A method, device and medium for evaluating network experience status

By constructing candidate association pairs and performing consistency checks and weighted fusion, the accuracy problem of network experience status evaluation in multi-source data environment is solved, and accurate evaluation of underlying network status and scene experience status is achieved.

CN123123210APending Publication Date: 2026-11-17JUQING NETWORK TECH (JINAN) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611354092.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-02
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing systems struggle to accurately assess user-facing network experience, especially in multi-source data environments where issues such as scattered data sources, duplicate records, inconsistent versions, and missing fields lead to high computational and storage resource consumption and distorted assessment results.

Method used

By generating range descriptions, extracting network measurement records and network experience index data, constructing candidate association pairs and performing consistency checks and association credibility calculations, performing hierarchical processing based on field availability masks and sample credibility, and combining network context for weighted fusion, a consistency relationship between the underlying network state and the scene experience state is generated.

Benefits of technology

In a multi-source data environment, reliable data pairings are selected to avoid the influence of low-quality samples, improve the accuracy of fusion state values, accurately identify the correlation between changes in the underlying network and changes in scene experience, and avoid distortion of evaluation results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123123210A_ABST
    Figure CN123123210A_ABST
Patent Text Reader

Abstract

This application discloses a method, device, and medium for evaluating network experience status, addressing the problem that existing systems struggle to accurately assess user-oriented network experience status. The method includes: extracting network measurement records and network experience index data from different data sources based on a range description; constructing candidate association pairs based on the correlation between the two; performing consistency checks and association credibility calculations on the candidate association pairs; dividing the candidate association pairs into credible samples of different credibility levels based on the obtained field availability mask and sample credibility; grouping the credible samples based on network context; weighting and fusing the network test index values ​​within each group based on the field availability mask and different credibility levels to obtain a fused state value for each group; generating the underlying network state and scenario experience state based on the fused state value, determining the consistency relationship between the two, and generating a network experience status evaluation result based on the consistency relationship.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of network performance measurement technology, and in particular to a method, device and medium for evaluating network experience status. Background Technology

[0002] With the diversification of internet services, users' perception of network experience no longer depends solely on network speed test results, but is influenced by multiple factors such as transmission capacity, interaction quality, stability, and latency jitter. Network experience status assessment has become an important technical means for operators to optimize network quality and pinpoint user experience bottlenecks.

[0003] Traditional network experience assessment methods typically involve directly collecting network measurement records and performing statistical analysis based on the collected metrics to determine network status. In real-world networks, measurement data sources are dispersed, with differences in time, field definitions, and formats. Furthermore, the collection process commonly suffers from duplicate records, version inconsistencies, and missing fields. In such multi-source data environments, existing methods usually treat all samples as equally reliable for statistical and fusion processing, and generally calculate and output the underlying physical network status and the user experience status independently. Network experience assessment relies on a large amount of raw data, and a complete assessment process requires re-analyzing and fusing all historical data, resulting in high computational and storage resource consumption.

[0004] This contradiction is particularly prominent in real-world network speed test platforms: As platforms continuously accumulate user speed test results under different contexts, speed test records and network experience index results commonly suffer from time synchronization issues, field version differences, and interference with individual measurements due to mechanisms such as separate storage links and asynchronous reporting. Consequently, existing systems struggle to accurately assess user-facing network experience. Summary of the Invention

[0005] This application provides a method, device, and medium for evaluating network experience status, which addresses the following technical problem: existing systems struggle to accurately evaluate user-oriented network experience status.

[0006] The embodiments of this application adopt the following technical solutions: This application provides a method for evaluating network experience status. The method includes: generating a range description based on a received report request; extracting network measurement records and network experience index data from different data sources based on the range description; constructing candidate association pairs based on the correlation between the network measurement records and the network experience index data; performing consistency checks and association credibility calculations on the candidate association pairs to determine the field availability mask and sample credibility corresponding to the candidate association pairs that pass the checks; dividing the candidate association pairs that pass the checks into credible samples of different credibility levels according to the field availability mask and sample credibility; grouping the credible samples based on the network context; weighting and fusing the network test index values ​​within each group based on the field availability mask and different credibility levels to obtain the corresponding fused state value for each group; generating the underlying network state and scene experience state based on the fused state value, determining the consistency relationship between the two, and generating a network experience status evaluation result based on the consistency relationship.

[0007] In one implementation of this application, candidate association pairs are constructed based on the correlation between network measurement records and network experience index data. Consistency checks and association credibility calculations are then performed on these candidate association pairs. Specifically, this includes: detecting whether there is a common session identifier between the network measurement records and the network experience index data; if a common session identifier exists, pairing the network measurement records and network experience index data corresponding to the same session identifier to obtain candidate association pairs; performing consistency checks on the candidate association pairs based on a first preset detection dimension set, and performing multi-dimensional weighted summation on the candidate association pairs that pass the consistency check based on the weights corresponding to each detection dimension to obtain association credibility; constructing a bipartite graph between the network measurement records and the network experience index data using association credibility as edge weights; performing maximum weight matching to obtain matching edges, and determining the matching edges whose association credibility is not less than a first threshold and whose difference between this association credibility and the association credibility of the second-highest candidate association pair is not less than a ambiguity resolution threshold as candidate association pairs that pass the check.

[0008] In one implementation of this application, after detecting whether there is a common session identifier between network measurement records and network experience index data, the method further includes: in the absence of a common session identifier, determining the consistency score of candidate association pairs on each dimension based on a second preset detection dimension; adaptively normalizing the preset weights based on the available flags of each dimension, and obtaining the association credibility by weighted averaging the consistency scores based on the normalized weights.

[0009] In one implementation of this application, determining the field availability mask and sample credibility of the candidate association pairs that have passed the verification specifically includes: performing availability verification on each field in the candidate association pairs that have passed the verification based on preset verification rules, and generating an availability mask for each field based on the verification results; wherein the availability mask is used to identify the validity of each field; obtaining the credibility components of the candidate association pairs; wherein the credibility components include at least association credibility, original data quality, measurement process quality, and network experience index confidence; and weighting the credibility components according to their respective availability flags and weights to obtain the sample credibility.

[0010] In one implementation of this application, trusted samples are grouped based on network context, and the network test index values ​​within each group are weighted and fused based on field availability masks and different trust levels to obtain the corresponding fusion state value for each group. Specifically, this includes: grouping trusted samples based on network context; for each group of trusted samples, normalizing the network test index values ​​within the group based on the baseline distribution of each network test index within the group; determining the fusion weight of each trusted sample based on sample trust, time decay weight, and anomaly handling weight; determining the field participation range of each trusted sample based on field availability masks and the trust level of each trusted sample; and weighting and fusing each index based on the fusion weight, the normalized network test index values, and the field participation range to obtain the corresponding fusion state value for each index.

[0011] In one implementation of this application, based on the fused state value, a bottom-level network state and a scene experience state are generated respectively, and the consistency relationship between the two is determined. Specifically, this includes: constructing a bottom-level network state vector based on the fused state value; wherein the bottom-level network state vector includes at least a transmission capability state, an interaction quality state, and a stability state, and the bottom-level network state vector corresponds to a field availability mask and a bottom-level state confidence score; projecting the bottom-level network state vector into a bottom-level support score for the target scene based on the bottom-level network state vector and the indicator dependency weights related to the target scene; obtaining the scene score and scene confidence score corresponding to the target scene; and jointly determining the bottom-level support score and scene score based on the bottom-level state confidence score, scene confidence score, and indicator coverage to obtain the consistency relationship corresponding to the target scene.

[0012] In one implementation of this application, the underlying support score and scene score are jointly determined based on the underlying state confidence, scene confidence, and indicator coverage to obtain the consistency relationship corresponding to the target scene. Specifically, this includes: when any one of the underlying state confidence, scene confidence, or indicator coverage is less than a preset threshold, the consistency relationship is determined to be insufficient evidence and does not form a strong conclusion; the underlying support score and scene score are compared with each preset state threshold, and the type of consistency relationship is determined based on the comparison results; the type includes at least the following: same-direction anomaly, scene deviation, scene tolerance, same-direction normal, and transitional state; for consistency relationships determined to be in a transitional state, if the preset downward trend condition is met within a consecutive preset number of windows, it is marked as a same-direction downward trend; if the preset score difference condition is met within a consecutive preset number of windows, a stable scene deviation prompt or a scene tolerance prompt is output.

[0013] In one implementation of this application, after generating the network experience status evaluation result based on the consistency relationship, the method further includes: generating a status fingerprint for each time window and its corresponding network context based on the consistency relationship; when new network measurement records or network experience index data arrive, determining the time window and network context to which the new data belongs, and updating the status fingerprint of the time window; generating a fact fingerprint based on the updated status fingerprint; wherein the fact fingerprint is used to identify the content of the fact unit; comparing the updated fact fingerprint with the fact fingerprint before the update; if the fact fingerprint has not changed, reusing the existing report paragraph; if the fact fingerprint has changed, regenerating the report paragraph corresponding to the changed fact fingerprint.

[0014] This application provides a network experience status assessment device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enable the at least one processor to: generate a range description based on a received report request, and extract network measurement records and network experience index data from different data sources based on the range description; construct candidate association pairs based on the correlation between the network measurement records and the network experience index data, perform consistency checks and correlation credibility calculations on the candidate association pairs, and determine the field availability mask and sample credibility corresponding to the candidate association pairs that pass the checks; divide the candidate association pairs that pass the checks into credible samples of different credibility levels according to the field availability mask and sample credibility; group the credible samples based on the network context, and perform weighted fusion of the network test index values ​​within each group based on the field availability mask and different credibility levels to obtain the fusion state value corresponding to each group; generate the underlying network state and scene experience state based on the fusion state value, and determine the consistency relationship between the two, and generate a network experience status assessment result based on the consistency relationship.

[0015] This application provides a non-volatile computer storage medium storing computer-executable instructions. These instructions are configured to: generate a range description based on a received report request; extract network measurement records and network experience index data from different data sources based on the range description; construct candidate association pairs based on the correlation between the network measurement records and the network experience index data; perform consistency checks and association credibility calculations on the candidate association pairs; determine the field availability mask and sample credibility corresponding to the candidate association pairs that pass the checks; classify the candidate association pairs that pass the checks into credible samples with different credibility levels based on the field availability mask and sample credibility; group the credible samples based on the network context; perform weighted fusion of network test index values ​​within each group based on the field availability mask and different credibility levels to obtain the corresponding fusion state value for each group; generate the underlying network state and scene experience state based on the fusion state value, determine the consistency relationship between the two, and generate a network experience state evaluation result based on the consistency relationship.

[0016] The above-mentioned technical solutions adopted in this application embodiment can achieve the following beneficial effects: This application embodiment extracts network measurement records and network experience index data from multiple data sources through range description, avoiding the introduction of irrelevant data and providing a clear and controlled data boundary for subsequent processing. By constructing candidate association pairs through correlation relationships and combining consistency checks and association credibility calculations, reliable data pairings can be screened out under complex conditions such as inconsistent multi-source data associations, repeated collections, and version differences. Simultaneously, field availability masks and sample credibility are generated, laying a quality foundation for subsequent hierarchical processing. Based on this, candidate association pairs are divided into credible samples of different credibility levels according to the field availability mask and sample credibility, enabling subsequent fusion to distinguish samples of different qualities and avoid excessive influence of low-quality samples on the results. Furthermore, credible samples are grouped based on network context, and the network test index values ​​within each group are weighted and fused using the field availability mask and different credibility levels. This allows the fusion process to simultaneously reflect sample credibility, field validity, and scenario characteristics, improving the accuracy of the fusion state value. Finally, based on the fused state values, the underlying network state and scene experience state are generated, and the consistency relationship between the two is determined. This enables accurate identification of the correlation between changes in the underlying network and changes in the scene experience, avoiding the distortion of evaluation results caused by independent judgment in traditional methods. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart of a network experience status evaluation method provided in this application embodiment; Figure 2 This application provides an overall architecture diagram of a network experience status evaluation system. Figure 3 This is a schematic diagram of the structure of a network experience status evaluation device provided in an embodiment of this application.

[0018] Figure label: 200: Network experience status assessment device; 201: Processor; 202: Memory. Detailed Implementation

[0019] This application provides a method, device, and medium for evaluating network experience status.

[0020] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0021] Figure 1 A flowchart of a network experience status evaluation method provided in this application embodiment is shown below. Figure 1 As shown, the network experience status evaluation method includes the following steps: S101. Based on the received report request, generate a range description, and extract network measurement records and network experience index data from different data sources based on the range description.

[0022] In one implementation of this application, the server receives a report request from the client and parses the user's identity from the login credentials carried in the report request. The server reads the product rule configuration to determine the start time, end time, list of allowed data sources, network type range, maximum number of samples, algorithm version number, data authorization level, and de-identification policy identifier for this report.

[0023] The above information is assembled into a scope description (Scope). The scope is stored in the form of a structured record, containing at least the following fields: user internal reference, start time, end time, list of allowed data sources, network type range, algorithm version, data authorization, and de-identification policy. This scope is written into the context container of this request. All subsequent data queries, sample calculations, state fusion, and Agent tool calls read the same scope from the context container and pass it through with the call.

[0024] The server uses the list of allowed data sources in the Scope as its entry point. It first queries the speed test data source, binding the query conditions to the user's internal reference, start time, end time, and network type range in the Scope to obtain a set of speed test sessions. Using the session identifier in the speed test session set as the key, it batch queries the Network Experience Index (NXI) data source for the corresponding Network Experience Index data and NXI reason details, and batch queries the optional context data source for the context records of the corresponding sessions.

[0025] Within the analysis service process, speed test records, network experience index data, and context records are merged according to session identifiers. Each merged record corresponds to a network measurement record and network experience index data pairing under a specific session identifier. The server performs field standardization on each merged record: mapping synonymous fields from different data sources to unified internal field names; converting client-reported latency fields to milliseconds according to conversion rules, while retaining the original field names, units, and conversion version numbers on the record; and normalizing the scores and reason details in the network experience index data using unified enumeration.

[0026] The server performs basic validity checks on each standardized record, determining the following items one by one: whether the numeric field can be parsed, whether the unit is known, whether the record time falls within the start and end time of the scope, whether the network type is within the scope's network type range, whether there are duplicate records under the same session identifier, and whether the client version and NXI version fall into the compatibility version table. The result of each check is written to the corresponding sample quality vector. If a numeric field cannot be parsed, the corresponding field's index in the sample quality vector is set to unavailable; if the versions are incompatible, the overall compatibility index is set to negative, and the record is not deleted.

[0027] S102. Based on the correlation between network measurement records and network experience index data, construct candidate association pairs, perform consistency checks and association credibility calculations on the candidate association pairs, and determine the field availability mask and sample credibility corresponding to the candidate association pairs that pass the checks.

[0028] In one implementation of this application, candidate association pairs are obtained by pairing network measurement records and network experience index data with common session identifiers. The association credibility is obtained by multi-dimensional weighted summation. A bipartite graph is constructed with the association credibility as the edge weight. Maximum weight matching is performed, and the matching edges with an association credibility not less than a first threshold and a difference between the association credibility and the second highest credibility not less than the ambiguity elimination threshold are determined as candidate association pairs that pass the test.

[0029] Specifically, the server extracts the speed test session identifier from the standardized network measurement records and network experience index data. It then compares each network measurement record with one piece of network experience index data to determine if they share the same speed test session identifier. If a network measurement record and a piece of network experience index data have the same speed test session identifier, they are paired to obtain a candidate association pair. It should be noted that the candidate association pair in this embodiment only indicates the possibility of an association between the two and is not equivalent to a reliable sample conclusion.

[0030] The server performs a multi-dimensional consistency check on each candidate association pair, and the check dimensions include: Time consistency: Whether the speed measurement time and the NXI generation or reporting time are within the allowable range; Network consistency: Whether the network access types recorded by both parties are consistent or compatible; Source consistency: Whether the client, SDK, or data source belongs to the same measurement link; Context consistency: Are there any obvious conflicts between devices, nodes, operators, or measurement environments? Version consistency: Whether the NXI result structure can interpret the speed measurement data for this version; Completeness: Whether the fields required to calculate a certain state exist.

[0031] The server calculates the association confidence level for candidate association pairs that pass the consistency check using the following formula: Qassoc_i=α1·K_i+α2·T_i+α3·N_i+α4·C_i+α5·V_i; Where K_i represents the association key consistency score, T_i represents the time consistency score, N_i represents the network type consistency score, C_i represents the context consistency score, V_i represents the version compatibility score, and α1 to α5 are the preset weights corresponding to each dimension, with the sum of the weights being 1.

[0032] In one embodiment, the scores for each dimension are determined according to the following quantification rules: Association key consistency Ki: When the network measurement record and the network experience index data have the same original session identifier, the score is 1.0; when there is a verifiable request identifier mapping between the two, the score is 0.8; when neither has a common key, it is marked as unusable, and the corresponding usability flag A_K is set to 0; when both have keys but the keys are inconsistent, it is judged as a hard conflict and excluded.

[0033] Network Consistency N_i: The score is 1.0 when both parties have the same network access type; the score is 0.7 when one party is another cellular network and the other party is 4G or 5G, which are compatible but not completely certain; the score is 0.4 when the network type of one party is unknown; the score is 0 and the network is excluded when one party is Wi-Fi and the other party is a cellular network, or when both parties are clearly 4G / 5G types but there is a conflict.

[0034] Context consistency C_i: Sub-consistency scores are calculated for anonymous terminals, operators, nodes, regions, and sources, with sub-dimension weights of 0.40, 0.20, 0.15, 0.10, and 0.15, respectively. After normalization by available fields, the scores are weighted and summed to obtain the context consistency score. The score is 1.0 when the sub-dimensions are completely identical, and 0.7 when they are compatible within the same category. Unknown dimensions are not included in the normalization. Conflicts in identity are identified as hard conflicts.

[0035] Version compatibility Vi: When both data are in the same data semantic version, the score is 1.0; when both data are in the same major version and the fields are backward compatible, the score is 0.8; when both data can be converted by a registered converter, the score is 0.6; when the version is unknown, the score is 0.4; when the field meanings conflict or cannot be converted, the score is 0 and the data is excluded.

[0036] Time consistency T_i: Calculated according to the time decay rule based on the time difference between the network measurement record and the network experience index data; in one embodiment, the time decay constant is 60 seconds, and the scores are approximately 1.000, 0.500, 0.250 and 0.031 when the time difference is 0 seconds, 60 seconds, 120 seconds and 300 seconds.

[0037] It should be noted that the embodiments of this application do not limit the above-mentioned quantization rules and values, and adjustments can be made according to actual conditions in application. The server-side scores, weights, time constants, and compatibility matrices of each detection dimension are bound to the algorithm version for storage, so that the same input yields a definite association result under the same algorithm version.

[0038] In one implementation of this application, in the absence of the common session identifier, the server takes the original speed test record r as the object and reads the record's time information, anonymous terminal identifier, network access type, data source, and version information from the standardization results. Taking the network experience index data e to be associated as the object, the server reads the corresponding time information, anonymous terminal identifier, network access type, data source, and version information from the standardization results.

[0039] Furthermore, the fields of network measurement records and network experience index data are converted to a unified standard: time is unified to UTC milliseconds; network access type is unified to one of Wi-Fi, 4G, 5G, other cellular, Ethernet, or unknown; device identifiers are converted to irreversible anonymous identifiers before entering the analysis service, and the anonymous identifiers are obtained by one-way transformation of the original device identifiers using the HMAC-SHA256 algorithm and a pre-set tenant key; client, SDK, and NXI versions are converted to version numbers in the version compatibility table; empty strings, illegal values, and uncollected fields are uniformly marked as missing and are not converted to the value 0.

[0040] The server executes candidate blocking, constructing a candidate association set based on the following conditions: the absolute value of the time difference between the network measurement record and the network experience index data is not greater than a preset maximum allowed value tau_max; the anonymous identities of both parties are compatible; the network types of both parties are compatible; the data sources of both parties are compatible; and the data versions of both parties are compatible. In one embodiment, tau_max can be 300 seconds.

[0041] The server performs hard conflict elimination on each candidate association pair. If any condition is met, the candidate association pair is directly excluded and not included in the subsequent scoring. Both parties have user or anonymous terminal identifiers, and the two are clearly inconsistent; One party uses Wi-Fi as its network access type, while the other uses a cellular network such as 4G or 5G. Both sides have clearly defined 4G or 5G types, but the types conflict; The data originates from different and non-recognizable product chains; The NXI architecture version cannot interpret this speed measurement data version; The time difference exceeds the preset maximum allowable value; Any record has been determined to be a duplicate or corrupted.

[0042] After the candidate blocking is executed, the range of data participating in subsequent calculations is limited to the same anonymous subject, similar time, and compatible network context. The complexity of full pairwise comparisons is reduced from O(m×n) to local comparisons, where m is the number of network measurement records and n is the number of network experience index data.

[0043] For candidate combination i excluded through hard conflict in the candidate association set, the association confidence is calculated using the following formula: Qassoc_i=Σ(α_d×A_i,d×S_i,d) / Σ(α_d×A_i,d); Wherein, d traverses the detection dimension set {K,T,N,C,V}, S_i,d is the score on the d-th dimension and takes a value between 0 and 1, A_i,d is the available flag of the d-th dimension, which takes a value of 1 when the field is comparable, and takes a value of 0 when neither side has the field. When the denominator is 0, the candidate combination cannot form a credible association. The denominator is renormalized to prevent unknown fields from being mistakenly treated as field conflicts.

[0044] When a common session identifier is absent, the available flag A_i,K for the association key consistency dimension is set to 0, and this dimension does not participate in the credibility calculation. The weights of the remaining dimensions are automatically normalized according to the above formula, with equivalent weights of approximately: time consistency 0.4167, network consistency 0.25, context consistency 0.25, and version compatibility 0.0833. It should be noted that this embodiment does not limit the quantification rules and values, and adjustments can be made according to actual conditions in application.

[0045] The server outputs the association credibility Qassoc_i of the candidate combinations, along with the scores of each dimension and available flags, to the matching stage.

[0046] In one implementation of this application, a bipartite graph is constructed between network measurement records and network experience index data within the same anonymous subject and time block, using association credibility as the edge weight. Maximum weight one-to-one matching is then performed to obtain matching edges. When the data volume is less than a preset size, the Hungarian algorithm is used for matching; during streaming processing, local maximum weight matching with backtracking is performed after sorting by event time.

[0047] The server performs a dual threshold determination for each matching edge: the association confidence of the matching edge must simultaneously satisfy a preset first threshold theta_assoc, and the difference between this association confidence and the association confidence of the second-highest candidate association pair on the same side must not be less than a preset ambiguity resolution threshold delta_ambiguity. In one embodiment, theta_assoc can be 0.75, and delta_ambiguity can be 0.10.

[0048] Matching edges with a correlation confidence level less than a preset first threshold are deemed to have insufficient correlation evidence; matching edges with a correlation confidence level less than the difference between the second-highest candidate correlation pair and a preset ambiguity resolution threshold are deemed to be candidate ambiguities. Neither type of matching edge is forcibly merged; they are retained as observation samples, and the reason for rejection, the number of candidates, and the highest correlation confidence level are recorded.

[0049] When multiple matching edges with the same association confidence exist, the unique optimal matching edge is selected in the following priority order: priority is given to candidate association pairs with smaller time differences between network measurement records and network experience index data; when the time differences are the same, priority is given to candidate association pairs with more comparable context fields; when the number of context fields is the same, priority is given to candidate association pairs with higher version compatibility levels; when the version compatibility levels are the same, priority is given to candidate association pairs with smaller lexicographical order of record stability identifiers.

[0050] The server will output the candidate association pairs that pass the verification, along with their association credibility and scores for each dimension, to the subsequent processing stage.

[0051] In one implementation of this application, based on preset testing rules, usability testing is performed on each field of the candidate association pairs that have passed the testing, and a usability mask is generated for each field based on the testing results; wherein, the usability mask is used to identify the validity of each field. The credibility components of the candidate association pairs are obtained; wherein, the credibility components include at least association credibility, original data quality, measurement process quality, and network experience index confidence. The credibility components are weighted and averaged according to their respective usability flags and weights to obtain the sample credibility.

[0052] Specifically, the candidate association pairs that pass the verification are read, along with their corresponding network measurement records and network experience index data. The server performs a field availability check, comparing each field in the candidate association pairs against preset verification rules. In this embodiment, the preset verification rules are set according to the field type: For core scalar fields, including download speed (download_mbps), upload speed (upload_mbps), idle latency (latency_idle_ms), and idle jitter (jitter_idle_ms), check whether they are finite non-negative numbers; fields that do not meet the requirements are marked as unavailable. For core time series fields, including downlink throughput sequence dl_series[], uplink throughput sequence ul_series[], sampling interval sampling_interval_ms, and sample timestamp, check whether the throughput sequence is non-empty and each valid sample is a finite non-negative number, and whether the sampling interval is greater than 0; fields that do not meet the requirements are marked as unusable. For the optional observation fields, including downlink latency_dl_ms, uplink latency_ul_ms, downlink jitter_dl_ms, uplink jitter_ul_ms, packet loss rate_pct, and packet loss flag has_packet_loss, check whether they pass the format and range checks; fields that do not meet the requirements are marked as unavailable.

[0053] The server generates an availability mask M_i (the field availability mask corresponding to the i-th candidate association pair) for each field based on the test results. Each field bit in the availability mask takes a value of 1 or 0. A value of 1 indicates that the field is valid and can participate in subsequent state calculations, while a value of 0 indicates that the field is unavailable and will not participate in subsequent statistics. In one embodiment, the generated availability mask is: M_i=[download=1,upload=1,latency=1,jitter=0,loss=0,videoScore=1,gamingScore=1]; Here, download represents the download speed field, upload represents the upload speed field, latency represents the latency field, jitter represents the jitter field, loss represents the packet loss rate field, videoScore represents the video scene score field, and gamingScore represents the gaming scene score field. This candidate association pair can participate in download, upload, latency, video, and game-related analyses, but cannot participate in jitter and packet loss statistics.

[0054] The server performs raw data quality calculations, which measure the reliability of each field in the network measurement records and network experience index data. For each field in a candidate association pair, the field quality q_i,f is determined according to the following rules: q_i,f is 1 if the field exists, is correctly formatted, and passes the range check; q_i,f is the valid point ratio r_valid if only a portion of the time-series field is valid; q_i,f is 0 if the field is missing or invalid. The server calculates the raw data quality Qraw_i using the following formula: Qraw_i=Σ(ω_f×A_i,f×q_i,f) / Σ(ω_f×A_i,f); Where ω_f is the preset weight of field f, A_i,f represents the usability flag of field f in the i-th candidate association pair as applicable to the current terminal and analysis purpose, and q_i,f represents the field quality of field f in the i-th candidate association pair. When a certain field cannot be reliably collected in the current terminal environment, the field is marked as inapplicable, and A_i,f is set to 0; or the final reliability is uniformly reduced by the web-side observation capability factor.

[0055] In one embodiment, the server calculates Qraw_i using the following formula: Ccore_i is set to 1 if all core scalars, uplink and downlink throughput sequences and sampling intervals meet the preset requirements; otherwise, Ccore_i is set to 0. The optional field integrity score Coptional_i is calculated according to the following rules and limited to between 0.50 and 1.00: the initial value is 1, and 0.08 is subtracted when latency_dl, latency_ul, jitter_dl, and jitter_ul are missing, and 0.10 is subtracted when packet loss is missing in the Client environment; where latency_dl represents the downlink latency field, latency_ul represents the uplink latency field, jitter_dl represents the downlink jitter field, jitter_ul represents the uplink jitter field, and Client environment represents the client software operating environment; Cseries_i = min(downlink valid point ratio, uplink valid point ratio) × sampling validity; Cseries_i represents the time sequence integrity score; the downlink valid point ratio represents the ratio of the number of valid sample points in the downlink throughput sequence to the total number of sample points; the uplink valid point ratio represents the ratio of the number of valid sample points in the uplink throughput sequence to the total number of sample points; sampling validity represents the degree of consistency between the actual sampling interval and the preset sampling interval. The closer the actual sampling interval is to the preset sampling interval, the closer this value is to 1.

[0056] Qraw_i=Ccore_i×Coptional_i×Cseries_i.

[0057] When Ccore_i is 0, the candidate association pair cannot be used as a complete and reliable sample; the fields that still meet the requirements are retained as partial samples or observational evidence under availability mask control.

[0058] Furthermore, the server performs measurement process quality calculations, which measure the stability and reliability of data acquisition during the measurement process. Specifically, the server first linearly resamples the uplink and downlink throughput sequences to a 100-millisecond sampling interval, resulting in a resampled sequence of length N. The first floor(N / 4) warm-up samples are deleted, where floor(N / 4) represents the integer part of N divided by 4. The last 75% is used as the measurement process quality window.

[0059] The server calculates d separately for each direction, where d represents the transmission direction and takes the value of downlink dl or uplink ul: spread_d=(P90_d-P10_d) / max(P50_d,0.1); dropRatio_d=count(x<0.6×P50_d) / N_d; slopeEnergy_d=[Σ|x_j-x_(j-1)| / N_d] / max(P50_d,0.1); Where, spread_d represents the dispersion index of direction d; P90_d represents the 90th percentile of the throughput sequence in direction d; P10_d represents the 10th percentile of the throughput sequence in direction d; P50_d represents the 50th percentile of the throughput sequence in direction d, i.e., the median; max(P50_d, 0.1) means taking the larger value between P50_d and 0.1 to prevent the denominator from being zero; dropRatio_d represents the proportion of low-speed samples in direction d; count(x < 0.6 × P50_d) represents the throughput sequence in direction d. The number of sample points whose values ​​are less than 0.6 times the median; N_d represents the total number of sample points in direction d within the quality window of the measurement process; slopeEnergy_d represents the fluctuation intensity of adjacent samples in direction d; Σ|x_j-x_(j-1)| represents the sum of the absolute values ​​of the differences between adjacent sample points in the throughput sequence of direction d; x_j represents the value of the j-th sample point in the throughput sequence of direction d; x_(j-1) represents the value of the (j-1)-th sample point in the throughput sequence of direction d; |x_j-x_(j-1)| represents the absolute value of the difference between adjacent sample points.

[0060] The server normalizes each metric into a penalty amount: Pspread_d=clamp((spread_d-0.60) / 0.80,0,1); Pdrop_d=clamp(dropRatio_d / 0.30,0,1); Pslope_d=clamp(slopeEnergy_d / 0.50,0,1); Where Pspread_d represents the dispersion penalty for direction d; clamp represents the operation of restricting the calculation result to between 0 and 1, taking 0 if less than 0 and 1 if greater than 1; Pdrop_d represents the penalty for the proportion of slow samples in direction d; and Pslope_d represents the penalty for the fluctuation intensity in direction d.

[0061] The server takes the worse direction from the uplink and downlink as the global value: Pspread=max(Pspread_dl,Pspread_ul); Pdrop=max(Pdrop_dl,Pdrop_ul); Pslope=max(Pslope_dl,Pslope_ul); Wherein, Pspread represents the global dispersion penalty; Pspread_dl represents the downlink dispersion penalty; Pspread_ul represents the uplink dispersion penalty; max represents the larger of the two; Pdrop represents the global low-speed sample ratio penalty; Pdrop_dl represents the downlink low-speed sample ratio penalty; Pdrop_ul represents the uplink low-speed sample ratio penalty; Pslope represents the global volatility intensity penalty; Pslope_dl represents the downlink volatility intensity penalty; and Pslope_ul represents the uplink volatility intensity penalty.

[0062] The server calculates using the following formula: Cstability=clamp(1-0.50×Pspread-0.30×Pdrop-0.20×Pslope,0.30,1); interference=0.60×Pdrop+0.40×Pspread; Cmeasurement=clamp(1-0.60×interference,0.30,1); Clength=min(Nsteady / Nmin,1); Wherein, Cstability represents the stability score, used to measure the stability of the throughput sequence; interference represents the interference level index, used to measure the level of interference encountered during the measurement process; Cmeasurement represents the measurement quality score, used to measure the overall reliability of the measurement process; Clength represents the effective duration factor; Nsteady represents the actual number of sample points within the measurement process quality window; Nmin represents the minimum number of sample points required, which is 20 in this embodiment; min(Nsteady / Nmin,1) means taking the smaller value between the ratio of Nsteady to Nmin and 1, that is, the steady-state window covers at least 2 seconds.

[0063] Server-side calculation of measurement process quality: Qprocess_i=Cstability×Cmeasurement×Clength.

[0064] When the throughput sequence is insufficient to calculate steady-state features, the server makes the process quality component unavailable and reduces the coverage of the corresponding field, and does not replace the missing time series evidence with the default full score.

[0065] When a switch between Wi-Fi and cellular networks, a foreground / background interruption resulting in a prolonged period without samples, or a test being actively cancelled is detected during the measurement process, the server sets a hard anomaly flag, and the candidate correlation pair will not be included in the core conclusion.

[0066] For specific scenarios, the server replaces the above max operation with scenario-oriented weighting: video scenarios take a downlink weight of 1 and an uplink weight of 0; video call scenarios take a downlink weight of 0.4 and an uplink weight of 0.6; AI dialogue and AI content generation scenarios take a downlink weight of 0.6 and an uplink weight of 0.4.

[0067] The server obtains the network experience index confidence score QNXI. When NXI is missing when calculating the underlying state, the server sets the available flag corresponding to QNXI to 0, and the remaining components are automatically normalized. When calculating the directly observed state of the scene, the NXI scene score and scene confidence score are necessary evidence. When they are missing, only the presumed state is output, and it is not used to impersonate the directly observed state.

[0068] The server obtains the association credibility Qassoc, which is the aforementioned association credibility calculation result.

[0069] The server uses the association confidence (Qassoc), raw data quality (Qraw), measurement process quality (Qprocess), and network experience index confidence (QNXI) as confidence components, and calculates the sample confidence (Q_i) using the following formula: Q_i=Σ(β_j×A_i,j×Q_i,j) / Σ(β_j×A_i,j); Where j iterates through {assoc, raw, process, NXI}, Q_i,j is the j-th confidence component, A_i,j is the available flag corresponding to the j-th component, and β_j is the preset weight of the j-th component. In one embodiment, the weights of each component are: association confidence is 0.30, raw data quality is 0.25, measurement process quality is 0.25, and network experience index confidence is 0.20.

[0070] S103. Based on the field availability mask and sample credibility, divide the candidate association pairs that pass the test into credibility samples of different credibility levels.

[0071] In one implementation of this application, the server classifies candidate association pairs according to the sample credibility Q_i and the availability mask M_i: When Qi ≥ 0.80, it is marked as a complete and reliable sample, which can participate in the calculation of the overall and scene states; When 0.60≤Q_i<0.80, it is marked as a partially reliable sample and only participates in the state calculation corresponding to the valid field with a label of 1 in the availability mask; When 0.40≤Q_i<0.60, it is marked as an observation sample and is not included in the core conclusion. It is used to describe insufficient data or abnormal distribution. When Q_i < 0.40, it is marked as an excluded sample and will not be included in the analysis.

[0072] When candidate association pairs have explicit identity conflicts, explicit network type conflicts, duplicate data, or unexplainable versions, the server directly marks them as excluded samples. This hard exclusion condition is not offset by other high scores.

[0073] S104. Based on the network context, trustworthy samples are grouped, and based on the field availability mask and different trust levels, the network test index values ​​within each group are weighted and fused to obtain the corresponding fusion state value for each group.

[0074] In one implementation of this application, trusted samples are grouped according to the network context, and the index values ​​within the group are normalized based on the baseline distribution of each network test index within the group; the fusion weight of each sample is determined by combining sample credibility, time decay weight and anomaly handling weight, and the field participation range of each sample is determined according to the field availability mask and credibility level; finally, the normalized index values ​​are weighted and fused according to the fusion weight to obtain the fusion state value corresponding to each index.

[0075] Specifically, the server reads the graded trusted samples and extracts network access type, operator information, node type, measurement period and network context parameters from the network measurement records corresponding to each trusted sample.

[0076] Trusted samples are grouped according to network access type, operator, node type, and measurement period. Each group of trusted samples corresponds to a network context c.

[0077] The server obtains the baseline distribution of the network context in which each group of trusted samples resides. Baseline sources include: the historical distribution of similar networks for the current user, preset technical thresholds for the same network type, statistical distribution of operators or node categories in the same region, and adaptive baselines obtained by combining multiple baselines according to their trustworthiness.

[0078] Within each group, each network test metric is normalized. Taking metric k of sample i as an example, the server calculates the normalized value Zrobust_i,k using the following formula: Zrobust_i,k=(X_i,k-Median_k,c) / (MAD_k,c+ε); Where X_i,k is the original value of index k for sample i, Median_k,c is the median of index k in context c, MAD_k,c is the absolute deviation of the median of index k in context c, and ε is a constant to prevent zero.

[0079] For metrics where lower values ​​are better, including latency, jitter, and packet loss rate, the server normalizes and unifies the direction so that higher standardized state values ​​always represent better network conditions.

[0080] The server calculates the fusion weight W_i for each sample within the same context and time window, using the following formula: W_i=Q_i×Wtime_i×Wanomaly_i; Where Q_i is the sample confidence level, Wtime_i is the time decay weight, and Wanomaly_i is the anomaly handling weight.

[0081] The time decay weight is calculated using the following formula: Wtime_i=exp(-λ×Δt_i); Where λ is the preset attenuation coefficient, and Δt_i is the time difference between the acquisition time of sample i and the current time.

[0082] The server reduces the anomaly handling weight for occasional extreme samples according to preset rules, and performs continuous anomaly judgment on continuous anomaly samples without directly deleting them from the sample set.

[0083] The server reads the field availability mask M_i and confidence level of each trusted sample to determine the field participation scope of each sample. When the corresponding bit of a field in the field availability mask of a certain indicator is 0, that indicator does not participate in the fusion of that sample; when the corresponding bit of a field in a certain indicator is 1, that indicator participates in the fusion of that sample. Excluded samples do not participate in grouping and fusion, observation samples do not participate in fusion, and some trusted samples only participate in fusion for fields marked as valid by the field availability mask.

[0084] The server performs weighted fusion on indicator k, and the fusion state value F_t,k is calculated using the following formula: F_t,k=Σ(W_i×Z_i,k×M_i,k) / Σ(W_i×M_i,k); Where Z_i,k is the normalized value of index k for sample i, M_i,k is the field availability mask corresponding to index k for sample i, and W_i is the fusion weight for sample i.

[0085] The server records the number of samples, total weight, data coverage, and result confidence for each fusion result. When the number of samples participating in the fusion is less than the preset minimum number of samples, the server outputs an insufficient evidence flag but does not output a definite status level.

[0086] This application also provides a unified core data model to support the execution of the above steps.

[0087] For each measurement, the server constructs a trusted network experience sample, denoted as S_i=<T_i,C_i,P_i,E_i,A_i,M_i,Q_i> Where T_i is the measurement time and its corresponding time window, C_i is the measurement context, including network type, operator, node, device, or source information, P_i is the underlying network measurement vector, including download, upload, latency, jitter, and packet loss, E_i is the business scenario experience vector, including scenario scores and scenario confidence, A_i is the set of experience reasons, including reason codes, associated metrics, and severity, M_i is the field availability mask, used to identify which fields are present and which are missing, and Q_i is the sample confidence. It should be noted that the field availability mask is used to preserve the semantics of missing fields, ensuring that samples only participate in the statistics corresponding to their valid fields, avoiding the misinterpretation of unmeasured fields as zero and thus contaminating other metrics.

[0088] For each time window, the server generates a network experience state, denoted as N_t=<P_t,E_t,R_t,V_t,D_t> Wherein, P_t represents the underlying network state, describing transmission capacity and interaction quality; E_t represents the scenario experience state, describing the availability and experience level of different business scenarios; R_t represents the consistency relationship between the underlying state and the scenario state; V_t represents the stability and change state; and D_t represents the data reliability of this window state. This model retains the two-layer structure and relationship between the underlying state and the scenario state, avoiding premature compression into a single total score. This allows it to identify states such as normal underlying metrics but a decline in specific scenarios, and a decline in underlying metrics but no significant impact on the scenario.

[0089] The deterministic analysis module outputs fact units and corresponding evidence units; the fact unit is represented as Fact_j=<FactId,Type,Scope,Value,Direction,Confidence,EvidenceIds> The evidence unit is represented as: Evidence_k=<EvidenceId,DataRange,Metric,Method,Result,ThresholdOrBaseline,SampleCount,Confidence,Version> The AI-generated structured statements reference one or more fact identifiers or evidence identifiers, which the server uses to perform automatic verification.

[0090] Among them, FactId is used to uniquely identify a fact unit, Type indicates the category to which the fact belongs, Scope indicates the time range, network context or scenario range in which the fact is effective, Value indicates the quantitative result or state value corresponding to the fact, Direction indicates the trend of change of the indicator or state corresponding to the fact, Confidence indicates the reliability of the fact, and EvidenceIds indicates the list of identifiers supporting the fact. EvidenceId is used to uniquely identify a piece of evidence unit, DataRange indicates the time range, network context range or sample range of the data on which the evidence is based, Metric indicates the underlying indicator or scenario indicator corresponding to the evidence, Method indicates the calculation or judgment method used to generate the evidence, Result indicates the calculation result or judgment conclusion corresponding to the evidence, ThresholdOrBaseline indicates the reference threshold or baseline value used when judging the evidence, SampleCount indicates the number of valid samples supporting the evidence, Confidence indicates the reliability of the evidence, and Version indicates the algorithm version, threshold version or baseline version number used to generate the evidence.

[0091] S105. Based on the fusion state value, generate the underlying network state and scene experience state, and determine the consistency relationship between the two, so as to generate the network experience state evaluation result based on the consistency relationship.

[0092] In one implementation of this application, the underlying network state vector in this embodiment consists of state components such as transmission capability, interaction quality, and stability. Each component is equipped with a field availability mask and an underlying state confidence level. After projecting this vector into an underlying support score according to the indicator dependency weights of the target scenario, the underlying support score and the scenario score are jointly determined based on the scenario score, scenario confidence level, and indicator coverage, thereby obtaining the consistency relationship corresponding to the target scenario.

[0093] Specifically, the server reads the corresponding fusion status value of each group and extracts the following status components from the fusion status value: download capability status p_dl, upload capability status p_ul, idle state latency status p_latency, idle state jitter status p_jitter_idle, downlink jitter status p_jitter_dl, uplink jitter status p_jitter_ul, packet loss status p_loss, queuing status p_queue, and stability status p_stability.

[0094] The server combines the above state components into the underlying network state vector P_t: P_t= <p_dl,p_ul,p_latency,p_jitter_idle,p_jitter_dl,p_jitter_ul,p_loss,p_queue,p_stability,M_P,D_P> The value range of each state component is [0, 100], and the direction is uniformly that the higher the value, the better the network status; M_P is the field availability mask, which is used to identify whether each state component is valid; D_P is the underlying state confidence.

[0095] The server decomposes the underlying network state into three state dimensions: transmission capability state, mainly composed of download capability state and upload capability state; interaction quality state, mainly composed of idle state latency state, idle state jitter state, downlink jitter state, uplink jitter state and packet loss state; and stability state, composed of the dispersion, abnormal frequency and continuous abnormal length in multiple time windows.

[0096] The effective sample size Neff is calculated for each state component using the following formula: Neff = [Σ(W_i)]² / Σ(W_i²); Where W_i is the sample fusion weight participating in the fusion of this state component.

[0097] The confidence level D_t,k of this state component is calculated using the following formula: D_t,k=weightedQuality_t,k×min(Neff / N_target,1); Wherein, weightedQuality_t,k is the weighted quality of the state component, and N_target is the preset number of target samples, which is 3 in one embodiment.

[0098] The confidence scores of each state component are combined into the underlying state confidence score D_P, and the coverage of the scene-related underlying indicators Coverage_t,s is calculated according to the proportion of the number of valid state components.

[0099] Obtain the scene score e_t,s and scene confidence D_E,t,s of the target scene, and encapsulate them as elements in the scene experience state vector E_t: <s,e_t,s,D_E,t,s,source_t,s,reasonSet_t,s> ; Where s is the target scene identifier, and its value can be video, video call, game, voice, AI dialogue or AI content generation; source_t,s is the source identifier of the scene score, and its value can be direct_observation or inferred; reasonSet_t,s is the set of deterministic reasons and evidence identifiers involved in the judgment.

[0100] When the scenario confidence D_E,t,s is lower than the preset confidence threshold or the scenario score is missing, the server infers the scenario state based on the underlying state vector, marks source_t,s as inferred, and lowers the scenario confidence.

[0101] Read the target scenario's metric dependency weights γ_s,k, and calculate the underlying support score B_t,s for each underlying metric using the following formula: B_t,s=Σ(γ_s,k×M_t,k×p_t,k) / Σ(γ_s,k×M_t,k); Where p_t,k is the value of the k-th state component in the underlying state vector, M_t,k is the field availability mask value of the state component, and γ_s,k is the dependency weight of the target scene s on the index k.

[0102] In one embodiment, the metric-dependent weight values ​​for each scenario are shown in Table 1:

[0103] Table 1 When the value of M_t,k for a certain indicator is 0, the indicator is not included in the summation of the numerator and denominator. The server only renormalizes the valid indicators with M_t,k=1, and at the same time reduces the coverage confidence of B_t,s.

[0104] The server calculates the score level of the underlying support B_t,s using the following formula: grade(B_t,s)=GOOD, when B_t,s≥70; grade(B_t,s)=TRANSITION, when 50≤B_t,s<70; grade(B_t,s)=BAD, when B_t,s<50.

[0105] The scene rating e_t,s is calculated using the same rules.

[0106] The server executes the following checks sequentially: The first step is to determine the consistency relationship R_t,s as EVIDENCE_INSUFFICIENT (insufficient evidence) when the confidence of the underlying state D_B,t,s is less than 0.60, or the confidence of the scenario D_E,t,s is less than 0.60, or the coverage_t,s is less than 0.60, and no strong conclusion is formed.

[0107] The second step involves performing a dual threshold comparison when both the underlying support score and the scenario score are not lower than 0.60 and the evidence is not insufficient. When B_t,s<50 and e_t,s<50, R_t,s=CO_DEGRADED (same direction anomaly); When B_t,s≥70 and e_t,s<50, R_t,s=SCENE_DEVIATION (Scene Deviation); When B_t,s<50 and e_t,s≥70, R_t,s=SCENE_TOLERANCE (scenario tolerance); When B_t,s≥70 and e_t,s≥70, R_t,s=CO_NORMAL (normal for both directions); In other cases, R_t,s = TRANSITION (transition state).

[0108] Third, when R_t,s=TRANSITION, the server performs a refined transition state determination: If both B_t,s and e_t,s decrease by at least 5 points relative to the previous window, and this occurs in two or more consecutive windows, they are marked as decreasing in the same direction, but are not promoted to CO_DEGRADED until both are below 50. If |B_t,s-e_t,s|≥15, then record the deviation direction as a candidate fact; when the condition is still met for two or more consecutive windows, output a stable scene deviation or scene tolerance prompt. If |B_t,s-e_t,s|<15, then no strong deviation conclusion will be output.

[0109] The server stores the consistency relationship R_t,s in the following structure: R_t,s=<relationType,B_t,s,e_t,s,D_B,t,s,D_E,t,s,Coverage_t,s,thresholdVersion,evidenceIds> ; Where relationType is the consistency relationship type, thresholdVersion is the threshold version number, and evidenceIds is the set of evidence identifiers supporting the judgment result. In one implementation of this application, the server divides the reporting period into multiple time windows and generates a state description N_t for each time window and network context.

[0110] Specifically, the server performs occasional anomaly detection for the state description of each time window. The detection criteria are: a certain indicator deviates significantly only in a single time window, returns to normal in the time windows before and after it, and the sample confidence level within that time window is lower than a preset confidence threshold or there is an anomaly in the measurement process. When the above conditions are met, the server marks the indicator as an occasional anomaly. Occasional anomalies are retained as evidence and do not directly change the long-term state level.

[0111] For each time window, a persistent anomaly determination is performed on the status description. The determination criteria are: the same indicator or the same scenario decreases in multiple consecutive time windows, or the anomaly ratio continuously exceeds a preset ratio threshold. When the above conditions are met, the server marks the indicator or scenario as persistently anomaly. Persistent anomalies are not deleted due to the use of robust statistics, and their anomaly handling weight increases with the length of the consecutive anomalies.

[0112] The occurrence times of anomalies are grouped, and recurring pattern recognition is performed. The determination criterion is that the anomaly repeatedly occurs within similar time periods. When the above condition is met, the server marks the anomaly as a periodic anomaly and includes the time condition in the factual scope of application of the anomaly.

[0113] Record the transitions of network experience status between the following states: normal, slightly abnormal, significantly abnormal, and insufficient evidence. The server records the start time, duration, recovery time, and affected scenarios for each transition.

[0114] Perform deterministic analysis to generate candidate causes. The server obtains the following criteria for judgment: underlying metrics that cause anomalies within the same time window, scene scores that decrease synchronously, reasons that recur in the NXI cause details, whether the anomaly only occurs in specific network type nodes, operators, or time periods, the consistency between the underlying state and the scene state, and whether the data credibility is sufficient to support cause judgment.

[0115] The server generates candidate causes based on the above criteria and binds these candidate causes to corresponding time windows, metrics, and NXI cause evidence. When only the scene score decreases while the underlying metrics show no obvious anomalies, the server generates a low-confidence candidate cause, indicating that the scene experience is inconsistent with the collected underlying metrics and that further detection paths or supplementary metrics are needed.

[0116] Network states, trends, anomalies, and candidate causes are transformed into fact units, and dependencies between these fact units are established. Each fact unit contains facts, evidence, scope, and confidence levels that have been deterministically calculated. The reporting context contains only fact units and their dependencies; the entire original historical record is not transmitted.

[0117] The server-side develops a presentation plan based on the reporting task type. The presentation plan first describes the overall status, then the changing trends, affected scenarios, candidate causes, and recommendations. The server reads factual details through controlled tools, which automatically append users and reporting scope, returning only the factual units for the current report.

[0118] The server outputs a collection of structured declarations. Each main declaration includes a declaration type (claimType), declaration text (text), fact identifiers (factIds), and confidence level (confidence). After the structured declarations pass validation, the server assembles them into a user-friendly natural language version.

[0119] The server performs the following six layers of validation on the AI ​​output in sequence: First-level structure validation: Check the declared JSON format, field types, required fields, enumeration values, and text length.

[0120] The second layer of reference validation checks whether the facts referenced in the declaration exist, whether they belong to the current user, whether they are in the current reporting period, and whether the declaration type is allowed.

[0121] The third layer of numerical verification: Check whether the values ​​appearing in the report come from the corresponding facts or evidence, and whether they are within the allowable formatting error range.

[0122] Fourth-level direction and range verification: When the statement contains descriptions such as rising, falling, continuous, or mainly occurring at night, check whether the direction, number of consecutive windows, and applicable range in the facts are consistent.

[0123] Fifth layer confidence check: When the confidence level of a fact is lower than a preset threshold, it is required to declare that uncertainty is used and no definitive cause is output; when the data is marked as insufficient evidence, supplementary measurement suggestions are output.

[0124] The sixth layer of security and privacy verification filters internal user identifiers, complete IP addresses, precise locations, raw keys, internal service addresses, and unauthorized device information.

[0125] The following actions are taken based on the failure type: For formatting errors, a request is made to regenerate the model along with the verification results; for statements with insufficient evidence, the statement is deleted while other compliant parts are retained; for cases where the model is unavailable, a rule-based report is generated based on the fact units and preset expression templates; for cases where the overall data is insufficient, a data quality description and supplementary measurement suggestions are output. All degradation results are tagged with the generation method, algorithm version, and confidence level. The degradation report uses the same batch of fact units, and its technical conclusions have the same data foundation as the AI ​​report.

[0126] The server stores three fingerprints for each time window and network context: window input fingerprint InputFP_t,c, state fingerprint StateFP_t,c, and fact and report dependency fingerprints FactFP and SectionFP. The window input fingerprint is used to determine whether the input or parameters of a given context window have changed; the state fingerprint is used to identify the content of the deterministic fusion state; and the fact and report dependency fingerprint is used to identify which facts and report paragraphs need to be regenerated, where t is the time window and c is the network context.

[0127] The server constructs the window input payload InputPayload_t,c, which is organized by the following fields: schemaVersion, algorithmVersion, parameterSetVersion, baselineVersion, contextId, windowStartUtcMs, windowEndUtcMs, and orderedSamples[]. Each element in orderedSamples[] must contain at least: sampleStableId, rawRecordVersion, rawContentDigest, nxiRecordVersion, nxiContentDigest, associationDecision, Qassoc, fieldMask, Qraw, Qprocess, QNXI, Qsample, sampleClass, and excludeReasonCodes.

[0128] Among them, sampleStableId is an irreversible HMAC of the database primary key and is not directly written to the device number or IP; at the same time, it saves the content summary of the original valid fields to identify error write-back; baselineVersion must participate in fingerprint calculation because changes in the context baseline will cause the same sample to obtain different normalization states; parameterSetVersion includes the weight, threshold, version compatibility table and scenario metric dependency table versions.

[0129] The server performs deterministic encoding before hashing, with the following encoding rules: characters are uniformly encoded in UTF-8; object keys are arranged lexicographically; the sample array is sorted by sampleStableId, and the evidence identifier array is sorted; time is uniformly encoded as UTCUnix millisecond integers; floating-point numbers are converted to integers with fixed precision, and scores and credibility are multiplied by 10000 and then rounded according to convention; -0 is uniformly represented as 0; illegal floating-point numbers are not written, and missing fields are explicitly written as null; strings are encoded in Unicode normalized form NFC; the order of database returns, JSON spaces, or the original order of object keys are not considered valid content. The server uses the JSONCanonicalizationScheme of RFC8785 or an internal specification encoder with equivalence rules.

[0130] Enter your fingerprint in the calculation window as follows: InputFP_t,c=HEX(SHA-256("NXI_INPUT_FP_V1\n"||CanonicalEncode(InputPayload_t,c))); CanonicalEncode represents normalized encoding, HEX represents hexadecimal string output, and fixed-field delimiter strings are used to avoid generating interchangeable fingerprints from identical byte strings used for different purposes.

[0131] The server constructs a state payload, StatePayload_t,c, which includes: InputFP_t,c, P_t, E_t, R_t, V_t, D_t, evidenceIds, and stateSchemaVersion.

[0132] The server calculates the status fingerprint using the following formula: StateFP_t,c=HEX(SHA-256("NXI_STATE_FP_V1\n"||CanonicalEncode(StatePayload_t,c))); Normal incremental change detection uses SHA-256; when cross-system verification is required to ensure that data has not been modified without authorization, the server uses the server key to calculate HMAC-SHA-256, and the key is not written into the fingerprint content.

[0133] When a new speed measurement record or NXI record arrives, the server performs an incremental update according to the following procedure: The server determines the time window and network context to which the data belongs based on the event time and network context, and only updates the trusted samples and state of the corresponding window. The server reconstructs the candidate associations and trusted sample set for the window, generating a new window input fingerprint. When the new window input fingerprint has not changed from the old fingerprint, the server reuses the existing state and ends the processing.

[0134] When the fingerprint entered in a new window changes, the server recalculates the P, E, R, V, and D of that window to generate a new state fingerprint. If the new state fingerprint remains unchanged, the server does not reconstruct the facts and reports and terminates the process.

[0135] When a new state fingerprint changes, the server only invalidates fact units that depend on the old state fingerprint. When late data crosses adjacent windows, the server recalculates not only the window to which the data belongs, but also the trend windows that depend on that window. For example, if the trend uses the most recent 7-day windows, when a certain day window changes, all 7-day trend facts covering that day window are added to the invalidation set.

[0136] Calculate the fact fingerprint (FactFP) and the report section fingerprint (SectionFP) using the following formulas: FactFP=SHA-256("NXI_FACT_FP_V1\n"||CanonicalEncode(<factType,scope,orderedStateFPs,factRuleVersion,thresholdVersion> )); SectionFP=SHA-256("NXI_SECTION_FP_V1\n"||CanonicalEncode(<orderedFactFPs,templateVersion,generationPolicyVersion,language> )); The model name or model version is stored as audit metadata; when the business requires the model to be regenerated after an upgrade, the model version is added to SectionFP; when it is only necessary to regenerate when the facts change, the model version is not added to the automatic expiration condition.

[0137] The server recalculates the affected trends, anomalies, and fact fingerprints, regenerating only the report segments that depend on the changed facts. When new data arrives, the server compares the old and new fact fingerprints: if there are no substantial changes, the original report is reused; if there are substantial changes, only the affected part of the report or the complete report is regenerated. Substantial changes include changes in status level, changes in trend direction, changes in principal cause, confidence levels crossing thresholds, or the addition of new persistent anomalies.

[0138] This application's embodiments, through field masking and sample grading, allow some missing records to still participate in effective indicator analysis without contaminating other statistical results with missing values. Compared to complete deletion or default value filling, this method improves the utilization rate of effective data and statistical accuracy. Context grouping, robust normalization, confidence weighting, and continuous anomaly identification collectively reduce misjudgments caused by mixed network environments and occasional extreme values, ensuring stable results when the same data is repeatedly calculated. The two-layer state and its consistency relationship can identify differences between underlying transmission capabilities, interaction quality, stability, and specific scenarios, providing more specific directions for locating network problems. Through time window state caching, fact dependence, and incremental updates, only the parts affected by newly added data are processed, and the model is invoked when facts change substantially, reducing database reads, repeated calculations, context transfers, and model calls. When the model fails, a rule-based report can be generated based on the same set of facts. The report saves data version, algorithm version, fact references, and generation method, allowing for the reproduction of the report's basis and supporting problem auditing.

[0139] Figure 2 This application provides an embodiment of a network experience status evaluation system architecture diagram, as shown below. Figure 2 As shown, users or business systems initiate requests to create reports, query status, and read reports through the interface layer and task control layer. This layer handles identity authentication, time range control, reporting tasks, version and idempotency, and permission isolation. Data is then extracted from network measurement data sources (speed / latency / jitter / packet loss / environment, etc.) and SEI data sources (scenario scoring / confidence / cause details, etc.). Subsequently, the data enters the trusted data layer for standardization, multi-dimensional consistency verification, missing data labeling, sample confidence calculation, and sample classification. The network state layer then performs context normalization, robust aggregation, two-layer state derivation, trend / anomaly analysis, and cause candidate identification, forming fact units, evidence units, scope of application, confidence, state version, and dependencies in the evidence fact layer. Finally, the constrained expression layer performs agent planning, structured generation, evidence binding, consistency verification, and degradation output, generating a structured report and readable text. It also supports state caching, affected time window updates, and incremental update mechanisms for traceable auditing.

[0140] Figure 3 This is a schematic diagram of the structure of a network experience status evaluation device provided in an embodiment of this application. Figure 3As shown, the network experience status assessment device 200 includes: at least one processor 201; and a memory 202 communicatively connected to the at least one processor 201; wherein the memory 202 stores instructions executable by the at least one processor 201, the instructions being executed by the at least one processor 201 to enable the at least one processor 201 to: generate a range description based on a received report request, and extract network measurement records and network experience index data from different data sources based on the range description; construct candidate association pairs based on the correlation between the network measurement records and the network experience index data, and perform consistency checks on the candidate association pairs. The process involves verifying and calculating the credibility of candidate association pairs, determining the field availability mask and sample credibility corresponding to the candidate association pairs that pass the verification, and classifying them into credible samples with different credibility levels based on the field availability mask and sample credibility. The credible samples are then grouped based on the network context, and the network test metric values ​​within each group are weighted and fused according to the field availability mask and different credibility levels to obtain the fused state value for each group. Based on the fused state value, the underlying network state and scene experience state are generated, and the consistency relationship between them is determined. A network experience state evaluation result is then generated based on this consistency relationship.

[0141] This application provides a non-volatile computer storage medium storing computer-executable instructions. These instructions are configured to: generate a range description based on a received report request; extract network measurement records and network experience index data from different data sources based on the range description; construct candidate association pairs based on the correlation between the network measurement records and the network experience index data; perform consistency checks and association credibility calculations on the candidate association pairs; determine the field availability mask and sample credibility corresponding to the candidate association pairs that pass the checks; classify the candidate association pairs that pass the checks into credible samples with different credibility levels based on the field availability mask and sample credibility; group the credible samples based on the network context; perform weighted fusion of network test index values ​​within each group based on the field availability mask and different credibility levels to obtain the corresponding fusion state value for each group; generate the underlying network state and scene experience state based on the fusion state value, determine the consistency relationship between the two, and generate a network experience state evaluation result based on the consistency relationship.

[0142] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments for devices and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0143] The above descriptions are merely embodiments of this application and are not intended to limit the scope of this application. For those skilled in the art, various modifications and variations can be made to the embodiments of this application. These modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions in the embodiments of this application.

Claims

1. A network experience state evaluation method, characterized in that, The method includes: Based on the received report request, a range description is generated to extract network measurement records and network experience index data from different data sources based on the range description; Based on the correlation between the network measurement records and the network experience index data, candidate association pairs are constructed, and consistency checks and association credibility calculations are performed on the candidate association pairs to determine the field availability mask and sample credibility corresponding to the candidate association pairs that pass the checks. Based on the field availability mask and sample credibility, the candidate association pairs that pass the test are divided into credibility samples of different credibility levels; The trusted samples are grouped based on the network context, and the network test index values ​​within each group are weighted and fused based on the field availability mask and different trust levels to obtain the corresponding fusion state value for each group. Based on the fusion state value, the underlying network state and scene experience state are generated, and the consistency relationship between the two is determined, so as to generate the network experience state evaluation result based on the consistency relationship.

2. The network experience status evaluation method of claim 1, wherein, The process of constructing candidate association pairs based on the correlation between the network measurement records and the network experience index data, and performing consistency checks and association credibility calculations on the candidate association pairs specifically includes: Detect whether there is a common session identifier between the network measurement records and the network experience index data; In the presence of a common session identifier, the network measurement records corresponding to the same session identifier are paired with the network experience index data to obtain the candidate association pairs; Based on the first preset detection dimension set, consistency detection is performed on the candidate association pairs, and based on the weights corresponding to each detection dimension, the candidate association pairs that pass the consistency detection are summed in multiple dimensions to obtain the association credibility. Using the correlation credibility as the edge weight, a bipartite graph is constructed between the network measurement records and the network experience index data; Perform maximum weight matching to obtain matching edges, and determine the matching edges that have passed the test as candidate association pairs, provided that the association confidence is not less than the first threshold and the difference between the association confidence and the association confidence of the second highest candidate association pair is not less than the ambiguity elimination threshold.

3. The network experience status evaluation method of claim 2, wherein, After detecting whether there is a common session identifier between the network measurement record and the network experience index data, the method further includes: In the absence of the common session identifier, the consistency score of the candidate association pair on each dimension is determined based on the second preset detection dimension. The preset weights are adaptively normalized based on the available flags of each dimension, and the association credibility is obtained by weighted averaging of the consistency scores based on the normalized weights.

4. The network experience status evaluation method of claim 1, wherein, The determination of the field availability mask and sample confidence corresponding to the candidate association pairs that have passed the test specifically includes: Based on preset verification rules, usability verification is performed on each field in the candidate association pairs that pass the verification, and a usability mask is generated for each field based on the verification results; wherein, the usability mask is used to identify the validity of each field; Obtain the confidence components of the candidate association pairs; wherein, the confidence components include at least the association confidence, the original data quality, the measurement process quality, and the network experience index confidence. The confidence score of the sample is obtained by weighting the confidence score components according to their respective available flags and weights.

5. The network experience status evaluation method according to claim 1, characterized in that, The process of grouping the trusted samples based on network context, and then weighting and fusing the network test metric values ​​within each group based on the field availability mask and different trust levels to obtain the corresponding fusion state value for each group, specifically includes: Based on the network context, the trusted samples are grouped; For each group of credible samples, the values ​​of each network test metric within the group are normalized based on the baseline distribution of each network test metric within the group. The fusion weight of each reliable sample is determined based on the sample reliability, time decay weight, and anomaly handling weight. Based on the field availability mask and the credibility level of each trusted sample, the field participation scope of each trusted sample is determined; Based on the fusion weights, the normalized network test metric values, and the field participation range, each metric is weighted and fused to obtain the fusion state value corresponding to each metric.

6. The network experience status evaluation method according to claim 1, characterized in that, Based on the fusion state value, the underlying network state and scene experience state are generated respectively, and the consistency relationship between the two is determined, specifically including: Based on the fusion state value, a bottom-level network state vector is constructed; wherein, the bottom-level network state vector includes at least transmission capability state, interaction quality state, and stability state, and the bottom-level network state vector corresponds to a field availability mask and a bottom-level state confidence. Based on the underlying network state vector and the index dependency weights related to the target scene, the underlying network state vector is projected into the underlying support score of the target scene. Obtain the scene score and scene confidence level corresponding to the target scene; Based on the underlying state confidence, the scenario confidence, and the indicator coverage, the underlying support score and the scenario score are jointly determined to obtain the consistency relationship corresponding to the target scenario.

7. The network experience status evaluation method according to claim 6, characterized in that, The step of jointly determining the underlying support score and the scene score based on the underlying state confidence, the scene confidence, and the indicator coverage to obtain the consistency relationship corresponding to the target scene specifically includes: When any one of the underlying state confidence, the scenario confidence, or the indicator coverage is less than a preset threshold, the consistency relationship is determined to be insufficient evidence and no strong conclusion is formed. The underlying support score and the scene score are compared with each preset state threshold, and the type of the consistency relationship is determined based on the comparison results; wherein, the type includes at least the same-direction anomaly, scene deviation, scene tolerance, same-direction normal and transition state; For consistency relationships that are determined to be in the transition state, if the preset downward trend condition is met within a consecutive preset number of windows, then they are marked as downward trends in the same direction. If the preset difference condition is met within a consecutive preset number of windows, a stable scene deviation prompt or a scene tolerance prompt will be output.

8. The network experience status evaluation method according to claim 1, characterized in that, After generating the network experience status evaluation result based on the consistency relationship, the method further includes: For each time window and its corresponding network context, a state fingerprint is generated based on the consistency relationship; When new network measurement records or network experience index data arrive, determine the time window and network context to which the new data belongs, and update the state fingerprint of that time window; Based on the updated state fingerprint, a fact fingerprint is generated; wherein the fact fingerprint is used to identify the content of the fact unit; The updated fact fingerprint is compared with the corresponding fact fingerprint before the update. If the fact fingerprint has not changed, the existing report paragraph is reused; if the fact fingerprint has changed, the report paragraph corresponding to the changed fact fingerprint is regenerated.

9. A network experience status assessment device, characterized in that, The device includes a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to perform the method according to any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are capable of performing the method described in any one of claims 1-8.