A data processing method and system for unified representation of multi-period time-series states and closed-loop hypothesis testing.
By employing a unified anchoring and equivalent indexing method in multi-cycle time-series state processing, the problem of unified encoding of multi-dimensional state variables is solved, achieving efficient data storage and retrieval, improving the reliability and manageability of multi-processing units, and supporting cross-cycle comparison and self-correction judgment verification.
Patent Information
- Application Number
- CN202611528215.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-30
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies cannot effectively map heterogeneous multidimensional state variables of the same observed object at different time granularities to a unified, ordered, and indexable coding space, resulting in high overhead for state data storage, transmission, and retrieval, and insufficient reliability and manageability of the output of multiple processing units.
By acquiring time-series data of the observed object at multiple time granularities, calculating the multidimensional state components at each time granularity, and using the benchmark value of the reference time granularity as a unified anchor, the position components of each time granularity are rewritten as comparison relation values, jointly mapped as signed state values, and establishing an equal value index. This enables the structured opinion data output of multiple processing units and the idempotent storage of hypothesis records, as well as automatic backfilling and iterative calibration of routing weights.
It realizes storable, indexable, synthesizable and a posteriori data objects with multi-period states, significantly reducing storage and transmission overhead, improving retrieval efficiency, ensuring cross-period comparability and synthesis reliability, and supporting a posteriori verification of judgments and self-correction of weights.
Smart Images

Figure CN123115266A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of data processing technology, specifically to a data processing method and system for unified representation of multi-period time-series states and closed-loop hypothesis verification. Background Technology
[0002] Monitoring the status of complex time-series systems (such as securities market targets, corporate operations, and price transmission in the industrial chain) requires simultaneously representing, storing, retrieving, and comparing the status of the same observed object at multiple time granularities (such as monthly, weekly, daily, and hourly charts).
[0003] One existing technology is a compressed transmission scheme for market data, such as CN119094787A, CN117709300A, CN117097809A, etc. It encodes and compresses the original market data stream or market information text in order to restore the original data. The compressed object does not contain any state semantics and does not support retrieval and comparison by state semantics.
[0004] The second existing technology is the symbolic representation of time series, typically the SAX method (Lin J., Keogh E. et al., A Symbolic Representation of Time Series, with Implications for StreamingAlgorithms, ACM DMKD 2003) and its improved family (ESAX, 1d-SAX, etc.), which aggregates a single numerical time series along the time axis into segments and maps them into symbols for similarity retrieval; patent application CN122660642A also discloses a lossy compression scheme that performs symbolic aggregation approximation on the sensing time series signal and extracts the state switching points from the symbol string. However, this type of scheme has the following drawbacks: First, symbolization is applied to a single numerical sequence along the time axis, and its piecewise averaging operation does not define multiple heterogeneous dimensions (such as direction, position, and fluctuation) with different dimensions at the same time, so it cannot be transferred to multidimensional joint encoding; Second, its symbols are unordered, and the distance between symbols needs to rely on table lookup definitions, so it cannot be directly compared and sorted; Third, it does not involve the state alignment of the same observation object at multiple time granularities.
[0005] Furthermore, market state detection methods based on Hidden Markov Models (HMMs) output probabilistic hidden state classifications, which do not constitute a unified state representation that can be sorted and indexed. In existing technologies, analysis conclusions at each time granularity are represented in their respective coordinate systems (support and resistance levels for each period, and index values in their respective dimensions). The same observed object cannot be directly compared, sorted, or indexed across monthly, weekly, daily, and hourly states. State retrieval across the entire object set can only perform row-by-row floating-point range comparisons, resulting in high storage and retrieval overhead. In existing multi-unit analyses, any unit can still output opinions even when evidence is missing; conflicts between units are not classified as structured problem types, and fusion weights rely on manual setting. The judgment records produced by the analysis system also lack idempotency indicators and automatic backfilling mechanisms, making it impossible to automatically measure the quality of judgments. Summary of the Invention
[0006] The technical problem to be solved
[0007] The technical problem to be solved by this invention is: how to map the heterogeneous multidimensional state variables of the same observed object at different time granularities to a unified, ordered, and indexable coding space, so as to reduce the storage, transmission and retrieval overhead of state data; and how to improve the reliability and manageability of the output of multi-processing units and realize the posterior verification closed loop of judgment records.
[0008] Technical solution
[0009] To address the aforementioned problems, this invention provides a data processing method for multi-period time-series states, comprising: acquiring time-series data of an observed object at multiple time granularities; calculating multidimensional state components at each time granularity; using a reference value of a base time granularity as a unified anchor, rewriting the positional components of each time granularity as comparison values between the unified anchor and the corresponding boundary of that time granularity; jointly mapping the multidimensional state components at the same time granularity into a signed state value, wherein the sign of the state value carries the direction, the absolute value carries the intensity, and the order relationship between state values is consistent with the order relationship of states; and writing the state values of each time granularity into a state table and establishing an equal-value index using the observed object identifier and time as keys. Further, it also includes: multiple processing units outputting structured opinion data based on the state table, mandatory abstention for lack of evidence, structured classification of conflicts, idempotent storage of hypothetical records and automatic backfilling of future data arrivals, and constrained iterative calibration of routing weights. This invention also provides a corresponding data processing system and a computer-readable storage medium (specific solutions are detailed in the claims).
[0010] It should be noted that the joint mapping of this invention is performed along the dimension axis, which belongs to a different problem domain than the piecewise compression along the time axis of the SAX family of methods: the piecewise aggregation mean of SAX is applied to a single numerical sequence, while the mean of heterogeneous dimensions with different dimensions such as "direction × position × fluctuation" is mathematically undefined. Therefore, the mathematical framework of SAX cannot be transferred to multidimensional joint coding.
[0011] Beneficial effects
[0012] This invention transforms multi-period states from incomparable graphical and textual conclusions into storable, indexable, synthesizable, and a posteriori data objects:
[0013] 1. Storage and transmission overhead is significantly reduced. Actual measurements on the applicant's real production data (approximately 7.49 million rows, 5,571 observation objects, 1,607 observation days, multi-period state overview, columnar storage format, experimental environment: single-machine columnar storage engine, experimental date: 2026-08-27) showed: physical volume decreased from 252.4MB to 28.2MB (a reduction of 8.93 times); daily full-object incremental synchronization package (5,549 objects) decreased from 229.7KB to 22.5KB after gzip processing (a reduction of 10.2 times).
[0014] 2. Improved retrieval efficiency: Status combination retrieval has been reduced from multiple floating-point range comparisons per row to a single character equality comparison, reducing the median time for full scan from 14.3ms to 4.0ms (a 3.58x speedup); the time for full object daily status scan has been reduced from 4.9ms to 2.9ms (a 1.69x speedup); and because the status values are ordered, comparison and sorting can be performed without table lookups.
[0015] 3. Cross-period comparability: Unified anchoring and alignment eliminate inconsistencies in the coordinate systems of different periods, making the states of the same observed object directly comparable across periods within the same coding space.
[0016] 4. Lossless and reversible: The status code can be reversibly decoded back to a combination of multi-dimensional components; the decoding and re-encoding of all 20 status values in the real data are verified back and forth, and 20 / 20 are consistent.
[0017] 5. Improved reliability of synthesis: Units lacking evidence are forcibly discarded by the machine, and the synthesis results are not contaminated by opinions without evidence; conflicts are classified into structured problem types, which can be tracked and studied.
[0018] 6. Judgment is posteriorly verifiable and weights are self-correcting: Assuming idempotent record entry and automatic result backfilling, the backfilling rate is a measurable indicator (backfilling rate = number of backfilled records ÷ number of records due for backfilling); routing weights are iterated based on posterior error constraints to achieve auditable self-correction without oscillation. Attached Figure Description
[0019] Figure 1 is a structural block diagram of the data processing system provided in an embodiment of the present invention;
[0020] Figure 2 is a flowchart of the data processing method provided in an embodiment of the present invention;
[0021] Figure 3 is a schematic diagram of the state panorama table structure and encoding mapping provided in an embodiment of the present invention. Detailed Implementation
[0022] Terminology conventions
[0023] - Period: The time granularity of the time series data of the observed object. In this embodiment, five periods are used as examples: monthly (MN1), weekly (W1), daily (D1), 4-hour (H4), and 1-hour (H1).
[0024] - State: The discrete values of the multidimensional state of a certain observed object at a certain period and a certain moment, which are jointly determined by four components: trend direction, relative position, fluctuation, contraction and expansion;
[0025] - Unified Anchor: The closing price of the base period (daily chart) is used as the unified anchor. The support and resistance boundaries of each period are calculated separately. The position status is generated by the comparison relationship between the unified anchor and the boundaries of each period.
[0026] - Status code: A signed hexadecimal single character (-F to F) obtained by combining the status codes;
[0027] - Status panorama table: A data table with (observation object, date) as the key and status codes and index components of each period as columns (hereinafter referred to as "status table" in the claims, the two are synonymous);
[0028] - Processing Unit: An independent computational unit that aggregates evidence from the state panorama table according to preset rules and outputs structured opinions;
[0029] - Hypothesis record: A structured record of a single judgment, including the observed object, date, opinions of each unit, composite conclusion, and unique idempotency identifier;
[0030] - Backfill: The action of automatically writing the result label into the corresponding hypothesis record after the actual arrival of the future observation data corresponding to the hypothesis record;
[0031] - Result label: Assuming the actual trend quantification value corresponding to the end of the backfill window (the calculation method can be found in the specific implementation of S11);
[0032] - Route weight: The fusion weight of opinions from each processing unit during synthesis.
[0033] System architecture (Figure 1)
[0034] The system comprises a data layer (multi-period time-series data acquisition, indicator component calculation), an observation layer (unified anchoring alignment, multi-dimensional state joint encoding, state panorama table and equivalent index), a processing layer (parallel opinions from multiple processing units, forced abstention due to lack of evidence, conflict classification and dynamic routing), and a feedback layer (assuming idempotent record storage, automatic backfilling upon future data arrival, and restricted calibration of routing weights); the functions and relationships of each module are consistent with those described in claims 8 and 9.
[0035] Method flow chart (Figure 2)
[0036] S1 Collect price and volume time series data for each observation object for five periods according to the object list and perform integrity verification; S2 Calculate the trend direction sign, support and resistance boundaries, fluctuation level, and contraction and expansion level for each (observation object, period); S3 Take the closing price of the benchmark period as a unified anchor and compare it with the support and resistance boundaries of each period to obtain the position sub-score of each period; S4 Combine the trend direction, position sub-score, fluctuation sub-score, and contraction sub-score into a state score according to formula (1), and then convert it into a state code through monotonic reversible mapping; S5 Write the state panorama table with (observation object, date) as the key and build an equivalent index; S6 Each processing unit produces opinion scores in parallel; S7 Determine whether the evidence of each unit is complete. If it is incomplete, the unit is forced to abstain and the weight is reset to zero (S7a); S8 Determine the routing weight according to the current state type and synthesize it according to formula (2). At the same time, classify the conflicts between units into structured problem types; S9 Write the synthesis conclusion into the hypothesis record library with the deterministic idempotency identifier; S10 / S10a Listen for the arrival of future observation data; S11 Upon arrival, the result label is automatically written; S12 adjusts the weights iteratively using the rolling posterior error according to formula (3), with the single adjustment range constrained by the upper limit of the proportion, and the calibration result is written back to S8.
[0037] Determining the support and resistance boundaries (specific implementation of S2)
[0038] As an alternative implementation, the support and resistance boundaries for each period are determined using the fractal point method: within the most recent N K-lines (N=120 in this embodiment), if the highest price of a certain K-line is simultaneously higher than the highest prices of the k K-lines to its left and right (k=3 in this embodiment), then this high point is a significant high point (resistance fractal); significant low points (support fractals) are symmetrically determined; the most recent significant high point is taken as the upper boundary of this period, and the most recent significant low point is taken as the lower boundary. Those skilled in the art will know that support and resistance boundaries can also be determined by other known methods such as previous highs and lows, and areas of high trading volume, all of which fall within the scope of this invention.
[0039] Determining the trend direction sign (specific implementation of S2)
[0040] As an alternative specific implementation, the trend direction symbol d is determined by the rise or fall of the latest closing price of the current period relative to the closing price of its previous period: +1 is taken for a rise, -1 is taken for a fall, and the direction symbol at the previous moment is inherited for a flat state. Those skilled in the art will appreciate that the trend direction can also be determined by other known trend determination methods such as the slope of the moving average, the positional relationship between the price and the reference moving average, all of which fall within the protection scope of the present invention.
[0041] Bucketing rule for position sub-score (specific implementation of S3)
[0042] Let the unified anchor be P, the upper boundary of a certain period be Ru, and the lower boundary be Rl, the relative position r=(P−Rl) / (Ru−Rl) is defined. The position sub-score v_pos takes an integer from 0 to 8, and the semantics of the nine buckets are as follows:
[0043] The specific determination of the nine bucket semantics is as follows: Bucket 0, deeply breaking below the lower boundary, P<Rl×0.98; Bucket 1, breaking below the lower boundary, Rl×0.98 ≤ P<Rl; Bucket 2, below the vicinity of the lower boundary, r ∈ [0, 0.125); Bucket 3, the lower quarter of the interval, r ∈[0.125, 0.25); Bucket 4, the lower half of the interval, r ∈ [0.25, 0.5); Bucket 5, the upper half of the interval, r ∈ [0.5,0.75); Bucket 6, the upper quarter of the interval, r ∈ [0.75, 0.875); Bucket 7, above the vicinity of the upper boundary, r ∈ [0.875,1); Bucket 8, breaking through the upper boundary, P ≥ Ru.
[0044] Joint encoding and mapping (specific implementation of S4)
[0045] All components are dimensionless integers and can be added with the same dimension:
[0046] s = d × (v_pos + v_vol + v_con) ……Equation (1)
[0047] Where d∈{−1,+1} is the trend direction symbol, v_pos∈[0,8] is the position sub-score, v_vol∈[0,4] is the volatility sub-score (five levels according to the historical quantile of the volatility of the current period), and v_con∈[0,3] is the contraction sub-score (four levels according to the historical quantile of the price amplitude of the current period, i.e., the normalized value of the difference between the highest price and the lowest price in the period relative to the unified anchor). The state sub-score s is an integer in the interval [−15,+15], with a total of 31 values.
[0048] The monotonically reversible mapping from state s to state code h is explicitly defined as: h = sign(s) × hex(|s|), where hex is a hexadecimal uppercase character (0-9, AF), and h = "0" when s = 0. The complete mapping is: −15→-F, −14→-E, …, −10→-A, −9→-9, …, −1→-1, 0→0, +1→1, …, +9→9, +10→A, …, +15→F. This mapping satisfies the following: the symbol carrying direction and absolute value are monotonically related to the state intensity, so the order relationship of the state code is the same as the order relationship of the state; and the inverse mapping is unique, and can be unambiguously decoded back to the component combination of (d, v_pos, v_vol, v_con) that satisfies equation (1). The state panorama table structure and the encoding mapping relationship are shown in Figure 3.
[0049] Multi-unit opinion synthesis (specific implementation of S8)
[0050] L = L0 + Σ ω_m × l_m...Equation (2)
[0051] Where l_m∈{−1,0,+1} is the opinion score of unit m, and the weight of abstaining units is reset to zero and does not participate in the summation; L0 is the prior of the synthesized opinion, and in this embodiment, L0=0. The initial determination of the routing weight is performed according to the mapping of "state type → weight reassembly". As a configurable specific implementation, the state type is divided into levels according to the degree of fluctuation:
[0052] The mapping for "State Type → Weighted Restructuring" is as follows: Stable type: Trend unit 0.25, Momentum unit 0.20, Fluctuation unit 0.10, Boundary unit 0.15, Risk unit 0.10, Cycle level unit 0.20; Regular type: Trend unit 0.20, Momentum unit 0.20, Fluctuation unit 0.15, Boundary unit 0.15, Risk unit 0.15, Cycle level unit 0.15; Extreme volatility type: Trend unit 0.10, Momentum unit 0.10, Fluctuation unit 0.20, Boundary unit 0.10, Risk unit 0.35, Cycle level unit 0.15.
[0053] Determination of completeness of unit evidence (S7): Each processing unit predeclares the evidence domain and applicable boundaries; the system determines the status of the data source according to the preset health index. The health index is defined as the product of data integrity rate and data timeliness rate (integrity rate = actual number of fields arrived / expected number of fields, timeliness rate = proportion of data batches arriving within the preset time limit). When the health index is lower than the preset threshold (0.8 in this embodiment), the evidence is determined to be incomplete.
[0054] Constrained calibration of route weights (specific implementation of S12)
[0055] ω_{m,t+1} = min[ ω_{m,t}(1+δ), max[ ω_{m,t}(1−δ), ω_{m,t} + η·e_{m,t} ] ]... Equation (3)
[0056] Where e_{m,t} is the rolling posterior error of unit m in round t (the rolling window takes the backfill results of the most recent 20 rounds), η is the calibration step size (0.05~0.2, 0.1 in this embodiment), δ is the upper limit ratio of the single adjustment amplitude (0.1~0.3, 0.2 in this embodiment), and the weight of each unit is always within the box constraint range [0.05, 0.60].
[0057] Definition of result labels and rolling posterior error (specific implementation of S11 / S12)
[0058] The result label is the quantified value of the actual trend corresponding to the assumed record at the end of the backfill window. In this embodiment, the result label is the direction of the observed object's interval return within the backfill window (the 5th / 20th trading day), with values of {−1, 0, +1} (decline / stable / rise). Those skilled in the art will understand that the result label can also use other calculable methods such as interval return values or return percentile levels, all of which fall within the scope of protection of this invention.
[0059] The rolling posterior error is calculated according to formula (4):
[0060] e_{m,t} = (1 / W) × Σ_k ( y_k − l_{m,k} )…Equation (4)
[0061] Where W is the length of the rolling window, y_k is the result label obtained in the kth round of refilling, and l_{m,k} is the opinion score of unit m in this round. When the unit opinion is consistently higher than the actual trend (l>y most often), e is negative, and equation (3) adjusts its weight downward; conversely, it is adjusted upward—the directionality of the error enables the calibration to have self-correcting capability.
[0062] Numerical walkthrough
[0063] Given v_pos=8, v_vol=2, v_con=0, d=+1, from equation (1) we get s=+10, and by mapping we get h="A"; given L0=0, three-unit opinion score +1 / −1 / 0, weight 0.45 / 0.30 / 0.25, from equation (2) we get L=0.15; given the unit opinion score sequence within the scrolling window of W=5 +1 / +1 / 0 / +1 / −1, and the backfill result label sequence +1 / 0 / 0 / +1 / −1, from equation (4) we get e=[(0)+(−1)+0+0+0] / 5=−0.2; then given ω_{m,t}=0.45, η=0.1, δ=0.2, the candidate update value is 0.45+0.1×(−0.2)=0.43, and the pruning interval is [0.36, [0.54], not exceeding the limit, therefore ω_{m,t+1}=0.43 (the weight is reduced due to persistently high opinion).
[0064] Deployment Architecture
[0065] The system is deployed in three stages: the data construction stage performs incremental state construction (S1-S5) on the entire observed object set at a preset frequency (e.g., after the close of each trading day), calculating and generating only the newly added state records for that period; the server receives the incremental state synchronization package generated by the data construction stage (daily full object incremental synchronization contains the status codes of approximately 5,549 objects, approximately 22.5KB after gzip compression), merges it into the state panorama table, and maintains the equivalent index; the query stage provides millisecond-level state retrieval based on the equivalent index for downstream processing units and the display stage. All three stages can be deployed on the same device or in a distributed manner, all falling within the protection scope of this invention.
[0066] Example 1: Multi-Period Status Monitoring of the Securities Market
[0067] A data processing system executes the above method (S1-S12) on approximately 5,500 securities after the market closes each day. Six processing units (trend, momentum, volatility, boundary, risk, and cycle level) generate opinions in parallel; assuming the record backfill window is set to the 5th / 20th trading day. Branch 1: If a unit's evidence domain relies on valuation data, and the health of this data source is below 0.8 on a given day, that unit is forcibly abstains, and the synthesis is completed by the other five units. Assume the record indicates the reason for abstention. Branch 2: When the daily status type is determined to be extreme volatility, the risk unit weight is increased to 0.35 according to the mapping; when the risk unit's opinion contradicts that of the majority of units, the conflict is classified as "risk falsification" and explicitly output, and the synthesis conclusion is downgraded. The technical effects are described in the experimental data of items 1, 2, and 4 of the beneficial effects section of the invention.
[0068] Example 2: Enterprise Operation Cycle Monitoring
[0069] The observation object is the enterprise; the cycle set is replaced with quarterly, annual, and industry prosperity cycles; the state components are replaced with growth direction, relative industry percentile, cash flow fluctuation level, and expansion / contraction level; the unified anchor is the enterprise's annual benchmark operating indicator; the processing unit is replaced with financial, market, organizational, and risk perspective units; the backfill window is set to quarterly operating data arrival. The rest of the process remains unchanged from equations (1) to (3).
[0070] Example 3: Investment Research Scenario
[0071] The observation object is the enterprise being evaluated; the unified anchor is the preset threshold and stage quantile of the key operating indicators; it is assumed that the record corresponds to each research conclusion, and the backfill anchor is set as the subsequent operating data disclosure event and the key indicator threshold achievement event, so the backfill window is divided into multiple intermediate backfills; the lack of evidence abandonment mechanism is used for due diligence data missing scenarios. The rest of the process is unchanged from equations (1) to (3).
[0072] Example of parameters (not intended to limit the scope of the claims)
[0073] The period set is {MN1, W1, D1, H4, H1}; the status code space is -F~F; the health threshold is 0.8; η=0.1; δ=0.2; the weight boundary is [0.05, 0.60]; the backfill window is 5 / 20 trading days; the rolling window is 20 rounds; L0=0.
Claims
1. A data processing method for multi-cycle time-series states, characterized in that, include: Acquire time-series data of the observed object at multiple time granularities, and calculate multidimensional state components at each time granularity. The multidimensional state components include at least trend direction components and position components. Using the reference value of the reference time granularity among the multiple time granularities as a unified anchor, the position components of each time granularity are rewritten as comparison relationship values between the unified anchor and the corresponding boundary of the time granularity, so that the position states of different time granularities are placed in the same comparison frame. The multidimensional state components at the same time granularity are jointly mapped into a signed state value, wherein the sign of the state value carries the direction, the absolute value carries the intensity, and the order relationship between the state values is consistent with the order relationship of the states represented by the multidimensional state components. Using the object identifier and time as keys, the state values at each time granularity are written into the state table, and an equal-value index is established for the column containing the state value.
2. The method according to claim 1, characterized in that, The joint mapping includes: combining the multidimensional state components into signed integer state components according to a preset synthesis formula, and then converting them into character-type state codes through monotonically reversible mapping; the method also includes: performing inverse mapping decoding on the character-type state codes to restore the combination of the multidimensional state components.
3. The method according to claim 1 or 2, characterized in that, The state table is a panoramic state table with the observed object identifier and date as keys and the state values at each time granularity as columns; the method further includes: responding to a retrieval request containing the target state value, and performing an iso-matching scan on the panoramic state table based on the iso-index.
4. The method according to claim 1, characterized in that, Also includes: Multiple processing units output structured opinion data based on the status values in the status table. When any processing unit declares missing evidence domain data or the health index of the corresponding data source is lower than a preset threshold, the opinion data of that processing unit is forcibly set to the abstention state and its synthesis weight is reset to zero.
5. The method according to claim 4, characterized in that, Also includes: For conflicts among the opinion data of the multiple processing units, structured classification tags are generated according to preset rules and written into the output data. The types of structured classification tags include at least risk falsification, time scale divergence, and overall and local divergence.
6. The method according to claim 4, characterized in that, Also includes: The opinion data and synthesis conclusions of each processing unit are written into the hypothesis record. The identifier of the hypothesis record is generated by the record content through deterministic rules. Repeated writing does not generate duplicate records. When the future observation data corresponding to the hypothesis record arrives, the result label is automatically written into the corresponding hypothesis record and the status of the hypothesis record is updated.
7. The method according to claim 6, characterized in that, Also includes: The rolling posterior error of each processing unit is calculated based on the backfill result label, and the composite weight of each processing unit is iteratively adjusted according to the rolling posterior error; wherein, the single adjustment range is constrained within the preset proportion band of the current weight, and the composite weight of each processing unit is within the preset boundary interval.
8. A data processing system for multi-cycle time-series states, characterized in that, include: The indicator component calculation module is used to acquire time series data of the observed object at multiple time granularities and calculate the multidimensional state components at each time granularity. The multidimensional state components include at least trend direction components and position components. The unified anchoring alignment module is used to rewrite the position components of each time granularity into comparison values between the unified anchor and the corresponding boundary of the time granularity, using the reference value of the reference time granularity as the unified anchor. A multidimensional state joint encoding module is used to jointly map the multidimensional state components at the same time granularity into a signed state value. The sign of the state value carries the direction, and the absolute value carries the intensity. The order relationship between the state values is consistent with the order relationship of the states represented by the multidimensional state components. The status panorama table and the equivalent index module are used to write the status values at each time granularity into the status table using the observed object identifier and time as keys, and to establish an equivalent index for the column containing the status values.
9. The system according to claim 8, characterized in that, Also includes: The processing layer module is used to output structured opinion data by multiple processing units based on the state table, and to forcibly set the opinion data of any processing unit to the abstention state and reset the synthesis weight to zero when the evidence domain data of any processing unit is missing or the health index of the data source is lower than a preset threshold, and to generate structured classification tags for opinion conflicts between processing units. The feedback layer module is used to write the opinion data and synthesis conclusions of each processing unit into the hypothesis record with deterministic idempotency, automatically backfill the result labels when future observation data arrives, and perform a single amplitude-limited iterative calibration of the synthesis weights of each processing unit based on the backfill results and the rolling posterior error. The system's deployment architecture includes: a data construction terminal that performs incremental state construction on the entire set of observed objects at a preset frequency, a server terminal that receives incremental state synchronization packets, and a query terminal that provides state retrieval based on the equivalent index.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 7.
Citation Information
Patent Citations
Market data compression method for market real-time network transmission
CN117097809A
Security market data compression and transmission algorithm based on JSON format
CN117709300A
Flow quotation data compression method and related equipment
CN119094787A
Internet of Things time series data lossy compression method based on symbolic aggregation approximation and variational auto-encoder
CN122660642A
