An adaptive decision analysis system and method based on artificial intelligence

Abstract

This invention discloses an adaptive decision analysis system and method based on artificial intelligence, belonging to the field of decision analysis technology. The invention collects and monitors variables, detects local extreme points of these variables, pairs adjacent extreme points and performs piecewise linear fitting; analyzes the features of extreme point segments to construct feature vectors; statistically analyzes the time offset distribution between extreme points of two variables, and analyzes typical delay times and causality strength coefficients; analyzes propagation costs, and outputs a queue of candidate intervention paths in ascending order of cost; collects corresponding data of historical control quantities and monitored variables, constructs and fits a second-order response surface model, analyzes the optimal adjustment value by solving for stationary points on the surface, and generates instruction elements with execution windows after boundary projection; analyzes the similarity between the current scenario and historical scenarios, and corrects the causality strength coefficient and the fitting coefficient of the second-order response surface model. This reduces the amount of data analysis, improves analysis efficiency, and enhances the robustness of the system.

Description

Technical Field

[0001] This invention relates to the field of decision analysis technology, specifically an adaptive decision analysis system and method based on artificial intelligence. Background Technology

[0002] In the field of industrial automation control, such as water treatment and power dispatch, systems typically need to make real-time decisions based on continuous time-series data collected by sensors.

[0003] In the signal trend extraction stage, most methods use fixed thresholds or simple moving averages to detect extreme points. Because signal fluctuation amplitudes vary significantly across different time periods, fixed thresholds are prone to generating numerous false detections under noise interference, while in flat regions they may miss true inflection points. Furthermore, they fail to establish a closed-loop quality feedback mechanism during piecewise linear fitting. In the model transfer and adaptation stage, existing solutions either employ offline training and online fixed global models, which cannot adapt to dynamic drift in operating conditions; or they introduce complex incremental learning or deep transfer networks, which place excessive demands on real-time computing resources and sample accumulation, making them difficult to deploy on industrial controllers.

[0004] Therefore, this invention discloses an adaptive decision analysis system and method based on artificial intelligence to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive decision analysis system and method based on artificial intelligence to solve the problems raised in the prior art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: an adaptive decision analysis method based on artificial intelligence, the method comprising the following steps: S1: Collect monitoring variables, detect local extreme points of the monitoring variables, pair adjacent extreme points and perform piecewise linear fitting; analyze the features of extreme point segments to construct feature vectors; S2: Statistically analyze the time offset distribution between the extreme points of two variables, analyze the typical delay time and causality strength coefficient; then analyze the propagation cost based on the causality strength coefficient and the distance between variables, and output the candidate intervention path queue in ascending order of cost. S3: Along the first segment of the candidate intervention path queue, collect the corresponding data of historical control variables and monitoring variables, construct and fit a second-order response surface model, analyze the optimal adjustment value by solving the surface stationary point, and generate instruction elements with execution windows after boundary projection. S4: Analyze the environmental variables before each instruction execution to generate a scene fingerprint, analyze the similarity between the current scene and historical scenes, select the nearest neighbor and historical scenes, and correct the causal strength coefficient and the fitting coefficient of the second-order response surface model.

[0007] S1 includes the following: S101: Monitoring variables are collected at fixed sampling intervals. The monitoring variables include controlled variables and disturbance variables. In the field of power dispatching, the controlled variables can be indicators reflecting the system output, such as grid frequency and bus voltage, while the disturbance variables can be external inputs or interference signals, such as load power. In the field of water dosing, the controlled variables can be indicators reflecting the system output, such as sedimentation tank effluent turbidity and filtered water turbidity, while the disturbance variables can be external inputs or interference signals, such as influent turbidity and influent flow rate. For each monitored variable, the local standard deviation within the time window is analyzed with each time point as the center. The product of the local standard deviation and the proportional coefficient is recorded as the amplitude threshold. The extreme values ​​within the monitored variable sequence are analyzed based on the amplitude threshold. The extreme values ​​include local maxima and local minima; the local maxima are all greater than the previous sampling point and the next sampling point, and the difference between the local maxima and the minimum value among several sampling points before and after is greater than the amplitude threshold; the local minima are all less than the previous sampling point and the next sampling point, and the difference between the maximum value among several sampling points before and after and the local minima is greater than the amplitude threshold. Arrange all detected extreme points in chronological order and record the time of occurrence and the corresponding value of the monitored variable; S102: The detected extreme points are paired alternately in chronological order. Each pair of extreme points consists of a local maximum and an adjacent local minimum. All original monitoring variables between each pair of extreme points are recorded as extreme point segments. Linear regression is performed on the extreme point segments to analyze the slope of the corresponding line and the sum of squared residuals. The sum of squared residuals is normalized. The normalized residual is equal to the original sum of squared residuals divided by the absolute value of the difference between the local maximum and the local minimum. If the normalized residual is greater than the fitting quality threshold, the time point with the largest absolute value of the residuals in the corresponding original monitoring variables is taken as the new extreme point. The pairing and fitting are recursively repeated until the sum of squared residuals corresponding to all extreme point segments is less than or equal to the fitting quality threshold. S103: Based on the time window, extract the segmented sequences of the controlled variable and the disturbance variable respectively; normalize the slope of the straight line corresponding to all extreme point segments within the time window and record it as the normalized change intensity; encode the change type of the extreme point segments; pair the disturbance extreme point segments with the controlled extreme point segments and calculate the difference at the start time; for each successfully matched segment pair, generate a feature vector; the feature vector includes the change type of the disturbance extreme point segment, the normalized change intensity of the disturbance extreme point segment, the change type of the controlled extreme point segment, the normalized change intensity of the controlled extreme point segment, and the start time difference between the controlled extreme point segment and the disturbance extreme point segment; The specific steps of pairing the perturbation extreme point segment with the controlled extreme point segment are as follows: for each perturbation extreme point segment, find the controlled extreme point segment that overlaps in time, has the same change type, and has the smallest absolute difference at the start time, and form a segment pair between the perturbation extreme point segment and the corresponding controlled extreme point segment; This application employs a dynamic amplitude threshold based on local standard deviation, rather than a fixed threshold. This adapts to the fluctuations in signal amplitude across different time periods, preventing noise from being misidentified as extreme points and improving the accuracy of trend extraction. By comparing the normalized residual with a fitting quality threshold, new extreme points are recursively inserted at the points of maximum residual until each segment achieves a linear fit. This mechanism ensures the accuracy of the piecewise representation in approximating the original signal, avoiding the loss of key change details due to overly coarse segmentation. The continuous time-series signal is transformed into a sparse sequence of extreme point segments, described by a small number of parameters such as slope, change type, and start time, significantly reducing the data volume for subsequent causal analysis while preserving the semantic information of trend changes.

[0008] Local extrema must be such that the difference between the local extremum and the maximum / minimum values ​​among several preceding and following sampling points is greater than an amplitude threshold, and the extremum itself must be greater than or less than its neighboring points. The requirement that the difference between the extremum and its neighboring extrema exceed the threshold effectively filters out false extrema caused by plateau regions or small fluctuations, retaining only physically significant turning points. By using the condition of "greater than preceding and following neighboring points," the detected maxima and minima are ensured to alternate strictly, preventing flat segments from being incorrectly labeled as extrema, thus making the subsequent pairing process stable and reliable.

[0009] S2 includes the following: S201: Select the segment pairs that have recently undergone several changes of the type of decrease from the successfully matched segment pairs, and extract the start time difference of the selected segment pairs as the delay set; after removing the maximum and minimum values ​​from the delay set, analyze the median of the remaining values ​​and record it as the typical delay time. For any pair of controlled variables and disturbance variables, the ratio of the number of successfully matched fragment pairs to the number of disturbance extreme point fragments with a decreasing change type is denoted as the causal strength coefficient; if the causal strength coefficient is less than the causal strength threshold, it is determined that there is no significant causal relationship between the corresponding disturbance variable and the controlled variable, and the corresponding causal edge is ignored in subsequent processing; S202: The monitoring points of the monitored variables are regarded as nodes of a directed graph; in the field of power dispatching, directed graph nodes include, but are not limited to, load power monitoring points, generator power angle monitoring points, bus voltage monitoring points, and power grid frequency monitoring points; in the field of water dosing, directed graph nodes include, but are not limited to, influent turbidity monitoring points, influent flow rate monitoring points, dosing amount monitoring points, sedimentation tank effluent turbidity monitoring points, and filtered water turbidity monitoring points. Based on common knowledge of the domain or system topology, the direction of directed edges is predefined, including but not limited to the direction from the influent flow monitoring point to the sedimentation tank effluent turbidity monitoring point, and the direction from the load power monitoring point to the power grid frequency monitoring point. Analyze the propagation cost of each pair of directed edges; the propagation cost is equal to the physical or electrical distance between the two monitoring points divided by the product of the causality strength coefficient and the typical delay time; Physical distance is used in the field of water chemical dosing; per-unit values ​​of line impedance or inter-node reactance values ​​are used in the field of power dispatching. Starting from the monitoring point of the disturbance variable, a depth-first or breadth-first traversal method is used to enumerate all paths to the monitoring point of the controlled variable without duplicate nodes. The total cost of the path is the sum of the costs of each edge. Paths with a total cost exceeding a certain multiple of the minimum cost are deleted, and the remaining paths are sorted in ascending order of total path cost to obtain a candidate intervention path queue. The current risk level is analyzed and calculated; based on the current risk level, the propagation cost of all paths containing adjustable points in the candidate intervention path queue is corrected; the corrected propagation cost is equal to the original propagation cost multiplied by the adjustment coefficient; the adjustment coefficient is less than 1. The adjustable points include, but are not limited to, the frequency adjustment point of the dosing pump, the tap of the adjustable transformer, and the valve opening adjustment point; After adjustment, the path queue is reordered in ascending order of the corrected total cost, and the path with the lowest cost is selected as the optimal intervention path.

[0010] This application directly calculates the median causal delay and intensity coefficient using change type and time offset, without requiring pre-set models or fitting parameters. It is entirely data-driven and highly interpretable. The propagation cost simultaneously considers sensor physical distance, causal intensity coefficient, and median delay, organically combining domain knowledge with data characteristics to make path ranking more practical for engineering applications. The path cost, including adjustable nodes, is dynamically adjusted based on the current risk level, allowing the system to prioritize actively intervened paths in high-risk situations, improving decision-making safety. Paths with a total cost exceeding 10 times the minimum cost are deleted, effectively controlling enumeration explosion and ensuring real-time performance.

[0011] S3 includes the following: S301: Mark the first adjustable point in the optimal intervention path; retrieve the corresponding monitoring points of the control quantity and the controlled variable of the first adjustable point from the historical database within a certain period of time; filter the collected data: retain only the data in which the historical control quantity is within a certain percentage range of the current value control quantity; Using the control variable u and the disturbance variable d as two independent variables, and the controlled variable as the dependent variable; with the current operating point (u0, d0) as the origin, u0 representing the current control variable and d0 representing the current disturbance variable, the independent variables are centered to construct a second-order response surface model: ; Where △u = u - u0, △d = d - d0; a0, a1, a2, a 11 a 12 a 22 These are the fitting coefficients; The fitting coefficients are solved by the least squares method, and the coefficients of determination are analyzed based on the well-fitted second-order response surface model. If the coefficient of determination is less than the threshold, expand the percentage range, re-filter the data and fit a second-order response surface model; if the coefficient of determination is still less than the threshold after expanding the percentage range, re-select the path with the second smallest cost until the coefficient of determination is less than the threshold. S302: Based on the fitted second-order response surface model, find the optimal adjustment value of the controlled variable when the current disturbance variable is fixed; The solution is as follows: Take the partial derivative of the second-order response surface model with respect to Δu, set it to zero, substitute Δd=0 into the equation, and we get a1+a 11 ×(u-u0)=0; Solve for the optimal adjustment value: If a 11 If ≠0, then the optimal adjustment value = u0 - a1 / (2a 11 ); If a 11 If a = 0, then the objective function is a monotonic function. In this case, within the allowable control range, the boundary value that minimizes the controlled variable is taken: if a1 is greater than 0, then the optimal adjustment value is equal to the minimum control value; if a is less than 0, then the optimal adjustment value is equal to the maximum control value; if a1 is equal to 0, then the optimal adjustment value is equal to the current control value.

[0012] If the optimal adjustment value exceeds the physical limit range [umin, umax], the optimal adjustment value is adjusted; the minimum physical limit value umin and the maximum physical limit value umax are determined by the physical characteristics of the actuator, such as the lower limit of the dosing pump frequency of 0Hz and the upper limit of 50Hz. If the optimal adjustment value is less than the minimum control value, then the projection to the optimal adjustment value is equal to the minimum control value; if the optimal adjustment value is greater than the maximum control value, then the projection to the optimal adjustment value is equal to the maximum control value. S303: Set the instruction duration and generate instruction elements; the instruction elements include an identifier of the adjustable point, the optimal adjustment value, the current time, and the effective expiration time of the instruction; Based on the fitted second-order response surface model, predict the expected value of the controlled variable ypred after the execution of the instruction: ypred = a0 + a1 × (optimal adjustment value - u0) + a 11 × (Optimal adjustment value - u0) 2 ; After the instruction is executed, the actual value of the controlled variable yreal is collected, and the sign of the expected deviation is recorded; This application employs a second-order response surface model instead of a fixed linear model, which can approximate the nonlinear characteristics of the dosing process and improve control accuracy. The analytical expression for the optimal control quantity is obtained directly by taking the partial derivative and setting it to zero, eliminating the need for iterative optimization, reducing computational complexity, and making it suitable for real-time decision-making. A fallback mechanism is defined for failures in response surface fitting: if the coefficient of determination is below a threshold, the data range is expanded for retry; if the threshold is still not met, the system switches to a suboptimal intervention path. This avoids using poorly fitted surface models for decision-making, preventing erroneous commands. Expanding the percentage range incorporates more samples, improving fitting stability while maintaining locality. When the current path cannot establish a reliable model, the system automatically switches to the second-lowest-cost path, ensuring the system always has a usable control scheme and improving system robustness.

[0013] S4 includes the following: S401: Before each instruction is executed, collect the environmental variable vector within the time window. In the water dosing scenario, the environmental variable vector = [mean influent turbidity, variance influent turbidity, mean flow rate, water temperature], and in the power dispatch scenario, the environmental variable vector = [mean load power, variance of load fluctuation, penetration rate of new energy]. Analyze the first-order moment mean vector and the second-order moment covariance matrix based on the number of sampling points within the time window, and concatenate the first-order moment mean vector and the second-order moment covariance matrix to form a scenario fingerprint. S402: Analyze the Mahalanobis distance between the current scene fingerprint and the historical scene fingerprints in the historical scene database. If the minimum Mahalanobis distance is greater than the Mahalanobis distance threshold, it is determined that there are no similar scenes and no migration is performed; otherwise, the causal strength coefficient and fitting coefficient of the historical scene are extracted. Analyze the standard deviations of the control variables and the controlled variables in the current and historical scenarios; correct the fitting coefficients. ; ; Among them, a （1，new） and a （11，new） a represents the corrected fitting coefficient. （1，i） and a （11，i） σ represents the historical fitting coefficient corresponding to historical scene i with the minimum Mahalanobis distance. （y，new） σ represents the standard deviation of the controlled variable values ​​in the current scenario. （u，new） σ represents the standard deviation of the control variables in the current scenario. （y，i） σ represents the standard deviation of the controlled variable values ​​corresponding to the historical scenario with the minimum Mahalanobis distance. （u，i） The standard deviation of the control variable corresponding to the historical scenario with the minimum Mahalanobis distance; The causal intensity coefficients for the current scene are directly adopted from the causal intensity coefficients for the historical scene corresponding to the minimum Mahalanobis distance; the corrected fitting coefficients are filled into the second-order response surface model of the current scene, replacing the original fitting coefficients.

[0014] This application uses first and second moments as statistical fingerprints of environmental variables, which concisely and fully describe the operating point and fluctuation characteristics of the scene, avoiding the need to store the original sequence. Based on the ratio of the standard deviations of the control and controlled variables in the new and historical scenes, the fitting coefficients are linearly scaled, allowing the transferred model to automatically adapt to the dimensions and sensitivity of the new scene without retraining.

[0015] An artificial intelligence-based adaptive decision analysis system is provided, which is applied to the aforementioned artificial intelligence-based adaptive decision analysis method. The system includes a data pairing analysis module, a propagation filtering module, a decision analysis module, and a scene matching correction module. The data pairing analysis module is used to collect monitoring variables, detect local extreme points of the monitoring variables, pair adjacent extreme points and perform piecewise linear fitting; analyze the features of extreme point segments to construct feature vectors; The propagation screening module is used to statistically analyze the time offset distribution between the extreme points of two variables, analyze the typical delay time and the causality strength coefficient, and then analyze the propagation cost based on the causality strength coefficient and the distance between variables, and output the candidate intervention path queue in ascending order of cost. The decision analysis module is used to collect corresponding data of historical control variables and monitoring variables along the first segment of the candidate intervention path queue, construct and fit a second-order response surface model, solve the surface stationary point analysis for the optimal adjustment value, and generate instruction elements with execution windows after boundary projection. The scene matching and correction module is used to analyze environmental variables before each instruction execution to generate scene fingerprints, analyze the similarity between the current scene and historical scenes, select the nearest neighbor and historical scenes, and correct the causal intensity coefficient and the fitting coefficient of the second-order response surface model.

[0016] The data pairing analysis module includes an extreme point filtering unit, an extreme point pairing unit, and a feature analysis unit. The extreme point screening unit is used to collect monitoring variables at fixed sampling intervals, analyze the local standard deviation within the time window centered on each time moment for each monitoring variable, and record the product of the local standard deviation and the proportional coefficient as the amplitude threshold. Based on the amplitude threshold, the extreme values ​​within the monitoring variable sequence are analyzed. The extreme point pairing unit is used to alternately pair the detected extreme points in chronological order; The feature analysis unit is used to extract segmented sequences of the controlled variable and the disturbance variable based on a time window; normalize the slope of the line corresponding to all extreme point segments within the time window and record it as the normalized change intensity; encode the change type of the extreme point segments; pair the disturbance extreme point segments with the controlled extreme point segments and calculate the difference at the starting time; and generate a feature vector for each successfully matched segment pair. The propagation screening module includes a causal strength analysis unit and a propagation path analysis unit; The causal strength analysis unit is used to filter out fragment pairs that have recently undergone several changes of the decreasing type from successfully matched fragment pairs, extract the start time difference of the filtered fragment pairs as the delay set, analyze the typical delay time of the delay set, and analyze the causal strength coefficient between the controlled variable and the disturbance variable. The propagation path analysis unit is used to treat the monitoring points of the monitored variables as nodes in a directed graph; pre-determine the direction of the directed edges based on domain common sense or system topology; analyze the propagation cost of each pair of directed edges; starting from the monitoring point of the disturbance variable, enumerate all paths to the monitoring point of the controlled variable without duplicate nodes using a depth-first or breadth-first traversal method, and the total cost of the path is the sum of the costs of each edge; delete paths whose total cost exceeds a certain multiple of the minimum cost, and sort the remaining paths in ascending order of total path cost to obtain a candidate intervention path queue; The decision analysis module includes a model fitting analysis unit, an optimal adjustment value analysis unit, and an instruction confirmation unit; The model fitting analysis unit is used to mark the first adjustable point in the optimal intervention path; retrieve the corresponding monitoring points of the control quantity and the controlled variable of the first adjustable point from the historical database within a certain period of time; screen the collected monitoring points; solve the fitting coefficients by the least squares method; analyze the determination coefficients based on the fitted second-order response surface model; and refit based on the determination coefficients. The optimal adjustment value analysis unit is used to solve for the optimal adjustment value of the controlled variable when the current disturbance variable is fixed, based on the fitted second-order response surface model; and to adjust the optimal adjustment value that exceeds the physical limit. The instruction confirmation unit is used to set the instruction duration and generate instruction elements; predict the expected value of the controlled variable after the instruction is executed based on the fitted second-order response surface model; collect the actual value of the controlled variable after the instruction is executed and record the sign of the expected deviation. The scene matching correction module includes a scene fingerprint analysis unit and a historical scene analysis correction unit; The scene fingerprint analysis unit is used to collect environmental variable vectors within a time window before each instruction is executed; and to analyze the scene fingerprint based on the environmental variable vectors. The historical scene analysis and correction unit is used to analyze the Mahalanobis distance between the current scene fingerprint and the historical scene fingerprint in the historical scene database; and to correct the fitting coefficient and causal strength coefficient of the current scene based on the historical scene.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: This application adopts a dynamic amplitude threshold based on local standard deviation, rather than a fixed threshold, which can adapt to the fluctuation amplitude of signals in different time periods, avoid noise being misjudged as extreme points, and improve the accuracy of trend extraction. By comparing the normalized residual with the fitting quality threshold, new extreme points are recursively inserted at the point of maximum residual until each segment of linear fitting meets the standard. This mechanism ensures the accuracy of the piecewise representation in approximating the original signal and avoids the loss of key change details due to overly coarse segmentation. Continuous time-series signals are transformed into sparse extreme point fragment sequences and described by a small number of parameters such as slope, change type, and start time, which significantly reduces the amount of data for subsequent causal analysis while preserving the semantic information of trend changes.

[0018] Local extrema must be such that the difference between the local extremum and the maximum / minimum values ​​among several preceding and following sampling points is greater than an amplitude threshold, and the extremum itself must be greater than or less than its neighboring points. The requirement that the difference between the extremum and its neighboring extrema exceed the threshold effectively filters out false extrema caused by plateau regions or small fluctuations, retaining only physically significant turning points. By using the condition of "greater than preceding and following neighboring points," the detected maxima and minima are ensured to alternate strictly, preventing flat segments from being incorrectly labeled as extrema, thus making the subsequent pairing process stable and reliable.

[0019] This application directly calculates the median causal delay and intensity coefficient using change type and time offset, without requiring pre-set models or fitting parameters. It is entirely data-driven and highly interpretable. The propagation cost simultaneously considers sensor physical distance, causal intensity coefficient, and median delay, organically combining domain knowledge with data characteristics to make path ranking more practical for engineering applications. The path cost, including adjustable nodes, is dynamically adjusted based on the current risk level, allowing the system to prioritize actively intervened paths in high-risk situations, improving decision-making safety. Paths with a total cost exceeding 10 times the minimum cost are deleted, effectively controlling enumeration explosion and ensuring real-time performance.

[0020] This application employs a second-order response surface model instead of a fixed linear model, which can approximate the nonlinear characteristics of the dosing process and improve control accuracy. The analytical expression for the optimal control quantity is obtained directly by taking the partial derivative and setting it to zero, eliminating the need for iterative optimization, reducing computational complexity, and making it suitable for real-time decision-making. A fallback mechanism is defined for failures in response surface fitting: if the coefficient of determination is below a threshold, the data range is expanded for retry; if the threshold is still not met, the system switches to a suboptimal intervention path. This avoids using poorly fitted surface models for decision-making, preventing erroneous commands. Expanding the percentage range incorporates more samples, improving fitting stability while maintaining locality. When the current path cannot establish a reliable model, the system automatically switches to the second-lowest-cost path, ensuring the system always has a usable control scheme and improving system robustness.

[0021] This application uses first and second moments as statistical fingerprints of environmental variables, which concisely and fully describe the operating point and fluctuation characteristics of the scene, avoiding the need to store the original sequence. Based on the ratio of the standard deviations of the control and controlled variables in the new and historical scenes, the fitting coefficients are linearly scaled, allowing the transferred model to automatically adapt to the dimensions and sensitivity of the new scene without retraining. Attached Figure Description

[0022] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating an adaptive decision analysis method based on artificial intelligence according to the present invention. Figure 2 This is a schematic diagram of the structure of an adaptive decision analysis system based on artificial intelligence according to the present invention. Detailed Implementation

[0023] 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.

[0024] Please see Figure 1 The present invention provides a technical solution: an adaptive decision analysis method based on artificial intelligence, the method comprising the following steps: S1: Collect monitoring variables, detect local extreme points of the monitoring variables, pair adjacent extreme points and perform piecewise linear fitting; analyze the features of extreme point segments to construct feature vectors; S1 includes the following: S101: Monitored variables are collected at fixed sampling intervals. The monitored variables include controlled variables and disturbance variables. In the field of power dispatching, the controlled variables can be indicators reflecting the system output, such as grid frequency and bus voltage, while the disturbance variables can be external inputs or interference signals, such as load power. In the field of water dosing, the controlled variables can be indicators reflecting the system output, such as sedimentation tank effluent turbidity and filtered water turbidity, while the disturbance variables can be external inputs or interference signals, such as influent turbidity and influent flow rate. For each monitored variable, the local standard deviation within the time window is analyzed with each time point as the center. The product of the local standard deviation and the proportional coefficient is recorded as the amplitude threshold. The extreme values ​​within the monitored variable sequence are analyzed based on the amplitude threshold. Extreme values ​​include local maxima and local minima; local maxima are all greater than the previous and next sampling points, and the difference between the local maximum and the minimum value among several sampling points before and after is greater than the amplitude threshold; local minima are all less than the previous and next sampling points, and the difference between the maximum value among several sampling points before and after and the local minima is greater than the amplitude threshold. Arrange all detected extreme points in chronological order and record the time of occurrence and the corresponding value of the monitored variable; S102: The detected extreme points are paired alternately in chronological order. Each pair of extreme points consists of a local maximum and an adjacent local minimum. All original monitoring variables between each pair of extreme points are recorded as extreme point segments. Linear regression is performed on the extreme point segments to analyze the slope of the corresponding line and the sum of squared residuals. The sum of squared residuals is normalized. The normalized residual is equal to the original sum of squared residuals divided by the absolute value of the difference between the local maximum and the local minimum. If the normalized residual is greater than the fitting quality threshold, the time point with the largest absolute value of the residuals in the corresponding original monitoring variables is taken as the new extreme point. The pairing and fitting are recursively repeated until the sum of squared residuals corresponding to all extreme point segments is less than or equal to the fitting quality threshold. S103: Based on the time window, extract the segmented sequences of the controlled variable and the disturbance variable respectively; normalize the slope of the straight line corresponding to all extreme point segments within the time window and record it as the normalized change intensity; encode the change type of the extreme point segments; pair the disturbance extreme point segments with the controlled extreme point segments and calculate the difference at the start time; for each successfully matched segment pair, generate a feature vector; the feature vector includes the change type of the disturbance extreme point segment, the normalized change intensity of the disturbance extreme point segment, the change type of the controlled extreme point segment, the normalized change intensity of the controlled extreme point segment, and the start time difference between the controlled extreme point segment and the disturbance extreme point segment; The specific process of pairing the perturbation extreme point segment with the controlled extreme point segment is as follows: for each perturbation extreme point segment, find the controlled extreme point segment that overlaps in time, has the same change type, and has the smallest absolute difference at the start time, and form a segment pair between the perturbation extreme point segment and the corresponding controlled extreme point segment. Example 1: In this example, the change type code τ = +1 when rising and the change type code τ = -1 when falling; S2: Statistically analyze the time offset distribution between the extreme points of two variables, analyze the typical delay time and causality strength coefficient; then analyze the propagation cost based on the causality strength coefficient and the distance between variables, and output the candidate intervention path queue in ascending order of cost. S2 includes the following: S201: From the successfully matched fragment pairs, select the fragment pairs that have recently undergone several changes of the decreasing type, and extract the starting time difference of the selected fragment pairs as the delay set; after removing the maximum and minimum values ​​from the delay set, analyze the median of the remaining values ​​and record it as the typical delay time. For any pair of controlled variables and disturbance variables, the ratio of the number of successfully matched fragment pairs to the number of disturbance extreme point fragments with a decreasing change type is denoted as the causal strength coefficient; if the causal strength coefficient is less than the causal strength threshold, it is determined that there is no significant causal relationship between the corresponding disturbance variable and the controlled variable, and the corresponding causal edge is ignored in subsequent processing; S202: The monitoring points of the monitored variables are regarded as nodes of a directed graph; in the field of power dispatching, the nodes of the directed graph include, but are not limited to, load power monitoring points, generator power angle monitoring points, bus voltage monitoring points, and power grid frequency monitoring points; in the field of water dosing, the nodes of the directed graph include, but are not limited to, influent turbidity monitoring points, influent flow rate monitoring points, dosing amount monitoring points, sedimentation tank effluent turbidity monitoring points, and filtered water turbidity monitoring points. Based on common knowledge of the domain or system topology, the direction of directed edges is predefined, including but not limited to the direction from the influent flow monitoring point to the sedimentation tank effluent turbidity monitoring point, and the direction from the load power monitoring point to the power grid frequency monitoring point. Analyze the propagation cost of each pair of directed edges; the propagation cost equals the physical or electrical distance between the two monitoring points divided by the product of the causality strength coefficient and the typical delay time. Physical distance is used in the field of water chemical dosing; per-unit values ​​of line impedance or inter-node reactance values ​​are used in the field of power dispatching. Starting from the monitoring point of the disturbance variable, a depth-first or breadth-first traversal method is used to enumerate all paths to the monitoring point of the controlled variable without duplicate nodes. The total cost of the path is the sum of the costs of each edge. Paths with a total cost exceeding a certain multiple of the minimum cost are deleted, and the remaining paths are sorted in ascending order of total path cost to obtain a candidate intervention path queue. Analyze and calculate the current risk level; based on the current risk level, adjust the propagation cost of all paths containing adjustable points in the candidate intervention path queue; the adjusted propagation cost is equal to the original propagation cost multiplied by the adjustment coefficient; the adjustment coefficient is less than 1; Adjustable points include, but are not limited to, the frequency adjustment point of the dosing pump, the tap of the adjustable transformer, and the valve opening adjustment point; Example 2: In this example, the risk level is quantified and graded by the distance between the current value of the monitored variable and the preset alarm threshold: when the distance is greater than 2 times the standard deviation, the risk level is 0, indicating normal; when the distance is between 1 and 2 times the standard deviation, the risk level is 1, indicating caution; when the distance is less than 1 time the standard deviation, the risk level is 2, indicating an alarm. After adjustment, the path queue is reordered in ascending order of the corrected total cost, and the path with the lowest cost is selected as the optimal intervention path.

[0025] S3: Along the first segment of the candidate intervention path queue, collect the corresponding data of historical control variables and monitoring variables, construct and fit a second-order response surface model, analyze the optimal adjustment value by solving the surface stationary point, and generate instruction elements with execution windows after boundary projection. S3 includes the following: S301: Mark the first adjustable point in the optimal intervention path; retrieve the corresponding monitoring points of the control quantity and the controlled variable of the first adjustable point within a certain period of time from the historical database; filter the collected data: retain only the data in which the historical control quantity is within a certain percentage range of the current value control quantity; Using the control variable u and the disturbance variable d as two independent variables, and the controlled variable as the dependent variable; with the current operating point (u0, d0) as the origin, u0 representing the current control variable and d0 representing the current disturbance variable, the independent variables are centered to construct a second-order response surface model: ; Where △u = u - u0, △d = d - d0; a0, a1, a2, a 11 a 12 a 22 These are the fitting coefficients; The fitting coefficients are solved by the least squares method, and the coefficients of determination are analyzed based on the well-fitted second-order response surface model. Example 3: In this implementation, the coefficient of determination R² = 1 - (sum of squares of residuals) / (sum of squares of total squares); If the coefficient of determination is less than the threshold, expand the percentage range, re-filter the data and fit a second-order response surface model; if the coefficient of determination is still less than the threshold after expanding the percentage range, re-select the path with the second smallest cost until the coefficient of determination is less than the threshold. S302: Based on the fitted second-order response surface model, find the optimal adjustment value of the controlled variable when the current disturbance variable is fixed; The solution is as follows: Take the partial derivative of the second-order response surface model with respect to Δu, set it to zero, substitute Δd=0 into the equation, and we get a1+a 11 ×(u-u0)=0; Solve for the optimal adjustment value: If a 11 If ≠0, then the optimal adjustment value = u0 - a1 / (2a 11 ); If a 11 If a1 = 0, then the objective function is a monotonic function. In this case, within the allowable control range, the boundary value that minimizes the controlled variable is taken: if a1 is greater than 0, then the optimal adjustment value is equal to the minimum control value; if a is less than 0, then the optimal adjustment value is equal to the maximum control value; if a1 equals 0, then the optimal adjustment value is equal to the current control value. If the optimal adjustment value exceeds the physical limit range [umin, umax], the optimal adjustment value is adjusted; the minimum physical limit value umin and the maximum physical limit value umax are determined by the physical characteristics of the actuator, such as the lower limit of the dosing pump frequency of 0Hz and the upper limit of 50Hz. If the optimal adjustment value is less than the minimum control value, then the projection to the optimal adjustment value is equal to the minimum control value; if the optimal adjustment value is greater than the maximum control value, then the projection to the optimal adjustment value is equal to the maximum control value. S303: Set the instruction duration and generate instruction elements; the instruction elements include the identifier of the adjustable point, the optimal adjustment value, the current time, and the effective expiration time of the instruction; Based on the fitted second-order response surface model, predict the expected value of the controlled variable ypred after the execution of the instruction: ypred = a0 + a1 × (optimal adjustment value - u0) + a 11 × (Optimal adjustment value - u0) 2 ; After the instruction is executed, the actual value of the controlled variable yreal is collected, and the sign of the expected deviation is recorded; Example 4, in this example, the expected deviation is esign=sign(yreal-ypred). S4: Analyze the environmental variables before each instruction execution to generate a scene fingerprint, analyze the similarity between the current scene and historical scenes, select the nearest neighbor and historical scenes, and correct the causal strength coefficient and the fitting coefficient of the second-order response surface model.

[0026] S4 includes the following: S401: Before each instruction is executed, collect the environmental variable vector within the time window. In the water dosing scenario, the environmental variable vector = [mean influent turbidity, variance influent turbidity, mean flow rate, water temperature], and in the power dispatch scenario, the environmental variable vector = [mean load power, variance of load fluctuation, penetration rate of new energy]. Based on the number of sampling points within the time window, analyze the first-order moment mean vector and the second-order moment covariance matrix, and concatenate the first-order moment mean vector and the second-order moment covariance matrix to form a scenario fingerprint. S402: Analyze the Mahalanobis distance between the current scene fingerprint and the historical scene fingerprints in the historical scene database. If the minimum Mahalanobis distance is greater than the Mahalanobis distance threshold, it is determined that there are no similar scenes and no migration is performed; otherwise, the causal strength coefficient and fitting coefficient of the historical scene are extracted. Analyze the standard deviations of the control variables and the controlled variables in the current and historical scenarios; correct the fitting coefficients. ; ; Among them, a （1，new） and a （11，new） a represents the corrected fitting coefficient. （1，i） and a （11，i） σ represents the historical fitting coefficient corresponding to the historical scene with the minimum Mahalanobis distance. （y，new） σ represents the standard deviation of the controlled variable values ​​in the current scenario. （u，new） σ represents the standard deviation of the control variables in the current scenario. （y，i） σ represents the standard deviation of the controlled variable values ​​corresponding to the historical scenario with the minimum Mahalanobis distance. （u，i） The standard deviation of the control variables corresponding to the historical scenario with the minimum Mahalanobis distance; The causal intensity coefficients for the current scene are directly adopted from the causal intensity coefficients for the historical scene corresponding to the minimum Mahalanobis distance; the corrected fitting coefficients are filled into the second-order response surface model of the current scene, replacing the original fitting coefficients.

[0027] Please see Figure 2 The present invention provides a technical solution: an adaptive decision analysis system based on artificial intelligence, which includes a data pairing analysis module, a propagation filtering module, a decision analysis module and a scene matching correction module; The data pairing analysis module is used to collect monitoring variables, detect local extreme points of the monitoring variables, pair adjacent extreme points and perform piecewise linear fitting; analyze the features of extreme point segments to construct feature vectors; The propagation screening module is used to statistically analyze the time offset distribution between the extreme points of two variables, analyze the typical delay time and the causality strength coefficient, and then analyze the propagation cost based on the causality strength coefficient and the distance between variables, and output the candidate intervention path queue in ascending order of cost. The decision analysis module is used to collect corresponding data of historical control variables and monitoring variables along the first segment of the candidate intervention path queue, construct and fit a second-order response surface model, analyze the optimal adjustment value by solving the surface stationary point, and generate instruction elements with execution windows after boundary projection. The scene matching correction module is used to analyze the environmental variables before each instruction execution to generate scene fingerprints, analyze the similarity between the current scene and historical scenes, select the nearest neighbor and historical scenes, and correct the causal strength coefficient and the fitting coefficient of the second-order response surface model.

[0028] The data pairing analysis module includes an extreme point filtering unit, an extreme point pairing unit, and a feature analysis unit; The extreme point screening unit is used to collect monitoring variables at fixed sampling intervals. For each monitoring variable, the local standard deviation within the time window is analyzed with each time point as the center. The product of the local standard deviation and the proportional coefficient is recorded as the amplitude threshold. The extreme values ​​within the monitoring variable sequence are analyzed based on the amplitude threshold. The extreme point pairing unit is used to alternately pair the detected extreme points in chronological order; The feature analysis unit is used to extract segmented sequences of controlled and disturbance variables based on time windows; normalize the slope of the straight line corresponding to all extreme point segments within the time window and record it as the normalized change intensity; encode the change type of the extreme point segments; pair the disturbance extreme point segments with the controlled extreme point segments and calculate the difference at the starting time; generate a feature vector for each successfully matched segment pair. The propagation screening module includes a causal strength analysis unit and a propagation path analysis unit; The causal strength analysis unit is used to filter out the fragment pairs that have recently undergone several changes of the decreasing type from the successfully matched fragment pairs, extract the start time difference of the selected fragment pairs and record it as the delay set; analyze the typical delay time of the delay set; and analyze the causal strength coefficient between the controlled variable and the disturbance variable. The propagation path analysis unit treats the monitoring points of the monitored variables as nodes in a directed graph; it pre-defines the directions of the directed edges based on domain common sense or system topology; it analyzes the propagation cost of each pair of directed edges; starting from the monitoring point of the disturbance variable, it enumerates all paths to the monitoring point of the controlled variable without duplicate nodes using a depth-first or breadth-first traversal method, and the total cost of the path is the sum of the costs of each edge; it deletes paths whose total cost exceeds a certain multiple of the minimum cost, and sorts the remaining paths in ascending order of total path cost to obtain a candidate intervention path queue. The decision analysis module includes a model fitting analysis unit, an optimal adjustment value analysis unit, and an instruction confirmation unit; The model fitting analysis unit is used to mark the first adjustable point in the optimal intervention path; retrieve the corresponding monitoring points of the control quantity and the controlled variable of the first adjustable point within a certain period of time from the historical database; screen the collected monitoring points; solve the fitting coefficients by the least squares method; analyze the coefficient of determination based on the fitted second-order response surface model; and refit based on the coefficient of determination. The optimal adjustment value analysis unit is used to solve for the optimal adjustment value of the controlled variable when the current disturbance variable is fixed, based on the fitted second-order response surface model; and to adjust the optimal adjustment value that exceeds the physical limit. The instruction confirmation unit is used to set the instruction duration and generate instruction elements; based on the fitted second-order response surface model, it predicts the expected value of the controlled variable after the instruction is executed; after the instruction is executed, it collects the actual value of the controlled variable and records the sign of the expected deviation. The scene matching correction module includes a scene fingerprint analysis unit and a historical scene analysis correction unit; The scene fingerprint analysis unit is used to collect the environment variable vector within a time window before each instruction is executed; and to analyze the scene fingerprint based on the environment variable vector. The historical scene analysis and correction unit is used to analyze the Mahalanobis distance between the current scene fingerprint and the historical scene fingerprints in the historical scene database; and to correct the fitting coefficient and causal strength coefficient of the current scene based on the historical scene.

[0029] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0030] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An adaptive decision analysis method based on artificial intelligence, characterized in that, The method includes the following steps: S1: Collect monitoring variables, detect local extreme points of the monitoring variables, pair adjacent extreme points and perform piecewise linear fitting; analyze the features of extreme point segments to construct feature vectors; S2: Statistically analyze the time offset distribution between the extreme points of two variables, analyze the typical delay time and causality strength coefficient; then analyze the propagation cost based on the causality strength coefficient and the distance between variables, and output the candidate intervention path queue in ascending order of cost. S3: Along the first segment of the candidate intervention path queue, collect the corresponding data of historical control variables and monitoring variables, construct and fit a second-order response surface model, analyze the optimal adjustment value by solving the surface stationary point, and generate instruction elements with execution windows after boundary projection. S4: Analyze the environmental variables before each instruction execution to generate a scene fingerprint, analyze the similarity between the current scene and historical scenes, select the nearest neighbor and historical scenes, and correct the causal strength coefficient and the fitting coefficient of the second-order response surface model.

2. The adaptive decision analysis method based on artificial intelligence according to claim 1, characterized in that: S1 includes the following: S101: Collect monitoring variables at fixed sampling intervals. The monitoring variables include controlled variables and disturbance variables. For each monitoring variable, analyze the local standard deviation within a time window centered on each time point. Record the product of the local standard deviation and the proportional coefficient as the amplitude threshold. Analyze the extreme values ​​within the monitoring variable sequence based on the amplitude threshold. The extreme values ​​include local maxima and local minima. Arrange all detected extreme value points in chronological order and record the occurrence time and corresponding monitoring variable values. S102: The detected extreme points are paired alternately in chronological order. Each pair of extreme points consists of a local maximum and an adjacent local minimum. All original monitoring variables between each pair of extreme points are recorded as extreme point segments. Linear regression is performed on the extreme point segments to analyze the slope of the corresponding line and the sum of squared residuals. The sum of squared residuals is normalized. The normalized residual is equal to the original sum of squared residuals divided by the absolute value of the difference between the local maximum and the local minimum. If the normalized residual is greater than the fitting quality threshold, the time point with the largest absolute value of the residuals in the corresponding original monitoring variables is taken as the new extreme point. The pairing and fitting are recursively repeated until the sum of squared residuals corresponding to all extreme point segments is less than or equal to the fitting quality threshold. S103: Based on the time window, extract the segmented sequences of the controlled variable and the disturbance variable respectively; normalize the slope of the line corresponding to all extreme point segments within the time window and record it as the normalized change intensity; encode the change type of the extreme point segments; pair the disturbance extreme point segments with the controlled extreme point segments and calculate the difference at the starting time; generate a feature vector for each successfully matched segment pair.

3. The adaptive decision analysis method based on artificial intelligence according to claim 2, characterized in that: The local maxima are all greater than the previous and next sampling points, and the difference between the local maxima and the minimum value among several sampling points before and after is greater than the amplitude threshold; the local minima are all less than the previous and next sampling points, and the difference between the maximum value among several sampling points before and after and the local minima is greater than the amplitude threshold.

4. The adaptive decision analysis method based on artificial intelligence according to claim 3, characterized in that: The specific steps of pairing the perturbation extreme point segment with the controlled extreme point segment are as follows: for each perturbation extreme point segment, find the controlled extreme point segment that overlaps in time, has the same change type, and has the smallest absolute difference at the start time, and form a segment pair between the perturbation extreme point segment and the corresponding controlled extreme point segment.

5. The adaptive decision analysis method based on artificial intelligence according to claim 2, characterized in that: S2 includes the following: S201: Select the segment pairs that have recently undergone several changes of the type of decrease from the successfully matched segment pairs, and extract the start time difference of the selected segment pairs as the delay set; after removing the maximum and minimum values ​​from the delay set, analyze the median of the remaining values ​​and record it as the typical delay time. For any pair of controlled variables and disturbance variables, the ratio of the number of successfully matched fragment pairs to the number of disturbance extreme point fragments with a decreasing change type is denoted as the causal strength coefficient; if the causal strength coefficient is less than the causal strength threshold, it is determined that there is no significant causal relationship between the corresponding disturbance variable and the controlled variable, and the corresponding causal edge is ignored in subsequent processing; S202: Treat the monitoring points of the monitored variables as nodes in a directed graph; pre-determine the direction of the directed edges based on domain common sense or system topology; Starting from the monitoring point of the disturbance variable, a depth-first or breadth-first traversal method is used to enumerate all paths to the monitoring point of the controlled variable without duplicate nodes. The total cost of the path is the sum of the costs of each edge. Delete paths whose total cost exceeds a certain multiple of the minimum cost, and sort the remaining paths in ascending order of total cost to obtain a queue of candidate intervention paths. Analyze and calculate the current risk level; Based on the current risk level, the propagation cost of all paths containing adjustable points in the candidate intervention path queue is adjusted; the adjusted propagation cost is equal to the original propagation cost multiplied by the adjustment coefficient. After adjustment, the path queue is reordered in ascending order of the corrected total cost, and the path with the lowest cost is selected as the optimal intervention path.

6. The adaptive decision analysis method based on artificial intelligence according to claim 5, characterized in that: S3 includes the following: S301: Mark the first adjustable point in the optimal intervention path; retrieve the corresponding monitoring points of the control quantity and the controlled variable of the first adjustable point from the historical database within a certain period of time; The collected data were filtered; the control variable and the disturbance variable were used as two independent variables, and the controlled variable was used as the dependent variable; the independent variables were centered, and a second-order response surface model was constructed. The fitting coefficients are solved by the least squares method, and the coefficients of determination are analyzed based on the well-fitted second-order response surface model. S302: Based on the fitted second-order response surface model, find the optimal adjustment value of the controlled variable when the current disturbance variable is fixed; If the optimal adjustment value exceeds the physical limit range, the optimal adjustment value is adjusted; if the optimal adjustment value is less than the minimum control value, the optimal adjustment value is projected to be equal to the minimum control value. If the optimal adjustment value is greater than the maximum control value, then the projection to the optimal adjustment value is equal to the maximum control value. S303: Set the instruction duration and generate instruction elements; the instruction elements include an identifier of the adjustable point, the optimal adjustment value, the current time, and the effective expiration time of the instruction; Based on the fitted second-order response surface model, predict the expected value of the controlled variable after the instruction is executed; after the instruction is executed, collect the actual value of the controlled variable and record the sign of the expected deviation.

7. The adaptive decision analysis method based on artificial intelligence according to claim 6, characterized in that: If the coefficient of determination is less than the threshold, expand the percentage range, re-filter the data, and fit a second-order response surface model; if the coefficient of determination is still less than the threshold after expanding the percentage range, then reselect the path with the second smallest cost until the coefficient of determination is less than the threshold.

8. The adaptive decision analysis method based on artificial intelligence according to claim 6, characterized in that: S4 includes the following: S401: Before each instruction is executed, collect the environmental variable vector within the time window; analyze the first-order moment mean vector and the second-order moment covariance matrix based on the number of sampling points within the time window, and concatenate the first-order moment mean vector and the second-order moment covariance matrix to form a scene fingerprint; S402: Analyze the Mahalanobis distance between the current scene fingerprint and the historical scene fingerprints in the historical scene database. If the minimum Mahalanobis distance is greater than the Mahalanobis distance threshold, it is determined that there are no similar scenes and no migration is performed. Otherwise, extract the causal strength coefficient and fitting coefficient of the historical scene; Analyze the standard deviation of the control variables and the controlled variables in the current and historical scenarios; Correct the fitting coefficients; The causal intensity coefficients for the current scene are directly adopted from the causal intensity coefficients for the historical scene corresponding to the minimum Mahalanobis distance; the corrected fitting coefficients are filled into the second-order response surface model of the current scene, replacing the original fitting coefficients.

9. An adaptive decision analysis system based on artificial intelligence, wherein the system is applied to the adaptive decision analysis method based on artificial intelligence as described in any one of claims 1-8, characterized in that, The system includes a data pairing and analysis module, a propagation and filtering module, a decision analysis module, and a scenario matching and correction module; The data pairing analysis module is used to collect monitoring variables, detect local extreme points of the monitoring variables, pair adjacent extreme points and perform piecewise linear fitting; analyze the features of extreme point segments to construct feature vectors; The propagation screening module is used to statistically analyze the time offset distribution between the extreme points of two variables, analyze the typical delay time and the causality strength coefficient, and then analyze the propagation cost based on the causality strength coefficient and the distance between variables, and output the candidate intervention path queue in ascending order of cost. The decision analysis module is used to collect corresponding data of historical control variables and monitoring variables along the first segment of the candidate intervention path queue, construct and fit a second-order response surface model, solve the surface stationary point analysis for the optimal adjustment value, and generate instruction elements with execution windows after boundary projection. The scene matching and correction module is used to analyze the environmental variables before each instruction execution to generate a scene fingerprint, analyze the similarity between the current scene and historical scenes, select the nearest neighbor and historical scenes, and correct the causal strength coefficient and the fitting coefficient of the second-order response surface model.

Images