A Service Quality Assessment Method and System Based on Multi-Source Data Fusion

Abstract

This invention relates to the field of service quality assessment technology, and discloses a service quality assessment method and system based on multi-source data fusion. The method includes acquiring service data from multiple data sources, specifically including service request time, service response latency, service error rate, and user feedback information; performing time and dimension alignment processing on these service data to form a unified formatted multidimensional dataset; performing feature clustering analysis based on the multidimensional dataset to extract target feature data for service quality assessment and anomaly indication data for dynamic monitoring; generating service quality assessment results using the target feature data, which includes service level ratings and potential risk indicators; identifying anomalies in service data updates based on the anomaly indication data, and dynamically correcting the service quality assessment results accordingly.

Description

Technical Field

[0001] This invention relates to the field of service quality assessment technology, specifically to a service quality assessment method and system based on multi-source data fusion. Background Technology

[0002] In today's rapidly developing digital services landscape, various service systems are expanding in scale and becoming increasingly complex, ranging from traditional IT operations and maintenance services to emerging cloud computing and IoT services. Users are paying increasing attention to service quality, making service quality assessment a crucial step in ensuring stable service operation and optimizing the service experience. However, the current service quality assessment process faces numerous challenges that urgently need to be addressed.

[0003] Service data sources are diverse. Different service modules, devices, and user interactions generate a large amount of data, which is stored in different systems or platforms. For example, service request time data may come from the front-end request log system, service response latency data may be collected by back-end server performance monitoring tools, service error rate data may be scattered across database error records and application logs, while user feedback information is mostly stored in customer service systems or user review platforms. These dispersed data sources make it difficult to form a unified analytical basis for service data. If evaluators want to obtain comprehensive service data, they often need to switch between multiple systems, which not only increases the complexity of data acquisition but may also lead to a lack of timeliness in the evaluation results due to time differences in data extraction.

[0004] Current technologies lack effective alignment mechanisms for processing multi-source service data. Different data sources generate data with varying time granularities; some systems record data in seconds, while others store data in minutes or even hours. This inconsistency in time dimensions can lead to biases in data correlation analysis. Furthermore, the definitions and statistical dimensions of data from different data sources are not uniform. For example, different monitoring tools may calculate "service response latency" differently; some only calculate the time it takes for the server to process requests, while others include data transmission time. This difference in dimensions makes it difficult to directly integrate multi-source data, thus affecting the accuracy of the evaluation results.

[0005] Traditional service quality assessment methods often rely on single-dimensional data or static feature analysis, making it difficult to comprehensively reflect the actual operational status of services. Some assessment methods only evaluate based on technical indicators such as service response latency or service error rate, ignoring the important subjective indicator of user feedback, resulting in a disconnect between assessment results and actual user experience. Furthermore, traditional methods lack the ability to dynamically monitor service data and identify anomalies. When service data fluctuates or becomes abnormal, the assessment results cannot be detected and adjusted in a timely manner, failing to reflect changes in service quality in real time, thus affecting the timeliness and effectiveness of service optimization decisions.

[0006] Existing assessment methods often generate evaluation results that only include simple scores, lacking effective identification and analysis of potential service risks. Various potential risks may exist during service operation. For example, the error rate of a service module may not have reached a threshold, but it shows a continuous upward trend. If such potential risks are not identified in time, it may lead to a significant decline in subsequent service quality. Traditional assessment methods cannot extract effective risk indication information from multi-source data, making it difficult to provide early warnings of potential service risks, which is detrimental to the stable operation and long-term optimization of the service system. Summary of the Invention

[0007] The purpose of this invention is to provide a service quality assessment method and system based on multi-source data fusion to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides a service quality assessment method based on multi-source data fusion, the method comprising:

[0009] The service data is obtained from multiple data sources, including service request time, service response latency, service error rate, and user feedback information.

[0010] The service data is processed for time alignment and dimension alignment to form a unified multidimensional dataset;

[0011] Based on the multidimensional dataset, feature clustering analysis is performed to extract target feature data for service quality assessment and anomaly indication data for dynamic monitoring;

[0012] The target feature data is used to generate service quality assessment results, which include service level scores and potential risk indicators.

[0013] The abnormality indication data is used to identify abnormal situations in service data updates, and the service quality assessment results are dynamically corrected based on the abnormal situations.

[0014] Preferably, the step of performing time alignment and dimension alignment on the service data to form a unified multidimensional dataset includes:

[0015] Integrate the time-series data of service request time, service response latency, service error rate, and user feedback information;

[0016] The integrated time-series data is mapped to the service instance dimension, time window dimension, and data source type dimension.

[0017] The multidimensional dataset is generated and used as input for the feature clustering analysis process.

[0018] Preferably, the feature clustering analysis based on the multidimensional dataset includes:

[0019] Convert the multidimensional dataset into a feature vector matrix;

[0020] Perform clustering operations on the feature vector matrix to output normal feature clusters and abnormal feature clusters;

[0021] The target feature data is extracted from the normal feature cluster, and the target feature data includes service stability indicators and response efficiency indicators;

[0022] The anomaly indication data is extracted from the anomaly feature cluster, and the anomaly indication data includes service interruption flags and latency deviation flags.

[0023] Preferably, generating service quality assessment results using the target feature data includes:

[0024] Calculate the service level score based on the aforementioned service stability and response efficiency indicators;

[0025] The potential risk indicators are generated by combining historical evaluation data from service instances;

[0026] Output the service quality assessment results to the service monitoring terminal.

[0027] Preferably, the step of identifying abnormal situations in service data updates based on the abnormality indication data includes:

[0028] Monitor the changing trends of the service interruption marker and latency deviation marker;

[0029] When the service interruption flag or delay deviation flag exceeds a preset threshold, the abnormal situation is determined;

[0030] The abnormal situation is input into the dynamic correction process.

[0031] Preferably, the dynamic correction of the service quality assessment result based on the abnormal situation includes:

[0032] Create a service quality assessment index, which stores the mapping relationship between service quality assessment instance identifiers and assessment metrics;

[0033] Update the evaluation metrics in the service quality evaluation index based on the aforementioned anomaly.

[0034] The service level score and potential risk indicators were recalculated using the updated service quality assessment index.

[0035] Generate revised service quality assessment results.

[0036] Preferably, the creation of the service quality assessment index includes:

[0037] Construct an evaluation indicator reference table, wherein the evaluation indicator reference table uses the service instance identifier as the key and the associated service level score and potential risk indicator as the value;

[0038] Configure an index update mechanism that triggers index reconstruction in response to the abnormal situation;

[0039] The service quality assessment index is used in the recalculation process.

[0040] Preferably, the method further includes:

[0041] Analyze the multi-source correlation in the anomaly indication data to identify cross-data source linkage anomalies;

[0042] When the linkage anomaly is confirmed, an evaluation and optimization strategy is executed, which includes adjusting the parameters of the feature clustering analysis process.

[0043] The feature clustering analysis process is re-executed based on the adjusted parameters;

[0044] The results of the re-execution are then input into the update process of the service quality assessment index.

[0045] Preferably, the execution evaluation optimization strategy includes:

[0046] Set up a service data sampling window for iterative training;

[0047] The performance metrics of the feature clustering analysis process are evaluated during the iterative training.

[0048] Optimize the granularity of clustering grouping operations based on the aforementioned performance metrics;

[0049] The optimized feature clustering analysis is then processed and re-executed.

[0050] Preferably, the present invention also includes a service quality assessment system based on multi-source data fusion and big data analysis. The system includes a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the service quality assessment method based on multi-source data fusion as described above.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] By acquiring service data from multiple data sources, covering key metrics such as service request time, service response latency, service error rate, and user feedback, this method overcomes the limitations of traditional evaluation methods that rely on a single data source. Traditional evaluations often depend on one or a few technical indicators, neglecting subjective user feedback, leading to a one-sided evaluation perspective. This method combines objective technical data with subjective user feedback, capturing various information during service operation from a more comprehensive angle. This provides a richer and more complete data foundation for subsequent evaluations, avoiding evaluation biases caused by missing or limited data, and ensuring that the evaluation results more closely reflect the actual operation of the service and the real user experience.

[0053] In the data processing stage, this method performs time and dimension alignment on the service data to form a unified multidimensional dataset, effectively solving a core challenge in the multi-source data fusion process. Multi-source data often originates from different systems or platforms, with significant differences in time granularity and data definitions. For example, the time intervals between data records from different data sources may range from seconds to hours, and the statistical ranges for the same indicator may also differ. These differences can lead to data that cannot be directly correlated for analysis, and may even produce contradictory analytical results. Time alignment, however, adjusts data with different time granularities to a unified time dimension, ensuring data comparability and correlation at the time level. Dimension alignment standardizes the definitions and statistical standards of data from various data sources, eliminating inconsistencies in data dimensions. This enables true fusion of multi-source data, forming a structurally unified and logically coherent dataset, laying a solid data foundation for subsequent feature analysis and evaluation, and reducing evaluation errors caused by differences in data formats or standards.

[0054] Feature clustering analysis is performed on a unified multidimensional dataset to extract target feature data and anomaly indication data. This process fully leverages the effective information within the multi-source fusion data. Feature clustering analysis can filter out target features that have a key impact on service quality assessment from massive amounts of multidimensional data based on the inherent correlation and feature similarity of the data. These target features can accurately reflect the core influencing factors of service quality, avoiding redundant analysis of irrelevant data during the assessment process and improving assessment efficiency. Simultaneously, the extracted anomaly indication data can capture abnormal fluctuations or potential abnormal trends in service data. Compared to traditional methods that can only identify anomalies after data exceeds a threshold, this method can detect abnormal signs in service operation earlier, providing effective data basis for subsequent dynamic monitoring and assessment correction. This allows service quality assessment to not only focus on the current state but also effectively control data change trends.

[0055] By leveraging target feature data to generate service quality assessment results that include service level ratings and potential risk indicators, the connotation and application value of the assessment results are enriched. Traditional assessment results are often presented in the form of a single rating, which can only roughly reflect the service quality level and cannot provide specific directions for service optimization. Service level ratings, on the other hand, can intuitively display the current overall quality level of the service, helping relevant personnel quickly grasp the service quality overview. Potential risk indicators can identify potential risk points in the service operation process from multi-source fusion data. For example, although the error rate of a certain service module is within the normal range, combined with the increasing trend of related complaints in user feedback, a potential risk indicator can be formed, suggesting that the module may have hidden problems. This assessment result, containing both aspects, not only makes the assessment conclusions more in-depth but also provides more specific guidance for service optimization decisions, helping staff to carry out targeted service improvement work.

[0056] This method identifies anomalies in service data updates based on anomaly indication data and dynamically corrects service quality assessment results, ensuring their real-time accuracy. During service operation, data is constantly updated, and various anomalies may occur, such as sudden increases in service response delays or a surge in negative user feedback. If assessment results cannot be adjusted in a timely manner to reflect these changes, they will become disconnected from the actual service status and lose their guiding significance. This method monitors anomalies in data updates in real time using anomaly indication data. Once an anomaly is detected, the assessment results are immediately corrected, ensuring that the assessment results always accurately reflect the current service quality status. This provides timely and reliable references for service monitoring, troubleshooting, and service optimization, helping the service system maintain a stable and efficient operating state and better meet user needs. Attached Figure Description

[0057] Figure 1This is a schematic diagram illustrating the working principle of the service quality assessment method based on multi-source data fusion described in this invention.

[0058] Figure 2 A flowchart for time alignment and dimension alignment processing;

[0059] Figure 3 The flowchart for feature clustering analysis processing;

[0060] Figure 4 A flowchart for dynamically revising service quality assessment results;

[0061] Figure 5 A flowchart optimized for multi-source correlation analysis and evaluation. Detailed Implementation

[0062] 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.

[0063] Please see Figure 1 This invention provides a service quality assessment method based on multi-source data fusion, the method comprising:

[0064] Service data from multiple data sources, including service request time, service response latency, service error rate, and user feedback, is acquired. This data undergoes time and dimensional alignment to form a unified multidimensional dataset. Feature clustering analysis is then performed on the multidimensional dataset to extract target feature data for service quality assessment and anomaly indication data for dynamic monitoring. Service quality assessment results, including service level scores and potential risk indicators, are generated using the target feature data. Anomalies in service data updates are identified based on the anomaly indication data, and the service quality assessment results are dynamically adjusted accordingly.

[0065] Example 1: See Figure 2Time alignment and dimension alignment are core steps in building a unified multidimensional dataset. The input to this process is raw service data streams from multiple independent systems. For example, service request time data might come from load balancer access logs, recorded in the format of timestamp, service instance ID, and request ID. Service response latency data might be provided by application performance monitoring tools, including request ID, response milliseconds, and success status code. Service error rate data might originate from an error tracking platform, aggregating the number of errors per minute for each service instance. User feedback information comes from the client's experience rating system, sent asynchronously as messages containing user ID, timestamp, rating value, and optional text comments.

[0066] These data sources inherently differ in collection frequency, time precision, and transmission latency. Load balancer logs might generate thousands of records per second, while user feedback might only have a few per minute. Application performance monitoring data might be granular with a five-second interval, while error tracking data might be aggregated once a minute. Time alignment processing first requires establishing a unified timeline benchmark for all data, such as Coordinated Universal Time (UTC). The timestamp in each data entry is extracted and converted to this benchmark time. For data without timestamps accurate to milliseconds, such as data only at the second or minute level, interpolation methods are used to distribute them to a finer granularity. For example, the number of errors aggregated in one minute is evenly distributed across each second within that minute, but this distribution is marked as an estimate for subsequent processing identification.

[0067] After initial timestamp standardization, the next step is data integration. Integration isn't simply putting all data into a single table; it involves association based on common keys such as request ID and service instance ID. For example, the request ID can be used to associate load balancer request records with application performance monitoring latency records. For data without a direct common key, such as user feedback unrelated to a specific request ID, time proximity matching is used. For instance, if a user feedback's timestamp is T, all requests from the same user within the time window [T-ΔT, T+ΔT] are matched, and the service instances of these requests are associated with the feedback scores.

[0068] After time-series data integration, the multi-dimensional mapping stage begins. Service instance dimension mapping assigns a unique dimension key to each service instance and records its static attributes, such as cluster affiliation, version number, and deployment region. All data related to this instance, regardless of its source, is associated through this dimension key. Time window dimension mapping divides the continuous time stream into non-overlapping fixed-interval windows, such as five-minute windows. Each window has a unique start timestamp as its key. All data records within a window are aggregated into summary data. For example, the request latency of a service instance within a window is aggregated into statistics such as average, 95th percentile, and maximum value. The number of errors is aggregated into a sum, and user ratings are aggregated into an average. Data source type dimension mapping assigns a category label to each original data source, such as "access logs," "performance metrics," "error metrics," and "user experience." Each summary data entry carries its source information to trace data lineage.

[0069] The resulting cube is a hierarchical data structure, with the top layer organized by time windows. Each time window object contains entries for all monitored service instances within that window. Each service instance entry contains all aggregated metric data for that window, along with annotations of the original data source for each metric. This dataset is typically stored in a columnar format for efficient querying and analysis. For example, it's possible to quickly retrieve the performance metric sequence for all windows of a service instance over the past hour, or obtain a snapshot of the error rate for all service instances at a specific point in time. This structured cube provides clear, well-organized, and semantically rich input for subsequent feature clustering analysis, enabling clustering algorithms to uncover patterns from multiple perspectives, including time, instance, and data source.

[0070] The entire processing flow is implemented using streaming frameworks such as Apache Flink or Spark Streaming. Data flows in in real time, and after time alignment, correlation integration, multidimensional mapping, and window aggregation, the resulting multidimensional dataset is updated in real time and pushed to the feature clustering analysis module. This design enables service quality assessment to be continuously performed based on near real-time data.

[0071] Example 2: See Figure 3 The multidimensional dataset serves as the input source, and it is organized in a three-dimensional structure according to service instances, time windows, and data source types. For example, in the service monitoring scenario of an e-commerce platform, a five-minute time window may contain data from hundreds of service instances, each instance corresponding to a set of aggregated metrics: request volume statistics from the load balancer, average response latency and P95 latency provided by application performance monitoring, the number of HTTP 500 errors recorded by the error tracking system, and the average rating value from the user feedback system.

[0072] The construction of the feature vector matrix begins by determining the vector dimensions. All metrics for each service instance within a single time window are flattened into a numerical vector. For example, the vector might contain eight dimensions: time window start timestamp (normalized to the 0-1 range), service instance ID (hash-encoded as a number), total requests, average latency, P95 latency, number of errors, average user rating, and rating sample size. Metrics with the same name from different data sources retain independent dimensions; for example, "Latency_Load Balancer" and "Latency_APM Monitoring" belong to two separate dimensions. The number of rows in the entire matrix is ​​the product of the number of service instances and the number of time windows, and the number of columns is the total number of defined feature dimensions.

[0073] The clustering operation employs an improved DBSCAN algorithm. This algorithm first calculates the standardized Euclidean distance between vectors and then applies z-score normalization to each dimension to eliminate dimensional differences. In the density parameter settings, the neighborhood radius ε is determined by analyzing the average distance distribution of historical data vectors, and the minimum sample size MinPts is set to 10% of the total number of time windows. In the e-commerce platform case, the clustering results for a certain time period might show that order service instances form a high-density core cluster, payment service instances form a medium-density cluster, while inventory service instances form discrete points due to data fluctuations.

[0074] The identification of normal feature clusters is based on the vector density and size within the cluster. When extracting service stability indicators from the core cluster, the coefficient of variation (standard deviation divided by the mean) of all vectors in the error count dimension of that cluster is calculated. This coefficient reflects the relative fluctuation of the service error rate. The response efficiency indicator is taken as the median value of the vectors in the P95 latency dimension within the cluster, representing the tail latency level in a typical scenario. Vectors of abnormal feature clusters are usually distributed outside the density threshold. The logic for generating service interruption markers is: when a vector's error count dimension value exceeds three times the standard deviation of the normal cluster in that dimension, the service instance is marked as having an interruption risk within that time window. The latency deviation marker calculates the percentage deviation of the vector's average latency dimension value from the mean of the normal cluster.

[0075] The service level rating calculation integrates multiple indicators. Service stability is weighted at 0.6, and response efficiency at 0.4. These two indicators are weighted and summed after min-max normalization to obtain a base rating of 0-100. For e-commerce payment services, an adjustment factor for user ratings is added: when a user rating is below 3 stars (out of 5), the base rating is forcibly reduced by 20%. The calculation of potential risk indicators incorporates time series analysis, taking the percentage difference between the current window's service stability indicator value and the moving average over the past 24 hours, and then overlaying it with the slope of the error rate trend line for the service instance over the past 7 days.

[0076] The output stage employs a hierarchical data structure. Each service instance generates a JSON object containing the following fields: instance ID, start and end times of the time window, service level score, potential risk level (low / medium / high), detailed indicator values ​​(including stability score and efficiency score), and an anomaly flag list (if present). These objects are grouped by time window and pushed to the monitoring terminal via a message queue. The terminal interface displays a heatmap of the service instances, with color depth representing risk level, and mouse hover revealing detailed score components.

[0077] The entire processing flow operates in micro-batch mode, triggering a computation task every five minutes. It reads the latest window data from the cube, performs a complete chain of feature vector transformation, cluster analysis, index calculation, and result output. Historical cluster centroids are cached to accelerate cluster allocation for new data. When the cluster profile coefficient of three consecutive windows falls below a threshold, the parameter tuning process is automatically triggered.

[0078] Example 3: See Figure 4 This process continuously receives anomaly indication data streams from the feature clustering analysis module. This data includes service interruption markers and latency deviation markers, each associated with a service instance identifier, a time window stamp, and numerical strength. Monitoring of service interruption markers focuses on their continuous occurrence patterns. The system maintains a sliding time window, defaulting to three consecutive monitoring periods (e.g., 15 minutes). When the total accumulated service interruption marker strength for a service instance within this window exceeds a threshold... When this occurs, an exception condition determination is triggered. The threshold is calculated as follows:

[0079] ;

[0080] in: This represents the historical average of the service outage flag intensity for this service instance over the same time period on the past 7 identical business days (e.g., 10:00 AM to 10:15 AM every Monday). This corresponds to the historical standard deviation. This dynamic threshold mechanism takes into account the periodicity of service load, avoiding false alarms caused by fixed thresholds during peak business periods.

[0081] Monitoring latency deviation markers analyzes their changing trends. The system calculates the first derivative of the latency deviation marker for each service instance, which is the difference between the marker value in the current time window and the marker value in the previous time window. When this derivative value is continuously positive and accumulates to exceed a set range, it indicates that latency is continuously deteriorating, and even if the absolute value has not reached its peak, it will be identified as a potential anomaly. This trend monitoring helps to detect gradual performance degradation, rather than just responding to sudden failures. Once an anomaly is identified, the system encapsulates it as a structured event object, which contains the anomaly type, the associated service instance identifier, the trigger time, the severity score, and related raw data references. This event object is immediately sent to the dynamic correction process.

[0082] The core of the dynamic correction process is maintaining a service quality assessment index, which is constructed in memory as a nested hash map structure. The outermost key is the service instance identifier, and its value is an embedded dictionary containing the latest assessment metrics. This dictionary stores, in key-value pairs, snapshots of the current service level score, potential risk metrics, service stability metrics, and response efficiency metrics, along with the corresponding version timestamps.

[0083] When an anomaly occurs, the correction process first locates the corresponding evaluation metric dictionary for that instance in the service quality assessment index based on the service instance identifier in the event. Then, update logic is executed according to the anomaly type. For service interruption anomalies, the system proportionally lowers the service stability metric value in the instance's dictionary based on the event's severity score. For latency deviation anomalies, the system primarily affects the response efficiency metric value. These adjustments are based on a predefined rule matrix, which defines the mapping relationship between different anomaly types, severity levels, and the correction magnitudes of various evaluation metrics.

[0084] After updating the evaluation metric dictionary, the system recalculates the service level rating and potential risk metrics using the updated metric values. The algorithm used for recalculation remains consistent with the initial evaluation module, but its input data reflects the latest anomalies. For example, for a service instance where latency deviation is caused by network jitter, its response efficiency metric will be lowered, resulting in a corresponding decrease in the recalculated service level rating, and the potential risk metric level may improve from "low" to "medium".

[0085] The revised service quality assessment results are generated. These results not only include the updated scores and metrics but also a correction log, recording the reason for the correction (i.e., the triggered exception event ID), the correction time, and the changes in metric values ​​before and after the correction, thus ensuring the transparency and traceability of the assessment process. The revised results are updated to the service quality assessment index in real time and simultaneously pushed to all service monitoring terminals that have subscribed to the service instance status, enabling operations personnel to immediately perceive dynamic changes in service status. The entire monitoring and correction cycle runs asynchronously and event-driven, ensuring that the system can respond to anomalies in near real-time.

[0086] Example 4: See Figure 5 The construction of the evaluation metric reference table is the core foundation. This table exists as a separate tablespace in the relational database, and its physical structure is designed with row storage mode to support high-frequency point queries. Each record corresponds to a snapshot of the evaluation status of a service instance at a specific point in time. The table contains the following key fields: primary key ID, service instance identifier, timestamp, service level score, potential risk level code, service stability metric value, response efficiency metric value, data version number, and record creation time. The service instance identifier and timestamp form a combined unique index, ensuring that only one valid record exists for the same instance at any given time. The potential risk level code uses an enumeration type, for example, using numbers 1, 2, and 3 to represent low, medium, and high risk, respectively. The data version number is used to implement an optimistic locking mechanism to ensure data consistency during concurrent updates.

[0087] The index update mechanism relies on an event listener architecture that subscribes to a topic called "Exception Events" in the message middleware. Whenever the dynamic correction module publishes an exception event, the listener triggers the index reconstruction process. This process first parses the exception event details from the transaction log, extracting key elements including the list of affected service instance identifiers, the time window of the exception event, the exception type, and the severity coefficient. Subsequently, the process generates a batch of Structured Query Language (SCL) statements based on these elements. These statements operate on all records in the evaluation metric reference table whose timestamps fall within the exception time window and whose service instance identifiers are listed in the affected instance list. The update logic follows a predefined set of rules; for example, for "service interruption" type exceptions, the service stability metric value of the relevant record is multiplied by a decay factor based on the severity coefficient. All update operations are completed within a single database transaction, ensuring either complete success or complete rollback. After the update is complete, the data version number is incremented, and a new timestamp is recorded.

[0088] The identification of linked anomalies is a multi-stage analysis process. The system maintains a sliding time window and continuously collects all marked anomaly indication data within that window. This data includes not only the anomaly markers themselves but also retains their complete data source context information. The analysis engine first groups these anomaly events by data source type, such as infrastructure monitoring, application performance monitoring, and end-user experience monitoring. Subsequently, the engine calculates the correlation between the occurrence of anomalies across different data source groups. The calculation is not based on simple time-of-occurrence matching but rather on overlap analysis based on event duration. For example, a "CPU utilization spike" event from infrastructure monitoring and an "increased request timeout rate" event from application performance monitoring are considered potentially correlated if their high-duration windows significantly overlap.

[0089] Table 1: Association Rules for Linked Abnormalities.

[0090]

[0091] When the identified association strength exceeds a preset confidence threshold and matches a rule pattern defined in the table, a linkage anomaly is confirmed. Upon confirmation, the system generates a linkage anomaly event, which includes the identifiers of all associated original anomaly events, a list of data source types involved, and a comprehensive severity score.

[0092] Implementing an evaluation and optimization strategy is a crucial step in addressing linked anomalies. The primary action of this strategy is adjusting the parameters of the feature clustering analysis. Specific adjustments include modifying the neighborhood search radius of the clustering algorithm and the minimum sample size for defining core points. In linked anomaly scenarios, because multiple data sources experience anomalies simultaneously, the overall density of the data distribution and the definition of outliers may change. Therefore, it is necessary to relax the definition conditions of core points to avoid incorrectly classifying normally functioning service instances as anomalies. The specific parameter adjustments are recommended using an analysis model based on historical linked anomaly cases.

[0093] Based on the adjusted parameters, the system will re-execute feature clustering analysis for the current time window and several affected historical time windows. This recalculation is localized, limited to the service instances and data time ranges directly or indirectly affected by the linked anomaly event, to control computational overhead. The re-execution process is identical to the initial clustering analysis, but uses new parameter settings to produce a revised set of feature cluster partitioning results.

[0094] The result of the re-execution is encapsulated as a data update event and input into the service quality assessment index update process. This process, in addition to updating the specific values ​​in the assessment metric reference table as in handling ordinary anomalies, includes an additional operation: recording in the extended metadata field of the relevant record that this update was caused by a recalculation triggered by a linked anomaly, and associating it with the corresponding linked anomaly event ID. This design ensures complete traceability of the evolution history of the assessment results, facilitating subsequent root cause analysis or auditing. The entire process, from event listening to index update completion, forms a closed-loop automated feedback and adjustment system capable of handling complex cross-data source anomaly scenarios.

[0095] Example 5: The service data sampling window is constructed using a stratified random sampling strategy. The system first extracts all records from the multidimensional dataset for the most recent 72 hours as a candidate pool. The candidate pool is stratified according to the business domain to which the service instances belong, for example, divided into order processing, payment clearing, and inventory management domains. Within each business domain, the records are further arranged in order of timestamps within the time window. Subsequently, the system randomly selects 5% of the service instances from each business domain and extracts complete time window data for these instances within the most recent 72 hours. This sampling method covers different business characteristics, preserves temporal continuity, and controls the size of the training data.

[0096] The iterative training process runs in an isolated computational sandbox environment. Each iteration consists of three main stages: parameter combination generation, clustering execution, and performance evaluation. The parameter combination generation stage employs a grid search strategy, with the system predefined to optimize a set of parameters and their value ranges. Core parameters include the neighborhood radius of the clustering algorithm, the minimum number of samples, and the distance metric. The search range for the neighborhood radius is dynamically determined based on the statistical quantiles of the distances between vectors in historical clustering results, while the search range for the minimum number of samples is linked to the percentage range of service instances in the sampled dataset. Each iteration selects a set of parameter combinations for testing.

[0097] The clustering execution phase converts the sampling window data into a feature vector matrix and runs the clustering algorithm using the currently tested parameter combinations. The algorithm outputs the cluster label and core point markers for each vector. The performance evaluation phase calculates two core performance metrics: the silhouette coefficient, which reflects the separation and compactness of clusters (a value closer to 1 indicates better clustering); and the intra-cluster distance ratio, which calculates the ratio of the average intra-cluster distance to the global average distance (a value below 1 indicates a compact intra-cluster structure). The system records the performance metric values ​​for each parameter combination. Parameter optimization decisions are based on the evaluation results of multiple iterations. The system maintains a parameter performance history table, accumulating records of parameter combinations and their performance metrics for each iteration. When a predetermined number of iterations is completed or the time limit is reached, the optimization engine analyzes this history table. The analysis process employs the Pareto optimality principle, selecting candidate combinations that are not comprehensively surpassed by other parameter combinations in both the silhouette coefficient and intra-cluster distance ratio dimensions. If multiple candidate combinations exist, the combination with the lowest historical usage frequency is selected to promote exploration diversity.

[0098] The optimized parameter combination is output to the configuration center for feature clustering analysis. The configuration center uses versioning management, generating a new configuration version number with each update while retaining historical versions. The system automatically pushes the new parameters to the online feature clustering analysis module, which applies the new parameters to perform clustering operations in the next processing cycle. To verify the generalization ability of the new parameters, the system performs shadow evaluations on the online clustering results within the following three time windows. Shadow evaluations use the same performance metrics, but the data comes from the complete dataset processed in real time, rather than the sampled data. If the shadow evaluation metrics outperform the old parameters, the new parameters are officially marked as a stable version; otherwise, a rollback mechanism is triggered, and the optimization process is restarted.

[0099] The metadata generated throughout the iterative training process is fully recorded, including details of the sampling window composition, the sequence of test parameter combinations, performance metrics for each iteration, the basis for the final optimization decision, and online validation results. These records are used for audit trails and also provide a reference for adjusting parameter ranges in subsequent optimization cycles. This closed-loop optimization mechanism enables feature clustering analysis to continuously adapt to changes in service data distribution, maintaining the accuracy of service quality assessment.

[0100] 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.

[0101] 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 service quality assessment method based on multi-source data fusion, characterized in that, include: Obtain service data from multiple data sources, including service request time, service response latency, service error rate, and user feedback information; The service data is processed for time alignment and dimension alignment to form a unified multidimensional dataset; Based on the multidimensional dataset, feature clustering analysis is performed to extract target feature data for service quality assessment and anomaly indication data for dynamic monitoring; The target feature data is used to generate service quality assessment results, which include service level scores and potential risk indicators. The abnormal situation in the service data update is identified based on the abnormality indication data, and the service quality assessment result is dynamically corrected based on the abnormal situation. The feature clustering analysis process based on the multidimensional dataset includes: Convert the multidimensional dataset into a feature vector matrix; Perform clustering operations on the feature vector matrix to output normal feature clusters and abnormal feature clusters; The target feature data is extracted from the normal feature cluster, and the target feature data includes service stability indicators and response efficiency indicators; The anomaly indication data is extracted from the anomaly feature cluster, and the anomaly indication data includes service interruption flags and latency deviation flags; The method further includes: Analyze the multi-source correlation in the anomaly indication data to identify cross-data source linkage anomalies; When the linkage anomaly is confirmed, an evaluation and optimization strategy is executed, which includes adjusting the parameters of the feature clustering analysis process. The feature clustering analysis process is re-executed based on the adjusted parameters; The results of the re-execution are then entered into the service quality assessment index update process.

2. The service quality assessment method based on multi-source data fusion according to claim 1, characterized in that, The process of performing time alignment and dimension alignment on the service data to form a unified multidimensional dataset includes: Integrate the time-series data of service request time, service response latency, service error rate, and user feedback information; The integrated time-series data is mapped to the service instance dimension, time window dimension, and data source type dimension. The multidimensional dataset is generated and used as input for the feature clustering analysis process.

3. The service quality assessment method based on multi-source data fusion according to claim 1, characterized in that, The process of generating service quality assessment results using the target feature data includes: Calculate the service level score based on the aforementioned service stability and response efficiency indicators; The potential risk indicators are generated by combining historical evaluation data from service instances; Output the service quality assessment results to the service monitoring terminal.

4. The service quality assessment method based on multi-source data fusion according to claim 1, characterized in that, The step of identifying abnormal situations in service data updates based on the anomaly indication data includes: Monitor the changing trends of the service interruption marker and latency deviation marker; When the service interruption flag or delay deviation flag exceeds a preset threshold, the abnormal situation is determined; The abnormal situation is input into the dynamic correction process.

5. The service quality assessment method based on multi-source data fusion according to claim 4, characterized in that, The dynamic correction of the service quality assessment results based on the abnormal situation includes: Create a service quality assessment index, which stores the mapping relationship between service quality assessment instance identifiers and assessment metrics; Update the evaluation metrics in the service quality evaluation index based on the aforementioned anomaly. The service level score and potential risk indicators were recalculated using the updated service quality assessment index. Generate revised service quality assessment results.

6. The service quality assessment method based on multi-source data fusion according to claim 5, characterized in that, The creation of the service quality assessment index includes: Construct an evaluation indicator reference table, wherein the evaluation indicator reference table uses the service instance identifier as the key and the associated service level score and potential risk indicator as the value; Configure an index update mechanism that triggers index reconstruction in response to the abnormal situation; The service quality assessment index is used in the recalculation process.

7. The service quality assessment method based on multi-source data fusion according to claim 1, characterized in that, The execution evaluation and optimization strategy includes: Set up a service data sampling window for iterative training; The performance metrics of the feature clustering analysis process are evaluated during the iterative training. Optimize the granularity of clustering grouping operations based on the aforementioned performance metrics; The optimized feature clustering analysis is then processed and re-executed.

8. A service quality assessment system based on multi-source data fusion and big data analysis, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the service quality assessment method based on multi-source data fusion as described in any one of claims 1 to 7.

Images