AU2020100700A4 - A Correction Method for Gas Sensor Based on Machine Learning - Google Patents

Abstract

This invention lies in the field of data processing technology, in particular to a gas sensor correction method based on machine learning. This invention consists of the following steps. Firstly, we collect the data of the outdoor sensor and the data of a nearby monitoring station at the same time. Secondly, use Python to format and process the data. Thirdly, build the model using gradient descent algorithms, and divide the selected and preprocessed data set into training set and test set. Finally, the training set is used for the training of model parameters. On this basis, the performance of the test set calculation model is applied. In general, the invented algorithms are concise but effective. The operation of the equipment is convenient and efficient. The experimental results have high accuracy and can be widely used in the correction of general gas sensors. In addition, our invention, the correction model, is universal and can be extended to other sensors.

Description

DESCRIPTION

TITLE

A Correction Method for Gas Sensor Based on Machine Learning

FIELD OF THE INVENTION

This invention lies in the field of data processing technology, in particular to a gas sensor correction method based on machine learning.

BACKGROUND OF THE INVENTION

In recent years, pollution of harmful gases such as CO, NO, O₃, SO₂ and so on has been increasing. At the atmospheric level, the spread of harmful gases causes severe harm to nature, traffic, construction and people's health. At the household level, there is also a crisis of harmful gas leaks, such as CO leak. Therefore, it is very important to monitor the concentration of harmful gases accurately and improve monitoring technology. Also, gas monitoring technology is of great significance to urban environmental planning and air pollution prevention.

Nowadays, there are two main methods to monitor the concentration of harmful gases. One is the long optical path method. Generally, several air monitoring stations are set up in the city, and automatic monitoring instruments are installed for continuous automatic monitoring. The monitoring results are taken back regularly by people assigned to analyze and obtain relevant data. This method has high monitoring accuracy. However, the cost of measuring instruments is too high, and it needs professional personnel to operate. Also, the whole monitoring period is too long.

The second monitoring method is the sensor method. With the development and use

2020100700 05 May 2020 of various natural gas, coal-gas and liquefied gas, the detection and control methods of related gases have been studied deeply at home and abroad. A variety of sensors for gas detection and analysis have been produced and widely used in gas detection and composition analysis in production and life. Among them, the most widely used are metal oxide semiconductor sensors and electrochemical sensors, both of which are relatively low cost and easy to operate. Metal oxide semiconductor sensor is a detection element made of metal semiconductor oxide or metal oxide material. When it interacts with gases, it will generate surface adsorption and reaction, which will cause characteristics of conductivity or voltammetry characteristics or surface potential change, and then the detection element can carry out gas concentration measurement. This sensor has relatively narrow linear detection range for common pollutants, and it can be easily affected by ambient temperature. Furthermore, this sensor has high power consumption. The electrochemical sensor used by our invention can be widely used for quantitative detection of various gases and has the following characteristics: zero power consumption, good linearity, strong output stability, high precision and integration. However, the electrochemical sensor also exists the phenomenon of temperature drift, which is affected by the temperature and humidity, resulting in the decline of accuracy.

Based on some problems in the operation of the current monitoring methods, relevant personnel put forward many solutions as follows:

A) The wireless calibration system of gas sensors is used in the coal mining industry, and the environmental parameters wireless monitoring device is used to monitor the

2020100700 05 May 2020 sensor outputs under different gas concentrations^[1].

B) In agricultural greenhouses, we usually use gas sensors to monitor gas content. The parameters are adjusted by means of ventilation and gas replenishment to make the sensor reach a suitable state. The standard gas is used for periodic calibration to determine the zero point of measurement and the change of error function, so as to reduce the system error and improve the accuracy of measurement^[2].

C) Many coal mines integrate infrared signal transmission module into gas sensor calibrator. In this way, automatic closed-loop calibration of various sensors is realized. It reduces the error of human operation and protects the life safety of the personnel^[3].

D) The absolute sensitivity of the sensor is corrected by the reciprocity method. In other words, electrical characteristics of the same set of sensors are compared on a semi-infinite volume test fast surface. Thus, the absolute sensitivity is measured. This method does not need to measure the displacement of the surface directly, so it is easy to operate.

E) Nonlinear treatment of metal oxide gas sensors. This method uses two techniques: data preprocessing and nonlinear calibration to deal with the nonlinearity of the sensor and the measurement circuit. Thus, the problem of the sensor in the quantitative measurement of gas concentration is solved^[4].

In order to obtain accurate gas monitoring data and reduce the adverse effects of many factors on the accuracy of data, the invention will provide a late-model correction method of the gas sensor. This method is based on machine learning. Through the gradient descent method, we build a linear correction model. Parameters

2020100700 05 May 2020 in this model are continuously updated iteratively in order to achieve the objective of minimizing loss function, high precision, high speed, high correction efficiency and strong universality.

In brief, our invention/model can effectively solve the current correction problems of electrochemical sensors and other gas sensors.

SUMMARY OF THE INVENTION

In order to solve the problems above properly, the methods proposed in our invention include the following steps (also shown in figure 1):

( 1 ) Data acquisition

Collect the data every other equal time during the experimental time. The collected data can be divided into two groups: one is the data measured by the sensor, and the other one is the data measured by the monitoring station near the sensor. Considering that the data measured by the monitoring station has a high accuracy, it can be regarded as true values of the gases’ concentration. There are some differences between the measured values of sensors and those of monitoring stations, which needs correcting by our correction model.

( 2 ) Data processing

1. Normalization: In our data, the dimensions of different indicators are different. Normalizing the data can eliminate the dimensional impact between indicators. We transform the original data’s values to the value between [0,1], so as to facilitate the analysis and comparison of data. The formula we used for normalization is:

X Xmin

X =------Xmin (1)

Xmax

2020100700 05 May 2020

2. Integration of sampling period: the sensor data and the detection station data need to be combined. Since each data has a date and an hour of sampling, the principle of combining data is to merge the same data of date and hour in the two groups of data to generate a new data set, which can be named as “data.txt”.

3. Set training set and test set: the data set generated above has immense number of samples. Some of those samples are used to train the parameter model theta, and the other data are used to calculate the performance of the model. We set num(trairi)·. num(test) = M\ N. For example, if there are 1000 samples in data.txt and Μ: N = 9:1 is set, 900 of them are used to train parameter models and 100 are used to calculate model performance.

(3) Build model

The invention is a correction method of gas sensors based on machine learning. We establish a linear regression model of hypothesis, h to correct the obtained sensor’s data.

X^ = (1, %2^l)' '^nb(2) /0°\ e = H¹(3) \θ_η/ hg(x^ = θ^τ * x®(4)

In the above formulas: i represents that this variable belongs to sample i. x® is the characteristic vector of the sample i, and y^ is the real value corresponding to χϋ Θ is a vector of parameters in hypothesis, h. hg(xL^ is the result calculated according to the hypothesis, h.

2020100700 05 May 2020 (4) Update parameters

In this program, we use gradient descent to train the model h which has been built in (3). Gradient descent is a kind of optimization algorithm that is simply used to find out the values of a function’s parameters that minimize a loss function as far as possible.

The procedure starts off with initial values of Θ, which are small and random, for the coefficient for the function. The loss of coefficients is evaluated by plugging them into the function and calculating the loss. Then calculate the derivative of the loss that has been worked out at last step and now that we know from the derivative which direction is downhill, we can update the coefficient values. We have to specify a learning rate parameter a that controls how much the coefficients can change on each update. The process cannot stop until the loss of the coefficients close enough to 0.0 to be good enough. When the model is updating the parameters at each step, it is important to determine the magnitude of the parameters’ updating. This magnitude is also called the learning rate, a. If the learning rate is set too small, the optimization process will be too slow to converge within the fixed iteration time, cycle; if the learning rate is too large, it will cause oscillation and it is difficult to approach the optimal solution. Even worse we will stay away from the optimal solution. Therefore, it is important to choose an appropriate learning rate. We calculate the parameter theta from the formula as follow:

m i=l ¹ (j = 1, 2, 3,-,n) m

i.e.:

i=l

Now we turn it into matrix form:

As the model we built:

h_g (x) = Θ^T * X where:

A x™ 1 r^<2) ... T^{(2) Λ}1 ^Λη

V X™ ... 4^m7

Owing that each x® corresponds to each y^ y (0 — ί' Ί yi γί ··» γΐ] a, — i ± ^AnJ

y(0 = (y(l) y(2) ...y(m))^T / y(l) \ y(2)

Thus, the formula that can be derived is as follow:

θ := Θ — a*X^T *e (7) (8) (9) where:

( θ^τ * x^¹) — j/¹) θ^τ * χ(²) — \θ^τ * x(^m) — y(^m) j (5) Calculate parameters

Formula (9) is the one that is used to calculate the parameters.

2020100700 05 May 2020

We use the training data to work out parameters and the test data is used to evaluate the performance of our model.

(6) Evaluate the model

The value of Loss illustrates how well our model fits the true data. Loss can be calculated through the following formula:

m

Loss = — * [h_e(x^) — y^]²#(10) m L-i i=l

DESCRIPTION OF DRAWING

Figure 1 shows the whole procedure of our correction model.

Figure 2 shows Loss of the correction model with different a versus the number of cycles using the training data.(data withdrawn from building the model)

Figure 3 shows the correction model performance for CO using the testing data.(data withheld from building the model)

DESCRIPTION OF PREFERRED EMBODIMENT

Stepl: Data acquisition

In this project, we collected two types of data. First, we placed the sensors near Wanliu Monitoring Station to detect the air quality and collect data. The second is true data, which is from Wanliu Monitoring Station. Among the data obtained, we selected the data of the created date, CO, temperature, and humidity for the next data

2020100700 05 May 2020 processing. Besides, we collect these two kinds of data of four months. In this way, we make sure we have enough samples to train and test our model.

Step2: Process the data (1) Sensor data processing: Since the data obtained by sensor monitoring including year, month, day, hour, minute, second, the date required for the project were divided into two parts: date and time. Date contained year, month, day, and time contained hours in 24-hour time system. Therefore, we needed to process the sensor data. We used the Inversed Order to get the value, to prevent that the change of the previous date will influence the later value to be wrong. In the end, we got date data and hour data.

(2) Numerical averaging: Because the sensor generates one set of data were in units of one second, and the data we ultimately use were in units of one hour, we needed to find the average value of all the values in each hour as every hour data. Took the data from 7:00 to 8:00 of 2017-10-10 as an example. According to the statistics result, the sensor read a total of 3600 pieces of data during this time period. We summed the temperature and humidity and CO in the 3600 pieces of data, and then divided by 3600, and got the average value in this hour as the value required for the model. Using this method, we processed all the data, and got the temperature and humidity CO numerical information in units of one hour.

(3) Normalization of values: Because the value of CO values in the monitoring station was an order of magnitude worse than the values of temperature and humidity,

2020100700 05 May 2020 when we processed the data later, the effect of temperature and humidity on the results is significant so that we couldn’t show the effect of CO values in the results. Consequently, we needed to normalize the data of temperature and humidity, so that the values of these two items were fixed between 0-1. The method of numerical normalization was to calculate the maximum and minimum values of one set of temperature values. Then, started with the first data: every temperature data subtracted the minimum temperature value and then divided it by the difference between the maximum temperature and the minimum temperature. We processed each value to get the normalized data result of the temperature data. Using this method, we normalized the CO values, temperature and humidity to make the three data have the same magnitude.

(4) Data rearrangement: This project collected four months of sensor data, so we collect these four months of data which obtained by the above three processes into one new sensor data file.

(5) Data integration: Take Wanliu monitoring station data as the true values and take sensor’s data as the measurement data. We combined values of the same date and hour in the two sets of data one by one, and then we get the final experimental data.

Step3: Select training and test data sets

The data is divided according to the ratio of 9: 1, i.e. M:N = 9:1.

Namely, the training data set accounts for 90% of the total data, and the test data set accounts for 10% of the total data.

2020100700 05 May 2020

Step4: Calculate parameters

In the optimization of the parameter sets, we chose the gradient descent algorithm to optimize the parameters.

<- 0 — a * X^T * e

Where Θ is parameter vector, a is learning rate, X^T * e is the gradient of the loss function. The optimization principle of the gradient descent algorithm is to start from initial value, iterate continuously, update the value of Θ, and minimize the objective function until convergence. We will iterate for cycle times. If the objective function (i.e. Loss) does not converge, the iteration will be stopped.

m

Loss = — * V¹ — yW]² i=l

The Loss function is the squared difference between the sample value and the existing standard value. The smaller the loss function, the better the trained model.

(1) Data acquisition

We use the sensor data in the training set as the sample matrix X, and the data of Wanliu Monitoring Station as standard value matrix Y. The number of parameters is determined by the quantity of sensor data. The initial value of Theta is defined as a single column Zero Vector Matrix Θ.

(2) Calculate the gradient

The training data set is substituted into the formula for calculation, and then gets the gradient and theta parameter values.

(3) Calculate Loss

2020100700 05 May 2020

We put the theta value into the Loss formula, and then adjust a and cyclein order to reach the optimal performance. Then we put the test set into Loss function and get the value of Loss function.

Figure 1 gives the scatterplot of the values of Loss calculated by our correction

1/2 model versus the different values of a. Black dots represent the values of Loss during the process of iteration. The scatterplot shows that when we choose a = 0.002 or 0.005 , Loss will not converge. Parameters a = 0.001 and cycles = 1000 are chosen because under this circumstance Loss converges quickly and after correction sensor’s values are close to the true values.

Figure 2 shows the performance of the correction model for CO using the testing data. Triangle dots represent data from Wanliu Environmental Monitoring Station, which can regard as true values of CO’s concentration. Round dots represent data from our sensor which has been already corrected.

After many trails (some of which are shown in Figure 1), we finally decide that:

a = 0.001 cycle = 1000

Under this circumstance, we achieve that:

Loss « 0.040085703245

Loss « 0.2002

Judging by the value of Loss, our correction model is highly accurate.

Claims (3)

1. CLAIMS:

   2020100700 05 May 2020

   1. A Correction Method for Gas Sensor Based on Machine Learning, characterized in that: traditional long optical path method is accurate but costly, while sensors are relatively low-cost but easily affected by ambient temperature and humidity, the invention proposes a highly operable, highly efficient, and highly accurate correction method for electrochemical sensors to monitor the concentration of harmful gases.
2. 2. According to method of claim 1, wherein makes full use of the sensor’s data and the monitor station’s data to implement the machine learning; use gradient descent method for correction and use the test data set to examine the accuracy of corrected data, the result is precise and reliable.
3. 3. According to method of claim 1, wherein uses relatively short code to build up the correction model, this simple model can be easily transplanted to other monitoring instruments after necessary modification of the program.