Uncertainty tracing and self-adaptive decision making system and method based on time sequence evolution causality graph
By using a temporal evolution causal graph uncertainty tracing and adaptive decision-making system, the problem of unstable causal relationship learning is solved, and stable causal identification and reliable decision-making are achieved under timestamp uncertainty. This optimizes the data acquisition strategy and improves the reliability and economy of industrial process decision-making.
Patent Information
- Application Number
- CN202511791795.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-01
- Publication Date
- 2026-03-17
AI Technical Summary
Existing causal reasoning models suffer from unstable causal relationship learning due to timestamp uncertainty in multi-source sensor data, affecting the reliability of decision-making in industrial processes, and lack specialized decomposition and tracing of time uncertainty.
An uncertainty tracing and adaptive decision-making system based on temporal evolution causal graphs is adopted. The system corrects timestamp deviation through a probabilistic time warp module, decomposes uncertainty sources through a dynamic uncertainty tracing module, adjusts decision thresholds through an adaptive decision-making module, and optimizes data collection strategies through a data collection planning module.
It can stably identify causal relationships, improve the reliability of decision-making, reduce resource consumption, enhance the economic efficiency of system operation, and maintain long-term coordination performance.
Smart Images
Figure CN121684028A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the interdisciplinary field of artificial intelligence and data science, and in particular relates to an uncertainty tracing and adaptive decision-making system and method based on temporal evolution causal graphs. Background Technology
[0002] With the development of the Industrial Internet and intelligent manufacturing, multi-source sensors are widely used to monitor complex industrial processes (such as semiconductor wafer processing and assembly line quality control). In these scenarios, the system relies on heterogeneous event data collected from multiple sensors (such as temperature, pressure, and optical sensors) distributed throughout the production equipment for real-time monitoring, causal analysis, and intelligent decision-making (such as fault warning and process parameter adjustment).
[0003] However, issues such as clock asynchrony, communication delays, differences in acquisition frequency, and network jitter exist during the acquisition of data from various sensors, leading to inherent uncertainties and systematic biases in the timestamp recording of events. For example, during wafer fabrication, the "process completed" events reported by the lithography and etching machines may experience time recording offsets due to the aforementioned reasons.
[0004] Traditional causal reasoning models (such as structural causal models and constraint-based causal discovery algorithms) typically assume that the temporal order of events is precisely known or deterministic. In real-world industrial scenarios with significant temporal noise, this strong assumption often fails. Directly applying these methods to learn causal relationships can lead to instability or even reversal, for example, misclassifying a "result" as a "cause." This directly reduces the reliability of downstream decisions based on such causal relationships, potentially causing false alarms, missed alarms, or suboptimal process adjustments, impacting production quality and efficiency.
[0005] On the other hand, current methods for handling uncertainty mostly focus on model output confidence or parameter variance, lacking a dedicated decomposition and tracing of time uncertainty, a key source. Furthermore, the system lacks an inherent mechanism to dynamically guide data acquisition strategies and decision threshold adjustments based on uncertainty. Therefore, to address these issues, the following solutions are proposed. Summary of the Invention
[0006] The purpose of this invention is to provide an uncertainty tracing and adaptive decision-making system and method based on time-series evolution causal graphs. By introducing a probabilistic time distortion and dynamic uncertainty tracing mechanism, it can effectively correct timestamp deviations in multi-source industrial data and analyze the sources of uncertainty, thus solving the problems of unstable causal relationship learning and reduced decision reliability caused by the uncertainty of time information in existing causal inference methods.
[0007] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:
[0008] This invention is an uncertainty tracing and adaptive decision-making system based on time-series evolution causal graphs, comprising:
[0009] The input module is used to receive event observation data from multiple data sources, where each event includes an event identifier, a timestamp of the source record, and event characteristics;
[0010] The probabilistic time warp module is used to estimate the true time posterior distribution of each event based on a hierarchical Bayesian model, and ensures that the true time series meets the predefined order relationship by introducing time series constraints between events as a regularization term.
[0011] The causal graph learning module is used to jointly infer the existence probability of edges and the posterior distribution of effects in the causal graph, taking the real-time posterior distribution as input, and supports incremental updates;
[0012] The dynamic uncertainty tracing module is used to perform multi-source variance decomposition on causal learning results, construct an uncertainty evolution tree, and trace the sources and contributions of temporal uncertainty, structural uncertainty, and sample uncertainty.
[0013] The adaptive decision-making module is used to dynamically adjust the decision threshold based on the uncertainty tracing results and to make conservative decisions by calculating the quantiles of the output posterior distribution.
[0014] The data collection planning module is used to evaluate the expected information gain of data collection actions based on the value information criterion and generate targeted collection instructions;
[0015] A versioned time sequence graph database is used to store historical versions of system states, supporting backtracking queries and incremental updates.
[0016] Furthermore, the probabilistic time warp module estimates the time offset of each data source through a hierarchical Bayesian model and applies identifiability constraints; when estimating the true time posterior distribution, it introduces inter-event temporal constraints as part of the loss function to ensure that the true time series of events conforms to a predefined order relationship.
[0017] Furthermore, when constructing the causal graph prior, the causal graph learning module modulates the existence probability of edges based on the real-time posterior distribution, making events with earlier time sequences more likely to act as cause nodes. The module also implements incremental updates of the causal graph through a sliding time window mechanism, updating only the nodes and edges affected by new data.
[0018] Furthermore, the dynamic uncertainty tracing module decomposes the total uncertainty into time uncertainty, structural uncertainty, and sample uncertainty using the variance decomposition method. The time uncertainty is further decomposed into source offset, sampling delay, and synchronization error. The module also constructs an uncertainty evolution tree to dynamically track the changes in the variance proportion of uncertainty sources with time as the horizontal axis.
[0019] Furthermore, the adaptive decision-making module achieves conservative decision thresholds by dynamically adjusting quantile parameters; when time uncertainty increases, the quantile parameters shift towards the tail of the distribution, so that the decision threshold is set based on the worst-case scenario; the output of the module includes binary decisions and associated confidence levels.
[0020] Furthermore, when calculating value information, the data collection planning module comprehensively considers the expected loss reduction and collection cost, and introduces a data source reputation scoring mechanism to penalize low-reputation sources; the module outputs machine-executable collection instructions, forming a closed-loop optimization.
[0021] An uncertainty tracing and adaptive decision-making method based on temporal evolution causal graphs, the method comprising the following steps:
[0022] Receive multi-source event observation data and perform standardized preprocessing;
[0023] The true temporal posterior distribution of events is estimated by a probabilistic time warp module, in which a hierarchical Bayesian model is used and inter-event temporal constraints are introduced as regularization terms.
[0024] Using the real-time posterior distribution as input, we learn the causal graph structure, jointly estimate the existence probability of edges and the posterior distribution of effects, and achieve incremental updates.
[0025] Uncertainty decomposition is performed on the causal learning results, an uncertainty evolution tree is constructed, and the contributions of multi-source uncertainty are traced.
[0026] The decision threshold is dynamically adjusted based on the uncertainty tracing results, and conservative decision-making is achieved through quantile calculation.
[0027] Based on value information criteria, plan data acquisition actions, generate targeted acquisition instructions, and execute them;
[0028] The system state is recorded in a versioned sequence graph database, supporting backtracking and incremental updates.
[0029] Furthermore, in the probabilistic time warp step, the true time posterior distribution is estimated by minimizing a loss function that includes temporal constraints, where the constraints ensure that the true time series of events conforms to a predefined order relationship.
[0030] Furthermore, in the causal graph learning step, the prior probability of the edge is modulated based on the real-time posterior distribution, making events with earlier time sequences more likely to be causes; when new data arrives, only the affected nodes and edges are updated.
[0031] Furthermore, in the data collection planning step, the expected information gain of the collection action is evaluated by calculating value information, and a data source reputation scoring mechanism is introduced to penalize low-reputation sources; after the collection instruction is executed, the new data is fed back to the input step, forming a system closed loop.
[0032] The present invention has the following beneficial effects:
[0033] 1. This invention employs a probabilistic time warp module and a hierarchical Bayesian model to jointly estimate the true posterior distribution of events in time. It also introduces inter-event temporal constraints as a regularization term, effectively correcting time offsets caused by clock asynchrony and communication delays in multi-source data. This approach reduces the interference of timestamp inconsistencies on causal structure learning, enabling the causal graph learning module to perform joint inference based on the time posterior, thus reducing the possibility of misjudging causal edges. Simultaneously, the incremental update mechanism only makes local adjustments to affected nodes, avoiding computational fluctuations caused by global reconstruction. Overall, this method can more stably identify causal relationships in environments with temporal noise, providing a more reliable basis for decision-making.
[0034] 2. The dynamic uncertainty tracing module of this invention analyzes the overall uncertainty into multi-source contributions such as time, structure, and samples through variance decomposition technology, and constructs an evolution tree to track its changing trend over time; the adaptive decision module dynamically adjusts the quantile threshold based on the uncertainty distribution, and automatically moves the decision boundary towards a conservative direction when the uncertainty is high; this mechanism enables the system to flexibly adjust the judgment criteria according to the uncertainty state, avoid decision deviations caused by environmental fluctuations, and thus maintain the stability of the output results under complex conditions, supporting the system to maintain coordinated performance in long-term operation.
[0035] 3. The data collection planning module of this invention evaluates the expected information gain and cost of collection actions based on the value information criterion, and introduces a data source reputation scoring mechanism to penalize sources with low consistency, prioritizing high-efficiency collection strategies. Through nested Monte Carlo simulation and variance reduction techniques, the system can efficiently calculate the potential value of different actions, avoiding redundant or inefficient data collection. This method achieves targeted collection planning, reduces unnecessary resource consumption, and ensures that new data makes a substantial contribution to narrowing key uncertainties, thereby improving the overall operational economy of the system.
[0036] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description
[0037] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0038] Figure 1 This is a schematic diagram of the uncertainty tracing and adaptive decision-making system based on temporal evolution causal graph of the present invention;
[0039] Figure 2 This is a flowchart illustrating the uncertainty tracing and adaptive decision-making method based on temporal evolution causal graphs of the present invention. Detailed Implementation
[0040] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0041] Please see Figure 1 As shown, this invention is an uncertainty tracing and adaptive decision-making system based on temporal evolution causal graphs, comprising:
[0042] The input module is used to receive event observation data from multiple data sources, where each event includes an event identifier, a timestamp of the source record, and event characteristics;
[0043] The probabilistic time warp module is used to estimate the true time posterior distribution of each event based on a hierarchical Bayesian model, and ensures that the true time series meets the predefined order relationship by introducing time series constraints between events as a regularization term.
[0044] The causal graph learning module is used to jointly infer the existence probability of edges and the posterior distribution of effects in a causal graph, taking the real-time posterior distribution as input, and supports incremental updates.
[0045] The dynamic uncertainty tracing module is used to perform multi-source variance decomposition on causal learning results, construct an uncertainty evolution tree, and trace the sources and contributions of temporal uncertainty, structural uncertainty, and sample uncertainty.
[0046] The adaptive decision-making module is used to dynamically adjust the decision threshold based on the uncertainty tracing results and to make conservative decisions by calculating the quantiles of the output posterior distribution.
[0047] The data collection planning module is used to evaluate the expected information gain of data collection actions based on the value information criterion and generate targeted collection instructions;
[0048] A versioned time sequence graph database is used to store historical versions of system states, supporting backtracking queries and incremental updates.
[0049] The probabilistic time warp module estimates the time offset of each data source using a hierarchical Bayesian model and imposes identifiability constraints. When estimating the true time posterior distribution, it introduces inter-event temporal constraints as part of the loss function to ensure that the true time series of events conforms to a predefined order relationship.
[0050] When constructing the causal graph prior, the causal graph learning module modulates the existence probability of edges based on the real-time posterior distribution, making events with earlier time sequences more likely to be causal nodes. The module also implements incremental updates of the causal graph through a sliding time window mechanism, updating only the nodes and edges affected by new data.
[0051] The dynamic uncertainty tracking module decomposes the total uncertainty into time uncertainty, structural uncertainty and sample uncertainty through the variance decomposition method. The time uncertainty is further decomposed into source offset, sampling delay and synchronization error. The module also constructs an uncertainty evolution tree to dynamically track the change in the variance proportion of uncertainty sources with time as the horizontal axis.
[0052] The adaptive decision-making module conservatively adjusts the quantile parameters to make the decision threshold conservative. When time uncertainty increases, the quantile parameters move towards the tail of the distribution, so that the decision threshold is set based on the worst case. The module's output includes binary decision and associated confidence.
[0053] When calculating value information, the data collection planning module comprehensively considers the expected loss reduction and collection cost, and introduces a data source reputation scoring mechanism to penalize low-reputation sources; the module outputs machine-executable collection instructions, forming a closed-loop optimization.
[0054] Please see Figure 2 As shown, an uncertainty tracing and adaptive decision-making method based on temporal evolution causal graphs includes the following steps:
[0055] Receive multi-source event observation data and perform standardized preprocessing;
[0056] The true temporal posterior distribution of events is estimated by a probabilistic time warp module, in which a hierarchical Bayesian model is used and inter-event temporal constraints are introduced as regularization terms.
[0057] Using the real-time posterior distribution as input, we learn the causal graph structure, jointly estimate the existence probability of edges and the posterior distribution of effects, and achieve incremental updates.
[0058] Uncertainty decomposition is performed on the causal learning results, an uncertainty evolution tree is constructed, and the contributions of multi-source uncertainty are traced.
[0059] The decision threshold is dynamically adjusted based on the uncertainty tracing results, and conservative decision-making is achieved through quantile calculation.
[0060] Based on value information criteria, plan data acquisition actions, generate targeted acquisition instructions, and execute them;
[0061] The system state is recorded in a versioned sequence graph database, supporting backtracking and incremental updates.
[0062] In the probabilistic time warp step, the true time posterior distribution is estimated by minimizing a loss function that includes temporal constraints, where the constraints ensure that the true time series of events conforms to a predefined order relationship.
[0063] In the causal graph learning step, the prior probability of an edge is modulated based on the real-time posterior distribution, making earlier chronological events more likely to be the cause; when new data arrives, only the affected nodes and edges are updated.
[0064] In the data collection planning step, the expected information gain of the collection action is evaluated by calculating value information, and a data source reputation scoring mechanism is introduced to penalize low-reputation sources; after the collection instruction is executed, the new data is fed back to the input step, forming a closed loop in the system.
[0065] One specific application of this embodiment is:
[0066] This embodiment is applied to the wafer fabrication process monitoring scenario in semiconductor manufacturing. In this scenario, multiple sensors (such as temperature, pressure, and optical sensors) monitor wafer fabrication steps (such as photolithography, etching, and deposition). However, due to sensor clock asynchrony, communication delays, and environmental noise, event timestamps have significant offsets and uncertainties, leading to errors in causal relationship learning, which in turn affects yield prediction and equipment fault diagnosis. This method achieves uncertainty tracing and adaptive decision-making through the following steps, all of which are automatically executed by computing devices without human intervention.
[0067] Step S1: Input multi-source event data
[0068] The system receives event observation data from multiple data sources during the semiconductor manufacturing process; the data format is as follows:
[0069] D={(e i ,t i,s ,x i ,s)|i=1,…,N,s∈S};
[0070] In the formula, e iFor event identifiers (such as "lithography complete", "etching begins"); t i,s For the timestamps recorded in data source s; x i For example, N is the event feature vector (e.g., sensor readings, device status); S is the total number of events; S is the data source set (e.g., sensor ID list); and s is the data source identifier.
[0071] The input module preprocesses the data, including format normalization, missing value marking, and outlier filtering, but does not perform time correction; the data is then transmitted to the probabilistic time warp module for further processing.
[0072] Step S2, Probabilistic Time Warp Module
[0073] This module employs a hierarchical Bayesian model to jointly estimate the time offset of each data source and the true time of events, and introduces inter-event timing constraints to address the problem that process steps in semiconductor manufacturing must be executed sequentially (e.g., photolithography must precede etching). The specific steps are as follows:
[0074] Step S21: Establish a hierarchical Bayesian model:
[0075] For each data source s, define a time offset δ s Prior distribution:
[0076]
[0077] In the formula, δ s μ is the time offset of the data source s. s The prior mean of the time offset of the data source s; μ is the prior variance of the time offset of the data source s; s and Calibration and initialization are based on historical data;
[0078] To ensure the model's identifiability, constraint ∑ is applied. s∈S δ s =0, meaning a reference sensor is used as the reference.
[0079] Step S22: Define the true time posterior distribution:
[0080] The real time of each event τ i The posterior distribution is:
[0081]
[0082] In the formula, p(τ) i ∣{t i,s},{δ s}) represents the posterior distribution of the true time given the observation timestamp and offset; p(τ) ip(t) represents the prior distribution over real time; i,s ∣τ i ,δ s Let be the likelihood function, assumed to be a Laplace distribution. i,s ∣τ i +δ s b), to robustly handle outliers, where b is the scaling parameter of the Laplace distribution;
[0083] Step S23: Introduce inter-event timing constraints: Based on semiconductor process knowledge, define an inter-event timing constraint set C; for example, for event u (lithography completion) and event v (etching start), the constraint is τ. u <τ v ;
[0084] In the posterior estimation, a constraint term is added as a regularization loss function to ensure that the true time series meets the process sequence. The loss function is defined as follows:
[0085]
[0086] In the formula, L({τ i},{δ s}) represents the loss function for variational inference; C represents the set of temporal constraints between events; u and v are both event indices; τ u ,τ v Let u and v be the actual times; max(0,τ) u -τ v ) 2 This is the penalty term for violating timing constraints; γ is the constraint strength parameter, set through cross-validation.
[0087] Using variational inference to minimize the loss function, we obtain an approximate posterior distribution q({τ)). i},{δ s Specifically, stochastic gradient descent is used to optimize the variational parameters to ensure that the constraint violation term approaches zero.
[0088] Step S24: Output the true time posterior: The module outputs the true time posterior distribution p(τ) for each event. i The parameters are Gaussian distribution parameters (mean and variance), which are used for subsequent causal graph learning.
[0089] Step S3, Cause-and-Effect Graph Learning Module
[0090] This module takes real-time posterior as input, learns a temporal causal graph, and dynamically updates the causal edge probabilities:
[0091] Step S31: Construct the prior of the causal graph: Define the prior distribution P(G) of the causal graph G, based on time order modulation: For any event pair (A, B), if τA <τ B Then the prior probability of edge A→B is higher, specifically:
[0092]
[0093] In the formula, P(A→B) is the probability of the existence of a causal edge from A to B; A and B are event nodes in the causal graph; logit -1 The function is the logistic sigmoid function; α is the sensitivity parameter. The posterior mean at real time;
[0094] Step S32: Jointly estimate the edge existence probability and effect: Given the real-time posterior p(τ) i ) and event characteristics x i The probability of the existence of edge A→B is calculated using Bayesian structure learning algorithms (such as dynamic linear models):
[0095] P edge (A→B)=p(A causes B∣{τ i},X)
[0096] In the formula, P edge (A→B) represents the posterior probability of the edge at a given time posterior and feature time.
[0097] Simultaneously estimate the posterior distribution of causal effects p(β) A→B ), where β A→B The magnitude of the effect from A to B;
[0098] Step S33, Incremental Update Mechanism: When new event data arrives, only the posterior distribution of the affected nodes and edges is updated to avoid full graph reconstruction; the update is based on a sliding time window, and the window size is adaptively adjusted according to the data frequency.
[0099] Step S34: Output causal graph: The module outputs the probability matrix and posterior distribution of the effect edges, which are stored in the time series graph database;
[0100] Step S4, Dynamic Uncertainty Traceability Module
[0101] This module performs uncertainty decomposition on the causal learning results and constructs an uncertainty evolution tree (USTT):
[0102] Step S41, Variance decomposition:
[0103] Based on the output variable Y (such as yield index), calculate the total variance and decompose it into multi-source contributions:
[0104]
[0105] In the formula, Var(Y) represents the total variance of the output variable Y; k (Y) represents the variance contribution of the k-th uncertainty source to Y; Cov k,l (Y) represents the covariance contribution of the kth and lth uncertainty sources to Y; k, l are the indices of the uncertainty sources: time uncertainty, structural uncertainty, and sample uncertainty.
[0106] The Sobol exponent method is used for decomposition, specifically by approximating the sensitivity exponents of each order through Monte Carlo sampling.
[0107] Step S42: Construct an uncertainty evolution tree:
[0108] The root node is Var(Y);
[0109] The first-level child nodes include:
[0110] Time uncertainty: further decomposed into source offset δ s Sampling delay Δt, synchronization error ε;
[0111] Structural uncertainty: decomposed into the probability P of edge existence edge variance and covariance;
[0112] Sample uncertainty includes missing data rate and observation noise;
[0113] Each leaf node corresponds to a specific source of uncertainty, and the node width is the variance percentage, which is dynamically updated over time.
[0114] Step S43: Output the traceability results:
[0115] The module outputs the distribution and evolution trend of uncertainty sources for adaptive decision-making and data acquisition planning;
[0116] Step S5, Adaptive Decision Module
[0117] This module adjusts the decision threshold based on the results of uncertainty tracing to ensure robustness:
[0118] Step S51: Calculate the conservative quantile threshold:
[0119] Decision threshold θ t Defined as the output posterior distribution F Y Quantiles:
[0120]
[0121] In the formula, This is the quantile function for the output variable Y; q is the conservative quantile parameter, initially set to 0.5 (median);
[0122] Based on the degree of time uncertainty σ τ (Mean of the posterior standard deviation at real time), dynamically adjusted q:
[0123] q = max(0.1, 0.5 - β × σ) τ );
[0124] In the formula, q is the conservative quantile parameter; β is the attenuation coefficient, calibrated using historical error rate; σ τ This represents the mean of the posterior standard deviation at real time.
[0125] Step S52: Execute a conservative decision: For the current event, if the predicted value of the output variable Y is lower than θ t If this occurs, an alarm will be triggered (such as a device malfunction warning);
[0126] The decision-making mechanism is based on the worst-case scenario to ensure that the error rate is controllable under high uncertainty.
[0127] Step S53: Output decision results: The module outputs a binary decision (normal / abnormal) and related confidence scores, and stores them in the database;
[0128] Step S6, Data Collection Planning Module
[0129] This module proactively plans data acquisition strategies based on the uncertainty traceability and Value Information (VOI) criteria:
[0130] Step S61, Calculated value information:
[0131] For each potential acquisition action a k (e.g., querying a specific sensor, increasing the sampling frequency), calculate its expected information gain:
[0132]
[0133] In the formula, VOI(a k ) is the data collection action a k Value information; a k This represents the k-th potential data collection action. Given an action a k The expected reduction in losses is conditional; ΔL is the expected reduction in losses (e.g., reduced cost of missed faults); C(a k The cost of data collection (e.g., computational resource consumption) is the cost of data acquisition.
[0134] Estimation using nested Monte Carlo simulations Furthermore, variance reduction techniques are employed to improve computational efficiency;
[0135] Step S62, Anti-manipulation Measures: Introduce a data source reputation scoring mechanism: a score r for each source s. sDynamic updates based on historical data consistency; in VOI calculation, a penalty term is applied to low-reputation sources: C(a k )←C(a k ) / r s ;
[0136] Step S63: Generate acquisition command: Select the action with the largest VOI, a * It generates machine-executable instructions (such as API calls and sensor calibration commands); after the instructions are executed, the new data is fed back to the input module, forming a closed-loop optimization.
[0137] Step S7: Versioned sequence diagram database
[0138] This module maintains a versioned time sequence diagram database, supporting full-process data traceability and incremental updates.
[0139] Step S71: Record version snapshots: After each causal graph update or decision adjustment, create a new version snapshot, including timestamp, causal graph state, uncertainty distribution, and decision log;
[0140] Step S72, Support backtracking query: Provide an API interface to allow retrieval of historical status by time range or event type for fault analysis and model auditing;
[0141] Step S73, Incremental Update Mechanism: Only store the difference data to optimize storage efficiency; old versions are automatically archived to retain the latest state in the scrolling window.
[0142] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.
[0143] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.
Claims
1. An uncertainty trace and adaptive decision system based on time-evolving causal graphs, characterized in that, The system comprises: An input module for receiving event observation data from multiple data sources, wherein each event comprises an event identifier, a timestamp of the source record, and event features; A probabilistic time warping module for estimating a real-time posterior distribution of each event based on a hierarchical Bayesian model, and ensuring that the real-time sequence satisfies a predefined order relationship by introducing inter-event timing constraints as regularization terms; A causal graph learning module for jointly inferring the existence probability of edges in the causal graph and the effect posterior distribution based on the real-time posterior distribution as input, and supporting incremental updates; A dynamic uncertainty tracing module for performing multi-source variance decomposition on the causal learning results, constructing an uncertainty evolution tree, and tracing the sources and contributions of temporal uncertainty, structural uncertainty, and sample uncertainty; An adaptive decision-making module for dynamically adjusting the decision threshold based on the uncertainty tracing results, and implementing conservative decision-making by calculating the quantile of the output posterior distribution; A data collection planning module for evaluating the expected information gain of data collection actions based on the value information criterion, and generating directional collection instructions; A versioned temporal graph database for storing historical versions of system states, supporting backtracking queries, and incremental updates.
2. The temporal-evolution-based causal diagram based uncertainty backtracking and adaptive decision system according to claim 1, characterized in that, The probabilistic time warping module estimates the time offset of each data source through a hierarchical Bayesian model and imposes distinguishability constraints; when estimating the real-time posterior distribution, inter-event timing constraints are introduced as part of the loss function to ensure that the event real-time sequence conforms to the predefined order relationship.
3. The temporal-evolution-based causal diagram based uncertainty backtracking and adaptive decision making system according to claim 1, wherein, The causal graph learning module modulates the existence probability of edges based on the real-time posterior distribution when constructing the causal graph prior, making events with earlier temporal order more likely to be cause nodes; the module also implements incremental updates of the causal graph through a sliding time window mechanism, updating only the nodes and edges affected by new data.
4. The temporal-evolution-based causality graph based uncertainty backtracking and adaptive decision system according to claim 1, characterized in that, The dynamic uncertainty tracing module decomposes total uncertainty into temporal uncertainty, structural uncertainty, and sample uncertainty through variance decomposition methods, where temporal uncertainty is further decomposed into source offset, sampling delay, and synchronization error; the module also constructs an uncertainty evolution tree to dynamically track the variance proportion changes of uncertainty sources along the time axis.
5. The temporal-evolution-based causal diagram based uncertainty backtracking and adaptive decision making system according to claim 1, wherein, The adaptive decision-making module implements the conservatism of the decision threshold by dynamically adjusting the quantile parameter; when temporal uncertainty increases, the quantile parameter moves to the tail of the distribution, making the decision threshold based on the worst-case scenario; the output of the module includes binary decisions and associated confidence levels.
6. The temporal-evolution-based causal diagram based uncertainty backtracking and adaptive decision making system according to claim 1, wherein, The data collection planning module considers both expected loss reduction and collection costs when calculating value information, and introduces a data source reputation scoring mechanism to penalize low-reputation sources; the module outputs machine-executable collection instructions, forming a closed-loop optimization.
7. Uncertainty tracing and adaptive decision-making method based on time-evolving causal graphs, characterized in that, The method comprises the following steps: Receiving multi-source event observation data and performing standardization preprocessing; Estimating the real-time posterior distribution of events through the probabilistic time warping module, which adopts a hierarchical Bayesian model and introduces inter-event timing constraints as regularization terms; Input real-time posterior distribution to learn causal graph structure, jointly estimate edge existence probability and effect posterior distribution, and realize incremental update; Decompose uncertainty of causal learning result, build uncertainty evolution tree, and trace contribution of multi-source uncertainty; Dynamically adjust decision threshold based on uncertainty tracing result, and realize conservative decision through quantile calculation; Plan data collection action based on value information criterion, generate directional collection instruction, and execute; Record system state in versioned time-series graph database, support backtracking and incremental update.
8. The method of claim 7, wherein, In the probabilistic time warping step, the real-time posterior distribution is estimated by minimizing a loss function containing temporal constraints, where the constraint term ensures that the real-time sequence of events conforms to the predefined order relationship.
9. The method of claim 7, wherein, In the causal graph learning step, the prior probability of the edge is modulated based on the real-time posterior distribution, so that the events with earlier time order are more likely to be the cause; when new data arrives, only the affected nodes and edges are updated.
10. The method of claim 7, wherein, In the data collection planning step, the expected information gain of the collection action is evaluated by value information calculation, and a data source reputation scoring mechanism is introduced to punish low-reputation sources; after the execution of the collection instruction, the new data is fed back to the input step, forming a closed loop system.