A method and system for multi-factor data analysis and job discrimination

Abstract

This application relates to the field of data analysis and pattern recognition, and discloses a multi-factor data analysis and job discrimination method and system. The method includes the following steps: constructing a spatiotemporal dynamic map of the monitoring system; based on historical map data, using a baseline prediction model to extrapolate the current natural evolution state; comparing the predicted state with the actual observed state, calculating the response residuals and matching them with job impulse templates to generate job hypotheses; extrapolating and verifying the future trajectories of each hypothesis to determine the optimal discrimination result; and performing causal decoupling based on the result to obtain cleaned data for model correction. This invention, by combining a discrimination scheme of competitive hypothesis generation and time-series verification, does not immediately make decisions regarding detected abnormal data, but instead generates multiple possible job hypotheses and extrapolates the impact trajectory of each hypothesis over subsequent time periods. By comparing these trajectories with real data, the optimal discrimination result is ultimately determined.

Description

Technical Field

[0001] This invention relates to the field of data analysis and pattern recognition, specifically to a method and system for multi-factor data analysis and task discrimination. Background Technology

[0002] With the continuous development of the Internet of Things, sensor networks, and big data technologies, various types of multi-element spatiotemporal data are being generated and accumulated at an astonishing rate in fields such as smart agriculture, intelligent manufacturing, and environmental protection. This data contains a wealth of information, comprehensively reflecting the operational status of physical systems, environmental changes, and potential external interventions. Therefore, in-depth and precise analysis of this massive amount of data is of indispensable practical significance for effectively understanding system behavior, timely detecting anomalies, and accurately attributing events.

[0003] Existing technologies have made some progress in multi-element spatiotemporal data processing. For example, some methods can achieve preliminary aggregation and storage of heterogeneous data; others can monitor data in real time by setting thresholds or based on statistical methods, and issue alarms when values ​​deviate from preset ranges, thereby assisting users in preliminary anomaly detection. These technologies have demonstrated their application value in data management and basic deviation detection.

[0004] However, existing technologies still have some significant shortcomings in practical applications. First, when faced with external tasks with similar initial characteristics, existing methods often struggle to make accurate distinctions. They rely heavily on static data pattern matching at the moment of event occurrence, neglecting the different impact patterns of different tasks in subsequent time series, leading to easily confused judgment results and insufficient reliability. Second, after initial training, most existing analysis models tend to have fixed internal parameters and understanding of system behavior. These models lack the ability to learn from and self-correct events confirmed in actual operation, making them unable to adapt to natural evolutionary drift or new task patterns that occur during long-term system operation, thus compromising long-term performance. Third, existing data analysis schemes have limitations in revealing the underlying causes of data changes. They often only provide statistical correlations between data, failing to clearly explain whether observed system changes originate from their own natural evolution or are caused by specific external tasks. Their internal discrimination logic is often opaque, weakening the physical interpretability and credibility of the analysis results. Therefore, those skilled in the art propose a multi-factor data analysis and task discrimination method and system to address these problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a multi-factor data analysis and job identification method and system, which solves problems such as difficulty in distinguishing job types, lack of adaptive updating of analysis models, and lack of causal explanation ability in existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides a method for multi-factor data analysis and job discrimination, comprising the following steps:

[0008] S1. Acquire multi-element data of the monitored physical system at continuous time points, construct a spatiotemporal dynamic map, and obtain spatiotemporal dynamic map data corresponding to the spatiotemporal dynamic map.

[0009] S2. Based on the historical time point data in the spatiotemporal dynamic map data, the predicted natural evolution state at the current time point is predicted through a preset baseline prediction model.

[0010] S3. Based on the predicted natural evolution state and the actual observation state at the current time point, calculate the response residual, and match the response residual with a preset task pulse template library to generate competitive task hypotheses.

[0011] S4. Perform time-series verification on the competitive job hypothesis, and determine an optimal hypothesis as the job discrimination result by deducing and comparing the future evolution trajectory;

[0012] S5. Based on the operation discrimination result, perform causal decoupling on the actual observation state to obtain purified data, and use the purified data to correct the baseline prediction model.

[0013] Preferably, in step S1, the steps of acquiring multi-element data of the monitoring physical system at continuous time points, constructing a spatiotemporal dynamic map, and obtaining spatiotemporal dynamic map data corresponding to the spatiotemporal dynamic map include: acquiring multi-element data of the monitoring physical system at continuous time points from an external monitoring device through a communication interface;

[0014] The physical space monitored by the monitoring physical system is divided into multiple monitoring units, and each monitoring unit is abstracted as a graph node. Edges between graph nodes are defined according to preset relationships between the monitoring units, and a graph structure is constructed based on the graph nodes and the edges between them. Here, each graph node is v. i i = 1, ..., N v N v The number of monitoring units, and all graph nodes constitute a graph node set. The edges between the graph nodes are ε, and the graph structure is as follows: The preset relationship is determined based on the actual relationship between the physical spaces corresponding to the monitoring units, including spatial adjacency relationship and functional association relationship;

[0015] The multi-element data corresponding to the monitoring unit at each time point are assigned as node features to the corresponding graph nodes to form the spatiotemporal dynamic map and obtain the spatiotemporal dynamic map data.

[0016] Preferably, in step S2, the step of predicting the predicted natural evolution state at the current time point using a preset baseline prediction model includes:

[0017] The spatiotemporal dynamic graph data within the historical time window is input into the spatiotemporal graph neural network, which serves as the baseline prediction model, to obtain the predicted natural evolution state at the current time point.

[0018] The baseline prediction model is as follows:

[0019] In the formula: X is the feature matrix for predicting the natural evolutionary state at the current time point t; t-K ,...,X t-1 Let A be the sequence of node feature matrices from historical time point (tK) to (t-1), and let A be the adjacency matrix in the spatiotemporal dynamic graph. evo These are the internally learnable parameters of the baseline prediction model.

[0020] Preferably, in step S3, the step of generating competitive job hypotheses by matching the response residual with a preset job impulse template library includes:

[0021] The similarity between the response residual and the pulse templates corresponding to various job types stored in the job pulse template library is calculated.

[0022] When the similarity between the response residual and the pulse template in the job pulse template library is greater than a preset similarity threshold, a competitive job hypothesis corresponding to the pulse template is generated;

[0023] Wherein, the response residual is X t To represent the actual observation state, This is the feature matrix for predicting the natural evolutionary state at the current time point t.

[0024] Preferably, in step S4, the step of performing time-series verification of the competing job hypothesis includes:

[0025] For each competing job hypothesis, based on the future evolution trajectory calculation model, by combining the baseline prediction model and the impulse response model corresponding to the baseline prediction model, the future evolution trajectory of the competing job hypothesis within a preset verification time window is superimposed and deduced.

[0026] Within the verification time window, each of the future evolution trajectories is continuously compared with the actual observation state, and the mismatch degree between each of the competing operation hypotheses and the actual observation state is calculated using the weighted mean square error formula.

[0027] The competitive job hypothesis that minimizes the mismatch value is determined as the optimal hypothesis;

[0028] The weighted mean square error formula is as follows:

[0029]

[0030] In the formula: L verif (H j Let H be the j-th competitive task. j The value of the mismatch with the actual observed state, where Np is the length of the validation time window, and w τ X is the time decay weight. t+τ Let be the feature matrix corresponding to the actual observed state at a future time (t+τ). Let ||·|| be the feature matrix corresponding to the future evolution trajectory of the j-th competitive task at time (t+τ). F It is the Frobenius norm.

[0031] Preferably, the future evolution trajectory calculation model is:

[0032]

[0033] In the formula: Let be the feature matrix corresponding to the future evolution trajectory of the j-th competitive task at time (t+τ); g represents the characteristic matrix corresponding to the predicted natural evolution state at a future time (t+τ). resp For impulse response model; I j Let τ be the pulse template corresponding to the j-th competitive task; τ is the future time offset from the current time t; θ resp,j Let be the parameters of the impulse response model for the j-th competitive task assumption.

[0034] Preferably, the future evolution trajectory calculation model is:

[0035]

[0036] In the formula: Let be the feature matrix corresponding to the future evolution trajectory of the j-th competitive task at time (t+τ); g represents the characteristic matrix corresponding to the predicted natural evolution state at a future time (t+τ). resp For impulse response model; I j Let τ be the pulse template corresponding to the j-th competitive task; τ is the future time offset from the current time t; θ resp,j Let be the parameters of the impulse response model for the j-th competitive task assumption.

[0037] Preferably, in step S5, the step of decoupling the causal relationship of the actual observed state to obtain cleaned data includes:

[0038] Using the impulse response model corresponding to the optimal assumption, the response impact of the operational assumption represented by the optimal assumption in future observation states is calculated;

[0039] The purification data is obtained by subtracting the impact of the response from the actual future observed state based on the purification data calculation formula.

[0040] Preferably, the purification data calculation formula is as follows:

[0041]

[0042] In the formula: Let X be the feature matrix corresponding to the cleaned data at a future time t+τ. t+τ Let g be the characteristic matrix corresponding to the actual observed state at the future time t+τ. resp For the impulse response model corresponding to the optimal assumption, I opt Let τ be the pulse template corresponding to the optimal assumption, τ be the future time offset from the current time t, and θ be the pulse template. resp,opt These are the parameters of the impulse response model based on the optimal assumption.

[0043] A second aspect of the present invention provides a multi-factor data analysis and job discrimination system, comprising:

[0044] The data construction module is used to acquire multi-element data of the monitored physical system at continuous time points, construct a spatiotemporal dynamic map, and obtain spatiotemporal dynamic map data corresponding to the spatiotemporal dynamic map.

[0045] The baseline prediction module is used to predict the predicted natural evolution state at the current time point based on the historical time point data in the spatiotemporal dynamic map data and through a preset baseline prediction model.

[0046] The hypothesis generation module is used to calculate the response residual based on the predicted natural evolution state and the actual observation state at the current time point, and to generate competitive job hypotheses by matching the response residual with a preset job pulse template library.

[0047] The hypothesis verification module is used to perform time-series verification of the competing job hypotheses. By deducing and comparing future evolution trajectories, an optimal hypothesis is determined as the job discrimination result.

[0048] The model correction module is used to perform causal decoupling on the actual observation state based on the operation discrimination result to obtain purified data, and to use the purified data to correct the baseline prediction model.

[0049] Preferably, the hypothesis verification module is used for:

[0050] For each competing job hypothesis, by combining the baseline prediction model and its corresponding impulse response model, its future evolution trajectory within a preset verification time window is deduced;

[0051] Within the verification time window, each of the future evolution trajectories is continuously compared with the actual observed state, and their respective mismatch degree is calculated.

[0052] The competitive job hypothesis with the smallest mismatch is determined as the optimal hypothesis.

[0053] This invention provides a method and system for multi-factor data analysis and job discrimination. It has the following beneficial effects:

[0054] 1. This invention combines a discrimination scheme of competitive hypothesis generation and time-series verification. For detected abnormal data, it does not make an immediate decision, but generates multiple possible operational hypotheses and extrapolates the impact trajectory of each hypothesis in the subsequent time. By comparing these trajectories with real data, the optimal discrimination result is finally determined. This method differs from the existing technology that performs static matching based solely on the features at the moment the event occurs, and solves the technical defects of the existing technology, which is difficult to accurately distinguish and has low reliability when facing operations with similar initial features.

[0055] 2. This invention, through a closed-loop feedback system based on causal decoupling, can extract the impact of the confirmed task from subsequent real data after each successful task identification, thereby refining data that reflects the pure natural evolution law of the system. This data is used in reverse to correct and optimize the baseline prediction model. This solution changes the defect of the existing technology where the analysis model becomes fixed after training, and solves the problems of its inability to learn from new events, its gradually outdated understanding of the system, and its declining reliability in long-term analysis.

[0056] 3. This invention deconstructs complex data changes into a clear superposition of baseline natural evolution and external task impulse response. The entire discrimination process follows this physical framework, from identifying residuals that deviate from natural evolution, to matching initial impulses representing tasks, and then to verifying subsequent response processes. It provides a complete attribution chain for discrimination, overcoming the shortcomings of existing technologies where data is directly input into the network for end-to-end analysis and the lack of transparency in the internal discrimination logic. It also solves the problem that existing technologies rely solely on statistical correlation of data and cannot establish true causal relationships, making their discrimination results difficult to understand and trust. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the system functional modules of the present invention;

[0058] Figure 2 This is a schematic diagram of the method flow of the present invention;

[0059] Figure 3 This is a schematic diagram illustrating the application of the present invention in ground-based smoke generator rain enhancement operations.

[0060] The module consists of: 10. Data construction module; 20. Baseline prediction module; 30. Hypothesis generation module; 40. Hypothesis verification module; and 50. Model calibration module. Detailed Implementation

[0061] The technical solutions in 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.

[0062] See attached document Figure 1 , Figure 1 This is a functional block diagram of a multi-factor data analysis and job discrimination system according to an embodiment of the present invention. The system can be deployed on one or more servers. Each server includes a processor, a memory, a communication interface, and a bus. The processor, memory, and communication interface communicate with each other via the bus. The memory stores computer programs, and the processor, when executing the computer programs stored in the memory, implements the methods described later.

[0063] In one specific embodiment, the multi-factor data analysis and job discrimination system includes a data construction module 10, a baseline prediction module 20, a hypothesis generation module 30, a hypothesis verification module 40, and a model calibration module 50.

[0064] The data construction module 10 is used to perform data acquisition and structured processing. This module acquires multi-factor data at continuous time points from external sensor networks, image acquisition devices, or other data sources via a communication interface. Subsequently, the module divides the monitored physical space into multiple monitoring units and abstracts each monitoring unit as a graph node. Edges between nodes are defined based on spatial adjacency or functional association between units, constructing a static graph structure. Finally, the data construction module 10 processes the multi-factor data acquired at each time point t into a node feature matrix X. t The output of this module is spatiotemporal dynamic spectrum data, which is then provided to the baseline prediction module 20 and the hypothesis generation module 30.

[0065] The baseline prediction module 20 has its input connected to the output of the data construction module 10. This module integrates a pre-trained or online-trainable baseline prediction model f. evo For example, a spatiotemporal graph neural network. The baseline prediction module 20 receives historical spatiotemporal dynamic graph data from the data construction module 10 and bases it on a data sequence (X) of K historical time points. t -K ,...,X t-1 Predict the natural evolution state at the current time point t. The output of this module is a prediction of the natural evolutionary state. It is then provided to the hypothesis generation module 30 and the hypothesis verification module 40.

[0066] Assume that the generation module 30 has its inputs connected to the outputs of the data construction module 10 and the baseline prediction module 20, respectively. This module first receives the actual observation state X at the current time point. t And predicting the state of natural evolution The response residual matrix R is obtained by calculating the difference between the two. t Next, the module will respond with the residual matrix R. t Matching with the internally stored job pulse template library I. This library stores pulse templates corresponding to various job types. When R t When the similarity with one or more pulse templates in the library meets a preset condition, the hypothesis generation module 30 generates a corresponding set of competing job hypotheses. The output of this module is the competitive job hypothesis, which is then provided to the hypothesis verification module 40.

[0067] The hypothesis verification module 40 has its inputs connected to the outputs of the hypothesis generation module 30 and the baseline prediction module 20, respectively. This module receives competing operational hypotheses. Then, for each of these hypotheses H j Perform timing verification. Specifically, this module calls the sequence verification method corresponding to hypothesis H. j The corresponding impulse response model gresp By combining the predictions of the future state by the baseline prediction module 20, the future evolution trajectory of this hypothesis within the future verification time window can be deduced.

[0068] Subsequently, the hypothesis verification module 40 continuously compares the deduced future evolutionary trajectories with the actual observed data within that time period, calculates the degree of mismatch for each, and determines the competitive job hypothesis with the smallest degree of mismatch as the optimal hypothesis, which is then output as the final job discrimination result. The output of this module is the job discrimination result, which is provided to the model calibration module 50.

[0069] The model calibration module 50 has its input connected to the output of the hypothesis verification module 40, and its output connected to the parameter storage section of the baseline prediction module 20. After receiving the job discrimination result, this module initiates a causal decoupling procedure. It utilizes the impulse response model g corresponding to the optimal hypothesis. resp The response effects caused by this operation are removed from subsequent actual observations to obtain cleaned data.

[0070] Finally, the model calibration module 50 uses the generated cleaned data as new training samples to adjust the baseline prediction model f in the baseline prediction module 20. evo The parameter θ evo Updates or fine-tuning are performed to complete the calibration of the model.

[0071] See attached document Figure 2 , Figure 2 This is a flowchart of a multi-factor data analysis and job determination method according to an embodiment of the present invention. The method can be executed by the system described in the foregoing embodiments. The method may include:

[0072] S1, construct spatiotemporal dynamic map data.

[0073] S2, predicts the natural evolution state at the current point in time based on historical data.

[0074] S3, calculate the response residuals and generate competing job hypotheses.

[0075] S4. Perform time-series verification of the competing job assumptions to determine the job discrimination results.

[0076] S5 performs causal decoupling and model feedback correction based on the job discrimination results.

[0077] The following content details the implementation methods for multi-factor data analysis and job discrimination:

[0078] In a specific embodiment of the present invention, the execution process of step S1 is described in detail. This step is performed by... Figure 1The data construction module 10 shown is executed with the goal of transforming raw, heterogeneous multi-element data into structured spatiotemporal dynamic map data that can be processed by subsequent analysis modules.

[0079] First, the system performs data acquisition and preprocessing. It acquires multi-factor data from external data sources at continuous time points for the monitored physical system via communication interfaces. This data may include, but is not limited to: multispectral or hyperspectral image data acquired by UAVs or satellites; environmental parameters (such as soil temperature, humidity, and conductivity) collected by sensor networks or ground-based meteorological station networks deployed in physical space; and elevation or surface deformation data acquired by radar equipment. Simultaneously, static data that does not change over time can also be acquired, such as soil type distribution maps and slope and aspect data. After acquiring the raw data, preprocessing operations are performed, including data cleaning to remove outliers, data alignment to ensure temporal and spatial registration of data from different sources, and data normalization to scale characteristic values ​​of different dimensions to a uniform numerical range.

[0080] Next, the system performs spatial abstraction and graph structure definition. The monitored physical space, such as a field or an industrial workshop, is divided into a graph structure consisting of N... v A grid or functional partition composed of monitoring units. Each monitoring unit is abstracted as a graph node v. i where i = 1, ..., N v All nodes constitute a node set. Subsequently, the connection relationships between nodes, i.e., edges ε, are defined based on the preset relationships between these monitoring units. These relationships can be spatial adjacency, meaning that nodes corresponding to units sharing a boundary have edges; or they can be functional associations, such as edges between nodes corresponding to units belonging to the same irrigation area or the same production process. This constructs a static graph structure. This graph structure can be accessed through an N. v ×N v The adjacency matrix A of dimension 1 is mathematically represented.

[0081] Finally, the system performs spatiotemporal feature fusion. At each time point t, for each graph node v i This involves integrating all multi-factor data belonging to the corresponding monitoring unit at that time point. For example, preprocessed values ​​such as the average spectral reflectance within the unit and the average sensor readings within the unit are concatenated with the unit's static data (such as soil type coding) to form a D-dimensional node feature vector. All N... v The stacking of the feature vectors of each node constitutes the node feature matrix X at time point t. t ,

[0082] By repeating the above process, a spatiotemporal dynamic atlas data sequence is finally generated, which consists of a static adjacency matrix A and a series of node feature matrices {X} ordered by time. 1 ,X 2 ,...,X t This structured data, composed of multiple elements, preserves the multi-element characteristic information, spatial topological relationships, and temporal evolution dynamics of the monitoring system, providing unified and standardized data input for the subsequent steps S2 to S5.

[0083] In a specific embodiment of the present invention, the execution process of step S2 is described in detail. This step is performed by... Figure 1 The baseline prediction module 20 shown is executed to model the endogenous dynamics of the system under conditions free from external operational intervention and to generate a dynamic benchmark for subsequent comparison.

[0084] The baseline prediction module 20 receives the output from step S1, i.e., the spatiotemporal dynamic map data. Specifically, its input is the sequence of node feature matrices (X) within the historical time window K. t-K ,...,X t-1 The graph topology is represented by an adjacency matrix A. K is a preset positive integer whose value is determined based on the dynamic characteristics of the monitored system.

[0085] Within this module, a baseline prediction model f is deployed. evo In a specific implementation, the model can employ a Spatiotemporal Graph Convolutional Network (STGCN) architecture. This network architecture, by combining graph convolutional layers and temporal convolutional layers, can simultaneously learn and represent spatial and temporal dependencies in the data. The graph convolutional layers are used to aggregate information from each node's neighboring nodes to capture spatial interactions; the temporal convolutional layers are used to analyze the patterns of evolution of each node's own features over time.

[0086] When performing forecasting, the baseline forecasting module 20 inputs historical data sequences into the baseline forecasting model f. evo In the model, the parameters θ learned internally are used to... evo The system performs forward propagation calculations on the input data and ultimately outputs a predicted natural evolution state at the current time point t. This state is expressed as an N... v ×D dimensional prediction feature matrix To characterize it. The entire prediction process can be represented by the following formula:

[0087]

[0088] In the formula: X is the feature matrix for predicting the natural evolutionary state at time point t; t-K ,...,Xt-1 Let A be the sequence of node feature matrices from historical time point (tK) to (t-1), and let A be the adjacency matrix in the spatiotemporal dynamic graph. evo These are the internally learnable parameters of the baseline prediction model.

[0089] The predicted natural evolutionary state output by this step This characterizes the state the system would have reached if it had evolved solely according to its inherent laws. This output will be passed to subsequent steps S3 and S4, serving as one of the core benchmarks for calculating the response residuals and predicting future evolution trajectories.

[0090] In a specific embodiment of the present invention, the execution process of step S3 is described in detail. This step is performed by... Figure 1 The hypothesis generation module 30 shown is executed to detect abnormal deviations in the system state and generate one or more operational hypotheses to be verified based on the characteristic patterns of the deviations.

[0091] Suppose that the generation module 30 receives two data at its input: one is the actual observed state at the current time point t, and the other is the node feature matrix X provided by the data construction module 10. t The other is the predicted natural evolutionary state at the same time point t, represented by the predicted feature matrix provided by the baseline prediction module 20. express.

[0092] This module first performs response residual calculation. It will calculate the actual observed state X. t Compared with the predicted natural evolutionary state Perform element-wise interpolation to obtain the response residual matrix R. t The calculation process can be expressed as:

[0093]

[0094] Response residual matrix R t Non-zero or large-valued elements in the graph indicate changes at the corresponding nodes that cannot be explained by the natural evolution of the system. These changes are potential signals of external intervention.

[0095] Next, the module will calculate the response residual matrix R. t Matching with the internally stored job pulse template library I. This library is a pre-defined data structure that stores pulse templates {I1, I2, ..., I...} corresponding to M different job types. M Each pulse template I j It is a feature vector that represents the typical impact pattern of the j-th operation on the node features at the moment it occurs.

[0096] The matching process is as follows: for the response residual matrix R t For nodes that exhibit significant changes, their corresponding residual feature vectors are compared with each template I in the job pulse template library I. j Perform similarity calculation. In a specific implementation, the cosine similarity algorithm can be used for similarity calculation.

[0097] Finally, this module performs the generation of competitive job hypotheses. The system sets a similarity threshold. When the residual feature vector of a node is similar to one or more impulse templates I... j ,I k When the similarity calculation results of ... all exceed the threshold, the system generates a competitive job hypothesis H corresponding to these templates. j H k All generated competitive job hypotheses constitute a competitive job hypothesis set. This set is the final output of this step and is passed to step S4 for subsequent timing verification.

[0098] In a specific embodiment of the present invention, the execution process of step S4 is described in detail. This step is performed by... Figure 1 The hypothesis verification module 40 shown is executed by dynamically extrapolating the future impact of competing job hypotheses and continuously comparing them with real observation data to determine the final job judgment result.

[0099] Hypothesis verification module 40 receives the set of competing job hypotheses from step S3. For each competitive job in the set, assume H j This module performs the extrapolation of the future response trajectory. The extrapolation process combines the predictions of the future state from the baseline prediction module 20 with the current assumption H. j The corresponding impulse response model g resp Specifically, this module calculates the feature matrix corresponding to the future evolution trajectory of each time point t+τ (where τ∈{1,...,Np}, and Np is the length of the future projection window) within a preset future verification time window. The calculation method is as follows:

[0100]

[0101] in: Let be the feature matrix corresponding to the future evolution trajectory of the j-th competitive task at time t+τ. g represents the feature matrix corresponding to the predicted natural evolution state at a future time t+τ. resp For impulse response model; I jLet τ be the pulse template corresponding to the j-th competitive task; τ is the future time offset from the current time t; θ resp,j Let be the parameters of the impulse response model for the j-th competitive task assumption.

[0102] In the process of deduction, predict the state of natural evolution. Based on available real historical data up to time t+τ-1, the baseline prediction model f of the baseline prediction module 20 is used. evo The obtained impulse response model g resp Then it learns and represents the specific task I. j After it occurs, how its effects propagate and decay within the system over time and space.

[0103] Next, the hypothesis verification module 40 continuously extrapolates the future evolution trajectory of each hypothesis within a preset verification time window. Compared with the actual observed state X at the same time point t +τ performs a rolling comparison. The comparison result is used to calculate the mismatch degree L. verif (H j The mismatch can be quantified using functions such as weighted mean square error. The specific calculation method is as follows:

[0104]

[0105] In the formula: L verif (H j Let H be the j-th competitive task. j The degree of mismatch, where Np is the length of the verification time window, and w τ Assigning a higher weight to recent observations as a time decay weight, X t+τ The feature matrix corresponding to the actual observed state at the future time t+τ. Let ||·|| be the feature matrix corresponding to the future evolution trajectory of the j-th competitive task at time t+τ. F It is the Frobenius norm.

[0106] As the verification time window progresses, the verification module 40 continuously updates the mismatch degree of each competing job hypothesis. When the verification window ends, or when the mismatch degree of a certain hypothesis is significantly lower than that of all other competing hypotheses and is below a preset discrimination threshold, the module determines the competing job hypothesis with the smallest mismatch degree as the optimal hypothesis and outputs it as the final job discrimination result. This output is then passed to step S5.

[0107] In a specific embodiment of the present invention, the execution process of step S5 is described in detail. This step is performed by... Figure 1The model correction module 50 shown is executed, and its core is to use the confirmed job discrimination results to decouple the system causally, and use this to purify the data to correct the baseline prediction model in reverse, so as to achieve adaptive optimization of the model.

[0108] The model calibration module 50 receives the job discrimination result from step S4, i.e., the specific job type and its occurrence information determined as the optimal hypothesis. This module first initiates the causal decoupling procedure. It calls the function corresponding to the optimal hypothesis (e.g., the one corresponding to impulse template I). opt The relevant impulse response model g resp This model can accurately calculate the specific impact of the confirmed operation on the actual observed state of the system at subsequent time points.

[0109] Next, the module performs response effect stripping. For each true observed state X within the future time window after job determination (e.g., from the current time t to t+Np). t +τ, the model correction module 50 will be based on the impulse response model g resp The calculated response impact of the operation at that point in time is from X t+τ Subtract from the middle. This operation yields a "cleaned" data sequence, whose feature matrix is ​​represented as follows: The purification data is calculated as follows:

[0110]

[0111] In the formula: Let X be the feature matrix corresponding to the cleaned data at a future time t+τ. t +τ is the feature matrix corresponding to the actual observed state at the future time t+τ, g resp For the impulse response model, I opt Let τ be the pulse template corresponding to the optimal assumption, τ be the future time offset from the current time t, and θ be the pulse template. resp,opt These are the parameters of the impulse response model based on the optimal assumptions.

[0112] Clean up data The specific impacts of identified operations are theoretically eliminated, providing a more accurate reflection of the system's natural evolutionary state during that period, free from specific external interventions. This purified data is considered a high-quality, noise-free sample.

[0113] Finally, the model calibration module 50 uses this cleaned data to adjust the baseline prediction model f in the baseline prediction module 20. evo Feedback correction is performed. Specifically, this module inputs the cleaned data as new training samples into f. evo During the training process, its internal learnable parameter θ evoPerform online fine-tuning or incremental training. For example, the gradient descent optimizer can be used to calculate the loss and update the model parameters based on cleaned data. This calibration process allows the baseline prediction model to continuously learn more accurate natural evolutionary patterns of the system, thereby providing more accurate natural evolutionary predictions in future analyses. This, in turn, improves the signal-to-noise ratio of the response residuals and the overall accuracy of job identification, forming a virtuous cycle.

[0114] To further illustrate the collaborative working process of the technical solution of this invention, an application scenario example in ground-based smoke generator rain enhancement operations will be used below.

[0115] See attached document Figure 3 , Figure 3 This is a schematic diagram illustrating the application of a ground-based smoke generator in rain enhancement operations according to an embodiment of the present invention. This embodiment uses a predetermined area affected by the operation as the monitoring target, aiming to accurately determine whether the ground-based smoke generator operation produces the expected rain enhancement effect using the method and system of the present invention, and to distinguish it from the natural evolution process of cloud systems within the area.

[0116] In step S1, the data construction module 10 first acquires multi-element data from weather radar, meteorological satellites, a network of automatic weather stations, and the flue gas furnace itself within the coverage area of ​​the operation. For example, weather radar provides three-dimensional radar reflectivity factor data with high spatiotemporal resolution; meteorological satellites provide data such as cloud top brightness temperature and cloud images; the network of automatic weather stations collects information such as wind field, temperature, and humidity at key locations; and the flue gas furnace operation log provides information such as the start and stop times of the operation, geographical location, and catalyst application dosage. The data construction module 10 divides the monitored three-dimensional atmospheric space into regular grid cells (voxels), with each voxel corresponding to a graph node. The edges between nodes can be defined based on wind field information, representing the physical relationships brought about by atmospheric motion. Then, the radar reflectivity, satellite observation characteristics, and interpolated meteorological elements such as temperature, humidity, and wind within each voxel at each time point are used as node features to construct spatiotemporal dynamic map data.

[0117] In step S2, the baseline prediction module 20 receives the constructed historical spatiotemporal dynamic map data. Through a pre-trained spatiotemporal neural network model, it learns and models the natural evolution patterns of cloud formation, development, movement, and dissipation within the target area under conditions without any human intervention. For example, based on historical weather data prior to the start of operations, the model can predict the natural evolution of cloud systems in the target area under the influence of existing weather systems over the next few hours, including changes in radar echo intensity, extent, and movement path. This prediction result constitutes a dynamic background meteorological field.

[0118] In step S3, it is assumed that after the start of the flue gas furnace operation, the generation module 30 continuously receives current real meteorological observation data (such as radar echo maps) and receives the predicted natural evolution state at the same time point from the baseline prediction module 20. It is assumed that an abnormal increase in radar echo intensity is observed in the downwind region of the operation point, significantly deviating from the purely natural evolution result predicted by the baseline model. The generation module 30 calculates this deviation as a response residual matrix. This module matches this response residual with its internal operation pulse template library. This library stores typical pulse patterns for various events, such as "successful flue gas furnace catalytic operation" and "local natural convection regeneration." The template for "successful flue gas furnace catalytic operation" may exhibit a gradually increasing linear or fan-shaped echo characteristic spreading downwind from the flue gas furnace point source. Since the observed echo enhancement pattern has a high similarity to the "successful flue gas furnace catalytic operation" template, and may also have some similarity to the "local natural convection regeneration" template, the generation module 30 generates two competing operation hypotheses: "flue gas furnace operation effective" and "natural convection development."

[0119] In step S4, the hypothesis verification module 40 receives the two competing operational hypotheses and performs time-series verification. For the hypothesis of "furnace operation in effect," its corresponding impulse response model (e.g., a simplified model coupling catalyst diffusion and cloud physics processes) will deduce its future impact trajectory: catalyst particles will continue to diffuse downwind under the influence of the wind field, gradually triggering water droplet condensation or ice crystal growth in the cloud along this path, causing the radar echo to continuously strengthen and expand along a specific corridor over the next 1-2 hours. For the hypothesis of "natural convection development," its impulse response model may deduce a convective cell that strengthens in place or evolves in different forms. In the subsequent verification time window, the hypothesis verification module 40 continuously compares the two deduced future evolution trajectories (i.e., the superposition of "baseline natural evolution + operational impact") with the actually observed radar echo sequence. If the actual radar echo enhancement area and intensity evolution process highly match the downwind impact corridor deduced by the "furnace operation in effect" hypothesis, then the mismatch degree of this hypothesis will be significantly lower than that of the "natural convection development" hypothesis. Finally, the hypothesis verification module 40 determines the hypothesis with the smallest mismatch, "furnace operation effective", as the optimal hypothesis and outputs the judgment result.

[0120] In step S5, the model correction module 50 receives the determination result of "smoke furnace operation effective". It initiates causal decoupling, using the impulse response model corresponding to the "smoke furnace operation effective" assumption to accurately calculate the additional impact of this rain enhancement operation on the target area's cloud system within the time period following the determination. For example, it determines which areas, at what time, and by how much the radar echo intensity increased. Subsequently, the model correction module 50 removes this "increment" caused by human intervention from the actual observation data (radar echo map) within that time period, thus obtaining a set of "purified" meteorological data. This purified data more accurately represents the purely natural evolution path that the cloud system would have followed under this weather background without artificial rain enhancement. Finally, the model correction module 50 uses this high-quality purified data as new training samples to provide feedback correction to the spatiotemporal neural network model in the baseline prediction module 20. In this way, the baseline model can continuously learn and correct its understanding of the regional natural weather evolution patterns, enabling it to provide a more accurate "background field" in future predictions, thereby improving the accuracy and reliability of subsequent operation determinations.

[0121] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for multi-factor data analysis and job discrimination, characterized in that, Includes the following steps: S1. Acquire multi-element data of the monitoring physical system at continuous time points, construct a spatiotemporal dynamic map, and obtain spatiotemporal dynamic map data corresponding to the spatiotemporal dynamic map; wherein, the multi-element data includes multispectral or hyperspectral image data, environmental parameter data, and elevation or surface deformation data; S2. Based on the historical time point data in the spatiotemporal dynamic map data, the predicted natural evolution state at the current time point is predicted through a preset baseline prediction model. S3. Based on the predicted natural evolution state and the actual observation state at the current time point, calculate the response residual, and match the response residual with a preset task pulse template library to generate competitive task hypotheses. S4. Perform time-series verification on the competitive job hypothesis, and determine an optimal hypothesis as the job discrimination result by deducing and comparing the future evolution trajectory; The steps for performing time-series verification of the competing job assumptions include: For each competing job hypothesis, based on the future evolution trajectory calculation model, by combining the baseline prediction model and the impulse response model corresponding to the baseline prediction model, the future evolution trajectory of the competing job hypothesis within a preset verification time window is superimposed and deduced. Within the verification time window, each of the future evolution trajectories is continuously compared with the actual observation state, and the mismatch degree between each of the competing operation hypotheses and the actual observation state is calculated using the weighted mean square error formula. The competitive job hypothesis that minimizes the mismatch value is determined as the optimal hypothesis; The weighted mean square error formula is as follows: ； In the formula: For the first A competitive job hypothesis The value of the degree of mismatch with the actual observed state. To verify the length of the time window, For time decay weight, For the future The feature matrix corresponding to the actual observed state at time 1. For the first A competitive job assumption in the future The feature matrix corresponding to the future evolution trajectory at time 1. It is the Frobenius norm; S5. Based on the operation discrimination result, perform causal decoupling on the actual observation state to obtain purified data, and use the purified data to correct the baseline prediction model.

2. The multi-factor data analysis and job discrimination method according to claim 1, characterized in that, In step S1, the steps of acquiring multi-element data of the monitoring physical system at continuous time points, constructing a spatiotemporal dynamic map, and obtaining spatiotemporal dynamic map data corresponding to the spatiotemporal dynamic map include: acquiring multi-element data of the monitoring physical system at continuous time points from an external monitoring device through a communication interface; The physical space monitored by the monitoring physical system is divided into multiple monitoring units, and each monitoring unit is abstracted as a graph node. Edges between graph nodes are defined according to preset relationships between the monitoring units, and a graph structure is constructed based on the graph nodes and the edges between them. The graph nodes are... , , The number of monitoring units, and all graph nodes constitute a graph node set. , The edges between the graph nodes are The graph structure is The preset relationship is determined based on the actual relationship between the physical spaces corresponding to the monitoring units, including spatial adjacency relationship and functional association relationship; The multi-element data corresponding to the monitoring unit at each time point are assigned as node features to the corresponding graph nodes to form the spatiotemporal dynamic map and obtain the spatiotemporal dynamic map data.

3. The multi-factor data analysis and job discrimination method according to claim 1, characterized in that, In step S2, the step of predicting the predicted natural evolution state at the current time point using a preset baseline prediction model includes: The spatiotemporal dynamic graph data within the historical time window is input into the spatiotemporal graph neural network, which serves as the baseline prediction model, to obtain the predicted natural evolution state at the current time point. The baseline prediction model is as follows: ; In the formula: At the current time point The feature matrix for predicting the natural evolutionary state; For historical time points arrive The sequence of node feature matrices, The adjacency matrix in the spatiotemporal dynamic graph is... These are the internally learnable parameters of the baseline prediction model.

4. The multi-factor data analysis and job discrimination method according to claim 1, characterized in that, In step S3, the step of generating competitive job hypotheses by matching the response residual with a preset job pulse template library includes: The similarity between the response residual and the pulse templates corresponding to various job types stored in the job pulse template library is calculated. When the similarity between the response residual and the pulse template in the job pulse template library is greater than a preset similarity threshold, a competitive job hypothesis corresponding to the pulse template is generated; Wherein, the response residual is , To represent the actual observation state, At the current time point The feature matrix that predicts the natural evolutionary state.

5. The multi-factor data analysis and job discrimination method according to claim 1, characterized in that, The model for calculating the future evolutionary trajectory is as follows: ； In the formula: For the first A competitive job assumption in the future The feature matrix corresponding to the future evolution trajectory at time t; For the future The feature matrix corresponding to the predicted natural evolutionary state at time step; It is an impulse response model; For the first The pulse template corresponding to each competitive task assumption; From the current moment The future time offset; In order to target the The parameters of the impulse response model under the assumption of a competitive task.

6. The multi-factor data analysis and job discrimination method according to claim 1, characterized in that, In step S5, the step of decoupling the causal relationship of the actual observed state to obtain cleaned data includes: Using the impulse response model corresponding to the optimal assumption, the response impact of the operational assumption represented by the optimal assumption in future observation states is calculated; The purification data is obtained by subtracting the impact of the response from the actual future observed state based on the purification data calculation formula.

7. The multi-factor data analysis and job discrimination method according to claim 6, characterized in that, The formula for calculating the purification data is: ； In the formula: For the future The feature matrix corresponding to the cleaned data at any given time. For the future The feature matrix corresponding to the actual observed state at time t. The impulse response model corresponding to the optimal assumption. The pulse template corresponding to the optimal assumption. From the current moment The future time offset, These are the parameters of the impulse response model based on the optimal assumption.

8. A multi-factor data analysis and job discrimination system, applied to the multi-factor data analysis and job discrimination method according to any one of claims 1-7, characterized in that, include: The data construction module is used to acquire multi-element data of the monitored physical system at continuous time points, construct a spatiotemporal dynamic map, and obtain spatiotemporal dynamic map data corresponding to the spatiotemporal dynamic map. The baseline prediction module is used to predict the predicted natural evolution state at the current time point based on the historical time point data in the spatiotemporal dynamic map data and through a preset baseline prediction model. The hypothesis generation module is used to calculate the response residual based on the predicted natural evolution state and the actual observation state at the current time point, and to generate competitive job hypotheses by matching the response residual with a preset job pulse template library. The hypothesis verification module is used to perform time-series verification of the competing job hypotheses. By deducing and comparing future evolution trajectories, an optimal hypothesis is determined as the job discrimination result. The hypothesis verification module is used for: For each competing job hypothesis, by superimposing the baseline prediction model and its corresponding impulse response model, its future evolution trajectory within a preset verification time window is deduced; Within the verification time window, each of the future evolution trajectories is continuously compared with the actual observed state, and their respective mismatch degree is calculated. The competitive job hypothesis with the minimum mismatch is determined as the optimal hypothesis; The model correction module is used to perform causal decoupling on the actual observation state based on the operation discrimination result to obtain purified data, and to use the purified data to correct the baseline prediction model.

Images