Intelligent ecological environmental protection monitoring method and system based on artificial intelligence

Abstract

The invention belongs to the technical field of environmental protection ecological monitoring, and particularly relates to an intelligent ecological environmental protection monitoring method and system based on artificial intelligence, and the method comprises the steps: integrating and constructing a multi-source data collection network, introducing a Q-Learning algorithm to dynamically adjust the sampling and cruising frequency, and achieving the multi-dimensional data collection; data abnormal values are identified and eliminated, and heterogeneous data are processed in a standardized mode; designing a cross-modal fusion network containing a coordinate attention mechanism, extracting various types of data features, and weighting by using an attention weight coefficient to generate a unified environment state vector; designing an intelligent analysis algorithm, intelligently identifying a pollution source, tracing the pollution source in combination with a graph attention network, and dynamically adjusting an early warning threshold in combination with a PPO algorithm; and optimizing a self-adaptive monitoring strategy, detecting a sensor fault through a self-encoder, and repairing missing data by adopting a K-nearest neighbor interpolation method to maintain data continuity. According to the invention, the coverage range of ecological monitoring can be effectively expanded, the data intelligent processing capability is enhanced, and the early warning accuracy and adaptability are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of environmental protection and ecology monitoring, and particularly relates to an intelligent ecological environmental protection monitoring method and system based on artificial intelligence. BACKGROUND

[0002] The development of environmental monitoring technology has always evolved in sync with human demand for ecological protection, from early manual sampling analysis to today's intelligent monitoring network. The technical system has formed a complex architecture with multiple dimensions and multiple levels.

[0003] In the field of ground monitoring networks, the iteration of sensor technology has driven the leap in data collection capabilities. The current mainstream monitoring terminals cover multiple parameters such as air, water quality, and soil. For example, the ultraviolet differential optical absorption spectroscopy (DOAS) technology used in air monitoring can simultaneously measure multiple pollutants such as SO2 and NOx, with an accuracy of ppb level. In water quality monitoring, ion-selective electrode-based sensors can measure pH, ammonia nitrogen, and other indicators in real time, with a response time controlled within 10 seconds. These sensors are connected to regional monitoring platforms through Internet of Things protocols (such as MQTT and CoAP), forming a dense network of data collection nodes. Taking a provincial ecological monitoring network as an example, more than 2000 ground stations deployed in the city have achieved 1-kilometer grid coverage in the urban built-up area, with data transmission delay less than 5 minutes, providing a foundation for environmental management.

[0004] Remote sensing monitoring technology plays an irreplaceable role in macro ecological condition assessment. Satellite remote sensing, with its advantages of wide range and periodic observation, has become the main means of monitoring land use change, vegetation coverage, and water eutrophication. For example, the sub-meter stereo mapping capability of Gaofen-7 satellite can accurately identify the spatial distribution of urban green space. Sentinel-5P satellite uses ultraviolet-visible near-infrared spectrometer to achieve monthly inversion of global atmospheric pollutant concentration, with a spatial resolution of 3.5 kilometers. Unmanned aerial vehicle remote sensing, as a supplement to satellite remote sensing, has shown outstanding performance in small-scale regional monitoring. Unmanned aerial vehicles equipped with hyperspectral cameras can obtain 5-nanometer-bandwidth images at a flight height of 100 meters, which can be used for fine mapping of farmland soil heavy metal pollution, with an identification accuracy improved by more than 30% compared with traditional methods.

[0005] Data processing and analysis techniques are also evolving. Early environmental data analysis relied heavily on statistical methods, such as describing the distribution of pollutant concentrations through mean and standard deviation, and exploring influencing factors through correlation analysis. With the rise of machine learning, algorithms such as support vector machines and random forests have been used for pollution prediction. For example, a research team used a random forest model to predict urban PM2.5 concentrations, achieving an accuracy rate of 78% for short-term predictions. In recent years, deep learning techniques have been applied to environmental monitoring. Convolutional neural network (CNN)-based image recognition models can automatically identify pollution sources in remote sensing images, and recurrent neural network (RNN)-based time series prediction models can improve the ability to predict the trend of pollutant concentration changes. Some regions have built environmental big data platforms that integrate monitoring data, pollution source data, and meteorological data, providing a data foundation for comprehensive analysis.

[0006] Despite the progress made in existing environmental monitoring technologies, there are still many shortcomings in practical applications, making it difficult to meet the increasingly complex ecological and environmental protection needs.

[0007] One of the prominent problems is the lack of coverage and coordination of the monitoring network. Ground sensor networks are limited by cost, and the deployment density is low in remote areas and ecologically sensitive areas, resulting in a large number of monitoring blind spots. For example, in the water quality monitoring of mountainous rivers, monitoring points can only be set up near a few towns, making it difficult to reflect the water quality status of the entire river basin. Although remote sensing technology has a wide coverage, satellite remote sensing has a contradiction between time resolution and spatial resolution. High spatial resolution satellites have a long revisit period, making it difficult to capture rapid changes in environmental factors. The endurance of unmanned aerial vehicle remote sensing is limited, and a single flight usually lasts no more than an hour, making it difficult to complete continuous monitoring of large areas. At the same time, ground monitoring and remote sensing monitoring data lack effective coordination, and the scale of ground point data and remote sensing area data does not match, resulting in poor data fusion effect and making it difficult to form a three-dimensional monitoring capability combining "point" and "area".

[0008] The low level of intelligence in data processing restricts the effectiveness of monitoring. Existing systems mainly rely on simple statistical analysis and threshold alarm for monitoring data, lacking deep mining and intelligent decision support capabilities. For example, when the concentration of pollutants exceeds the standard, the system can only send an alarm signal, but cannot automatically analyze the reasons for exceeding the standard and trace the source of pollution. The application of machine learning algorithms also has limitations. Most models are trained for a single pollutant and a single region, with poor generalization ability. When the monitoring area or pollutant type changes, the model needs to be retrained, and the adaptability is not strong. In addition, environmental monitoring data is multi-source and heterogeneous, including structured data from sensors, image data from remote sensing, and video data from video surveillance. Existing technologies cannot effectively integrate these data, resulting in the failure to fully release the value of data.

[0009] Another shortcoming of the existing technology is the weak prediction and early warning capability. Environmental problems often have complexity and suddenness, and need to be predicted in advance and warned in time to gain time for emergency disposal. However, the existing prediction models are mostly based on the trend extrapolation of historical data, and have insufficient prediction capability for sudden pollution events. For example, in sudden water pollution events, the existing models are difficult to accurately predict the diffusion path and impact range of pollutants, resulting in delayed emergency response. At the same time, the setting of early warning threshold mostly uses fixed values, without considering the differences in different regions and different seasons, which is easy to cause false positives and false negatives. Statistics of an environmental protection department shows that the false positive rate of its monitoring system is as high as 20%, which not only increases the burden of staff, but also reduces the credibility of the system.

[0010] The lack of autonomy and adaptability of the system also affects the continuity and effectiveness of the monitoring. The existing monitoring systems are mostly fixed deployment, and lack dynamic adjustment capability. When the environmental conditions of the monitoring area change, such as the addition of pollution sources and changes in ecological landscape, the system cannot automatically adjust the monitoring parameters and strategies. Device maintenance also mainly relies on manual work, and when the sensor fails, it is difficult to find and repair in time, resulting in data loss. The operation data of a certain monitoring network shows that the data loss rate caused by equipment failure reaches 15%, which seriously affects the continuity and integrity of the data. In addition, the standards of monitoring systems in different regions and different departments are not the same, and the data is difficult to share and interoperate, forming an "information island", which restricts the coordination and integrity of environmental management.

[0011] The existence of these defects makes it difficult for the existing environmental monitoring technology to meet the fine and intelligent needs of ecological environment protection in the new era, and it is urgent to develop intelligent ecological environmental protection monitoring methods and systems based on artificial intelligence to improve the comprehensiveness, accuracy and intelligent level of monitoring, and provide stronger technical support for ecological environment governance.

[0012] An ecological monitoring method and system for water and soil conservation are disclosed in prior art CN120278478A. The system includes a vegetation evaluation module, a soil evaluation module, a water flow and sediment evaluation module, and an ecological comprehensive evaluation module. The vegetation evaluation module includes a vegetation data collection unit and a vegetation data analysis unit. The soil evaluation module includes a soil data collection unit and a soil analysis unit. The system monitors and dynamically responds to three key ecological elements: vegetation, soil, and water flow. When any indicator exceeds the preset risk threshold, the system can issue corresponding dispatch instructions in time to take necessary protective measures to prevent potential ecological problems from worsening. When all indicators are within the safe range, the system can accurately determine the overall ecological condition by constructing an ecological comprehensive evaluation coefficient to make a decision on whether to issue a warning dispatch. However, the system has relatively low intelligence, lacks autonomy and adaptability in different scenarios, and has relatively poor warning capability. In prior art CN120014813A, a forest ecological monitoring method and system based on sound analysis are disclosed. The system includes obtaining real-time continuous recordings of the forest monitoring area and embedding timestamp and geographic location information in the recordings to ensure that all recording data have accurate spatiotemporal identifiers. The system uses short-time energy calculation to remove silent parts of the recording data and dynamically sets an energy threshold to retain valid signals to obtain valid audio data. The system converts valid audio data into a time-frequency matrix through STFT and adjusts filter parameters based on real-time weather data to perform dynamic noise reduction processing and obtain noise-reduced frequency spectrum. The system extracts MFCC features using a mel filter bank transform and DCT to generate standardized feature vectors. The system uses a pre-trained CNN and LSTM to perform deep learning classification on the MFCC features, sets statistical thresholds based on historical data, detects abnormalities, and finally generates output data containing classification results, activity intensity, and abnormal warning information. However, this invention only considers sound features, has low data utilization, and is only suitable for forests with less noise interference and different ecological scenarios with distinct sound features. SUMMARY

[0013] To solve the above problems, the present application provides an intelligent ecological environmental protection monitoring method and system based on artificial intelligence to solve the problems existing in the above prior art and achieve wide coverage, high precision, intelligence, and certain adaptive ability of ecological environmental protection monitoring.

[0014] To achieve the above purpose, the present application provides the following technical solutions:

[0015] An intelligent ecological environmental protection monitoring method based on artificial intelligence is applied to an intelligent ecological environmental protection monitoring system based on artificial intelligence, which includes the following steps:

[0016] S1: Construct a multi-source data acquisition network, integrate satellite remote sensing equipment, multiple types of unmanned aerial vehicles (carrying hyperspectral and thermal infrared cameras), ground sensor arrays (atmospheric micro-stations, water quality monitoring stations), and underwater vehicles; synchronize time through the Beidou timing system, use the UTM projection coordinate system as a spatial reference, introduce the Q-Learning algorithm to dynamically adjust the sampling and cruising frequency, and achieve multi-dimensional data acquisition;

[0017] S2: Data preprocessing and fusion, use the improved Z-Score algorithm to identify and eliminate outliers, use the FLAASH model to perform radiation and geometric correction on remote sensing images, and process heterogeneous data through min-max standardization; design a cross-modal fusion network with coordinate attention mechanism, extract features of each type of data, and use attention weight coefficients to generate a unified environmental state vector;

[0018] S3: Design intelligent analysis algorithm, identify pollution sources based on improved YOLOv8 model; construct regional topology graph, combine graph attention network to trace pollution sources, solve diffusion equation by finite difference method to simulate diffusion path; use Attention-LSTM model to predict pollutant concentration, combine PPO algorithm to dynamically adjust warning threshold;

[0019] S4: Optimize adaptive monitoring strategy, construct device scheduling model based on deep Q network, take grid cell state, device location and power as input, generate cruise path of unmanned aerial vehicle and underwater vehicle; detect sensor failure through autoencoder, repair missing data using K-nearest neighbor interpolation method to maintain data continuity.

[0020] Further, the step S1 comprises:

[0021] S11: Satellite remote sensing equipment deployment and parameter configuration, select Sentinel-2 and Gaofen-6 satellites, set 10-meter spatial resolution for Sentinel-2 satellite, 5-day revisit period, set 8-meter spatial resolution for Gaofen-6 satellite, 4-day revisit period, collect multi-spectral data in the wavelength range of 400-2500nm, focus on monitoring macro indicators such as vegetation coverage and water area; simultaneously access atmospheric pollutant data from Sentinel-5P satellite, set satellite resolution to 3.5 kilometers, realize regional scale pollution distribution monitoring;

[0022] S12: Multi-type unmanned aerial vehicle deployment and parameter configuration, deploy different types of quadcopters in ecologically sensitive areas, carry 5nm band width hyperspectral camera and 640x512 resolution thermal infrared camera, cruise every 2 hours, collect 1-meter resolution ground reflectance and temperature data, capture micro-scale ecological changes;

[0023] S13: Deployment and parameter configuration of underwater vehicles. Deploy autonomous underwater vehicles (AUVs) in the seabed, rivers, lakes and reservoirs, equipped with acoustic Doppler current profilers (ADCP) and water quality sensors to collect water depth, current velocity and chlorophyll a concentration data in real time, with a cruising range covering more than 90% of the water area.

[0024] S14: Design of a spatiotemporal synchronization mechanism to establish a unified spatiotemporal benchmark and solve the problem of time and space matching of multi-source data. Specific steps include:

[0025] S141: Time synchronization, all devices use BeiDou time synchronization, and the time error is controlled within 10ms; timestamp correction is performed on satellite remote sensing data to make the imaging time accurate to the second level;

[0026] S142: Spatial registration, using the UTM projection coordinate system, uniformly transforms UAV and ground sensor data to this coordinate system; remote sensing images are corrected through ground control points (GCP) to make the spatial deviation of data from different sources less than 1 meter;

[0027] S143: Dynamic sampling strategy design, based on reinforcement learning algorithm (Q-Learning) to dynamically adjust the sampling frequency. When an anomaly occurs in the monitoring area, the drone's cruise frequency and sensor sampling frequency are automatically increased. In non-anomaly areas, the normal frequency is maintained to save energy.

[0028] In a preferred embodiment, the air pollutant data mentioned in step S11 includes NO2 and SO2 concentrations.

[0029] In a preferred embodiment, the ecologically sensitive area mentioned in step S12 includes nature reserves and wetlands.

[0030] Further, step S2 includes:

[0031] S21: Outlier removal. An improved Z-Score algorithm is used to identify outlier data. The calculation expression is:

[0032]

[0033] Where, x i For the original data, μ is the mean and σ is the standard deviation; when |Z i When | > 3.29, corresponding to a 99.9% confidence interval, it is determined to be an outlier, and the gap is filled by linear interpolation.

[0034] S22: Remote sensing data correction. Radiometric correction is performed on satellite and UAV imagery to eliminate atmospheric scattering effects. Several corrections are applied to the satellite and UAV imagery. The surface reflectance is calculated using the FLAASH model. The calculation expression is as follows:

[0035]

[0036] Wherein, p is the surface reflectivity, L is the apparent radiation brightness, d is the distance between the earth and the sun, E0 is the solar radiation at the top of the atmosphere, and θ is the solar zenith angle.

[0037] S23: Data standardization, mapping data of different dimensions to the interval [0, 1], using the min-max standardization formula for data standardization, and the calculation expression is:

[0038]

[0039] Wherein, x min , x max are the minimum and maximum values in the data set, respectively;

[0040] S24: Design a cross-modal fusion network (CMFN) based on attention mechanism to realize deep fusion of heterogeneous data, and the specific steps include:

[0041] S241: Feature extraction layer design, for remote sensing images, ResNet-50 network is used to extract spatial features, and a 512-dimensional vector is outputted;

[0042] For sensor data, a 1D convolution layer with a size of 3 is used to extract the time sequence features of the sensor data, and a 256-dimensional vector is outputted;

[0043] For spectral data, PCA dimension reduction is used to retain 95% of the variance, and then a fully connected layer is used to convert it into a 256-dimensional feature vector;

[0044] S242: Attention fusion layer design, calculate the weight coefficient α i of different modal features, the calculation expression is:

[0045]

[0046] Wherein, f i is the feature vector of the i-th modal, and s(·) is the attention score function, which is composed of a single hidden layer neural network; the fusion feature vector is obtained by weighted summation, and the calculation expression is:

[0047]

[0048] S243: Output layer design, input the fusion features into the fully connected layer, and output a unified environmental state vector, including 128-dimensional features, covering multi-dimensional information of atmosphere, water quality and vegetation.

[0049] Further, the step S3 includes:

[0050] S31: Pollution source intelligent recognition model design, build pollution source recognition algorithm based on improved YOLOv8, realize automatic positioning of industrial chimney, sewage discharge port pollution source, specific steps include:

[0051] S311: Data set construction, collect 100,000 remote sensing images containing various pollution sources and unmanned aerial images, label target categories and bounding boxes, divide training set, validation set and test set according to the ratio of 8:1:1;

[0052] S312: Model design and improvement, add coordinate attention module between the output end of YOLOv8 backbone architecture C3 module and the down sampling convolution layer, enhance the recognition ability of small targets, the coordinate attention module calculation expression is:

[0053] CA(F) = Conv(σ(MLP(AvgPool(F))))

[0054] Where, F is the input feature map, AvgPool is the average pooling, MLP is the multilayer perceptron, σ is the Sigmoid activation function, Conv is the convolution operation; The C3 module outputs the feature map F, the coordinate attention module respectively performs global average pooling along the x-axis and y-axis on the output feature map F, and obtains two 1D feature vectors; Then the two 1D feature vectors are spliced, and then compressed through 1x1 convolution, and then BatchNorm and ReLU activation; Then the activated spliced vector is split into two independent branches, which are respectively restored through 1x1 convolution, and then the channel number is restored through Sigmoid activation to obtain the x-axis and y-axis attention weights; Finally, the weights are multiplied by the original feature map F, and the enhanced feature map F' is output;

[0055] S313: Model training, place the YOLOv8 model designed and improved in step S312 on the training data set and validation data set constructed in step S311 for model training, use SGD optimizer, learning rate is set to 0.01, momentum is 0.9, upper limit of training generation is set to 50, model training is completed, model performance is tested on test set, and model with good test performance is selected for deployment;

[0056] S32: Build pollution source tracing model combined with graph neural network (GNN), divide the monitoring area into 100x100 meter grid units, build regional topology graph G=(V,E), where V is the grid node, E is the connection relationship between nodes, use graph attention network (GAT) to learn node features:

[0057]

[0058] where is the feature vector of node i, N(i) is the neighbor node set, is the attention coefficient calculated by the self-attention mechanism, and W is the weight matrix; the node with the highest anomaly degree is selected as the most likely pollution source by ranking the anomaly degrees of the node features;

[0059] S33: Diffusion simulation, based on the pollutant diffusion equation, combined with real-time meteorological and hydrological data, to predict the pollutant diffusion range, the prediction calculation expression is:

[0060]

[0061] where c is the pollutant concentration, t is the time, u and v are the flow velocities in x and y directions respectively, and D is the diffusion coefficient;

[0062] S34: Adaptive prediction and early warning model construction, design an LSTM model based on time series attention mechanism (Attention-LSTM) for pollutant concentration prediction and dynamic early warning, the specific steps include:

[0063] S341: Prediction model design, the model input is the environmental state vector in the past n hours, the sample interval can be adjusted according to the actual situation, the model output is the pollutant concentration prediction value in the future m hours, and the model structure includes:

[0064] LSTm layer, set 256 hidden units to capture time series features and calculate hidden states:

[0065] h t =LSTM(x t ,h t-1 )

[0066] where x t is the input vector at time t;

[0067] Attention layer, calculate the weight of hidden state at each time Output weighted hidden state

[0068] Output layer, output the prediction value through the fully connected layer

[0069] S342: Model training, using Adam optimizer, learning rate set to 0.001, training iteration set to 100, loss function using root mean square error loss (RMSE), calculation expression is:

[0070]

[0071] where N is the sample number, i.e. the total number of pollutant concentration prediction samples participating in the calculation; y iis the actual observation value of the ith sample, i.e. the real pollutant concentration value; is the predicted value of the ith sample, i.e. the pollutant concentration prediction result output by the model;

[0072] S343: Dynamic early warning mechanism design, using the proximal policy optimization (PPO) reinforcement learning algorithm to dynamically adjust the early warning threshold, and defining state variables s t , actions a t , and rewards r t respectively;

[0073] The state variable needs to be converted into a feature vector that can be processed by a neural network, where c t is the current pollutant concentration, and the average concentration value in the last n hours is taken according to actual needs; m t is the weather condition, including wind speed, humidity, air pressure, temperature, and light intensity dimension data indicators; is the season, using one-hot encoding form;

[0074] The action a t is the adjusted early warning threshold;

[0075] The reward function r t has the following calculation expression:

[0076] r t = r correct - λ false r false - λ miss r miss

[0077] Where r correct is a correct warning, i.e. the prediction exceeds the standard and the actual value exceeds the standard, which scores 10 points / time; r false is a false alarm, i.e. the prediction exceeds the standard but the actual value does not exceed the standard, which deducts 5 points / time; r miss is a missed alarm, i.e. the actual value exceeds the standard but is not predicted, which deducts 20 points / time; λ false and λ miss are the penalty weights for false alarms and missed alarms respectively, and their size relationship can be set according to actual needs;

[0078] The core objective function selects the clipped surrogate objective function, and the calculation expression is:

[0079]

[0080] Where, is the probability ratio of the new and old strategies; is the advantage function estimate, which is used to measure the advantage of action a t relative to the average level; and ò is the clipping coefficient, which is used to limit r t(θ) is in the range of to avoid too large policy update;

[0081] S344: Initialize the policy network, use a 2-layer fully connected neural network to construct the policy network of PPO reinforcement learning algorithm, the hidden layer dimension is 128, the activation function uses ReLU activation function, and the output layer is the mean and standard deviation of Gaussian distribution; the initial policy is initialized based on historical data;

[0082] S345: Collect trajectory data by interaction sampling, run the early warning system based on the current early warning threshold, and use the current policy π θ Interact with the environment, continue for T steps, and collect trajectory data:

[0083] τ={(s0,a0,r0),(s1,a1,r1),...,(s T ,a T ,r T )}

[0084] After each step, calculate the discounted cumulative return where γ is the discount factor, emphasizing recent rewards;

[0085] S346: Advantage function estimation, use generalized advantage estimation to calculate the advantage function

[0086] δ t =r t +γV(s t+1 )-V(s t )

[0087] where V(s t ) is the value network, which estimates the state value; λ is a control parameter that controls the balance between bias and variance; the advantage function reflects the "additional income of using action a t compared to the average policy";

[0088] S347: Policy optimization update, shuffle the trajectory data and divide it into mini-batches, and repeatedly update the policy network. In each iteration update, calculate the ratio of new and old policies r t (θ) and the clipping target L CLIP (θ), minimize the negative target function by using the Adam optimizer with a learning rate of 3e -4 , and maximize the reward; then, update the value network in synchronization, using MSE(V(s t ), G t ) as the loss function to ensure that the value estimation is consistent with the actual return;

[0089] S348: Convergence judgment and strategy deployment, when the average reward fluctuation of 5 consecutive iterations is less than 5%, and the early warning accuracy on the validation set is stable at more than 90%, stop iteration; deploy the final strategy network to the early warning system to output dynamic threshold in real time, and re-sample data every 24 hours for fine-tuning.

[0090] Further, the step S4 comprises:

[0091] S41: Device scheduling model design, a device scheduling algorithm based on deep Q network (DQN) is constructed to optimize the cruise path of the unmanned aerial vehicle and the AUV, and the specific steps comprise:

[0092] S411: State space design, define s=(c, p, e), wherein c is the current pollutant concentration distribution, p is the device position, and e is the remaining power;

[0093] S412: Action space design, the actions of the unmanned aerial vehicle and the AUV include "going forward", "going backward", "turning left", "turning right", "going up", and "going down" in six directions, including forward, backward, upward, and downward distances which can be determined by the user according to actual needs, and the turning angle can also be determined by the user according to actual needs;

[0094] S413: Reward function design, the reward function is set as:

[0095] r=λ1Δc-λ2d-λ3e

[0096] Wherein, Δc is the area of the newly discovered pollution area, d is the moving distance, e is the power consumption, and λ1, λ2, and λ3 are weight coefficients;

[0097] S414: Model training, an experience replay mechanism is used, the upper limit of training iteration is set to 5000, and the training is stopped when the pollution area coverage rate of the unmanned aerial vehicle is improved to 95%;

[0098] S42: Sensor fault self-diagnosis and repair, a sensor fault detection algorithm based on an autoencoder is designed to realize real-time fault identification and automatic repair, and the specific steps comprise:

[0099] S421: Autoencoder architecture design, the autoencoder adopts a symmetrical encoder-decoder structure, and is designed for the characteristics of sensor time series data, and the specific hierarchical structure comprises:

[0100] Input layer, the input of the input layer is the monitoring data in the continuous historical time period of the sensor, which is sampled according to the fixed sampling frequency of f minutes / time according to the user's demand, forming an n-dimensional input vector;

[0101] The encoder architecture design includes two hidden layers. The first hidden layer adopts a full connection layer, and the activation function is set to ReLU activation function, which is used to extract the basic time sequence characteristics of the data. The weight is initialized by He normal distribution, and the bias term is initialized to 0.1. The second hidden layer is a bottleneck layer, which also adopts a full connection layer, and the activation function is set to LeakyReLU activation function with a slope of 0.01.

[0102] The decoder architecture design includes one hidden layer and one output layer, both of which adopt a full connection layer architecture. The first hidden layer is symmetrical to the first hidden layer of the encoder, and the activation function is ReLU activation function, which is used to gradually restore the feature dimension. The output layer has the same dimension as the input layer, and the activation function is Linear linear activation function, which outputs the reconstructed sensor data vector.

[0103] S422: Autoencoder architecture training, using mean square error (MSE) loss function, the calculation expression is:

[0104]

[0105] Where x is the original input, i is the reconstructed output of the autoencoder network;

[0106] The optimizer is Adam optimizer, the learning rate is set to 0.001, the autoencoder is trained using normal sensor fault-free historical sample data, the upper limit of the iteration optimization round is set to 50, and the batch data size is 32. When the validation set MSE decreases by less than 1e -6 for 5 consecutive iteration optimization rounds, stop training;

[0107] S423: Calculate reconstruction error Where x is the n-dimensional data vector collected by the sensor in real time; is the reconstructed output vector of the autoencoder for x; x i is the i-th element of x, i.e. the sensor sampling value at a certain time; is the i-th element of , i.e. the corresponding sampling value at a certain time; e is the reconstruction error, which reflects the deviation of real-time data from the normal mode;

[0108] S424: Threshold τ definition and calculation, based on the reconstruction error distribution of normal data to determine the threshold τ, the specific steps are:

[0109] Collect k pieces of sensor data under fault-free state, covering different weather and meteorological conditions;

[0110] According to the reconstruction error calculation method given in step S423, calculate the reconstruction error e k ​(k = 1, 2, …, 1000)

[0111] Threshold τ = μ e + 3σ e , where μ e is the mean of normal data error, and σ e is the standard deviation.

[0112] S425: Fault determination, real-time calculation of reconstruction error e of sensor data; if e > τ, determine that the sensor is faulty; if e ≤ τ, determine that the data is normal, and directly used for subsequent analysis.

[0113] S426: Data repair, for the data of the faulty sensor, use K-Nearest Neighbor (KNN) interpolation method to repair, select the weighted average value of the synchronous data of the five nearest normal sensors as the replacement value of the faulty abnormal data, and the calculation expression is:

[0114]

[0115] wherein d i and d j are the spatial Euclidean distances between the faulty sensor and the normal sensor, which are obtained by the latitude and longitude coordinates according to the spherical distance formula of the earth.

[0116] In one preferred embodiment, the current pollution concentration distribution described in step S411 considers a 100x100 grid.

[0117] In another aspect, an intelligent ecological environmental protection monitoring system based on artificial intelligence is provided, which is applied to any one of the intelligent ecological environmental protection monitoring methods based on artificial intelligence. The intelligent ecological environmental protection monitoring system based on artificial intelligence comprises:

[0118] System architecture, the system architecture adopts an "edge-cloud" collaborative architecture, including an edge layer, a cloud end, and a communication layer; the edge layer is deployed on sensor nodes and unmanned aerial vehicles, runs lightweight intelligent algorithms, and reduces data transmission volume; the cloud end is deployed on a server cluster, runs complex algorithms, and processes massive data through a distributed computing framework (Spark); the communication layer adopts a combination of 5G mobile communication and satellite communication, 5G covers urban areas, and satellite communication covers remote mountainous areas and oceans.

[0119] Data acquisition module, the data acquisition module is a development device driver interface, supports 15 kinds of sensor protocols, and is used for realizing automatic access and analysis of data.

[0120] Intelligent analysis module, the intelligent analysis module integrates pollution source identification, traceability, and prediction algorithms, and provides visual analysis results.

[0121] An adaptive control module automatically generates device scheduling instructions according to the analysis results, and sends the device scheduling instructions to the unmanned aerial vehicle and the sensor device through an API interface;

[0122] A user interaction module develops a Web and mobile application, and supports environmental data query, early warning information receiving, and manual intervention operation functions.

[0123] In a preferred embodiment, the lightweight intelligent algorithm of the system architecture includes a pollution source identification algorithm and a data preprocessing algorithm, and the complex algorithm includes a multi-modal fusion algorithm and a prediction model.

[0124] In a preferred embodiment, the visualization analysis result of the intelligent analysis module includes a pollution diffusion thermodynamic diagram and a warning timeline.

[0125] Compared with the prior art, the present application has the following beneficial effects:

[0126] 1. The present application breaks through the limitations of single technology coverage and data island problems in traditional monitoring through multi-source data collaborative collection and intelligent fusion mechanism, realizes full-dimensional stereoscopic perception, and theoretically solves the inherent contradiction between monitoring range and precision.

[0127] 2. The present application realizes adaptive decision and resource scheduling by means of reinforcement learning, and gets rid of the dependence on artificial experience, and theoretically constructs a dynamic response intelligent monitoring closed loop, and improves the autonomous optimization ability of the system in a complex environment. BRIEF DESCRIPTION OF DRAWINGS

[0128] Figure 1 A flowchart of an intelligent ecological environmental protection monitoring method based on artificial intelligence is provided.

[0129] Figure 2 A system structure diagram of an intelligent ecological environmental protection monitoring system based on artificial intelligence is provided. DETAILED DESCRIPTION

[0130] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0131] In the description of the present application, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined with "first", "second" can explicitly or implicitly include one or more of the features. In the description of the present application, the meaning of "multiple" is two or more, unless otherwise explicitly specified.

[0132] In the description of the present application, the term "for example" is used to indicate "as an example, illustration or description". Any embodiment described as "for example" in the present application is not necessarily interpreted as more preferred or more advantageous than other embodiments. The following description is given in order to enable any person skilled in the art to implement and use the present application. In the following description, details are listed for the purpose of explanation. It should be understood that those skilled in the art can recognize that the present application can be implemented without using these specific details. In other examples, well-known structures and processes will not be described in detail to avoid unnecessary details making the description of the present application obscure. Therefore, the present application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope in accordance with the principles and characteristics disclosed in the present application.

[0133] In an embodiment of the present application, a smart ecological environmental protection monitoring method based on artificial intelligence can specifically refer to the accompanying Figure 1 , applied to a smart ecological environmental protection monitoring system based on artificial intelligence, specifically comprising:

[0134] S1: Construct a multi-source data acquisition network, integrate satellite remote sensing equipment, multiple types of unmanned aerial vehicles (carrying hyperspectral and thermal infrared cameras), ground sensor arrays (atmospheric micro stations, water quality monitoring stations), and underwater vehicles; synchronize time through the Beidou timing system, use the UTM projection coordinate system spatial reference, introduce the Q-Learning algorithm to dynamically adjust the sampling and cruising frequency, and realize multi-dimensional data acquisition;

[0135] S2: Data preprocessing and fusion, use the improved Z-Score algorithm to identify and eliminate outliers, use the FLAASH model to perform radiation and geometric correction on remote sensing images, and process heterogeneous data through min-max standardization; design a cross-modal fusion network containing a coordinate attention mechanism, extract features of each type of data, and use attention weight coefficients to generate a unified environmental state vector;

[0136] S3: Design an intelligent analysis algorithm, identify pollution sources based on the improved YOLOv8 model; construct a regional topology graph, trace the source of pollution in combination with the graph attention network, and solve the diffusion equation to simulate the diffusion path through the finite difference method; use the Attention-LSTM model to predict the concentration of pollutants, and dynamically adjust the warning threshold in combination with the PPO algorithm;

[0137] S4: Optimize adaptive monitoring strategy, construct device scheduling model based on deep Q network, take grid unit state, device position and power as input, generate cruise path of unmanned aerial vehicle and underwater vehicle; detect sensor failure through self-encoder, repair missing data by K nearest neighbor interpolation method, maintain data continuity.

[0138] Referring to the drawings Figure 2 The application also provides an intelligent ecological environmental protection monitoring system based on artificial intelligence, which comprises a system architecture, a data acquisition module, an intelligent analysis module, an adaptive control module and a user interaction module.

[0139] The system architecture adopts an edge-cloud collaborative architecture and comprises an edge layer, a cloud and a communication layer; the edge layer is arranged on the sensor nodes and the unmanned aerial vehicle, runs a lightweight intelligent algorithm and reduces data transmission amount; the cloud is arranged on a server cluster, runs a complex algorithm and processes massive data through a distributed computing framework (Spark); the communication layer adopts a combination of 5G mobile communication and satellite communication, 5G covers urban areas and satellite communication covers remote mountainous areas and oceans.

[0140] The data acquisition module is a device driver interface developed to support 15 kinds of sensor protocols and is used for realizing automatic access and analysis of data;

[0141] The intelligent analysis module integrates pollution source identification, traceability and prediction algorithms and provides visual analysis results;

[0142] The adaptive control module automatically generates device scheduling instructions according to analysis results and sends the instructions to the unmanned aerial vehicle and the sensor device through an API interface;

[0143] The user interaction module develops a Web and mobile terminal application and supports environmental data query, early warning information receiving and manual intervention operation functions.

[0144] In a specific embodiment, to solve the problems of insufficient coverage, poor collaboration, poor data processing capability, low intelligent level, weak early warning capability, and insufficient autonomy and adaptability in existing ecological environmental protection monitoring methods, an ecological environmental protection monitoring personnel proposes an intelligent ecological environmental protection monitoring method based on artificial intelligence. First, the ecological environmental protection monitoring personnel constructs a multi-source data acquisition network, integrates satellite remote sensing equipment, multiple types of unmanned aerial vehicles, ground sensor arrays, and underwater vehicles, synchronizes time through the Beidou timing system, uses the UTM projection coordinate system as a spatial reference, and introduces the Q-Learning algorithm to dynamically adjust the sampling and cruising frequency, achieving multi-dimensional data acquisition. Second, the ecological environmental protection monitoring personnel identifies and removes outliers by improving the Z-Score algorithm, performs radiation and geometric correction using the FLAASH model, and processes heterogeneous data through min-max standardization to achieve data preprocessing. Meanwhile, a cross-modal fusion network with a coordinate attention mechanism is designed to generate a unified environmental state vector, achieving multi-source data fusion. Third, the ecological environmental protection monitoring personnel designs an intelligent analysis algorithm, identifies pollution sources based on an improved YOLOv8 model, traces the source of pollution using a graph attention network, simulates the diffusion path by solving the diffusion equation using the finite difference method, predicts the concentration of pollutants using an Attention-LSTM model, and dynamically adjusts the early warning threshold using the PPO algorithm. Finally, the ecological environmental protection monitoring personnel optimizes the adaptive monitoring strategy, constructs a device scheduling model based on a deep Q network, takes the state of the grid cell, the location of the device, and the power as input, generates the cruising path of the unmanned aerial vehicle and underwater vehicle, detects sensor failures using an autoencoder, repairs missing data using the K-Nearest Neighbor interpolation method, and maintains data continuity.

[0145] Embodiment 1

[0146] In one embodiment, to solve the problems of insufficient coverage, poor collaboration, and single data source in existing ecological environmental protection monitoring methods, an ecological environmental protection monitoring personnel constructs a multi-source data acquisition network to achieve multi-dimensional data acquisition. The specific steps include:

[0147] S11: Satellite remote sensing equipment deployment and parameter configuration, select Sentinel-2 and Gaofen-6 satellites, set the Sentinel-2 satellite to 10-meter spatial resolution and 5-day revisit period, set the Gaofen-6 satellite to 8-meter spatial resolution and 4-day revisit period, collect multi-spectral data with a wavelength range of 400-2500 nm, and focus on monitoring macro indicators such as vegetation coverage and water area; simultaneously access atmospheric pollutant data from the Sentinel-5P satellite, set the satellite resolution to 3.5 kilometers, and achieve regional-scale pollution distribution monitoring;

[0148] S12: Deployment and parameter configuration of various types of drones. Different models of quadcopter drones are deployed in ecologically sensitive areas, equipped with a 5nm bandwidth hyperspectral camera and a 640×512 resolution thermal infrared camera. They cruise once every 2 hours to collect surface reflectance and temperature data with a resolution of 1 meter, capturing micro-scale ecological changes.

[0149] S13: Deployment and parameter configuration of underwater vehicles. Deploy autonomous underwater vehicles (AUVs) in the seabed, rivers, lakes and reservoirs, equipped with acoustic Doppler current profilers (ADCP) and water quality sensors to collect water depth, current velocity and chlorophyll a concentration data in real time, with a cruising range covering more than 90% of the water area.

[0150] S14: Design of a spatiotemporal synchronization mechanism to establish a unified spatiotemporal benchmark and solve the problem of time and space matching of multi-source data. Specific steps include:

[0151] S141: Time synchronization, all devices use BeiDou time synchronization, and the time error is controlled within 10ms; timestamp correction is performed on satellite remote sensing data to make the imaging time accurate to the second level;

[0152] S142: Spatial registration, using the UTM projection coordinate system, uniformly transforms UAV and ground sensor data to this coordinate system; remote sensing images are corrected through ground control points (GCP) to make the spatial deviation of data from different sources less than 1 meter;

[0153] S143: Dynamic sampling strategy design, based on reinforcement learning algorithm (Q-Learning) to dynamically adjust the sampling frequency. When an anomaly occurs in the monitoring area, the drone's cruise frequency and sensor sampling frequency are automatically increased. In non-anomaly areas, the normal frequency is maintained to save energy.

[0154] Example 2

[0155] In one embodiment, to address the problem of poor data processing capabilities in existing ecological and environmental monitoring methods, ecological and environmental monitoring personnel preprocessed and fused multi-source data. Specific steps included:

[0156] S21: Outlier removal. An improved Z-Score algorithm is used to identify outlier data. The calculation expression is:

[0157]

[0158] Where, x i For the original data, μ is the mean and σ is the standard deviation; when |Z i When | > 3.29, corresponding to a 99.9% confidence interval, it is determined to be an outlier, and the gap is filled by linear interpolation.

[0159] S22: Remote sensing data correction, satellite and unmanned aerial vehicle image radiation correction, eliminate the influence of atmospheric scattering, and several corrections are made to satellite and unmanned aerial vehicle image. FLAASH model is used to calculate the ground reflectivity, and the calculation expression is:

[0160]

[0161] Where, p is the ground reflectivity, L is the apparent radiation brightness, d is the distance from the earth to the sun, E0 is the solar radiation at the top of the atmosphere, and θ is the solar zenith angle;

[0162] S23: Data standardization, mapping different dimensional data to the interval [0, 1], using min-max standardization formula for data standardization, the calculation expression is:

[0163]

[0164] Where, x min , x max are the minimum and maximum values in the data set, respectively;

[0165] S24: Design of cross-modal fusion network (CMFN) based on attention mechanism, realize the deep fusion of heterogeneous data, the specific steps include:

[0166] S241: Feature extraction layer design, for remote sensing image, ResNet-50 network is used to extract spatial features, output 512-dimensional vector;

[0167] For sensor data, 1D convolution layer with size 3 is used to extract time series features of sensor data, output 256-dimensional vector;

[0168] For spectral data, PCA dimension reduction is used to retain 95% variance, and then full connection layer is used to convert to 256-dimensional feature vector;

[0169] S242: Attention fusion layer design, calculate the weight coefficient a i of different modal features, the calculation expression is:

[0170]

[0171] Where, f i is the feature vector of the i-th modal, s(·) is the attention score function, which is composed of a single hidden layer neural network; the fusion feature vector is obtained by weighted summation, and the calculation expression is:

[0172]

[0173] S243: Output layer design, input fusion features into a fully connected layer, output a unified environmental state vector including 128-dimensional features covering atmospheric, water quality, and vegetation multi-dimensional information.

[0174] Embodiment 3

[0175] In one embodiment, to solve the problems of low intelligent level and weak early warning capability existing in the existing ecological environmental protection monitoring method, the ecological environmental protection monitoring personnel designs an intelligent analysis algorithm, and the specific steps include:

[0176] S31: Pollution source intelligent recognition model design, build a pollution source recognition algorithm based on improved YOLOv8, realize automatic positioning of industrial chimneys and sewage discharge port pollution sources, and the specific steps include:

[0177] S311: Data set construction, collect 100,000 remote sensing images and unmanned aerial images containing various pollution sources, label the target category and boundary box, and divide the training set, validation set and test set according to the ratio of 8:1:1;

[0178] S312: Model design and improvement, add a coordinate attention module between the output end of the backbone architecture C3 module of YOLOv8 and the down-sampling convolution layer to enhance the recognition ability of small targets, and the calculation expression of the coordinate attention module is:

[0179] CA(F) = Conv(σ(MLP(AvgPool(F))))

[0180] Wherein, F is the input feature map, AvgPool is the average pooling, MLP is the multi-layer perception, σ is the Sigmoid activation function, Conv is the convolution operation; the C3 module outputs a feature map F, the coordinate attention module performs global average pooling along the x-axis and y-axis on the output feature map F respectively to obtain two 1D feature vectors; then the two 1D feature vectors are spliced, and then compressed through 1x1 convolution, and then BatchNorm and ReLU activation; then the spliced vector after activation is split into two independent branches, which are restored through 1x1 convolution, and then the x-axis and y-axis attention weights are obtained through Sigmoid activation; finally, the weights are multiplied by the original feature map F to output the enhanced feature map F';

[0181] S313: Model training, place the YOLOv8 model designed and improved in step S312 on the training data set and validation data set constructed in step S311 to perform model training, use the SGD optimizer, set the learning rate to 0.01, the momentum to 0.9, and the upper limit of the training generation number to 50, test the model performance on the test set after the model training is completed, and select the model with good test performance for deployment;

[0182] S32: Construct a pollution source tracing model combined with a graph neural network (GNN), divide the monitoring area into 100x100 meter grid units, construct a regional topology graph G=(V, E), where V is the grid node, E is the connection relationship between nodes, and learn the node features using a graph attention network (GAT):

[0183]

[0184] where is the feature vector of node i, N(i) is the neighbor node set, is the attention coefficient, which is calculated by the self-attention mechanism, W is the weight matrix; Sort the node features by anomaly degree, and select the node with the highest anomaly degree as the most likely pollution source;

[0185] S33: Diffusion simulation, based on the pollutant diffusion equation, combined with real-time meteorological and hydrological data, predict the pollutant diffusion range, the prediction calculation expression is:

[0186]

[0187] where c is the pollutant concentration, t is the time, u and v are the flow velocities in x and y directions respectively, and D is the diffusion coefficient;

[0188] S34: Adaptive prediction and early warning model construction, design an LSTM model based on time series attention mechanism (Attention-LSTM) for pollutant concentration prediction and dynamic early warning, the specific steps include:

[0189] S341: Prediction model design, the model input is the environmental state vector in the past n hours, the sample interval can be adjusted according to the actual situation, the model output is the pollutant concentration prediction value in the future m hours, the model structure includes:

[0190] LSTm layer, set 256 hidden units to capture time series features and calculate hidden states:

[0191] h t =LSTM(x t ,h t-1 )

[0192] where x t is the input vector at time t;

[0193] Attention layer, calculate the weight of hidden state at each time Output weighted hidden state

[0194] Output layer, output the prediction value through the fully connected layer

[0195] S342: Model training, using Adam optimizer, learning rate set to 0.001, training iteration set to 100, loss function using root mean square error loss (RMSE), calculation expression is:

[0196]

[0197] Wherein, is the sample number, that is, the total number of pollutant concentration prediction samples participating in the calculation; y i is the actual observation value of the i-th sample, that is, the true pollutant concentration value; is the prediction value of the i-th sample, that is, the pollutant concentration prediction result output by the model;

[0198] S343: Dynamic early warning mechanism design, using proximal policy optimization (PPO) reinforcement learning algorithm to dynamically adjust the early warning threshold, defining state variables s t , action a t , reward r t respectively;

[0199] State variables need to be converted into feature vectors that can be processed by neural networks, wherein, c t is the current pollutant concentration, and the average concentration value in the last n hours is taken according to actual needs; m t is the weather condition, including wind speed, humidity, air pressure, temperature, and light intensity dimension data indicators; is the season, using one-hot encoding form;

[0200] Action a t is the adjusted early warning threshold;

[0201] Reward function r t calculation expression is:

[0202] r t = r correct -λ false r false -λ miss r miss

[0203] Wherein, r correct is correct warning, that is, prediction exceeds and actual exceeds, 10 points / time; r false is false alarm, that is, prediction exceeds but actual does not exceed, minus 5 points / time; r miss is a missed call, that is, actual exceeds but not predicted, minus 20 points / time; λ false and λ miss are penalty weights for false alarm and missed call respectively, and the size relationship between the two can be set according to actual needs;

[0204] The core objective function selects the clipping substitute objective function, and the calculation expression is:

[0205]

[0206] Wherein, is the probability ratio of the new and old strategies; is the advantage function estimate, which is used to measure the advantage of action a t over the average level; ò is the clipping coefficient, which is used to limit r t (θ) in the range of to avoid the policy update being too large;

[0207] S344: Initialize the policy network, use a 2-layer fully connected neural network to construct the policy network of the PPO reinforcement learning algorithm, the hidden layer dimension is 128, the activation function uses the ReLU activation function, and the output layer is the mean and standard deviation of the Gaussian distribution; The initial policy is initialized based on historical data;

[0208] S345: Interact with the sampling to collect trajectory data, run the early warning system based on the current early warning threshold, and use the current policy π θ to interact with the environment for T steps to collect trajectory data:

[0209] τ={(s0,a0,r0),(s1,a1,r1),...,(s T ,a T ,r T )}

[0210] After each step, calculate the discounted cumulative return where γ is the discount factor, emphasizing recent rewards;

[0211] S346: Advantage function estimation, use generalized advantage estimation to calculate the advantage function

[0212] δ t =r t +γV(s t+1 )-V(s t )

[0213] Wherein, V(s t ) is the value network, which estimates the state value; λ is a control parameter that controls the balance between bias and variance; The advantage function reflects the "additional income of using action a t compared to the average strategy";

[0214] S347: Policy optimization update, shuffle the trajectory data and divide it into mini-batches, and repeatedly update the policy network. In each iteration update, calculate the ratio of the new and old strategies r t(θ) and the clipping target L CLIP (θ) by an Adam optimizer with a learning rate of 3e -4 -1maximize the reward by minimizing the negative target function with an Adam optimizer with a learning rate of 3e t -1maximize the reward by minimizing the negative target function with an Adam optimizer with a learning rate of 3e t -1maximize the reward by minimizing the negative target function with an Adam optimizer with a learning rate of 3e

[0215] S348: Convergence judgment and strategy deployment, when the average reward fluctuation of 5 consecutive iterations is less than 5%, and the early warning accuracy on the validation set is stable at more than 90%, stop iteration; deploy the final strategy network to the early warning system to output dynamic threshold in real time, and re-sample data every 24 hours for fine-tuning.

[0216] Embodiment 4

[0217] In one embodiment, to solve the problem of poor adaptability of existing ecological environmental protection monitoring methods, the ecological environmental protection monitoring personnel proposes an optimized adaptive monitoring strategy, the specific steps of which include:

[0218] S41: Equipment scheduling model design, build a device scheduling algorithm based on deep Q network (DQN), optimize the cruise path of the unmanned aerial vehicle and AUV, the specific steps of which include:

[0219] S411: State space design, define s=(c,p,e), where c is the current pollutant concentration distribution, p is the device position, and e is the remaining power;

[0220] S412: Action space design, the actions of the unmanned aerial vehicle and AUV include "forward", "backward", "left turn", "right turn", "up", "down" in 6 directions, including forward, backward, upward, and downward distance, which can be determined by the user according to actual needs, and the turning angle can also be determined by the user according to actual needs;

[0221] S413: Reward function design, the reward function is set as:

[0222] r=λ1Δc-λ2d-λ3e

[0223] Where Δc is the area of newly discovered pollution area, d is the moving distance, e is the power consumption, and λ1, λ2, and λ3 are weight coefficients;

[0224] S414: Model training, use the experience replay mechanism, set the upper limit of training iterations to 5000, and stop training when the pollution area coverage rate of the unmanned aerial vehicle improves to 95%;

[0225] S42: Sensor fault self-diagnosis and repair, design a sensor fault detection algorithm based on autoencoder, realize real-time fault identification and automatic repair, the specific steps include:

[0226] S421: Autoencoder architecture design, the autoencoder adopts a symmetrical encoder-decoder structure, and is designed for the characteristics of sensor time series data, and the specific hierarchical structure includes:

[0227] Input layer, the input of the input layer is the monitoring data of the sensor in the continuous historical time period, which is sampled according to the fixed sampling frequency of f minutes / time according to the user demand, forming an n-dimensional input vector;

[0228] Encoder architecture design, the encoder architecture contains two hidden layers, the first hidden layer adopts a full connection layer, and the activation function is set to ReLU activation function, which is used to extract the basic time series features of the data. The weight initialization adopts He normal distribution, and the bias term is initialized to 0.1; the second hidden layer is a bottleneck layer, which also adopts a full connection layer, and the activation function is set to LeakyReLU activation function with a slope of 0.01;

[0229] Decoder architecture design, the decoder includes a hidden layer and an output layer, both of which adopt full connection layer architecture; the first hidden layer is symmetrical with the first hidden layer of the encoder, and the activation function is ReLU activation function, which is used to gradually restore the feature dimension; the output layer is consistent with the input layer in dimension, and the activation function is Linear linear activation function, which outputs the reconstructed sensor data vector;

[0230] S422: Autoencoder architecture training, using mean square error (MSE) loss function, the calculation expression is:

[0231]

[0232] Where, x i is the original input, is the autoencoder network reconstruction output;

[0233] Optimizer Adam optimizer, learning rate set to 0.001, use normal sensor fault-free historical sample data to train autoencoder, the upper limit of iteration optimization rounds is set to 50, and the batch data size is 32, when the validation set MSE decreases by less than 1e -6 Stop training when 5 consecutive iteration optimization rounds;

[0234] S423: Calculate reconstruction error Where, x is the n-dimensional data vector collected by the sensor in real time; is the reconstruction output vector of the autoencoder for x; x i is the i-th element of x, that is, the sensor sampling value at a certain time; is the ith element of the reconstructed corresponding time sampling value; e is the reconstruction error, reflecting the deviation of real-time data from the normal mode;

[0235] S424: Threshold τ definition and calculation, based on the reconstruction error distribution of normal data to determine the threshold τ, the specific steps are:

[0236] Collect k pieces of sensor data under normal fault-free state, covering different weather, meteorological conditions;

[0237] According to the reconstruction error calculation method given in step S423, the reconstruction error e of each piece of sensor data is calculated k (k = 1, 2,..., 1000);

[0238] The threshold τ = μ e + 3σ is calculated by using the 3σ criterion e , wherein μ e is the mean of the normal data error, and σ e is the standard deviation;

[0239] S425: Fault determination, real-time calculation of the reconstruction error e of the sensor data; if e > τ, it is determined that the sensor is faulty; if e ≤ τ, it is determined that the data is normal, and is directly used for subsequent analysis;

[0240] S426: Data repair, for the data of the faulty sensor, the K nearest neighbor (KNN) interpolation method is used for repair, and the weighted average value of the same period data of the five nearest normal sensors is calculated as the replacement value of the faulty abnormal data, and the calculation expression is:

[0241]

[0242] , wherein d i and d j are the spatial Euclidean distances between the faulty sensor and the normal sensor, which are obtained by the latitude and longitude coordinates according to the earth spherical distance formula.

[0243] Although the preferred embodiments of the present application have been described, those skilled in the art can make further changes and modifications to these embodiments once they know the basic inventive concept. Therefore, the appended claims are intended to be interpreted as including all changes and modifications falling within the scope of the present application.

[0244] Obviously, those skilled in the art can make various modifications and variations to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. An intelligent ecological environmental protection monitoring method based on artificial intelligence, characterized in that, Comprising the following steps: S1: Constructing a multi-source data acquisition network, integrating satellite remote sensing equipment, multiple types of unmanned aerial vehicles, ground sensor arrays, and underwater vehicles; synchronizing time through the Beidou timing system, using the UTM projection coordinate system as a spatial reference, introducing the Q-Learning algorithm to dynamically adjust the sampling and cruising frequency, and achieving multi-dimensional data acquisition; S2: Data preprocessing and fusion, using the improved Z-Score algorithm to identify and eliminate outliers, using the FLAASH model for radiometric and geometric correction of remote sensing images, and processing heterogeneous data through min-max standardization; designing a cross-modal fusion network with coordinate attention mechanism, extracting features from each type of data, and generating a unified environmental state vector using attention weight coefficients; S3: Designing intelligent analysis algorithms, identifying pollution sources based on the improved YOLOv8 model, constructing regional topological graphs, tracing the source of pollution using graph attention networks, and simulating diffusion paths by solving diffusion equations using the finite difference method; using the Attention-LSTM model to predict pollutant concentrations and dynamically adjusting the warning threshold using the PPO algorithm; S4: Optimizing adaptive monitoring strategies, constructing a device scheduling model based on deep Q networks, using grid cell state, device location, and power as inputs to generate unmanned aerial vehicle and underwater vehicle cruising paths; detecting sensor failures using autoencoders and repairing missing data using K-nearest neighbor interpolation to maintain data continuity. 2.The intelligent ecological environmental protection monitoring method based on artificial intelligence according to claim 1, characterized in that, The step S1 comprises: S11: Satellite remote sensing equipment deployment and parameter configuration, selecting Sentinel-2 and Gaofen-6 satellites, setting the Sentinel-2 satellite to 10-meter spatial resolution with a 5-day revisit period, and the Gaofen-6 satellite to 8-meter spatial resolution with a 4-day revisit period, collecting multi-spectral data in the 400-2500nm band range, focusing on monitoring macro indicators such as vegetation coverage and water area; simultaneously accessing Sentinel-5P satellite atmospheric pollutant data with a satellite resolution of 3.5 kilometers to achieve regional-scale pollution distribution monitoring; S12: Multi-type unmanned aerial vehicle deployment and parameter configuration, deploying different types of quadcopters in ecologically sensitive areas, equipped with 5nm band width hyperspectral cameras and 640x512 resolution thermal infrared cameras, cruising every 2 hours to collect 1-meter resolution ground reflectance and temperature data, capturing micro-scale ecological changes; S13: Underwater vehicle deployment and parameter configuration, deploying autonomous underwater vehicles in sea, river, lake, and reservoir waters, equipped with acoustic Doppler current profilers and water quality sensors, collecting real-time water depth, flow rate, and chlorophyll-a concentration data, with a cruising range covering over 90% of the water area; S14: Designing a time and space synchronization mechanism, establishing a unified time and space reference to solve the time and space matching problem of multi-source data. 3.The intelligent ecological environmental protection monitoring method based on artificial intelligence according to claim 2, characterized in that, The step S14 comprises: S141: Time synchronization, all devices use Beidou timing, with a time error controlled within 10ms; timestamp correction is performed on satellite remote sensing data, with imaging time accurate to the second level; S142: Spatial registration, convert the data of unmanned aerial vehicle and ground sensor to the UTM projection coordinate system; correct the remote sensing image through ground control points to make the spatial deviation of data from different sources less than 1 meter; S143: Dynamic sampling strategy design, based on reinforcement learning algorithm to dynamically adjust the sampling frequency, when the monitoring area appears abnormal, automatically improve the unmanned aerial vehicle cruising frequency and sensor sampling frequency, and keep the normal frequency in the non-abnormal area to save energy consumption.

4. The intelligent ecological environmental protection monitoring method based on artificial intelligence according to claim 1, characterized in that, The step S2 comprises: S21: Outlier rejection, using the improved Z-Score algorithm to identify abnormal data, the calculation expression is: Where x i is the original data, μ is the mean, and σ is the standard deviation; when |Z i | > 3.29, corresponding to a 99.9% confidence interval, it is determined to be an outlier, and the gap is filled by linear interpolation. S22: Remote sensing data correction, radiation correction is performed on satellite and unmanned aerial vehicle images to eliminate the influence of atmospheric scattering, and several corrections are performed on satellite and unmanned aerial vehicle images, FLAASH model is used to calculate the ground reflectivity, the calculation expression is: Wherein, ρ is the ground reflectivity, L is the apparent radiation brightness, d is the distance from the earth to the sun, E0 is the solar irradiance at the top of the atmosphere, θ is the solar zenith angle; S23: Data standardization, mapping data of different dimensions to the interval [0, 1], using the min-max standardization formula to standardize the data, the calculation expression is: where x min , x max are the minimum and maximum values in the dataset, respectively; S24: Design of cross-modal fusion network based on attention mechanism, realize the deep fusion of heterogeneous data.

5. The intelligent ecological environmental protection monitoring method based on artificial intelligence according to claim 4, characterized in that, The step S24 comprises: S241: Feature extraction layer design, for remote sensing image, using ResNet-50 network to extract spatial features, outputting 512-dimensional vector; For sensor data, 1D convolution layer with size of 3 is used to extract time series features of sensor data, outputting 256-dimensional vector; For spectral data, PCA dimension reduction is used to retain 95% variance, and then full connection layer is used to convert to 256-dimensional feature vector; S242: Attention fusion layer design, calculate the weight coefficient a of different modal features i The calculation expression is: where f i is the feature vector of the i-th modality, s(·) is an attention score function, which is composed of a single hidden layer neural network; the fusion feature vector is obtained by weighted summation, and the calculation expression is: S243: Output layer design, input the fusion features into the full connection layer, output the unified environmental state vector, including 128-dimensional features, covering atmospheric, water quality, vegetation multidimensional information.

6. The intelligent ecological environmental protection monitoring method based on artificial intelligence according to claim 1, characterized in that, The step S3 comprises: S31: Pollution source intelligent identification model design, build pollution source identification algorithm based on improved YOLOv8, realize automatic positioning of industrial chimney and sewage discharge pollution source; S32: Combine graph neural network to build pollution tracing model, divide the monitoring area into 100*100 meter grid units, build regional topology graph G=(V,E), where V is the grid node, E is the connection relationship between nodes, use graph attention network to learn node features: Wherein, is the feature vector of node i, N(i) is the neighbor node set, is the attention coefficient, which is calculated by self-attention mechanism, W is the weight matrix; select the node with the highest abnormal degree as the most possible pollution source through the abnormal degree sorting of node features; S33: Diffusion simulation, based on the pollutant diffusion equation, combined with real-time meteorological and hydrological data, predict the diffusion range of pollutants, the prediction calculation expression is: Wherein, c is the concentration of pollutants, t is the time, u and v are the flow velocities in x and y directions respectively, D is the diffusion coefficient; S34: Adaptive prediction and early warning model construction, design LSTM model based on time series attention mechanism for pollutant concentration prediction and dynamic early warning.

7. The intelligent ecological environmental protection monitoring method based on artificial intelligence according to claim 6, characterized in that, The step S31 comprises: S311: Data set construction, 100,000 remote sensing images and unmanned aerial images containing various types of pollution sources are collected, target categories and boundary boxes are labeled, and training set, validation set and test set are divided according to the ratio of 8:1:1; S312: Model design and improvement, a coordinate attention module is added between the output end of the backbone architecture C3 module of YOLOv8 and the down-sampling convolution layer, which enhances the recognition ability of small targets, and the calculation expression of the coordinate attention module is: CA(F) = Conv(σ(MLP(AvgPool(F)))) Wherein, F is the input feature map, AvgPool is the average pooling, MLP is the multi-layer perception, σ is the Sigmoid activation function, and Conv is the convolution operation; the C3 module outputs a feature map F, the coordinate attention module performs global average pooling along the x-axis and y-axis on the output feature map F to obtain two 1D feature vectors; then the two 1D feature vectors are spliced, and then compressed through 1x1 convolution, and then BatchNorm and ReLU activation; then the spliced vector after activation is split into two independent branches, which are restored through 1x1 convolution, and then the x-axis and y-axis attention weights are obtained through Sigmoid activation; finally, the weights are multiplied with the original feature map F, and the enhanced feature map F' is output; S313: Model training, the YOLOv8 model designed and improved in step S312 is placed on the training data set and validation data set constructed in step S311 for model training, the SGD optimizer is adopted, the learning rate is set to 0.01, the momentum is 0.9, the upper limit of training generation number is set to 50, the model performance is tested on the test set after the model training is completed, and the model with good test performance is selected for deployment; 8.The intelligent ecological environmental protection monitoring method based on artificial intelligence according to claim 6, characterized in that, The step S34 comprises: S341: Prediction model design, the model input is the environmental state vector of the past n hours, the sample interval can be adjusted according to the actual situation, the model output is the pollutant concentration prediction value of the future m hours, and the model structure comprises: LSTm layer, 256 hidden units are set to capture time sequence features and calculate hidden state: h t = LSTM(x t ,h t-1 ) where x t is the input vector at time t; attention layer, which computes the weight of the hidden state at each time output weighted hidden state output layer, which outputs a prediction value through a fully connected layer S342: Model training, the Adam optimizer is adopted, the learning rate is set to 0.001, the training generation number is set to 100, and the root mean square error loss (RMSE) is used as the loss function, and the calculation expression is: wherein N is the number of samples, i.e. the total number of pollutant concentration prediction samples participating in the calculation; y i yi is the actual observation value of the i-th sample, i.e. the true pollutant concentration value; pi is the prediction value of the i-th sample, i.e. the pollutant concentration prediction result output by the model; S343: Dynamic early warning mechanism design, using proximal policy optimization (PPO) reinforcement learning algorithm to dynamically adjust the early warning threshold, respectively define state variables s t , action a t , reward r t ; State variables Need to be converted into a neural network can handle the feature vector, where c t The current pollutant concentration, according to the actual need to take the average concentration value in the last n hours; m t The weather conditions include wind speed, humidity, air pressure, temperature, and light intensity dimension data indicators. Season, using one-hot encoding form; Action a t is the adjusted early warning threshold; Reward function r t The computational expression is: r t = r correct - λ false r false - λ miss r miss wherein, r correct 10 points / time for correct warning, i.e. predicting exceeding and actually exceeding; false 5 points / time for false alarm, i.e. predicting exceeding but not actually exceeding; miss 20 points / time for missing alarm, i.e. actually exceeding but not predicting; false and λ miss The penalty weights for false alarm and missing alarm are obviously different, and the size relationship between the two can be set according to actual needs. The core target function selects the clipping substitute target function, and the calculation expression is: where, is the probability ratio of the new and old policies; is the advantage function estimate, measuring the action a t relative to the mean level of advantage; ò is a clipping coefficient to limit r t (θ) is in the range to avoid excessively large policy updates; S344: Initialization strategy network, a 2-layer fully connected neural network is used to constitute the strategy network of PPO reinforcement learning algorithm, the hidden layer dimension is 128, the activation function adopts ReLU activation function, and the output layer is the mean and standard deviation of Gaussian distribution; the initial strategy is initialized based on historical data; S345: Interactively sample the trajectory data, run the warning system based on the current warning threshold, and use the current policy π θ Interact with the environment for T steps, collect trajectory data: τ = {(s0, a0, r0), (s1, a1, r1),..., (s T , a T , r T )} At the end of each step, the discounted cumulative return is computed where γ is a discount factor that emphasizes recent rewards; S346: Estimation of the log-rank statistic, using the generalized log-rank statistic delta t = r t + gamma * V(s t+1 ) - V(s t ) where V(s t ) is the value network that estimates the state value; λ is a control parameter that controls the balance between bias and variance; and the advantage function reflects the "extra reward of taking action a t over the average policy". S347: policy optimization update, shuffle the trajectory data and divide into mini-batches, iteratively update the policy network; in each iteration, calculate the ratio of new and old policy r t (θ) and the clipping target L CLIP (θ); minimize the negative target function and maximize the reward by an Adam optimizer with a learning rate of 3e -4 ; then, update the value network in sync, taking the MSE(V(s t ), G t ) as the loss function to ensure that the value estimate is consistent with the actual return S348: Convergence judgment and strategy deployment, when the average reward fluctuation of 5 consecutive iterations is less than 5%, and the early warning accuracy on the validation set is stable at more than 90%, stop iteration; deploy the final strategy network to the early warning system to output dynamic thresholds in real time, and re-sample data every 24 hours for fine-tuning. 9.The intelligent ecological environmental protection monitoring method based on artificial intelligence according to claim 1, characterized in that, The step S4 comprises: S41: Equipment scheduling model design, a deep Q network (DQN) based equipment scheduling algorithm is constructed to optimize the cruise path of the unmanned aerial vehicle and the AUV, and the specific steps comprise: S411: State space design, define s=(c, p, e), wherein c is the current pollutant concentration distribution, p is the equipment position, and e is the remaining power; S412: Action space design, the actions of the unmanned aerial vehicle and the AUV include "forward", "backward", "left turn", "right turn", "upward", "downward" six directions, including the distance of forward movement, backward movement, upward movement and downward movement which can be determined by the user according to actual needs, and the turning angle can also be determined by the user according to actual needs; S413: Reward function design, the reward function is set as: r=λ1Δc-λ2d-λ3e Wherein, Δc is the area of the newly discovered pollution area, d is the moving distance, e is the power consumption, and λ1, λ2, λ3 are weight coefficients; S414: Model training, an experience replay mechanism is used, and the upper limit of training iteration is set to 5000, and the training is stopped when the pollution area coverage rate of the unmanned aerial vehicle is improved to 95%; S42: Sensor fault self-diagnosis and repair, a sensor fault detection algorithm based on an autoencoder is designed to realize real-time fault identification and automatic repair, and the specific steps comprise: S421: Autoencoder architecture design, the autoencoder adopts a symmetrical encoder-decoder structure, and is designed according to the characteristics of sensor time series data, and the specific hierarchical structure comprises: Input layer, the input of the input layer is the monitoring data of the sensor in a continuous historical time period, which is sampled according to a fixed sampling frequency of f minutes / time according to user needs to form an n-dimensional input vector; Encoder architecture design, the encoder architecture includes two hidden layers, the first hidden layer adopts a full connection layer, and the activation function is set to ReLU activation function, which is used to extract the basic time sequence characteristics of the data. The weight is initialized by He normal distribution, and the bias term is initialized to 0.1; the second hidden layer is a bottleneck layer, which also adopts a full connection layer, and the activation function is set to LeakyReLU activation function with a slope of 0.01; Decoder architecture design, the decoder includes a hidden layer and an output layer, both of which adopt a full connection layer architecture; the first hidden layer is symmetrical to the first hidden layer of the encoder, and the activation function is ReLU activation function, which is used to gradually restore the feature dimension; the output layer is consistent with the input layer in dimension, and the activation function is Linear linear activation function, which outputs the reconstructed sensor data vector; S422: Autoencoder architecture training, a mean square error loss function is used, and the calculation expression is: where x i is the original input, is the autoencoder network reconstruction output; The optimizer selects the Adam optimizer, the learning rate is set to 0.001, the self-encoder is trained using the historical sample data of the normal sensor without failure, the upper limit of the iteration optimization round is set to 50, the batch data size is 32, and the training is stopped when the verification set MSE decreases by less than 1e -6 -4 for 5 consecutive iteration optimization rounds -6 S423: Calculate reconstruction error where x is the n-dimensional data vector collected by the sensor in real time; is the reconstruction output vector of the autoencoder for x; x i is the i-th element of x, i.e., the sensor sampling value at a certain time; is the i-th element of is the i-th element of the reconstructed sampling value at the corresponding time; e is the reconstruction error, reflecting the deviation of real-time data from the normal mode; S424: Threshold τ definition and calculation, the threshold τ is determined based on the reconstruction error distribution of normal data, and the specific steps are: Collect k pieces of sensor data in fault-free state, covering different weather, meteorological conditions; The reconstruction error e of each piece of sensor data is calculated according to the reconstruction error calculation method given in step S423 k (k = 1, 2,..., 1000); The threshold τ = μ + 3σ is calculated using the 3σ criterion e + 3σ e where μ e is the mean of the error of the normal data and σ e is the standard deviation; S425: Fault determination, real-time calculation of reconstruction error e of sensor data; if e>τ, determine that the sensor is faulty; if e≤τ, determine that the data is normal, and directly used for subsequent analysis; S426: Data repair, for the data of the faulty sensor, adopt K-neighbor interpolation method to repair, select 5 nearest normal sensors of the same period data, calculate the weighted average value as the replacement value of the abnormal data, the calculation expression is: where d i and d j are the spatial Euclidean distances between the faulty sensor and the normal sensors, obtained by the longitude and latitude coordinates according to the Earth's spherical distance formula.

10. An intelligent ecological environmental protection monitoring system based on artificial intelligence, characterized in that, The intelligent ecological environmental protection monitoring system based on artificial intelligence comprises the intelligent ecological environmental protection monitoring method based on artificial intelligence according to any one of claims 1-9. The system architecture adopts an edge-cloud collaborative architecture, including an edge layer, a cloud end and a communication layer; the edge layer is deployed on the sensor nodes and the unmanned aerial vehicle, runs a lightweight intelligent algorithm, and reduces the data transmission amount; the cloud end is deployed on a server cluster, runs a complex algorithm, and processes massive data through a distributed computing framework; the communication layer adopts a combination of 5G mobile communication and satellite communication, 5G covers urban areas, and satellite communication covers remote mountainous areas and oceans; The data acquisition module is a development device driver interface, supports 15 kinds of sensor protocols, and is used for realizing automatic access and analysis of data; The intelligent analysis module integrates pollution source identification, traceability and prediction algorithms, and provides visual analysis results; The adaptive control module automatically generates device scheduling instructions according to the analysis results, and sends the instructions to the unmanned aerial vehicle and the sensor device through an API interface; The user interaction module develops a Web and mobile terminal application, supports environmental data query, early warning information receiving and manual intervention operation functions.

Images