A network resource quality evaluation method based on multi-modal data

Abstract

The application relates to the technical field of network resource evaluation, and discloses a network resource quality evaluation method based on multi-modal data. The method acquires multi-modal network data, collects key performance indicators and user behavior data of network resources, associates data points, calculates quality evaluation values, generates a resource quality evaluation set, and then sorts adjacent resource pairs according to network topology distance based on the quality evaluation values and resource identifiers in the set to obtain a quality consistency partition annotation set. Then, resource points in the quality consistent section are extracted, performance indicator time series and user behavior sequences are acquired, change rates are compared in combination with resource load data, influence degrees under network condition fluctuations are evaluated, and influence degree analysis results are generated. Abnormal points with influence values greater than average influence benchmark values and located in quality conflict sections are identified to form a resource abnormal point set. Detailed information of the abnormal points is acquired, abnormal points with potential diffusion risks are marked, and a network resource quality evaluation report is generated.

Description

Technical Field

[0001] This invention relates to the field of network resource assessment technology, specifically to a method for assessing network resource quality based on multimodal data. Background Technology

[0002] With the rapid development of internet technology, the scale of network resources has grown exponentially, encompassing various types such as data storage nodes, computing resources, and content distribution nodes. Their quality directly affects user access experience, business service stability, and overall network operating efficiency. Currently, network resource quality assessment mainly relies on single-dimensional data analysis, such as judging resource quality solely through key performance indicators or user behavior data. This single-dimensional assessment method has significant limitations.

[0003] When relying solely on key performance indicators, it is difficult to reflect the actual quality differences perceived by users. However, if users have a poor access experience due to unreasonable content loading logic, a single performance indicator evaluation may misjudge resource quality. On the other hand, when relying solely on user behavior data, it is impossible to distinguish whether the differences in user behavior are caused by quality problems of the resource itself or by fluctuations in the external network environment, and it is easy to misattribute the impact of the external environment to resource quality problems.

[0004] Existing assessment methods lack consideration for network topology. When analyzing the quality relationships between adjacent network resources, they fail to incorporate topological distances for correlation analysis, making it impossible to accurately identify quality-consistent and conflicting segments. For example, when some resources exhibit quality anomalies within the same network topology area, existing methods struggle to quickly pinpoint whether the anomalies are spreading among adjacent resources or whether they are isolated or regionally correlated. Furthermore, when assessing the impact of network condition fluctuations on resource quality, existing methods do not combine performance indicator time series, user behavior sequences, and resource load data for analysis. This makes it difficult to accurately quantify the degree of impact of fluctuations on different resources, resulting in difficulties in predicting resource quality change trends in advance when the network environment changes, and also hindering the effective identification of potential anomalies.

[0005] Existing methods, after identifying resource anomalies, lack the ability to mark the potential spread risk of these anomalies. They can only simply list anomaly information, failing to provide network operations and maintenance personnel with targeted risk prevention and control strategies. This leads to low operational response efficiency, difficulty in quickly curbing the spread of anomalies, and consequently affects the stable operation of the entire network. These problems collectively result in insufficient accuracy, comprehensiveness, and practicality of existing network resource quality assessment methods, failing to meet the needs for accurate resource quality assessment and efficient operation and maintenance in complex network environments. Summary of the Invention

[0006] The purpose of this invention is to provide a network resource quality assessment method based on multimodal data to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides a network resource quality assessment method based on multimodal data, the method comprising:

[0008] Acquire multimodal network data, collect key performance indicators of network resources and user behavior data, associate data points and calculate corresponding quality assessment values ​​to generate a resource quality assessment set;

[0009] Based on the quality assessment values ​​and resource identifiers in the resource quality assessment set, adjacent resource pairs are sorted by network topology distance, and quality-consistent and quality-conflicting segments in adjacent resource pairs are marked to obtain a quality consistency partition label set.

[0010] Extract resource points located in the quality-consistent segment from the quality-consistent partition annotation set, obtain the time series of performance indicators and user behavior sequences, compare the rate of change with resource load data, assess the degree of impact under network condition fluctuations, and generate impact degree analysis results.

[0011] Identify outliers among the resource points whose impact values ​​are greater than the average impact benchmark value and are located in the quality conflict zone in the impact degree analysis results, and form a set of resource outliers;

[0012] Obtain all anomalies and their detailed information from the resource anomaly set, mark anomalies with potential spread risks, and generate a network resource quality assessment report.

[0013] Preferably, the resource quality assessment set includes quality assessment values, resource identifiers, and normalized performance indicators; the quality consistency partition labeling set specifically includes quality consistency segment labels, quality conflict segment labels, and the difference rate of quality assessment values ​​between adjacent resources; the impact degree analysis results include the impact degree of performance degradation rate on resources, the impact degree of user behavior change rate on resources, and a comparison of load impact response under each network fluctuation condition; the resource anomaly point set includes anomaly point location information, anomaly point performance fluctuation characteristics, and anomaly point load and traffic fluctuation ratio; and the network resource quality assessment report includes an anomaly point list and anomaly point multi-indicator joint judgment label.

[0014] Preferably, acquiring multimodal network data includes:

[0015] Collect key performance indicator data and user behavior data of network resources, obtain resource identifiers and corresponding timestamps, and record the collection results as two data items: performance factors and behavior factors, and obtain resource data groups;

[0016] Based on the performance factor and behavior factor data in the resource data group, normalization processing is performed respectively. The normalized results are then correlated with resource identifiers. The weighted average of the normalized performance value and the normalized behavior value is calculated as the quality assessment value, and a resource quality assessment set is generated.

[0017] Preferably, the step of sorting adjacent resource pairs by network topology distance based on the quality assessment values ​​and resource identifiers in the resource quality assessment set includes:

[0018] Extract the quality assessment values ​​and corresponding resource identifiers from the resource quality assessment set, identify the positional relationship of all resource points in the network based on the resource identifiers, call the resource point coordinate set, calculate and sort the distances of resource points on the network based on the adjacent distance threshold, and generate an adjacent resource point distance sorting sequence.

[0019] Based on the adjacent resource point distance sorting sequence, calculate the quality assessment value difference rate between each pair of adjacent resource points, and integrate them to generate a quality assessment value difference rate sequence.

[0020] Based on the quality assessment value difference rate sequence, the performance change direction and behavior change direction between adjacent resource points are extracted. Each pair of resource points is classified and labeled according to whether the change trends in the two directions are consistent. The segments with consistent directions and conflicting directions are recorded and grouped respectively to obtain a quality consistency partition label set.

[0021] Preferably, the step of extracting resource points located in the quality-consistent segment of the quality-consistent partition annotation set and obtaining the performance index time series and user behavior series includes:

[0022] Based on the quality consistency partition label set, the segments marked as quality consistency are filtered out, and the performance data and user behavior data of each resource point within the time period are detected. The data are arranged in chronological order to form a performance time series and a behavior time series, and an environmental change time series set is generated.

[0023] Based on the environmental change time series set, the performance degradation rate and behavior change rate between consecutive time nodes in the time series of each resource point are calculated. The performance degradation rate and behavior change rate are compared side by side under the same load conditions. By jointly analyzing the two types of rate indicators, the numerical relationship between the change magnitude under the load is identified. The impact value series of each resource point is integrated to establish the impact degree analysis results.

[0024] Preferably, the outliers identified in the impact analysis results, which are resource points with impact values ​​greater than the average impact benchmark and located in the quality conflict zone, include:

[0025] Based on the impact analysis results, resource points with impact response values ​​greater than the average response baseline value and resource points in the quality conflict zone are selected. Continuous resource usage records of resource points are extracted in chronological order, and resource load and traffic data corresponding to each time node are collected to generate a continuous resource record set.

[0026] The continuous resource record set is called to extract the load value and flow value of the resource point at two consecutive time nodes, calculate the load change ratio and flow change ratio respectively, and integrate them into the load change ratio sequence and flow change ratio sequence to establish a resource fluctuation change dataset;

[0027] Based on the resource fluctuation change dataset, network condition amplitude data and user fluctuation data for the corresponding time period are extracted. It is determined whether the load change ratio and traffic change ratio both exceed the set fluctuation identification threshold. It is also determined whether the network condition amplitude and user fluctuation both exceed the disturbance judgment threshold. Time nodes that meet the conditions are marked as anomalies, and a set of resource anomaly points is generated.

[0028] Preferably, the generated network resource quality assessment report includes:

[0029] Obtain all locations and corresponding identification information in the set of resource anomalies, calculate the joint risk assessment value of each anomaly, and establish a joint risk assessment value sequence.

[0030] Based on the joint risk judgment value sequence, points with impact response values ​​greater than the impact response risk threshold, quality assessment values ​​lower than the quality benchmark value, and directional consistency labels as conflict segments are selected. The corresponding point numbers, location identifiers, and their respective partitions are extracted and marked as detection anomalies with a risk of spread. Points that meet the joint conditions are partitioned and output in a structured format to generate a network resource quality assessment report.

[0031] Preferably, the method further includes:

[0032] During the resource assessment process, network response data is captured in real time to generate feedback logs containing resource location offsets and status change information;

[0033] Extract the abnormal event features from the feedback log, and perform similarity matching between the abnormal event features and the historical evaluation case library to generate adaptive adjustment instructions;

[0034] The calculation parameters of the quality assessment value are dynamically updated based on the adaptive adjustment instruction;

[0035] The updated parameters are re-injected into the evaluation process, and the quality consistency partition label set is recalculated.

[0036] Preferably, the method further includes:

[0037] Construct a distributed evaluation architecture and use the distributed evaluation architecture to parse the format differences of different data sources;

[0038] The multimodal network data is converted into evaluation data in a unified format;

[0039] The time-series relationships and anomaly handling context of the data are preserved during the transformation process;

[0040] The required optimization parameters for the evaluation process are injected to generate an evaluation data stream that meets the conditions for distributed execution.

[0041] Preferably, the method further includes:

[0042] Configure multimodal data acquisition channels and data processing channels according to network resource types;

[0043] The multimodal data acquisition channel and data processing channel are configured in parallel to create an efficient data evaluation link.

[0044] Compared with the prior art, the beneficial effects of the present invention are:

[0045] This network resource quality assessment method based on multimodal data breaks through the limitations of traditional single-dimensional data assessment by integrating multimodal network data for evaluation. This method simultaneously collects key performance indicators of network resources and user behavior data, correlates these two types of data, and calculates quality assessment values. The resulting resource quality assessment set can simultaneously reflect both objective resource performance and user subjective perception, avoiding the problems of relying solely on performance indicators while ignoring user experience, or relying solely on user behavior data while failing to distinguish the influence of the external environment. This makes the resource quality assessment results more closely aligned with actual application scenarios and more accurately reflects the true quality status of resources.

[0046] In terms of analysis combined with network topology, this method sorts adjacent resource pairs by network topological distance based on quality assessment values ​​and resource identifiers, and marks quality-consistent and quality-conflicting segments. The resulting quality consistency partition label set can clearly present the quality correlation of resources in different topological locations. This analysis method combined with topological distance enables operations and maintenance personnel to quickly identify the consistency and conflict of resource quality within the same topological region. For example, when the quality assessment values ​​of adjacent resources tend to be consistent within a certain topological segment, it can be judged that the resource quality in that area is stable; while when the quality assessment values ​​of adjacent resources differ significantly, quality conflicts can be detected in a timely manner, providing direction for further locating the cause of the conflict, effectively improving the ability to identify the regional correlation of resource quality.

[0047] In assessing the impact of network condition fluctuations, this method extracts resource points in segments with consistent quality, obtains time series of performance indicators and user behavior sequences, and compares the rate of change with resource load data. The resulting impact analysis accurately quantifies the impact of network condition fluctuations on the quality of different resources. This multi-sequence combined with load data analysis method clearly distinguishes whether resource quality changes are caused by internal performance issues or external network fluctuations. For example, when network conditions fluctuate, if the rate of change of a resource's performance indicators is strongly correlated with the rate of change of load data, and the user behavior sequence changes synchronously, it can be determined that the fluctuation has a significant impact on that resource. Conversely, it can be determined that the resource quality changes are more caused by internal factors. This precise impact analysis provides a clear basis for subsequent targeted resource optimization.

[0048] In terms of anomaly identification and risk labeling, this method identifies anomalies whose impact values ​​exceed the average impact benchmark and are located within quality conflict zones, forming a resource anomaly set. This ensures the accuracy of anomaly identification, excluding non-critical anomalies with low impact while focusing on high-impact anomalies within quality conflict zones. This avoids the problems of excessive anomaly identification or omission of critical anomalies in traditional methods. Furthermore, after obtaining detailed information on anomalies, this method labels anomalies with potential propagation risks. The generated network resource quality assessment report not only includes anomaly information but also clearly identifies the propagation risk of anomalies. This allows operations and maintenance personnel to focus on high-propagation-risk anomalies without having to analyze risks one by one from a large number of anomalies, significantly improving the targeting and efficiency of operations and maintenance responses and reducing the impact of anomaly propagation on the overall stable operation of the network. Attached Figure Description

[0049] Figure 1 This is a schematic diagram illustrating the working principle of the network resource quality assessment method based on multimodal data described in this invention.

[0050] Figure 2 A schematic diagram illustrating the working principle of a resource quality assessment set;

[0051] Figure 3 This is a schematic diagram illustrating the working principle of the resource quality assessment set.

[0052] Figure 4 A schematic diagram illustrating the working principle of the analysis results for identifying the degree of influence. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figure 1 This invention provides a method for network resource quality assessment based on multimodal data, the method comprising:

[0055] By acquiring multimodal network data, key performance indicators of network resources and user behavior data are collected. Data points are correlated and corresponding quality assessment values ​​are calculated to generate a resource quality assessment set. Based on the quality assessment values ​​and resource identifiers in the resource quality assessment set, adjacent resource pairs are sorted by network topology distance, and quality-consistent and quality-conflict segments in adjacent resource pairs are marked to obtain a quality consistency partition label set. Resource points located in quality-consistent segments in the quality consistency partition label set are extracted, and performance indicator time series and user behavior series are obtained. The rate of change is compared with resource load data to assess the impact of network condition fluctuations and generate impact degree analysis results. Anomalies are identified among resource points whose impact values ​​are greater than the average impact benchmark value and are located in quality-conflict segments in the impact degree analysis results, forming a resource anomaly point set. All anomalies in the resource anomaly point set and their detailed information are obtained, anomalies with potential spread risks are marked, and a network resource quality assessment report is generated.

[0056] Example 1: See Figure 2 This involves the detailed definition and processing flow of the resource quality assessment set and its related data structures. The resource quality assessment set includes quality assessment values, resource identifiers, and normalized performance indicators. The quality consistency partition label set is further refined into quality consistency segment labels, quality conflict segment labels, and the difference rate of quality assessment values ​​between adjacent resources. The impact analysis results must cover the impact of performance degradation rate on resources, the impact of user behavior change rate on resources, and a comparison of load impact response under each network fluctuation condition. The structure of the resource anomaly point set includes anomaly point location information, anomaly point performance fluctuation characteristics, and anomaly point load and traffic fluctuation ratio. The final generated network resource quality assessment report must include an anomaly point list and a joint judgment label for multiple anomaly indicators.

[0057] The process of acquiring multimodal network data begins with the collection of key performance indicators (KPIs) of network resources and user behavior data. Resource identifiers and their corresponding timestamps are precisely acquired, and the collection results are systematically recorded as two core data sets: performance factors and behavioral factors, thus forming the initial resource data set. Performance factors specifically include quantitative indicators such as bandwidth utilization, latency, and packet loss rate, while behavioral factors encompass user-side metrics such as user request frequency, session duration, and data transmission volume. Data collection is performed through distributed sensor nodes strategically positioned along the network's critical paths, capturing performance and behavioral data in real time, once per second. The collected raw data is transmitted to the central processing unit for further processing via a high-throughput message queue.

[0058] The central processing unit performs a rigorous verification and cleaning process on the received data. Verification includes data range checks, format conformity verification, and logical consistency verification to ensure data integrity and reliability. The cleaning process aims to remove invalid data and outliers, such as extreme values ​​outside the reasonable range or incorrectly formatted records, thereby improving data quality. The cleaned, valid data is stored in a dedicated time-series database, which optimizes the storage and retrieval performance of time-series data, supporting subsequent normalization processing and analytical calculations.

[0059] Based on the performance factor and behavioral factor data in the resource data set, normalization processing is performed separately. Performance factors are processed using the maximum-minimum normalization method, linearly transforming the original data to the interval [0,1]. Behavioral factors are processed using the Z-score normalization method, conforming to a standard normal distribution with a mean of 0 and a standard deviation of 1. A strict correspondence is established between the normalized results and resource identifiers, ensuring that each data point is traceable to its source resource. Subsequently, the weighted average of the normalized performance and behavioral values ​​is calculated, which is the final quality assessment value. In the weighted average calculation, the weight coefficient for performance factors is set to 0.6, and the weight coefficient for behavioral factors is set to 0.4. These weight coefficients can be dynamically adjusted based on the specific type of network resource through historical data statistical analysis. The calculation process is executed using the MapReduce programming model in a distributed computing framework to achieve efficient parallel computing. The final generated resource quality assessment set is stored in the form of a structured data table, containing fields such as resource identifier, quality assessment value, normalized performance index, and timestamp, and is written to the distributed file system for subsequent module calls.

[0060] The formula for calculating the quality assessment value is as follows:

[0061]

[0062] in: Indicates the quality assessment value. The weighting coefficients of the performance factors. This represents the normalized performance value. The weighting coefficients of the behavioral factors. This represents the normalized behavior value.

[0063] Resource identifiers are generated using a globally unique identifier format, ensuring that each network resource has a unique identifier and avoiding any potential conflicts or confusion. Timestamp records use an internationally standard time format with millisecond-level precision, providing a high-precision time reference for time series analysis. Data acquisition channels are configured according to network resource types; for example, independent acquisition channels are established for network devices, user terminals, and application servers. Data processing channels are also configured in parallel according to the characteristics of the processing tasks, forming an efficient data evaluation chain. Throughout the acquisition and processing process, the data maintains its time series relationship and anomaly handling context, ensuring the consistency and accuracy of subsequent analysis.

[0064] During the conversion of multimodal data into unified evaluation data, the system parses the format differences between different data sources. These differences may manifest in data encoding methods, sampling frequencies, or data structures. Through a format adapter component, the system can handle various common data formats, including JSON, XML, and CSV, and convert them into a unified Avro binary format. The unified evaluation data format includes necessary fields such as data values, timestamps, and data source identifiers, ensuring data consistency and interoperability. During the conversion process, the time-series relationship of the data is maintained by strictly preserving the timestamp sequence and data point order. Anomaly handling context is recorded through additional metadata, including data quality flags, processing history logs, and anomaly tags, providing a complete traceability basis for subsequent anomaly diagnosis and handling.

[0065] Optimized parameters, such as calculation weights, threshold settings, and algorithm parameters, are dynamically injected into the evaluation process through a configuration service. These parameters are adaptively adjusted based on historical operational data and real-time network conditions to optimize the accuracy of the evaluation results. The evaluation data stream is managed through a distributed message flow platform, using Kafka as the message middleware to support high-throughput and low-latency data transmission, meeting the performance requirements of large-scale network data processing. The data acquisition channel uses lightweight proxy nodes deployed near the data source to reduce network transmission overhead and latency. The data processing channel utilizes in-memory computing technology to accelerate data transformation and feature calculation, further improving processing efficiency. The evaluation data stream adopts a partitioned storage strategy, supporting rapid retrieval and analysis by time range and resource type. The entire architecture design supports horizontal scaling, allowing for flexible responses to data volume growth by adding computing nodes, ensuring the system's ability to handle large-scale network data. Finally, all processed data and metadata are integrated into a resource quality evaluation set, forming a complete, consistent, and efficiently accessible data foundation, providing reliable data support for subsequent quality consistency analysis, impact assessment, and anomaly identification.

[0066] Example 2: See Figure 3This involves partitioning and labeling to ensure network resource quality consistency. The process extracts quality assessment values ​​and corresponding resource identifiers from the generated resource quality assessment set. The resource identifiers use a globally unique identifier format to ensure that each resource point has a unique and identifiable characteristic. The physical location and logical connection information of the resource points are obtained from the network topology database to establish a resource point coordinate set. This coordinate set not only contains the geographical coordinates of the resource points but also records their logical positional relationships within the network topology, forming a complete network space mapping.

[0067] Resource point distances are calculated based on an adjacent distance threshold, which is dynamically adjusted according to network size and resource distribution density. The distance calculation employs the shortest path algorithm from graph theory to calculate the network topology distance between resource points. Multiple factors, including network hop count, transmission latency, and bandwidth capacity, are considered during the calculation process to ensure accuracy. When generating the adjacent resource point distance ranking sequence, an optimized ranking algorithm is used to sort resource point pairs in ascending order of distance values ​​for easier subsequent processing. A relative difference calculation method is used to calculate the quality assessment value difference rate between each pair of adjacent resource points. This calculation considers the ratio of the absolute difference in the quality assessment values ​​of two resource points to their average value, expressing the degree of difference as a percentage. The difference rate calculation process fully considers the distribution characteristics of quality assessment values ​​to avoid extreme values ​​having an excessive impact on the calculation results. The generated quality assessment value difference rate sequence contains difference information for all adjacent resource point pairs, providing a data foundation for subsequent analysis.

[0068] When extracting the direction of performance and behavior changes between adjacent resource points, a time-series data analysis method is used. The direction of performance changes is determined by analyzing the changing trends of performance indicators over time, while the direction of behavior changes is identified by identifying patterns in user behavior data. A sliding window technique is used to determine the direction, analyzing the characteristics of data changes within a continuous time interval. The changing trends of each pair of resource points in both directions are compared, and classification and labeling are performed based on the degree of consistency of the trends.

[0069] The directional consistency classification employs a binary labeling system, using specific numerical values ​​to represent consistent and conflicting states. The classification process considers factors such as the magnitude and duration of change trends to ensure the accuracy of the labeling results. Segments with consistent directions are grouped and recorded, detailing their starting resource point, ending resource point, and consistency strength index. Segments with conflicting directions are similarly grouped and recorded, with the conflict type and degree labeled. These grouped records form a complete quality consistency partition label set. The quality consistency partition label set is stored and represented in a graphical structure, where nodes represent resource points in the network, and edges represent the adjacency relationships between resource points. Edge attributes include information such as the difference rate of quality assessment values ​​and directional consistency labels. The graphical structure uses an adjacency list storage format, facilitating fast traversal and query operations. This storage format supports efficient range queries and adjacency relationship retrieval, providing convenience for subsequent analysis.

[0070] The entire processing is executed within a distributed graph computing framework, leveraging parallel computing capabilities to handle large-scale network topology data. The framework employs a master-slave architecture, with the master node responsible for task scheduling and data distribution, and slave nodes executing the specific computational tasks. This architecture fully utilizes the advantages of distributed computing, improving processing efficiency. A real-time monitoring mechanism is implemented during the computation process to track the status and progress of each computing node, ensuring the successful completion of computational tasks.

[0071] The generated quality-consistent partition annotation set is periodically persisted to a distributed database, supporting historical data backtracking and trend analysis. The database adopts a columnar storage format, optimizing the storage and retrieval performance of large-scale labeled data. The data persistence process includes integrity verification and consistency checks to ensure the accuracy and reliability of the stored data. The annotation set data can be used in subsequent processing stages such as quality impact analysis and anomaly detection. During processing, the system monitors network status changes in real time and dynamically adjusts distance thresholds and difference rate calculation methods. Monitoring data includes information such as network topology changes, resource point status updates, and fluctuations in quality assessment values. Based on the monitoring results, the system adaptively updates calculation parameters and algorithm configurations to ensure the timeliness and accuracy of partition annotation results. This dynamic adjustment mechanism enables the system to adapt to changes in the network environment and maintain the reliability of evaluation results. Directional consistency classification adopts a streaming processing approach, processing and labeling new data immediately upon arrival. The streaming processing engine adopts an event-driven architecture, receiving and processing data update requests in real time. Processing results are promptly updated to the quality-consistent partition annotation set to maintain data timeliness. Data quality control measures are implemented during streaming processing, including data validity checks and error recovery mechanisms, to ensure the accuracy of processing results.

[0072] The final generated quality-consistent partition annotation set contains complete network resource quality distribution information, reflecting the quality relationship patterns between resource points in the network. This information provides crucial data for identifying quality anomalies and assessing network quality. The annotation set data is provided externally through a standard interface, supporting various query and analysis operations. The data access interface adopts a RESTful architecture, providing flexible data retrieval and manipulation capabilities.

[0073] Example 3: See Figure 4 This study focuses on analyzing the impact of network condition fluctuations and accurately identifying resource anomalies. The process begins with filtering a quality-consistent partitioning label set, extracting resource points marked as quality-consistent segments. These resource points represent areas in the network topology where performance changes and user behavior changes are aligned, making them ideal observation targets for analyzing the impact of network fluctuations. For each target resource point, the system acquires complete performance indicator time series and user behavior series within a specific time period. The time period is set to a fixed-length window, defaulting to ten minutes. Time series data is extracted from a high-precision time series database, with a strict one-second sampling interval to ensure fine data granularity. The performance indicator time series includes continuous records of key performance parameters such as bandwidth utilization, latency, and packet loss rate, while the user behavior series covers the time-series changes in behavioral indicators such as user request frequency, session duration, and data transmission volume. These series together constitute an environmental change time series set, providing a foundational dataset for subsequent rate analysis.

[0074] Based on a time series set of environmental changes, the system calculates the rate of change between consecutive time points. The performance degradation rate is calculated by analyzing the decrease in performance indicators between adjacent sampling points, specifically representing the negative change in performance indicators per unit time. The behavior change rate is determined by measuring the change in user behavior indicators within the same time interval, including both positive growth and negative decrease patterns. The rate calculation employs a discrete difference method, and considering the smoothness requirements of the time series, a sliding window averaging process is implemented to suppress noise interference. During the calculation, the system compares the performance degradation rate and the behavior change rate under the same load conditions. Load conditions are quantified using resource utilization indicators; the same load refers to resource utilization being within the same preset percentage range, such as a medium-high load range of 70%-80%. This load alignment mechanism eliminates the interference of differences in resource usage intensity on the rate comparison.

[0075] By jointly analyzing two types of indicators—performance degradation rate and behavior change rate—the system identifies the comprehensive impact of load changes. This analysis employs a multivariate correlation model to examine the collaborative change patterns and mutual constraints among the rate indicators. The degree of impact is quantified using the following formula:

[0076]

[0077] in: Indicates the impact value. Weighting coefficients representing the rate of performance degradation. This indicates the rate of performance degradation. Weighting coefficients representing the rate of change in behavior. This represents the rate of change of behavior. The weighting coefficients are dynamically adjusted based on real-time load levels; the weight of the performance degradation rate increases under high load conditions, while the weight of the rate of change of behavior increases under low load conditions. The calculation results form a sequence of impact values ​​for each resource point, which are then integrated to establish a complete analysis of the network's impact.

[0078] The process of identifying outliers employs a multi-condition joint screening mechanism. First, based on the impact analysis results, resource points whose impact response values ​​exceed the average response baseline are extracted. The average response baseline is calculated by statistically analyzing historical impact values ​​using a moving average algorithm with a rolling time window; the window length is set to thirty minutes by default. Simultaneously, resource points located in quality conflict zones are screened, as these areas exhibit a characteristic of performance changes and user behavior changes not moving in the same direction. The results of this dual screening form a candidate outlier set.

[0079] For candidate anomalies, the system extracts their continuous resource usage records in chronological order, generating a continuous resource record set. This dataset is extracted from resource monitoring logs and includes core fields such as millisecond-level timestamps, resource load values ​​(e.g., CPU utilization, memory usage), and network traffic values. After accessing the continuous resource record set, the system calculates the load change ratio and traffic change ratio between adjacent time points. The load change ratio is defined as the ratio of the current load value to the load value of the previous sampling point, and the traffic change ratio is calculated using the same method. The calculation results form a load change ratio sequence and a traffic change ratio sequence, which together constitute the resource fluctuation change dataset.

[0080] Based on the resource fluctuation dataset, the system extracts network condition fluctuation data and user fluctuation data for the corresponding time periods. Network condition fluctuation data includes underlying network metrics such as network bandwidth change rate and latency jitter, while user fluctuation data includes application-layer metrics such as user access volume change rate and request frequency fluctuation rate. Anomaly detection employs a hierarchical threshold system: First, it checks whether the load change ratio and traffic change ratio simultaneously exceed the set fluctuation identification thresholds. These thresholds are set differently based on resource type, with a default threshold of 1.5 for the load change ratio and 2.0 for the traffic change ratio. Second, it verifies whether both network condition fluctuations and user fluctuations exceed the disturbance judgment threshold, which is set based on steady-state network statistics, typically twice the historical standard deviation. When a resource point simultaneously meets both of these conditions, its corresponding time point is marked as an anomaly. All anomaly information is integrated to form a resource anomaly set, including core attributes such as anomaly location coordinates, anomaly occurrence time, performance fluctuation characteristics, and load / traffic fluctuation ratios.

[0081] Time-series data is consumed in real-time via a distributed stream processing engine, and the rate calculation module is deployed as a parallel processing unit. Anomaly detection thresholds are dynamically refreshed, automatically updating baseline values ​​hourly based on network operating status. Events marked as anomalies trigger real-time alarms, and the entire set of anomalies is stored in a time-series analysis database, supporting historical anomaly pattern backtracking and network vulnerability assessment. The resource fluctuation dataset uses a columnar storage format to optimize query efficiency, and the anomaly determination logic is implemented through a rule engine, supporting dynamic hot updates of threshold parameters. The final generated set of resource anomalies provides core input for the network quality assessment report; the spatiotemporal distribution characteristics of the anomalies reflect potential fault areas and performance bottlenecks in the network.

[0082] Example 4: Focusing on adaptive optimization of the generation and evaluation process of network resource quality assessment reports. This process begins with in-depth analysis of the resource anomaly set. The system acquires the identification information and detailed attribute data of all anomalies in the set, including core fields such as anomaly location coordinates, timestamps, performance fluctuation characteristics, and load-traffic fluctuation ratio. For each anomaly, the system calculates a joint risk assessment value, which is derived through a weighted sum of multi-dimensional indicators. The calculation model integrates three core parameters: impact response value, quality assessment value, and directional consistency label, with weight allocation dynamically set based on the network risk model. The calculation results form a joint risk assessment value sequence, serving as the basis for subsequent screening.

[0083] Based on this sequence, the system performs multi-condition joint screening: extracting points whose impact response values ​​exceed the impact response risk threshold, which is derived by training a machine learning model on historical anomaly data; screening points whose quality assessment values ​​are lower than the quality benchmark value, which is taken from the historical median of the resource quality assessment set; and marking points with location direction consistency labels as conflict zones. Points where these three conditions intersect are identified as high-risk anomalies. The system extracts the number, physical location identifier, and network partition information of these points, while simultaneously assessing their diffusion risk attributes. The diffusion risk is determined based on the network connection density and traffic mutation pattern of the anomaly: if an anomaly has high-traffic interaction with more than 20 neighboring nodes, and its traffic pattern undergoes a mutation of more than 50% when the anomaly occurs, it is marked as having diffusion risk.

[0084] Locations meeting all criteria are output in a structured format. The report adopts a hierarchical architecture: the top layer is a summary list of anomalies, the middle layer contains detailed diagnostic data, and the bottom layer provides handling recommendations. The multi-indicator joint judgment label for anomalies uses a three-digit coding system: the first digit indicates the risk level (levels 1-3), the second digit indicates the scope of impact (single point / region / global), and the last digit indicates the urgency of handling (immediate / 24 hours / observation). The complete report is packaged in JSON format and transmitted to the network operations center via an encrypted channel. During the evaluation process, the system simultaneously implements an adaptive optimization mechanism. Real-time network response data is captured through an event listening interface, including resource location offsets (such as IP changes caused by server migration) and status change information (such as alarm codes triggered by device failures). This data is recorded as a structured feedback log, containing fields such as event time, resource ID, change type, and scope of impact. The system extracts the feature vector of abnormal events from the logs, including dimensions such as event type code, duration, and number of affected nodes.

[0085] The historical evaluation case library stores feature data of abnormal cases processed within the past six months, organized using a time-series graph database. The similarity matching process calculates the cosine similarity between the current event feature vector and the case library vectors, with a matching threshold set to 0.85. When the similarity exceeds the threshold, the system generates an adaptive adjustment instruction set, including the following operations: adjusting the performance factor weights in the quality evaluation value calculation (fluctuating between 0.5 and 0.7), updating the difference rate threshold for determining quality conflict zones (dynamic range of 5%-15%), and modifying the weight coefficient allocation rules in the impact value calculation formula.

[0086] The parameter dynamic update adopts an incremental learning mode, with each adjustment not exceeding 10% of the original value. The updated parameters are immediately injected into the evaluation process, triggering a recalculation of the quality consistency partition annotation set. The recalculation process prioritizes data from the most recent hour, employing an incremental update strategy: only the network regions most affected by parameter changes (such as adjacent resource pairs with the top 10% difference rate changes) are locally recalculated, avoiding the resource consumption of full calculation. The newly generated annotation set is compared with historical versions, and areas of significant change trigger a secondary diagnostic process.

[0087] Table 1: Joint Risk Assessment Table for Resource Anomalies

[0088] Anomaly ID Position coordinates Joint Risk Value Quality assessment value Direction labels Risk level Propagation risk labeling EP-2115 10.3.5.17:80 86.7 0.32 conflict Level 3 yes EP-2116 10.3.5.18:443 73.2 0.41 conflict Level 2 no EP-2117 10.3.6.22:22 92.1 0.28 conflict Level 3 yes EP-2118 10.3.7.31:3306 68.9 0.47 Consistent Level 1 no EP-2119 10.3.8.45:8080 95.4 0.25 conflict Level 3 yes

[0089] Referring to Table 1, the report generation module integrates a visualization component to map high-risk anomalies onto a network topology map, displaying the risk distribution using a heatmap. The propagation risk path analysis employs a graph propagation algorithm to predict the physical links and logical connections through which anomalies may spread. The handling suggestion library contains combined solutions for different risk labels, such as automatically associating "isolation-diagnosis-traffic migration" processes for anomalies with level 3 risk and propagation markers. The final report includes a digital signature and timestamp to ensure the integrity of the audit trail.

[0090] Feedback log analysis employs a streaming processing framework, with event feature extraction time controlled within 200 milliseconds. The case library matching service is deployed as a distributed microservice, supporting thousands of matching requests per second. Parameter update instructions are managed through version control, generating a configuration snapshot with each update and supporting rollback in abnormal states. Recomputation task scheduling uses a priority queue, prioritizing computation tasks in critical business areas. This entire optimization mechanism forms a closed-loop control system, continuously improving evaluation accuracy.

[0091] Example 5: Constructing a distributed evaluation architecture to address the complex processing needs of heterogeneous data sources in a network environment. This architecture adopts microservice design principles, dividing into four core components: data acquisition service, data processing service, evaluation calculation service, and distributed storage service. The data acquisition service is deployed at network edge nodes, capturing raw multimodal data in real time through lightweight agents. These data sources vary significantly, including SNMP traps generated by network devices, application logs generated by user terminals, and performance metrics output by server monitoring systems, with data encoding methods covering binary streams, text formats, and structured messages.

[0092] The format parsing process is implemented through a pluggable adapter mechanism. Each adapter is designed for a specific data format; for example, the JSON adapter handles nested data structures returned by REST APIs, the XML adapter parses network configuration documents, and the CSV adapter transforms flat log files. During deep parsing, the adapters extract key fields: timestamps are uniformly formatted according to the ISO 8601 standard, data values ​​are type-converted according to metadata definitions, and data source identifiers are appended with physical location and application type tags. The parsing process maintains the time-series characteristics of the original data, strictly preserving the chronological order of data points. Anomaly handling context information is embedded as metadata, recording the data validation status, missing value handling methods, and format conversion markers during the parsing process.

[0093] The unified data transformation uses Avro serialization format for output. Its binary structure consists of three parts: a header storing the schema definition, data blocks carrying the actual values, and a tail retaining exception context metadata. The transformed data stream is injected into a Kafka message queue and partitioned into topics according to resource type. Each topic is configured with an independent consumer group, and data processing service instances act as consumers, pulling data streams in parallel. The data stream passes through a pipelined processing channel: the first-level channel performs data cleaning, filtering invalid timestamps and values ​​outside the reasonable range; the second-level channel performs feature extraction, calculating statistical indicators and derived features; and the third-level channel performs normalization operations required for quality assessment.

[0094] Multimodal data acquisition channels are configured independently based on resource characteristics. Network device channels are configured with a high sampling frequency (5 times per second) to capture millisecond-level performance fluctuations; user terminal channels employ an event-triggered mode to record user interaction behavior; application server channels are configured with adaptive sampling, increasing the acquisition density during peak load periods. The outputs of all channels converge to a central stream processing platform, ensuring that data with the same resources is routed to the same processing node via partition keys. Data processing channels are deployed in containers, with each channel running in an independent Pod, enabling cross-node communication through a service mesh. Data transmission between channels uses the gRPC protocol, with binary encoding reducing serialization overhead.

[0095] The physical layer employs RDMA network technology to accelerate data transmission between nodes, and a data locality strategy ensures that computational tasks are preferentially scheduled to data storage nodes. The logical layer implements data compression, using Delta-of-Delta encoding combined with the Zstandard compression algorithm for time-series data. The stream processing layer is configured with a backpressure mechanism to dynamically adjust the data processing rate to match the system load. Optimization parameters are dynamically injected through a configuration center, including feature calculation window size, normalization parameter thresholds, and evaluation algorithm coefficients. Parameter updates are released in versioned fashion, with a canary release mechanism to control the scope of impact of changes.

[0096] The distributed storage service employs a hybrid architecture: hot data is stored in an in-memory database, supporting real-time evaluation with millisecond-level response times; warm data is written to a columnar storage database to optimize aggregation query efficiency; and cold data is archived to an object storage system. The data partitioning strategy combines time dimensions with resource spatial distribution; time partitions are divided by hour, and spatial partitions are calculated based on resource topology location hashes. The storage layer provides a unified interface for evaluation calculation services, supporting time-range retrieval, precise resource ID queries, and topology region scanning.

[0097] The monitoring system tracks key metrics such as data processing latency, CPU load, and queue backlog, automatically triggering horizontal scaling of container instances. When peak data input exceeds a preset threshold, temporary processing nodes are automatically deployed and added to the consumer group; resources are automatically reclaimed after the peak. A service discovery mechanism maintains a dynamic node list, and clients access the service cluster through a load balancer. A failover strategy ensures automatic switching to a backup instance in case of a single point of failure, and a data checkpoint mechanism guarantees the persistence of the streaming processing state. The architecture supports cross-datacenter deployment, routing evaluation tasks to the nearest available region through global load balancing, and geographically distributed data replicas meet low-latency access requirements. Evaluation results are output to a unified message bus for downstream systems to subscribe to and use.

[0098] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0099] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for network resource quality assessment based on multi-modal data, characterized in that, The method comprises: acquiring multi-modal network data, collecting key performance indicators and user behavior data of network resources, associating data points and calculating corresponding quality evaluation values, and generating a resource quality evaluation set; wherein the acquisition of multi-modal network data comprises: collecting key performance indicator data and user behavior data of network resources, obtaining resource identifiers and corresponding timestamps, and recording the collection results as performance factors and behavior factors, and obtaining a resource data group; based on the performance factor and behavior factor data in the resource data group, respectively normalize the results, establish a corresponding relationship between the normalized results and the resource identifiers, calculate the weighted average of the normalized performance value and the normalized behavior value as the quality evaluation value, and generate a resource quality evaluation set; based on the quality evaluation values and resource identifiers in the resource quality evaluation set, sort the adjacent resource pairs according to the network topology distance, mark the quality consistent and quality conflict sections in the adjacent resource pairs, and obtain a quality consistency partition annotation set; wherein the sorting of adjacent resource pairs according to the network topology distance based on the quality evaluation values and resource identifiers in the resource quality evaluation set comprises: extracting the quality evaluation values and corresponding resource identifiers in the resource quality evaluation set, identifying the location relationship of all resource points in the network according to the resource identifiers, calling a resource point coordinate set, taking the adjacent distance threshold as the benchmark, calculating and sorting the distances of the resource points on the network, and generating an adjacent resource point distance sorting sequence; based on the adjacent resource point distance sorting sequence, calculate the quality evaluation value difference rate between each pair of adjacent resource points, and integrate to generate a quality evaluation value difference rate sequence; based on the quality evaluation value difference rate sequence, extract the performance change direction and behavior change direction between adjacent resource points, classify and label according to whether each pair of resource points is consistent in the two direction trends, record and group the sections with consistent and conflicting directions respectively, and obtain a quality consistency partition annotation set; extracting the resource points in the quality consistent section in the quality consistency partition annotation set, obtaining the performance indicator time sequence and user behavior sequence, comparing the change rate combined with the resource load data, evaluating the influence degree under the network condition fluctuation, and generating an influence degree analysis result; identify the abnormal points in the resource points with an influence value greater than the average influence benchmark value and in the quality conflict section in the influence degree analysis result, and form a resource abnormal point set; obtain all abnormal points and their detailed information in the resource abnormal point set, mark the abnormal points with potential diffusion risk, and generate a network resource quality evaluation report. 2.The network resource quality evaluation method based on multi-modal data according to claim 1, characterized in that, The resource quality evaluation set includes quality evaluation values, resource identifiers, and normalized performance indicators, the quality consistency partition annotation set specifically includes quality consistent section annotations, quality conflict section annotations, and adjacent resource pair quality evaluation value difference rates, the influence degree analysis result includes performance decline rate influence on resources, user behavior change rate influence on resources, and load influence response comparison under each network fluctuation condition, and the resource anomaly point set includes anomaly point position information, anomaly point performance amplitude characteristics, and anomaly point load and traffic fluctuation ratio. 3.The network resource quality evaluation method based on multi-modal data according to claim 1, characterized in that, The resource points located in the quality consistent section in the quality consistency partition annotation set are extracted, and performance indicator time series and user behavior sequences are obtained, including: According to the quality consistency partition annotation set, the sections marked as quality consistent are filtered, the performance data and user behavior data in each resource point time period are detected, the performance time series and behavior time series are arranged in time sequence to form an environment change time series set, and the environment change time series set is generated; Based on the environment change time series set, the performance decline rate and behavior change rate between consecutive time nodes in each resource point time series are calculated, the performance decline rate and behavior change rate are compared under the same load condition, the numerical relationship of the change amplitude under the load is identified through joint analysis of the two types of rate indicators, the influence value sequence of each resource point is integrated, and the influence degree analysis result is established. 4.The method for network resource quality evaluation based on multi-modal data according to claim 1, characterized in that, The anomaly point in the resource point with an influence value greater than the average influence reference value and located in the quality conflict section in the influence degree analysis result includes: According to the influence degree analysis result, the resource points with an influence response value greater than the average response reference value and located in the quality conflict section are filtered, the continuous resource use records of the resource points are extracted in time sequence, the resource load and traffic data corresponding to each time node are collected, and a continuous resource record set is generated; The continuous resource record set is called to extract the load value and traffic value of the resource point at two consecutive time nodes, respectively, to calculate the load change ratio and the traffic change ratio, respectively, to integrate the load change ratio sequence and the traffic change ratio sequence, and to establish a resource fluctuation change data set; Based on the resource fluctuation change data set, network condition amplitude data and user fluctuation data corresponding to the time period are extracted, it is judged whether the load change ratio and the traffic change ratio exceed the set fluctuation recognition threshold value, it is judged whether the network condition amplitude and the user fluctuation exceed the disturbance judgment threshold value at the same time, the time nodes meeting the conditions are marked as anomaly points, and a resource anomaly point set is generated. 5.The network resource quality evaluation method based on multi-modal data according to claim 1, characterized in that, The network resource quality evaluation report includes: All points in the resource anomaly point set and corresponding identification information are obtained, a joint risk judgment value of each anomaly point is calculated, and a joint risk judgment value sequence is established; Based on the joint risk decision value sequence, the point position with an impact response value greater than an impact response risk threshold, a quality evaluation value lower than a quality benchmark value, and a direction consistency label as a conflict section is screened out, the corresponding point position number, position identifier, and belonging partition are extracted, and the point position is marked as a detection anomaly and has a diffusion risk. The point position meeting the joint condition is output in a structured format according to the partition, and a network resource quality evaluation report is generated. 6.The method for network resource quality evaluation based on multi-modal data according to claim 1, characterized in that, The method further comprises: Real-time capture of network response data during resource evaluation, generating a feedback log containing resource position offset and state change information; Extracting abnormal event features from the feedback log and performing similarity matching with the historical evaluation case library to generate adaptive adjustment instructions; Based on the adaptive adjustment instructions, dynamically update the calculation parameters of the quality evaluation value; Re-inject the updated parameters into the evaluation process to recalculate the quality consistency partition label set.

7. The network resource quality assessment method based on multi-modal data according to claim 6, characterized in that, The method further comprises: Building a distributed evaluation architecture and resolving format differences of different data sources through the distributed evaluation architecture; Convert the multi-modal network data into evaluation data in a unified format; Preserve the time sequence relationship and abnormal processing context of the data during the conversion process; Inject the optimization parameters required by the evaluation process to generate an evaluation data stream that meets the distributed execution conditions. 8.The network resource quality evaluation method based on multi-modal data according to claim 1, characterized in that, The method further comprises: According to the network resource type, configure multi-modal data collection channels and data processing channels; Parallelly set the multi-modal data collection channels and data processing channels to create an efficient data evaluation link.

Images