Water level and water quality monitoring method based on AI technology energizing intelligent coupling pivot rod

Abstract

The invention provides a water level and water quality monitoring method based on an AI technology energizing intelligent central pivot rod, which takes an AI intelligent central pivot sensing rod as a core and integrates various components such as a radar static pressure composite monitoring water level terminal, a water quality five-parameter monitoring terminal, a camera acquisition terminal and a cloud analysis full-amount meteorological sensing station. Through specific structure assembly and field installation, the device is deployed in different water areas such as inland rivers, lakes and rivers. The system collects data according to a set time interval and records environment and scene characteristic parameters. After collected data are preprocessed, adaptive lightweight model combinations are selected according to different scenes and data characteristics, a training set and a test set are divided for model training and fusion, and then the data are deployed to a monitoring terminal and verified by the test set. The system realizes intelligent data processing by means of the AI technology, has flexibility and reliability, can efficiently and accurately collect and stably transmit data, provides accurate technical support for water level and water quality monitoring, and has great significance for flood control and water body ecological protection.

Description

Technical Field

[0001] The present invention relates to the fields of water conservancy informatization and hydrological monitoring information, and in particular to a method for monitoring water level and water quality using an intelligent connection central rod empowered by AI technology. Background Art

[0002] In the context of global climate change, the impact of the El Niño phenomenon is intensifying, and extreme weather events are becoming more frequent. In recent years, extremely heavy rains have occurred frequently, posing huge challenges to river and urban flood control work. Under the impact of extreme precipitation, many shortcomings of the existing flood control system have gradually emerged, making it difficult to withstand such a high-intensity test, resulting in a sharp increase in the risk of urban waterlogging and seriously threatening the safety of residents' lives and property. At the same time, the water ecosystem has also suffered unprecedented damage, the river ecological balance has been broken, biodiversity has been damaged, and the ecological service function has declined. In this situation, the situation of urban inland river water quality governance has become increasingly severe and has become an important issue that needs to be solved urgently.

[0003] With the acceleration of the urban development process, the urban commercial layout has undergone profound changes. A large number of stores have been transformed into restaurants, but the construction of related infrastructure has not kept up. Due to the lack of effective kitchen waste water isolation and connection facilities, the kitchen waste water generated by many restaurants is illegally connected to the rainwater pipe or even directly discharged into the rainwater inlet, and the untreated sewage flows into the inland river wantonly, becoming one of the important pollution sources for the deterioration of the inland river water quality. At the same time, although the problem of industrial sewage being secretly discharged is not common, due to its high pollutant concentration and complex composition, once it occurs, it will cause great harm to the water environment. The combined effect of these industrial sewage and illegally discharged catering sewage has seriously aggravated the degree of deterioration of the inland river water quality.

[0004] Due to the limitations of early planning and construction in old urban areas, the construction of infrastructure lags behind, and the progress of rainwater and sewage pipe network renovation is slow. So far, rainwater and sewage diversion has not been achieved. During each rainfall, domestic sewage will pour into the inland river in large quantities along with the rainwater, further increasing the pollution load of the inland river. In addition, there is currently a lack of effective water quality detection means for the initial rain on roads, making it difficult to accurately judge its discharge destination, which makes the management of inland river water quality face greater difficulties and it is impossible to take targeted treatment measures in a timely manner.

[0005] In terms of the operation and management of urban rivers, the existing water level and water quality monitoring equipment has obvious deficiencies. The equipment layout is sparse, and it is impossible to achieve comprehensive coverage monitoring of the river, resulting in the lack of water level and water quality data in some areas; at the same time, the data accuracy is poor, making it difficult to meet the needs of real-time and accurate regulation of water level, flow velocity and flow rate. This has caused the water flow in some rivers and lakes to be sluggish, the water self-purification ability and pollutant diffusion ability to decrease, sediment and pollutants to accumulate and deposit at the bottom of the river, and the water quality of the inland river to continue to deteriorate. In extreme weather such as heavy rain, this situation is also extremely likely to cause waterlogging disasters, seriously affecting the normal operation of the city and the lives of residents.

[0007] In summary, there are many problems in the field of water body monitoring and treatment at present. Traditional monitoring and treatment methods are difficult to meet the actual needs. There is an urgent need for an efficient and accurate water level and water quality monitoring technology to improve the monitoring ability of the water environment, timely discover and solve water pollution problems, and ensure flood control safety and sustainable utilization of water resources. Summary of the Invention

[0008] The present invention provides a method for monitoring water level and water quality based on an AI technology empowered intelligent connection central rod, which is characterized by including;

[0009] Construct a monitoring system composed of an on-line monitoring device, a network communication system and a remote monitoring platform, wherein the on-line monitoring device includes an AI intelligent connection central sensing rod, a control terminal and a network communication terminal;

[0010] According to different scenarios, the AI intelligent connection central sensing rod selects different lightweight model combinations;

[0011] Install the AI intelligent connection central sensing rod by using specific structure assembly and on-site installation methods in different scenarios;

[0012] Collect data at set time intervals according to the AI intelligent connection central sensing rod in different scenarios, and record relevant environmental parameters and characteristic parameters of specific scenarios;

[0013] Preprocess the collected data, including removing outliers, setting a reasonable data range to eliminate obvious error data points, using a filtering algorithm to remove noise interference, and normalizing or standardizing the data to ensure that data of different parameters are in the same scale range;

[0014] Divide the preprocessed data into a training set and a test set, train each model using the training set data, and adjust the model parameters to minimize the prediction error of the model on the training set;

[0015] Determine a fusion strategy for the lightweight model combination, determine the fusion weights, and perform model fusion; and

[0016] Deploy the fused model to the corresponding monitoring terminal, verify the model using the test set data, and evaluate the accuracy and stability of the model.

[0017] In an embodiment of the present invention, the selection of different lightweight model combinations by the AI intelligent connection central sensing rod according to different scenarios includes:

[0018] The radar static pressure composite monitoring water level terminal carries different combinations of lightweight models in different waters and performs fusion processing on water pressure and water depth data;

[0019] The five-parameter water quality monitoring terminal is equipped with a combined lightweight model in different water areas to correct measurement deviations and predict water quality change trends;

[0020] The camera acquisition terminal is equipped with a combined lightweight model in different water areas to achieve image feature extraction, target recognition, and dynamic change analysis:

[0021] The Yunxi full-scale meteorological sensing station is equipped with a combined lightweight model of the TensorFlow Lite lightweight model and the PyTorch Mobile lightweight model to respectively perform real-time feature extraction and short-term time series prediction on meteorological data, providing a meteorological background for water level and water quality monitoring; and

[0022] The Yunxi optical charging energy fine control terminal is equipped with a combined lightweight model of the image recognition MobileNetV3, the target detection YOLO-Nano, and the environmental data prediction LSTM-Light. It uses MobileNetV3 to identify the surface state of the battery panel, YOLO-Nano to monitor surrounding environmental threats, and LSTM-Light to predict energy supply and demand by combining environmental and device data to optimize energy regulation.

[0023] In an embodiment of the present invention, the combined lightweight model carried by the radar static pressure composite water level monitoring terminal in different water areas includes:

[0024] In inland rivers, a combined lightweight model of a convolutional neural network model of deep learning and a weighted average model of traditional algorithms is carried;

[0025] In small watershed rivers, a combined lightweight model of a model of the machine learning algorithm support vector machine SVM and an ARIMA model of time series analysis is carried;

[0026] In lakes, a combined lightweight model of a neural network model and a model of empirical formula and statistical analysis is carried;

[0027] In reservoirs, a combined lightweight model of a model based on physical principles and a model of machine learning is carried; and

[0028] In rivers, a combined lightweight model of a model based on acoustic principles and a model based on hydraulic principles is carried.

[0029] In an embodiment of the present invention, the combined lightweight model carried by the five-parameter water quality monitoring terminal in different water areas includes:

[0030] In inland rivers, a combined lightweight model of a physicochemical empirical model and an LSTM model of deep learning is carried;

[0031] In small watershed rivers, a combined lightweight model of a water quality model based on the material balance principle and a convolutional neural network model of deep learning is carried;

[0032] Deploy a combined lightweight model of a random forest model and a decision tree model on a lake;

[0033] Deploy a combined lightweight model of a lightweight sensing model based on data validity identification and repetitive screening and a lightweight model of knowledge distillation on a reservoir; and

[0034] Deploy a combined lightweight model of a lightweight water quality prediction model based on a convolutional neural network and a lightweight anomaly detection model based on a long short-term memory network on a river.

[0035] In an embodiment of the present invention, the camera acquisition terminal deploying the combined lightweight model in different waters includes:

[0036] Deploy a combined lightweight model of a target detection lightweight model based on YOLO and a semantic segmentation lightweight model based on U-Net on an inland river;

[0037] Deploy a combined lightweight model of a target detection lightweight model based on Faster R-CNN and a semantic segmentation lightweight model based on MobileNetV3 on a small watershed river channel;

[0038] Deploy a combined lightweight model of a GhostNet lightweight model and a MobileMamba lightweight model on a lake;

[0039] Deploy a combined lightweight model of a YOLOv5 lightweight model and a UNet++ lightweight model on a reservoir; and

[0040] Deploy a combined lightweight model of an EfficientDe lightweight model and a Swin Transformer Lite lightweight model on a river.

[0041] The present invention also provides a water level and water quality monitoring system based on AI technology empowering the intelligent connection central rod, which is characterized by including:

[0042] An AI intelligent connection central sensing rod, arranged in inland rivers, small watershed river channels, lakes, reservoirs, and river waters, for real-time collection of water levels, water quality, flow velocity, flow rate, rainfall, and video data of rivers, inland rivers, lakes, and reservoirs, including:

[0043] A radar static pressure composite water level monitoring terminal, using radar and static pressure sensing technologies, and combining corresponding lightweight models according to different water area scenarios to monitor water area data;

[0044] A water quality five-parameter monitoring terminal, having a variety of water quality sensors, and selecting a matching combination of water quality sensors and lightweight models according to different water area environments to realize real-time monitoring and analysis of five water quality parameters;

[0045] The camera acquisition terminal selects cameras according to different scenarios, and is equipped with corresponding lightweight model combinations to collect video image data for identifying target objects in the water area and monitoring abnormal water surface conditions;

[0046] The cloud analysis full - scale meteorological sensing station is equipped with a lightweight model combination to collect and process meteorological data such as rainfall, wind speed, wind direction, and humidity; and

[0047] The lightweight model set, according to different customer requirements, environmental scenarios, and data characteristics, the AI intelligent connection central sensing pole is equipped with a variety of lightweight models;

[0048] The network communication system uses NBIOT, 4G, and 5G technologies to transmit the collected data to the remote monitoring platform; and

[0049] The remote monitoring platform presents the collected data to users in real - time.

[0050] In an embodiment of the present invention, the five - parameter water quality monitoring terminal includes:

[0051] The first module is used for the inland water area scenario and includes one or more of a pH sensor, a dissolved oxygen sensor, a conductivity sensor, a turbidity sensor, and a temperature sensor, and is equipped with a lightweight model combination of a physical - chemical experience model and a deep - learning LSTM model;

[0052] The second module is used for the small - watershed river channel water area scenario and includes one or more of a pH sensor, a dissolved oxygen sensor, a conductivity sensor, a turbidity sensor, and a chemical oxygen demand sensor, and is equipped with a lightweight model combination of a water quality model based on the material balance principle and a deep - learning convolutional neural network model;

[0053] The third module is used for the lake water area scenario and includes one or more of a pH sensor, a dissolved oxygen sensor, a conductivity sensor, a turbidity sensor, and a temperature sensor, and is equipped with a lightweight model combination of a random forest model and a decision tree model;

[0054] The fourth module is used for the reservoir water area scenario and includes one or more of a pH sensor, a dissolved oxygen sensor, a conductivity sensor, a turbidity sensor, and a temperature sensor, and is equipped with a lightweight model combination of a lightweight sensing model based on data validity identification and repetitive screening and a lightweight model of knowledge distillation; and

[0055] The fifth module is used for the river water area scenario and includes one or more of a pH sensor, a dissolved oxygen sensor, a conductivity sensor, a turbidity sensor, and a temperature sensor, and is equipped with a lightweight water quality prediction model of a convolutional neural network and a lightweight anomaly detection model of a long - short - term memory network combination lightweight model.

[0056] In an embodiment of the present invention, the combined lightweight model of the random forest model and the decision tree model includes the following formula:

[0057] where S DO is the dissolved oxygen saturation, C DO is the actually measured dissolved oxygen concentration, is the saturated dissolved oxygen concentration under given water temperature, air pressure and salinity conditions;

[0058] K = k*c, where k is a proportionality constant related to factors such as the type of ions and temperature, K is the conductivity, and c is the ion concentration;

[0059] pH = -log 10 [H + , where [H + is the hydrogen ion concentration.

[0060] In an embodiment of the present invention, the combined lightweight model of the lightweight sensing model based on data validity identification and repetitive screening and the knowledge distillation lightweight model includes the following formula:

[0061] is used to calculate the difference between two probability distributions and measure the difference in the output distributions of the student model and the teacher model in knowledge distillation, where P is the soft label probability distribution of the teacher model and Q is the probability distribution predicted by the student model;

[0062] is used to calculate the hard label loss, where Y i is the one-hot encoding of the true label, is the predicted probability of the student model.

[0063] In an embodiment of the present invention, the set of lightweight models includes:

[0064] A water pollution monitoring model configured to judge the degree and type of water pollution and evaluate the water body health status based on the data of the five-parameter water quality monitoring terminal;

[0065] A hydrodynamic model configured to fuse flow velocity and flow rate data, simulate the water flow state of rivers and lakes, and predict water level changes and water flow directions;

[0066] A target recognition model configured to identify floating objects and ship target objects in the water area based on the data collected by the camera terminal;

[0067] A meteorological impact model configured to analyze the impact of meteorological factors on the water area environment by integrating rainfall, air pressure, temperature and humidity data;

[0068] The multi-modal detection model is configured to adopt DeepSeek-R1-Distill to support the fusion detection of cameras, radar waves, and ultrasonic waves.

[0069] The time series prediction model is configured to adopt DeepSeek-V3 to process rainfall, temperature, humidity, and water quality data and grasp the time-varying trend of the data; and

[0070] The anomaly classification model is configured to use TinyBERT to process the static pressure sensor data to detect abnormal pressure data.

[0071] The present invention empowers the intelligent connection central rod with AI technology as the core hardware of the system, which reserves rich interfaces and integrates data collection, transmission, and processing functions. The device is equipped with a camera acquisition terminal, various types of sensors, and an intelligent rainfall workstation, and is equipped with a five-parameter water quality monitoring terminal. The five-parameter water quality monitoring terminal can flexibly select and match sensors and carry lightweight models according to different water environments such as inland rivers, lakes, and rivers. It is built with a 4G high-gain and Beidou positioning module and has an all-round sensing ability. It can accurately match the terminal and sensors according to customer needs and environmental scenarios to ensure the efficient and accurate collection of water level, flow velocity, and water quality in different scenarios and environments. Brief Description of the Drawings

[0072] Figure 1 The flowchart shows the method for monitoring water level and water quality based on the intelligent connection central rod empowered by AI technology in an embodiment of the present invention;

[0073] Figure 2 The schematic diagram shows the water level and water quality monitoring system based on the intelligent connection central rod empowered by AI technology in an embodiment of the present invention. Detailed Description of the Embodiments

[0074] In the following description, the present invention is described with reference to the embodiments. However, those skilled in the art will recognize that the embodiments can be implemented without one or more specific details or in combination with other alternative and / or additional methods, materials, or components. In other cases, well-known structures, materials, or operations are not shown or described in detail to avoid obscuring the inventive points of the present invention. Similarly, for the purpose of explanation, specific quantities, materials, and configurations are set forth to provide a comprehensive understanding of the embodiments of the present invention. However, the present invention is not limited to these specific details.

[0075] In this specification, the reference to "an embodiment" or "the embodiment" means that the specific features, structures, or characteristics described in connection with the embodiment are included in at least one embodiment of the present invention. The phrase "in an embodiment" appearing throughout this specification does not necessarily refer to the same embodiment.

[0076] In the present invention, each embodiment is only intended to illustrate the solution of the present invention and should not be construed as restrictive.

[0077] In addition, the numbering of the steps of each method of the present invention does not limit the execution order of the method steps. Unless otherwise specified, the method steps can be executed in different orders.

[0078] The present invention will be further described below in conjunction with the specific embodiments with reference to the accompanying drawings.

[0079] Figure 1 The flowchart of the method for water level and water quality monitoring based on the AI - empowered intelligent connection central rod in an embodiment of the present invention is shown.

[0080] As Figure 1 shown, the method for water level and water quality monitoring based on the AI - empowered intelligent connection central rod includes:

[0081] Construct system 100: Build a monitoring system composed of an on - line monitoring device, a network communication system, and a remote monitoring platform. Among them, the on - line monitoring device includes an AI intelligent connection central sensing rod, a control terminal, and a network communication terminal. The AI intelligent connection central sensing rod is the core device for data collection. The control terminal is responsible for issuing commands and regulating the operation of the sensing rod. The network communication terminal undertakes the key task of data transmission to ensure smooth information interaction among all parts, jointly forming an organic whole.

[0082] Select lightweight model 110: According to the data characteristics and monitoring requirements in different scenarios, combined with the performance of the AI intelligent connection central sensing rod, select a suitable model combination from a variety of lightweight models. For example, for the radar - static pressure composite water level monitoring terminal, according to the different situations of inland rivers, small - watershed rivers, lakes, reservoirs, and rivers, respectively select the combination of the convolutional neural network (CNN) model of deep learning and the weighted average model of traditional algorithms, the model of the support vector machine (SVM) of machine learning algorithms and the ARIMA model of time - series analysis, etc.; for the water quality five - parameter monitoring terminal, select the combination of the physical - chemical empirical model and the LSTM model of deep learning in inland rivers, and select the combination of the water quality model based on the material balance principle and the convolutional neural network (CNN) model of deep learning in small - watershed rivers, etc., to ensure that the model can accurately process and analyze data in different scenarios.

[0083] On-site installation 120: According to the different geographical characteristics and environmental requirements of inland rivers, small watershed rivers, lakes, reservoirs, and river waters, specific structures designed specifically are assembled respectively, and the deployment of the AI intelligent connection center sensing pole is completed through precise on-site installation methods. The specific structure is a combination of specific terminals and lightweight models. For different water areas, the present invention provides different combinations of terminals and lightweight models. The sensing pole integrates a variety of functional modules, including a radar static pressure composite water level monitoring terminal, a water quality five-parameter monitoring terminal, a camera acquisition terminal, a cloud analysis full-scale meteorological sensing station, and a cloud analysis optical charging energy precision control terminal. These modules cooperate with each other to achieve synchronous acquisition of various types of data such as water level, water quality, flow velocity and flow rate, rainfall, and video, as well as monitoring of meteorological data and intelligent management of energy, meeting the monitoring requirements of different water area environments.

[0084] Data acquisition 130: The AI intelligent connection center sensing pole continuously and stably acquires water level, water quality, flow velocity and flow rate, rainfall, and video data at set time intervals in different scenarios. At the same time, relevant environmental parameters during data acquisition, such as water temperature, air pressure, wind speed, wind direction, etc., and characteristic parameters of specific scenarios, such as river channel slope, lake basin shape, reservoir storage capacity curve, etc., are recorded in detail. These rich data provide a comprehensive information basis for subsequent analysis and processing, helping to accurately grasp the changes in the water area environment.

[0085] Data preprocessing 140: Preprocess the acquired data to improve data quality and usability. Specifically, it includes: setting a reasonable data range, removing obvious error data points, and eliminating outliers; using filtering algorithms, such as moving average filtering, median filtering, or Kalman filtering, to remove noise interference in the data; performing normalization or standardization processing on the data, converting data of different parameters to the same scale range, ensuring the accuracy and stability of subsequent model processing, and making the data more compliant with the requirements of model training and analysis.

[0086] Model training 150: Divide the preprocessed data into a training set and a test set according to a certain ratio. Use the training set data to train each selected model respectively. By adjusting the parameters of the model, such as the number of layers and nodes of the neural network, the kernel function type and parameters of the SVM, the order of the ARIMA model, etc., make the prediction error of the model on the training set reach the minimum. During the training process, optimization algorithms, such as stochastic gradient descent, Adagrad, etc., are used to continuously optimize the model performance and improve the model's fitting ability and prediction accuracy for the data.

[0087] Model Fusion 160: For the selected combination of lightweight models, determine an appropriate fusion strategy, and determine the fusion weights through multiple experiments, cross-validation, or optimization algorithms. Common fusion strategies include weighted average, voting method, etc. The trained multiple models are fused according to the fusion strategy, integrating the advantages of each model to obtain a fusion model with better performance, enabling it to more accurately reflect the data characteristics and laws, and improving the accuracy and reliability of monitoring.

[0088] Model Deployment and Evaluation 170: Deploy the fused model to the corresponding monitoring terminals, such as radar static pressure composite monitoring water level terminals, water quality five-parameter monitoring terminals, etc. Use the test set data to verify the deployed model, and comprehensively evaluate the accuracy and stability of the model by calculating evaluation indicators such as root mean square error (RMSE), mean absolute error (MAE), intersection over union (IoU), etc. According to the verification results, further adjust and optimize the model, such as fine-tuning model parameters, reselecting fusion weights, etc., to ensure that the model can operate stably and accurately in the actual monitoring scenario, providing reliable technical support for water level and water quality monitoring.

[0089] Figure 2 Fig. shows a schematic diagram of a water level and water quality monitoring system based on an AI technology empowered intelligent connection central rod in an embodiment of the present invention.

[0090] As Figure 2 shown, the water level and water quality monitoring system based on an AI technology empowered intelligent connection central rod includes:

[0091] The AI intelligent connection central sensing rod 200, as the core hardware of the system, reserves rich interfaces and integrates data collection, transmission, and processing functions. The device is equipped with various types of sensors, and can flexibly select and configure sensors and lightweight models according to different water area environments such as inland rivers, lakes, and rivers to ensure efficient data collection and transmission.

[0092] The network communication system 300 uses NBIOT, 4G, 5G technologies and the OneNet protocol, and through the NBIOT, 4G, 5G base stations, transmits the monitoring data collected by the lightweight model carried by the AI intelligent connection central sensing rod 200 to the remote monitoring platform 400 in real time through the OneNet mobile Internet of Things open platform or the telecommunications AEP platform;

[0093] The remote monitoring platform 400 presents the collected data to the user in real time;

[0094] The AI intelligent connection central sensing rod includes:

[0095] The control module 210 is configured to select the appropriate terminal device and lightweight model according to customer requirements and environmental scenarios.

[0096] Radar and static pressure composite monitoring water level terminal 221 uses radar (ultrasonic) and static pressure sensing technologies to accurately measure the water level height. At the same time, it can obtain flow velocity and flow rate data and also provide auxiliary data for water quality monitoring. For example, in river monitoring, through the reflection of radar waves and the principle of static pressure, it can accurately sense the subtle changes in water level and water flow velocity. Water level data is obtained using the principle of radar electromagnetic wave reflection ranging and the pressure-water depth conversion principle of the static pressure sensor. For different water area hydrological characteristics, different combinations of lightweight models are carried to fuse the two types of data, eliminate environmental interference, and improve measurement accuracy.

[0097] Water quality five-parameter monitoring terminal 222 uses technologies such as the electrode method and light scattering, and is equipped with five sensors among pH sensor, dissolved oxygen sensor, conductivity sensor, turbidity sensor, temperature sensor, and chemical oxygen demand sensor. Different modules are selected according to different water areas, and it can monitor key water quality parameters such as the acidity and alkalinity of water bodies, dissolved oxygen content, ion concentration, turbidity, and water temperature in real time. Based on the differences in pollution sources and ecology in different water areas, a combined lightweight model is carried to correct measurement deviations and predict the water quality change trend.

[0098] Camera acquisition terminal 223 obtains water area images through an image sensor. According to monitoring targets (floating objects, biological activities, etc.) and environmental conditions, a combined lightweight model is carried to achieve image feature extraction, target recognition, and dynamic change analysis.

[0099] Cloud analysis full-scale meteorological sensing station 224 is equipped with a combination of lightweight models such as TensorFlow Lite and PyTorch Mobile. Meteorological parameters are collected through various sensors, and the TensorFlow Lite and PyTorch Mobile combined lightweight models are used to perform real-time feature extraction and short-term time series prediction on meteorological data respectively, providing a meteorological background for water level and water quality monitoring.

[0100] Beidou positioning module 225 is built with a high-precision Beidou positioning chip, providing accurate positioning services for the AI intelligent connection center sensing pole and obtaining the geographical location information of the device. On the one hand, it is convenient for geospatial analysis and management of monitoring data, and on the other hand, it plays an important role in device tracking, maintenance, etc. For example, it can accurately determine the position of the sensing pole in large-area water area monitoring.

[0101] Lightweight model set 226. According to different customer needs, environmental scenarios, and data characteristics, the AI intelligent connection center sensing pole can flexibly carry a variety of lightweight models to efficiently process and deeply analyze the collected data, improving the intelligent level and monitoring accuracy of the monitoring system.

[0102] The network communication module 230, based on NBIOT, 4G, and 5G technologies, follows the OneNet protocol to establish a stable connection with the network communication system, transmits the data collected by the AI intelligent connection center sensing pole to the remote monitoring platform in real time, and at the same time receives the instructions from the remote monitoring platform to ensure smooth two-way interaction of data and ensure the timeliness and accuracy of data transmission.

[0103] The cloud analysis optical charging energy precise control terminal 240 uses solar energy and an external replaceable energy storage battery as the power supply source, has the ability to adapt to the voltages of various sensors, and stably powers each module of the AI intelligent connection center pole. It realizes energy conversion and storage through solar panels and storage batteries, uses MobileNetV3 to identify the surface state of the battery panel, YOLO-Nano to monitor surrounding environmental threats, and LSTM-Light to predict energy supply and demand by combining environmental and device data to optimize energy regulation.

[0104] In an embodiment of the present invention, the steps and methods of the lightweight model of the combined radar static pressure composite monitoring water level terminal 221 with a convolutional neural network (CNN) model of deep learning and a weighted average model of traditional algorithms on inland rivers are as follows:

[0105] Data collection: Collect inland river water level data through radar (ultrasonic) and static pressure sensors at a set frequency.

[0106] Data preprocessing: Process the collected data such as denoising and filtering. In this embodiment, moving average filtering is used to remove high-frequency noise.

[0107] Model selection and training: Select a CNN model to process radar (ultrasonic) data and a weighted average model to process static pressure data. Use the preprocessed data to train the two models respectively, and adjust the parameters to minimize the loss function.

[0108] Deployment and testing: Deploy the fusion model to the water level monitoring terminal, conduct on-site testing and verification, and adjust and optimize according to the results.

[0109] In this embodiment, the parameters and data required for the lightweight combination of a convolutional neural network (CNN) model of deep learning and a weighted average model of traditional algorithms are as follows:

[0110] Sensor parameters: Include the transmission power and receiving sensitivity of the radar (ultrasonic) sensor, the accuracy, resolution, etc. of the static pressure sensor.

[0111] Environmental parameters: The water temperature, air pressure, water flow velocity, surface wind force, etc. of the inland river.

[0112] Historical data: Historical water level data, historical meteorological data, etc. of the inland river, used for model training and verification.

[0113] The relevant lightweight formulas for the combination of the convolutional neural network (CNN) model of deep learning and the weighted average model of traditional algorithms are as follows:

[0114] Radar (ultrasonic) water level calculation: Among them, h1 is the water level height calculated by radar (ultrasonic), v is the wave speed, and t is the propagation time.

[0115] Static pressure water level calculation: Among them, h2 is the water level height calculated by static pressure, p is the pressure, ρ is the density of water, and g is the acceleration due to gravity.

[0116] Fusion calculation: H = w1×h1 + w2×h2, where H is the final water level height, w1 and w2 are the weights of the radar (ultrasonic) model result and the static pressure model result respectively, and w1 + w2 = 1. The weights can be determined through multiple experiments or optimization algorithms.

[0117] In an embodiment of the present invention, the steps and methods for the lightweight model of the combined model of the machine learning algorithm support vector machine (SVM) and the autoregressive integrated moving average (ARIMA) model carried by the radar static pressure composite monitoring water level terminal in the small watershed river channel are as follows:

[0118] Data collection: Using radar (ultrasonic) sensors and static pressure sensors, continuously collect water level-related data of the small watershed river channel at regular time intervals (such as once every 10 minutes).

[0119] Data preprocessing: Clean the collected data, remove outliers, for example, by setting a reasonable data range to eliminate obvious error data points; use a filtering algorithm (such as median filtering) to remove noise interference and improve data quality.

[0120] Model training, model selection: Select a model suitable for processing such data according to the characteristics of radar (ultrasonic) data. In this embodiment, it is the SVM model, and select the ARIMA model for static pressure data.

[0121] Parameter setting: Set initial parameters for the selected models, such as the kernel function type and parameters of SVM, and the order of the ARIMA model.

[0122] Training data division: Divide the preprocessed data into a training set and a test set, generally divided according to a ratio of 70% - 30% or 80% - 20%.

[0123] Training process: Use the training set data to train the two models respectively, and continuously adjust the model parameters to minimize the prediction error of the models on the training set.

[0124] Model fusion: Fusion the two trained models according to the set fusion rule (such as weighted average) to obtain the final water level prediction model.

[0125] Model Deployment and Verification: Deploy the fused model into the microprocessor of the radar-static pressure composite monitoring water level terminal, and use the test set data to verify the model, evaluate the accuracy and stability of the model, such as calculating indicators like root mean square error (RMSE), mean absolute error (MAE), etc. According to the verification results, fine-tune and optimize the model parameters or fusion weights.

[0126] In this embodiment, the parameters and data required for the combination of the model with support vector machine (SVM) and autoregressive integrated moving average (ARIMA) model for time series analysis are as follows:

[0127] Sensor Parameters: Transmission frequency, beam angle, measurement accuracy of the radar (ultrasonic) sensor; measuring range, accuracy, zero drift coefficient, etc. of the static pressure sensor. These parameters determine the quality and accuracy of the data collected by the sensor.

[0128] Small Watershed River Channel Characteristic Parameters: Slope, roughness coefficient, cross-sectional shape and size, catchment area, etc. of the river channel. These parameters help to understand the water flow characteristics and water level change laws of the small watershed river channel, and play an important role in the training and calibration of the model.

[0129] Environmental Parameters: Meteorological data such as rainfall, evaporation, temperature, air pressure, etc. in the small watershed. These environmental factors will affect the water level change in the small watershed, and as auxiliary data can improve the prediction accuracy of the model.

[0130] Historical Water Level Data: Historical water level data of the small watershed river channel, preferably including data under different seasons and different rainfall conditions, for model training and verification, enabling the model to learn the characteristics and laws of water level changes.

[0131] In this embodiment, the relevant formulas of the SVM model and the ARIMA model for time series analysis are as follows:

[0132] Radar (Ultrasonic) Water Level Calculation: Where h1 is the water level height calculated by radar (ultrasonic), v is the wave speed, and t is the propagation time.

[0133] Static Pressure Water Level Calculation: Where h2 is the water level height calculated by static pressure, p is the pressure, ρ is the density of water, and g is the acceleration due to gravity.

[0134] Fusion Calculation: H = w1×h1 + w2×h2, where H is the final water level height, w1 and w2 are the weights of the radar (ultrasonic) model result and the static pressure model result respectively, and w1 + w2 = 1. The weights can be determined through multiple experiments or optimization algorithms.

[0135] In an embodiment of the present invention, the steps and methods for the lightweight model of the combined model of the neural network-based model, empirical formula, and statistical analysis model in the radar-static pressure composite monitoring water level terminal are as follows:

[0136] Data collection: Use radar (ultrasonic) sensors and static pressure sensors to collect lake water level data at set time intervals (such as 30 minutes). At the same time, record meteorological data at the collection moment, such as wind speed, wind direction, air temperature, etc.

[0137] Data preprocessing: Clean the collected data to remove outliers caused by sensor failures, abnormal reflections, etc. Use filtering algorithms (such as Kalman filtering) to remove noise and improve data quality.

[0138] Model selection: For radar (ultrasonic) data, select a lightweight model based on neural networks, such as MobileNet, etc.; for static pressure data, select a model based on statistical analysis and empirical formulas.

[0139] Parameter setting: Set initial parameters for the selected models, such as the number of layers and nodes of the neural network, and the correlation coefficients of the statistical model, etc.

[0140] Training data division: Divide the preprocessed data into a training set and a test set according to a ratio of 70% - 30% or 80% - 20%.

[0141] Model training: Use the training set data to train the two models separately, and continuously adjust the model parameters to minimize the prediction error of the models on the training set.

[0142] Model fusion: Fusion the two trained models according to the set fusion rules (such as weighted average) to obtain the final water level prediction model.

[0143] Model deployment and verification: Deploy the fused model to the microprocessor of the monitoring water level terminal, use the test set data to verify the model, and evaluate the accuracy and stability of the model. According to the verification results, fine-tune and optimize the model parameters or fusion weights.

[0144] In this embodiment, the parameters and data required for the combined lightweight model of the neural network-based model, empirical formula, and statistical analysis model are as follows:

[0145] Sensor parameters: Transmission power, receiving sensitivity, beam width of the radar (ultrasonic) sensor; accuracy, resolution, temperature compensation coefficient, etc. of the static pressure sensor. These parameters determine the quality and accuracy of the data collected by the sensors.

[0146] Lake characteristic parameters: average water depth, maximum water depth, water surface area, lake basin shape, slope, etc. These parameters help to understand the water level change law of the lake and play an important role in the training and calibration of the model.

[0147] Environmental parameters: meteorological data such as wind speed, wind direction, air temperature, air pressure, precipitation, evaporation, etc. in the lake area. These environmental factors will affect the lake water level and can improve the prediction accuracy of the model as auxiliary data.

[0148] Historical water level data: historical water level data of the lake, preferably covering data under different seasons and different meteorological conditions, used for the training and verification of the model to enable the model to learn the characteristics and laws of water level changes.

[0149] In this embodiment, the relevant formulas of the model loaded with neural network and the model of empirical formula and statistical analysis are as follows:

[0150] Radar (ultrasonic) water level calculation: Where h1 is the water level height calculated by radar (ultrasonic), v is the wave speed, and t is the propagation time.

[0151] Static pressure water level calculation: Where h2 is the water level height calculated by static pressure, p is the pressure, ρ is the density of water, and g is the acceleration due to gravity.

[0152] Fusion calculation: H = w1×h1 + w2×h2, where H is the final water level height, w1 and w2 are the weights of the radar (ultrasonic) model result and the static pressure model result respectively, and w1 + w2 = 1. The weights can be determined through multiple experiments or optimization algorithms.

[0153] In an embodiment of the present invention, the steps and methods of the lightweight model of the combined physical principle model and machine learning model carried by the radar static pressure composite monitoring water level terminal in the reservoir are as follows:

[0154] Data collection: At regular time intervals, use radar (ultrasonic) and static pressure sensors to collect reservoir water level data, and at the same time record data such as reservoir inflow and outflow.

[0155] Data preprocessing: Remove outliers, use methods such as smoothing filtering to denoise the data, and perform normalization or standardization processing on the data

[0156] Model selection and parameter setting: Select a suitable lightweight model, select a convolutional neural network model for radar (ultrasonic) data, select a linear regression model for static pressure data, and set initial parameters.

[0157] Model training: Divide the data into a training set and a test set, train the model with the training set, and adjust the parameters to minimize the loss function.

[0158] Model fusion: Determine the fusion strategy, such as weighted average fusion, and determine the weights through experiments or optimization algorithms.

[0159] Model deployment and verification: Deploy the model to the monitoring terminal, verify it with the test set, and fine-tune and optimize according to the results.

[0160] In this embodiment, the models incorporating physical principles and machine learning combine to lightweight the required parameters and data of the model.

[0161] Sensor parameters: The frequency and emission angle of the radar (ultrasonic) sensor, the range and accuracy of the static pressure sensor, etc.

[0162] Reservoir characteristic parameters: The reservoir storage curve, dam height, water surface area, average water depth, reservoir bottom slope, etc.

[0163] Environmental parameters: Meteorological data such as air temperature, air pressure, wind speed, and wind direction around the reservoir, as well as parameters such as water temperature and water quality of the water body.

[0164] Operation data: Operation data such as the historical water level data, inflow rate, outflow rate, and water storage volume of the reservoir.

[0165] In this embodiment, the relevant formulas of the models incorporating physical principles and machine learning are as follows:

[0166] Radar (ultrasonic) water level calculation: Where h1 is the water level height calculated by the radar (ultrasonic), v is the wave speed, and t is the propagation time.

[0167] Static pressure water level calculation: Where h2 is the water level height calculated by the static pressure, p is the pressure, ρ is the density of water, and g is the acceleration due to gravity.

[0168] Fusion calculation: H = w1×h1 + w2×h2, where H is the final water level height, w1 and w2 are the weights of the radar (ultrasonic) model result and the static pressure model result respectively, and w1 + w2 = 1. The weights can be determined through multiple experiments or optimization algorithms.

[0169] In an embodiment of the present invention, the steps and methods for the radar-static pressure composite monitoring water level terminal to lightweight the model by combining the models incorporating acoustic principles and hydraulic principles in rivers are as follows:

[0170] Install sensors: At appropriate positions in the river, such as the shore or bridge pier, install the radar (ultrasonic) and static pressure sensors as required to ensure their stability and appropriate measurement range.

[0171] Data acquisition: At a set time interval, such as 15 minutes, acquire the radar (ultrasonic) and static pressure data, and at the same time record data such as the water flow velocity and sediment concentration.

[0172] Data preprocessing: Eliminate obvious outliers, remove noise using a filtering algorithm, and normalize or standardize the data.

[0173] Model selection and parameter setting: Select a suitable model. For example, use the LSTM model in deep learning for radar (ultrasonic) data and a simplified model based on physical equations for static pressure data, and set the initial parameters of the model.

[0174] Model training: Divide the training set and the test set, train the model with the training set, and adjust the parameters to minimize the error of the model on the training set.

[0175] Model fusion (when there are two models): Determine the fusion algorithm and weights according to the characteristics of the river and the data characteristics. For example, optimize the weights using a genetic algorithm.

[0176] Model deployment and verification: Deploy the model to the monitoring terminal, verify it with the test set, and adjust and optimize according to the results.

[0177] In this embodiment, the parameters and data required for the lightweight model combined with the model based on acoustic principles and the model based on hydraulic principles are as follows:

[0178] Sensor parameters: The frequency, transmit power, and receive sensitivity of the radar (ultrasonic) sensor, and the accuracy, response time, etc. of the static pressure sensor.

[0179] River characteristic parameters: The riverbed slope, river width, cross-sectional shape, roughness coefficient, etc. of the river.

[0180] Environmental parameters: Meteorological data such as air temperature, air pressure, precipitation, wind speed, and wind direction, as well as water quality parameters such as water temperature and pH value.

[0181] Flow parameters: Flow velocity, flow rate, sediment concentration, etc.

[0182] In this embodiment, the relevant formulas for the lightweight model combined with the model based on acoustic principles and the model based on hydraulic principles are as follows:

[0183] Radar (ultrasonic) water level calculation: Where h1 is the water level height calculated by radar (ultrasonic), v is the wave speed, and t is the propagation time.

[0184] Static pressure water level calculation: Where h2 is the water level height calculated by static pressure, p is the pressure, ρ is the density of water, and g is the acceleration due to gravity.

[0185] Fusion calculation: H = w1×h1 + w2×h2, where H is the final water level height, w1 and w2 are the weights of the radar (ultrasonic) model result and the static pressure model result respectively, and w1 + w2 = 1. The weights can be determined through multiple experiments or optimization algorithms.

[0186] In an embodiment of the present invention, five water quality sensors are selected and matched for the inland river in the water quality five-parameter monitoring terminal: a pH sensor is used to detect acidity and alkalinity; a dissolved oxygen sensor is used to monitor the dissolved oxygen content in water; a conductivity sensor can reflect the ion concentration in water; a turbidity sensor is used to measure the degree of scattering and absorption of light by suspended particles in the water body to characterize the turbidity of the water body; a temperature sensor is used to measure the water temperature. The steps and methods of the lightweight model combining the physical and chemical empirical model and the deep learning LSTM model are as follows:

[0187] First, install the sensors. Install the five sensors at appropriate positions in the inland river to ensure full contact with the water body and no interference from water flow, etc. Then collect data at regular time intervals, and remove outliers, filter noise, and normalize the data. Next, select appropriate physical and chemical empirical models and LSTM models and set parameters, train the two models respectively with the collected data, determine an appropriate fusion strategy to fuse the two, and finally deploy them to the monitoring terminal, and verify and optimize with the data not involved in training.

[0188] The required parameters and data include:

[0189] Sensor parameters, including range, accuracy, response time, etc.

[0190] Inland river environmental parameters, such as water flow velocity, water depth, water temperature, air temperature, air pressure, etc.

[0191] Historical water quality data, including the five parameters of pH, dissolved oxygen, etc. of the inland river in the past and other water quality index data. Relevant data such as the distribution of pollution sources, surrounding land use types, and population density are also required.

[0192] In this embodiment, the relevant formulas of the lightweight model combining the physical and chemical empirical model and the deep learning LSTM model are as follows:

[0193] pH=-log 10 [H + , where [H + is the hydrogen ion concentration.

[0194] Dissolved oxygen saturation where S DO is the dissolved oxygen saturation, C DO is the actually measured dissolved oxygen concentration, is the saturated dissolved oxygen concentration under given water temperature, air pressure and salinity conditions.

[0195] The relationship between conductivity and ion concentration conforms to Kohlrausch's law, K v =∑ i c i λ i , K is the conductivity, c i is the ion concentration, λi is the ionic molar conductivity.

[0196] In an embodiment of the present invention, the five-parameter water quality monitoring terminal selects and matches five water quality sensors in the small watershed river: The pH sensor is used to detect the acidity and alkalinity. The dissolved oxygen sensor is used to monitor the dissolved oxygen content in the water. The conductivity sensor can reflect the ion concentration in the water. The turbidity sensor is used to measure the degree of light scattering and absorption by suspended particles in the water body, characterizing the water turbidity. The chemical oxygen demand (COD) sensor: can quickly determine the content of reducible substances that can be oxidized in the water body, reflecting the degree of organic pollution of the water body. The steps and methods of the lightweight model combining the water quality model based on the material balance principle and the convolutional neural network (CNN) model of deep learning are as follows:

[0197] First, reasonably install the five sensors at different positions in the small watershed river, ensuring full contact with the water body and no interference. Collect data at regular time intervals, and perform preprocessing on the data such as denoising, removing outliers, and normalization. Select and adjust the parameters of the model based on the material balance principle and the CNN model according to the characteristics of the small watershed, train the two models respectively with the preprocessed data, fuse the results of the two models by means of weighted fusion, etc., deploy them to the monitoring terminal, verify with an independent data set and optimize according to the results.

[0198] Required parameters and data: Parameters such as the range, accuracy, and response time of the sensors. Environmental parameters such as the water flow velocity, flow rate, water depth, water temperature, air temperature, and air pressure in the small watershed. Historical water quality data includes pH, DO, turbidity, conductivity, COD, and other related index data, as well as the distribution of pollution sources, land use types, soil types, precipitation runoff data, etc. in the watershed.

[0199] The relevant formulas of the lightweight model combining the water quality model based on the material balance principle and the convolutional neural network (CNN) model of deep learning are as follows:

[0200] pH = -log 10 [H + , where [H + is the hydrogen ion concentration.

[0201] Dissolved oxygen saturation where S DO is the dissolved oxygen saturation, C DO is the actually measured dissolved oxygen concentration, is the saturated dissolved oxygen concentration under given water temperature, air pressure, and salinity conditions.

[0202] The relationship between conductivity and ion concentration conforms to Kohlrausch's law, K v = ∑ i c i λ i , K is the conductivity, ci is the ionic concentration, and λ i is the molar ionic conductivity.

[0203] The COD calculation often uses the relevant formula of the potassium dichromate method V0 and V1 are the volumes of ammonium ferrous sulfate consumed in the titration of the blank and the sample respectively, c is the concentration of ammonium ferrous sulfate, and V is the volume of the water sample.

[0204] In an embodiment of the present invention, the five-parameter water quality monitoring terminal selects and matches five water quality sensors in the lake: the pH sensor is used to detect the acidity and alkalinity, the dissolved oxygen sensor is used to monitor the dissolved oxygen content in the water, the conductivity sensor can reflect the ionic concentration in the water, the turbidity sensor is used to measure the degree of scattering and absorption of light by suspended particles in the water body, which characterizes the turbidity of the water body, and the temperature sensor is used to measure the water temperature.

[0205] The steps and methods for loading the combined lightweight model of the random forest model and the decision tree model are as follows:

[0206] Data collection: Reasonably arrange the five water quality sensors at different positions in the lake (such as the center of the lake, the inlet of the lake, the outlet of the lake, etc.), and collect data at regular time intervals (such as every hour or every half day). At the same time, record the environmental parameters when collecting data, such as air temperature, air pressure, wind speed, wind direction, etc.

[0207] Data preprocessing: Clean the collected original data to remove outliers and noise; perform normalization processing on the data to convert the data of different parameters to the same scale range for subsequent model processing.

[0208] Model training: For the model based on physical and chemical principles, determine the relevant physical and chemical parameters and boundary conditions according to the specific situation of the lake, and use historical data to calibrate and verify the model parameters. For the machine learning-based random forest model, divide the preprocessed data into a training set and a test set. Use the training set to train the model, adjust the model parameters (such as the number of decision trees, the depth of the tree, etc.) to improve the prediction performance of the model. Evaluate the generalization ability of the model through methods such as cross-validation.

[0209] Model fusion: Fusion of the two trained models can adopt methods such as weighted average, determine the weights according to the performance of the two models on the test set, so that the fused model can more accurately predict the lake water quality parameters.

[0210] Model deployment: Deploy the fused model to the hardware device of the five-parameter water quality monitoring terminal to achieve real-time monitoring and prediction of the lake water quality.

[0211] Model Evaluation and Update: Regularly evaluate the model using new monitoring data, and adjust the model's parameters or structure in a timely manner according to the evaluation results to adapt to the changes in lake water quality.

[0212] Parameters and data required for the lightweight model with the combination of random forest model and decision tree model:

[0213] Sensor parameters: including the range, accuracy, resolution, response time, calibration coefficient, etc. of each sensor. These parameters are crucial for accurately obtaining water quality data.

[0214] Lake environmental parameters: such as the area, water depth, lake basin shape, water flow velocity (if there is water flow), vertical distribution of water temperature, air temperature, air pressure, precipitation, evaporation, wind speed, wind direction, etc.

[0215] Historical water quality data: Long-term accumulated lake water quality data, including five parameters and other related water quality indicators (such as total nitrogen, total phosphorus, chemical oxygen demand, etc.). These data are the basis for model training and verification.

[0216] Pollution source data: Understand the pollution source situation around the lake, such as industrial wastewater discharge, domestic sewage discharge, agricultural non-point source pollution (fertilizer and pesticide usage, etc.), atmospheric deposition, etc. These data can help analyze the reasons for water quality changes.

[0217] Biological data: Information on the types, quantities, and distributions of aquatic organisms in the lake, because biological activities can also affect water quality.

[0218] The relevant formulas for the lightweight model with the combination of random forest model and decision tree model are as follows:

[0219] pH = -log 10 [H + , where [H + is the hydrogen ion concentration.

[0220] Dissolved oxygen saturation where S DO is the dissolved oxygen saturation, C DO is the actually measured dissolved oxygen concentration, is the saturated dissolved oxygen concentration under given water temperature, air pressure, and salinity conditions.

[0221] The relationship between conductivity and ion concentration conforms to Kohlrausch's law, K v = ∑ i c i λ i , K is the conductivity, c i is the ion concentration, λ i is the ionic molar conductivity.

[0222] In an embodiment of the present invention, five water quality sensors are selected and matched in the reservoir for the water quality five-parameter monitoring terminal: a pH sensor is used to detect acidity and alkalinity, a dissolved oxygen sensor is used to monitor the dissolved oxygen content in water, a conductivity sensor can reflect the ion concentration in water, a turbidity sensor is used to measure the degree of scattering and absorption of light by suspended particles in the water body to characterize the water body turbidity, and a temperature sensor is used to measure the water temperature.

[0223] The steps and methods for loading the lightweight model combination of the lightweight sensing model based on data validity identification and repetitive elimination and the lightweight model of knowledge distillation are as follows:

[0224] Parameter and data preparation: Define parameters such as the range, accuracy, and output signal of the sensors, and prepare the original data collected by the sensors, the preprocessed data, etc. Also, define the hyperparameters of the lightweight model, such as the temperature parameter in knowledge distillation, the weight of the loss function, etc.

[0225] Hardware and software integration: Connect the sensors to the water quality five-parameter monitoring terminal through corresponding interfaces to ensure stable connection and normal signal transmission. In the software system of the monitoring terminal, import the algorithm program of the lightweight model and set the relevant parameters and operating environment of the model.

[0226] Model training and calibration: Use a large amount of reservoir water quality data collected by the sensors to train the lightweight model, adjust the parameters of the model so that it can accurately process and predict water quality data. According to the actual reservoir water quality situation and known standard data, calibrate the sensors and the model to ensure the accuracy of measurement and prediction.

[0227] Testing and optimization: Conduct actual on-site tests to check the accuracy, stability of the data, and the performance of the model. According to the test results, optimize and adjust the system, such as adjusting the position of the sensors, optimizing the parameters of the model, etc.

[0228] The parameters and data required for loading the lightweight model combination of the lightweight sensing model based on data validity identification and repetitive elimination and the lightweight model of knowledge distillation are as follows:

[0229] Sensor-related parameters: Include the range, accuracy, resolution, response time, operating temperature range, supply voltage, etc. of each sensor.

[0230] Reservoir water quality and environmental data: In addition to the pH value, dissolved oxygen, conductivity, turbidity, and temperature data collected by the sensors, the water level and flow data of the reservoir, as well as environmental parameters such as the temperature, air pressure, and light intensity at the location of the reservoir are also required.

[0231] Model training data: If a lightweight model based on knowledge distillation is adopted, a teacher model with excellent performance but high computational cost is required, as well as a large amount of labeled or unlabeled water quality data for training the teacher model and the student model. For a lightweight sensing model based on data validity identification and repetitive elimination, a large amount of historical water quality sensing data sets are required to establish the model.

[0232] The lightweight formula for the combination of a lightweight sensing model for data validity identification and repetitive elimination and a lightweight model for group knowledge distillation is as follows:

[0233] Kullback-Leibler (KL) divergence formula: It is used to calculate the difference between two probability distributions and measures the difference between the output distributions of the student model and the teacher model in knowledge distillation. The formula is where P is the soft label probability distribution of the teacher model and Q is the probability distribution predicted by the student model.

[0234] Cross-entropy loss formula: It is used to calculate the hard label loss. The formula is where Y i is the one-hot encoding of the true label, and is the predicted probability of the student model.

[0235] In an embodiment of the present invention, the water quality five-parameter monitoring terminal selects five water quality sensors in the river: a pH sensor is used to detect acidity and alkalinity, a dissolved oxygen sensor is used to monitor the dissolved oxygen content in water, a conductivity sensor can reflect the ion concentration in water, a turbidity sensor is used to measure the degree of scattering and absorption of light by suspended particles in the water body to characterize the water body turbidity, and a temperature sensor is used to measure the water temperature.

[0236] The steps and methods for the combination lightweight model of a lightweight water quality prediction model with a convolutional neural network and a lightweight anomaly detection model with a long short-term memory network (LSTM) are as follows:

[0237] Parameter and data preparation: Define parameters such as the range, accuracy, and output signal of the sensor, and prepare the original data collected by the sensor, the preprocessed data, etc. Determine the hyperparameters of the lightweight model, such as the learning rate, the number of iterations, and the convolutional kernel size.

[0238] Hardware and software integration: Connect the sensor to the water quality five-parameter monitoring terminal through the corresponding interface to ensure a stable connection and normal signal transmission. In the software system of the monitoring terminal, import the algorithm program of the lightweight model and set the relevant parameters and operating environment of the model.

[0239] Model Training and Calibration: Use a large amount of river water quality data collected by sensors to train the lightweight model and adjust its parameters so that it can accurately process and predict water quality data. Calibrate the sensors and the model according to the actual river water quality situation and known standard data to ensure the accuracy of measurement and prediction.

[0240] Testing and Optimization: Conduct actual on-site tests to check the accuracy, stability of the data, and the performance of the model. Optimize and adjust the system according to the test results, such as adjusting the position of the sensors, optimizing the parameters of the model, etc.

[0241] The parameters and data required for the lightweight model composed of the lightweight water quality prediction model based on convolutional neural network and the lightweight anomaly detection model based on long short-term memory network (LSTM) are as follows:

[0242] Sensor-related parameters: The range, accuracy, resolution, response time, operating temperature range, supply voltage, etc. of each sensor.

[0243] River water quality and environmental data: In addition to the pH value, dissolved oxygen, conductivity, turbidity, and temperature data collected by the sensors, the water level and flow rate data of the river, as well as environmental parameters such as air temperature, air pressure, and illumination at the location of the river are also required.

[0244] Model training data: For the lightweight water quality prediction model based on convolutional neural network and the lightweight anomaly detection model based on LSTM, a large amount of historical water quality data and corresponding labels (such as water quality grades, whether abnormal, etc.) are required for training and verification.

[0245] The formula for the lightweight model composed of the lightweight water quality prediction model based on convolutional neural network and the lightweight anomaly detection model based on long short-term memory network (LSTM) is as follows:

[0246] Mean Squared Error (MSE) formula: Used to measure the difference between the model's predicted value and the true value, and the formula is where y i is the true value, is the predicted value, and n is the number of samples.

[0247] Accuracy formula: For the anomaly detection model, the accuracy can be expressed as where TP is the true positive, TN is the true negative, FP is the false positive, and FN is the false negative.

[0248] In an embodiment of the present invention, the camera acquisition terminal selects a domestic product in the inland river, which is based on the 28+nm Stack process technology, with 50 million pixels, 1 / 1.28 inches, and a pixel size of 1.22μm. Through PixGain The technology has a maximum dynamic range of 88 dB for dual frames and can effectively suppress motion artifacts. It adopts technology with a read noise (RN) < 1e-, which can clearly capture suspended substances in the river channel and water quality anomalies, and has good imaging effects under different lighting conditions.

[0249] In this embodiment, the steps and methods for loading a combined lightweight model of a target detection lightweight model based on YOLO (You Only Look Once) and a semantic segmentation lightweight model based on U-Net are as follows:

[0250] Parameter and data preparation: Define parameters such as the resolution, frame rate, and sensitivity of the camera sensor. Prepare a large amount of image and video data containing suspended substances in the river channel, water quality anomalies, etc. for model training, and annotate information such as the target positions and semantic categories in the data. Determine model hyperparameters such as the learning rate and number of iterations.

[0251] Hardware and software integration: Connect the camera sensor to the acquisition terminal to ensure that the interfaces match and the communication is normal. In the operating system of the acquisition terminal, install the software framework and libraries required for model operation, and import the code and weight files of the lightweight model.

[0252] Model training and calibration: Use the prepared annotated data to train the model, and adjust the model parameters through algorithms such as backpropagation. According to the actual situation of the inland river, calibrate the model, such as adjusting parameters such as the threshold, to make the model adapt to the inland river environment.

[0253] Testing and optimization: Conduct tests at the inland river site to check the clarity of image and video acquisition, and the accuracy of model detection and segmentation. According to the test results, optimize the installation position and angle of the camera, and adjust the model parameters to improve the system performance.

[0254] Parameters and data required for the combined lightweight model of the target detection lightweight model based on YOLO (You Only Look Once) and the semantic segmentation lightweight model based on U-Net:

[0255] Camera sensor parameters: Resolution, frame rate, sensitivity, dynamic range, focal length, field of view angle, etc.

[0256] Inland river environment data: River channel width, water depth, water flow velocity, lighting conditions, weather conditions, etc.

[0257] Model training data: A large amount of inland river images and videos, as well as corresponding target annotation information, semantic segmentation labels, etc.

[0258] Formulas related to the combined lightweight model of the target detection lightweight model based on YOLO (You Only Look Once) and the semantic segmentation lightweight model based on U-Net:

[0259] Intersection over Union (IoU) formula: It is used to measure the overlap degree between the predicted bounding box and the ground truth bounding box in object detection.

[0260] Dice coefficient formula: It evaluates the similarity between the predicted result and the ground truth label in semantic segmentation.

[0261] In an embodiment of the present invention, the main technical indicators of the Hikvision DS-2CD2T46WDV3-I5 camera for the camera acquisition terminal in the small watershed river are as follows: Pixel and resolution: 4 million pixels, maximum resolution 2560×1440. Sensor: 1 / 3" Progressive Scan CMOS, minimum illumination color 0.005Lux@(F1.2, AGC ON). Wide dynamic range: 120dB wide dynamic range. Lens: The focal length has multiple options such as 2.8mm, 4mm, 6mm, etc., and the maximum aperture number is F1.6. Supplementary light: Infrared supplementary light, wavelength 850nm, maximum irradiation distance 50m. Video: The main stream supports H.265 / H.264, and the frame rate is 25fps at 50Hz. Audio: 1 built-in microphone, supporting multiple audio compression standards. Network: 1 RJ45 10M / 100M adaptive network port, supporting multiple network protocols. Power supply: DC 12V±25% or PoE 802.3af, Class 3 power supply. Protection: IP67 protection level, operating temperature -30℃~60℃. It is suitable for the field environment of small watershed rivers.

[0262] The steps and methods for loading the combined lightweight model of the object detection lightweight model based on Faster R-CNN and the semantic segmentation lightweight model based on MobileNetV3 are as follows:

[0263] Parameter and data preparation: Define parameters such as the resolution, frame rate, aperture, and focal length of the camera. Collect image and video data of the small watershed river under different water levels, water flow velocities, and weather conditions, and label information such as river obstacles and water surface anomalies. Determine the model hyperparameters, such as the learning rate and batch size.

[0264] Hardware and software integration: Connect the camera to the acquisition terminal to ensure correct line connection and normal communication. Install relevant software drivers on the acquisition terminal, configure the operating environment, and import the code and weight files of the lightweight model.

[0265] Model training and calibration: Use the labeled data to train the model and make the model converge by adjusting the parameters. Calibrate the model according to the actual situation of the small watershed river, such as adjusting the detection threshold, etc.

[0266] Testing and Optimization: Conduct tests at the small watershed river site to check the image and video acquisition quality and the model detection accuracy. According to the test results, adjust the camera angle and optimize the model parameters to improve the system performance.

[0267] The combined lightweight model with a target detection lightweight model based on Faster R-CNN and a semantic segmentation lightweight model based on MobileNetV3 is as follows:

[0268] The formula for mean average precision (mAP): Used to evaluate the performance of the target detection model, where AP I is the average precision for each class, and n is the number of classes.

[0269] The formula for pixel accuracy (PA): Used for semantic segmentation model evaluation,

[0270]

[0271] In an embodiment of the present invention, the technical indicators of the camera acquisition terminal in the lake are as follows: Pixel specification: Generally 48 million pixels. Photosensitive element size: 1 / 2 inch. Resolution: Can reach about 6000×8000 pixels. Aperture: Usually equipped with a lens to achieve a large aperture, such as about f / 1.7. Focus: Supports fast focusing technologies such as phase focusing. Video shooting: Supports high-frame-rate video shooting, such as 4K 60fps or even higher. Sensitivity: The ISO range is generally between 100-3200, and can be extended in some scenarios. Protection: IP67 protection level, operating temperature -30°C to 60°C. Dynamic range: Has a good dynamic range and can capture rich details under different lighting conditions.

[0272] In this embodiment, the steps and methods for the combined lightweight model with GhostNet lightweight model and MobileMamba lightweight model are as follows:

[0273] Parameter and Data Preparation: Define parameters such as the resolution, frame rate, sensitivity, aperture, and focal length of the camera. Collect a large amount of image and video data of the lake under different seasons, weather, and lighting conditions, and mark the position, type, area of aquatic plants on the lake surface, and information such as abnormal conditions on the water surface. Determine the model hyperparameters, such as the learning rate, batch size, number of iterations, etc.

[0274] Hardware and Software Integration: Connect the camera to the acquisition terminal and ensure a stable connection. In the operating system of the acquisition terminal, install the software framework and libraries required for model operation, and import the code and weight files of the lightweight model.

[0275] Model Training and Calibration: The model is trained using labeled data and converges by adjusting parameters. According to the actual situation of the lake, the model is calibrated, such as adjusting detection thresholds, segmentation parameters, etc.

[0276] Testing and Optimization: Conduct tests at the lake site to check the quality of image and video acquisition and the accuracy of model detection. According to the test results, adjust the installation position and angle of the camera, optimize model parameters, and improve system performance.

[0277] In this embodiment, the parameters and data required for the combined lightweight model of the GhostNet lightweight model and the MobileMamba-based lightweight model:

[0278] Camera Parameters: Resolution, frame rate, sensitivity, dynamic range, aperture, focal length, field of view angle, etc.

[0279] Lake Environmental Data: Lake area, average water depth, water flow velocity, wind direction and speed, light intensity, weather conditions, etc.

[0280] Model Training Data: A large amount of lake image and video data, as well as the labeled information of aquatic plants and water surface conditions on the lake surface.

[0281] The relevant formulas for the GhostNet lightweight model and the MobileMamba-based lightweight model are as follows:

[0282] Formula for the GhostNet lightweight model:

[0283] True Convolution: Let the input feature map be X ∈ R H×W×C , and the true convolution kernel be W real ∈ R K×K×C×C′ , and the original feature map X real is obtained through the true convolution operation. The formula is X real = W real * X, where * represents the convolution operation X real ∈ R H′×W′×C″ , and H′, W′ are the height and width of the convolved feature map.

[0284] Pseudo Convolution: Use the pseudo convolution kernel W ghost ∈ R 1×1×C′×C″ to operate on X real to generate the pseudo feature map X ghost , and the formula is X ghost = W ghost * X real , X ghost ∈ R H′×W′×C″ .

[0285] Output feature map: Concatenate the original feature map and the pseudo-feature map to obtain the final output Y of the Ghost module. The formula is Y = concat(X real , X ghost ), where Y ∈ R H′×W′×(C′×C″) .

[0286] MobileMamba lightweight model formula:

[0287] State update: s t = f(s t -1, x t ), where s_t is the state at the current time t, s t -1 is the state at the previous time, x t is the input at the current time, and f is the state update function, which generally involves matrix multiplication and non-linear activation functions.

[0288] Output calculation: Y t = g(s t ), where Y t is the output at the current time, and g is the output function, which may include operations such as linear transformation and normalization.

[0289] In MobileMamba, these functions f and g are implemented using lightweight convolutions and linear transformations to reduce the computational complexity and the number of parameters.

[0290] In an embodiment of the present invention, the technical indicators of the camera acquisition terminal in the reservoir camera are as follows: Sensor: 1 / 2.7-inch CMOS, Pixel and Resolution: 4 million pixels, 2688×1520 resolution, Scanning Method and Electronic Shutter: Progressive Scanning, 1 / 3s - 1 / 100000s can be manually or automatically adjusted, Illuminance: 0.002Lux (color mode); 0.0002Lux (black and white mode); 0Lux (fill light on), Signal-to-Noise Ratio: >56dB, Fill Light: 1 infrared lamp, Fill Light Distance: 50m, Lens: Fixed Focus, M12 Interface, Focal Length: 3.6mm, Aperture: F1.6, Horizontal Field of View Angle: 84°, Video Compression: H.265, H.264, H.264H, H.264B, MJPEG, Video Frame Rate: 50Hz main stream 2688×1520@25fps, sub-stream 704×576@25fps, Wide Dynamic Range: 120dB, Network: 1 RJ-45 network port, supporting 10M / 100M network, supporting multiple network protocols, Power Supply: DC12V (±30%), POE (802.3af), Working Environment: Temperature -40°C to +60°C, Humidity ≤95%, Protection Grade: IP67.

[0291] In this embodiment, the steps and methods of loading the lightweight model combined with the YOLOv5 lightweight model and the UNet++ lightweight model:

[0292] Parameter and data preparation: Define parameters such as the resolution, frame rate, aperture, focal length, and field of view angle of the camera. Collect image and video data of the reservoir under different seasons, water levels, and weather conditions, and annotate information such as water surface changes and pollutant locations. Determine model hyperparameters, such as the learning rate and weight decay coefficient.

[0293] Hardware and software integration: Correctly connect the camera to the acquisition terminal and power it on. Install the operating system and relevant software drivers on the acquisition terminal, set up the deep learning framework environment, and import the code and pre-trained weights of the lightweight model.

[0294] Model training and calibration: Train the model using the annotated data, and adopt an optimization algorithm to update the model parameters to make them converge. According to the actual scenario of the reservoir, adjust parameters such as the threshold and scale of the model for calibration.

[0295] Testing and optimization: Conduct tests at the reservoir site, observe the quality of the images and videos captured by the camera, and check the accuracy of the model in detecting water surface changes and pollutants. According to the test results, adjust the position and angle of the camera, further optimize the model parameters, and improve the system performance.

[0296] Parameters and data required for the combined lightweight model of YOLOv5 lightweight model and UNet++ lightweight model:

[0297] Camera parameters: Resolution, frame rate, sensitivity, shutter speed, anti-shake performance, etc.

[0298] Reservoir environment parameters: Reservoir capacity, water level change range, inflow rate, outflow rate, surrounding topography and landforms, wind direction, etc.

[0299] Model training data: Reservoir image and video data containing different water surface states, different pollutant types and locations, and corresponding annotation information.

[0300] Formula for the combined lightweight model of YOLOv5 lightweight model and UNet++ lightweight model:

[0301] Formula for Recall: Used to evaluate the detection ability of the object detection model for positive samples, where TP is the true positive and FN is the false negative.

[0302] Formula for Structural Similarity Index (SSIM): Used to measure the similarity between images, where μ x 、μ y are the means, σ X 、σ Y are the standard deviations, σ XY is the covariance, and C1, C2 are constants.

[0303] In an embodiment of the present invention, the camera acquisition terminal has the following technical indicators for river cameras: Imaging: The fog penetration function can cope with bad weather, and the wide dynamic range can adapt to complex lighting. Intelligence: Automatically track moving ships and lock the shooting in real time. Video: Support H.265 and H.264 compression, with frame rates of 25fps or 30fps. Interface: RJ45 network port, optional Wi-Fi, and multiple network protocols. Environment: -40°C to +60°C, humidity ≤ 95%RH. Power supply: DC12V or POE, and the protection level is IP67.

[0304] The steps and methods for loading the lightweight model combined with the EfficientDe lightweight model and the Swin Transformer Lite lightweight model are as follows:

[0305] Parameter and data preparation: Define parameters such as the resolution, frame rate, aperture, focal length, field of view angle, and pan-tilt rotation angle of the camera. Collect image and video data of ships passing by in the river at different times, different weather conditions, and different water flow speeds, and mark information such as the ship position, type, water surface change characteristics, and surrounding environment conditions. Determine the model hyperparameters, such as the initial learning rate, momentum, etc.

[0306] Hardware and software integration: Install the camera in a suitable position and connect it to the acquisition terminal to ensure normal power supply and network transmission. Install the operating system and camera driver on the acquisition terminal, build a deep learning running environment, and import the code and weight files of the lightweight model.

[0307] Model training and calibration: Use the labeled data to train the model and update the model parameters through algorithms such as backpropagation. According to the actual river scenario, adjust parameters such as the detection threshold and classification confidence of the model for calibration.

[0308] Testing and optimization: Conduct tests on the river site, check the quality of the images and videos collected by the camera, and evaluate the detection accuracy of the model for ships and water surface changes, etc. According to the test results, adjust the installation height and angle of the camera, optimize the model parameters, and improve the system performance.

[0309] The parameters and data required for the EfficientDe lightweight model and the Swin Transformer Lite lightweight model are as follows:

[0310] Camera parameters: Resolution, frame rate, sensitivity, dynamic range, anti-shake performance, pan-tilt control parameters, etc.

[0311] River environment parameters: River width, water depth, water flow speed, flow direction, water quality conditions, distribution of surrounding buildings, etc.

[0312] Model training data: A large number of river images and video data containing ships, water surface changes, and surrounding environments, as well as corresponding annotation information.

[0313] The relevant formulas of the EfficientDe lightweight model and the Swin Transformer Lite lightweight model are as follows:

[0314] The loss function formula of EfficientDet is:

[0315] L = L cls + λreg

[0316] where L is the total loss, and L cls is the classification loss, which is used to measure the difference between the predicted class and the true class. Usually, Focal Loss is used to handle the problem of unbalanced positive and negative samples. The formula is FL (t) = -α t (1 - p t ) γ log(p t ), where p t is the predicted probability, and α t and γ are hyperparameters. L cls is the regression loss, which is used to measure the difference between the predicted bounding box and the true bounding box. Usually, Smooth L1 Loss is used. The formula is λ is a hyperparameter used to balance the classification loss and the regression loss.

[0317] Key formulas of Swin Transformer

[0318] Window Self-Attention (WSA): where Q, K, and V are the query matrix, key matrix, and value matrix respectively, and d is the feature dimension. In Swin Transformer Lite, this calculation is performed within each window.

[0319] Shifted Window Self-Attention (SWSA): By shifting the window, there is overlap between adjacent windows, thus fusing more context information. The calculation method is similar to WSA, but the window position has changed.

[0320] Multi-Layer Perceptron (MLP): MLP(x) = GELU(xW1 + b1)W2 + b2, where x is the input feature, W1 and W2 are weight matrices, b1 and b2 are bias terms, and GELU is the activation function.

[0321] In an embodiment of the present invention, the Yunxi full-scale meteorological sensing station 224 (covering meteorological data such as rainfall, wind speed, wind direction, and humidity) is an intelligent meteorological device based on technologies such as the Internet of Things and sensors. It can comprehensively and real-time sense and collect meteorological data such as rainfall, wind speed, wind direction, and humidity, and through data transmission and processing technologies, provide data support for meteorological monitoring, analysis, and early warning, etc. It can be widely used in multiple fields such as meteorology, agriculture, and transportation. The steps and methods for loading the combined lightweight model of the TensorFlow Lite lightweight model and the PyTorch Mobile lightweight model are as follows:

[0322] Model selection and evaluation: According to the specific application scenarios and requirements of the meteorological station, evaluate the performance of TensorFlow Lite and PyTorch Mobile in meteorological data processing tasks, such as accuracy, inference speed, etc.

[0323] Data preparation: Collect and organize data such as rainfall, wind speed, wind direction, and humidity collected by the meteorological station, perform preprocessing such as cleaning and annotation, and divide the training set, validation set, and test set.

[0324] Model configuration and parameter adjustment: For the selected model, adjust the hyperparameters according to the characteristics of meteorological data, such as learning rate, number of iterations, etc.

[0325] Model training and optimization: Use the prepared data to train the model, and use optimization algorithms to minimize the loss function and improve the model performance.

[0326] Model deployment and integration: Deploy the trained model to the hardware device or computing platform of the Yunxi full-scale meteorological sensing station, and integrate it with the data collection, transmission, and processing system of the meteorological station.

[0327] Parameters and data required for loading the combined lightweight model of the TensorFlow Lite lightweight model and the PyTorch Mobile lightweight model:

[0328] Model parameters: Such as the weight parameters, hyperparameters, etc. of the TensorFlow Lite and PyTorch Mobile models.

[0329] Meteorological data: Including historical and real-time rainfall, wind speed, wind direction, humidity data, as well as metadata such as time and location related to these data.

[0330] Hardware parameters: Parameters such as the computing power, memory size, and storage capacity of the sensing station hardware device, which are used to adapt to model deployment.

[0331] Formulas related to the combined lightweight model of the TensorFlow Lite lightweight model and the PyTorch Mobile lightweight model:

[0332] Wind speed and wind force conversion formula: Where V is the wind speed (m /

[0333] s), and B is the wind force.

[0334] Humidity-related formula: Relative humidity Where e is the actual water vapor pressure, E is the saturated water vapor pressure, and the saturated water vapor pressure is related to temperature. The Magnus formula is commonly used for calculation E0 is the saturated water vapor pressure at the reference temperature, T is the temperature, and a and b are constants.

[0335] Rainfall calculation is usually based on data such as the volume of rainwater collected by sensors: Generally, rainwater is collected by a rain gauge for a certain period of time, and the rainfall is represented by measuring the depth of the rainwater. For example, if the depth of the rainwater collected in a certain period is h millimeters, then the rainfall in that period is h millimeters.

[0336] In an embodiment of the present invention, the cloud analysis optical charging energy precision control terminal uses solar energy and an external replaceable energy storage battery as the power supply source, and can continuously and stably provide power support for the AI intelligent connection central rod to achieve intelligent management and control. It adapts to the voltages of various sensors, facilitating the connection of various devices such as external cameras, ultrasonic sensors, radar wave sensors, static pressure sensors, water quality five-parameter terminals, rain gauges integrating temperature, wind speed, and humidity monitoring functions, and Beidou positioning modules, 4G high-gain modules, etc., greatly expanding its application functions.

[0337] The steps and methods for it to load the combined lightweight models of image recognition MobileNetV3, object detection YOLO-Nano, and environmental data prediction LSTM-Light are as follows:

[0338] Hardware evaluation and model screening: Analyze in detail the hardware parameters of the cloud analysis optical charging energy precision control terminal, including the computing power of the processor (such as FLOPS per second), memory size, storage capacity, etc. According to the hardware resource status, combined with the data processing requirements of various sensors, select a suitable model combination from the above lightweight models.

[0339] Model optimization: Quantization: Use a deep learning framework, such as TensorFlow Lite, to convert the 32-bit floating-point parameters in the model into 8-bit integer types. For example, through the linear quantization formula Q = round(S×F + Z), where Q is the quantized value, S is the quantization scale factor, F is the original floating-point value, and Z is the quantization zero point, so as to reduce the model storage requirements and computational complexity.

[0340] Pruning: Remove the connections and neurons in the model that have less impact on the final output result. For example, in a convolutional neural network, evaluate the weights of the convolutional kernels, cut off the connections with smaller weight values, and reduce the model complexity.

[0341] Model integration: Interface adaptation: Write code to implement the docking of the model with various sensor data interfaces. For example, for camera data, write data reading and preprocessing functions to convert the image data into the format required by the model input.

[0342] System integration: Integrate the optimized model into the operating system of the cloud analysis optical charging energy control terminal to ensure that the model can run stably in the hardware environment of the terminal and work in coordination with the data upload module.

[0343] The parameters and data required for the lightweight model combined with Image Recognition MobileNetV3, Object Detection YOLO-Nano, and Environmental Data Prediction LSTM-Light are as follows:

[0344] Sensor data: Camera: A large amount of labeled image data, including images of different scenes and objects, and the annotation content covers information such as object categories and positions.

[0345] Ultrasonic and radar wave sensors: Historical data containing information such as the distance and speed of the target object, as well as the corresponding target attribute annotations.

[0346] Static pressure sensor: Measurement data at different pressure values and the corresponding actual pressure environment description.

[0347] Water quality five-parameter terminal: Historical measurement data of water quality parameters (such as dissolved oxygen, acidity and alkalinity, etc.), and the environmental information of the sampling points.

[0348] Rainfall workstation: Long-term series of temperature, wind speed, and humidity measurement data.

[0349] Beidou positioning module: Historical positioning data, including information such as longitude, latitude, and timestamp.

[0350] Hardware parameters: 1. Processor: Model, main frequency, number of cores, FLOPS, etc. 2. Memory: Capacity, read and write speed. 3. Storage: Capacity, storage type (such as flash memory, hard disk). 4. Sensor characteristic parameters: 5. Measurement range: Such as the maximum detection distance of the radar wave sensor. 6. Accuracy: Such as the measurement accuracy of each parameter in the water quality five-parameter terminal. 7. Resolution: Such as the distance resolution of the ultrasonic sensor.

[0351] The relevant formulas for the lightweight model combined with Image Recognition MobileNetV3, Object Detection YOLO-Nano, and Environmental Data Prediction LSTM-Light are as follows:

[0352] Battery capacity calculation:

[0353] Formula: C = P total ×t×k / V

[0354] C is the battery capacity (Ah); P total is the total power consumption of the terminal device (W), and the power consumption of each module (such as RTU system, sensor, data upload module, etc.) needs to be comprehensively considered; t is the power supply time (h), here it is 20 days, that is, 20×24 = 480h; k is the safety factor, and its value range is generally between 1.2 - 1.5; V is the battery working voltage (V). Through this formula, the battery capacity meeting the 20-day power supply requirement can be calculated.

[0355] Model quantization formula (taking symmetric quantization as an example):

[0356] Formula: q = round(r / s), r = q×s

[0357] q is the quantized integer value, r is the original floating-point value, and s is the quantization step. During the quantization process, floating-point data is converted into integer data through this formula to reduce storage and calculation overhead.

[0358] The AI intelligent connection center sensing pole 200 is the core hardware of this system, with rich interfaces reserved, integrating data collection, transmission, and processing functions. The device is equipped with a camera acquisition terminal 223, various types of sensors, and an intelligent workstation, and is equipped with a five-parameter water quality monitoring terminal 222. It can flexibly select sensors and carry lightweight models according to different water environments such as inland rivers, lakes, and rivers. Its built-in network communication module 230 with a 4G high-gain chip and Beidou positioning module 225 have all-round sensing capabilities; through the cloud analysis optical charging energy fine control terminal 240, the output voltage and current of each terminal and sensor can be intelligently allocated. It can accurately match terminals and sensors according to customer needs and environmental scenarios to ensure efficient data collection and stable transmission.

[0359] In an embodiment of the present invention, the lightweight model set 226 further includes:

[0360] A water pollution monitoring model, which judges the degree and type of water pollution based on the data of the five-parameter water quality monitoring terminal and evaluates the water body health status;

[0361] A hydrodynamic model, which integrates flow velocity and flow rate data, simulates the water flow state of rivers and lakes, and predicts water level changes and water flow directions;

[0362] A target recognition model, which identifies floating objects and ship target objects in the water area based on the data of the camera acquisition terminal;

[0363] A meteorological impact model, which comprehensively analyzes the impact of meteorological factors on the water environment by integrating rainfall, air pressure, temperature, and humidity data;

[0364] The multimodal detection model uses DeepSeek-R1-Distill technology to support camera, radar wave, and ultrasonic fusion detection.

[0365] The time series prediction model uses DeepSeek-V3 technology to process rainfall, temperature, humidity, and water quality data to grasp the temporal trend of the data;

[0366] The anomaly classification model uses TinyBERT to process static pressure sensor data to find anomalies in pressure data.

[0367] The loading method and steps are as follows:

[0368] Hardware adaptation check: Before the model is installed, the hardware configuration of the AI ​​intelligent connection center perception rod is fully checked, including memory, storage capacity, and processor performance, to ensure that it can stably support the operation of the model. At the same time, carefully confirm that the connection between each sensor and the device is normal and the data transmission is stable.

[0369] Model download and storage: Download the required lightweight model files from the professional model library or development platform, store them in the designated area, and reasonably plan the storage path for subsequent calls.

[0370] Model compression: Use the TensorRT quantization tool to convert the FP32 model to FP16 or INT8, reducing the amount of computation and storage requirements and improving the operating efficiency of the model.

[0371] Interface adaptation: With the help of Modbus RTU protocol (function code 0x03 is used to read register data, and 0x06 is used to write single register data), data interaction between sensors and models is realized.

[0372] Model configuration and parameter adjustment: According to the characteristics of different environmental waters and the type of data collected by sensors, the model parameters are adjusted in a targeted manner. For example, when monitoring lakes, the hydrodynamic model parameters are optimized based on the characteristics of the stable water environment; when monitoring inland rivers, the water pollution monitoring model parameters are adjusted according to the characteristics of the water flow and the pollution status.

[0373] Model integration and testing: Integrate the lightweight model with parameter configuration into the control module of the AI ​​intelligent central perception rod for joint debugging and testing. By inputting simulated data or a small amount of actual collected data, check the model operation status and data processing accuracy, and troubleshoot and solve problems in a timely manner.

[0374] Deployment verification: Perform deployment verification to test the model inference delay (target ≤200ms) and memory usage (≤512MB) to ensure that the system meets performance standards and is stable and reliable in actual operation.

[0375] The required parameters and data are as follows:

[0376] Original sensor data: image data collected by camera acquisition terminals, reflected wave data of radar wave sensors, echo time data of ultrasonic sensors, pressure data of static pressure sensors, flow velocity data of flow rate sensors, rainfall data of rainfall sensors, air pressure data of air pressure sensors, temperature and humidity data of temperature and humidity sensors, as well as data such as dissolved oxygen, pH value, temperature, chemical oxygen demand (COD), turbidity, ammonia nitrogen, biochemical oxygen demand, and suspended solids (SS) of the five-parameter water quality monitoring terminal. These original data are the basis for model operation and analysis.

[0377] Environmental parameters: including water area types (inland rivers, lakes, rivers, etc.), geographical location information (obtained through Beidou positioning modules), surrounding topographical and geomorphic data, etc. These parameters help the model more accurately adapt to different environmental scenarios and improve the pertinence and accuracy of data processing.

[0378] Device parameters: hardware parameters of the AI intelligent connection center sensing pole (memory size, processor model and performance parameters, etc.), communication parameters (frequency band, communication protocol, etc. of the 4G high-gain module). These parameters are used to ensure the stable operation of the model on the device and the smooth transmission of data.

[0379] In this embodiment, the combined lightweight model fusion and optimization strategy is as follows:

[0380] Model fusion: In the data preprocessing stage, the data required by the multi-modal detection model, time series prediction model, and anomaly classification model are unified and integrated to reduce repeated acquisition and transmission. For example, align the camera image data, water quality data, static pressure data, etc. according to the time stamp to form a comprehensive data set, providing comprehensive information input for each model.

[0381] Parameter collaborative optimization: Dynamically adjust the model parameters according to the performance of different models in actual operation. For example, during the flood season, increase the weight of the time series prediction model for rainfall data, and at the same time optimize the sensitivity of the multi-modal detection model for detecting floating objects on the water surface to adapt to the complex and changeable water environment.

[0382] Resource scheduling optimization: Use the cloud analysis optical charging energy precision control terminal 240 to dynamically allocate voltage and current according to the resource requirements (such as computing resources, memory resources) during model operation. During the peak period of model inference, prioritize ensuring the models with large computing resource requirements to avoid system performance degradation caused by resource competition.

[0383] Some of the formulas involved in this embodiment are as follows:

[0384] Data transmission rate formula: Where R represents the data transfer rate (bps), S represents the amount of data transferred (bit), and T represents the transfer time (s). This formula is used to measure the efficiency of the device in the data transfer process.

[0385] Formulas related to sensor data calculation: Different sensors have their own data calculation logics. For example, the formula for calculating the flow rate by a flow velocity and flow rate sensor may be Q = v × A, where Q is the flow rate (m 3 / s), v is the flow velocity (m / s), and A is the cross-sectional area of the flow passage (m 2 ).

[0386] Although the above-described embodiments have been described, those skilled in the art can make additional changes and modifications once they learn the basic creative concept. Therefore, the above description is only an embodiment of the present invention, and does not limit the patent protection scope of the present invention. Any equivalent structural or equivalent process transformation made by using the content of the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A method for monitoring water level and water quality by using an intelligent central rod enabled by AI technology, characterized in that: include; Construct a monitoring system consisting of an online monitoring device, a network communication system and a remote monitoring platform, wherein the online monitoring device includes an AI intelligent central sensing rod, a control terminal and a network communication terminal; According to different scenarios, the AI ​​intelligent connection center sensing rod selects different lightweight model combinations; In different scenarios, specific structural assembly and on-site installation methods are used to install the AI ​​intelligent hub sensing pole; The AI ​​intelligent central sensing rod collects data at set time intervals in different scenarios and records relevant environmental parameters and characteristic parameters of specific scenarios; Preprocess the collected data, including removing outliers, setting a reasonable data range to eliminate obviously erroneous data points, using filtering algorithms to remove noise interference, and normalizing or standardizing the data to ensure that data of different parameters are in the same scale range; The preprocessed data is divided into a training set and a test set. The training set data is used to train each model separately. The model parameters are adjusted to minimize the prediction error of the model on the training set. Determine the fusion strategy and fusion weight for the lightweight model combination and perform model fusion; as well as The fused model is deployed to the corresponding monitoring terminal, and the model is verified using the test set data to evaluate the accuracy and stability of the model.

2. According to claim 1, the method for monitoring water level and water quality by using an intelligent central rod enabled by AI technology is characterized in that: The AI ​​intelligent connection central sensing rod according to different scenarios selects different lightweight model combinations including: The radar static pressure composite monitoring water level terminal is equipped with different combined lightweight models in different waters to perform fusion processing on water pressure and water depth data; The five-parameter water quality monitoring terminal is equipped with a combined lightweight model in different waters to correct measurement deviations and predict water quality change trends; The camera acquisition terminal is equipped with a combined lightweight model in different waters to achieve image feature extraction, target recognition and dynamic change analysis: The Yunxi full-scale meteorological sensing station is equipped with a combination of lightweight models of TensorFlow Lite and PyTorch Mobile, which respectively perform real-time feature extraction and short-term time series prediction on meteorological data, providing meteorological background for water level and water quality monitoring; and Yunxi Photovoltaic Charging Energy Precision Control Terminal is equipped with a lightweight model combining MobileNetV3 for image recognition, YOLO-Nano for target detection and LSTM-Light for environmental data prediction. It uses MobileNetV3 to identify the surface status of the solar panels, YOLO-Nano to monitor surrounding environmental threats, and LSTM-Light to combine environmental and equipment data to predict energy supply and demand and optimize energy regulation.

3. The method for monitoring water level and water quality using an intelligent central rod based on AI technology according to claim 2 is characterized in that: The radar static pressure composite water level monitoring terminal is equipped with a combination of lightweight models in different waters, including: A lightweight model combining a deep learning convolutional neural network model and a weighted average model of a traditional algorithm is used in the inland river; A lightweight model combining the machine learning algorithm vector machine (SVM) model and the time series analysis ARIMA model in small watershed rivers; A lightweight model combining a model equipped with a neural network and a model using empirical formulas and statistical analysis in lakes; A lightweight model that combines a physics-based model with a machine learning model on a reservoir; and A lightweight model is built by combining a model based on acoustic principles and a model based on hydraulic principles in rivers.

4. The method for monitoring water level and water quality by using an intelligent central rod enabled by AI technology according to claim 2 is characterized in that: The five-parameter water quality monitoring terminal is equipped with a combination of lightweight models in different waters, including: A lightweight model combining physical and chemical empirical models and deep learning LSTM models in inland rivers; A lightweight model combining a water quality model based on the material balance principle and a deep learning convolutional neural network model in a small river basin; A lightweight model combining a random forest model and a decision tree model is installed on the lake; A lightweight model combining a lightweight sensing model based on data validity identification and duplication elimination and a lightweight model based on knowledge distillation is installed in the reservoir; and A lightweight model is built by combining a lightweight water quality prediction model based on a convolutional neural network and a lightweight anomaly detection model based on a long short-term memory network in rivers.

5. According to claim 2, the method for monitoring water level and water quality by using an intelligent central rod enabled by AI technology is characterized in that: The camera acquisition terminal is equipped with a combined lightweight model in different waters, including: A lightweight model combining a YOLO-based object detection lightweight model and a U-Net-based semantic segmentation lightweight model is installed in the inland river; A lightweight model combining a Faster R-CNN-based target detection lightweight model and a MobileNetV3-based semantic segmentation lightweight model is installed in small watershed rivers; A lightweight model combining GhostNet lightweight model and MobileMamba lightweight model is installed on the lake; Install a lightweight model based on a combination of the YOLOv5 lightweight model and the UNet++ lightweight model in the reservoir; and Jianghe is equipped with a lightweight model based on the combination of the EfficientDe lightweight model and the Swin Transformer Lite lightweight model.

6. A water level and water quality monitoring system based on AI technology empowering the intelligent central pole, characterized in that: include: AI intelligent connection center sensing rods are deployed in inland rivers, small watershed rivers, lakes, reservoirs, and river waters to collect water level, water quality, flow rate, rainfall and video data of rivers, inland rivers, lakes and reservoirs in real time, including: The radar and static pressure composite monitoring water level terminal uses radar and static pressure sensing technology to monitor water area data according to different water area scenarios and combined with corresponding lightweight models; The water quality five-parameter monitoring terminal has a variety of water quality sensors. According to different water environments, matching water quality sensors and lightweight model combinations are selected to achieve real-time monitoring and analysis of five water quality parameters; Camera acquisition terminal, which selects cameras according to different scenarios and carries corresponding lightweight model combinations to collect video image data to identify target objects in the water and monitor abnormal conditions on the water surface; Yunxi full-scale weather sensing station, equipped with a lightweight model combination to collect and process rainfall, wind speed, wind direction, and humidity meteorological data; and A collection of lightweight models. According to different customer needs, environmental scenarios and data characteristics, the AI ​​intelligent hub sensing rod is equipped with a variety of lightweight models; Network communication system, NBIOT, 4G, 5G technology will transmit the collected data to the remote monitoring platform; and Remote monitoring platform presents collected data to users in real time.

7. The water level and water quality monitoring system based on AI technology empowering the intelligent link central pole according to claim 6 is characterized in that: The water quality five-parameter monitoring terminal includes: The first module is used for inland water scenarios and includes one or more of a pH sensor, a dissolved oxygen sensor, a conductivity sensor, a turbidity sensor, and a temperature sensor, and is equipped with a lightweight model combining a physical and chemical empirical model and a deep learning LSTM model; The second module is used for small river basin water scene and includes one or more of pH sensor, dissolved oxygen sensor, conductivity sensor, turbidity sensor, chemical oxygen demand sensor, and a lightweight model combining a water quality model based on material balance principle and a deep learning convolutional neural network model; The third module is used for lake water scenarios and includes one or more of a pH sensor, a dissolved oxygen sensor, a conductivity sensor, a turbidity sensor, and a temperature sensor, and is equipped with a lightweight model combining a random forest model and a decision tree model; A fourth module, which is used in a reservoir water scene and includes one or more of a pH sensor, a dissolved oxygen sensor, a conductivity sensor, a turbidity sensor, and a temperature sensor, and is equipped with a lightweight sensing model based on data validity identification and repeatability screening and a lightweight model combination lightweight model of knowledge distillation; and The fifth module is used for river water scenarios and includes one or more of a pH sensor, a dissolved oxygen sensor, a conductivity sensor, a turbidity sensor, and a temperature sensor, and is equipped with a lightweight water quality prediction model of a convolutional neural network and a lightweight anomaly detection model of a long short-term memory network.

8. The water level and water quality monitoring system based on AI technology empowering intelligent central pole according to claim 7 is characterized in that: The lightweight model combining the random forest model and the decision tree model includes the following formula: Where S DO is the dissolved oxygen saturation, C DO is the actual measured dissolved oxygen concentration, is the saturated dissolved oxygen concentration under given water temperature, air pressure and salinity conditions; K = k*c, where k is a proportional constant, which is related to factors such as the type of ions and temperature, K is conductivity, and c is ion concentration; pH = -log 10 [H + ], where [H + ] is the hydrogen ion concentration.

9. The water level and water quality monitoring system based on AI technology empowering intelligent central pole according to claim 7 is characterized in that: The lightweight model based on data validity identification and repetitive screening and the lightweight model of knowledge distillation combined lightweight model includes the following formula: Used to calculate the difference between two probability distributions, measuring the difference between the output distribution of the student model and the teacher model in knowledge distillation, where P is the soft label probability distribution of the teacher model and Q is the probability distribution predicted by the student model; Used to calculate the hard label loss, where Y i is the one-hot encoding of the true label, is the predicted probability of the student model.

10. The water level and water quality monitoring system based on AI technology empowering intelligent central pole according to claim 6 is characterized in that: The lightweight model set includes: The water pollution monitoring model is configured to determine the degree and type of water pollution based on the data from the five-parameter water quality monitoring terminal and assess the health of the water body; The hydrodynamic model is configured to integrate velocity and flow data, simulate the flow state of rivers and lakes, and predict water level changes and flow direction; The target recognition model is configured to collect terminal data based on the camera and identify floating objects and ship target objects in the water area; The meteorological impact model is configured to integrate rainfall, air pressure, temperature and humidity data to analyze the impact of meteorological factors on the water environment; The multimodal detection model is configured to use DeepSeek-R1-Distill to support camera, radar wave, and ultrasonic fusion detection. The time series prediction model is configured to use DeepSeek-V3 to process rainfall, temperature, humidity, and water quality data to grasp the temporal trend of the data; and The anomaly classification model is configured to use TinyBERT to process the static pressure sensor data to find anomalies in the pressure data.

Images