CN117310088B - Intelligent CO 2 Sensor system and method of operation thereof - Google Patents

Intelligent CO 2 Sensor system and method of operation thereof Download PDF

Info

Publication number
CN117310088B
CN117310088B CN202311091366.XA CN202311091366A CN117310088B CN 117310088 B CN117310088 B CN 117310088B CN 202311091366 A CN202311091366 A CN 202311091366A CN 117310088 B CN117310088 B CN 117310088B
Authority
CN
China
Prior art keywords
sensor
tpa
data
convolution
time step
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active
Application number
CN202311091366.XA
Other languages
Chinese (zh)
Other versions
CN117310088A (en
Inventor
冯天梁
熊行创
韩永刚
赵若凡
刘子龙
金尚忠
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
National Institute of Metrology
Original Assignee
National Institute of Metrology
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by National Institute of Metrology filed Critical National Institute of Metrology
Priority to CN202311091366.XA priority Critical patent/CN117310088B/en
Publication of CN117310088A publication Critical patent/CN117310088A/en
Application granted granted Critical
Publication of CN117310088B publication Critical patent/CN117310088B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Classifications

    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/0004—Gaseous mixtures, e.g. polluted air
    • G01N33/0009—General constructional details of gas analysers, e.g. portable test equipment
    • G01N33/0027—General constructional details of gas analysers, e.g. portable test equipment concerning the detector
    • G01N33/0036—General constructional details of gas analysers, e.g. portable test equipment concerning the detector specially adapted to detect a particular component
    • G01N33/004—CO or CO2
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/0004—Gaseous mixtures, e.g. polluted air
    • G01N33/0006—Calibrating gas analysers
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/0004—Gaseous mixtures, e.g. polluted air
    • G01N33/0009—General constructional details of gas analysers, e.g. portable test equipment
    • G01N33/0062—General constructional details of gas analysers, e.g. portable test equipment concerning the measuring method or the display, e.g. intermittent measurement or digital display
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01N—INVESTIGATING OR ANALYSING MATERIALS BY DETERMINING THEIR CHEMICAL OR PHYSICAL PROPERTIES
    • G01N33/00—Investigating or analysing materials by specific methods not covered by groups G01N1/00 - G01N31/00
    • G01N33/0004—Gaseous mixtures, e.g. polluted air
    • G01N33/0009—General constructional details of gas analysers, e.g. portable test equipment
    • G01N33/007—Arrangements to check the analyser

Landscapes

  • Chemical & Material Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Medicinal Chemistry (AREA)
  • Food Science & Technology (AREA)
  • Combustion & Propulsion (AREA)
  • Physics & Mathematics (AREA)
  • Analytical Chemistry (AREA)
  • Biochemistry (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Immunology (AREA)
  • Pathology (AREA)
  • Testing Or Calibration Of Command Recording Devices (AREA)

Abstract

The invention provides an intelligent CO 2 Sensor system and method of operation thereof, the system comprising CO 2 Sensor node for atmospheric airCO of (c) 2 Monitoring and calibrating monitoring data by utilizing a TPA-GRU model; CO 2 Reference station for CO in the atmosphere 2 Monitoring; data communication network for communicating CO 2 Reference station and CO 2 Monitoring data of the sensor nodes are sent to an edge computing server; and the edge computing server is used for preprocessing the received monitoring data and training the TPA-GRU model based on the preprocessing result. The invention can dynamically capture the interaction characteristics among a plurality of sensors in the sensor node in a single time step and the time sequence change characteristics of each sensor in a time window by adopting the TPA-GRU model, thereby leading the calibrated CO to be 2 The sensor maintains good accuracy and stability in complex, dynamic scenarios.

Description

Intelligent CO 2 Sensor system and method of operation thereof
Technical Field
The invention belongs to CO 2 Sensor technical field especially relates to an intelligent CO 2 A sensor system and a method of operating the same.
Background
Acceleration of industrialization and urbanization to carbon dioxide CO 2 The discharge amount is continuously increased, which poses a threat to the environment and human health. For real-time monitoring and control of CO 2 Is disposed of in critical areas by government and related institutions 2 And (5) a monitoring station. However, conventional high precision CO 2 The reference station has high construction cost, difficult maintenance and huge volume, and is not suitable for large-scale deployment and application. Thereby CO 2 The sensor is generated, but CO 2 The sensor is limited by factors such as principle, cost and the like, has relatively poor monitoring accuracy, and needs to be calibrated after deployment to improve the monitoring accuracy.
In CO 2 During the calibration process of the sensor, weather and other gas sensors are generally introduced as auxiliary variablesWith CO 2 The sensors together form a CO 2 Sensor node, CO 2 Sensor node and high-precision CO 2 Monitoring CO of a device 2 The CO-location of the reference station collects monitoring data for generating a calibration model that reflects the CO 2 Sensor monitoring value and CO 2 Mapping relation of reference station monitoring values. After calibration is completed, CO 2 The monitoring data of each sensor in the sensor nodes is input into a calibration model, and the model outputs CO in the nodes 2 Calibration values of the sensor. But CO 2 Sensor and CO 2 The monitored values of both reference stations change over time, both exhibit a linear or nonlinear relationship that is not unity, so to further improve the performance of the calibration model, researchers have begun to use algorithms based on deep learning on CO 2 The sensor is calibrated.
The existing calibration model sensor based on time sequence focuses more on the hidden relation of characteristic data acquired by sensors among various step sizes, and in a single time step, a plurality of sensors acquire data, so that the single time step contains a plurality of pieces of information, the hidden relation among the plurality of sensors is ignored by a self-attention mechanism in the sensor, and noise in the data acquired by the plurality of sensors cannot be avoided; at the same time, the self-attention mechanism of the method "averages" the information for a plurality of time steps and cannot detect a time pattern that is favorable for the sensor calibration values. The cyclic window skip module in another calibration model deep cm must manually adjust the skip length to match the period of a particular contaminant, and furthermore deep cm does not take into account the effect of the variation law of individual features over the time step on other features, particularly the target contaminant feature.
Thus using the two calibration models above 2 The sensor has the problem of insufficient accuracy and stability when in a complex and dynamic environment.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides an intelligent CO 2 A sensor system and a method of operating the same,thereby realizing CO 2 Calibration of the sensor such that the calibrated CO 2 The sensor is deployed in a complex and dynamic environment, still has higher stability and accuracy, and can generate an accurate monitoring value for reference by people.
In order to achieve the above purpose, the invention adopts the following technical scheme:
the scheme provides an intelligent CO 2 A sensor system, comprising:
CO 2 sensor node for CO in the atmosphere 2 Monitoring, loading and running a trained TPA-GRU model, and calibrating the monitoring data using the TPA-GRU model, wherein the CO 2 The sensor node is arranged at the CO 2 A reference station is beside;
CO 2 reference station for CO in the atmosphere 2 Monitoring;
data communication network for communicating CO 2 Reference station and CO 2 Monitoring data of the sensor nodes are sent to an edge computing server;
and the edge computing server is used for preprocessing the received monitoring data and training the TPA-GRU model based on the preprocessing result.
The beneficial effects of the invention are as follows: the invention uses CO 2 The sensor node is arranged at the CO 2 By reference station, use it to monitor CO in the atmosphere 2 The interaction characteristics among a plurality of sensors inside the sensor node in a single time step and the time sequence change characteristics of each sensor in a time window can be dynamically captured by using the TPA-GRU model to calibrate the monitoring data, so that the calibrated CO 2 The sensor maintains good accuracy and stability in complex, dynamic scenarios.
Further, the CO 2 The sensor node includes:
a controller for setting CO 2 Sampling frequency of sensor, temperature sensor, humidity sensor and pressure sensor, and storing CO 2 Sensor, temperature sensor, humidity sensor and pressure sensor are produced by AD conversion elementMonitoring data, and controlling communication of the networking element with the edge computing server and loading and running the trained TPA-GRU model;
the networking element is used for carrying out data communication with the edge computing server by adopting the MQTT protocol;
AD conversion element for converting CO 2 Analog signals generated by the reaction of the sensor, the temperature sensor, the humidity sensor and the pressure sensor with the environment are converted into digital monitoring data, and the digital monitoring data are stored in the controller.
The beneficial effects of the above-mentioned further scheme are: CO 2 The sensor node is a sensor array comprising a plurality of sensors capable of monitoring the target gas CO 2 Other environmental variables, such that the TPA-GRU model takes into account more environmental factors, implementing the CO 2 The sensor is calibrated more accurately. Furthermore, networking elements employing the MQTT protocol ensure stable and efficient data communications, particularly in low bandwidth or unstable deployment environments.
Still further, the CO 2 The reference station includes:
an in-station server for transmitting data with the edge computing server and storing CO 2 Monitoring data generated by the monitoring equipment and periodically or selectively developing the monitoring data to an edge computing server;
CO 2 monitoring device for CO in atmosphere 2 Monitoring and storing the data in an in-station server.
The beneficial effects of the above-mentioned further scheme are: CO by in-station server 2 The reference station realizes the centralized storage and management of data, and meanwhile, the in-station server can select to send data to the edge computing server periodically or according to the need, so that the data transmission is more flexible and efficient.
Still further, the TPA-GRU model comprises:
GRU gating circulation unit module for according to CO 2 Obtaining a hidden characteristic of a data set generated by a sensor node, wherein the hidden characteristic is divided into a past time step hidden characteristic and a T time step hidden characteristic;
a Conv2D two-dimensional convolution module for calculating a plurality of convolution values according to the past time step, and combining the convolution values to form a convolution tensor, wherein each convolution valueCorresponding CO 2 Time sequence change characteristics of the sensor, the temperature sensor, the humidity sensor and the pressure sensor in a time window;
a TPA time dimension attention module for receiving the convolution tensor and the T time step hidden characteristic, and a scoring function for determining the time step hidden characteristic according to the CO 2 The i' th sensor in the sensor node scores the calibration value and the relation with the hidden characteristic of the current time step and outputs the weight alpha through an activation function i” ；
An MLP multi-layer sensor module for sensing the weight alpha i” And convolution tensor, calculating to obtain weight vector v t And integrate the T-th time step hidden feature and the weight vector v t And outputting the calibration value of the T time step.
The beneficial effects of the above-mentioned further scheme are: compared with an LSTM long-term memory network, the GRU gating circulating unit has fewer parameters and high performance, and is more suitable for terminal deployment; in the use scene of the invention, compared with a high-speed network of a high-speed network, the MLP multi-layer sensor has simpler structure, low resource consumption and high operation efficiency.
The invention provides an intelligent CO 2 A method of operating a sensor system, comprising the steps of:
s1, CO 2 Sensor node deployment to CO 2 Beside a reference station and using CO 2 Sensor and CO 2 Reference station to CO in the atmosphere 2 Monitoring and storing;
s2, CO 2 Sensor node and CO 2 The data monitored by the reference station are sent to an edge computing server;
s3, preprocessing the monitoring data by utilizing an edge computing server to obtain a training set and a testing set;
s4, training the TPA-GRU model with the initial weight preset by using a training set, and testing the TPA-GRU model by using a testing set to obtain a trained TPA-GRU model;
s5, transmitting the trained TPA-GRU model to the CO through a data communication network by utilizing an edge computing server 2 Sensor node and use CO 2 The sensor node loads and runs the trained TPA-GRU model;
and S6, calibrating the monitoring data by utilizing the TPA-GRU model, and outputting a calibration value.
The beneficial effects of the invention are as follows: the invention uses CO 2 The sensor node is arranged at the CO 2 By reference station, use it to monitor CO in the atmosphere 2 The interaction characteristics among a plurality of sensors inside the sensor node in a single time step and the time sequence change characteristics of each sensor in a time window can be dynamically captured by using the TPA-GRU model to calibrate the monitoring data, so that the calibrated CO 2 The sensor maintains good accuracy and stability in complex and dynamic scenes, thereby realizing the aim of CO 2 Calibration of the sensor such that the calibrated CO 2 The sensor is deployed in a complex and dynamic environment, still has higher stability and accuracy, and can generate an accurate monitoring value for reference by people.
Further, the step S3 includes the steps of:
s301, merging the monitoring data by utilizing an edge computing server according to the same timestamp, deleting time step data comprising null values and outliers, and constructing a data set D;
s302, dividing the data set D into a training set and a testing set.
The beneficial effects of the above-mentioned further scheme are: the edge computing server automatically merges the monitoring data according to the same time stamp, so that errors caused by manually merging the data are eliminated; and deleting the time step data of the null value and the abnormal value, so that the quality of the data set is improved, and the robustness of the model is enhanced.
Still further, the step S4 includes the steps of:
s401 extracting continuous tau time step CO from training set 2 Monitoring data of the sensor node;
s402, inputting the extracted monitoring data into a pre-set TPA-GRU model with initial weight to obtain calibration values of tau time steps
S403, measuring the calibration value by using the loss functionWith corresponding CO 2 Data y collected by reference station τ Is the gap between (1);
the expression of the loss function is as follows:
wherein L is MAE (Θ) represents a loss function, Θ represents a parameter that can be learned, n represents a total time step;
s404, minimizing the gap through an optimization algorithm Adam;
s405, judging whether the minimized gap is unchanged in a preset training round, finishing training of the TPA-GRU model, saving the network weight with the minimum loss function value, and entering a step S406, otherwise, returning to the step S401;
s406, testing the TPA-GRU model by using the test set to obtain a trained TPA-GRU model.
The beneficial effects of the above-mentioned further scheme are: the MAE loss function gives the same weight to each sample error, which makes the model more focused on those samples with larger errors; the optimization algorithm Adam enables the model to converge more quickly, and the generalization capability of the model can be improved; the stopping condition of the training round is set to ensure that the model is not excessively trained, so that the overfitting is avoided, and the waste of calculation force resources is avoided; after training, the test set is used for testing, so that the effectiveness and accuracy of the model are ensured.
Still further, the step S6 includes the steps of:
s601, inputting CO of tau time step including T time step and history time step 2 Monitoring data by the sensor nodes to generate a data set;
s602, inputting the generated data set into a GRU (gate-controlled loop) unit module to obtain hidden features;
s603, dividing the hidden feature into a past time step hidden feature and a T time step hidden feature;
s604, transmitting the hidden characteristic of the past time step to a Conv2D two-dimensional convolution module, and carrying out convolution operation with k filters in the Conv2D two-dimensional convolution module to obtain a plurality of convolution valuesWherein each convolution value->Corresponding CO 2 The time sequence change characteristics of each sensor in the sensor nodes in a time window;
wherein h is i',(t-τ-1+l) Representing the hidden form characteristic of the monitoring value of the ith sensor in the CO2 sensor node at the t-tau-1+l time step, and x represents convolution operation and C j,T-τ+l A j-th convolution filter with the length T is represented, and l represents a first convolution operation sequence in a time sequence window;
s605, combining the convolution values to form a convolution tensorAnd inputting the convolution tensor and the T-th time step hidden feature to the TPA time dimension attention module;
s606, a scoring function in the TPA time dimension attention module is used for scoring the data according to the CO 2 The ith sensor in the sensor node scores the relation between the calibration value and the hidden characteristic of the current time step and outputs the weight alpha through a sigmoid activation function i” ；
The expression of the scoring function is as follows:
wherein f (·) represents a scoring function, h T Representing a T-th time step concealment feature, W a Weights representing scoring functions;
the weight alpha i” The expression of (2) is as follows:
wherein sigmoid (·) represents a sigmoid activation function;
s607 according to the weight alpha i” Convolution tensorCalculating to obtain a weight vector v t ：
Wherein m represents CO converted by GRU (grid-controlled loop) unit module 2 The hidden characteristic quantity of the sensor in the sensor node, i represents the product operation order of the ith row convolution tensor and the corresponding ith weight;
s608, integrating the T-th time step hidden feature and the weight vector v by using the MLP multi-layer sensor module t Outputting the calibration value of the T time step
Wherein W is h And W is v All represent weights, and MLP (·) represents the MLP multi-layer perceptron.
The beneficial effects of the above-mentioned further scheme are: by combining the TPA time dimension attention module and the Conv2D two-dimensional convolution module, the interaction and time sequence change characteristics among different sensors are accurately captured; the MLP multi-layer sensor module further integrates and outputs the calibration value, and the processing mode not only improves the model performance, but also simplifies the processing flow, thereby being beneficial to improving the real-time performance and the corresponding speed of the system.
Drawings
Fig. 1 is a schematic diagram of a system structure according to the present invention.
FIG. 2 is a schematic diagram of TPA-GRU model structure in the present invention.
FIG. 3 is a flow chart of a method of operation of the present invention.
Detailed Description
The following description of the embodiments of the present invention is provided to facilitate understanding of the present invention by those skilled in the art, but it should be understood that the present invention is not limited to the scope of the embodiments, and all the inventions which make use of the inventive concept are protected by the spirit and scope of the present invention as defined and defined in the appended claims to those skilled in the art.
Example 1
As shown in FIG. 1, the present invention provides an intelligent CO 2 A sensor system, comprising:
CO 2 sensor node for CO in the atmosphere 2 Monitoring, loading and running a trained TPA-GRU model, and calibrating the monitoring data using the TPA-GRU model, wherein the CO 2 The sensor node is arranged at the CO 2 A reference station is beside;
CO 2 reference station for CO in the atmosphere 2 Monitoring;
data communicationA communication network for communicating CO 2 Reference station and CO 2 Monitoring data of the sensor nodes are sent to an edge computing server;
and the edge computing server is used for preprocessing the received monitoring data and training the TPA-GRU model based on the preprocessing result.
As shown in FIG. 1, the CO 2 The sensor node includes:
a controller for setting CO 2 Sampling frequency of sensor, temperature sensor, humidity sensor and pressure sensor, and storing CO 2 Monitoring data generated by the sensors, temperature sensors, humidity sensors and pressure sensors through the AD conversion element, and controlling communication of the networking element and the edge computing server, and loading and running the trained TPA-GRU model;
the networking element is used for carrying out data communication with the edge computing server by adopting the MQTT protocol;
AD conversion element for converting CO 2 Analog signals generated by the reaction of the sensor, the temperature sensor, the humidity sensor and the pressure sensor with the environment are converted into digital monitoring data, and the digital monitoring data are stored in the controller.
As shown in FIG. 1, the CO 2 The reference station includes:
an in-station server for transmitting data with the edge computing server and storing CO 2 Monitoring data generated by the monitoring equipment and periodically or selectively developing the monitoring data to an edge computing server;
CO 2 monitoring device for CO in atmosphere 2 Monitoring and storing the data in an in-station server.
As shown in fig. 2, the TPA-GRU model includes:
GRU gating circulation unit module for according to CO 2 Obtaining a hidden characteristic of a data set generated by a sensor node, wherein the hidden characteristic is divided into a past time step hidden characteristic and a T time step hidden characteristic;
a Conv2D two-dimensional convolution module for calculating a plurality of convolution values according to the past time step and for calculating each volumeThe product values are combined to form a convolution tensor, wherein each convolution valueCorresponding CO 2 Time sequence change characteristics of the sensor, the temperature sensor, the humidity sensor and the pressure sensor in a time window;
a TPA time dimension attention module for receiving the convolution tensor and the T time step hidden characteristic, and a scoring function for determining the time step hidden characteristic according to the CO 2 The ith sensor in the sensor node scores the calibration value and the relation with the hidden characteristic of the current time step and outputs the weight alpha through an activation function i” ；
An MLP multi-layer sensor module for sensing the weight alpha i” And convolution tensor, calculating to obtain weight vector v t And integrate the T-th time step hidden feature and the weight vector v t And outputting the calibration value of the T time step.
In this embodiment, as shown in FIGS. 1-2, CO is used 2 Sensor nodes are arranged in CO 2 The reference station 20 is located beside (i.e. CO-located) to collect monitoring data, the monitoring data of the two are sent to the edge computing server through the data communication network, the edge computing server pre-processes the received monitoring data and divides the monitoring data into a training set and a testing set, the training set and the testing set are utilized to train and optimize the pre-configured initialized TPA-GRU model, the trained TPA-GRU model is generated and stored, and the trained TPA-GRU model is sent to the CO 2 Sensor node, CO 2 The sensor nodes load and run the trained TPA-GRU model, followed by CO 2 The sensor node will be sensitive to CO in the atmosphere 2 Monitoring and CO using TPA-GRU model 2 The sensor is calibrated. The calibration process comprises the following steps: first, a TPA-GRU model utilizes a GRU gating cycle unit module to extract CO in a single time step 2 Interaction characteristics of each sensor in the sensor node are extracted through a Conv2D two-dimensional convolution module, time sequence change characteristics of each sensor in a time window are extracted, and a TPA time dimension attention module gives different degrees of influence of each characteristic on a calibration resultWeighting, and finally integrating the characteristics through an MLP (multi-layer perceptron) module to generate CO (carbon monoxide) 2 Calibration values of the sensor.
As shown in fig. 2, the TPA-GRU model according to the present invention is composed of a GRU-gated circulation unit module, a Conv2D two-dimensional convolution module, a TPA time dimension attention module, and an MLP multi-layer sensor module.
The GRU gating cycle unit module is of a single-layer GRU structure and consists of 1 GRU unit, wherein the GRU unit comprises 1 update gate and 1 reset gate. Updating the gate controls how the hidden state of the previous time step remains to the current time step, and resetting the gate controls how new input information is combined with the hidden state of the previous time step. This mechanism can effectively extract the interaction characteristics between multiple sensors inside the sensor node in a single time step.
Conv2D two-dimensional convolution module: comprises 1 two-dimensional convolution layer comprising 32 convolution filters, the convolution kernel having a size of 1 x 12. The module can effectively extract the time sequence change characteristics of each sensor in a time window.
TPA time dimension attention module: based on the attention mechanism, the system consists of an attention weight calculation layer and a weighted characteristic integration layer. The input features of each time step are assigned weights, the weight values depending on how well the input features match the attention parameters inside the module. Thus, the module can give different weights according to the extent to which the row vector affects the calibration result.
MLP multilayer perceptron module: comprises a linear layer and a Relu activation function for non-linearly transforming the input features and outputting final calibration values.
As shown in FIG. 1, is an intelligent and precise CO 2 Sensor system schematic diagram, system can be divided into CO 2 Sensor node, CO 2 Reference station and edge calculation server.
The function of each structure is described below:
CO 2 the sensor nodes include controllers (e.g., MSP430, PIC controller, arduino, raspberry)Pi, etc., is critical to control the operation of the sensor, store and process the data generated by the sensor, and have a certain computational load and run the trained TPA-GRU model, CO 2 Sensors, temperature sensors, humidity sensors, pressure sensors, AD conversion elements, and networking elements.
In this embodiment, the controller is responsible for controlling the CO 2 The system comprises sensors, temperature sensors, humidity sensors, pressure sensors, AD conversion elements and networking elements, and is characterized in that the sampling frequency of each sensor is set, monitoring data generated by each sensor through the AD conversion elements is stored, the networking elements and an edge computing server are controlled to carry out data transmission, and the system has enough calculation force loading and runs a trained TPA-GRU model.
The AD conversion element is responsible for converting CO 2 The analog voltage generated by the reaction of the sensor, the temperature sensor, the humidity sensor and the pressure sensor with the environment is converted into a digital signal.
The networking element adopts MQTT protocol to be responsible for data communication with the edge computing server, and MQTT (message queue telemetry transmission) is a lightweight communication protocol based on a publish/subscribe mode, and is suitable for remote or network bandwidth-smaller scenes. In terms of data transmission, it will CO 2 Data generated by the sensor nodes are sent through being issued to a specific MQTT theme; in terms of data reception, it subscribes to a specific topic to receive trained TPA-GRU models from edge computing servers. Both the transmission and reception of data takes place over a data communications network.
In this embodiment, CO 2 The reference station comprises high precision CO 2 Monitoring equipment (such as high-cost measuring instruments based on cavity-surrounding spectral absorption technology, tuned diode laser absorption spectroscopy) and in-station servers.
The in-station server is responsible for data transmission with the edge computing server and stores high-precision CO 2 Monitoring values generated by the monitoring device and periodically or selectively sending data to the edge computing server.
Edge computing server is responsible for CO 2 Sensor node, CO 2 Reference stationAnd transmitting data, receiving monitoring data, preprocessing and dividing the monitoring data, and training a pre-configured initialization TPA-GRU model with enough power until convergence.
The invention uses CO 2 The sensor node is arranged at the CO 2 By reference station, use it to monitor CO in the atmosphere 2 The interaction characteristics among a plurality of sensors inside the sensor node in a single time step and the time sequence change characteristics of each sensor in a time window can be dynamically captured by using the TPA-GRU model to calibrate the monitoring data, so that the calibrated CO 2 The sensor maintains good accuracy and stability in complex and dynamic scenes, thereby realizing the aim of CO 2 Calibration of the sensor such that the calibrated CO 2 The sensor is deployed in a complex and dynamic environment, still has higher stability and accuracy, and can generate an accurate monitoring value for reference by people.
Example 2
As shown in FIG. 3, the present invention provides an intelligent CO 2 A method of operating a sensor system, the method comprising:
s1, CO 2 Sensor node deployment to CO 2 Beside a reference station and using CO 2 Sensor and CO 2 Reference station to CO in the atmosphere 2 Monitoring and storing;
s2, CO 2 Sensor node and CO 2 The data monitored by the reference station are sent to an edge computing server;
s3, preprocessing the monitoring data by utilizing an edge computing server to obtain a training set and a testing set, wherein the implementation method is as follows:
s301, merging the monitoring data by utilizing an edge computing server according to the same timestamp, deleting time step data comprising null values and outliers, and constructing a data set D;
s302, dividing the data set D into a training set and a testing set;
s4, training the TPA-GRU model with the initial weight set in advance by utilizing a training set, and testing the TPA-GRU model by utilizing a testing set to obtain a trained TPA-GRU model, wherein the implementation method is as follows:
s401 extracting continuous tau time step CO from training set 2 Monitoring data of the sensor node;
s402, inputting the extracted monitoring data into a pre-set TPA-GRU model with initial weight to obtain calibration values of tau time steps
S403, measuring the calibration value by using the loss functionWith corresponding CO 2 Data y collected by reference station τ Is the gap between (1);
s405, judging whether the minimized gap is unchanged in a preset training round, finishing training of the TPA-GRU model, saving the network weight with the minimum loss function value, and entering a step S406, otherwise, returning to the step S401;
s406, testing the TPA-GRU model by using a test set to obtain a trained TPA-GRU model;
s5, transmitting the trained TPA-GRU model to the CO through a data communication network by utilizing an edge computing server 2 Sensor node and use CO 2 The sensor node loads and runs the trained TPA-GRU model;
s6, utilizing a TPA-GRU model to calibrate the monitoring data and outputting a calibration value, wherein the implementation method is as follows:
s601, inputting CO of tau time step including T time step and history time step 2 Monitoring data by the sensor nodes to generate a data set;
s602, inputting the generated data set into a GRU (gate-controlled loop) unit module to obtain hidden features;
s603, dividing the hidden feature into a past time step hidden feature and a T time step hidden feature;
s604, transmitting the hidden characteristic of the past time step to Conv2D two-dimensional convolution module, and carrying out convolution operation with k filters in Conv2D two-dimensional convolution module to obtain a plurality of convolution valuesWherein each convolution value->Corresponding CO 2 The time sequence change characteristics of each sensor in the sensor nodes in a time window;
wherein h is i',(t-τ-1+l) Representing the hidden form characteristic of the monitoring value of the ith sensor in the CO2 sensor node at the t-tau-1+l time step, and x represents convolution operation and C j,T-τ+l A j-th convolution filter with the length T is represented, and l represents a first convolution operation sequence in a time sequence window;
s605, combining the convolution values to form a convolution tensorAnd inputting the convolution tensor and the T-th time step hidden feature to the TPA time dimension attention module;
s606, a scoring function in the TPA time dimension attention module is used for scoring the data according to the CO 2 The ith sensor in the sensor node scores the relation between the calibration value and the hidden characteristic of the current time step and outputs the weight alpha through a sigmoid activation function i” ；
The expression of the scoring function is as follows:
wherein f (·) represents a scoring function, h T Representing a T-th time step concealment feature, W a Weights representing scoring functions;
the weight alpha i” The expression of (2) is as follows:
wherein sigmoid (·) represents a sigmoid activation function;
s607 according to the weight alpha i” Convolution tensorCalculating to obtain a weight vector v t ：
Wherein m represents CO converted by GRU (grid-controlled loop) unit module 2 The hidden characteristic quantity of the sensor in the sensor node, i represents the product operation order of the ith row convolution tensor and the corresponding ith weight;
s608, integrating the T-th time step hidden feature and the weight vector v by using the MLP multi-layer sensor module t Outputting the calibration value of the T time step
Wherein W is h And W is v All represent weights, and MLP (·) represents the MLP multi-layer perceptron.
In this embodiment, CO 2 Sensor node and CO 2 Co-locating and collecting data by a reference station: CO 2 Sensor node deployment at CO 2 Beside a reference station (also called CO-location), both simultaneously for CO in the atmosphere 2 Monitoring is carried out for a certain period of time and monitoring data (CO 2 The sensor, the temperature sensor, the humidity sensor and the pressure sensor are connected through AD conversion elementsThe part converts the analog signal into digital signal monitoring data and stores the digital signal monitoring data in a storage unit in the controller, and the CO 2 Reference station through high precision CO 2 The monitoring device obtains the monitoring data and stores the data in the in-station server).
In this embodiment, the collected data is sent to an edge computing server: after a period of monitoring, the monitoring data of the two are sent to an edge computing server, and CO 2 The sensor nodes communicate with the edge computing server via a data communication network using networking elements, CO 2 The reference station communicates with the edge computing server over a data communication network using an in-station server.
In the embodiment, the edge computing server preprocesses the monitoring data, namely, merging the data, deleting null values and abnormal values, and dividing the data into a training set and a testing set: the edge computing server combines the received monitoring data according to the same time stamp, deletes time step data containing null value and abnormal value to ensure the stability of the trained calibration model, and constructs a data set D= { X, Y }, wherein X= { X 1 ,x 2 ,x 3 ,…,x n Is CO 2 The sensor node collects data, x n Representing the monitoring data acquired at the nth time step, and x n ＝{x 1,n ,x 2,n ,x 3,n ,…,x d,n X, where x d,n Representing the nth time step CO 2 Acquisition data of the d-th sensor of the sensor node, y= { Y 1 ,y 2 ,y 3 ,…,y n Is CO 2 Monitoring data collected by a reference station, wherein y n Representing the nth time step CO 2 The reference station collects monitoring data and divides the data set D into a training set and a testing set.
In this embodiment, the edge computing server uses the dataset to train and optimize the pre-configured TPA-GRU model: in the edge computation server, the TPA-GRU model with initialization weights has been pre-configured (using a deep learning framework such as TensorFlow, pyTorch, keras, etc.), and the training set is used to train the TPA-GRU model. At each training time, successive τ time steps are extracted from the training setLong CO 2 Sensor nodes collect monitoring dataIn this embodiment, the parameter τ is set to 30./>Inputting TPA-GRU model, TPA-GRU model outputs calibration value of τ time step +.>Loss function measurement calibration value->With corresponding CO 2 Data y collected by reference station τ By using an optimization function to minimize the loss, updating the parameters of the TPA-GRU model by back propagation, thereby completing one training. The training process will continue until the performance of the model reaches a preset standard, or the error has not significantly degraded, at which point the model has converged. When training is carried out, an optimization function is Adam, the initial learning rate is set to be 0.003, the batch processing size is 32, and in order to prevent the model from being over-fitted, a dropout method is set, and the dropout probability is 0.5. Using the mean absolute error (mean absolute error, MAE) as a loss function, defined as:
wherein L is MAE (Θ) represents a loss function, Θ represents a parameter that can be learned, and n represents the total time step.
The loss function is used for measuring the difference between the calibration value and the true value generated by the TPA-GRU model in the TPA-GRU model training stage, minimizing the difference through an optimization algorithm Adam and updating the parameters of the model through back propagation. Adam automatically halves the learning rate whenever the current minimum of the trained penalty values does not change within the last 8 training rounds. And stopping training and storing the network weight with the minimum loss function value if the current minimum value of the loss values in the 25 training rounds does not change or the total training round number exceeds 250. After model training is completed, the performance of the model will be evaluated using the test set to ensure that the performance of the model in an unknown environment is stable.
In this embodiment, the TPA-GRU model after preservation training is sent to the CO 2 Sensor node: after the TPA-GRU model training and testing process is finished, automatically storing the trained TPA-GRU model, and transmitting the TPA-GRU model to the CO through a data communication network by an edge computing server 2 And a sensor node.
In this embodiment, CO 2 The sensor node loads and runs the trained TPA-GRU model: CO 2 The sensor node loads and operates the received trained TPA-GRU model on a controller, and the controller is preconfigured with the environment required for operating the TPA-GRU model, and the CO 2 The sensor nodes continuously collect and store the monitoring data.
In this embodiment, TPA-GRU model calibrated CO 2 And (3) outputting a calibration result according to the monitoring value of the sensor: monitoring data is input into a TPA-GRU model, and the TPA-GRU model outputs CO 2 Calibration values of the sensor. Specifically, it is assumed that CO at time T is to be calibrated 2 The monitoring value of the sensor only needs to input the CO comprising the T time step and the historical time step which is up to tau time steps 2 Sensor node monitoring data, generating an input datasetWherein { x T-τ+1 ，x T-τ+2 ，x T-τ+3 ，…，x T-1 Is historical time step data, x T For the T-th time step data, in this example, τ=30 (consistent with the size set during TPA-GRU model training) is set, and the TPA-GRU model calibration CO is shown in fig. 1 2 The specific process of the monitoring value of the sensor is as follows:
first, input data setThe GRU is input into a gating circulating unit module, and the gating circulating unit module outputs hidden form characteristics +.>Where m represents a hidden feature number, τ represents a time span, and m=12 is set in this embodiment;
hidden form features to be outputDividing into past time step hidden featuresAnd T-th time step hidden feature h T ∈R m×1 Past time step concealment feature +.>In the incoming Conv2D two-dimensional convolution module, k filters C in the module are used for filtering j ∈R 1×(T-1) Convolving j=1, 2, 3..k, wherein T-1 is the time window length, and in this embodiment, setting t=τ, the convolution value +.>
Wherein,representation->The convolution value obtained by the cross-correlation operation of the ith row and the jth filter.
The input information is further refined through Conv2D two-dimensional convolution module operation, and each convolution value generated corresponds to CO 2 Time sequence variation of each sensor in sensor node in time windowFeatures. Each convolution valueCombining to form a convolution tensor H C ∈R m×k In the ith row vector +.>For example, will->Concealing feature h from T-th time step T ∈R m×1 Inputting into a TPA time dimension attention module, wherein a scoring function f (·) in the module is based on CO 2 The degree of influence of the i' th sensor in the sensor node on the final calibration result, and the hidden characteristic h with the current time step T Scoring the relationship of (2); then outputs the weight alpha with the value range between 0 and 1 through the sigmoid activation function i” And the subsequent calculation is convenient. The scoring function is calculated as follows:
/>
wherein W is a The weights representing the scoring functions, sigmoid (·) representing the sigmoid activation functions.
All weights α to be generated next i” And pair convolution tensorMultiplying and adding to obtain weight vector v t ：
Wherein m represents CO converted by GRU (grid-controlled loop) unit module 2 The hidden characteristic quantity of the sensor in the sensor node, i represents the product operation order of the ith row convolution tensor and the corresponding ith weight;
integrating the T-th time step hidden feature h T And weight vector v t Features are further integrated through an MLP (multi-layer perceptron) module, and finally CO is output 2 Calibration value of the T-th time step of a sensor
Wherein W is h ，W v Representing different weights, MLP (·) represents an MLP multi-layer perceptron.
The invention has the beneficial effects that:
(1) Compared with a calibration model SensorFormer, the method has the remarkable advantages of real-time performance and high accuracy:
the sensor former introduces the monitoring values of the sensor at the past, current and future moments, so that the value at the current moment to be calibrated needs to bear a certain delay on the premise of ensuring the calibration precision;
the self-attention mechanism of the SensorFormer model focuses more on the dependency between time steps, and the special attention mechanism used in the present invention focuses on CO during time variation 2 The dependency relationship among different types of sensors in the sensor node is more suitable for the calibration task of multiple sensors, and has higher accuracy;
the complexity of the self-attention mechanism in the sensor former model is higher, and the self-attention mechanism can not be applied to an edge system with lower calculation power on the premise of not simplifying the structure of the sensor former model, and the special attention mechanism adopted by the invention is relatively lower, so that the sensor former model can be smoothly applied to the edge system.
(2) Compared with the method for converting the original data into the hidden features by using the feedforward neural network in the sensor former model, the method for converting the hidden features by using the gating circulating unit module and the convolution module comprises richer detail information and time sequence change information, and the accuracy of the calibrated monitoring value is further improved.
(3) Compared with a calibration model deep CM, the method has the remarkable advantages of high accuracy and stability:
in the circulation window jumping module of the deep CM model, different atmospheric pollutants show different periodicity, so in order to match the cycle of specific atmospheric pollutants, the jumping length of the module must be manually adjusted, and the invention can learn the cycle mode by itself, thus being more intelligent and more accurate;
the greater the periodic regularity of the pollutant concentration, the higher the accuracy of the deep CM model, however, in the scene of irregular variation of the pollutant concentration, such as monitoring the CO in the tail gas of the vehicle in the urban road 2 Is used for monitoring the emission of CO in a crop greenhouse 2 The effect is not satisfactory at the content of (3). The method has higher accuracy in monitoring the periodic pollutants and good stability in non-periodic scenes.
Of course, the present invention is capable of other various embodiments and its several details are capable of modification and variation in light of the present invention by one of ordinary skill in the art without departing from the spirit and scope of the present invention, as defined in the appended claims.

Claims (6)

1. Intelligent CO 2 A sensor system, comprising:
CO 2 sensor node for CO in the atmosphere 2 Monitoring, loading and running a trained TPA-GRU model, and calibrating the monitoring data using the TPA-GRU model, wherein the CO 2 The sensor node is arranged at the CO 2 A reference station is beside;
CO 2 reference station for CO in the atmosphere 2 Monitoring;
data communication network for communicating CO 2 Reference station and CO 2 Monitoring data of the sensor nodes are sent to an edge computing server;
the edge computing server is used for preprocessing the received monitoring data and training the TPA-GRU model based on the preprocessing result;
the TPA-GRU model comprises:
GRU gating circulation unit module for according to CO 2 Obtaining hidden features from a dataset generated by a sensor node, wherein the hidden features are divided into a past time step hidden feature and a second time step hidden featureTA time step concealment feature;
a Conv2D two-dimensional convolution module for calculating a plurality of convolution values according to the past time step, and combining the convolution values to form a convolution tensor, wherein each convolution valueCorresponding CO 2 Time sequence change characteristics of the sensor, the temperature sensor, the humidity sensor and the pressure sensor in a time window;
TPA time dimension attention module for receiving convolution tensor and thTConcealing features in time steps, from scoring functions, according to CO 2 Sensor node NoThe individual sensors score the calibration values and the relation to the hidden feature of the current time step and output weights via an activation function>；
An MLP multi-layer sensor module for sensing the weightAnd convolution tensor, calculating to obtain weight vector +.>And integrate the firstTTime step concealment feature and weight vector +.>Output the firstTCalibration values for time steps.
2. The intelligent CO of claim 1 2 Sensor system, characterized in that the CO 2 The sensor node includes:
a controller for setting CO 2 Sampling frequency of sensor, temperature sensor, humidity sensor and pressure sensor, and storing CO 2 Monitoring data generated by the sensors, temperature sensors, humidity sensors and pressure sensors through the AD conversion element, and controlling communication of the networking element and the edge computing server, and loading and running the trained TPA-GRU model;
the networking element is used for carrying out data communication with the edge computing server by adopting the MQTT protocol;
AD conversion element for converting CO 2 Analog signals generated by the reaction of the sensor, the temperature sensor, the humidity sensor and the pressure sensor with the environment are converted into digital monitoring data, and the digital monitoring data are stored in the controller.
3. The intelligent CO of claim 1 2 Sensor system, characterized in that the CO 2 The reference station includes:
an in-station server for transmitting data with the edge computing server and storing CO 2 Monitoring data generated by the monitoring equipment and periodically or selectively transmitting the monitoring data to an edge computing server;
CO 2 monitoring device for CO in atmosphere 2 Monitoring and storing the data in an in-station server.
4. Intelligent CO 2 A method of operating a sensor system, comprising the steps of:
s1, CO 2 Sensor node deployment to CO 2 Reference toBeside the station and using CO 2 Sensor and CO 2 Reference station to CO in the atmosphere 2 Monitoring and storing;
s2, CO 2 Sensor node and CO 2 The data monitored by the reference station are sent to an edge computing server;
s3, preprocessing the monitoring data by utilizing an edge computing server to obtain a training set and a testing set;
s4, training the TPA-GRU model with the initial weight preset by using a training set, and testing the TPA-GRU model by using a testing set to obtain a trained TPA-GRU model;
the step S4 includes the steps of:
s401 extracting continuous from training setCO of individual time steps 2 Monitoring data of the sensor node;
s402, inputting the extracted monitoring data into a pre-set TPA-GRU model with initial weight to obtainCalibration value for the individual time steps +.>；
S403, measuring the calibration value by using the loss functionWith corresponding CO 2 Data collected by reference station->Is the gap between (1);
the expression of the loss function is as follows:
wherein,representing a loss function->Representing a parameter that can be learned, < >>Representing a total time step;
s404, minimizing the gap through an optimization algorithm Adam;
s405, judging whether the minimized gap is unchanged in a preset training round, finishing training of the TPA-GRU model, saving the network weight with the minimum loss function value, and entering a step S406, otherwise, returning to the step S401;
s406, testing the TPA-GRU model by using a test set to obtain a trained TPA-GRU model;
s5, transmitting the trained TPA-GRU model to the CO through a data communication network by utilizing an edge computing server 2 Sensor node and use CO 2 The sensor node loads and runs the trained TPA-GRU model;
and S6, calibrating the monitoring data by utilizing the TPA-GRU model, and outputting a calibration value.
5. The intelligent CO of claim 4 2 A method of operating a sensor system, characterized in that said step S3 comprises the steps of:
s301, merging the monitoring data by utilizing an edge computing server according to the same timestamp, deleting time step data comprising null values and outliers, and constructing a data set D;
s302, dividing the data set D into a training set and a testing set.
6. The intelligent CO of claim 4 2 A method of operating a sensor system, wherein S6 comprises the steps of:
s601, input includesTTime step and historical time stepTime-step CO 2 Monitoring data by the sensor nodes to generate a data set;
s602, inputting the generated data set into a GRU (gate-controlled loop) unit module to obtain hidden features;
s603, dividing the hidden feature into a past time step hidden feature and a thTA time step concealment feature;
s604, transmitting the hidden characteristic of the past time step into the Conv2D two-dimensional convolution module and connecting the hidden characteristic with the Conv2D two-dimensional convolution modulekThe filters perform convolution operation to obtain a plurality of convolution valuesWherein each convolution value +>Corresponding CO 2 The time sequence change characteristics of each sensor in the sensor nodes in a time window;
wherein,representing +.f. in CO2 sensor node>The individual sensor is at->Hidden form feature of the monitored value at time step, < >>Representing convolution operation,/->Indicate->Length of->Is a convolution filter of->Representing +.>A secondary convolution operation sequence;
s605, combining the convolution values to form a convolution tensorAnd the convolution tensor is combined with the firstTThe time step hiding feature is input to the TPA time dimension attention module;
s606, a scoring function in the TPA time dimension attention module is used for scoring the data according to the CO 2 Sensor node NoThe individual sensors score the calibration values and the relation with the hidden features of the current time step and output weights +_ via a sigmoid activation function>；
The expression of the scoring function is as follows:
wherein,representing a scoring function->Represent the firstTTime step hidden feature,/->Weights representing scoring functions;
the weight isThe expression of (2) is as follows:
wherein,representing a sigmoid activation function;
s607, according to the weightAnd convolution tensor->Calculating to obtain weight vector->：
Wherein,mrepresenting CO converted by GRU (gate-controlled loop) unit module 2 The number of hidden features of the sensor in the sensor node,irepresent the firstiLine convolution tensor and corresponding firstThe product operation sequence of the weights;
S608、integration of the first with MLP multilayer perceptron moduleTTime step concealment feature and weight vectorOutput the firstTCalibration value for time step->：
Wherein,and->All represent weights, ++>Representing an MLP multi-layer perceptron.
CN202311091366.XA 2023-08-28 2023-08-28 Intelligent CO 2 Sensor system and method of operation thereof Active CN117310088B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311091366.XA CN117310088B (en) 2023-08-28 2023-08-28 Intelligent CO 2 Sensor system and method of operation thereof

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311091366.XA CN117310088B (en) 2023-08-28 2023-08-28 Intelligent CO 2 Sensor system and method of operation thereof

Publications (2)

Publication Number Publication Date
CN117310088A CN117310088A (en) 2023-12-29
CN117310088B true CN117310088B (en) 2024-04-02

Family

ID=89280104

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311091366.XA Active CN117310088B (en) 2023-08-28 2023-08-28 Intelligent CO 2 Sensor system and method of operation thereof

Country Status (1)

Country Link
CN (1) CN117310088B (en)

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114548555A (en) * 2022-02-22 2022-05-27 大连理工大学 Axial flow compressor stall surge prediction method based on deep autoregressive network
CN114609329A (en) * 2022-01-28 2022-06-10 西安电子科技大学 Gas monitoring system based on sensor networking under industrial environment
CN114818515A (en) * 2022-06-24 2022-07-29 中国海洋大学 Multidimensional time sequence prediction method based on self-attention mechanism and graph convolution network
CN116612664A (en) * 2023-05-06 2023-08-18 中国地质大学（武汉） Ship traffic flow prediction method based on improved space-time diagram attention neural network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114609329A (en) * 2022-01-28 2022-06-10 西安电子科技大学 Gas monitoring system based on sensor networking under industrial environment
CN114548555A (en) * 2022-02-22 2022-05-27 大连理工大学 Axial flow compressor stall surge prediction method based on deep autoregressive network
CN114818515A (en) * 2022-06-24 2022-07-29 中国海洋大学 Multidimensional time sequence prediction method based on self-attention mechanism and graph convolution network
CN116612664A (en) * 2023-05-06 2023-08-18 中国地质大学（武汉） Ship traffic flow prediction method based on improved space-time diagram attention neural network

Also Published As

Publication number Publication date
CN117310088A (en) 2023-12-29

Similar Documents

Publication Publication Date Title
US20220214322A1 (en) Air pollutants concentration forecasting method and apparatus and storage medium
CN110909909A (en) Short-term traffic flow prediction method based on deep learning and multi-layer spatiotemporal feature map
CN113446988B (en) Airport runway pavement state monitoring system and method based on cloud edge fusion framework
CN109473119B (en) Acoustic target event monitoring method
Do et al. Graph-deep-learning-based inference of fine-grained air quality from mobile IoT sensors
CN112257911B (en) TCN multivariate time sequence prediction method based on parallel space-time attention mechanism
Vijayakumar et al. RETRACTED ARTICLE: Artificial intelligence-based agriculture automated monitoring systems using WSN
CN113988263A (en) Knowledge distillation-based space-time prediction method in industrial Internet of things edge equipment
CN115128978A (en) Internet of things environment big data detection and intelligent monitoring system
CN115759461A (en) Internet of things-oriented multivariate time sequence prediction method and system
CN112749663A (en) Agricultural fruit maturity detection system based on Internet of things and CCNN model
CN116720156A (en) Weather element forecasting method based on graph neural network multi-mode weather data fusion
CN117310088B (en) Intelligent CO 2 Sensor system and method of operation thereof
CN116341763B (en) Air quality prediction method
CN115801152B (en) WiFi action recognition method based on hierarchical transformer model
CN114386672B (en) Environment big data Internet of things intelligent detection system
CN116822592A (en) Target tracking method based on event data and impulse neural network
CN115062764B (en) Intelligent illuminance adjustment and environmental parameter Internet of things big data system
CN114417242B (en) Big data detection system for livestock and poultry activity information
CN114995248A (en) Intelligent maintenance and environmental parameter big data internet of things system
CN115168327A (en) Large-scale data space-time prediction method based on multilayer tree long-short term memory network
CN113962432A (en) Wind power prediction method and system integrating three-dimensional convolution and light-weight convolution threshold unit
CN115016275B (en) Intelligent feeding and livestock house big data Internet of things system
CN117934963B (en) Gas sensor drift compensation method
Jin et al. Research on multi-sensor information fusion algorithm of fan detection robot based on improved BP neural network

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant