An intelligent management system for landscaping engineering projects based on big data analysis

Abstract

The application discloses a kind of based on big data analysis's landscape engineering project intelligent management system, comprising: collection and object mapping module, collection engineering multi-source data and establish object mapping relationship generation uniform event flow dataset;Phase division and feature construction module, construction phase division is carried out to work surface and constructs stage joint feature matrix;Seedling state modeling module, generate continuous time state sequence;NOTEARS causal modeling module, improved NOTEARS model is used to construct stage engineering-physiological double-layer directed causal diagram;Intervention variable calculation module, calculate path total effect and generate intervention priority sequence;Measure generation module, determine intervention measure set;Execution backwrite update module, form difference sample and execute causal diagram incremental update.The application improves the decision accuracy and structural stability of landscape engineering management.

Description

Technical Field

[0001] This invention relates to the fields of machine learning and project management technology, and in particular to an intelligent management system for landscaping engineering projects based on big data analysis. Background Technology

[0002] With the continuous expansion of landscaping projects and the advancement of smart city construction, landscaping projects generate a large amount of construction records, seedling batch information, IoT sensor data, and image data during the stages of construction preparation, soil treatment, planting, maintenance, and acceptance. Existing landscaping project management systems typically focus on progress management or single data monitoring, mainly using tabular statistics, threshold alarms, or simple regression analysis to assess the project status. However, in complex, multi-stage, and multi-variable coupled scenarios, it is difficult to reveal the causal relationship between project variables and the physiological state of seedlings, and it is impossible to form an interpretable basis for intervention decisions. In practical applications, construction event data, environmental monitoring data, and seedling growth data are stored in different systems, lacking a unified object mapping mechanism, making it difficult to accurately establish spatial and temporal correlations between data, affecting the reliability of subsequent analysis results.

[0003] For evaluating the effectiveness of garden maintenance, existing technologies mostly rely on empirical rules or single-factor statistical analysis methods. They formulate adjustment strategies by calculating the correlation between variables such as irrigation volume and fertilization frequency and survival rate and growth indicators. However, correlation analysis cannot distinguish causal directions, struggles to avoid confounding effects between variables, and is prone to directional misjudgments in multivariate coupling scenarios. Furthermore, existing machine learning models often employ black-box prediction methods, lacking acyclic constraint structure learning and domain-specific embedding mechanisms. This fails to translate garden maintenance standards into structural constraints for model optimization, resulting in causal graphs with unstable directions or inconsistencies with engineering principles. In addition, after intervention, existing systems typically only record the results, lacking a mechanism to write execution records and retest data back to the model for incremental updates, hindering continuous correction and iterative optimization of the model structure.

[0004] Therefore, how to provide an intelligent management system for landscaping projects based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent management system for landscaping engineering projects based on big data analysis. The invention achieves spatiotemporal alignment and structured integration of multi-source data for landscaping projects through unified event flow modeling and phased feature construction; it introduces an improved NOTEARS model that integrates domain standards and a virtual perturbation direction freezing mechanism to enhance the directional stability and structural consistency of causal structure learning; and it generates intervention priorities based on path effect calculation and combines them with an execution write-back incremental update mechanism to form a closed-loop iterative causal-driven engineering management model.

[0006] According to an embodiment of the present invention, a smart management system for landscaping engineering projects based on big data analysis includes:

[0007] The data acquisition and object mapping module is used to acquire multi-source data for the project, establish object mapping relationships, and generate a unified event stream dataset.

[0008] The phase division and feature construction module is used to divide the work surface into construction phases based on the unified event flow dataset, and construct the work surface feature vector, seedling batch feature vector and sensor point feature vector in the corresponding phase to form a phased joint feature matrix.

[0009] The seedling state modeling module is used to construct a set of latent variables of seedling physiological state based on the staged joint feature matrix and generate the corresponding continuous time state sequence.

[0010] The NOTEARS causal modeling module is used to perform acyclic constraint optimization based on the continuous time state sequence using the improved NOTEARS model, and introduces a direction freezing mechanism based on virtual perturbation inversion to construct a staged engineering-physiology bilayer directed causal graph.

[0011] The intervention variable calculation module is used to calculate the total path effect from controllable engineering variables to target indicators in the staged engineering-physiology bilayer directed causal graph, and generate an intervention priority sequence.

[0012] The measure generation module is used to determine a set of intervention measures based on the intervention priority sequence and constraints.

[0013] The write-back update module is used to write the execution records and retest data corresponding to the set of intervention measures into the unified event stream dataset to form differential samples before and after execution, and to perform incremental updates on the staged engineering-physiology two-layer directed causal graph.

[0014] Optionally, the acquisition and object mapping module includes:

[0015] Receive multi-source engineering data, which includes construction event data, seedling procurement and quarantine data, planting and maintenance operation data, irrigation and environmental monitoring data, and image data output by UAVs or ground acquisition terminals, and encode them uniformly according to timestamps, spatial coordinates and project numbers;

[0016] Based on the spatial coordinates and the boundary information of the engineering drawings, the construction area is divided into polygonal grids. Each grid cell is defined as a work surface object, and a unique work surface identifier is assigned to each work surface object.

[0017] Based on the batch number, place of origin and specification parameters in the seedling delivery note, quarantine record and planting record, the seedling samples are aggregated into seedling batch objects, and a unique seedling batch identifier is generated for each seedling batch object.

[0018] Based on the installation coordinates of irrigation valves, soil sensors and meteorological monitoring stations, spatial clustering of monitoring equipment is performed, the area where the cluster center is located is determined as the sensing point object, and a unique sensing point identifier is generated for each sensing point object.

[0019] A spatial intersection operation is performed based on the polygonal range of the work surface object and the spatial coordinates of the sensor point object to establish a mapping table of the coverage relationship between the sensor point object and the work surface object;

[0020] Based on the overlap calculation of the planting record time of seedling batch objects and the construction time window execution time of work surface objects, a mapping table of the placement relationship between seedling batch objects and work surface objects is established.

[0021] The multi-source data of the project is written into the event record unit according to the corresponding work surface identifier, seedling batch identifier and sensor point identifier. Each event record unit includes the occurrence time, spatial location, object identifier, event type and attribute fields, thus forming a unified event flow dataset.

[0022] Optionally, the stage division and feature construction module includes:

[0023] Based on the event type field and timestamp field of each event record in the unified event flow dataset, the same work surface is sorted on a continuous time axis, and the construction stage interval is divided according to the preset event coding range of construction preparation events, soil treatment events, planting events, maintenance events and acceptance events.

[0024] When the proportion of planting event records to all event records in a continuous time interval exceeds a preset proportion threshold, and there is a seedling batch delivery relationship mapping record, the corresponding time interval will be marked as the planting stage.

[0025] When the number of maintenance operation event records exceeds the preset frequency threshold within a continuous time interval, and there are no new seedling batch release records, the corresponding time interval will be marked as the maintenance stage.

[0026] Based on the time intervals of each stage, the average, maximum, minimum and time change rate of the numerical data associated with the work surface object are calculated within each time interval; the frequency per unit area is calculated for the count data; and the cumulative duration is calculated for the duration data.

[0027] The specifications, sampling records and distribution quantities of seedling batches within the corresponding time intervals are statistically summarized to generate seedling batch feature vectors.

[0028] The sensor sequence data of the sensor point object within the corresponding time interval is segmented and aggregated according to a fixed time step, and the mean and standard deviation of each segment are calculated and then spliced ​​to generate the sensor point feature vector.

[0029] The feature vectors of the work surface, the seedling batch, and the sensor location are concatenated according to a unified field order to form a staged joint feature matrix with fixed dimensions.

[0030] Optionally, the seedling status modeling module includes:

[0031] Based on the staged joint feature matrix, numerical fields and operational fields related to seedling growth are selected. The numerical fields include soil moisture content, soil electrical conductivity, air temperature, relative humidity and rainfall. The operational fields include planting time, irrigation records and maintenance records.

[0032] The selected fields are aligned with the seedling batch identifier and the work area identifier in time, the data is mapped to a continuous time axis with a fixed time step, and missing numerical fields are filled with linear interpolation of adjacent time steps, and missing work fields are filled with the most recent valid record.

[0033] Define a set of latent variables for the physiological state of seedlings, including water stress index, salt stress index and growth recovery index, and establish calculation rules from fields to latent variables;

[0034] For each time step, the corresponding latent variable values ​​are generated according to the calculation rules, and the latent variable values ​​are truncated based on the historical sample statistical interval.

[0035] The hidden variable values ​​at each time step are arranged in chronological order to generate a continuous time state sequence corresponding to the work surface and the seedling batch.

[0036] Optionally, the NOTEARS causal modeling module includes:

[0037] The improved NOTEARS model introduces a domain knowledge constraint embedding mechanism and a direction freezing mechanism based on virtual perturbation inversion on the basis of the original acyclic constraint optimization structure;

[0038] The improved NOTEARS model concatenates the staged joint feature matrix with the continuous time state sequence of the corresponding time step according to the work surface identifier and the seedling batch identifier to form the intra-stage joint input matrix;

[0039] A weight matrix is ​​constructed based on the joint input matrix within the stage, and each element in the weight matrix represents the strength of the directed influence between the corresponding variables.

[0040] Construct a joint optimization objective function, which includes a residual squared loss term, an acyclic constraint term, and a regularization term;

[0041] The residual squared loss term is obtained by weighting and summing the values ​​of all variable nodes except the corresponding variable node itself for each variable node using the weight coefficients of the corresponding column in the weight matrix, and then calculating the squared difference between the predicted value and the actual value, and summing them over all samples.

[0042] The acyclic constraint term generates an acyclic detection matrix by performing matrix exponentiation on the weight matrix, and calculates the difference between the trace value of the acyclic detection matrix and the number of variables as the acyclic constraint function value.

[0043] The regularization term is obtained by summing the absolute values ​​of all elements in the weight matrix;

[0044] The joint optimization objective function is solved by an augmented Lagrange iteration method. In each iteration, the weight matrix parameters, Lagrange multipliers and penalty coefficients are updated until the value of the acyclic constraint function is less than the preset convergence threshold.

[0045] The domain knowledge constraint embedding mechanism maps the clauses of garden maintenance specifications into variable pairs, constructs a set of prohibited edges, a set of required edges, a set of symbolic constraints, and a set of order constraints, and adds the squared terms of the weights of prohibited edges, the squared terms of the differences when the weights of required edges are below the lower limit, the squared terms of the differences when the symbolic direction is violated, and the squared terms of the weights of the reverse edges of the order relationship to the joint optimization objective function.

[0046] The direction freezing mechanism based on virtual perturbation inversion applies a fixed-amplitude numerical perturbation to each variable in the joint input matrix during the iterative update of the weight matrix. While keeping the values ​​of other variables unchanged, perturbation samples are constructed. These perturbation samples are then substituted into the linear combination calculation process corresponding to the weight matrix. The response changes between variables are compared. When a perturbation of variable A causes a change in variable B, but a perturbation of variable B does not cause a change in variable A, the direction from A to B is marked as a definite direction. During the iteration process, the corresponding reverse weight is kept at zero, and only the weights of the definite direction are updated, generating a staged engineering-physiology two-layer directed causal graph.

[0047] Optionally, the intervention variable calculation module includes:

[0048] Extract the set of engineering variable nodes with control attribute identifiers from the phased engineering-physiology two-layer directed causal graph;

[0049] For each engineering variable node in the set of engineering variable nodes, based on the weight matrix of the phased engineering-physiology two-layer directed causal graph, all directed paths between the engineering variable node and the target index node are traversed and calculated. The path traversal is expanded layer by layer according to the connection relationship with non-zero edge weights until the target index node is reached.

[0050] For each directed path, the path effect value is calculated. The path effect value is obtained by multiplying the corresponding weights of each edge on the path. When multiple paths exist, the path effect values ​​are accumulated to obtain the total path effect value from the corresponding controllable engineering variable to the target indicator.

[0051] The absolute values ​​of the total path effect are calculated and sorted according to their numerical values ​​to generate an intervention priority sequence;

[0052] When the absolute value of the total path effect of a certain engineering variable node is less than the preset minimum effect threshold of 0.02, the corresponding engineering variable node is removed from the intervention priority sequence, forming a set of intervention variables and a corresponding priority sequence that meet the threshold condition.

[0053] Optionally, the measure generation module includes:

[0054] Based on the set of intervention variables and their corresponding priority sequences, the value range, historical statistical mean, and allowable adjustment range of each engineering variable at the corresponding time step in the current work surface are obtained.

[0055] For each engineering variable node in the priority sequence, the adjustment direction is determined according to the side weight sign of the corresponding target index in the weight matrix. When the side weight is positive, the variable value is adjusted in a decreasing direction, and when the side weight is negative, the variable value is adjusted in an increasing direction.

[0056] The variable values ​​are tiered according to a preset adjustment step size, which is 0.05 times the current value of the variable, and multiple candidate adjustment values ​​are generated within the allowable adjustment range.

[0057] Substitute each candidate adjustment value into the linear combination calculation process corresponding to the weight matrix to calculate the predicted change of the target indicator.

[0058] Under the premise of meeting the resource constraints and construction time constraints, the predicted changes corresponding to each candidate adjustment value are ranked, and the candidate adjustment value ranked first is selected as the intervention value of the corresponding engineering variable.

[0059] Repeat the above process for all engineering variables in the intervention priority sequence to form an intervention set, which includes variable name, adjustment direction, adjustment magnitude and implementation time interval.

[0060] Optionally, the write-back update module includes:

[0061] Obtain the on-site execution records corresponding to the set of intervention measures. The execution records include the actual adjusted values ​​of engineering variables, implementation timestamps, work area identifiers, and seedling batch identifiers.

[0062] Obtain retest data after the implementation of intervention measures. The retest data includes the observed values ​​of the target indicators and their corresponding timestamps after the implementation time interval.

[0063] The execution records and retest data are written into the unified event stream dataset in chronological order. The original event records are matched and verified based on the work area identifier, seedling batch identifier and timestamp. For records with conflicting matching results, duplicate records are removed and fields are overwritten.

[0064] The difference between the target indicator values ​​in the continuous time state sequence before the implementation of the intervention measures and the target indicator values ​​in the retest data is calculated to form a difference sample before and after the implementation. The difference sample includes the adjustment range of engineering variables and the change in target indicators.

[0065] The differential samples are appended to the time step positions corresponding to the staged joint feature matrix and the continuous time state sequence to form an expanded sample set;

[0066] While keeping the frozen directional relationships unchanged, incremental optimization and updates are performed on the weight matrix based on the expanded sample set. The update process uses the residual squared loss term, the acyclic constraint term, and the domain knowledge constraint term. The absolute value of the update increment of each weight element is limited to no more than 0.05, generating the updated staged engineering-physiology two-layer directed causal graph.

[0067] The beneficial effects of this invention are:

[0068] This invention achieves consistent alignment of construction event data, seedling batch data, environmental monitoring data, and image data across time and space dimensions by constructing a unified event flow dataset and object mapping mechanism. It combines stage division and feature construction processes to form a structured, staged joint feature matrix. Addressing the challenges of scattered multi-source data, ambiguous stage boundaries, and difficulties in unified modeling of variable relationships in landscape engineering, this invention proposes a seedling physiological latent variable modeling strategy based on continuous time state sequences. This strategy transforms water stress index, salt stress index, and growth recovery index into calculable dynamic state variables, significantly improving the ability to characterize the changing trends of seedling growth processes. In the causal modeling stage, an improved NOTEARS model is introduced. Through a continuously differentiable acyclic constraint optimization structure combined with a domain knowledge constraint embedding mechanism, landscape maintenance standards are transformed... Transforming the penalty term into a differentiable term for joint optimization effectively enhances the consistency between the causal graph structure and engineering laws. Simultaneously, a direction-freezing mechanism based on virtual perturbation inversion is designed. During iteration, the direction of variables is determined and locked through bidirectional perturbation response, reducing directional oscillations and the generation of structurally unstable edges, thus improving the convergence stability of causal structure learning. In the intervention decision-making stage, the total path effect of controllable engineering variables on the target index is quantified and ranked through path effect multiplication and accumulation calculation. Combined with constraints, hierarchical candidate measures are generated, achieving refined intervention decisions driven by the causal structure. Finally, by executing the write-back update module to construct differential samples and perform incremental optimization updates, the model continuously corrects edge weights and structural parameters in actual engineering feedback, realizing causal-driven, dynamic iteration, and interpretable intelligent management of landscaping projects. Attached Figure Description

[0069] 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:

[0070] Figure 1 This is a schematic diagram of the structure of an intelligent management system for landscaping projects based on big data analysis proposed in this invention.

[0071] Figure 2 This is a schematic diagram of the improved NOTEARS model structure in the intelligent management system for landscaping engineering projects based on big data analysis proposed in this invention.

[0072] Figure 3 This is a flowchart illustrating the virtual disturbance inversion direction freezing mechanism in an intelligent management system for landscaping projects based on big data analysis, as proposed in this invention. Detailed Implementation

[0073] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0074] refer to Figures 1-3 A smart management system for landscaping projects based on big data analytics, comprising:

[0075] The data acquisition and object mapping module is used to acquire multi-source data for the project, establish object mapping relationships, and generate a unified event stream dataset.

[0076] The phase division and feature construction module is used to divide the work surface into construction phases based on the unified event flow dataset, and construct the work surface feature vector, seedling batch feature vector and sensor point feature vector in the corresponding phase to form a phased joint feature matrix.

[0077] The seedling state modeling module is used to construct a set of latent variables of seedling physiological state based on a staged joint feature matrix and generate the corresponding continuous time state sequence.

[0078] The NOTEARS causal modeling module is used to perform acyclic constraint optimization based on continuous time state sequences using an improved NOTEARS model, and introduces a direction freezing mechanism based on virtual perturbation inversion to construct a phased engineering-physiology bilayer directed causal graph.

[0079] The intervention variable calculation module is used to calculate the total path effect from controllable engineered variables to target indicators in a staged engineering-physiology bilayer directed causal graph, and generate an intervention priority sequence.

[0080] The measure generation module is used to determine the set of intervention measures based on the intervention priority sequence and constraints.

[0081] The write-back update module is used to write the execution records and retest data corresponding to the set of intervention measures into a unified event stream dataset, forming differential samples before and after execution, and to perform incremental updates on the staged engineering-physiological two-layer directed causal graph.

[0082] In this embodiment, the acquisition and object mapping module includes:

[0083] The system receives multi-source engineering data, including construction event data, seedling procurement and quarantine data, planting and maintenance operation data, irrigation and environmental monitoring data, and image data output by drones or ground acquisition terminals. The data is then uniformly encoded according to timestamps, spatial coordinates, and project numbers.

[0084] Based on the spatial coordinates and the boundary information of the engineering drawings, the construction area is divided into polygonal grids. Each grid cell is defined as a work surface object, and a unique work surface identifier is assigned to each work surface object.

[0085] Based on the batch number, place of origin and specification parameters in the seedling delivery note, quarantine record and planting record, the seedling samples are aggregated into seedling batch objects, and a unique seedling batch identifier is generated for each seedling batch object.

[0086] Based on the installation coordinates of irrigation valves, soil sensors and meteorological monitoring stations, spatial clustering of monitoring equipment is performed, the area where the cluster center is located is determined as the sensing point object, and a unique sensing point identifier is generated for each sensing point object.

[0087] A spatial intersection operation is performed based on the polygonal range of the work surface object and the spatial coordinates of the sensor point object to establish a mapping table of the coverage relationship between the sensor point object and the work surface object;

[0088] Based on the overlap calculation of the planting record time of seedling batch objects and the construction time window execution time of work surface objects, a mapping table of the placement relationship between seedling batch objects and work surface objects is established.

[0089] The multi-source data of the project is written into the event record unit according to the corresponding work surface identifier, seedling batch identifier and sensor point identifier. Each event record unit includes the occurrence time, spatial location, object identifier, event type and attribute fields, thus forming a unified event flow dataset.

[0090] In this implementation, the polygon mesh subdivision is based on the engineering drawing coordinate system and then undergoes a unified coordinate transformation. The construction area is then divided according to the preset mesh side length, and overlaps and holes are eliminated through spatial topology verification. The spatial intersection operation adopts a combination of point-on-polygon determination and buffer expansion to project the sensor coordinates onto the working surface plane coordinate system and calculate the inclusion relationship. The temporal overlap calculation uses the planting record start and end time and the construction time window to determine the interval intersection. When the intersection duration is greater than the preset minimum duration, the placement relationship is established. The event recording unit uses sequential writing and unique primary key constraints to ensure the consistency and traceability of object mapping.

[0091] In this embodiment, the stage division and feature construction module includes:

[0092] Based on the event type field and timestamp field of each event record in the unified event flow dataset, the same work surface is sorted on a continuous time axis, and the construction stage interval is divided according to the preset event coding range of construction preparation events, soil treatment events, planting events, maintenance events and acceptance events.

[0093] When the proportion of planting event records to all event records in a continuous time interval exceeds a preset proportion threshold, and there is a seedling batch delivery relationship mapping record, the corresponding time interval will be marked as the planting stage.

[0094] When the number of maintenance operation event records exceeds the preset frequency threshold within a continuous time interval, and there are no new seedling batch release records, the corresponding time interval will be marked as the maintenance stage.

[0095] Based on the time intervals of each stage, the average, maximum, minimum and time change rate of the numerical data associated with the work surface object are calculated within each time interval; the frequency per unit area is calculated for the count data; and the cumulative duration is calculated for the duration data.

[0096] The specifications, sampling records and distribution quantities of seedling batches within the corresponding time intervals are statistically summarized to generate seedling batch feature vectors.

[0097] The sensor sequence data of the sensor point object within the corresponding time interval is segmented and aggregated according to a fixed time step, and the mean and standard deviation of each segment are calculated and then spliced ​​to generate the sensor point feature vector.

[0098] The feature vectors of the work surface, the seedling batch, and the sensor location are concatenated according to a unified field order to form a staged joint feature matrix with fixed dimensions.

[0099] In this implementation, the proportion threshold and frequency threshold are determined based on historical project sample statistics. First, the stage labels and event distribution of completed projects are statistically analyzed, and the proportion range and unit time frequency range of each type of event in the corresponding stage are calculated. The proportion threshold is taken as the median of the range of 0.6 to 0.8, which is 0.7, and the frequency threshold is taken as 0.7 of the average frequency per unit time. A calibration is performed based on the data of 14 consecutive days after the start of a new project. The construction preparation stage is determined by the continuous occurrence of preparation events and the absence of material delivery records. The soil treatment stage is determined by the frequency of soil improvement or land preparation events exceeding 0.7 times the average frequency per unit time. The planting stage is determined by the proportion of planting events being greater than 0.7 and the existence of batch delivery records. The maintenance stage is determined by the frequency of maintenance events exceeding 0.7 times the average frequency per unit time and the absence of new batch delivery records. The acceptance stage is determined by the continuous occurrence of acceptance events and the duration being greater than 3 days.

[0100] In this embodiment, the seedling status modeling module includes:

[0101] Based on the staged joint feature matrix, numerical fields and operational fields related to seedling growth are selected. The numerical fields include soil moisture content, soil electrical conductivity, air temperature, relative humidity and rainfall. The operational fields include planting time, irrigation records and maintenance records.

[0102] The selected fields are aligned with the seedling batch identifier and the work area identifier in time, the data is mapped to a continuous time axis with a fixed time step, and missing numerical fields are filled with linear interpolation of adjacent time steps, and missing work fields are filled with the most recent valid record.

[0103] Define a set of latent variables for the physiological state of seedlings, including water stress index, salt stress index and growth recovery index, and establish calculation rules from fields to latent variables;

[0104] For each time step, the corresponding latent variable values ​​are generated according to the calculation rules, and the latent variable values ​​are truncated based on the historical sample statistical interval.

[0105] The hidden variable values ​​at each time step are arranged in chronological order to generate a continuous time state sequence corresponding to the work surface and the seedling batch.

[0106] In this implementation, the calculation rules for fields to latent variables are implemented using a normalized weighted combination method. First, the soil moisture content, air temperature, soil electrical conductivity, cumulative planting time, and cumulative irrigation record values ​​are normalized to their minimum and maximum values, converting each field into a value between 0 and 1. The water stress index is calculated by multiplying the normalized soil moisture content by a weight of 0.6 and the air temperature by a weight of 0.4. The salt stress index is directly used as the index value based on the normalized soil electrical conductivity value. The growth recovery index is calculated by multiplying the normalized cumulative planting time by a weight of 0.5 and the cumulative irrigation record value by a weight of 0.5. The calculation results of each index are all limited to the range of 0 to 1.

[0107] In this embodiment, the NOTEARS causal modeling module includes:

[0108] The improved NOTEARS model introduces a domain knowledge constraint embedding mechanism and a direction freezing mechanism based on virtual perturbation inversion on the basis of the original acyclic constraint optimization structure;

[0109] The improved NOTEARS model concatenates the staged joint feature matrix with the continuous time state sequence of the corresponding time step according to the work surface identifier and the seedling batch identifier to form the intra-stage joint input matrix;

[0110] A weight matrix is ​​constructed based on the joint input matrix within the stage, and each element in the weight matrix represents the strength of the directed influence between the corresponding variables.

[0111] Construct a joint optimization objective function, which includes a residual squared loss term, an acyclic constraint term, and a regularization term;

[0112] The residual squared loss term is obtained by weighting and summing the values ​​of all variable nodes except the corresponding variable node itself for each variable node using the weight coefficients of the corresponding column in the weight matrix, and then calculating the squared difference between the predicted value and the actual value, and summing them over all samples.

[0113] The acyclic constraint term generates an acyclic detection matrix by performing matrix exponentiation on the weight matrix, and calculates the difference between the trace value of the acyclic detection matrix and the number of variables as the acyclic constraint function value.

[0114] The regularization term is obtained by summing the absolute values ​​of all elements in the weight matrix;

[0115] The joint optimization objective function is solved by an augmented Lagrange iteration method. In each iteration, the weight matrix parameters, Lagrange multipliers and penalty coefficients are updated until the value of the acyclic constraint function is less than the preset convergence threshold.

[0116] The domain knowledge constraint embedding mechanism maps the clauses of garden maintenance specifications into variable pairs, constructs a set of prohibited edges, a set of required edges, a set of symbolic constraints, and a set of order constraints, and adds the squared terms of the weights of prohibited edges, the squared terms of the differences when the weights of required edges are below the lower limit, the squared terms of the differences when the symbolic direction is violated, and the squared terms of the weights of the reverse edges of the order relationship to the joint optimization objective function.

[0117] The direction freezing mechanism based on virtual perturbation inversion applies a fixed-amplitude numerical perturbation to each variable in the joint input matrix during the iterative update of the weight matrix. While keeping the values ​​of other variables unchanged, perturbation samples are constructed. These perturbation samples are then substituted into the linear combination calculation process corresponding to the weight matrix. The response changes between variables are compared. When a perturbation of variable A causes a change in variable B, but a perturbation of variable B does not cause a change in variable A, the direction from A to B is marked as a definite direction. During the iteration process, the corresponding reverse weight is kept at zero, and only the weights of the definite direction are updated, generating a staged engineering-physiology two-layer directed causal graph.

[0118] In this implementation, the acyclic constraint of the joint optimization objective function is preset to a convergence threshold of 1x10. -5The regularization coefficient is set to 0.1 to ensure graph sparsity. In the augmented Lagrange iteration solution, the initial penalty coefficient is set to 1.0, the update growth factor is set to 10, and the maximum number of iterations is limited to 1000. For the direction freezing mechanism, the magnitude of the virtual perturbation is set to 0.1 times the standard deviation of the corresponding variable, and the threshold for judging response changes is set to 0.05. In the domain knowledge constraint embedding mechanism, the penalty weights for violations of prohibited edges and mandatory edges are both set to 100 to ensure that business logic takes precedence over data fitting.

[0119] The improved NOTEARS model maintains the same basic structure as the original NOTEARS model. Both models take the variable sample matrix as input, construct a weight matrix to represent the strength of the directed influence between variables, characterize the deviation between the predicted and the true values ​​through the residual squared loss term, constrain the weight matrix to satisfy the directed acyclic structure through a continuously differentiable acyclic constraint function, and solve the joint optimization objective function using an augmented Lagrangian iterative method to obtain the causal connection relationship between variables and the corresponding edge weights.

[0120] The improved NOTEARS model introduces a domain knowledge constraint embedding mechanism and a direction freezing mechanism based on virtual perturbation inversion on the basis of the above structure. The domain knowledge constraint embedding mechanism maps the clauses of the garden maintenance standard into variable pairs and transforms prohibited edges, mandatory edges, symbolic constraints, and order constraints into differentiable penalty terms added to the joint optimization objective function; the direction freezing mechanism applies a fixed amplitude perturbation to the variables during the iteration process, calculates the perturbation propagation response, and determines the one-way relationship based on the difference between the two-way responses, fixing the reverse weight to zero in subsequent optimizations;

[0121] Through the above improvements, the model maintains a continuously differentiable optimization framework while integrating landscape expertise with a data-driven learning process. It also progressively locks the causal direction during the training phase, reducing directional oscillations and the generation of unstable edges. This results in a staged engineering-physiological bilayer directed causal graph structure that better conforms to maintenance rules and has higher structural consistency and iterative convergence stability.

[0122] In this embodiment, the intervention variable calculation module includes:

[0123] Extract the set of engineering variable nodes with control attribute identifiers from the phased engineering-physiology two-layer directed causal graph;

[0124] For each engineering variable node in the set of engineering variable nodes, based on the weight matrix of the phased engineering-physiology two-layer directed causal graph, all directed paths between the engineering variable node and the target index node are traversed and calculated. The path traversal is expanded layer by layer according to the connection relationship with non-zero edge weights until the target index node is reached.

[0125] For each directed path, the path effect value is calculated. The path effect value is obtained by multiplying the corresponding weights of each edge on the path. When multiple paths exist, the path effect values ​​are accumulated to obtain the total path effect value from the corresponding controllable engineering variable to the target indicator.

[0126] The absolute values ​​of the total path effect are calculated and sorted according to their numerical values ​​to generate an intervention priority sequence;

[0127] When the absolute value of the total path effect of a certain engineering variable node is less than the preset minimum effect threshold of 0.02, the corresponding engineering variable node is removed from the intervention priority sequence, forming a set of intervention variables and a corresponding priority sequence that meet the threshold condition.

[0128] In this implementation, the control attribute identifiers of engineering variable nodes are pre-written into the data structure during the variable dictionary construction phase, and Boolean tags are used to distinguish between adjustable variables and observed variables. Path traversal is performed sequentially using a directed acyclic graph after topological sorting. First, the graph structure is topologically sorted to generate a node sequence, and then depth-first search is used to calculate all paths from engineering variable nodes to target indicator nodes according to the node sequence. To avoid the number of paths growing exponentially with the node size, the extension is terminated when the path length exceeds the preset maximum level of 5. Symbol information is retained during the path effect calculation process, and positive and negative effects are recorded separately before sorting to ensure the consistency of intervention direction determination.

[0129] In this embodiment, the measure generation module includes:

[0130] Based on the set of intervention variables and their corresponding priority sequences, the value range, historical statistical mean, and allowable adjustment range of each engineering variable at the corresponding time step in the current work surface are obtained.

[0131] For each engineering variable node in the priority sequence, the adjustment direction is determined according to the side weight sign of the corresponding target index in the weight matrix. When the side weight is positive, the variable value is adjusted in a decreasing direction, and when the side weight is negative, the variable value is adjusted in an increasing direction.

[0132] The variable values ​​are tiered according to a preset adjustment step size, which is 0.05 times the current value of the variable, and multiple candidate adjustment values ​​are generated within the allowable adjustment range.

[0133] Substitute each candidate adjustment value into the linear combination calculation process corresponding to the weight matrix to calculate the predicted change of the target indicator.

[0134] Under the premise of meeting the resource constraints and construction time constraints, the predicted changes corresponding to each candidate adjustment value are ranked, and the candidate adjustment value ranked first is selected as the intervention value of the corresponding engineering variable.

[0135] Repeat the above process for all engineering variables in the intervention priority sequence to form an intervention set, which includes variable name, adjustment direction, adjustment magnitude and implementation time interval.

[0136] In this implementation, the allowable adjustment range is determined based on the 5% to 95% percentile of the historical value distribution of engineering variables, and the upper limit is adjusted in conjunction with the rated capacity of on-site equipment; resource constraints are determined by cumulatively calculating the resource consumption corresponding to the adjusted variables and comparing it with the upper limit of the stage budget; construction time constraints are achieved by limiting the planned implementation time interval to the current construction stage time window; when multiple engineering variables are adjusted simultaneously, a sequential update method is adopted, after the intervention value of the previous variable is determined, the updated variable value is written into the joint input matrix, and then the predicted change is calculated for the next variable to ensure the consistency and executability of the combination of measures.

[0137] In this embodiment, the write-back update module includes:

[0138] Obtain the on-site execution records corresponding to the set of intervention measures. The execution records include the actual adjusted values ​​of engineering variables, implementation timestamps, work area identifiers, and seedling batch identifiers.

[0139] Obtain retest data after the implementation of intervention measures. The retest data includes the observed values ​​of the target indicators and their corresponding timestamps after the implementation time interval.

[0140] The execution records and retest data are written into the unified event stream dataset in chronological order. The original event records are matched and verified based on the work area identifier, seedling batch identifier and timestamp. For records with conflicting matching results, duplicate records are removed and fields are overwritten.

[0141] The difference between the target indicator values ​​in the continuous time state sequence before the implementation of the intervention measures and the target indicator values ​​in the retest data is calculated to form a difference sample before and after the implementation. The difference sample includes the adjustment range of engineering variables and the change in target indicators.

[0142] The differential samples are appended to the time step positions corresponding to the staged joint feature matrix and the continuous time state sequence to form an expanded sample set;

[0143] While keeping the frozen directional relationships unchanged, incremental optimization and updates are performed on the weight matrix based on the expanded sample set. The update process uses the residual squared loss term, the acyclic constraint term, and the domain knowledge constraint term. The absolute value of the update increment of each weight element is limited to no more than 0.05, generating the updated staged engineering-physiology two-layer directed causal graph.

[0144] In this implementation, the target index value before implementation is selected as the state value of the most recent complete time step before the implementation time stamp of the intervention measure. The retest data is the average value of the three consecutive time steps after the implementation time stamp. When the differential samples are written into the extended sample set, they are marked according to the construction stage and only participate in the weight update of the corresponding stage subgraph. The incremental optimization update adopts the local gradient update method under fixed direction constraints, and only the weight elements that have direct or indirect path connection with the intervened variable participate in the calculation, so as to control the update range and improve the iteration stability.

[0145] Example 1: To verify the feasibility of this invention in practice, it was applied to an ecological landscape improvement project in a new urban area. The project covers a total green area of ​​approximately 126,000 square meters, including tree planting areas, shrub color blocks, and lawn restoration areas, divided into 18 work areas, with a construction period of 6 months. During implementation, the project generated a large amount of construction event data, seedling batch information, irrigation and environmental monitoring data, and drone aerial image data. Traditional management methods mainly rely on manual inspections and experience-based judgment, with responses to abnormal seedling growth typically delayed by 5 to 10 days. Furthermore, it is difficult to determine whether the problem is caused by insufficient irrigation, increased soil salinity, or inadequate maintenance frequency, resulting in insufficient decision-making basis and unclear intervention directions.

[0146] In this embodiment, data is first integrated into the construction log system, seedling procurement system, IoT sensing platform, and UAV image acquisition terminal via the data acquisition and object mapping module. Construction event data, seedling batch data, soil moisture content, soil conductivity, temperature, humidity, and image features are uniformly encoded and mapped to 18 work surface objects according to timestamps and spatial coordinates. Subsequently, construction stages are divided based on event types such as construction preparation, planting, and maintenance. Within each stage, work surface feature vectors, seedling batch feature vectors, and sensor point feature vectors are constructed, forming a 128-dimensional staged joint feature matrix. Through the seedling state modeling module, variables such as soil moisture content, temperature, and conductivity are mapped to water stress index, salt stress index, and growth recovery index, with a time step of 1 day, forming a continuous time state sequence.

[0147] In the causal modeling stage, an improved NOTEARS model was used to perform structural learning on 42 variable nodes across 18 work surfaces. Constraints from garden maintenance standards, such as "irrigation affects soil moisture content" and "increased soil conductivity increases salt stress," were incorporated. A virtual perturbation inversion mechanism was used to lock the variable orientation. With training data consisting of the first 90 days of historical data, the model converged on the 120th iteration, with the acyclic constraint function value decreasing to 8.7 × 10⁻⁹, representing an approximately 23% improvement in structural stability compared to when the orientation freezing mechanism was not introduced. In the generated staged engineering-physiology bilayer directed causal graph, the edge weight of irrigation intensity on the water stress index was -0.42, the edge weight of soil conductivity on the salt stress index was 0.51, and the edge weight of maintenance frequency on the growth recovery index was 0.36.

[0148] During the calculation of intervention variables and the generation of measures, the system performed path effect calculations with "seedling survival rate" as the target indicator. The results showed that the total path effect value for irrigation intensity was -0.38, maintenance frequency was -0.21, and fertilization frequency was -0.08. Variables below the threshold of 0.02 were automatically removed. The system generated candidate adjustment schemes based on the weight symbols and allowable adjustment ranges. Under the premise that resource constraints did not exceed the stage budget of 120,000 yuan, it was determined that irrigation intensity would be increased by 5% for three work areas and maintenance frequency would be increased by 10% for two work areas. A retest was conducted within 7 days after the intervention. The average seedling survival rate increased from 91.2% to 95.6%, the water stress index decreased by an average of 0.14, and the salt stress index remained stable below 0.23.

[0149] To verify the effectiveness of the method of the present invention, the system of the present invention is compared with the traditional system, and the comparative experiment is shown in Table 1:

[0150] Table 1. Comparison of the Implementation Effects of Intelligent Management in Ecological Landscape Enhancement Projects

[0151]

[0152] As can be seen from the comparison results in Table 1 above, the system of this invention outperforms the traditional system in both engineering management effectiveness and model performance. The average survival rate of seedlings increased from 91.2% to 95.6%, an increase of 4.4 percentage points, indicating that the intervention measures were more precise and effective; the water stress index decreased from 0.37 to 0.23, and the salt stress index decreased from 0.29 to 0.23, indicating a significant improvement in the physiological state of the seedlings. The predicted RMSE of the target indicators decreased from 0.082 to 0.057, with an error reduction of approximately 30%, reflecting a significant improvement in the accuracy of causal modeling; the structural stability score of the causal graph increased from 0.71 to 0.87, indicating enhanced structural consistency. The average duration of anomaly response was shortened from 6.8 days to 1.9 days, and the budget overrun rate decreased from 8.5% to 2.1%, demonstrating a simultaneous improvement in decision-making efficiency and resource control capabilities.

[0153] As can be seen from the above embodiments, the present invention can achieve structured integration of multi-source data and causal relationship modeling in real landscape engineering scenarios, significantly improving the accuracy and response efficiency of intervention decisions. Furthermore, the invention achieves continuous optimization of the model through the execution of a write-back incremental update mechanism, verifying the feasibility and practical application value of the present invention in landscape greening project management.

[0154] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A large data analysis-based intelligent management system for landscaping engineering projects, characterized in that, include: The data acquisition and object mapping module is used to acquire multi-source data for the project, establish object mapping relationships, and generate a unified event stream dataset. The phase division and feature construction module is used to divide the work surface into construction phases based on the unified event flow dataset, and construct the work surface feature vector, seedling batch feature vector and sensor point feature vector in the corresponding phase to form a phased joint feature matrix. The seedling state modeling module is used to construct a set of latent variables of seedling physiological state based on the staged joint feature matrix and generate the corresponding continuous time state sequence. The NOTEARS causal modeling module is used to perform acyclic constraint optimization based on the continuous time state sequence using the improved NOTEARS model, and introduces a direction freezing mechanism based on virtual perturbation inversion to construct a staged engineering-physiology bilayer directed causal graph. The intervention variable calculation module is used to calculate the total path effect from controllable engineering variables to target indicators in the staged engineering-physiology bilayer directed causal graph, and generate an intervention priority sequence. The measure generation module is used to determine a set of intervention measures based on the intervention priority sequence and constraints. The write-back update module is used to write the execution records and retest data corresponding to the set of intervention measures into the unified event stream dataset to form differential samples before and after execution, and to perform incremental updates on the staged engineering-physiology two-layer directed causal graph. The NOTEARS causal modeling module includes: The improved NOTEARS model introduces a domain knowledge constraint embedding mechanism and a direction freezing mechanism based on virtual perturbation inversion on the basis of the original acyclic constraint optimization structure; The improved NOTEARS model concatenates the staged joint feature matrix with the continuous time state sequence of the corresponding time step according to the work surface identifier and the seedling batch identifier to form the intra-stage joint input matrix; A weight matrix is ​​constructed based on the joint input matrix within the stage, and each element in the weight matrix represents the strength of the directed influence between the corresponding variables. Construct a joint optimization objective function, which includes a residual squared loss term, an acyclic constraint term, and a regularization term; The residual squared loss term is obtained by weighting and summing the values ​​of all variable nodes except the corresponding variable node itself for each variable node using the weight coefficients of the corresponding column in the weight matrix, and then calculating the squared difference between the predicted value and the actual value, and summing them over all samples. The acyclic constraint term generates an acyclic detection matrix by performing matrix exponentiation on the weight matrix, and calculates the difference between the trace value of the acyclic detection matrix and the number of variables as the acyclic constraint function value. The regularization term is obtained by summing the absolute values ​​of all elements in the weight matrix; The joint optimization objective function is solved by an augmented Lagrange iteration method. In each iteration, the weight matrix parameters, Lagrange multipliers and penalty coefficients are updated until the value of the acyclic constraint function is less than the preset convergence threshold. The domain knowledge constraint embedding mechanism maps the clauses of garden maintenance specifications into variable pairs, constructs a set of prohibited edges, a set of required edges, a set of symbolic constraints, and a set of order constraints, and adds the squared terms of the weights of prohibited edges, the squared terms of the differences when the weights of required edges are below the lower limit, the squared terms of the differences when the symbolic direction is violated, and the squared terms of the weights of the reverse edges of the order relationship to the joint optimization objective function. The direction freezing mechanism based on virtual perturbation inversion applies a fixed-amplitude numerical perturbation to each variable in the joint input matrix during the iterative update of the weight matrix. While keeping the values ​​of other variables unchanged, perturbation samples are constructed. These perturbation samples are then substituted into the linear combination calculation process corresponding to the weight matrix. The response changes between variables are compared. When a perturbation of variable A causes a change in variable B, but a perturbation of variable B does not cause a change in variable A, the direction from A to B is marked as a definite direction. During the iteration process, the corresponding reverse weight is kept at zero, and only the weights of the definite direction are updated, generating a staged engineering-physiology two-layer directed causal graph.

2. The intelligent management system for landscaping projects based on big data analysis according to claim 1, characterized in that, The acquisition and object mapping module includes: Receive multi-source engineering data, which includes construction event data, seedling procurement and quarantine data, planting and maintenance operation data, irrigation and environmental monitoring data, and image data output by UAVs or ground acquisition terminals, and encode them uniformly according to timestamps, spatial coordinates and project numbers; Based on the spatial coordinates and the boundary information of the engineering drawings, the construction area is divided into polygonal grids. Each grid cell is defined as a work surface object, and a unique work surface identifier is assigned to each work surface object. Based on the batch number, place of origin and specification parameters in the seedling delivery note, quarantine record and planting record, the seedling samples are aggregated into seedling batch objects, and a unique seedling batch identifier is generated for each seedling batch object. Based on the installation coordinates of irrigation valves, soil sensors and meteorological monitoring stations, spatial clustering of monitoring equipment is performed, the area where the cluster center is located is determined as the sensing point object, and a unique sensing point identifier is generated for each sensing point object. A spatial intersection operation is performed based on the polygonal range of the work surface object and the spatial coordinates of the sensor point object to establish a mapping table of the coverage relationship between the sensor point object and the work surface object; Based on the overlap calculation of the planting record time of seedling batch objects and the construction time window execution time of work surface objects, a mapping table of the placement relationship between seedling batch objects and work surface objects is established. The multi-source data of the project is written into the event record unit according to the corresponding work surface identifier, seedling batch identifier and sensor point identifier. Each event record unit includes the occurrence time, spatial location, object identifier, event type and attribute fields, thus forming a unified event flow dataset.

3. The intelligent management system for landscaping projects based on big data analysis according to claim 1, characterized in that, The stage division and feature construction module includes: Based on the event type field and timestamp field of each event record in the unified event flow dataset, the same work surface is sorted on a continuous time axis, and the construction stage interval is divided according to the preset event coding range of construction preparation events, soil treatment events, planting events, maintenance events and acceptance events. When the proportion of planting event records to all event records in a continuous time interval exceeds a preset proportion threshold, and there is a seedling batch delivery relationship mapping record, the corresponding time interval will be marked as the planting stage. When the number of maintenance operation event records exceeds the preset frequency threshold within a continuous time interval, and there are no new seedling batch delivery records, the corresponding time interval will be marked as the maintenance stage. Based on the time intervals of each stage, the average, maximum, minimum and time change rate of the numerical data associated with the work surface object are calculated within each time interval; the frequency per unit area is calculated for the count data; and the cumulative duration is calculated for the duration data. The specifications, sampling records and distribution quantities of seedling batches within the corresponding time intervals are statistically summarized to generate seedling batch feature vectors. The sensor sequence data of the sensor point object within the corresponding time interval is segmented and aggregated according to a fixed time step, and the mean and standard deviation of each segment are calculated and then spliced ​​to generate the sensor point feature vector. The feature vectors of the work surface, the seedling batch, and the sensor location are concatenated according to a unified field order to form a staged joint feature matrix with fixed dimensions.

4. The intelligent management system for landscaping projects based on big data analysis according to claim 1, characterized in that, The seedling status modeling module includes: Based on the staged joint feature matrix, numerical fields and operational fields related to seedling growth are selected. The numerical fields include soil moisture content, soil electrical conductivity, air temperature, relative humidity and rainfall. The operational fields include planting time, irrigation records and maintenance records. The selected fields are aligned with the seedling batch identifier and the work area identifier in time, the data is mapped to a continuous time axis with a fixed time step, and missing numerical fields are filled with linear interpolation of adjacent time steps, and missing work fields are filled with the most recent valid record. Define a set of latent variables for the physiological state of seedlings, including water stress index, salt stress index and growth recovery index, and establish calculation rules from fields to latent variables; For each time step, the corresponding latent variable values ​​are generated according to the calculation rules, and the latent variable values ​​are truncated based on the historical sample statistical interval. The hidden variable values ​​at each time step are arranged in chronological order to generate a continuous time state sequence corresponding to the work surface and the seedling batch.

5. The intelligent management system for landscaping projects based on big data analysis according to claim 1, characterized in that, The intervention variable calculation module includes: Extract the set of engineering variable nodes with control attribute identifiers from the phased engineering-physiology two-layer directed causal graph; For each engineering variable node in the set of engineering variable nodes, based on the weight matrix of the phased engineering-physiology two-layer directed causal graph, all directed paths between the corresponding engineering variable node and the target index node are traversed and calculated. The path traversal is expanded layer by layer according to the connection relationship with non-zero edge weights until the target index node is reached. For each directed path, the path effect value is calculated. The path effect value is obtained by multiplying the corresponding weights of each edge on the path. When multiple paths exist, the path effect values ​​are accumulated to obtain the total path effect value from the corresponding controllable engineering variable to the target indicator. The absolute values ​​of the total path effect are calculated and sorted according to their numerical values ​​to generate an intervention priority sequence; When the absolute value of the total path effect of a certain engineering variable node is less than the preset minimum effect threshold of 0.02, the corresponding engineering variable node is removed from the intervention priority sequence, forming a set of intervention variables and a corresponding priority sequence that meet the threshold condition.

6. The intelligent management system for landscaping projects based on big data analysis according to claim 1, characterized in that, The measure generation module includes: Based on the set of intervention variables and their corresponding priority sequences, the value range, historical statistical mean, and allowable adjustment range of each engineering variable at the corresponding time step in the current work surface are obtained. For each engineering variable node in the priority sequence, the adjustment direction is determined according to the side weight sign of the corresponding target index in the weight matrix. When the side weight is positive, the variable value is adjusted in a decreasing direction, and when the side weight is negative, the variable value is adjusted in an increasing direction. The variable values ​​are tiered according to a preset adjustment step size, which is 0.05 times the current value of the variable, and multiple candidate adjustment values ​​are generated within the allowable adjustment range. Substitute each candidate adjustment value into the linear combination calculation process corresponding to the weight matrix to calculate the predicted change of the target indicator. Under the premise of meeting the resource constraints and construction time constraints, the predicted changes corresponding to each candidate adjustment value are ranked, and the candidate adjustment value ranked first is selected as the intervention value of the corresponding engineering variable. Repeat the above process for all engineering variables in the intervention priority sequence to form an intervention set, which includes variable name, adjustment direction, adjustment magnitude and implementation time interval.

7. The intelligent management system for landscaping projects based on big data analysis according to claim 1, characterized in that, The write-back update module includes: Obtain the on-site execution records corresponding to the set of intervention measures. The execution records include the actual adjusted values ​​of engineering variables, implementation timestamps, work area identifiers, and seedling batch identifiers. Obtain retest data after the implementation of intervention measures. The retest data includes the observed values ​​of the target indicators and their corresponding timestamps after the implementation time interval. The execution records and retest data are written into the unified event stream dataset in chronological order. The original event records are matched and verified based on the work area identifier, seedling batch identifier and timestamp. For records with conflicting matching results, duplicate records are removed and fields are overwritten. The difference between the target indicator values ​​in the continuous time state sequence before the implementation of the intervention measures and the target indicator values ​​in the retest data is calculated to form a difference sample before and after the implementation. The difference sample includes the adjustment range of engineering variables and the change in target indicators. The differential samples are appended to the time step positions corresponding to the staged joint feature matrix and the continuous time state sequence to form an expanded sample set; While keeping the frozen directional relationships unchanged, incremental optimization and updates are performed on the weight matrix based on the expanded sample set. The update process uses the residual squared loss term, the acyclic constraint term, and the domain knowledge constraint term. The absolute value of the update increment of each weight element is limited to no more than 0.05, generating the updated staged engineering-physiology two-layer directed causal graph.

Images