Landscape engineering maintenance management system and method based on deep learning

Abstract

The invention relates to the technical field of landscape engineering maintenance management, in particular to a landscape engineering maintenance management system and method based on deep learning. According to the invention, deep learning and landscape engineering maintenance management are fused, a multi-dimensional heterogeneous sensing network sensing layer is constructed, environmental data are efficiently collected and preprocessed, and a solid foundation is provided for analysis; the data layer module is in wireless connection and performs feature extraction; the model layer realizes five-level classification prediction of plant diseases based on a deep learning algorithm; the decision-making layer generates a maintenance scheme score sequence, and the feedback layer collects an implementation result optimization model to form closed-loop management; the system realizes conversion from passive response to active prediction, significantly improves maintenance accuracy and perspectiveness, and brings revolutionary progress for landscape engineering maintenance management.

Description

Technical Field

[0001] The present invention relates to the technical field of landscape engineering maintenance management, and particularly to a landscape engineering maintenance management system and method based on deep learning. Background Art

[0002] With the acceleration of the urbanization process, the maintenance management of urban landscape engineering faces many challenges. Traditional landscape engineering maintenance management mainly relies on manual experience judgment. This method is not only inefficient but also greatly affected by subjective factors, making it difficult to cope with complex and changeable maintenance environments. Especially in the prevention of plant diseases, a passive response strategy is often adopted, and treatment is carried out after the diseases occur, resulting in waste of maintenance resources and decline in landscape quality.

[0003] In the prior art, there are already systems that use Internet of Things technology for landscape monitoring. However, these systems generally stay at the level of data collection and simple monitoring, lacking in-depth intelligent analysis and prediction functions. At the same time, although some systems have tried to introduce machine learning technology for disease identification, it is usually for the identification of diseases that have already occurred, rather than predictive prevention, and lacks an effective maintenance decision support mechanism.

[0004] Therefore, there is an urgent need to develop a system that can realize intelligent prediction and decision-making for landscape engineering maintenance based on deep learning technology, achieve the transformation from passive response to active prevention, and improve the efficiency and quality of landscape maintenance. Summary of the Invention

[0005] The present invention provides a landscape engineering maintenance management system and method based on deep learning, aiming to solve the problems of relying on manual experience, lacking predictability and precise decision support in traditional landscape engineering maintenance management.

[0006] The present invention proposes a landscape engineering maintenance management system based on deep learning, including:

[0007] A perception layer module for collecting monitoring data of the landscape environment through a multi-dimensional heterogeneous sensing network and sending it to a cloud server; a data layer module, wirelessly connected to the perception layer module, for preprocessing and feature extraction of the monitoring data; a model layer module, connected to the data layer module, for constructing a plant disease prediction model based on a deep learning algorithm and generating a prediction result of maintenance requirements; a decision layer module, connected to the model layer module, for generating a scoring and ranking of maintenance plans based on the prediction result of maintenance requirements; a feedback layer module, connected to the decision layer module, for collecting the results of maintenance implementation and optimizing the plant disease prediction model.

[0008] Preferably, the perception layer module includes: an intelligent environment sensor unit for collecting air humidity, light intensity, soil moisture, and temperature data; an edge computing unit connected to the intelligent environment sensor unit for preliminarily filtering and detecting anomalies in the collected environmental data; and a communication unit connected to the edge computing unit for transmitting the filtered environmental data to a cloud server through the Internet of Things.

[0009] Preferably, the data layer module includes: a data cleaning unit for processing outliers, missing values, and duplicate values in the monitoring data; a multi-source data fusion unit connected to the data cleaning unit for integrating environmental data, historical maintenance records, and plant growth status data; and a feature engineering unit connected to the multi-source data fusion unit for generating feature vectors through time-series feature extraction and spatial feature correlation mining.

[0010] Preferably, the model layer module includes: a training unit for constructing and training a deep learning model based on the feature vectors; a prediction unit connected to the training unit for inputting plant environmental data into the trained deep learning model to generate plant disease prediction results; and an evaluation unit connected to the prediction unit for evaluating the model performance through a confusion matrix and F1 score.

[0011] Preferably, the prediction unit classifies the occurrence of plant diseases into 5 categories: slight, slight to moderate, moderate to severe, severe, and very severe, and outputs the corresponding disease occurrence probabilities.

[0012] Preferably, the decision layer module includes: a solution generation unit for creating multiple optional maintenance solutions based on the maintenance requirement prediction results; a scoring unit connected to the solution generation unit for scoring the optional maintenance solutions through the linear weighting method, and the scoring formula is: maintenance solution score = historical usage scoring weight * historical usage score * (1 - disease recurrence rate 1) * (1 - disease recurrence rate 2) + rationality scoring weight * rationality score + cost weight * cost score; and a sorting unit connected to the scoring unit for sorting the maintenance solutions according to the maintenance solution scores.

[0013] Preferably, the feedback layer module includes: an effect evaluation unit for quantitatively evaluating the actual effects after the implementation of maintenance measures; a knowledge precipitation unit connected to the effect evaluation unit for constructing the maintenance experience and data analysis results into a structured knowledge graph; and a model optimization unit connected to the knowledge precipitation unit for continuously optimizing the parameters of the deep learning model based on the maintenance implementation effects.

[0014] Preferably, it further includes a visualization module connected to the decision-making layer module for displaying the system status and maintenance suggestions through a graphical interface. Among them, the visualization module includes a front-end user interface and a back-end Web server.

[0015] Preferably, the deep learning algorithm includes a convolutional neural network, a recurrent neural network, a temporal-spatial two-stream neural network, and a hybrid attention mechanism network. Among them, the hybrid attention mechanism network combines spatial attention and channel attention to enhance the sensitivity of the model to changes in key environmental factors.

[0016] A landscape engineering maintenance management method based on deep learning includes the following steps: collecting monitoring data of the landscape environment through a multi-dimensional heterogeneous sensor network and sending it to a cloud server; preprocessing and feature extraction of the monitoring data to generate feature vectors; constructing and training a deep learning model based on the feature vectors to form a plant disease prediction model; using the plant disease prediction model to analyze the real-time collected environmental data to generate a maintenance requirement prediction result; creating multiple optional maintenance plans based on the maintenance requirement prediction result; scoring and sorting the optional maintenance plans by the linear weighted method to generate an optimal maintenance decision; executing the optimal maintenance decision and collecting the maintenance implementation result; optimizing the plant disease prediction model based on the maintenance implementation result to achieve continuous learning and evolution of the system.

[0017] The beneficial effects of the present invention include:

[0018] 1) By combining deep learning technology with landscape engineering maintenance management, the present invention realizes the transformation from passive response to active prediction, significantly improving the accuracy and forward-looking of maintenance management.

[0019] 2) The perception layer design of the present invention combining a multi-dimensional heterogeneous sensor network and an edge computing unit realizes high-quality collection and preprocessing of environmental data, providing a reliable data basis for subsequent analysis.

[0020] 3) The deep learning model of the present invention conducts five-level classification prediction for plant diseases, with the prediction accuracy increased by more than 30% compared with traditional methods, and potential diseases can be predicted 7-14 days in advance.

[0021] 4) The present invention innovatively introduces a maintenance plan scoring mechanism based on three dimensions of historical effect, rationality, and cost, making the decision-making process more scientific and quantitative.

[0022] 5) The closed-loop feedback mechanism of the present invention realizes continuous improvement of the system's intelligence level through continuous learning and knowledge precipitation, with strong adaptability.

[0023] 6) The application of the present invention can reduce water resource consumption by 30 - 40% and fertilizer usage by 25%, while improving the plant health condition by 15%, showing significant ecological and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is the overall system architecture of the present invention;

[0025] Figure 2 It is the structural diagram of the perception layer module of the present invention;

[0026] Figure 3 It is the structural diagram of the data layer module of the present invention;

[0027] Figure 4 It is the structure of the model layer module of the present invention;

[0028] Figure 5 It is the structural diagram of the decision layer module of the present invention;

[0029] Figure 6 It is the structural diagram of the feedback layer module of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0030] Please refer to the attached Figure 1-6 , The system of the present invention adopts a hierarchical architecture design, including five functional layers: the perception layer, the data layer, the model layer, the decision layer, and the feedback layer, to realize the full-process intelligent management from environmental data collection to the generation of maintenance decisions. The system analyzes and processes the collected multi-source heterogeneous data through deep learning technology, constructs a plant disease prediction model, generates accurate maintenance decisions based on the prediction results, and continuously optimizes the prediction model through a closed-loop feedback mechanism to improve the intelligent level of the system.

[0031] The present invention provides a landscape engineering maintenance management system based on deep learning, including: a perception layer module 1, used for collecting monitoring data of the landscape environment through a multi-dimensional heterogeneous sensing network and sending it to the cloud server; a data layer module 2, wirelessly connected to the perception layer module 1, used for preprocessing and feature extraction of the monitoring data; a model layer module 3, connected to the data layer module 2, used for constructing a plant disease prediction model based on deep learning algorithms and generating a prediction result of maintenance requirements; a decision layer module 4, connected to the model layer module 3, used for generating a scoring and ranking of maintenance plans based on the prediction result of maintenance requirements; a feedback layer module 5, connected to the decision layer module 4, used for collecting the results of maintenance implementation and optimizing the plant disease prediction model.

[0032] Preferably, the perception layer module 1 includes: an intelligent environment sensor unit 11 for collecting air humidity, light intensity, soil moisture, and temperature data; an edge computing unit 12 connected to the intelligent environment sensor unit 11 for preliminarily filtering and anomaly detecting the collected environmental data; and a communication unit 13 connected to the edge computing unit 12 for transmitting the filtered environmental data to the cloud server through the Internet of Things.

[0033] Preferably, the data layer module 2 includes: a data cleaning unit 21 for processing outliers, missing values, and duplicate values in the monitoring data; a multi-source data fusion unit 22 connected to the data cleaning unit 21 for integrating environmental data, historical maintenance records, and plant growth status data; and a feature engineering unit 23 connected to the multi-source data fusion unit 22 for generating feature vectors by extracting time-series features and associating and mining spatial features.

[0034] Preferably, the model layer module 3 includes: a training unit 31 for constructing and training a deep learning model based on the feature vectors; a prediction unit 32 connected to the training unit 31 for inputting plant environmental data into the trained deep learning model to generate plant disease prediction results; and an evaluation unit 33 connected to the prediction unit 32 for evaluating the model performance through a confusion matrix and an F1 score.

[0035] In a preferred embodiment of the present invention, the prediction unit 32 classifies the occurrence of plant diseases into 5 categories: slight, slight to moderate, moderate to severe, severe, and critical, and outputs the corresponding disease occurrence probabilities.

[0036] Preferably, the decision-making layer module 4 includes: a solution generation unit 41 for creating multiple optional maintenance solutions based on the maintenance requirement prediction results; and a scoring unit 42 connected to the solution generation unit 41 for scoring the optional maintenance solutions by the linear weighting method. The scoring formula is:

[0037] Maintenance solution score = (historical usage scoring weight × historical usage score × (1 - disease recurrence rate 1) × (1 - disease recurrence rate 2)) + (rationality scoring weight × rationality score) + (cost weight × cost score),

[0038] Among them, the historical usage scoring weight represents the weight coefficient of the historical maintenance effect in the scoring, and its general value range is from 0.4 to 0.6; the historical usage score represents the evaluation of the historical usage effect of the plan, and its value range is from 0 to 10; the disease recurrence rate 1 and the disease recurrence rate 2 respectively represent the short-term and long-term disease recurrence probabilities, and their value ranges are from 0 to 1; the rationality scoring weight represents the weight coefficient of the plan rationality in the scoring, and its general value range is from 0.2 to 0.4; the rationality score represents the scientific rationality evaluation of the plan design, and its value range is from 0 to 10; the cost weight represents the weight coefficient of the cost factor in the scoring, and its general value range is from 0.1 to 0.3; the cost score represents the cost economy evaluation of the plan, and its value range is from 0 to 10. Preferably, the present invention sets the historical usage scoring weight to 0.5, the rationality scoring weight to 0.3, and the cost weight to 0.2. This weight configuration can balance the three factors of historical experience, plan rationality, and economic cost.

[0039] In addition, the decision-making layer module 4 further includes a sorting unit 43, connected to the scoring unit 42, for sorting the maintenance plans according to the scores of the maintenance plans.

[0040] Preferably, the feedback layer module 5 includes: an effect evaluation unit 51, for quantitatively evaluating the actual effect after the implementation of the maintenance measures; a knowledge precipitation unit 52, connected to the effect evaluation unit 51, for constructing the maintenance experience and data analysis results into a structured knowledge graph; a model optimization unit 53, connected to the knowledge precipitation unit 52, for continuously optimizing the deep learning model parameters based on the maintenance implementation effect.

[0041] In another preferred embodiment of the present invention, the system further includes: a visualization module 6, connected to the decision-making layer module 4, for displaying the system status and maintenance suggestions through a graphical interface; wherein, the visualization module 6 includes a front-end user interface 61 and a back-end Web server 62.

[0042] Preferably, the deep learning algorithm includes: a convolutional neural network, a recurrent neural network, a temporal-spatial two-stream neural network, and a hybrid attention mechanism network; wherein, the hybrid attention mechanism network combines spatial attention and channel attention to enhance the sensitivity of the model to changes in key environmental factors.

[0043] The present invention also provides a landscape engineering maintenance management method based on deep learning, comprising the following steps: collecting monitoring data of the landscape environment through a multi-dimensional heterogeneous sensing network and sending it to a cloud server; preprocessing and feature extraction of the monitoring data to generate feature vectors; constructing and training a deep learning model based on the feature vectors to form a plant disease prediction model; using the plant disease prediction model to analyze the real-time collected environmental data to generate a maintenance requirement prediction result; creating multiple optional maintenance plans based on the maintenance requirement prediction result; scoring and ranking the optional maintenance plans by the linear weighted method to generate an optimal maintenance decision; executing the optimal maintenance decision and collecting the maintenance implementation result; optimizing the plant disease prediction model based on the maintenance implementation result to achieve continuous learning and evolution of the system.

[0044] The present invention will be further described in detail below with reference to the drawings and embodiments.

[0045] Embodiment 1: System overall architecture

[0046] As Figure 1 shown, the landscape engineering maintenance management system based on deep learning provided by the present invention adopts a hierarchical architecture design, including five core functional modules: a perception layer module 1, a data layer module 2, a model layer module 3, a decision layer module 4, and a feedback layer module 5, as well as a visualization module 6. A closed loop of data flow is formed between the modules to realize the full-process intelligent management from environmental perception to maintenance decision-making.

[0047] The perception layer module 1 is the basic data collection unit of the system, mainly responsible for collecting monitoring data of the landscape environment through a multi-dimensional heterogeneous sensing network. As Figure 2 shown, the perception layer module 1 includes:

[0048] An intelligent environment sensor unit 11, used to collect environmental parameters such as air humidity, light intensity, soil moisture, and temperature. In the preferred embodiment of the present invention, the sensors are not simply evenly distributed, but are optimized according to the characteristics of the landscape area, the density of plant species, and the microclimate characteristics to form a "dense-sparse-dense" three-dimensional monitoring grid. For example, in the disease-prone area or the area where rare plants are concentrated, the sensor density can reach 1 per 10 square meters, while in the general area, it can be reduced to 1 per 50 square meters. This differential layout strategy can optimize the system cost while ensuring the monitoring effect.

[0049] The edge computing unit 12, connected to the intelligent environment sensor unit 11, is used to perform preliminary filtering and anomaly detection on the collected environmental data. The edge computing unit 12 realizes the front-end preprocessing of data and can screen out obviously abnormal data points, such as obviously unreasonable values like the temperature sensor reading suddenly jumping from 25°C to 85°C. The edge computing unit 12 adopts lightweight anomaly detection algorithms, such as the 3σ rule based on statistics, and marks the values outside the range of the mean ± 3 times the standard deviation as anomalies. In practical applications, the system will set different anomaly judgment thresholds for different environmental parameters. For example, the effective change range of temperature is -10°C to 45°C, the effective range of humidity is 0% to 100%, and the effective range of light intensity is 0 to 120,000 lux, etc.

[0050] The communication unit 13, connected to the edge computing unit 12, is used to transmit the filtered environmental data to the cloud server through Internet of Things technology. The communication unit 13 supports a variety of communication protocols, including but not limited to low-power wide-area network technologies such as LoRa, NB-IoT, ZigBee, etc., and flexibly selects according to the communication requirements of different landscape environments. In actual deployment, the system will select a suitable combination of communication technologies according to the number and distribution density of sensor nodes. For example, NB-IoT technology is preferred in open areas, and ZigBee networking technology is preferred in densely wooded areas to ensure the reliability and real-time nature of data transmission.

[0051] The data layer module 2 is wirelessly connected to the perception layer module 1 and is responsible for preprocessing and feature extraction of the monitoring data. As Figure 3 shown, the data layer module 2 includes:

[0052] The data cleaning unit 21 is used to process outliers, missing values, and duplicate values in the monitoring data. For outliers, the present invention adopts an adaptive data cleaning algorithm and uses a differentiated cleaning strategy according to the characteristics of different types of sensor data. For example, for temperature and humidity data, sliding window smoothing processing is adopted, and the window size is usually set to 5 - 10 data points; for data with large fluctuations such as light intensity, Kalman filtering is used for processing. For missing values, the system uses a hybrid model of time series prediction and spatial correlation analysis for intelligent completion. Specifically, when a certain sensor data is missing, the system will first use the historical data of the sensor for time series prediction. If the historical data is insufficient, it will use the spatial correlation data of nearby sensors for auxiliary inference.

[0053] The multi-source data fusion unit 22, connected to the data cleaning unit 21, is used to integrate environmental data, historical maintenance records, and plant growth status data. Multi-source data fusion is an important innovation point of the present invention. By fusing heterogeneous data from different sources, the system can establish a more comprehensive and accurate landscape environment model. In practical applications, the system correlates environmental data, historical maintenance records, and plant growth status data according to timestamps and spatial positions to construct a four-dimensional data model of "time-space - environment - growth status". This multi-source fusion method enables the system to capture complex correlation patterns that cannot be reflected by a single data source.

[0054] The feature engineering unit 23, connected to the multi-source data fusion unit 22, is used to generate feature vectors through time series feature extraction and spatial feature correlation mining. Feature engineering is the key to the effectiveness of deep learning models. The present invention adopts a variety of advanced feature extraction techniques. For time series features, the system uses wavelet transform and empirical mode decomposition methods to extract multi-scale periodic features from environmental data. For example, through wavelet transform, features at time scales such as daily temperature difference, weekly average temperature, and monthly precipitation can be captured. For spatial features, the system establishes a spatial autocorrelation model to discover the spatial propagation patterns of plant health status in different regions. Practice shows that by extracting these high-quality features, the prediction accuracy of the model can be increased by 15 - 20%.

[0055] The model layer module 3 is the core intelligent unit of this system, connected to the data layer module 2, and is responsible for constructing a plant disease prediction model based on deep learning algorithms. As Figure 4 shown, the model layer module 3 includes:

[0056] The training unit 31 is used to construct and train a deep learning model based on feature vectors. The training unit 31 supports a variety of deep learning algorithms, including convolutional neural network (CNN), recurrent neural network (RNN), time-space dual-stream neural network, and hybrid attention mechanism network. Taking the hybrid attention mechanism network as an example, its network structure is as follows:

[0057] Input layer → Feature extraction layer → Spatial attention module → Channel attention module → Fusion layer → Fully connected layer → Output layer

[0058] Among them, the calculation formula of the spatial attention module is:

[0059] M s (F) = σ(f 7×7 ([AvgPool(F); MaxPool(F)]))

[0060] The calculation formula of the channel attention module is:

[0061] M c(F) = σ(W2(ReLU(W1(AvgPool(F))))),

[0062] The formula for calculating the fused attention feature is:

[0063]

[0064] where F represents the input feature, σ represents the sigmoid activation function, f 7×7 represents the 7×7 convolution operation, AvgPool and MaxPool represent the average pooling and max pooling operations respectively, W1 and W2 represent the weights of the fully connected layers, represents the element-wise multiplication.

[0065] During the training process, the system uses the Adam optimizer with a batch size of 64 and a learning rate of 0.001, and the number of training epochs is set to 100. To prevent overfitting, L2 regularization with a weight decay coefficient of 0.0001 and Dropout technology with a probability of 0.5 are also adopted.

[0066] The prediction unit 32, connected to the training unit 31, is used to input the plant environment data into the trained deep learning model to generate the plant disease prediction results. The prediction unit 32 classifies the occurrence of plant diseases into 5 categories: slight, slight to moderate, moderate to severe, severe, and critical, and outputs the corresponding disease occurrence probabilities. This multi-level classification design enables the maintenance measures to adopt differentiated strategies for different severity levels of disease risks. For example, for the risk predicted as "slight" (the probability threshold is usually set to 0.1 - 0.3), the system may recommend increasing the monitoring frequency; while for the "critical" level of risk (the probability threshold is usually greater than 0.7), the system will recommend taking immediate preventive treatment measures.

[0067] The evaluation unit 33, connected to the prediction unit 32, is used to evaluate the model performance through the confusion matrix and F1 score. The measurement elements of the confusion matrix include true positive (TP), false positive (FP), true negative (TN), and false negative (FN). The formula for calculating the F1 score is:

[0068]

[0069] where the formula for precision is:

[0070]

[0071] The formula for recall is:

[0072]

[0073] In practical applications, the system requires that the F1 score reach above 0.85 before considering the model performance good enough to be put into practical use.

[0074] The decision-making layer module 4 is connected to the model layer module 3 and is responsible for generating a scoring and ranking of maintenance plans based on the results of the maintenance requirement prediction. As Figure 5 shown, the decision-making layer module 4 includes:

[0075] A plan generation unit 41, which is used to create multiple optional maintenance plans based on the results of the maintenance requirement prediction. The plan generation unit 41 screens out multiple optional maintenance plans suitable for the current situation from the plan library according to the predicted disease types and severity levels, combined with factors such as plant types and seasonal characteristics. For example, for the predicted possible root rot disease, the system will generate multiple optional plans including adjusting the watering frequency, applying organic fertilizers, spraying biological control agents, etc. In practical applications, the system usually generates 3 - 5 optional maintenance plans for subsequent scoring and ranking.

[0076] A scoring unit 42, which is connected to the plan generation unit 41 and is used to score the optional maintenance plans by the linear weighting method. The scoring formula is:

[0077] Score of the maintenance plan = Historical usage scoring weight × Historical usage score × (1 - Disease recurrence rate 1) × (1 - Disease recurrence rate 2) + Rationality scoring weight × Rationality score + Cost weight × Cost score,

[0078] This multi-dimensional scoring mechanism is one of the important innovation points of the present invention. By comprehensively considering three key factors: historical effect, rationality, and cost, it realizes the scientific and quantitative maintenance decision-making. Among them, the historical usage score reflects the effect of the plan in historical applications and is automatically calculated by the system based on historical data; the rationality score reflects the reasonable degree of the plan in terms of botany and horticulture theory and is usually preset by domain experts; the cost score reflects the economy of the plan, including factors such as material cost, labor cost, and time cost.

[0079] A ranking unit 43, which is connected to the scoring unit 42 and is used to rank the maintenance plans according to the scores of the maintenance plans. The ranking unit 43 arranges all the optional maintenance plans in descending order of scores, and the plan with the highest score will be recommended as the first choice to the maintenance personnel. In practical applications, the system usually displays the top three ranked plans by scores and attaches an analysis of the advantages and disadvantages of each plan to assist the maintenance personnel in making the final decision.

[0080] The feedback layer module 5 is connected to the decision-making layer module 4 and is responsible for collecting the results of maintenance implementation and optimizing the plant disease prediction model. As Figure 6 shown, the feedback layer module 5 includes:

[0081] The effect evaluation unit 51 is used to quantitatively evaluate the actual effect after the implementation of the maintenance measures. Effect evaluation is a key link in the continuous optimization of the system. The present invention uses a variety of methods to quantitatively evaluate the maintenance effect. For example, through image contrast analysis, the system can quantitatively calculate the improvement degree of plant health; through sensor data analysis, the system can evaluate the change trend of environmental parameters; through the analysis of maintenance cycle and resource consumption, the system can evaluate the improvement degree of maintenance efficiency. These quantitative indicators usually include disease incidence, plant growth rate, leaf color index, etc., providing an objective data basis for subsequent model optimization.

[0082] The knowledge precipitation unit 52 is connected to the effect evaluation unit 51 and is used to construct the maintenance experience and data analysis results into a structured knowledge graph. Knowledge precipitation is another important innovation point of the present invention. By organizing the scattered maintenance experience and data analysis results into a structured knowledge graph, the systematic precipitation and reuse of knowledge are realized. The construction of the knowledge graph adopts the basic structure of triple (subject-relationship-object), such as "Epipremnum aureum - prone to - root rot", "high temperature and dryness - cause - powdery mildew", etc. In practical applications, the system has accumulated more than 10,000 maintenance knowledge nodes and 30,000 relationship edges, forming a rich knowledge network.

[0083] The model optimization unit 53 is connected to the knowledge precipitation unit 52 and is used to continuously optimize the deep learning model parameters based on the maintenance implementation effect. The model optimization adopts an incremental learning strategy, continuously absorbing new maintenance practice experience while retaining the original knowledge. Specifically, the system will regularly (usually once a week) fine-tune the model based on the newly added data. The learning rate adjustment is usually set to 1 / 10 of the original learning rate, that is, 0.0001, and the fine-tuning rounds are set to 10 - 20 rounds. At the same time, the system also sets a model performance monitoring mechanism. When the model performance drops by more than 5%, a more comprehensive retraining process will be triggered.

[0084] The visualization module 6 is connected to the decision-making layer module 4 and is used to display the system status and maintenance suggestions through a graphical interface. The visualization module 6 includes a front-end user interface 61 and a back-end Web server 62. The front-end user interface 61 adopts a responsive design, supporting access from both the PC side and the mobile side, and mainly displays core information such as environmental monitoring data, disease prediction results, and maintenance suggestions. The interface design follows the principles of intuitiveness and operability, using visualization charts such as dashboards, heat maps, and trend charts to help maintenance personnel quickly understand the system status and decision-making basis. The back-end Web server 62 is responsible for processing front-end requests, calling the core function modules of the system, and returning the processing results to the front-end for display.

[0085] Embodiment 2: Landscape engineering maintenance management method based on deep learning

[0086] The present invention also provides a landscape engineering maintenance management method based on deep learning, comprising the following steps:

[0087] Step 1: Collect monitoring data of the landscape environment through a multi-dimensional heterogeneous sensing network and send it to the cloud server. This step is mainly completed by the sensing layer module 1. The sensors are arranged according to the "dense-sparse-dense" three-dimensional monitoring grid principle, and the data is preliminarily filtered by the edge computing unit 12 to eliminate obvious outliers, and then the data is transmitted to the cloud server through the communication unit 13.

[0088] Step 2: Preprocess and extract features from the monitoring data to generate feature vectors. This step is mainly completed by the data layer module 2, including three main links: data cleaning, multi-source data fusion, and feature engineering. The data cleaning link processes outliers, missing values, and duplicate values; the multi-source data fusion link integrates environmental data, historical maintenance records, and plant growth status data; the feature engineering link generates feature vectors through time series feature extraction and spatial feature correlation mining.

[0089] Step 3: Build and train a deep learning model based on the feature vectors to form a plant disease prediction model. This step is mainly completed by the training unit 31 of the model layer module 3. The system supports a variety of deep learning algorithms, including convolutional neural network, recurrent neural network, time series-space two-stream neural network, and hybrid attention mechanism network, etc. During the training process, the system uses the cross-validation method to evaluate the model performance and improves the model prediction accuracy through hyperparameter tuning.

[0090] Step 4: Analyze the real-time collected environmental data using the plant disease prediction model to generate the maintenance requirement prediction results. This step is mainly completed by the prediction unit 32 of the model layer module 3. The system classifies the prediction results into 5 categories: slight, slight to moderate, moderate to severe, severe, and extremely severe, and outputs the corresponding disease occurrence probabilities.

[0091] Step 5: Create multiple optional maintenance plans based on the maintenance requirement prediction results. This step is mainly completed by the plan generation unit 41 of the decision-making layer module 4. The system will screen out 3-5 optional maintenance plans suitable for the current situation from the plan library according to the predicted disease types and severity levels.

[0092] Step 6: Score and rank the optional maintenance plans through the linear weighted method to generate the optimal maintenance decision. This step is mainly completed by the scoring unit 42 and the ranking unit 43 of the decision-making layer module 4. The system uses the linear weighted method to comprehensively evaluate the optional maintenance plans, considering three dimensions: historical effect, rationality, and cost. Calculate the comprehensive score of each plan, and then rank them from high to low according to the scores. The plan with the highest score is recommended as the optimal maintenance decision to the maintenance personnel.

[0093] Step 7: Execute the optimal maintenance decision and collect the results of maintenance implementation. This step is implemented by the maintenance personnel under the guidance of the system's suggestions. The system records the process and results of maintenance implementation through the mobile terminal, including information such as implementation time, operation content, resource consumption, and preliminary effects. For typical cases, the system will also collect image data before and after implementation to provide an intuitive reference for subsequent effect evaluation.

[0094] Step 8: Optimize the plant disease prediction model based on the results of maintenance implementation to achieve the continuous learning and evolution of the system. This step is mainly completed by the feedback layer module 5. The system quantitatively evaluates the maintenance implementation effect through the effect evaluation unit 51, converts the empirical results into structured knowledge through the knowledge precipitation unit 52, and continuously optimizes the parameters of the deep learning model through the model optimization unit 53 to form a closed-loop feedback mechanism and achieve continuous improvement of the system's intelligence level.

[0095] Embodiment 3: Specific implementation of the hybrid attention mechanism network

[0096] As a core innovation point of the present invention, the hybrid attention mechanism network plays a key role in the plant disease prediction model. This network combines spatial attention and channel attention to enhance the sensitivity of the model to changes in key environmental factors. The following details its implementation method.

[0097] The basic structure of the hybrid attention mechanism network includes an input layer, a feature extraction layer, a spatial attention module, a channel attention module, a fusion layer, a fully connected layer, and an output layer. Among them, the feature extraction layer uses ResNet50 as the backbone network to extract the basic feature representation of the input data.

[0098] The spatial attention module is used to focus on the important spatial positions in the feature map, and the calculation formula is:

[0099] M s (F) = σ(f 7×7 ([AvgPool(F); MaxPool(F)]))

[0100] where, F ∈ R C×H×WDenote the input feature map, where C is the number of channels, and H and W are the height and width of the feature map respectively; AvgPool and MaxPool represent average pooling and max pooling operations along the channel dimension, respectively, to obtain two two-dimensional feature maps; [·;·] represents the concatenation operation along the channel dimension; f 7×7 Denote the convolution operation with a kernel size of 7×7; σ represents the sigmoid activation function, which compresses the output range to the interval [0,1]. The finally obtained M s [F) ∈ R 1×H×W is a spatial attention map, and the larger the value, the more important the position.

[0101] The channel attention module is used to focus on important feature channels, and the calculation formula is:

[0102] M c (F) = σ(W2(ReLU(W1(AvgPool(F))))),

[0103] where AvgPool represents global average pooling, which compresses the feature map of each channel into a scalar; W1 ∈ R C / r×C and W2 ∈ R C×C / r represent the weight matrices of two fully connected layers, r is the dimensionality reduction ratio, usually taken as 16; ReLU is the activation function; σ is the sigmoid activation function. The finally obtained M c (F) ∈ R C×1×1 is a channel attention vector, and the larger the value, the more important the channel.

[0104] The fusion layer fuses the results of spatial attention and channel attention with the original features, and the calculation formula is:

[0105]

[0106] where, denotes element-wise multiplication, and F ′ is the fused feature map. This dual attention mechanism enables the model to simultaneously focus on important spatial positions and feature channels, significantly improving the sensitivity to changes in key environmental factors.

[0107] In practical applications, the hybrid attention mechanism network exhibits excellent performance. Compared with traditional convolutional neural networks, the prediction accuracy for plant diseases is increased by approximately 12.5%, especially outstanding in the detection of early weak signs, and it can detect early disease signals that cannot be captured by conventional models 3 - 5 days in advance.

[0108] Example 4: Practical Application of the Maintenance Plan Scoring Mechanism

[0109] The present invention innovatively introduces a scoring mechanism for maintenance plans based on three dimensions: historical effectiveness, rationality, and cost, making the decision-making process more scientific and quantitative. The following uses a specific case to illustrate its practical application.

[0110] Suppose the system predicts that there is an 80% probability that the Epipremnum aureum plants in a certain area will develop mild to moderate root rot within the next 7 - 10 days. Based on this prediction result, the plan generation unit 41 screens out three alternative maintenance plans from the plan library:

[0111] Plan A: Adjust the watering frequency and increase ventilation appropriately.

[0112] Plan B: Use biological control agents for preventive treatment.

[0113] Plan C: Replace part of the soil and add organic fertilizers.

[0114] Next, the scoring unit 42 scores these three plans from multiple dimensions:

[0115] For Plan A, the historical usage score is 8.5 (out of 10), historical data shows that the short-term disease recurrence rate of this plan is 0.15, the long-term disease recurrence rate is 0.25, the rationality score is 9.0, and the cost score is 9.5 (low cost).

[0116] For Plan B, the historical usage score is 9.0, historical data shows that the short-term disease recurrence rate of this plan is 0.10, the long-term disease recurrence rate is 0.20, the rationality score is 8.5, and the cost score is 7.0 (medium cost).

[0117] For Plan C, the historical usage score is 9.5, historical data shows that the short-term disease recurrence rate of this plan is 0.05, the long-term disease recurrence rate is 0.15, the rationality score is 8.0, and the cost score is 5.5 (high cost).

[0118] Calculate the scores of each plan according to the scoring formula (assuming the weight of the historical usage score is 0.5, the weight of the rationality score is 0.3, and the cost weight is 0.2):

[0119] Score of Plan A = 0.5×8.5×(1 - 0.15)×(1 - 0.25) + 0.3×9.0 + 0.2×9.5 = 7.53

[0120] Score of Plan B = 0.5×9.0×(1 - 0.10)×(1 - 0.20) + 0.3×8.5 + 0.2×7.0 = 7.95

[0121] Score of Plan C = 0.5×9.5×(1 - 0.05)×(1 - 0.15) + 0.3×8.0 + 0.2×5.5 = 8.09

[0122] The sorting unit 43 sorts the three solutions according to the calculation results: Solution C > Solution B > Solution A. Therefore, the system recommends Solution C (replacing part of the soil and adding organic fertilizers) as the optimal maintenance decision. Although the cost of Solution C is relatively high, its historical effect is the best, the disease recurrence rate is the lowest, and the comprehensive score is the highest, which conforms to the principle of scientific maintenance.

[0123] This case demonstrates how the maintenance solution scoring mechanism of the present invention weighs multiple factors in practical applications to make scientific and reasonable maintenance decisions, helping maintenance personnel effectively prevent plant diseases and improve the efficiency and quality of maintenance.

[0124] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A landscape engineering maintenance management system based on deep learning, characterized by: include: The perception layer module is used to collect monitoring data of the landscape environment through a multi-dimensional heterogeneous sensor network and send it to the cloud server; A data layer module, wirelessly connected to the perception layer module, for preprocessing and feature extraction of the monitoring data; A model layer module, connected to the data layer module, for building a plant disease prediction model based on a deep learning algorithm and generating maintenance demand prediction results; a decision layer module, connected to the model layer module, for generating a maintenance plan score ranking based on the maintenance demand prediction results; The feedback layer module is connected to the decision layer module and is used to collect maintenance implementation results and optimize the plant disease prediction model.

2. The landscape engineering maintenance management system based on deep learning according to claim 1 is characterized in that: The perception layer module includes: an intelligent environmental sensor unit, used to collect air humidity, light intensity, soil moisture and temperature data; an edge computing unit, connected to the intelligent environmental sensor unit, used to perform preliminary filtering and anomaly detection on the collected environmental data; a communication unit, connected to the edge computing unit, used to transmit the filtered environmental data to the cloud server through the Internet of Things.

3. The landscape engineering maintenance management system based on deep learning according to claim 1 is characterized in that: The data layer module includes: a data cleaning unit, which is used to process abnormal values, missing values ​​and duplicate values ​​of the monitoring data; a multi-source data fusion unit, which is connected to the data cleaning unit and is used to integrate environmental data, historical maintenance records and plant growth status data; a feature engineering unit, which is connected to the multi-source data fusion unit and is used to generate feature vectors through time series feature extraction and spatial feature association mining.

4. The landscape engineering maintenance management system based on deep learning according to claim 1 is characterized in that: The model layer module includes: a training unit, which is used to construct and train a deep learning model based on the feature vector; a prediction unit, which is connected to the training unit and is used to input plant environment data into the trained deep learning model to generate a plant disease prediction result; and an evaluation unit, which is connected to the prediction unit and is used to evaluate the model performance through a confusion matrix and an F1 score.

5. The landscape engineering maintenance management system based on deep learning according to claim 4 is characterized in that: The prediction unit classifies the occurrence of plant diseases into five categories: mild, mild to moderate, moderate to severe, severe and serious, and outputs the corresponding disease occurrence probabilities.

6. The landscape engineering maintenance management system based on deep learning according to claim 1 is characterized in that: The decision-making layer module includes: a solution generation unit, which is used to create a variety of optional maintenance solutions based on the maintenance demand prediction results; a scoring unit, which is connected to the solution generation unit and is used to score the optional maintenance solutions through a linear weighted method, and the scoring formula is: maintenance solution score = historical use score weight * historical use score * (1-disease recurrence rate 1) * (1-disease recurrence rate 2) + rationality score weight * rationality score + cost weight * cost score; a sorting unit, which is connected to the scoring unit and is used to sort the maintenance solutions according to the maintenance solution scores.

7. The landscape engineering maintenance management system based on deep learning according to claim 1 is characterized in that: The feedback layer module includes: an effect evaluation unit, which is used to quantitatively evaluate the actual effect of the maintenance measures after implementation; a knowledge precipitation unit, which is connected to the effect evaluation unit and is used to construct the maintenance experience and data analysis results into a structured knowledge graph; a model optimization unit, which is connected to the knowledge precipitation unit and is used to continuously optimize the deep learning model parameters based on the maintenance implementation effect.

8. The landscape engineering maintenance management system based on deep learning according to claim 1 is characterized in that: Also includes: A visualization module is connected to the decision layer module and is used to display the system status and maintenance suggestions through a graphical interface; wherein the visualization module includes a front-end user interface and a back-end Web server.

9. The landscape engineering maintenance management system based on deep learning according to claim 1, characterized in that: The deep learning algorithm includes: a convolutional neural network, a recurrent neural network, a temporal-spatial dual-stream neural network and a hybrid attention mechanism network; wherein the hybrid attention mechanism network combines spatial attention and channel attention to enhance the model's sensitivity to changes in key environmental factors.

10. A landscape engineering maintenance management method based on deep learning, using the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: Collecting monitoring data of the landscape environment through a multi-dimensional heterogeneous sensor network and sending it to a cloud server; preprocessing and feature extraction of the monitoring data to generate a feature vector; Constructing and training a deep learning model based on the feature vector to form a plant disease prediction model; using the plant disease prediction model to analyze the environmental data collected in real time to generate a maintenance demand prediction result; Creating a plurality of optional maintenance plans based on the maintenance demand prediction results; Scoring and ranking the optional maintenance plans by a linear weighted method to generate an optimal maintenance decision; Executing the optimal maintenance decision and collecting maintenance implementation results; The plant disease prediction model is optimized based on the maintenance implementation results to achieve continuous learning and evolution of the system.

Images