CN102914623B - Fusing method of temperature compensation of humidity sensor - Google Patents

Abstract

The invention relates to a fusing method of temperature compensation of a humidity sensor. A characteristic function y=f(x) of the humidity sensor is given, wherein y represents the output of the humidity sensor, x represents the input of the humidity sensor, a temperature parameter is provided additionally, when t0<t1<t2<t3, according to the characteristics that a characteristic curve of the humidity sensor is linear in the middle and is nonlinear at two ends under the influence of the temperature, through a linear least square method, a linear equation f2(x, t) influenced by the temperature within a range of [t1, t2] is obtained; and when t0<t<t1 or t2<t<t3, error compensation caused by the temperature influence is carried out on f1(x, t) and f3 (x, t) through an RBF (Radial Basis Function) neutral network. The fusing method has the beneficial effect that an actual value which is closer to an output value of the humidity sensor is finally obtained.

Description

A kind of fusion method of humidity sensor temperature compensation

Technical field

The invention belongs to the soft compensation technique of sensor field, be specifically related to the application of method for self-adaption amalgamation in humidity sensor temperature compensation that combine with least square of a kind of RBF neural network.

Background technology

Air humidity is an important parameter in meteorological observation, and meteorological department has dropped into the automatic weather station of service operation at present, and the warm and humid integrated transducer of HMP45D that mostly adopts Finland Vaisala company to produce is measured air humidity.HMP45D Temperature Humidity Sensor is the thin polymer film capacitive transducer of new generation with HUMICAP technology, in the process of application, by linear relationship, calculate relative humidity size, yet in practical business process, there is the measurement of humidity under high/low temperature condition, can not meet the problem of measuring accuracy and reliability.Therefore, the non-linear effects that compensation temperature produces, becomes the problem that sensor research staff and worker in meteorology are concerned about the most.Non-linear effects humidity sensor being produced in order to eliminate temperature, researcher is mainly round hardware compensating and two kinds of methods of software compensation both at home and abroad.Hardware compensating method, owing to being subject to the impact of the factors such as electronic devices and components drift in circuit, causes whole measuring system poor reliability and precision low, cannot accomplish omnidistance compensation, and in actual use procedure, application is restricted.Software compensation method has multivariate regression analytic approach, yet when application multivariate regression analytic approach is carried out Data Fusion, need to solve large-scale matrix equation, even equation is ill-condition equation sometimes, cannot solve, and fitting precision is lower, often can not meets application request; Utilize the non-linear mapping capability of BP neural network also can realize the temperature compensation of humidity sensor, but there are some problems in the method, be easily absorbed in local extremum in application process, make failure to train, and when data volume is larger, pace of learning is slower, and fitting precision is lower.

Summary of the invention

The object of the invention is to overcome the deficiency of above prior art, and the fusion method of the humidity sensor temperature compensation that a kind of error is less is provided, and specifically has following technical scheme to realize:

The fusion method of described humidity sensor temperature compensation, the characteristic function of establishing humidity sensor is y=f (x), x ∈ [δ ₀, δ _h], wherein y is the output of sensor, and x is the input of sensor, and input reference signal is δ _h-δ ₀, it is characterized in that, set up temperature parameters t, linear in the middle of the family curve according to humidity sensor under temperature impact presents, the nonlinear feature in two ends, makes t0<t1<t2<t3, obtains formula (1),

$y=f(x,t)=\left\{\begin{array}{c}{f}_1(x,t),t\&Element;[t_0,t_1]\\ f_2(x,t),t\&Element;[t_1,t_2]\\ f_3(x,t),t\&Element;[t_2,t_3]\end{array}\right.---\left(1\right)$

Wherein, the characteristic function curve of humidity sensor is in input signal range ability, to be the function of continuous monotone variation; Work as t1<t<t2, by linear least square, obtain f ₂(x, t) is at the linear equation of [t1, t2] interior temperature influence; As t0<t<t1 or t2<t<t3, by RBF neural network to f ₁(x, t) and f ₃(x, t) carries out being affected and the error compensation brought by temperature.

The further design of the fusion method of described humidity sensor temperature compensation is, work as t1<t<t2, in the described linear equation of finally trying to achieve, add that parametric coefficients ξ compensation temperature produces minor impact, according to the output after A/D conversion, directly with straight-line equation, try to achieve corresponding input signal x, simplify the processing mode of output signal.

The further design of the fusion method of described humidity sensor temperature compensation is, described error compensation of being undertaken by RBF neural network, comprises the steps:

1) netinit, chooses n training sample at random as cluster centre c _i(i=1,2 ..., n);

2) the training sample set of input is pressed to the grouping of arest neighbors rule, then readjust cluster centre, the c finally obtaining _icenter for network function;

3) variance while solving network function and be Gaussian function i=1,2 ..., n, wherein c _maxfor the ultimate range between selected center;

4) calculate the weights between hidden layer and output layer (p=1,2 ..., P; I=1,2 ..., n);

5). by formula calculation of performance indicators J, then carries out weights adjustment, wherein, and y _kdesirable output, for reality output, k=1,2 ... L, L, M is empirical value;

6). judge the performance index J≤ζ that whether satisfies condition, ζ is empirical value, if meet, finishes training, otherwise makes J=0 return to step 2, repeats above-mentioned training process, until satisfy condition.

The present invention is the family curve under the impact of temperature according to humidity sensor, one section presents linearity, two ends present nonlinear feature, for the less linearity range of temperature impact, utilize the method for least square to simulate straight-line equation, add corresponding parametric coefficients, the impact of compensation temperature, make input input be good linear relationship, for the larger characteristic two ends of temperature impact, utilize RBF neural network model to realize the temperature compensation of humidity sensor, make input and output also be good linear relationship, the output valve that finally obtains humidity sensor approaches actual value more.

Accompanying drawing explanation

Fig. 1 is humidity sensor temperature compensation principle;

Fig. 2 is the impact of temperature on humidity sensor measurement result;

Fig. 3 is linear least square fitting a straight line schematic diagram;

Fig. 4 is the humidity sensor input-output curve after temperature compensation.

Embodiment

Below in conjunction with accompanying drawing, the present invention program is elaborated.

The fusion method of described humidity sensor temperature compensation, the characteristic function of establishing humidity sensor is y=f (x), x ∈ [δ ₀, δ _h], wherein y is the output of sensor, and x is the input of sensor, and input reference signal is δ _h-δ ₀, it is characterized in that, set up temperature parameters t, linear in the middle of the family curve according to humidity sensor under temperature impact presents, the nonlinear feature in two ends, makes t0<t1<t2<t3, obtains formula (1),

$y=f(x,t)=\left\{\begin{array}{c}{f}_1(x,t),t\&Element;[t_0,t_1]\\ f_2(x,t),t\&Element;[t_1,t_2]\\ f_3(x,t),t\&Element;[t_2,t_3]\end{array}\right.---\left(1\right)$

Wherein, the characteristic function curve of humidity sensor is in input signal range ability, to be the function of continuous monotone variation; Work as t1<t<t2, by linear least square, obtain f ₂(x, t) is at the linear equation of [t1, t2] interior temperature influence; As t0<t<t1 or t2<t<t3, by RBF neural network to f ₁(x, t) and f ₃(x, t) carries out being affected and the error compensation brought by temperature.

Work as t1<t<t2, in the described linear equation of finally trying to achieve, add that parametric coefficients ξ compensation temperature produces minor impact, according to the output after A/D conversion, directly with straight-line equation, try to achieve corresponding input signal x, simplify the processing mode of output signal.

The error compensation of being undertaken by RBF neural network, comprises the steps:

1) netinit, chooses n training sample at random as cluster centre c _i(i=1,2 ..., n);

2) the training sample set of input is pressed to the grouping of arest neighbors rule, then readjust cluster centre, the c finally obtaining _icenter for network function;

3) variance while solving network function and be Gaussian function i=1,2 ..., n, wherein c _maxfor the ultimate range between selected center;

4) calculate the weights between hidden layer and output layer (p=1,2 ..., P; I=1,2 ..., n);

5). by formula $J=\underset{k=1}{\overset{L}{\mathrm{\&Sigma;}}}{\left|\right|y_k-\widehat{y_k}\left|\right|}_2=\frac{1}{2}\underset{l=1}{\overset{L}{\mathrm{\&Sigma;}}}\underset{p=1}{\overset{M}{\mathrm{\&Sigma;}}}{(y_{\mathrm{pk}}-\widehat{y_{\mathrm{pk}}})}^2$ Y _kdesirable output, for reality output, k=1,2 ... L, L, M is empirical value, then calculation of performance indicators J carries out weights adjustment, wherein, y _kdesirable output, for reality output, k=1,2 ... L, L, M is empirical value;

6). judge the performance index J≤ζ that whether satisfies condition, ζ is empirical value, if meet, finishes training, otherwise makes J=0 return to step 2, repeats above-mentioned training process, until satisfy condition.

The experimental data that table 1 is actual measurement gained, the measurement error value under condition of different temperatures is that measured value deducts standard value, and the measurement error value in different temperatures situation is depicted as to smooth curve, the then impact of analysis temperature on humidity sensor.Resulting under different temperatures and different humidity condition the Curve of the Measuring Error of humidity sensor, as shown in Figure 2.

Table 1

Fig. 3 utilizes linear least square method to carry out fitting a straight line to good linearity range, and the straight-line equation obtaining is that y=1.0293*x-1.6659 adds minor impact error amount 0.8, and the input and output after over-fitting present good linear relationship.Certainly in actual application process, need to choose different values for different sensor model numbers, then just can calculate the actual humidity value in this segment according to the straight-line equation of matching, compensation precision is higher.

Fig. 4 utilizes the method for RBF neural network to carrying out temperature Compensation Study between inelastic region, temperature range two ends, using the part sample in experimental data as training sample, part sample is as test sample book, the data in input vector using temperature value and relative humidity measurement value, relative humidity standard value is as desired output.Sample value is imported to neural network, and input node gets 4, and hidden neuron number gets 10, and hidden neuron adopts Gaussian function, and output layer neuron adopts purelin Linear function, and the humidity value to humidity sensor in nonlinear temperature interval carries out temperature compensation.The input and output that can obtain after neural networks compensate present good linear relationship, can access the humidity value that more approaches actual value.

In order to check the effect of the blending algorithm that the present invention proposes, this algorithm and least square polynomial fit and BP neural net method are compared to analysis and research.In experiment simulation process, we find when temperature range larger samples number is more, single BP neural net method learning training speed is obviously slack-off, it is large that error of fitting also becomes, easily be absorbed in local minimum, affect compensation effect, and in the process of least square polynomial solving, when the highest power item value of polynomial expression is larger, fitting precision is relatively high, but two ends present oscillatory occurences, and along with the change of data volume is large, speed is slack-off gradually, and it is large that error also becomes gradually, and it is undesirable that fitting effect becomes.The method that the present invention proposes is owing to being that matching is carried out respectively in by stages, and compensation effect can reach satisfied precision.

Table 2 has provided the humidity match value after compensation in employing BP neural network and the p-30 ° of C ~ 0 ° C of least square polynomial method and 20 ° of C ~ 50 ° C temperature ranges, the effect temperature compensation of contrast this paper method, can draw, at temperature characteristics, present in obvious linearity and nonlinear situation, the model of temperature compensation of utilizing algorithm that least square and RBF neural network combine to set up, simple, precision is higher, and network learning and training speed is fast.

Table 2

The self-adaptation blending algorithm that the present invention utilizes RBF neural network to combine with least square, the impact that effectively compensation temperature produces humidity sensor, the outstanding advantages of the method is simple simultaneously, matching training speed is fast, compensation precision is high, can, for the temperature compensation of humidity sensor in business, greatly improve accuracy of measurement and the reliability of sensor.

Claims (1)

1. a fusion method for humidity sensor temperature compensation, the characteristic function of establishing humidity sensor is y=f (x), x ∈ [δ ₀, δ _h], wherein y is the output of sensor, and x is the input of sensor, and input reference signal is δ _h-δ ₀, it is characterized in that, set up temperature parameters t, the family curve according to humidity sensor under temperature impact, makes t0<t1<t2<t3, obtains formula (1),

$y=f(x,t)=\left\{\begin{array}{c}{f}_1(x,t),t\&Element;[t_0,t_1]\\ f_2(x,t),t\&Element;[t_1,t_2]\\ f_3(x,t),t\&Element;[t_2,t_3]\end{array}\right.---\left(1\right)$

Wherein, the characteristic function curve of humidity sensor is in input signal range ability, to be the function of continuous monotone variation; Work as t1<t<t2, by linear least square, obtain f ₂(x, t) is at the linear equation of [t1, t2] interior temperature influence; As t0<t<t1 or t2<t<t3, by RBF neural network to f ₁(x, t) and f ₃(x, t) carries out being affected and the error compensation brought by temperature;

Work as t1<t<t2, in the described linear equation of finally trying to achieve, add parametric coefficients ξ compensation temperature, according to the output after A/D conversion, directly with straight-line equation, try to achieve corresponding input signal x;

Described error compensation of being undertaken by RBF neural network, comprises the steps:

1). netinit, choose at random n training sample as cluster centre c _i(i=1,2 ..., n);

2). the training sample set of input is pressed to the grouping of arest neighbors rule, then readjust cluster centre, the c finally obtaining _icenter for network function;

3). the variance when solving network function and being Gaussian function c wherein _maxfor the ultimate range between selected center;

4). calculate the weights between hidden layer and output layer $w=\exp \left(\frac{n}{c_{\max }^2}{\left|\right|x_p-c_i\left|\right|}^2\right),(p=\mathrm{1,2},...,P;i=\mathrm{1,2},..,n);$

5). by formula calculation of performance indicators J, then carries out weights adjustment, wherein, and y _kdesirable output, for reality output, k=1,2 ... L, L, M is empirical value;

6). judge whether performance index satisfy condition for empirical value, if meet, finish training, otherwise make J=0 return to step 2, repeat above-mentioned training process, until satisfy condition.