Culture solution ion concentration measuring device

Abstract

To measure each ion concentration of a plurality of ions in a non-destructive manner and in a short time.SOLUTION: A device 1 for measuring an ion concentration of each of a plurality of ions contained in a culture solution 51, comprises: a property sensor 2 that is immersed in the culture solution, and detects an electrical conductivity, a hydrogen ion concentration index and a temperature of the culture solution; a transmission probe 3 immersed in the culture solution; a near-infrared spectrometer 4 that generates spectral data based on an output signal of the transmission probe; and a calculation unit that processes the spectral data obtained from the near-infrared spectrometer, and applies the processed spectral data, and the electrical conductivity data, hydrogen ion concentration index data and temperature data obtained from the property sensor to a calibration curve model constructed using a machine learning model, to calculate each ion concentration.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present disclosure relates to an apparatus for measuring ion concentrations in a culture medium. [Background technology]

[0002] In a plant factory, leafy vegetables such as lettuce are generally grown hydroponically in an indoor facility using an artificial light source and a culture solution. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 11-196693 Summary of the Invention [Problem to be solved by the invention]

[0004] In hydroponic plant production, it is important to control the concentration of each of the multiple ions contained in the nutrient solution. Generally, measurement of each ion concentration requires dedicated measuring equipment such as ion chromatography or inductively coupled plasma optical emission spectroscopy (ICP-OES). However, measurements using these instruments are destructive and require a relatively long time.

[0005] The present disclosure has been devised in light of the above circumstances, and its purpose is to provide an ion concentration measuring device for a culture medium that can measure the ion concentration of each of a plurality of ions non-destructively and in a short period of time. [Means for solving the problem]

[0006] According to one aspect of the present disclosure, An apparatus for measuring the concentration of each of a plurality of ions contained in a culture solution, a property sensor immersed in the culture solution for detecting the electrical conductivity, pH and temperature of the culture solution; a transmission probe immersed in the culture medium; a near-infrared spectrometer that generates spectral data based on the output signal of the transmission probe; a calculation unit that processes the spectral data obtained from the near-infrared spectrometer, and applies the processed spectral data, the electrical conductivity data, the hydrogen ion concentration exponent data, and the temperature data obtained from the property sensor to a calibration curve model constructed by a machine learning model to calculate the concentration of each ion. The present invention provides an apparatus for measuring ion concentrations in a culture medium.

[0007] Preferably, the calculation unit converting the spectral data obtained from the near-infrared spectrometer into absorbance data; resampling the absorbance-converted spectral data to make wavelength intervals constant; smoothing and differentiating the resampled spectral data; correcting the smoothed and differentiated spectral data based on the temperature detected by the property sensor; Among the corrected spectral data, a plurality of data having a predetermined wavelength highly correlated with the concentration of the ion to be measured are obtained; Converting the acquired plurality of spectral data of a predetermined wavelength into uncorrelated variables The spectral data is processed by The calculation unit generates an input data set by adding the electrical conductivity data and the hydrogen ion concentration exponent data to the processed spectral data, and applies the generated input data set to the calibration curve model to calculate the concentration of each ion. [Effects of the Invention]

[0008] According to the present disclosure, the concentration of each of a plurality of ions can be measured non-destructively and in a short time. [Brief explanation of the drawings]

[0009] [Figure 1] 1A and 1B are schematic diagrams illustrating a plant factory and an ion concentration measuring device according to the present embodiment. [Figure 2] 1 is a table conceptually showing training data and test data. [Figure 3] 10 is a flowchart showing a procedure for creating training data. [Figure 4] FIG. 10 is a diagram showing the absorbance processing. [Figure 5] FIG. 10 is a diagram illustrating a resampling process. [Figure 6] FIG. 10 is a diagram illustrating smoothing and differentiation processing. [Figure 7] FIG. 10 is a diagram illustrating a temperature correction process. [Figure 8] FIG. 10 is a diagram illustrating wavelength selection. [Figure 9] FIG. 10 is a diagram illustrating a decorrelation process. [Figure 10] FIG. 1 shows training data and test data. [Figure 11] FIG. 10 is a diagram showing the calculation results of each ion concentration. [Figure 12] FIG. 1 is a diagram showing an image of GBRT learning. [Figure 13] 1 is a graph showing the accuracy between true values ​​and measured values. [Figure 14] The results of examining the accuracy index of the measurement values ​​are shown below. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. Note that the present disclosure is not limited to the following embodiments.

[0011] [Device configuration] This embodiment relates to a device for measuring the concentrations of multiple ions contained in a nutrient solution in a plant factory. The plants (crops or agricultural products) being cultivated are, for example, leafy vegetables such as lettuce, and are hydroponically grown in the plant factory. The plant factory is installed in an indoor facility and includes an artificial light source such as an LED that irradiates the plants with light, a nutrient solution tank in which the plant roots are immersed, an air conditioning system that conditions the factory interior, and a nutrient solution control system that controls the concentrations of components in the nutrient solution. Plants absorb water by having their roots absorb the nutrient solution in the nutrient solution tank.

[0012] A plant factory and an ion concentration measuring device according to this embodiment are schematically shown in Figure 1. In the plant factory, a plurality of (two in the illustrated example) culture solution tanks 52 each containing a culture solution 51 are installed, and plants 53 such as vegetables are hydroponically cultivated in the culture solution tanks 52.

[0013] The ion concentration measuring device 1 includes a property sensor 2, a transmission probe 3, a near-infrared (NIR) spectrometer 4, and a personal computer 5. The property sensor 2 is provided in each culture solution tank 52 and immersed in the culture solution 51. The property sensor 2 is configured to detect the electrical conductivity EC, hydrogen ion concentration exponent pH, and temperature T of the culture solution 51. The multiple (two) property sensors 2 are each connected to a personal computer main body 8 of the personal computer 5.

[0014] The transmission probe 3 is also provided in each culture solution tank 52 and immersed in the culture solution 51. The transmission probe 3 is configured to measure the transmission characteristics, absorption characteristics, etc. of the culture solution 51, which is the measurement target. Light is supplied to the transmission probe 3 from a halogen light source 6, which serves as a light source. The transmission probe 3 has the function of irradiating the culture solution 51 with light and receiving the light that has passed through the culture solution 51 by reflecting it.

[0015] In this embodiment, the multiple (two) transmission probes 3 are connected to the near-infrared spectrometer 4 via a probe switching device 7. The probe switching device 7 receives a switching signal from the personal computer main body 8 of the personal computer 5, and outputs only the output signal of one transmission probe 3 corresponding to this switching signal to the near-infrared spectrometer 4. In other words, measurement of the transmission characteristics, etc. of the culture solution 51 is performed for each culture solution tank 52.

[0016] The near-infrared spectrometer 4 is configured to generate spectral data based on the output signal of the transmission probe 3. The generated spectral data is output to the personal computer main body 8.

[0017] The personal computer 5 has a personal computer main body 8 equipped with a CPU, memory, etc., a keyboard 9 and a mouse 10 for inputting data to the personal computer main body 8, and a monitor 11 for outputting data from the personal computer main body 8. The personal computer main body 8 constitutes a calculation unit 12, the keyboard 9 and the mouse 10 constitute an input unit 13, and the monitor 11 constitutes an output unit 14.

[0018] The calculation unit 12, consisting of the personal computer main body 8, is configured to process the spectral data obtained from the near-infrared spectrometer 4, and to apply the processed spectral data, electrical conductivity data (hereinafter referred to as EC data), hydrogen ion concentration exponent data (hereinafter referred to as pH data), and temperature data obtained from the property sensor 2 to a calibration curve model constructed by a machine learning model, thereby calculating the concentration of each ion in the culture solution.

[0019] Here, the calculation process during production before the ion concentration measuring device 1 is completed and the calculation process during use after completion are almost the same except for some minor differences. Therefore, for convenience, the calculation process during production will be described first, and then the calculation process during use will be described.

[0020] [Calculation process during production] Figure 2 shows a conceptual diagram of the training data used in machine learning training. i (where i is an integer greater than or equal to 2) training data are grouped into Sample1, Sample2, ... Sample i It is expressed as λ1, λ2, λ m (where m is an integer of 2 or greater) is the selected wavelength, which will be described later. One set of temperature data, one set of EC data, one set of pH data, and m sets of selected wavelength data constitutes one training data. Note that the training data shown here is merely conceptual.

[0021] 3 shows the procedure for creating one training data set. First, in step S101, the calculation unit 12 converts the spectral data obtained from the near-infrared spectrometer 4 into absorbance data. Next, in step S102, the calculation unit 12 performs resampling on the absorbance-converted spectral data to make the wavelength intervals uniform. Next, in step S103, the calculation unit 12 smooths and differentiates the resampled spectral data. Next, in step S104, the calculation unit 12 corrects the smoothed and differentiated spectral data based on the temperature T detected by the property sensor 2.

[0022] Next, in step S105, a plurality of wavelengths having a high correlation with the concentration of the ion to be measured are selected and acquired from the corrected spectral data. The wavelengths selected here are the aforementioned selected wavelengths λ1, λ2, . . . λ. m During production, the developer selects and determines the optimal wavelength for each training data. After training using multiple training data, the selected wavelength is finally determined, and the selected wavelength is stored in the calculation unit 12, and the selected wavelength becomes a predetermined value. During use, calculations are performed using the data on the selected wavelength of this predetermined value.

[0023] Next, in step S106, the calculation unit 12 converts the multiple spectral data for the selected wavelengths into uncorrelated variables. However, this decorrelation process may be performed only under specific conditions. For example, as will be described in detail later, it may be performed only when the calibration curve model is multiple regression or linear support vector machine (SVM) and the correlation between variables is high (depending on the ion species). In other words, the decorrelation process may be omitted except under specific conditions.

[0024] The spectral data processing is completed through steps S101 to S106. During production, parameters are adjusted or adapted in the smoothing and differentiation process of step S103, the temperature correction process of step S104, and the decorrelation process of step S106, just as with the selected wavelength. The finally determined parameter values ​​are stored in the calculation unit 12 and become predetermined values. During use, calculations are performed using these predetermined parameter values.

[0025] Next, in step S107, the calculation unit 12 adds the EC data and pH data to the processed spectral data to create training data, which will be input data for the calibration curve model in a later step.

[0026] Each step will be explained below. Figure 4 shows the absorbance process in step S101. Here, the process removes the effects of light source fluctuations from the sample spectrum. (A) shows the intensity of the light source I0, (B) shows the intensity of the sample spectrum I, and (C) shows the absorbance Abs value versus wavelength. The absorbance is calculated using the following equation (1):

[0027]

number

[0028] 5 shows the resampling process in step S102. Before the process shown in (A), the wavelength intervals between measurement points of the absorbance-converted spectral data are not constant. Therefore, as shown in (B) after the process, this is adjusted to a constant (2.000) across the entire wavelength range (900 to 1700 nm in this embodiment). The resampling process is performed using cubic spline interpolation.

[0029] Figure 6 shows the smoothing and differentiation process in step S103. (A) shows the state before processing, and (B) shows the state after processing. Here, the effects of noise originating from the power supply and other sources are reduced, smoothed, and differentiation is performed to extract the change portion. A Savitzky-Golay filter is used for this smoothing and differentiation process. The parameters of this filter are adjusted and adapted during production, and are set to fixed, predetermined values ​​during use.

[0030] Figure 7 shows the temperature correction process in step S104. (A) shows the spectrum data of the same sample acquired at different temperatures (20°C, 25°C, and 30°C) before processing. (B) shows the spectrum data after processing, where the effects of these temperatures have been removed.

[0031] Near-infrared (NIR) spectra are strongly affected by temperature, so correction is performed here to remove this temperature effect, improving measurement accuracy. The corrected intensity is calculated using the following equation (2): Corrected intensity = Uncorrected intensity - Temperature × Correction coefficient (2)

[0032] The correction coefficient is the slope of a linear regression analysis of the temperature and intensity for each wavelength, and is adjusted and adapted during fabrication and becomes a fixed, predetermined value during use.

[0033] 8 shows the wavelength selection process in step S105. Here, a plurality of wavelengths that have a high correlation with the concentration of the ion to be measured are selected from the spectrum data after temperature correction processing.

[0034] (A) shows the relationship between the wavelength of the spectral data and the correlation coefficient CC. The larger the absolute value of the correlation coefficient CC, the stronger the correlation between the wavelength and the concentration of the ion being measured. Therefore, here, wavelengths where the absolute value of the correlation coefficient CC is equal to or greater than a threshold value are selected. The wavelengths of each plot shown in (A) are the selected wavelengths.

[0035] For example, if we focus on the plot surrounded by circle a in (A) (wavelength = approximately 1230 nm, correlation coefficient CC = approximately -0.6), the relationship between ion concentration and intensity at the wavelength of this plot is as shown in (B), and it can be seen that there is a generally negative correlation.

[0036] During fabrication, trial and error was performed using a large amount of training data, and as a result, optimal multiple selection wavelengths (λ1, λ2, λ m These selected wavelength values ​​are then stored in the calculation unit 12, and when the device is subsequently used, calculations are performed based on the stored selected wavelength values.

[0037] FIG. 9 shows the decorrelation process in step S106. Here, in order to prevent overlearning of the learning model, intensity information with high correlation between wavelengths is converted into decorrelation variables. (A) shows the training data (Sample 1, Sample 2, ... Sample i ) for each selected wavelength (λ1, λ2, . . . λ m ) is shown. The wavelength intensities of (A) are expressed as uncorrelated variables (S1, S2, ...S n , where n=m), the result of the conversion is shown in (B). The conversion formula for converting the wavelength intensity into a non-correlated variable is expressed as the following formula (3).

[0038]

number

[0039] The second term on the left side is a correction coefficient for each selected wavelength, which is obtained by principal component analysis of the training data. During production, this correction coefficient is adjusted and adapted, and the finally determined value of the correction coefficient is stored in the calculation unit 12. Then, during subsequent use, calculations are performed based on the stored value of the correction coefficient.

[0040] FIG. 10 shows the training data created in step S107. EC data and pH data are added to the decorrelation-processed data of FIG. 9(B) to create training data as shown in FIG. 10. This training data differs from the training data of FIG. 2 in that it has been decorrelated, and it is the training data of FIG. 10 that is actually used. This training data will be input data for the calibration curve model in a later step. Therefore, the training data shown in FIG. 10 is collectively referred to as an input data set. By creating training data and an input data set by adding EC data and pH data to the spectral data, measurement accuracy can be improved.

[0041] Figure 11 shows the calculation results (output) when the input data set is applied (input) to the calibration curve model and the concentration of each ion is calculated. Here, nitrate ion (NO3), potassium ion (K), phosphate ion (PO4), and iron ion (Fe) are shown as examples of ions.

[0042] The calibration curve model is constructed by a machine learning model. The machine learning model is, for example, multiple regression, elastic net multiple regression (or simply elastic net), partial least squares regression (PLS), gradient boosting regression tree (GBRT), or linear support vector machine (linear SVM). In this embodiment, R (4.1.1) is used as the processing language, and MLmetrics (1.1.1), xgboost (1.5.0.1), data.table (1.14.2), Matrix (1.3-4), glmnet (4.3-3), kernlab (0.9-29), and pls (2.8-0) are used as packages.

[0043] Figure 12 shows an image of GBRT learning. By repeating learning using each training data in the input dataset, the error between the true value and the calculated value (or estimated value) of the target ion concentration gradually decreases, and the calculated value approaches the true value. Therefore, by performing learning a sufficient number of times, it is possible to construct a highly accurate machine learning model, i.e., a calibration curve model, that can obtain calculated values ​​close to the true value.

[0044] For example, if the machine learning model is GBRT, adjustment and adaptation of the number of iterative learning iterations and the depth of the decision tree, which are hyperparameters, are mainly performed. Once the values ​​of the hyperparameters are finally determined, the values ​​are stored in the calculation unit 12 and become predetermined values. The calibration curve model is then essentially completed. Thereafter, when the model is used, calculations are performed based on the stored values.

[0045] The simplest calibration curve is represented by a two-dimensional graph, but the calibration curve of this embodiment has the property of defining a large number of values ​​for a large number of values, making it difficult to illustrate.

[0046] In the above-described manufacturing process of the ion concentration measuring device 1, the adjustment, adaptation or development of each value is carried out based on the true value separately measured using ion chromatography or ICP.

[0047] [Calculation process when using] Next, the calculation process when the ion concentration measuring device 1 is used after completion will be described.

[0048] The calculation process during use is almost the same as the calculation process during manufacture described above, with the exception of a few minor differences. Roughly speaking, the only difference is that the aforementioned "training" is replaced with "testing." "Testing" here refers to testing when an actual measurement is performed using the completed ion concentration measuring device 1. For practical purposes, "testing" can be interpreted as "measurement."

[0049] 2 represents test data when in use, that is, test data actually obtained by the ion concentration measuring device 1 when measuring the concentrations of each ion.

[0050] 3 also shows the procedure for creating one piece of test data. Here, in step S105, a wavelength has already been selected and stored in the calculation unit 12, so the wavelength selection step is omitted. Instead, in step S105, a plurality of pieces of spectral data corrected in step S104, each of which corresponds to a predetermined wavelength, i.e., the selected wavelength, are acquired. In the next step S106, the acquired plurality of pieces of spectral data for the selected wavelength are subjected to decorrelation processing.

[0051] Finally, in step S107, test data such as that shown in Fig. 10 is generated. This test data is compiled into an input data set. When this input data set is applied to the calibration curve model to calculate the concentration of each ion, the calculation results shown in Fig. 11 are obtained.

[0052] [Accuracy verification] Next, the accuracy of each ion concentration measured by the ion concentration measuring device 1 was verified, and the results will be described.

[0053] FIG. 13 is a graph showing the accuracy between the true value (horizontal axis) and the measured value (vertical axis) measured by the present device 1 for the nitrate ion concentration.

[0054] Line a is the 100% line where the two are perfectly consistent. White circles indicate samples when training data is used, and black squares indicate samples when test data is used. As shown in the figure, the samples are generally clustered around the 100% line a, indicating that the device 1 of this embodiment has relatively high accuracy. The number of samples is 72 for training data and 14 for test data. The evaluation index value, which is the coefficient of determination (R2), is 0.9963 for training data and 0.9652 for test data. The mean absolute percentage error (MAPE) is 0.0193 for training data and 0.0753 for test data.

[0055] Figure 14 shows the results of examining the values ​​of several accuracy indices for the measured values ​​of several ion concentrations in several test data sets. Nitrate ions, potassium ions, phosphate ions, magnesium ions, and iron ions are shown as examples of ions to be measured. The evaluation indices shown are the coefficient of determination (R2) and mean absolute percentage error (MAPE). The machine learning models shown are multiple regression, elastic net, PLS, GBRT, and linear SVM. As can be seen from the figure, accuracy is generally good.

[0056] As described above, according to the ion concentration measuring device 1 of this embodiment, the concentration of each ion in the culture solution 51 is calculated using a calibration curve model based on the output signals from the property sensor 2 and the transmission probe 3, so that the concentration of each ion can be measured non-destructively and in a short time. Furthermore, the spectral data obtained from the near-infrared spectrometer 4 is applied to the calibration curve model after being processed using steps S101 to S106 in Fig. 3, so that the measurement accuracy can be improved.

[0057] Although the embodiments of the present disclosure have been described in detail above, various other embodiments and modifications of the present disclosure are conceivable. For example, the ions to be measured may be other ions, such as ammonium ions, calcium ions, or manganese ions.

[0058] The embodiments of the present disclosure are not limited to the above-described embodiments, and all modifications, applications, and equivalents encompassed within the spirit of the present disclosure as defined by the claims are included in the present disclosure. Therefore, the present disclosure should not be interpreted as being limited, and can be applied to any other technology that falls within the spirit of the present disclosure. [Explanation of symbols]

[0059] 1. Ion concentration measuring device 2. Property sensor 3 Transmission probe 4 Near-infrared spectrometer 12 Arithmetic section 51 Culture solution

Claims

1. An apparatus for measuring the concentration of each of a plurality of ions contained in a culture solution, a property sensor immersed in the culture solution for detecting the electrical conductivity, pH and temperature of the culture solution; a transmission probe immersed in the culture medium; a near-infrared spectrometer that generates spectral data based on the output signal of the transmission probe; a calculation unit that processes the spectral data obtained from the near-infrared spectrometer using the temperature detected by the property sensor, and applies the processed spectral data, the electrical conductivity data and the hydrogen ion concentration exponent data obtained from the property sensor to a calibration curve model constructed by a machine learning model to calculate the concentration of each ion. An apparatus for measuring ion concentration in a culture solution.

2. The calculation unit converting the spectral data obtained from the near-infrared spectrometer into absorbance data; resampling the absorbance-converted spectral data to make wavelength intervals constant; smoothing and differentiating the resampled spectral data; correcting the smoothed and differentiated spectral data based on the temperature detected by the property sensor; Among the corrected spectral data, a plurality of data having a predetermined wavelength highly correlated with the concentration of the ion to be measured are obtained; Converting the acquired plurality of spectral data of a predetermined wavelength into uncorrelated variables The spectral data is processed by The calculation unit generates an input data set by adding the electrical conductivity data and the hydrogen ion concentration exponent data to the processed spectral data, and applies the generated input data set to the calibration curve model to calculate each ion concentration. The device for measuring ion concentrations in a culture solution according to claim 1.

Images