Sensor device

Abstract

To provide a sensor device capable of enhancing the estimation accuracy for an unknown signal.SOLUTION: A sensor device comprises a plurality of sensors and an online estimator. Each of the plurality of sensors are so configured to receive input of an input signal which includes unknown input and known input. The online estimator comprises an arithmetic unit which performs first arithmetic operation to sequentially update a response coefficient of the plurality of sensors relative to each of the unknown input, and second arithmetic operation to obtain a linear sum of estimated values of the unknown input estimated by each of the plurality of sensors. The sensor device uses the updated response coefficient to update the estimated value of the unknown input.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a sensor device. [Background technology]

[0002] As equipment becomes more automated, the importance of sensors that measure the environment around the equipment is increasing. For example, sensor fusion has been attracting attention in recent years. Sensor fusion is a technology that integrates information obtained from multiple sensors and performs composite analysis to process information according to the purpose.

[0003] For example, Patent Document 1 discloses a composite sensor including an acceleration sensor and an angular velocity sensor. The composite sensor described in Patent Document 1 corrects the output signal of the angular velocity sensor based on the output signals of multiple acceleration sensors, thereby improving the accuracy of estimating the angular velocity.

[0004] Also, for example, Patent Document 2 discloses a liquid amount measurement system. The liquid amount measurement system described in Patent Document 2 evaluates parameters of the liquid in a container using the likelihood of a multiple model configured with a plurality of sensors. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] International Publication No. 2020 / 158485 [Patent Document 2] WO 02 / 35185 Summary of the Invention [Problem to be solved by the invention]

[0006] The systems of Patent Documents 1 and 2 aim to improve the measurement accuracy of an input signal to be measured. However, when multiple signals are superimposed on the input signal to be measured, it is difficult to accurately isolate and estimate only the necessary information from the input signal including these multiple signals.

[0007] The present disclosure has been made in consideration of the above circumstances, and aims to provide a sensor device that can extract necessary information from an input signal containing multiple pieces of information, and improve the accuracy of estimating the necessary information. [Means for solving the problem]

[0008] In order to solve the above problems, the present disclosure provides the following means.

[0009] (1) A sensor device according to a first aspect includes a plurality of sensors and an online estimator. Each of the plurality of sensors is configured to receive an input signal including an unknown input and a known input. The online estimator has a calculation unit that performs a first calculation for sequentially updating a response coefficient of each of the plurality of sensors with respect to the unknown input, and a second calculation for obtaining a linear sum of an estimated value of the unknown input estimated by each of the plurality of sensors. The sensor device updates the estimated value of the unknown input using the updated response coefficient.

[0010] (2) In the sensor device according to the above aspect, each sensor constituting the plurality of sensors may be connected to the other sensors, and the input signal may interact between the plurality of sensors.

[0011] (3) The sensor device according to the above aspect may further include a feature conversion unit, wherein weights are applied to the input signals input to the plurality of sensors by the feature conversion unit.

[0012] (4) The sensor device according to the above aspect may further include a feature conversion unit, wherein input signals input to the plurality of sensors may be nonlinearly converted by a basis function in the feature conversion unit.

[0013] (5) In the sensor device according to the above aspect, the first calculation may be performed according to a distributed Kalman filter.

[0014] (6) In the sensor device according to the above aspect, in the first calculation, updated response coefficients are calculated for each of the plurality of sensors, and the updated response coefficients may be calculated by multiplying a difference between an observed value of the input signal and a product of the response coefficient before the update and an estimated value of the unknown input before the update by a coefficient, and adding the result to the response coefficient before the update.

[0015] (7) In the sensor device according to the above aspect, the estimate of the unknown input may be updated in each of the plurality of sensors, and the updated estimate of the unknown input may be obtained by removing the component of the known input from the observed value of the input signal and dividing the result by the updated response coefficient.

[0016] (8) In the sensor device according to the above aspect, in the second calculation, a weight may be applied to each of the estimated values ​​of the unknown input.

[0017] (9) The sensor device according to the above aspect may further include a known-input measurement sensor that independently measures the known input. Effect of the Invention

[0018] A sensor device according to the present disclosure can extract necessary information from an input signal that includes a plurality of pieces of information, and can improve the accuracy of estimating the necessary information. [Brief description of the drawings]

[0019] [Figure 1] FIG. 1 is a schematic diagram of a sensor device according to a first embodiment. [Diagram 2]FIG. 13 is a diagram illustrating the relationship between observed values, unknown inputs, and known inputs. [Diagram 3] FIG. 2 shows a flow diagram of a data processing method of the sensor device according to the embodiment. [Figure 4] 11 shows a simulation result of the sensor device according to the present embodiment. [Diagram 5] FIG. 13 is a schematic diagram of a sensor device according to a first modified example. [Figure 6] FIG. 13 is a schematic diagram of a sensor device according to a second modified example. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0020] The present embodiment will be described in detail below with reference to the drawings as appropriate. The drawings used in the following description may show characteristic parts in an enlarged scale for the sake of convenience in order to make the characteristics easier to understand, and the dimensional ratios of each component may differ from the actual ones. The materials, dimensions, etc. exemplified in the following description are merely examples, and the present invention is not limited thereto. They may be modified as appropriate within the scope of the effects of the present invention.

[0021] "First embodiment" 1 is a schematic diagram of a sensor device 100 according to the first embodiment. The sensor device 100 includes a plurality of sensors 10, an online estimator 20, and a known input measurement sensor 30.

[0022] The multiple sensors 10 may be any sensors. The sensor 10 may be, for example, a CO2 sensor. Each of the multiple sensors 10 receives the same input signal S in is input. In Fig. 1, only three of the n sensors 10 (n is a natural number) are illustrated. The number of sensors 10 is arbitrary.

[0023] Input signal S inincludes unknown input and known input. The unknown input is a signal representing information that is the subject of sensing in the sensor 10, and is an unknown input value from which information cannot be obtained except by the sensor 10. The known input is a signal representing information that is not the subject of sensing in the sensor 10, and is an input value from which information can be obtained independently by another different sensor (e.g., known input measurement sensor 30). For example, the unknown input is CO2 concentration, and the known input is temperature. The known input does not need to be one type of physical quantity, and may include multiple physical quantities such as temperature, humidity, and air pressure.

[0024] The known input measurement sensor 30 is an independent sensor for measuring a known input. The known input measurement sensor 30 measures a parameter that varies due to the environment in which the multiple sensors 10 are installed. The known input measurement sensor 30 shown in FIG. 1 is a temperature sensor that measures temperature T(t). The known input measurement sensor 30 may be a sensor that measures a physical quantity other than temperature, or may be multiple sensors that measure multiple physical quantities. The known input measurement sensor 30 may be located inside the sensor device 100 as shown in FIG. 1, or may be located outside the sensor device 100.

[0025] Input signal S in is observed as an observation value s(t) at each sensor 10. For example, the sensor 10 receives an input signal S in is converted to a voltage value and observed. The observed value s(t) is not limited to a voltage, and may be any physical quantity. Each observed value s(t) has an unknown input u(t) and a known input T(t). For ease of understanding, the following description will be given taking an example in which the known input is only temperature, but the known input may be something other than temperature, or may include multiple physical quantities.

[0026] FIG. 2 is a diagram showing the relationship between the observed value s(t), the unknown input u(t), and the known input T(t). In FIG. 2, the top graph shows the time change of the standard value of the strength of the observed value s(t), the center graph shows the time change of the standard value of the strength of the known input T(t), and the bottom graph shows the time change of the standard value of the strength of the unknown input u(t). The observed value s(t) is a superposition of the unknown input u(t) and the known input T(t). What is actually observed by the sensor 10 is the observed value s(t), and it is difficult to sequentially separate and measure the unknown input u(t) and the known input T(t). As shown in FIG. 2, it is often difficult to read the behavior of the unknown input u(t) from the observed value s(t). On the other hand, what we actually want to measure is the unknown input u(t). Therefore, it is necessary to estimate the unknown input u(t) from the observed value s(t).

[0027] Each sensor 10 has a response coefficient α. The response coefficient α is a parameter corresponding to the sensitivity of the sensor 10. Even when the sensors 10 are manufactured under the same conditions, the response coefficient α may be different, and the response coefficient α is a value unique to each sensor 10. For example, the response coefficient of the (i-1)th sensor 10 is α i-1 The response coefficient of the i-th sensor 10 is α i The response coefficient of the (i+1)th sensor 10 is α i+1 When expressing the response coefficient of any one of the multiple sensors 10, the response coefficient may be represented by α without a subscript.

[0028] Each sensor 10 obtains an estimated value u^(t) of the unknown input using, for example, the observed value s(t) and a response coefficient α. As described above, the response coefficient α differs for each sensor 10. Therefore, the estimated value u^(t) of the unknown input obtained by each sensor 10 also differs for each sensor 10. For example, the estimated value of the unknown input obtained by the (i-1)th sensor 10 is u^ i-1 In (t), the estimated unknown input obtained by the i-th sensor 10 is u^ i At (t), the estimated unknown input obtained from the i+1th sensor 10 is u^ i+1(t). When expressing an estimate of an unknown input obtained by any one of the multiple sensors 10, the estimate of the unknown input may be expressed as u^(t) without a subscript.

[0029] The estimated value u^(t) of the unknown input is obtained, for example, by a calculation unit provided in each sensor 10. The calculation unit is, for example, a microcontroller or a processor. The calculation unit has, for example, a CPU that performs calculation processing, a ROM that stores programs, and a RAM that is a working memory when the CPU performs calculation processing. Here, an example is shown in which the sensor 10 has the calculation unit, but the calculation unit may be external to the sensor 10. For example, this calculation may be performed by the calculation unit of the online estimator 20, or by a separately provided calculator.

[0030] Here, a method for determining the estimated value u^(t) of the unknown input using the observed value s(t) and the response coefficient α will be described.

[0031] The observed value s(t) can be expressed as a function of the unknown input u(t) and the known input T(t), as s(t) = F(u(t), T(t)). Assuming that the response components from second order onwards in the observed value s(t) are sufficiently small, the observed value is described by the partial differential coefficient when this function is first differentiated. The function is expressed by the partial differential coefficient in the following equation (1).

[0032]

number

[0033] Here, if dF / du is α and dF / dT is β, this is expressed by the forward model shown in the following equation (2).

[0034]

number

[0035] In equation (2), α is the response coefficient of the sensor 10. Since the sensitivity of the sensor 10 depends on temperature, the response coefficient α also depends on temperature.

[0036] β is a parameter that is found by calibrating the sensor 10 in an environment where the unknown input u(t) is constant. For example, β can be found by placing the sensor 10 in an environment where the CO2 concentration is constant and measuring the signal (e.g., voltage value) output from the sensor 10 while changing the temperature. In other words, β is a constant.

[0037] The estimated value u^(t) of the unknown input can be obtained as an inverse model expressed by the following equation (3) by transforming equation (2).

[0038]

number

[0039] In equation (3), s(t) is an observed value, β is a constant, and T(t) is a value obtained by the known input measurement sensor 30. Once the response coefficient α is determined, the estimated value u^(t) of the unknown input can be obtained from equation (3). The response coefficient α is a value that is successively updated by the online estimator 20 described later, and is successively obtained in response to changes in the surrounding environment (e.g., temperature). The estimated value u^(t) of the unknown input obtained by each sensor 10 is sent to the online estimator 20. It is assumed that the response coefficient β is obtained in advance by prior measurement.

[0040] The online estimator 20 has a calculation unit that performs the first calculation and the second calculation. The calculation unit is, for example, a microcontroller or a processor. The calculation unit has, for example, a CPU that performs the calculation process, a ROM that stores a program, and a RAM that is a working memory when the CPU performs the calculation process.

[0041] The first calculation is a calculation for sequentially updating the response coefficient α for the unknown input in each of the multiple sensors 10. The second calculation is a calculation for obtaining a linear sum of the estimated values ​​u^(t) of the unknown input estimated in each of the multiple sensors 10.

[0042] First, the first calculation will be explained. The response coefficient α changes depending on the surrounding environment such as temperature. Therefore, the response coefficient α needs to be updated sequentially. The first calculation is performed, for example, according to a distributed Kalman filter. For information on the distributed Kalman filter, please refer to Takaba Joji, "Measurement and Control", Vol. 56, No. 12, December 2017 issue.

[0043] An example of a specific procedure of the first calculation will be described. In the distributed Kalman filter, the observation value vector y t =Cx t +v t The observation vector y t corresponds to a set of observation values ​​s(t) observed by each sensor 10. Also, the state vector x t corresponds to a set of response coefficients α to be updated for each sensor 10, (α1, ...α i-1 , α i , α i+1 …α n The observation matrix C corresponds to a set of estimates u^(t) of the unknown inputs at each sensor 10, denoted as (u^(t)1, ...u^(t) i-1 , u^(t) i , u^(t) i+1 …u^(t) n ) is expressed as the observation noise v t represents Gaussian noise. Furthermore, βT in equation (2) is the offset of the observation vector, which is obtained from the known input. Even if a known offset is added to the observation vector, estimation using the distributed Kalman filter is possible without loss of generality.

[0044] In a distributed Kalman filter, the state vector x t can be updated using the following formula (4): x^ t|t is the updated state vector, and x^t|t-1 is the state vector before the update, and K t is the Kalman gain. The Kalman gain is a coefficient expressed by the following equation (5). P in equation (5) t|t-1 is the covariance matrix. R is 1 in this case.

[0045]

number

[0046]

number

[0047] In this embodiment, the state vector to be estimated is not the unknown input itself to be estimated, but the response coefficient of the unknown input with respect to the observed value. In this way, the unknown input can be estimated. If the state vector to be estimated is the unknown input itself, it is difficult to estimate it well. This is because it is necessary to accurately model the dynamics that the unknown input follows, and in reality, this is not always possible. Therefore, the above formula (4) is replaced with values ​​according to this embodiment and expressed as the following formula (6). In formula (6), α' is the response coefficient after the update, α is the response coefficient before the update, u^(t) is the estimated value of the unknown input before the update, and K t is the Kalman gain. The Kalman gain is calculated by multiplying the covariance matrix by the estimate of the unknown input before the update.

[0048]

number

[0049] By using equation (6), the updated response coefficient α' can be obtained for each of the multiple sensors. As shown in equation (6), the updated response coefficient α' is calculated by multiplying the difference between the observed value s(t) of the input signal and the product of the response coefficient α before the update and the estimated value u^(t) of the unknown input before the update by a coefficient (Kalman gain K t) and add this value to the response coefficient α before the update.

[0050] The value of the response coefficient α′ updated in the first calculation of the online estimator 20 is transmitted to each sensor 10, and the response coefficient is updated in each sensor 10. For example, the response coefficient of the (i-1)th sensor 10 is α i-1 From α' i-1 The response coefficient of the i-th sensor 10 is updated to α i From α' i The response coefficient of the (i+1)th sensor 10 is updated to α i+1 From α' i+1 When expressing the updated response coefficient of any one of the multiple sensors 10, the updated response coefficient may be denoted by α′ without a subscript.

[0051] Next, the second calculation of the online estimator 20 will be described. In the second calculation, a linear sum of the estimated value u^(t) of the unknown input estimated by each of the multiple sensors is calculated. The second calculation is performed every time the estimated value u^(t) of the unknown input is updated. The update of the estimated value u^(t) of the unknown input will be described later.

[0052] When performing the second calculation, a weight w may be applied to the estimated value u^(t) of the unknown input. A different weight w is applied to each of the estimated values ​​u^(t) of the unknown input obtained by each sensor 10. For example, the estimated value u^(t) of the unknown input obtained by the (i-1)th sensor 10 may be i-1 (t) has weight i-1 is applied, and the estimated value u^ of the unknown input obtained by the i-th sensor 10 is i (t) has weight i is applied, and the estimated value u^ of the unknown input obtained by the i+1th sensor 10 is i+1 (t) has weight i+1 is applied. When expressing one of a plurality of weights, a subscript may be omitted and it may be expressed as w.

[0053] The weight w is applied so as to minimize the posterior variance of the estimated value output from the sensor device 100. That is, a large weight w is applied to a signal of high importance, and a small weight w is applied to a signal of low importance. The value of the weight w may be updated at any time.

[0054] By performing the second calculation, a linear sum of the estimated values ​​u^(t) of the unknown inputs obtained by the respective sensors 10 is output as an estimated value from the sensor device 100. This estimated value is the input signal S in The results show a high degree of agreement with the results obtained when only unknown input is extracted from the

[0055] Here, the estimated value from the sensor device 100 is sequentially updated. This is because the estimated value u^(t) of the unknown input is sequentially updated. The estimated value u^(t) of the unknown input is updated using the updated response coefficient α'. This calculation may be performed by the calculation unit of the sensor 10, by the calculation unit of the online estimator 20, or by another calculation unit in the sensor device 100.

[0056] The updated estimate of the unknown input u^(t) can be found by substituting the updated response coefficient α' into equation (3). As described above, the response coefficient α before the update becomes the updated response coefficient α' by performing the first calculation. In equation (3), s(t) is the observed value, β is a constant, and T(t) is a value obtained by the known input measurement sensor 30. The updated estimate of the unknown input u^(t) can be found by substituting the response coefficient α' into equation (3). The updated estimate of the unknown input u^(t) can be found by subtracting the known input component βT(t) from the observed value s(t) of the input signal and dividing it by the updated response coefficient α'.

[0057] As described above, the sensor device 100 according to this embodiment sequentially updates the estimated value output from the sensor device 100 by repeating the update of the response coefficient α and the update of the estimated value u^(t) of the unknown input.

[0058] That is, the data processing method using the sensor device 100 of this embodiment includes a step of updating the response coefficients α of multiple sensors 10, a step of updating the estimated value u^(t) of the unknown input using the updated response coefficients α', and a step of calculating the linear sum of the updated estimated value u^(t) of the unknown input.

[0059] FIG. 3 shows a flow diagram of a data processing method of the sensor device 100 according to this embodiment. First, in a preparation step S0, an equation showing the relationship between the observed value s(t), the unknown input u(t), and the known input T(t) is derived. The preparation step S0 has a forward model setting step S01 and an inverse model derivation step S02. In the forward model setting step S01, a forward model represented by equation (2) is set. The forward model is found by differentiating a function representing the observed value s(t) once. In the inverse model derivation step S02, an inverse model represented by equation (3) is derived. The inverse model is found by algebraically transforming the forward model.

[0060] Next, in the response coefficient updating step S1, the response coefficient α is updated. The response coefficient updating step S1 is performed according to the first calculation described above. The response coefficient updating step S1 is performed using a distributed Kalman filter.

[0061] Next, in the estimated value updating step S2, the estimated value u^(t) of the unknown input is updated by substituting the updated response coefficient α' into the inverse model expressed by equation (3).

[0062] Next, in the observation matrix setting step S3, the observation matrix C is reset. The observation matrix C corresponds to a set of estimated values ​​u^(t) of unknown inputs at each sensor 10. In the observation matrix setting step S3, each of the estimated values ​​constituting the observation matrix C is changed to a value after update in the estimated value updating step S2. As shown in equations (4) and (6), the observation matrix C is necessary to update the response coefficient α.

[0063] The response coefficient updating step S1, the estimated value updating step S2, and the observation matrix setting step S3 are repeated.

[0064] In the linear sum deriving step S4, a linear sum of the estimated value u^(t) of the unknown input is obtained. The linear sum deriving step S4 is performed according to the second calculation described above. The linear sum deriving step S4 is performed every time the estimated value u^(t) of the unknown input is updated in the estimated value updating step S2.

[0065] The sensor device 100 according to this embodiment treats the response coefficient α of each sensor 10 as a state variable, and sequentially estimates the response coefficient α. If the target to be obtained is an estimated value output from the sensor device 100, this estimated value is generally treated as a state variable, but the sensor device 100 according to this embodiment indirectly obtains an estimated value u^(t) of the unknown input by using the state variable as the response coefficient α. By performing such indirect estimation, the input signal S in As a result, the sensor device 100 according to the present embodiment can improve the estimation accuracy of estimating an unknown input from an input signal including an unknown input and a known input.

[0066] FIG. 4 shows simulation results of the sensor device 100 according to this embodiment. The top graph in FIG. 4 shows signal changes corresponding to the observed values ​​s(t) at each sensor 10. The second graph from the top in FIG. 4 shows signal changes corresponding to temperature changes around the sensor 10, corresponding to fluctuations in the known input. The third graph from the top in FIG. 4 shows signal changes corresponding to an unknown input, assuming a periodically fluctuating signal. The bottom graph in FIG. 4 corresponds to the estimated values ​​u^(t) of the unknown input output from each sensor 10.

[0067] 4, when the sensor device 100 according to the present embodiment is used, the estimated value u^(t) of the unknown input converges toward a predetermined value. In other words, when the sensor device 100 according to the present embodiment is used, it can be said that the unknown input can be appropriately estimated from the observed value s(t) showing complex fluctuations.

[0068] Up to this point, one embodiment of the present disclosure has been described using the sensor device 100 shown in FIG. 1, but the present disclosure is not limited to this example and various modifications are possible.

[0069] For example, Fig. 5 is a schematic diagram of a sensor device 101 according to a first modification. In the sensor device 101 shown in Fig. 5, each of the sensors 10 constituting a plurality of sensors 10 is connected to the other sensors 10. An input signal S in The input signal S applied to the reservoir is a signal that is transmitted to the other sensors 10. ... in changes as it propagates through the reservoir, adding information about the response coefficient α. The reservoir projects the input signal into a multidimensional nonlinear space.

[0070] Input signals S between multiple sensors 10 in When interacting, the input signal S in The feature amount is extracted, and the estimation accuracy of the estimated value output from the sensor device 101 is improved.

[0071] 6 is a schematic diagram of a sensor device 102 according to a second modification. The sensor device 102 differs from the sensor device 100 in that it further includes a feature conversion unit 40. The feature conversion unit 40 converts an input signal S in Each of the sensors 10 receives an input signal vS in is applied.

[0072] The weight v is changed, for example, by learning. When the weight v is optimized by learning, a large input signal can be applied to a sensor that contributes highly to the estimation of the unknown input. In other words, the input signal can be given characteristics. in By applying the feature amount to the estimated value output from the sensor device 102, the estimation accuracy is improved.

[0073] The feature conversion unit 40 converts the input signal S in In the application of pattern recognition and machine learning, some preprocessing is often performed on the original data to extract features. in By performing nonlinear conversion, the estimation accuracy of the estimated value output from the sensor device 102 is improved. [Explanation of symbols]

[0074] 10 Sensors 20 Online Estimator 30 Known Input Measurement Sensor 40 Feature conversion section 100, 101, 102 Sensor device s(t) Observation S in Input Signal u(t) unknown input u^(t) Estimate of unknown input T(t) known input α, α' response coefficients

Claims

1. a plurality of sensors; and an online estimator; Each of the plurality of sensors is configured to receive an input signal including an unknown input and a known input; The online estimator a first operation for sequentially updating a response coefficient for each of the plurality of sensors to the unknown input; a second calculation for obtaining a linear sum of the estimated values ​​of the unknown inputs estimated by each of the plurality of sensors; The sensor device updates the estimates of the unknown inputs at each sensor using the updated response coefficients.

2. Each of the sensors constituting the plurality of sensors is connected to other sensors; The sensor device of claim 1 , wherein the input signal interacts between the plurality of sensors.

3. Further comprising a feature conversion unit, The sensor device according to claim 1 , wherein the input signals input to the plurality of sensors are weighted by the feature conversion unit.

4. Further comprising a feature conversion unit, The sensor device according to claim 1 , wherein the input signals input to the plurality of sensors are nonlinearly transformed by the feature transformation unit using basis functions.

5. The sensor device of claim 1 , wherein the first calculation is performed according to a distributed Kalman filter.

6. In the first calculation, an updated response coefficient is obtained for each of the plurality of sensors; 2. The sensor device according to claim 1, wherein the response coefficient after the update is obtained by multiplying a difference between an observed value of the input signal and a product of the response coefficient before the update and an estimated value of the unknown input before the update by a coefficient and adding the result to the response coefficient before the update.

7. an estimate of the unknown input is updated at each of the plurality of sensors; The sensor device according to claim 1 , wherein the updated estimated value of the unknown input is obtained by removing a component of the known input from the observed value of the input signal and dividing the result by the updated response coefficient.

8. The sensor device of claim 1 , wherein in the second operation, a weight is applied to each of the estimates of the unknown input.

9. Further comprising a known input measurement sensor; The sensor device of claim 1 , wherein the known-input measurement sensor independently measures the known input.

Images