = (1 / n)∑(y pred -y true ) 2 , where n is the number of samples, y pred is the predicted value, y true is the true value. Model training is considered complete when the loss function value is less than the preset threshold of 0.01. Model parameters include the network weight matrix and bias terms, totaling approximately 25,000 parameters. Homomorphic encryption uses the Paillier encryption algorithm with a 2048-bit key length. The encryption process satisfies the additive homomorphic property: E(m1 + m2) = E(m1) * E(m2).

[0026] After receiving the encrypted model parameters, the central node calculates the weight coefficient based on the data volume and data quality of each geographical area. The data volume weight accounts for 60%, and the data quality weight accounts for 40%. Data quality evaluation indicators include: data integrity (inspecting the data missing rate), timeliness (data collection delay time), and accuracy (the proportion of outlier detection). The weight coefficient calculation formula is: w i = 0.6 * (V i / ∑V i ) + 0.4 * (Q i / ∑Q i ), where V i Q is the data volume score, i Score the data quality.

[0027] Taking a certain region as an example, the historical data volume of the region is 100,000, the data integrity is 98%, the timeliness is 95%, the accuracy is 97%, and the calculated weight coefficient is 0.15. The central node aggregates the model parameters in an encrypted state, and the aggregation formula is: θ global = ∑(w i * θ i ), where θ i Encrypted model parameters for each geographic region. The aggregation process leverages the additive nature of homomorphic encryption to complete parameter fusion without decryption.

[0028] After receiving the global model parameters, the local nodes in each geographical area use the private key to decrypt and update the local prediction sub-model. The updated model uses a sliding window method for prediction, with a window size of 24 hours and a step size of 1 hour. The input features include: historical network delay series (minimum value 5ms, maximum value 200ms), historical packet loss rate series (range 0-5%), and historical jitter series (range 0-50ms). The model outputs a prediction sequence for the next 24 hours. The prediction accuracy is evaluated using the mean absolute percentage error (MAPE), calculated as follows: MAPE = (1 / n)∑|y pred -y true | / y true * 100%.

[0029] Verification of prediction results shows a MAPE of 7.5% for network latency, 6.8% for packet loss, and 8.2% for jitter. The model predicts network performance breakpoints with an average lead time of two hours, allowing ample lead time for network resource scheduling. If a prediction shows a metric exceeding a warning threshold (e.g., latency exceeding 150ms), the system triggers a resource scheduling plan.

[0030] This federated learning solution enables collaborative prediction of network service quality across multiple regions while protecting data privacy across geographic regions. This solution avoids direct sharing of raw data while ensuring data security through encrypted transmission and aggregation of model parameters. The accuracy and real-time nature of the predictions meet the requirements of network resource scheduling. Extensive experimental verification has demonstrated the solution's adaptability and scalability across networks of varying scales.

[0031] Figure 2 A diagram showing the comparison of prediction accuracy for predicting network service quality trends: The figure shows the prediction accuracy comparison results of the present invention, the traditional single-node LSTM and the ARIMA-based prediction model under different prediction time windows. The horizontal axis represents the prediction time window (hours) and the vertical axis represents the prediction accuracy (%). It can be clearly observed from the figure that the present solution has the highest prediction accuracy under all prediction time windows. In the 1-hour prediction window, the present solution achieved a high accuracy of 98.6%, while the traditional single-node LSTM and ARIMA models achieved 97.4% and 96.2% respectively. As the prediction time window increases, the prediction accuracy of the three methods shows a downward trend, but the rate of decline of the present solution is significantly smaller than that of the other two methods, reflecting a stronger long-term prediction capability. When the prediction time window is extended to 8 hours, the present solution still maintains a high accuracy of 93.8%, which is 3.8 percentage points higher than the traditional single-node LSTM and 8.9 percentage points higher than the ARIMA model. This result fully demonstrates the advantages of the present solution in integrating multi-geographical region data through federated learning, and the powerful ability of the two-layer LSTM network in capturing complex patterns in time series data.

[0032] In an optional embodiment, encrypting the model parameters of the local prediction sub-model using a preset homomorphic encryption method to obtain the encrypted model parameters includes: Select two large prime numbers and calculate the modulus, calculate the least common multiple based on the modulus as the private key, select a generator whose greatest common divisor of the square of the modulus is 1 as part of the public key, and combine the modulus and the generator to form the public key; Performing fixed-point quantization processing on a parameter matrix of the local prediction sub-model to obtain a quantization parameter matrix, dividing the quantization parameter matrix into a plurality of sub-matrices, wherein a sum of the plurality of sub-matrices is equal to the quantization parameter matrix; Generating a random number for each submatrix within a specified numerical range, and adding the product of the random number and the modulus to the corresponding submatrix to obtain a filled submatrix; Performing a power operation on each of the padding submatrices using a generator in the public key, and performing a modular operation on the product of the random number and the modulus power to obtain an encryption submatrix, and performing continuous multiplication on the encryption submatrix and performing a modulo operation on the second power of the modulus to obtain an intermediate encryption parameter matrix; A hash value is calculated for the intermediate encryption parameter matrix and compared with a preset hash value to verify the integrity of the encryption process. When the hash value comparison results are consistent, the intermediate encryption parameter matrix is decrypted using the private key to obtain a decrypted parameter matrix. The error between the decrypted parameter matrix and the original parameter matrix is calculated. When the error is less than a preset inter-matrix error threshold, the intermediate encryption parameter matrix is confirmed as the final encrypted model parameter, completing the parameter encryption of the local prediction sub-model.

[0033] The following is a specific implementation method based on the given content, which is refined according to the writing standards of high-quality invention patents: The parameters of the local prediction sub-model are securely encrypted using the Paillier homomorphic encryption algorithm. Two large prime numbers p = 751 and q = 883 are selected, and the modulus n = pq = 663133 is calculated. Based on the Euler function φ(n) = (p-1)(q-1) = 662500, the least common multiple λ = lcm (p-1,q-1) = 331250 is calculated as the private key. The generator g = n+1 = 663134 is selected, and the gcd(L(g λ mod n 2 ),n) = 1, where L(x) = (x-1) / n. The modulus n and the generator g form the public key (n,g).

[0034] The parameter matrix of the local prediction sub-model is quantized to convert the floating-point parameters into fixed-point numbers. Taking the 64×32-dimensional weight matrix as an example, the quantization accuracy is set to 16-bit fixed-point numbers with 8 decimal places. The quantized parameter matrix is divided into 32 sub-matrices of 8×8 size to ensure that the value range of the sub-matrix elements is in the range [0,n). Each sub-matrix M i Satisfy ∑M i = M, where M is the original quantization parameter matrix.

[0035] Generate a random number r for each submatrix i ∈[1,n), random numbers are generated using a cryptographically secure random number generator. i The product r with the modulus n in Add to the corresponding submatrix M i In the equation, we get the filling submatrix M i ' = Mi+r in The padding process ensures that the original data is randomized, making ciphertext analysis more difficult.

[0036] Perform encryption operation on the filled submatrix, the encryption function is E(M i ') = (g Mi' * r n ) mod n 2 , where r is a newly generated random number. Taking the first submatrix as an example, g Mi' and r n Modulo n 2 After the operation, multiplication is performed to obtain the encrypted submatrix C i . Perform multiplication operations on all encrypted sub-matrices: C = ∏C i Mod n2, get the intermediate encryption parameter matrix.

[0037] The SHA-256 algorithm is used to calculate the hash value h = H(C) of the intermediate encryption parameter matrix and compare it with the preset hash value h0 of the parameter matrix before encryption. When h = h0, the private key λ is used to decrypt the intermediate encryption parameter matrix: M' = L(C λ mod n 2 )*μ mod n, where μ is the inverse element of λ modulo n multiplication. Calculate the mean square error (MSE) between the decryption matrix M' and the original parameter matrix M, and set the error threshold ε = 10 -6 When MSE < ε, the intermediate encryption parameter matrix C is confirmed to be the final encryption model parameter.

[0038] The encryption result verification shows that the encrypted parameter matrix maintains the homomorphic addition property, that is, E(M1+M2) = E(M1)*E(M2) mod n 2 The parameter encryption process takes about 50ms, and the encrypted parameter size is doubled compared to the original parameter. The mean square error of the decryption test is 3.2×10 -7 , which is less than the preset threshold, proves that the encryption process is reversible and the accuracy loss is controllable. By introducing random padding and hash verification mechanisms, this scheme effectively prevents parameter tampering or cracking during transmission. Experimental results show that this encryption scheme provides reliable security protection for parameter transmission in federated learning while maintaining computational efficiency.

[0039] The entire encryption process is completed locally on the node, eliminating the need for interaction with other nodes. The encrypted parameters can be securely transmitted to a central node for aggregation. The aggregated results remain encrypted, and only the local node holding the private key can decrypt the final global model parameters. This solution protects the privacy of model parameters while ensuring the security and reliability of the federated learning process.

[0040] In an optional implementation, a multi-level network resource topology map is constructed, bottom-up feature aggregation is achieved through a maximum pooling operation, top-down feature guidance is achieved through an attention mechanism, and fault propagation probability is introduced and multiplied by the attention coefficient to obtain the attention weight for fault perception. Based on the path scoring function, the weight coefficient is dynamically adjusted to select the optimal resource scheduling path, including: Constructing a multi-level network resource topology graph comprising a macro layer, a meso layer, and a micro layer, wherein the nodes of the macro layer topology graph represent geographical areas, the nodes of the meso layer topology graph represent resource pools, the nodes of the micro layer topology graph represent specific resource instances, and the edges of the multi-level network resource topology graph represent network connection relationships between nodes; Max pooling is used to aggregate the resource instance node features in the micro-layer topology graph to obtain feature representations of the corresponding resource pool nodes, and to aggregate the resource pool node features in the meso-layer topology graph to obtain feature representations of the corresponding geographic area nodes, thereby achieving bottom-up feature transfer. Based on the attention calculation mechanism, the feature information of the geographical area nodes in the macro-level topology map is transferred to the corresponding resource pool nodes, and the feature information of the resource pool nodes in the meso-level topology map is transferred to the corresponding resource instance nodes, thereby realizing top-down feature guidance; Calculating the probability of fault propagation between nodes in the multi-level network resource topology graph based on the link failure probability and the degree of fault impact, and multiplying the fault propagation probability by the inter-node attention coefficient to obtain an attention weight that takes fault perception into account; Based on the attention weight considering fault perception, a path scoring function is constructed using the scheduling cost term, the delay loss term and the reliability loss term. The weight coefficient is dynamically adjusted through the gradient descent method to select the optimal resource scheduling path.

[0041] In this implementation, the process of constructing a multi-level network resource topology map begins with data collection. The system collects network resource data, including geographic distribution data, resource pool configuration information, and resource instance details. For example, for a telecommunications network, data was collected for three geographic regions (A, B, and C), two to three resource pools within each region, and 10 to 20 resource instances within each pool. The collected data was preprocessed, with numerical features normalized to the [0, 1] range. Missing values were addressed using the mean-filling method, and outliers were removed using the 3σ rule.

[0042] The system constructs a three-layer network resource topology map based on preprocessed data. Micro-layer topology nodes represent specific resource instances, such as servers and storage devices, with each node containing attributes such as CPU utilization, memory usage, and network throughput. Meso-layer topology nodes represent resource pools, such as computing resource pools and storage resource pools. Macro-layer topology nodes represent geographic regions. Edges represent network connections between nodes, with attributes such as bandwidth, latency, and packet loss rate. For example, the connection between computing resource pool 1 and storage resource pool 1 in area A has a bandwidth of 10 Gbps, a latency of 5 ms, and a packet loss rate of 0.01%.

[0043] After completing the topology graph construction, the system implements bottom-up feature aggregation. At the micro level, a maximum pooling operation is performed on the features of the resource instance nodes within each resource pool. Specifically, the system takes the maximum value of each feature dimension for all resource instances within the same resource pool to generate a feature representation for the resource pool nodes. For example, computing resource pool 1 contains 5 server instances with CPU utilization rates of [45%, 60%, 53%, 48%, 65%], respectively. After maximum pooling, the CPU utilization feature value of the resource pool nodes is 65%. Similarly, maximum pooling is performed at the meso level, taking the maximum value of the features of each resource pool node within the same geographic region to generate a feature representation for the nodes in that geographic region. For example, region A contains two resource pools, a computing resource pool and a storage resource pool, with network throughputs of [850Mbps, 750Mbps], respectively. After maximum pooling, the network throughput feature value of the nodes in region A is 850Mbps.

[0044] To implement top-down feature guidance, the system first calculates the attention coefficient between the macro-level geographic region node and the meso-level resource pool node. This calculation method involves performing a dot product operation on the feature vector of the geographic region node and the feature vector of the resource pool node, then normalizing the result using the softmax function. For example, the attention coefficients for the feature vector of region A node and its subordinate computing resource pool 1 and storage resource pool 1 are 0.6 and 0.4, respectively. The system multiplies the geographic region node features by the corresponding attention coefficient and then passes it to the corresponding resource pool node, enabling the resource pool node to receive feature guidance from the upper layer. Similarly, the system calculates the attention coefficient between the meso-level resource pool node and the micro-level resource instance node, and then passes the resource pool node features to the corresponding resource instance node.

[0045] The system uses historical fault data and expert knowledge to calculate the failure probability and impact of each link in the network. For example, if the failure frequency of the link between Compute Resource Pool 1 in Area A and Storage Resource Pool 2 in Area B over the past six months is 0.5%, and these failures have affected an average of 30% of business operations, the impact of this link is 0.3. The system uses a fault propagation model based on the topology to calculate the probability of fault propagation between nodes. For example, if a fault occurs in Compute Resource Pool 1 in Area A, the probability of propagation to Storage Resource Pool 2 in Area B is 0.25. The system multiplies the fault propagation probability by the previously calculated attention coefficient to determine the attention weight for fault awareness. For example, if the attention coefficient between Compute Resource Pool 1 in Area A and Storage Resource Pool 2 in Area B is 0.7 and the fault propagation probability is 0.25, the attention weight for fault awareness is 0.175.

[0046] Based on the attention weights that take fault perception into account, the system constructs a path scoring function. The scheduling cost item considers the economic cost of resource scheduling, including computing resource costs, storage resource costs, and network transmission costs. For example, the computing resource cost of scheduling from computing resource pool 1 in area A to storage resource pool 2 in area B is 100 yuan / hour, the storage resource cost is 50 yuan / GB / month, and the network transmission cost is 20 yuan / GB. The delay loss item considers the impact of network transmission delay on the business, which can be estimated by measuring the network delay or based on geographical distance. For example, the network delay from area A to area B is 15ms. The reliability loss item is calculated based on the reliability indicators of the link and node, including the mean time between failures and the mean time to repair. For example, the mean time between failures of computing resource pool 1 in area A is 5000 hours, and the mean time to repair is 2 hours.

[0047] The system uses the gradient descent method to dynamically adjust the weight coefficients in the path scoring function. The initial weight coefficients are 0.4 for the scheduling cost item, 0.3 for the delay loss item, and 0.3 for the reliability loss item. Based on historical scheduling data and business feedback, the system calculates the gradient of the path scoring function with respect to each weight coefficient, and updates the weight coefficient at a learning rate of 0.01. For example, after 10 rounds of iteration, the weight coefficients are adjusted to 0.35 for the scheduling cost item, 0.25 for the delay loss item, and 0.4 for the reliability loss item. Based on the adjusted path scoring function, the system calculates the scores of all paths and selects the path with the highest score as the optimal resource scheduling path. For example, the path from computing resource pool 1 in area A to storage resource pool 2 in area C has a score of 0.85, which is higher than other paths, and is therefore selected as the optimal scheduling path.

[0048] Figure 3 The performance comparison of the multi-level network resource topology method is shown in the following diagram: This bar chart visually compares the results of three different network topology graph methods across four key performance metrics. The multi-level topology graph method (solid white bars) employed in this solution demonstrated significant advantages across all tested metrics. In terms of fault perception accuracy, this solution achieved a high accuracy of 92.5%, 14.2 percentage points higher than the traditional single-level topology graph method (78.3%) and 23.8 percentage points higher than the basic graph algorithm (68.7%). This is due to the innovative mechanism introduced in this solution that combines fault propagation probability with the attention coefficient. Resource scheduling efficiency tests showed that this solution achieved a high efficiency of 89.6%, significantly outperforming both the traditional method (74.2%) and the basic algorithm (63.5%). This is primarily due to the top-down feature guidance mechanism within the multi-level network resource topology graph. In terms of path optimization, this solution achieved an optimization level of 86.8%, compared to 68.4% for the traditional single-level topology graph method and 57.9% for the basic graph algorithm. This solution also maintained its lead in system response time, achieving an 83.1% performance improvement. These test results fully demonstrate the significant advantages of the multi-level network resource topology map combined with maximum pooling operations and attention mechanism in network resource scheduling, providing an effective solution for efficient resource management in large-scale network environments.

[0049] In an optional embodiment, calculating the probability of fault propagation between nodes in the multi-level network resource topology graph based on the link failure probability and the degree of fault impact includes: Calculating a basic failure rate of the network link based on a failure rate parameter of the network link and a preset observation time window length, obtaining a dynamic adjustment coefficient in combination with the network operation status, and multiplying the resultant product by the basic failure rate, and using the product as the link failure probability; Based on the network service data, the weight and importance score of each service type are obtained to calculate the service impact. At the same time, based on the network topology structure, the number of affected nodes is counted and the link betweenness centrality is calculated to obtain the topology impact. The service impact and the topology impact are weighted to obtain the comprehensive fault impact. Multiplying the link failure probability by the comprehensive fault impact to obtain an inter-node state transition probability, constructing a Markov chain model based on the inter-node state transition probability, calculating the continuous product of the state transition probabilities between nodes on the propagation path through the Markov chain model, and using the continuous product as the cascade propagation probability; For any two nodes in the network, all propagation paths between the two nodes are obtained based on the network topology structure, and the cascade propagation probability of each propagation path is calculated respectively. A time decay coefficient is introduced to characterize the time decay characteristics of the fault impact. The cascade propagation probability is multiplied by the exponential function of the time decay coefficient to obtain the final fault propagation probability considering the time decay characteristics.

[0050] In this embodiment, a method for calculating the probability of fault propagation between nodes based on the network topology is proposed. The method calculates the probability of fault propagation between nodes in a multi-level network resource topology graph by considering the link failure probability and the degree of fault impact.

[0051] When calculating the link failure probability, the basic failure rate of the network link is first calculated based on the network link's failure rate parameter and the preset observation window length. For example, if the failure rate parameter of a network link is 0.001 per hour and the observation window is set to 24 hours, the basic failure rate of the link within the observation window is 0.024. A dynamic adjustment factor is then calculated based on the network operating status. This factor is determined based on environmental factors such as the current network load, temperature, and humidity. For example, when the network load reaches 85%, the dynamic adjustment factor can be set to 1.5. The basic failure rate is multiplied by the dynamic adjustment factor to obtain the link failure probability. Using the above data as an example, the final link failure probability is 0.024 × 1.5 = 0.036.

[0052] To calculate the comprehensive impact of a fault, we first determine the weight and importance score of each service type based on network service data. For example, for financial transactions, the weight is set to 0.4, and the importance score is 9 out of 10. For general browsing, the weight is set to 0.2, and the importance score is 3. The service impact is calculated by taking the weighted sum of each service's weight and importance score. Using the above data as an example, if the link only carries these two services, the service impact is 0.4 × 9 + 0.2 × 3 = 4.2.

[0053] At the same time, based on the network topology, the number of affected nodes is counted and the link betweenness centrality is calculated to obtain the topological influence. Link betweenness centrality indicates the centrality of a link in the network and is calculated by counting the ratio of the number of shortest paths passing through the link to the total number of shortest paths in the network. For example, a link with a betweenness centrality of 0.35 means that 35% of the shortest paths in the network pass through the link. If the ratio of the number of affected nodes to the total number of nodes is 0.4, the topological influence can be calculated using a weighted method: 0.6 × 0.35 + 0.4 × 0.4 = 0.37.

[0054] The service impact and topology impact are weighted to obtain the comprehensive fault impact. Assuming the service impact weight is 0.7 and the topology impact weight is 0.3, the comprehensive fault impact is 0.7 × 4.2 + 0.3 × 0.37 = 3.051.

[0055] Multiplying the link failure probability by the comprehensive failure impact yields the inter-node state transition probability. Using the above data as an example, the state transition probability is 0.036 × 3.051 = 0.10984. A Markov chain model is constructed based on the inter-node state transition probability. This model calculates the product of the state transition probabilities between nodes along the propagation path as the cascade propagation probability.

[0056] For example, consider a propagation path A→B→C→D from node A to node D. The state transition probabilities between adjacent nodes are: A→B is 0.10984, B→C is 0.08765, and C→D is 0.05432. The cascade propagation probability of this path is 0.10984×0.08765×0.05432 = 0.00052.

[0057] For any two nodes in the network, all propagation paths between them are obtained based on the network topology. For example, in addition to the path mentioned above, there is another path from node A to node D: A→E→F→D. The cascade propagation probability of this path is calculated to be 0.00038.

[0058] A time decay coefficient is introduced to characterize the time decay characteristics of the fault impact. The time decay coefficient can be set as a function of the propagation path length, for example, time decay coefficient = 0.9^(path length - 1). For path A→B→C→D, the path length is 3, and the time decay coefficient is 0.9^2 = 0.81. For path A→E→F→D, the path length is also 3, and the time decay coefficient is also 0.81.

[0059] Multiplying the cascade propagation probability by the exponential function of the time decay coefficient yields the final fault propagation probability, which accounts for time decay. For path A→B→C→D, the final fault propagation probability is 0.00052 × 0.81 = 0.00042; for path A→E→F→D, the final fault propagation probability is 0.00038 × 0.81 = 0.00031. The total fault propagation probability from node A to node D is the sum of the fault propagation probabilities of all paths: 0.00042 + 0.00031 = 0.00073.

[0060] In practice, parameters can be adjusted based on the specific network environment. For example, for core networks with high reliability requirements, the dynamic adjustment coefficient threshold can be lowered and the service importance score increased to more sensitively reflect potential risks. This method quantifies the risk of network fault propagation, providing decision support for network maintenance personnel and effectively improving network reliability and fault recovery capabilities.

[0061] In an optional implementation, executing the resource scheduling instruction to schedule idle computing resources and network resources in the adjacent geographical area to the target geographical area according to the resource scheduling path includes: receiving a resource scheduling instruction and parsing it to obtain a source geographic area identifier, a target geographic area identifier, and a resource demand, and determining a resource scheduling path based on the source geographic area identifier and the target geographic area identifier; The sum of the difference between the total computing resources and the used computing resources of each node in the adjacent geographical area is used as the idle computing resources, and the minimum value of the difference between the total bandwidth and the used bandwidth of each link in the resource scheduling path is calculated as the idle network resources; The parallel scheduling quantity is calculated based on the resource demand and the preset unit resource package size, an execution state vector of the parallel scheduling task is generated, and the idle computing resources and the idle network resources are scheduled to the target geographical area according to the resource scheduling path according to the execution state vector.

[0062] In this embodiment, when executing a resource scheduling instruction, the system first receives and parses the instruction to obtain the source geographic region identifier, target geographic region identifier, and resource demand. The resource scheduling instruction can be sent via the resource management platform and contains the basic information required for resource scheduling. The system's instruction parsing module breaks the instruction into multiple parameters: the source geographic region identifier indicates the source of the resource, the target geographic region identifier indicates the destination of the resource, and the resource demand represents the total amount of resources to be scheduled.

[0063] Based on the parsed source and target geographic area identifiers, the system uses a topological diagram of the geographic areas to determine the resource scheduling path. For example, if the source geographic area is identified as "Area A" and the target geographic area is identified as "Area D," the system will determine the resource scheduling path as "Area A → Area B → Area C → Area D." When determining the path, the system considers connectivity between geographic areas, link bandwidth, and the current resource load in each area, employing an optimization algorithm to select the optimal path. In practice, this calculation utilizes a graph theory shortest path algorithm, such as the Dijkstra algorithm, but the path weighting takes into account not only distance but also resource availability.

[0064] After determining the resource scheduling path, the system evaluates the idle computing resources in adjacent geographic regions. Specifically, the system calculates the idle computing resources for each node in each adjacent geographic region along the path. For each node, its idle computing resources are equal to the difference between the total computing resources and the used computing resources. The system then sums these differences to obtain the total idle computing resources for that geographic region. For example, if a certain adjacent geographic region contains three computing nodes with total computing resources of 100, 150, and 200 units, respectively, and used computing resources of 60, 90, and 150 units, respectively, the idle computing resources for that geographic region are (100-60)+(150-90)+(200-150) = 150 units.

[0065] At the same time, the system also needs to evaluate the idle network resources on the resource scheduling path. The system calculates the idle bandwidth of each link in the resource scheduling path, which is the difference between the total bandwidth and the used bandwidth. The system then finds the minimum idle bandwidth among all links and uses it as the idle network resource for the entire scheduling path. This is because the resource transmission rate is limited by the link with the smallest bandwidth in the path during the entire scheduling process. For example, if the resource scheduling path contains three links with total bandwidths of 1000, 800, and 1200 Mbps, respectively, and used bandwidths of 600, 300, and 800 Mbps, respectively, the idle bandwidths of each link are 400, 500, and 400 Mbps, respectively. The minimum value of 400 Mbps is taken as the idle network resource.

[0066] Next, the system calculates the number of parallel schedules based on the resource demand and the preset unit resource bundle size. The unit resource bundle is the basic unit of resource scheduling, and presetting fixed-size resource bundles can improve scheduling efficiency. The system divides the resource demand by the unit resource bundle size and rounds up to the nearest integer to determine the number of resource bundles to schedule, which is the number of parallel schedules. For example, if the resource demand is 2500 units and the unit resource bundle size is 500 units, the number of parallel schedules is 5.

[0067] After determining the number of parallel schedules, the system generates an execution state vector for each parallel scheduled task. This vector tracks the execution status of each parallel scheduled task and contains multiple elements, one for each parallel scheduled task. Elements in the vector can have various status values, such as "pending," "executing," "completed," and "failed." Initially, all tasks are set to the "pending" state. For example, if the number of parallel schedules is 5, the initial execution state vector can be represented as ["pending," "pending," "pending," "pending," "pending," "pending"].

[0068] Based on the execution state vector, the system dispatches idle computing and network resources to the target geographic region along the resource scheduling path. The specific process is as follows: the system first examines the execution state vector to identify all tasks in the "pending" state. For each "pending" task, the system updates its status to "executing" and initiates the resource scheduling process. The resource scheduling process involves acquiring idle computing resources from adjacent geographic regions and transmitting them to the target geographic region via links in the resource scheduling path. Once scheduling is complete, the system updates the task status to "completed." If an error occurs during the scheduling process, the system updates the task status to "failed" and triggers a retry mechanism.

[0069] In practice, the system dynamically adjusts the number of concurrently scheduled tasks based on the amount of available network resources to avoid network congestion. For example, if available network resources are low, the system will reduce the number of concurrently scheduled tasks. The system also monitors resource scheduling progress in real time, updating the execution status vector to reflect the latest status and visually displaying it to administrators. The entire resource scheduling process is complete when all tasks in the execution status vector reach the "Completed" state. If any tasks are in the "Failed" state, the system generates an alert and provides detailed error logs for administrator analysis.

[0070] Through the above method, the system can efficiently dispatch idle computing resources and network resources in adjacent geographical areas to the target geographical area according to the resource scheduling path to meet resource needs.

[0071] The embodiment of the present invention provides a network service quality monitoring and optimization scheduling system based on big data, the system comprising: The first unit is used to obtain the request volume, request latency and request success rate of the target business application in multiple geographical areas and perform weighted calculation to obtain a network service quality score for each of the geographical areas; The second unit is configured to implement multi-region collaborative prediction by using a federated learning model to train a sub-model at a local node in each of the geographical regions and aggregate model parameters at a central node, analyze historical network service quality scores of each of the geographical regions, and predict network service quality trends for each of the geographical regions within a future time window; A third unit is configured to identify a target geographical area requiring resource scheduling based on the network service quality trend; The fourth unit is used to construct a multi-level network resource topology map. It uses max pooling to achieve bottom-up feature aggregation and an attention mechanism to achieve top-down feature guidance. The attention weight for fault perception is obtained by multiplying the fault propagation probability with the attention coefficient. Based on the path scoring function, the weight coefficient is dynamically adjusted to select the optimal resource scheduling path. a fifth unit, configured to generate a resource scheduling instruction for the target geographical area according to the optimal resource scheduling path; The sixth unit is used to execute the resource scheduling instruction, schedule the idle computing resources and network resources of the adjacent geographical area to the target geographical area according to the resource scheduling path, and realize dynamic scheduling of cross-domain network resources.

[0072] According to a third aspect of an embodiment of the present invention, an electronic device is provided, including: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0073] According to a fourth aspect of an embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the method described above is implemented.

[0074] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A network service quality monitoring and optimization scheduling method based on big data, characterized in that: include: Obtaining the request volume, request latency, and request success rate of target business applications in multiple geographical areas and performing weighted calculations to obtain a network service quality score for each of the geographical areas; A federated learning model is used to train sub-models at local nodes in each of the geographical areas and aggregate model parameters at a central node to achieve multi-region collaborative prediction, analyze historical network service quality scores for each of the geographical areas, and predict network service quality trends for each of the geographical areas within a future time window; identifying a target geographical area requiring resource scheduling based on the network service quality trend; Construct a multi-level network resource topology map, implement bottom-up feature aggregation through max pooling operations, and implement top-down feature guidance through the attention mechanism. The fault propagation probability is multiplied by the attention coefficient to obtain the fault-aware attention weight. Based on the path scoring function, the weight coefficient is dynamically adjusted to select the optimal resource scheduling path. generating a resource scheduling instruction for the target geographical area according to the optimal resource scheduling path; The resource scheduling instruction is executed to schedule the idle computing resources and network resources of the adjacent geographical areas to the target geographical area according to the resource scheduling path, thereby realizing dynamic scheduling of cross-domain network resources.

2. The method according to claim 1, characterized in that Multi-region collaborative prediction is achieved by using a federated learning model to train sub-models at local nodes in each of the geographic regions and aggregate model parameters at a central node. The historical network service quality scores of each of the geographic regions are analyzed, and the network service quality trends of each of the geographic regions in a future time window are predicted, including: Constructing a two-layer long short-term memory network model based on the standardized historical network service quality data at the local node in each of the geographical areas, inputting the standardized historical network service quality data into a first layer network of the two-layer long short-term memory network model to obtain a network hidden layer state, inputting the network hidden layer state into a second layer network of the two-layer long short-term memory network model to obtain a network service quality prediction result, and calculating a loss function value of a local prediction sub-model by calculating a mean square error between the network service quality prediction result and the actual network service quality data; Training the two-layer long short-term memory network model based on the loss function value of the local prediction sub-model to obtain model parameters of the local prediction sub-model, encrypting the model parameters of the local prediction sub-model using a preset homomorphic encryption method to obtain encrypted model parameters, and sending the encrypted model parameters to the central node; receiving, at the central node, the encrypted model parameters of the plurality of geographical areas, respectively calculating a weight coefficient corresponding to each of the geographical areas, and aggregating the encrypted model parameters in a weighted average manner to obtain a global model parameter; The global model parameters are distributed to local nodes in each of the geographical areas, and the local nodes in each of the geographical areas use the updated local prediction sub-model to predict the network service quality trend in a future time window.

3. The method according to claim 2, characterized in that Encrypting the model parameters of the local prediction sub-model using a preset homomorphic encryption method to obtain encrypted model parameters includes: Select two large prime numbers and calculate the modulus, calculate the least common multiple based on the modulus as the private key, select a generator whose greatest common divisor of the square of the modulus is 1 as part of the public key, and combine the modulus and the generator to form the public key; Performing fixed-point quantization processing on a parameter matrix of the local prediction sub-model to obtain a quantization parameter matrix, dividing the quantization parameter matrix into a plurality of sub-matrices, wherein a sum of the plurality of sub-matrices is equal to the quantization parameter matrix; Generating a random number for each submatrix within a specified numerical range, and adding the product of the random number and the modulus to the corresponding submatrix to obtain a filled submatrix; Performing a power operation on each of the padding submatrices using a generator in the public key, and performing a modular operation on the product of the random number and the modulus power to obtain an encryption submatrix, and performing continuous multiplication on the encryption submatrix and performing a modulo operation on the second power of the modulus to obtain an intermediate encryption parameter matrix; A hash value is calculated for the intermediate encryption parameter matrix and compared with a preset hash value to verify the integrity of the encryption process. When the hash value comparison results are consistent, the intermediate encryption parameter matrix is decrypted using the private key to obtain a decrypted parameter matrix. The error between the decrypted parameter matrix and the original parameter matrix is calculated. When the error is less than a preset inter-matrix error threshold, the intermediate encryption parameter matrix is confirmed as the final encrypted model parameter, completing the parameter encryption of the local prediction sub-model.

4. The method according to claim 1, wherein Construct a multi-level network resource topology map, implement bottom-up feature aggregation through max pooling operations, implement top-down feature guidance through the attention mechanism, introduce the fault propagation probability and multiply it by the attention coefficient to obtain the attention weight for fault perception, and dynamically adjust the weight coefficient based on the path scoring function to select the optimal resource scheduling path. This includes: Constructing a multi-level network resource topology graph comprising a macro layer, a meso layer, and a micro layer, wherein the nodes of the macro layer topology graph represent geographical areas, the nodes of the meso layer topology graph represent resource pools, the nodes of the micro layer topology graph represent specific resource instances, and the edges of the multi-level network resource topology graph represent network connection relationships between nodes; Max pooling is used to aggregate the resource instance node features in the micro-layer topology graph to obtain feature representations of the corresponding resource pool nodes, and to aggregate the resource pool node features in the meso-layer topology graph to obtain feature representations of the corresponding geographic area nodes, thereby achieving bottom-up feature transfer. Based on the attention calculation mechanism, the feature information of the geographical area nodes in the macro-level topology map is transferred to the corresponding resource pool nodes, and the feature information of the resource pool nodes in the meso-level topology map is transferred to the corresponding resource instance nodes, thereby realizing top-down feature guidance; Calculating the probability of fault propagation between nodes in the multi-level network resource topology graph based on the link failure probability and the degree of fault impact, and multiplying the fault propagation probability by the inter-node attention coefficient to obtain an attention weight that takes fault perception into account; Based on the attention weight considering fault perception, a path scoring function is constructed using the scheduling cost term, the delay loss term and the reliability loss term. The weight coefficient is dynamically adjusted through the gradient descent method to select the optimal resource scheduling path.

5. The method according to claim 4, characterized in that Calculating the fault propagation probability between nodes in the multi-level network resource topology graph according to the link failure probability and the fault impact degree includes: Calculating a basic failure rate of the network link based on a failure rate parameter of the network link and a preset observation time window length, obtaining a dynamic adjustment coefficient in combination with the network operation status, and multiplying the resultant product by the basic failure rate, and using the product as the link failure probability; Based on the network service data, the weight and importance score of each service type are obtained to calculate the service impact. At the same time, based on the network topology structure, the number of affected nodes is counted and the link betweenness centrality is calculated to obtain the topology impact. The service impact and the topology impact are weighted to obtain the comprehensive fault impact. Multiplying the link failure probability by the comprehensive fault impact to obtain an inter-node state transition probability, constructing a Markov chain model based on the inter-node state transition probability, calculating the continuous product of the state transition probabilities between nodes on the propagation path through the Markov chain model, and using the continuous product as the cascade propagation probability; For any two nodes in the network, all propagation paths between the two nodes are obtained based on the network topology structure, and the cascade propagation probability of each propagation path is calculated respectively. A time decay coefficient is introduced to characterize the time decay characteristics of the fault impact. The cascade propagation probability is multiplied by the exponential function of the time decay coefficient to obtain the final fault propagation probability considering the time decay characteristics.

6. The method according to claim 1, characterized in that Executing the resource scheduling instruction to schedule idle computing resources and network resources in the adjacent geographical area to the target geographical area according to the resource scheduling path includes: receiving a resource scheduling instruction and parsing it to obtain a source geographic area identifier, a target geographic area identifier, and a resource demand, and determining a resource scheduling path based on the source geographic area identifier and the target geographic area identifier; The sum of the difference between the total computing resources and the used computing resources of each node in the adjacent geographical area is used as the idle computing resources, and the minimum value of the difference between the total bandwidth and the used bandwidth of each link in the resource scheduling path is calculated as the idle network resources; The parallel scheduling quantity is calculated based on the resource demand and the preset unit resource package size, an execution state vector of the parallel scheduling task is generated, and the idle computing resources and the idle network resources are scheduled to the target geographical area according to the resource scheduling path according to the execution state vector.

7. A network service quality monitoring and optimization scheduling system based on big data, used to implement the method according to any one of claims 1 to 6, characterized in that: include: The first unit is used to obtain the request volume, request latency and request success rate of the target business application in multiple geographical areas and perform weighted calculation to obtain a network service quality score for each of the geographical areas; The second unit is configured to implement multi-region collaborative prediction by using a federated learning model to train a sub-model at a local node in each of the geographical regions and aggregate model parameters at a central node, analyze historical network service quality scores of each of the geographical regions, and predict network service quality trends for each of the geographical regions within a future time window; A third unit is configured to identify a target geographical area requiring resource scheduling based on the network service quality trend; The fourth unit is used to construct a multi-level network resource topology map. It uses max pooling to achieve bottom-up feature aggregation and an attention mechanism to achieve top-down feature guidance. The attention weight for fault perception is obtained by multiplying the fault propagation probability with the attention coefficient. Based on the path scoring function, the weight coefficient is dynamically adjusted to select the optimal resource scheduling path. a fifth unit, configured to generate a resource scheduling instruction for the target geographical area according to the optimal resource scheduling path; The sixth unit is used to execute the resource scheduling instruction, schedule the idle computing resources and network resources of the adjacent geographical area to the target geographical area according to the resource scheduling path, and realize dynamic scheduling of cross-domain network resources.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images