CN113108949B - Model fusion-based sonde temperature sensor error prediction method - Google Patents
Model fusion-based sonde temperature sensor error prediction method Download PDFInfo
- Publication number
- CN113108949B CN113108949B CN202110279691.3A CN202110279691A CN113108949B CN 113108949 B CN113108949 B CN 113108949B CN 202110279691 A CN202110279691 A CN 202110279691A CN 113108949 B CN113108949 B CN 113108949B
- Authority
- CN
- China
- Prior art keywords
- model
- prediction
- error
- data
- layer
- 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
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01K—MEASURING TEMPERATURE; MEASURING QUANTITY OF HEAT; THERMALLY-SENSITIVE ELEMENTS NOT OTHERWISE PROVIDED FOR
- G01K15/00—Testing or calibrating of thermometers
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01K—MEASURING TEMPERATURE; MEASURING QUANTITY OF HEAT; THERMALLY-SENSITIVE ELEMENTS NOT OTHERWISE PROVIDED FOR
- G01K13/00—Thermometers specially adapted for specific purposes
- G01K13/02—Thermometers specially adapted for specific purposes for measuring temperature of moving fluids or granular materials capable of flow
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01W—METEOROLOGY
- G01W1/00—Meteorology
- G01W1/08—Adaptations of balloons, missiles, or aircraft for meteorological purposes; Radiosondes
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Abstract
The invention discloses a method for predicting an error of a sonde temperature sensor based on model fusion, and belongs to the technical field of sensors. The invention comprises the following steps: constructing an overhead meteorological detection data set; constructing corresponding influence factors by utilizing the characteristic engineering; designing an enhanced deep neural network model; and designing a sensor error prediction method based on model fusion. The method fully utilizes the advantages of machine learning in error prediction, takes residual error as an idea, designs a prediction model aiming at the error of the sonde temperature sensor, and effectively improves the measurement accuracy of the sensor.
Description
Technical Field
The invention relates to the technical field of sensors, in particular to a method for predicting an error of a temperature sensor of a sonde based on model fusion.
Background
At present, in the field of high-altitude meteorological detection, a qualitative analysis method is mainly adopted for error prediction of a temperature sensor, and a temperature sensor error correction table is constructed by measuring measurement errors of the temperature sensor under meteorological conditions of different longitudes, latitudes, altitudes and the like, so that the error prediction of the sensor is realized. However, the table lookup method only provides an approximate error range, and does not accurately error data, and due to the influence of manufacturing raw materials, processes and the like of the sensor, the simple table lookup method cannot meet the requirement of the measurement accuracy of the sensor.
With the development of computer simulation technology, the error quantitative analysis through fluid mechanics simulation becomes a main research method in the field of high-altitude detection. The method is characterized in that a temperature sensor simulation mode diagram is constructed by designing a fluid dynamics model, the measurement errors of the sensor are calculated, the regression analysis is carried out on the measurement errors by a least square method, the error prediction of the temperature sensor is realized, the method is always kept in simulation data, and the method is not applied to the actual environment and lacks of certain practical factors.
Machine learning is used as a novel information processing method, a new solution is brought to the field of high-altitude meteorological detection, the error prediction of the temperature sensor can be realized by constructing a model fusion-based air-detecting temperature sensor error prediction model and learning data acquired by the sensor in the actual meteorological detection process, and the measurement accuracy of the sensor is improved.
Disclosure of Invention
The invention provides a method for predicting errors of a temperature sensor of a sonde based on model fusion, aiming at the problem that the sensor has measurement errors due to interference of factors such as solar radiation, cloud entering and cloud exiting in the process of high-altitude meteorological detection. The method combines a deep neural network, a Gaussian function, a support vector machine, an XGboost and a logistic regression, designs a prediction method for the error of the temperature sensor of the sonde based on a residual error idea, and realizes error compensation by predicting the error of the temperature sensor, so that the measurement accuracy of the temperature sensor is further improved, and the requirement of the high-altitude meteorological detection field on the measurement accuracy of the sensor is met.
In order to achieve the purpose, the invention adopts the following technical scheme:
a method for predicting an error of a sonde temperature sensor based on model fusion mainly comprises the following steps:
step 1: construction of high altitude weather detection data set
When constructing the sensor error prediction model, first, sensor data acquisition is required. The used data is based on data actually measured in the high-altitude detection process, and the study objects are a temperature sensor to be repaired and an international standard temperature sensor. And the sonde data transmission system transmits the measured high altitude meteorological element data every 1 s. The constructed data set has 18 kinds of data related to the measurement error of the sensor.
And 2, step: designing an enhanced deep neural network model
For error prediction of a temperature sensor of a sonde, due to the fact that the error prediction has high prediction precision requirements, the conventional deep neural network structure cannot be used for predicting the error with high precision. Therefore, a Gaussian function is proposed to replace a Sigmoid function as an activation function of a network hidden layer, so as to further extract the relation between different characteristic values and improve the prediction performance of the model.
And 3, step 3: sensor error prediction method based on model fusion
The original data set is firstly divided into 3 sub-data sets by 3-fold cross validation, and then the divided data is input into the prediction model of the layer 1, and the respective prediction results are output. And finally, according to the residual error idea, taking the output value of the layer 1 as the input of the layer 2, and training the learning algorithm of the prediction model of the layer 2 to learn the residual error between the predicted value of the first layer and the final standard value so as to further improve the prediction precision.
In the model construction, a support vector machine model (SVM) based on a hyperplane, an XGboost model based on a tree principle, a deep neural network model (DNN) based on a neuron and a logistic regression model (LR) based on a weighting are selected. On the basis of SVM, DNN and XGboost prediction results, an LR model is trained, different model prediction results are subjected to weighted analysis, the parts with better extracted characteristics of the models are captured, and the parts with poor performance of the models are discarded, so that the prediction results are effectively optimized, and the final prediction accuracy is improved.
Compared with the prior art, the invention has the following advantages:
1. compared with the existing table look-up method and computer model simulation, the method has the advantages that the sensor error prediction model is constructed by collecting data in the actual high-altitude meteorological detection process, so that the method can be applied to actual engineering.
2. Compared with the traditional single model, the method has the advantages that the residual error is taken as the idea, the model fusion method is adopted, the error prediction precision of the model is enhanced, and the sensor performance is improved.
Drawings
FIG. 1 is a diagram of the overall architecture of the sonde temperature sensor error prediction based on model fusion
FIG. 2 is a visual representation of support vector machine sensor error prediction
FIG. 3 is a visual diagram of error prediction of an XGboost sensor
Table 1 is a sonde data acquisition table
Table 2 is a deep neural network error prediction comparison table
Table 3 is a comparison table of the predicted results of each model
Detailed Description
The method realizes the error prediction of the sonde temperature sensor based on model fusion. The specific method adopted by the invention will be described in detail below with reference to the accompanying drawings.
Fig. 1 shows an overall error prediction architecture diagram of a sonde temperature sensor based on model fusion, which mainly comprises four parts, namely a deep neural network, a support vector machine, an XGBoost and a logistic regression.
Deep neural network
According to the basic structure of the deep neural network, a 5-layer network is selected for training, one input layer, three hidden layers and one output layer. From the collected data, the time, longitude, latitude, altitude and other 14 features of the used features can be known. According to the related knowledge, the constructed characteristics comprise 3 characteristics such as solar altitude angle, lift-off speed grouping and the like, namely, a data set comprises 17 characteristics related to the measurement error of the sensor, and the content of the data set is shown in table 1.
According to the data set, 17 neurons are arranged in the input layer of the deep neural network; in order to realize better error correction capability, the hidden layer is designed to be 20-30-10; after the prediction of the deep neural network, the corresponding prediction error is output, so that the output layer has 1 neuron, namely the network structure is 17-20-30-10-1.
In order to verify the influence of the Gaussian function on the prediction result, the traditional Sigmoid function and the Gaussian function are respectively adopted for the activation function to carry out a comparison test. Inputting the same test data into a deep neural network, training the network to achieve final convergence, and counting the predicted error values under different models, wherein the used evaluation indexes comprise three indexes, namely root Mean Square Error (MSE) for measuring the overall distribution condition of the errors, average error (ME) for measuring the size of the predicted errors and standard deviation (delta) for measuring the dispersion condition of the errors. The Sigmoid function and gaussian function predictions are shown in table 2.
Support vector machine
The kernel function of the support vector machine designed by the invention is a radial basis function, and the kernel function coefficient is 10 -3 The penalty factor is 10. The same test data is input into a support vector machine, the model is trained to achieve final convergence, and a prediction result visualization graph is shown in fig. 2.
(III) XGboost
The XGboost model designed by the invention has the tree depth of 3, the tree number of 500 and the learning rate of 0.01. The same test data is input into the XGboost model, the model is trained to achieve final convergence, and a visual graph of a prediction result is shown in FIG. 3.
(IV) fusion model
And combining DNN, XGboost, SVM and LR according to the residual error idea to construct a fusion prediction model. On the basis of SVM, DNN and XGboost prediction results, an LR model is trained, different model prediction results are subjected to weighted analysis, the parts with better extracted characteristics of the models are captured, and the parts with poor performance of the models are discarded, so that the prediction results are effectively optimized, the final prediction precision is improved, and the prediction precision of the models is shown in table 3.
Table 1 sonde data acquisition table
Parameter(s) | Numerical value | Parameter(s) | Numerical value |
time/H | 18:58:05 | Distance/m | 0.148 |
Measuring barometric pressure/hPa | 992.59 | Longitude/degree | 112.7879 |
Measurement of temperature/. Degree.C | 30.94 | Latitude/degree | 28.1095 |
Measurement of humidity/%) | 68 | Azimuth angle/° | 76.4 |
Dew point/. Degree.C | 24.34 | Elevation angle/° | 1.6 |
Altitude/m | 119.3 | Standard time/H | 39483 |
Wind speed/m/s | 5.35 | Standard temperature/. Degree.C | 31.2 |
Wind direction/° c | 128.1 |
TABLE 2 deep neural network error prediction contrast table
Function(s) | Root mean square error | Mean error of | Standard deviation of |
Sigmoid function | 0.13 | -0.0110 | 0.3393 |
Gaussian function | 0.11 | -0.005 | 0.3369 |
TABLE 3 comparison table of prediction results of each model
Model (model) | Root mean square error | Mean error of | Standard deviation of |
Before correction | 0.50 | -0.5547 | 0.4392 |
DNN | 0.13 | -0.011 | 0.3393 |
DNN + Gaussian function | 0.11 | -0.005 | 0.3369 |
LR | 0.21 | 0.1837 | 0.3618 |
SVM | 0.15 | 0.1499 | 0.3512 |
XGBoost | 0.07 | -0.0042 | 0.2288 |
Model fusion | 0.04 | 8.70E-18 | 0.2054 |
Claims (1)
1. A method for predicting the error of a sonde temperature sensor based on model fusion is characterized by comprising the following steps: comprises the following steps:
step 1: constructing an overhead meteorological detection data set;
when constructing a sensor error prediction model, firstly, acquiring sensor data; the used data is based on actually measured data in the high-altitude detection process, and the study objects are a temperature sensor to be repaired and an international standard temperature sensor; the sonde data transmission system transmits the measured high altitude meteorological element data every 1 s; the constructed data set has 18 data related to the measurement error of the sensor;
and 2, step: designing an enhanced deep neural network model;
for error prediction of the temperature sensor of the sonde, a Gaussian function is used as an activation function of a network hidden layer instead of a Sigmoid function so as to extract the relation between different characteristic values and improve the prediction performance of the model;
and step 3: designing a sensor error prediction method based on model fusion;
firstly, dividing an original data set into 3 sub-data sets through 3-fold cross validation, then inputting the divided data into a prediction model of a layer 1, and outputting respective prediction results; finally, according to the residual error idea, taking the output value of the 1 st layer as the input of the 2 nd layer, training the learning algorithm of the 2 nd layer prediction model, and enabling the learning algorithm to learn the residual error between the first layer prediction value and the final standard value so as to improve the prediction precision;
when the model is constructed, selecting a support vector machine model SVM based on a hyperplane, an XGboost model based on a tree principle, a deep neural network model DNN based on a neuron and a weighted logistic regression model LR;
acquiring 15 characteristic meteorological element characteristics of time, distance, air pressure, longitude, temperature, latitude, humidity, azimuth angle, dew point, elevation angle, altitude, standard time, wind speed, standard temperature and wind direction measured in the actual high-altitude meteorological detection process;
constructing influence factors of measurement errors of the sensor in the high-altitude meteorological detection process, wherein the influence factors comprise 5 characteristics of a solar altitude angle, a lift-off speed, lift-off speed grouping, cloud entering and cloud exiting;
combining the 15 collected original characteristics related to the measurement error of the temperature sensor with the 5 construction characteristics to construct an overhead meteorological detection data set;
for a hidden layer activation function in the deep neural network, a Gaussian function is used for replacing a Sigmoid function, and the prediction precision of the model on the measurement error of the temperature sensor is improved;
on the aspect of data set division, 3-fold cross validation is adopted to divide training data in a data set into a training set 1, a training set 2 and a training set 3, two parts of the training set are used as a training set of a certain model, and the other part of the training set is used as a test set; performing non-repeated three-time division on the data set, respectively performing one-time prediction on each part, wherein the predicted model is obtained by training the rest two parts, the obtained final prediction result is also complete data set data, and the part of data is sent to a second-time prediction model for training prediction; according to the residual error idea, comparing the error between the output value of the first layer and the standard value and taking the error as the input data of the second layer model; the second layer model selects a weighted-based logistic regression model LR to learn the residual error between the predicted value and the standard value of the first layer; performing weighted analysis on the prediction results of the models in the first layer, and capturing the parts of the models with good extraction characteristics;
and combining the deep neural network, the XGboost, the support vector machine and the logistic regression by taking the residual error as an idea to construct a two-stage prediction model and improve the prediction precision of the model.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110279691.3A CN113108949B (en) | 2021-03-16 | 2021-03-16 | Model fusion-based sonde temperature sensor error prediction method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202110279691.3A CN113108949B (en) | 2021-03-16 | 2021-03-16 | Model fusion-based sonde temperature sensor error prediction method |
Publications (2)
Publication Number | Publication Date |
---|---|
CN113108949A CN113108949A (en) | 2021-07-13 |
CN113108949B true CN113108949B (en) | 2022-11-04 |
Family
ID=76711502
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202110279691.3A Active CN113108949B (en) | 2021-03-16 | 2021-03-16 | Model fusion-based sonde temperature sensor error prediction method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN113108949B (en) |
Families Citing this family (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN116268545B (en) * | 2023-05-23 | 2023-08-15 | 深圳市鸿云智科技有限公司 | Diagnosis method and calibration method for cigarette weight detection |
Family Cites Families (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102567391B (en) * | 2010-12-20 | 2014-05-07 | 中国移动通信集团广东有限公司 | Method and device for building classification forecasting mixed model |
CN102194057B (en) * | 2011-05-05 | 2013-02-06 | 浙江宜景环保科技有限公司 | Method for updating model of waste plastic oil refining combustion optimization |
CN109902346B (en) * | 2019-01-24 | 2023-04-07 | 东南大学 | Neural network-based regional weighted average temperature information acquisition method |
CN110220725B (en) * | 2019-05-30 | 2021-03-23 | 河海大学 | Subway wheel health state prediction method based on deep learning and BP integration |
AU2020104000A4 (en) * | 2020-12-10 | 2021-02-18 | Guangxi University | Short-term Load Forecasting Method Based on TCN and IPSO-LSSVM Combined Model |
-
2021
- 2021-03-16 CN CN202110279691.3A patent/CN113108949B/en active Active
Also Published As
Publication number | Publication date |
---|---|
CN113108949A (en) | 2021-07-13 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN109308522B (en) | GIS fault prediction method based on recurrent neural network | |
CN114092832B (en) | High-resolution remote sensing image classification method based on parallel hybrid convolutional network | |
CN112766549A (en) | Air pollutant concentration forecasting method and device and storage medium | |
CN110307982B (en) | Bearing fault classification method based on CNN and Adaboost | |
CN110333076B (en) | Bearing fault diagnosis method based on CNN-Stacking | |
CN111737912B (en) | MWHTS (metal wrap through) simulated bright temperature calculation method based on deep neural network | |
CN111737913B (en) | MWHTS clear sky observation bright temperature selection method based on cloud water content inversion | |
CN105223241A (en) | A kind of compensation method of humidity sensor | |
CN112036075A (en) | Abnormal data judgment method based on environmental monitoring data association relation | |
CN110968069A (en) | Fault prediction method of wind generating set, corresponding device and electronic equipment | |
CN111638034B (en) | Strain balance temperature gradient error compensation method and system based on deep learning | |
CN113255972B (en) | Short-term rainfall prediction method based on Attention mechanism | |
CN113743013A (en) | XGboost-based temperature prediction data correction method | |
CN112232543A (en) | Multi-site prediction method based on graph convolution network | |
CN114444378A (en) | Short-term power prediction method for regional wind power cluster | |
CN113505923A (en) | Regional power grid short-term load prediction method and system | |
CN113984969A (en) | Air quality prediction method and system based on multi-source space-time data fusion | |
CN113108949B (en) | Model fusion-based sonde temperature sensor error prediction method | |
CN114818579A (en) | Analog circuit fault diagnosis method based on one-dimensional convolution long-short term memory network | |
CN114676779A (en) | Atmospheric quality monitoring data calibration method and system based on robust width network | |
CN114066070A (en) | Weather forecasting method based on artificial intelligence and numerical calculation | |
CN113657023A (en) | Near-surface ozone concentration inversion method based on combination of machine learning and deep learning | |
CN113642255A (en) | Photovoltaic power generation power prediction method based on multi-scale convolution cyclic neural network | |
CN116312860B (en) | Agricultural product soluble solid matter prediction method based on supervised transfer learning | |
CN112329334A (en) | MWHTS and MWTS-II fusion inversion sea surface air pressure method based on simulated brightness temperature |
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 | ||
TR01 | Transfer of patent right |
Effective date of registration: 20221104 Address after: 100012 817, Floor 8, No. 101, Floor 3 to 8, Building 17, Rongchuang Road, Chaoyang District, Beijing Patentee after: Beijing Ge Lei Information Technology Co.,Ltd. Address before: 100124 No. 100 Chaoyang District Ping Tian Park, Beijing Patentee before: Beijing University of Technology |
|
TR01 | Transfer of patent right |